MyArxiv
Computation and Language 133
☆ SRUM: Fine-Grained Self-Rewarding for Unified Multimodal Models
Recently, remarkable progress has been made in Unified Multimodal Models (UMMs), which integrate vision-language generation and understanding capabilities within a single framework. However, a significant gap exists where a model's strong visual understanding often fails to transfer to its visual generation. A model might correctly understand an image based on user instructions, yet be unable to generate a faithful image from text prompts. This phenomenon directly raises a compelling question: Can a model achieve self-improvement by using its understanding module to reward its generation module? To bridge this gap and achieve self-improvement, we introduce SRUM, a self-rewarding post-training framework that can be directly applied to existing UMMs of various designs. SRUM creates a feedback loop where the model's own understanding module acts as an internal ``evaluator'', providing corrective signals to improve its generation module, without requiring additional human-labeled data. To ensure this feedback is comprehensive, we designed a global-local dual reward system. To tackle the inherent structural complexity of images, this system offers multi-scale guidance: a \textbf{global reward} ensures the correctness of the overall visual semantics and layout, while a \textbf{local reward} refines fine-grained, object-level fidelity. SRUM leads to powerful capabilities and shows strong generalization, boosting performance on T2I-CompBench from 82.18 to \textbf{88.37} and on T2I-ReasonBench from 43.82 to \textbf{46.75}. Overall, our work establishes a powerful new paradigm for enabling a UMMs' understanding module to guide and enhance its own generation via self-rewarding.
comment: 20 pages, 8 figures, webpage can be seen in https://waynejin0918.github.io/srum_web/
☆ Cost Analysis of Human-corrected Transcription for Predominately Oral Languages
Creating speech datasets for low-resource languages is a critical yet poorly understood challenge, particularly regarding the actual cost in human labor. This paper investigates the time and complexity required to produce high-quality annotated speech data for a subset of low-resource languages, low literacy Predominately Oral Languages, focusing on Bambara, a Manding language of Mali. Through a one-month field study involving ten transcribers with native proficiency, we analyze the correction of ASR-generated transcriptions of 53 hours of Bambara voice data. We report that it takes, on average, 30 hours of human labor to accurately transcribe one hour of speech data under laboratory conditions and 36 hours under field conditions. The study provides a baseline and practical insights for a large class of languages with comparable profiles undertaking the creation of NLP resources.
comment: 6 pages, 1 figure
☆ Content Anonymization for Privacy in Long-form Audio
Voice anonymization techniques have been found to successfully obscure a speaker's acoustic identity in short, isolated utterances in benchmarks such as the VoicePrivacy Challenge. In practice, however, utterances seldom occur in isolation: long-form audio is commonplace in domains such as interviews, phone calls, and meetings. In these cases, many utterances from the same speaker are available, which pose a significantly greater privacy risk: given multiple utterances from the same speaker, an attacker could exploit an individual's vocabulary, syntax, and turns of phrase to re-identify them, even when their voice is completely disguised. To address this risk, we propose new content anonymization approaches. Our approach performs a contextual rewriting of the transcripts in an ASR-TTS pipeline to eliminate speaker-specific style while preserving meaning. We present results in a long-form telephone conversation setting demonstrating the effectiveness of a content-based attack on voice-anonymized speech. Then we show how the proposed content-based anonymization methods can mitigate this risk while preserving speech utility. Overall, we find that paraphrasing is an effective defense against content-based attacks and recommend that stakeholders adopt this step to ensure anonymity in long-form audio.
☆ Dr.LLM: Dynamic Layer Routing in LLMs
Large Language Models (LLMs) process every token through all layers of a transformer stack, causing wasted computation on simple queries and insufficient flexibility for harder ones that need deeper reasoning. Adaptive-depth methods can improve efficiency, but prior approaches rely on costly inference-time search, architectural changes, or large-scale retraining, and in practice often degrade accuracy despite efficiency gains. We introduce Dr.LLM, Dynamic routing of Layers for LLMs, a retrofittable framework that equips pretrained models with lightweight per-layer routers deciding to skip, execute, or repeat a block. Routers are trained with explicit supervision: using Monte Carlo Tree Search (MCTS), we derive high-quality layer configurations that preserve or improve accuracy under a compute budget. Our design, windowed pooling for stable routing, focal loss with class balancing, and bottleneck MLP routers, ensures robustness under class imbalance and long sequences. On ARC (logic) and DART (math), Dr.LLM improves accuracy by up to +3.4%p while saving 5 layers per example on average. Routers generalize to out-of-domain tasks (MMLU, GSM8k, AIME, TruthfulQA, SQuADv2, GPQA, PIQA, AGIEval) with only 0.85% accuracy drop while retaining efficiency, and outperform prior routing methods by up to +7.7%p. Overall, Dr.LLM shows that explicitly supervised routers retrofit frozen LLMs for budget-aware, accuracy-driven inference without altering base weights.
comment: 17 pages, Under submission
☆ Language Models Model Language
Linguistic commentary on LLMs, heavily influenced by the theoretical frameworks of de Saussure and Chomsky, is often speculative and unproductive. Critics challenge whether LLMs can legitimately model language, citing the need for "deep structure" or "grounding" to achieve an idealized linguistic "competence." We argue for a radical shift in perspective towards the empiricist principles of Witold Ma\'nczak, a prominent general and historical linguist. He defines language not as a "system of signs" or a "computational system of the brain" but as the totality of all that is said and written. Above all, he identifies frequency of use of particular language elements as language's primary governing principle. Using his framework, we challenge prior critiques of LLMs and provide a constructive guide for designing, evaluating, and interpreting language models.
☆ Hey, wait a minute: on at-issue sensitivity in Language Models
Evaluating the naturalness of dialogue in language models (LMs) is not trivial: notions of 'naturalness' vary, and scalable quantitative metrics remain limited. This study leverages the linguistic notion of 'at-issueness' to assess dialogue naturalness and introduces a new method: Divide, Generate, Recombine, and Compare (DGRC). DGRC (i) divides a dialogue as a prompt, (ii) generates continuations for subparts using LMs, (iii) recombines the dialogue and continuations, and (iv) compares the likelihoods of the recombined sequences. This approach mitigates bias in linguistic analyses of LMs and enables systematic testing of discourse-sensitive behavior. Applying DGRC, we find that LMs prefer to continue dialogue on at-issue content, with this effect enhanced in instruct-tuned models. They also reduce their at-issue preference when relevant cues (e.g., "Hey, wait a minute") are present. Although instruct-tuning does not further amplify this modulation, the pattern reflects a hallmark of successful dialogue dynamics.
comment: 10 pages, 5 figures, 3 tables. See https://github.com/sangheek16/hey-wait-a-minute for code and data
☆ Which Word Orders Facilitate Length Generalization in LMs? An Investigation with GCG-Based Artificial Languages EMNLP 2025
Whether language models (LMs) have inductive biases that favor typologically frequent grammatical properties over rare, implausible ones has been investigated, typically using artificial languages (ALs) (White and Cotterell, 2021; Kuribayashi et al., 2024). In this paper, we extend these works from two perspectives. First, we extend their context-free AL formalization by adopting Generalized Categorial Grammar (GCG) (Wood, 2014), which allows ALs to cover attested but previously overlooked constructions, such as unbounded dependency and mildly context-sensitive structures. Second, our evaluation focuses more on the generalization ability of LMs to process unseen longer test sentences. Thus, our ALs better capture features of natural languages and our experimental paradigm leads to clearer conclusions -- typologically plausible word orders tend to be easier for LMs to productively generalize.
comment: EMNLP 2025 Main Conference
☆ Omni-Captioner: Data Pipeline, Models, and Benchmark for Omni Detailed Perception
Fine-grained perception of multimodal information is critical for advancing human-AI interaction. With recent progress in audio-visual technologies, Omni Language Models (OLMs), capable of processing audio and video signals in parallel, have emerged as a promising paradigm for achieving richer understanding and reasoning. However, their capacity to capture and describe fine-grained details remains limited explored. In this work, we present a systematic and comprehensive investigation of omni detailed perception from the perspectives of the data pipeline, models, and benchmark. We first identify an inherent "co-growth" between detail and hallucination in current OLMs. To address this, we propose Omni-Detective, an agentic data generation pipeline integrating tool-calling, to autonomously produce highly detailed yet minimally hallucinatory multimodal data. Based on the data generated with Omni-Detective, we train two captioning models: Audio-Captioner for audio-only detailed perception, and Omni-Captioner for audio-visual detailed perception. Under the cascade evaluation protocol, Audio-Captioner achieves the best performance on MMAU and MMAR among all open-source models, surpassing Gemini 2.5 Flash and delivering performance comparable to Gemini 2.5 Pro. On existing detailed captioning benchmarks, Omni-Captioner sets a new state-of-the-art on VDC and achieves the best trade-off between detail and hallucination on the video-SALMONN 2 testset. Given the absence of a dedicated benchmark for omni detailed perception, we design Omni-Cloze, a novel cloze-style evaluation for detailed audio, visual, and audio-visual captioning that ensures stable, efficient, and reliable assessment. Experimental results and analysis demonstrate the effectiveness of Omni-Detective in generating high-quality detailed captions, as well as the superiority of Omni-Cloze in evaluating such detailed captions.
comment: https://github.com/ddlBoJack/Omni-Captioner
☆ Generation Space Size: Understanding and Calibrating Open-Endedness of LLM Generations
Different open-ended generation tasks require different degrees of output diversity. However, current LLMs are often miscalibrated. They collapse to overly homogeneous outputs for creative tasks and hallucinate diverse but incorrect responses for factual tasks. We argue that these two failure modes are unified by, and can both be addressed by, the notion of effective generation space size (GSS) -- the set of semantically distinct outputs a model considers for a prompt. We present GSSBench, a task suite of prompt pairs with ground-truth GSS relationships to assess different metrics and understand where models diverge from desired behavior. We find that hallucination detection metrics, particularly EigenScore, consistently outperform standard diversity and uncertainty quantification metrics, while using only model internals, providing interpretable insights into a model's internal task representations. We demonstrate three applications of GSS: (1) detecting prompt ambiguity and predicting clarification questions for better grounding, (2) interpreting overthinking and underthinking in reasoning models, and (3) steering models to expand their generation space to yield high-quality and diverse outputs.
☆ Who is a Better Matchmaker? Human vs. Algorithmic Judge Assignment in a High-Stakes Startup Competition
There is growing interest in applying artificial intelligence (AI) to automate and support complex decision-making tasks. However, it remains unclear how algorithms compare to human judgment in contexts requiring semantic understanding and domain expertise. We examine this in the context of the judge assignment problem, matching submissions to suitably qualified judges. Specifically, we tackled this problem at the Harvard President's Innovation Challenge, the university's premier venture competition awarding over \$500,000 to student and alumni startups. This represents a real-world environment where high-quality judge assignment is essential. We developed an AI-based judge-assignment algorithm, Hybrid Lexical-Semantic Similarity Ensemble (HLSE), and deployed it at the competition. We then evaluated its performance against human expert assignments using blinded match-quality scores from judges on $309$ judge-venture pairs. Using a Mann-Whitney U statistic based test, we found no statistically significant difference in assignment quality between the two approaches ($AUC=0.48, p=0.40$); on average, algorithmic matches are rated $3.90$ and manual matches $3.94$ on a 5-point scale, where 5 indicates an excellent match. Furthermore, manual assignments that previously required a full week could be automated in several hours by the algorithm during deployment. These results demonstrate that HLSE achieves human-expert-level matching quality while offering greater scalability and efficiency, underscoring the potential of AI-driven solutions to support and enhance human decision-making for judge assignment in high-stakes settings.
comment: 17 Pages, 2 figures
☆ Demystifying Hybrid Thinking: Can LLMs Truly Switch Between Think and No-Think?
Hybrid thinking enables LLMs to switch between reasoning and direct answering, offering a balance between efficiency and reasoning capability. Yet our experiments reveal that current hybrid thinking LLMs only achieve partial mode separation: reasoning behaviors often leak into the no-think mode. To understand and mitigate this, we analyze the factors influencing controllability and identify four that matter most: (1) larger data scale, (2) using think and no-think answers from different questions rather than the same question, (3) a moderate increase in no-think data number, and (4) a two-phase strategy that first trains reasoning ability and then applies hybrid think training. Building on these findings, we propose a practical recipe that, compared to standard training, can maintain accuracy in both modes while significantly reducing no-think output length (from $1085$ to $585$ on MATH500) and occurrences of reasoning-supportive tokens such as ``\texttt{wait}'' (from $5917$ to $522$ on MATH500). Our findings highlight the limitations of current hybrid thinking and offer directions for strengthening its controllability.
comment: 10 pages, 6 figures
☆ The Role of Parametric Injection-A Systematic Study of Parametric Retrieval-Augmented Generation
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by retrieving external documents. As an emerging form of RAG, parametric retrieval-augmented generation (PRAG) encodes documents as model parameters (i.e., LoRA modules) and injects these representations into the model during inference, enabling interaction between the LLM and documents at parametric level. Compared with directly placing documents in the input context, PRAG is more efficient and has the potential to offer deeper model-document interaction. Despite its growing attention, the mechanism underlying parametric injection remains poorly understood. In this work, we present a systematic study of PRAG to clarify the role of parametric injection, showing that parameterized documents capture only partial semantic information of documents, and relying on them alone yields inferior performance compared to interaction at text level. However, these parametric representations encode high-level document information that can enhance the model's understanding of documents within the input context. When combined parameterized documents with textual documents, the model can leverage relevant information more effectively and become more robust to noisy inputs, achieving better performance than either source alone. We recommend jointly using parameterized and textual documents and advocate for increasing the information content of parametric representations to advance PRAG.
☆ Reasoning Pattern Matters: Learning to Reason without Human Rationales
Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities under the widely adopted SFT+RLVR paradigm, which first performs Supervised Fine-Tuning (SFT) on human-annotated reasoning trajectories (rationales) to establish initial reasoning behaviors, then applies Reinforcement Learning with Verifiable Rewards (RLVR) to optimize the model using verifiable signals without golden rationales. However, annotating high-quality rationales for the SFT stage remains prohibitively expensive. This paper investigates when and how rationale annotation costs can be substantially reduced without compromising reasoning performance. We identify a broad class of problems, termed patterned reasoning tasks, where reasoning follows a fixed, procedural strategy consistent across instances. Although instances vary in content such as domain knowledge, factual information, or numeric values, the solution derives from applying a shared reasoning pattern. We argue that the success of SFT+RLVR on such tasks primarily stems from its ability to enable models to internalize these reasoning patterns. Using numerical semantic matching as a representative task, we provide both causal and behavioral evidence showing that reasoning patterns rather than the quantity or quality of rationales are the key determinant of performance. Building on these insights, we propose Pattern-Aware LLMs as Rationale AnnOtators (PARO), a simple yet effective framework that enables LLMs to generate rationales aligned with task-specific reasoning patterns without requiring human rationale annotations. Experiments show that PARO-generated rationales achieve comparable SFT+RLVR performance to human rationales that are 10 times larger. These results suggest that large-scale human rationale annotations can be replaced with LLM-based automatic annotations requiring only limited human supervision over reasoning patterns.
comment: Submitted to Frontiers of Computer Science
☆ COSTAR-A: A prompting framework for enhancing Large Language Model performance on Point-of-View questions
Large Language Models (LLMs) are highly sensitive to prompt design, and making optimized prompting techniques is crucial for generating consistent, high-quality outputs. In this study, we introduce COSTAR-A, a novel prompt engineering framework that enhances the existing COSTAR method, which stands for Context, Objective, Style, Tone, Audience, and Response, by adding the 'Answer' component at the end. We demonstrate that while the original COSTAR framework improves prompt clarity and aligns outputs for larger LLMs, its performance is less consistent with smaller, locally optimized models, particularly in tasks that require more directive or constrained outputs. Through a series of controlled prompt-output assessments with smaller (at most 8 billion parameters), fine-tuned models, we found that COSTAR-A can enhance the output structure and decisiveness of localized LLMs for certain tasks, although its effectiveness varies across models and use cases. Notably, the Llama 3.1-8B model exhibited performance improvements when prompted with COSTAR-A compared to COSTAR alone. These findings emphasize the adaptability and scalability of COSTAR-A as a prompting framework, particularly in computationally efficient AI deployments on resource-constrained hardware.
comment: 20 pages, 2 figures
☆ ACADATA: Parallel Dataset of Academic Data for Machine Translation
We present ACADATA, a high-quality parallel dataset for academic translation, that consists of two subsets: ACAD-TRAIN, which contains approximately 1.5 million author-generated paragraph pairs across 96 language directions and ACAD-BENCH, a curated evaluation set of almost 6,000 translations covering 12 directions. To validate its utility, we fine-tune two Large Language Models (LLMs) on ACAD-TRAIN and benchmark them on ACAD-BENCH against specialized machine-translation systems, general-purpose, open-weight LLMs, and several large-scale proprietary models. Experimental results demonstrate that fine-tuning on ACAD-TRAIN leads to improvements in academic translation quality by +6.1 and +12.4 d-BLEU points on average for 7B and 2B models respectively, while also improving long-context translation in a general domain by up to 24.9% when translating out of English. The fine-tuned top-performing model surpasses the best propietary and open-weight models on academic translation domain. By releasing ACAD-TRAIN, ACAD-BENCH and the fine-tuned models, we provide the community with a valuable resource to advance research in academic domain and long-context translation.
☆ StyleDecipher: Robust and Explainable Detection of LLM-Generated Texts with Stylistic Analysis
With the increasing integration of large language models (LLMs) into open-domain writing, detecting machine-generated text has become a critical task for ensuring content authenticity and trust. Existing approaches rely on statistical discrepancies or model-specific heuristics to distinguish between LLM-generated and human-written text. However, these methods struggle in real-world scenarios due to limited generalization, vulnerability to paraphrasing, and lack of explainability, particularly when facing stylistic diversity or hybrid human-AI authorship. In this work, we propose StyleDecipher, a robust and explainable detection framework that revisits LLM-generated text detection using combined feature extractors to quantify stylistic differences. By jointly modeling discrete stylistic indicators and continuous stylistic representations derived from semantic embeddings, StyleDecipher captures distinctive style-level divergences between human and LLM outputs within a unified representation space. This framework enables accurate, explainable, and domain-agnostic detection without requiring access to model internals or labeled segments. Extensive experiments across five diverse domains, including news, code, essays, reviews, and academic abstracts, demonstrate that StyleDecipher consistently achieves state-of-the-art in-domain accuracy. Moreover, in cross-domain evaluations, it surpasses existing baselines by up to 36.30%, while maintaining robustness against adversarial perturbations and mixed human-AI content. Further qualitative and quantitative analysis confirms that stylistic signals provide explainable evidence for distinguishing machine-generated text. Our source code can be accessed at https://github.com/SiyuanLi00/StyleDecipher.
☆ Reasoning in the Dark: Interleaved Vision-Text Reasoning in Latent Space
Multimodal reasoning aims to enhance the capabilities of MLLMs by incorporating intermediate reasoning steps before reaching the final answer. It has evolved from text-only reasoning to the integration of visual information, enabling the thought process to be conveyed through both images and text. Despite its effectiveness, current multimodal reasoning methods depend on explicit reasoning steps that require labor-intensive vision-text annotations and inherently introduce significant inference latency. To address these issues, we introduce multimodal latent reasoning with the advantages of multimodal representation, reduced annotation, and inference efficiency. To facilicate it, we propose Interleaved Vision-Text Latent Reasoning (IVT-LR), which injects both visual and textual information in the reasoning process within the latent space. Specifically, IVT-LR represents each reasoning step by combining two implicit parts: latent text (the hidden states from the previous step) and latent vision (a set of selected image embeddings). We further introduce a progressive multi-stage training strategy to enable MLLMs to perform the above multimodal latent reasoning steps. Experiments on M3CoT and ScienceQA demonstrate that our IVT-LR method achieves an average performance increase of 5.45% in accuracy, while simultaneously achieving a speed increase of over 5 times compared to existing approaches. Code available at https://github.com/FYYDCC/IVT-LR.
☆ Teaching Language Models to Faithfully Express their Uncertainty
Large language models (LLMs) often miscommunicate their uncertainty: repeated queries can produce divergent answers, yet generated responses are typically unhedged or hedged in ways that do not reflect this variability. This conveys unfaithful information about the uncertain state of the LLMs' knowledge, creating a faithfulness gap that affects even strong LLMs. We introduce Faithful Uncertainty Tuning (FUT): a fine-tuning approach that teaches instruction-tuned LLMs to express uncertainty faithfully without altering their underlying answer distribution. We construct training data by augmenting model samples with uncertainty hedges (i.e. verbal cues such as 'possibly' or 'likely') aligned with sample consistency, requiring no supervision beyond the model and a set of prompts. We evaluate FUT on open-domain question answering (QA) across multiple models and datasets. Our results show that FUT substantially reduces the faithfulness gap, while preserving QA accuracy and introducing minimal semantic distribution shift. Further analyses demonstrate robustness across decoding strategies, choice of hedgers, and other forms of uncertainty expression (i.e. numerical). These findings establish FUT as a simple and effective way to teach LLMs to communicate uncertainty faithfully.
☆ VISaGE: Understanding Visual Generics and Exceptions EMNLP 2025
While Vision Language Models (VLMs) learn conceptual representations, in the form of generalized knowledge, during training, they are typically used to analyze individual instances. When evaluation instances are atypical, this paradigm results in tension between two priors in the model. The first is a pragmatic prior that the textual and visual input are both relevant, arising from VLM finetuning on congruent inputs; the second is a semantic prior that the conceptual representation is generally true for instances of the category. In order to understand how VLMs trade off these priors, we introduce a new evaluation dataset, VISaGE, consisting of both typical and exceptional images. In carefully balanced experiments, we show that conceptual understanding degrades when the assumption of congruency underlying the pragmatic prior is violated with incongruent images. This effect is stronger than the effect of the semantic prior when querying about individual instances.
comment: EMNLP 2025
☆ BoN Appetit Team at LeWiDi-2025: Best-of-N Test-time Scaling Can Not Stomach Annotation Disagreements (Yet)
Test-time scaling is a family of techniques to improve LLM outputs at inference time by performing extra computation. To the best of our knowledge, test-time scaling has been limited to domains with verifiably correct answers, like mathematics and coding. We transfer test-time scaling to the LeWiDi-2025 tasks to evaluate annotation disagreements. We experiment with three test-time scaling methods: two benchmark algorithms (Model Averaging and Majority Voting), and a Best-of-N sampling method. The two benchmark methods improve LLM performance consistently on the LeWiDi tasks, but the Best-of-N method does not. Our experiments suggest that the Best-of-N method does not currently transfer from mathematics to LeWiDi tasks, and we analyze potential reasons for this gap.
☆ When Personalization Tricks Detectors: The Feature-Inversion Trap in Machine-Generated Text Detection
Large language models (LLMs) have grown more powerful in language generation, producing fluent text and even imitating personal style. Yet, this ability also heightens the risk of identity impersonation. To the best of our knowledge, no prior work has examined personalized machine-generated text (MGT) detection. In this paper, we introduce \dataset, the first benchmark for evaluating detector robustness in personalized settings, built from literary and blog texts paired with their LLM-generated imitations. Our experimental results demonstrate large performance gaps across detectors in personalized settings: some state-of-the-art models suffer significant drops. We attribute this limitation to the \textit{feature-inversion trap}, where features that are discriminative in general domains become inverted and misleading when applied to personalized text. Based on this finding, we propose \method, a simple and reliable way to predict detector performance changes in personalized settings. \method identifies latent directions corresponding to inverted features and constructs probe datasets that differ primarily along these features to evaluate detector dependence. Our experiments show that \method can accurately predict both the direction and the magnitude of post-transfer changes, showing 85\% correlation with the actual performance gaps. We hope that this work will encourage further research on personalized text detection.
☆ SMEC: Rethinking Matryoshka Representation Learning for Retrieval Embedding Compression EMNLP2025
Large language models (LLMs) generate high-dimensional embeddings that capture rich semantic and syntactic information. However, high-dimensional embeddings exacerbate computational complexity and storage requirements, thereby hindering practical deployment. To address these challenges, we propose a novel training framework named Sequential Matryoshka Embedding Compression (SMEC). This framework introduces the Sequential Matryoshka Representation Learning(SMRL) method to mitigate gradient variance during training, the Adaptive Dimension Selection (ADS) module to reduce information degradation during dimension pruning, and the Selectable Cross-batch Memory (S-XBM) module to enhance unsupervised learning between high- and low-dimensional embeddings. Experiments on image, text, and multimodal datasets demonstrate that SMEC achieves significant dimensionality reduction while maintaining performance. For instance, on the BEIR dataset, our approach improves the performance of compressed LLM2Vec embeddings (256 dimensions) by 1.1 points and 2.7 points compared to the Matryoshka-Adaptor and Search-Adaptor models, respectively.
comment: Accepted by EMNLP2025
☆ Resource-sensitive but language-blind: Community size and not grammatical complexity better predicts the accuracy of Large Language Models in a novel Wug Test
The linguistic abilities of Large Language Models are a matter of ongoing debate. This study contributes to this discussion by investigating model performance in a morphological generalization task that involves novel words. Using a multilingual adaptation of the Wug Test, six models were tested across four partially unrelated languages (Catalan, English, Greek, and Spanish) and compared with human speakers. The aim is to determine whether model accuracy approximates human competence and whether it is shaped primarily by linguistic complexity or by the quantity of available training data. Consistent with previous research, the results show that the models are able to generalize morphological processes to unseen words with human-like accuracy. However, accuracy patterns align more closely with community size and data availability than with structural complexity, refining earlier claims in the literature. In particular, languages with larger speaker communities and stronger digital representation, such as Spanish and English, revealed higher accuracy than less-resourced ones like Catalan and Greek. Overall, our findings suggest that model behavior is mainly driven by the richness of linguistic resources rather than by sensitivity to grammatical complexity, reflecting a form of performance that resembles human linguistic competence only superficially.
☆ Probing Latent Knowledge Conflict for Faithful Retrieval-Augmented Generation
Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm to enhance the factuality of Large Language Models (LLMs). However, existing RAG systems often suffer from an unfaithfulness issue, where the model's response contradicts evidence from the retrieved context. Existing approaches to improving contextual faithfulness largely rely on external interventions, such as prompt engineering, decoding constraints, or reward-based fine-tuning. These works treat the LLM as a black box and overlook a crucial question: how does the LLM internally integrate retrieved evidence with its parametric memory, particularly under knowledge conflicts? To address this gap, we conduct a probing-based analysis of hidden-state representations in LLMs and observe three findings: knowledge integration occurs hierarchically, conflicts manifest as latent signals at the sentence level, and irrelevant context is often amplified when aligned with parametric knowledge. Building on these findings, we propose CLEAR (Conflict-Localized and Enhanced Attention for RAG), a framework that (i) decomposes context into fine-grained sentence-level knowledge, (ii) employs hidden-state probing to localize conflicting knowledge, and (iii) introduces conflict-aware fine-tuning to guide the model to accurately integrate retrieved evidence. Extensive experiments across three benchmarks demonstrate that CLEAR substantially improves both accuracy and contextual faithfulness, consistently outperforming strong baselines under diverse conflict conditions. The related resources are available at https://github.com/LinfengGao/CLEAR.
☆ PRoH: Dynamic Planning and Reasoning over Knowledge Hypergraphs for Retrieval-Augmented Generation
Knowledge Hypergraphs (KHs) have recently emerged as a knowledge representation for retrieval-augmented generation (RAG), offering a paradigm to model multi-entity relations into a structured form. However, existing KH-based RAG methods suffer from three major limitations: static retrieval planning, non-adaptive retrieval execution, and superficial use of KH structure and semantics, which constrain their ability to perform effective multi-hop question answering. To overcome these limitations, we propose PRoH, a dynamic Planning and Reasoning over Knowledge Hypergraphs framework. PRoH incorporates three core innovations: (i) a context-aware planning module that sketches the local KH neighborhood to guide structurally grounded reasoning plan generation; (ii) a structured question decomposition process that organizes subquestions as a dynamically evolving Directed Acyclic Graph (DAG) to enable adaptive, multi-trajectory exploration; and (iii) an Entity-Weighted Overlap (EWO)-guided reasoning path retrieval algorithm that prioritizes semantically coherent hyperedge traversals. Experiments across multiple domains demonstrate that PRoH achieves state-of-the-art performance, surpassing the prior SOTA model HyperGraphRAG by an average of 19.73% in F1 and 8.41% in Generation Evaluation (G-E) score, while maintaining strong robustness in long-range multi-hop reasoning tasks.
☆ Tokenization Disparities as Infrastructure Bias: How Subword Systems Create Inequities in LLM Access and Efficiency
Tokenization disparities pose a significant barrier to achieving equitable access to artificial intelligence across linguistically diverse populations. This study conducts a large-scale cross-linguistic evaluation of tokenization efficiency in over 200 languages to systematically quantify computational inequities in large language models (LLMs). Using a standardized experimental framework, we applied consistent preprocessing and normalization protocols, followed by uniform tokenization through the tiktoken library across all language samples. Comprehensive tokenization statistics were collected using established evaluation metrics, including Tokens Per Sentence (TPS) and Relative Tokenization Cost (RTC), benchmarked against English baselines. Our cross-linguistic analysis reveals substantial and systematic disparities: Latin-script languages consistently exhibit higher tokenization efficiency, while non-Latin and morphologically complex languages incur significantly greater token inflation, often 3-5 times higher RTC ratios. These inefficiencies translate into increased computational costs and reduced effective context utilization for underrepresented languages. Overall, the findings highlight structural inequities in current AI systems, where speakers of low-resource and non-Latin languages face disproportionate computational disadvantages. Future research should prioritize the development of linguistically informed tokenization strategies and adaptive vocabulary construction methods that incorporate typological diversity, ensuring more inclusive and computationally equitable multilingual AI systems.
comment: 6 pages 4 figures
☆ LLM-REVal: Can We Trust LLM Reviewers Yet?
The rapid advancement of large language models (LLMs) has inspired researchers to integrate them extensively into the academic workflow, potentially reshaping how research is practiced and reviewed. While previous studies highlight the potential of LLMs in supporting research and peer review, their dual roles in the academic workflow and the complex interplay between research and review bring new risks that remain largely underexplored. In this study, we focus on how the deep integration of LLMs into both peer-review and research processes may influence scholarly fairness, examining the potential risks of using LLMs as reviewers by simulation. This simulation incorporates a research agent, which generates papers and revises, alongside a review agent, which assesses the submissions. Based on the simulation results, we conduct human annotations and identify pronounced misalignment between LLM-based reviews and human judgments: (1) LLM reviewers systematically inflate scores for LLM-authored papers, assigning them markedly higher scores than human-authored ones; (2) LLM reviewers persistently underrate human-authored papers with critical statements (e.g., risk, fairness), even after multiple revisions. Our analysis reveals that these stem from two primary biases in LLM reviewers: a linguistic feature bias favoring LLM-generated writing styles, and an aversion toward critical statements. These results highlight the risks and equity concerns posed to human authors and academic research if LLMs are deployed in the peer review cycle without adequate caution. On the other hand, revisions guided by LLM reviews yield quality gains in both LLM-based and human evaluations, illustrating the potential of the LLMs-as-reviewers for early-stage researchers and enhancing low-quality papers.
☆ MoBiLE: Efficient Mixture-of-Experts Inference on Consumer GPU with Mixture of Big Little Experts SP
Mixture-of-Experts (MoE) models have recently demonstrated exceptional performance across a diverse range of applications. The principle of sparse activation in MoE models facilitates an offloading strategy, wherein active experts are maintained in GPU HBM, while inactive experts are stored in CPU DRAM. The efficacy of this approach, however, is fundamentally constrained by the limited bandwidth of the CPU-GPU interconnect. To mitigate this bottleneck, existing approaches have employed prefetching to accelerate MoE inference. These methods attempt to predict and prefetch the required experts using specially trained modules. Nevertheless, such techniques are often encumbered by significant training overhead and have shown diminished effectiveness on recent MoE models with fine-grained expert segmentation. In this paper, we propose MoBiLE, a plug-and-play offloading-based MoE inference framework with \textit{mixture of big-little experts}. It reduces the number of experts for unimportant tokens to half for acceleration while maintaining full experts for important tokens to guarantee model quality. Further, a dedicated fallback and prefetching mechanism is designed for switching between little and big experts to improve memory efficiency. We evaluate MoBiLE on four typical modern MoE architectures and challenging generative tasks. Our results show that MoBiLE achieves a speedup of 1.60x to 1.72x compared to the baseline on a consumer GPU system, with negligible degradation in accuracy.
comment: Accepted to ASP-DAC 2026
☆ Fine-grained Analysis of Brain-LLM Alignment through Input Attribution
Understanding the alignment between large language models (LLMs) and human brain activity can reveal computational principles underlying language processing. We introduce a fine-grained input attribution method to identify the specific words most important for brain-LLM alignment, and leverage it to study a contentious research question about brain-LLM alignment: the relationship between brain alignment (BA) and next-word prediction (NWP). Our findings reveal that BA and NWP rely on largely distinct word subsets: NWP exhibits recency and primacy biases with a focus on syntax, while BA prioritizes semantic and discourse-level information with a more targeted recency effect. This work advances our understanding of how LLMs relate to human language processing and highlights differences in feature reliance between BA and NWP. Beyond this study, our attribution method can be broadly applied to explore the cognitive relevance of model predictions in diverse language processing tasks.
☆ Simple Projection Variants Improve ColBERT Performance
Multi-vector dense retrieval methods like ColBERT systematically use a single-layer linear projection to reduce the dimensionality of individual vectors. In this study, we explore the implications of the MaxSim operator on the gradient flows of the training of multi-vector models and show that such a simple linear projection has inherent, if non-critical, limitations in this setting. We then discuss the theoretical improvements that could result from replacing this single-layer projection with well-studied alternative feedforward linear networks (FFN), such as deeper, non-linear FFN blocks, GLU blocks, and skip-connections, could alleviate these limitations. Through the design and systematic evaluation of alternate projection blocks, we show that better-designed final projections positively impact the downstream performance of ColBERT models. We highlight that many projection variants outperform the original linear projections, with the best-performing variants increasing average performance on a range of retrieval benchmarks across domains by over 2 NDCG@10 points. We then conduct further exploration on the individual parameters of these projections block in order to understand what drives this empirical performance, highlighting the particular importance of upscaled intermediate projections and residual connections. As part of these ablation studies, we show that numerous suboptimal projection variants still outperform the traditional single-layer projection across multiple benchmarks, confirming our hypothesis. Finally, we observe that this effect is consistent across random seeds, further confirming that replacing the linear layer of ColBERT models is a robust, drop-in upgrade.
☆ Beating Harmful Stereotypes Through Facts: RAG-based Counter-speech Generation
Counter-speech generation is at the core of many expert activities, such as fact-checking and hate speech, to counter harmful content. Yet, existing work treats counter-speech generation as pure text generation task, mainly based on Large Language Models or NGO experts. These approaches show severe drawbacks due to the limited reliability and coherence in the generated countering text, and in scalability, respectively. To close this gap, we introduce a novel framework to model counter-speech generation as knowledge-wise text generation process. Our framework integrates advanced Retrieval-Augmented Generation (RAG) pipelines to ensure the generation of trustworthy counter-speech for 8 main target groups identified in the hate speech literature, including women, people of colour, persons with disabilities, migrants, Muslims, Jews, LGBT persons, and other. We built a knowledge base over the United Nations Digital Library, EUR-Lex and the EU Agency for Fundamental Rights, comprising a total of 32,792 texts. We use the MultiTarget-CONAN dataset to empirically assess the quality of the generated counter-speech, both through standard metrics (i.e., JudgeLM) and a human evaluation. Results show that our framework outperforms standard LLM baselines and competitive approach, on both assessments. The resulting framework and the knowledge base pave the way for studying trustworthy and sound counter-speech generation, in hate speech and beyond.
☆ A large-scale, unsupervised pipeline for automatic corpus annotation using LLMs: variation and change in the English consider construction
As natural language corpora expand at an unprecedented rate, manual annotation remains a significant methodological bottleneck in corpus linguistic work. We address this challenge by presenting a scalable, unsupervised pipeline for automating grammatical annotation in voluminous corpora using large language models (LLMs). Unlike previous supervised and iterative approaches, our method employs a four-phase workflow: prompt engineering, pre-hoc evaluation, automated batch processing, and post-hoc validation. We demonstrate the pipeline's accessibility and effectiveness through a diachronic case study of variation in the English consider construction. Using GPT-5 through the OpenAI API, we annotate 143,933 sentences from the Corpus of Historical American English (COHA) in under 60 hours, achieving 98%+ accuracy on two sophisticated annotation procedures. Our results suggest that LLMs can perform a range of data preparation tasks at scale with minimal human intervention, opening new possibilities for corpus-based research, though implementation requires attention to costs, licensing, and other ethical considerations.
☆ Vision Language Models Map Logos to Text via Semantic Entanglement in the Visual Projector
Vision Language Models (VLMs) have achieved impressive progress in multimodal reasoning; yet, they remain vulnerable to hallucinations, where outputs are not grounded in visual evidence. In this paper, we investigate a previously overlooked setting: logo hallucination, where models generate brand names or textual content despite logos containing no visible words. Using curated splits of pure symbols, hybrids, and text-bearing logos, as well as the challenging Hard-60 subset, we systematically measure hallucination across leading VLMs. We further probe robustness through nine structured perturbations and show that hallucinations persist even under strong distortions, with occlusion exposing the sharpest weaknesses. Embedding-level analysis with open-weight LLaVA demonstrates that hallucination is tied to a small subset of projector dimensions, and targeted ablation substantially reduces errors while preserving OCR accuracy. Together, these findings reveal that VLMs often rely on symbolic priors rather than genuine glyph perception, particularly for iconic circular logos, and that projector subspaces play a decisive role in this failure mode. Our work contributes both a novel diagnostic lens and actionable mitigation insights, highlighting projector disentanglement and OCR-guided decoding as promising directions for building more trustworthy multimodal systems.
☆ Chinese ModernBERT with Whole-Word Masking
Encoder-only Transformers have advanced along three axes -- architecture, data, and systems -- yielding Pareto gains in accuracy, speed, and memory efficiency. Yet these improvements have not fully transferred to Chinese, where tokenization and morphology differ markedly from English. We introduce Chinese ModernBERT, a from-scratch Chinese encoder that couples: (i) a hardware-aware 32k BPE vocabulary tailored to frequent Chinese affixes/compounds, lowering the embedding budget; (ii) whole-word masking (WWM) with a dynamic masking curriculum (30% -> 15%) to align task difficulty with training progress; (iii) a two-stage pre-training pipeline that extends the native context from 1,024 to 8,192 tokens using RoPE and alternating local/global attention; and (iv) a damped-cosine learning-rate schedule for stable long-horizon optimization. We pre-train on ~1.2T Chinese tokens from CCI3-HQ, CCI4 (Chinese), and Cosmopedia-Chinese. On CLUE, Chinese ModernBERT is competitive with strong Chinese encoders under a unified fine-tuning protocol. Under bf16 it achieves high long-sequence throughput while maintaining strong short-sequence speed, reflecting benefits from budget allocation and attention design. To probe retrieval-oriented quality, we add a small amount of open contrastive data: fine-tuning on SimCLUE (~3M pairs) improves further when adding T2Ranking (~2M), reaching 0.505 (Pearson) / 0.537 (Spearman) on the SimCLUE test set. Under this open-data setting, Chinese ModernBERT surpasses Qwen-0.6B-embedding on SimCLUE, suggesting a clear scaling path for STS with additional curated pairs. We will release tokenizer and weights to facilitate reproducible research.
☆ Shallow Robustness, Deep Vulnerabilities: Multi-Turn Evaluation of Medical LLMs NeurIPS 2025
Large language models (LLMs) are rapidly transitioning into medical clinical use, yet their reliability under realistic, multi-turn interactions remains poorly understood. Existing evaluation frameworks typically assess single-turn question answering under idealized conditions, overlooking the complexities of medical consultations where conflicting input, misleading context, and authority influence are common. We introduce MedQA-Followup, a framework for systematically evaluating multi-turn robustness in medical question answering. Our approach distinguishes between shallow robustness (resisting misleading initial context) and deep robustness (maintaining accuracy when answers are challenged across turns), while also introducing an indirect-direct axis that separates contextual framing (indirect) from explicit suggestion (direct). Using controlled interventions on the MedQA dataset, we evaluate five state-of-the-art LLMs and find that while models perform reasonably well under shallow perturbations, they exhibit severe vulnerabilities in multi-turn settings, with accuracy dropping from 91.2% to as low as 13.5% for Claude Sonnet 4. Counterintuitively, indirect, context-based interventions are often more harmful than direct suggestions, yielding larger accuracy drops across models and exposing a significant vulnerability for clinical deployment. Further compounding analyses reveal model differences, with some showing additional performance drops under repeated interventions while others partially recovering or even improving. These findings highlight multi-turn robustness as a critical but underexplored dimension for safe and reliable deployment of medical LLMs.
comment: Dataset and code: https://huggingface.co/datasets/dynamoai-ml/MedQA-USMLE-4-MultiTurnRobust ; https://github.com/bmanczak/MedQA-MultiTurnRobustness Accepted as a poster at NeurIPS 2025 Workshop on GenAI for Health: Potential, Trust, and Policy Compliance
☆ DSAS: A Universal Plug-and-Play Framework for Attention Optimization in Multi-Document Question Answering NeurIPS 2025
While large language models (LLMs) show considerable promise across various fields, they have notable limitations in handling multi-document question answering (Multi-doc QA) tasks. The first challenge is long-range dependency modeling, where LLMs struggle to focus on key information in long texts, which weakens important semantic connections. Second, most LLMs suffer from the ''lost-in-the-middle'' issue, where they have difficulty processing information in the middle of long inputs. Current solutions either truncate global dependencies or demand costly finetuning, ultimately lacking a universal and simple solution for these challenges. To resolve these limitations, we propose Dual-Stage Adaptive Sharpening (DSAS) containing two modules. (i) The Contextual Gate Weighting (CGW) module alleviates ''lost-in-the-middle'' by assessing paragraph relevance through layer-wise attention tracking and position-aware weighting. (ii) The Reciprocal Attention Suppression (RAS) module enhances focus on critical paragraphs by suppressing information exchange between key and irrelevant texts, thus mitigating the limitations in long-range dependency modeling. Notably, DSAS functions as a plug-and-play solution requiring no architectural modifications or extra training parameters. Extensive experiments on four benchmarks demonstrate DSAS's efficacy across mainstream LLMs (Llama, Qwen, Mistral, and Deepseek), with an average F1-score improvement of 4.2% in Multi-doc QA tasks on Llama-3.1-8B-Instruct and Qwen2.5-14B-Instruct. Ablation studies confirm the essential contributions of both the CGW and RAS modules. In addition, detailed discussions in the Appendix further validate the robustness and scalability of DSAS.
comment: 27 pages, has been accepted by NeurIPS 2025
☆ Analysing Moral Bias in Finetuned LLMs through Mechanistic Interpretability
Large language models (LLMs) have been shown to internalize human-like biases during finetuning, yet the mechanisms by which these biases manifest remain unclear. In this work, we investigated whether the well-known Knobe effect, a moral bias in intentionality judgements, emerges in finetuned LLMs and whether it can be traced back to specific components of the model. We conducted a Layer-Patching analysis across 3 open-weights LLMs and demonstrated that the bias is not only learned during finetuning but also localized in a specific set of layers. Surprisingly, we found that patching activations from the corresponding pretrained model into just a few critical layers is sufficient to eliminate the effect. Our findings offer new evidence that social biases in LLMs can be interpreted, localized, and mitigated through targeted interventions, without the need for model retraining.
comment: Preprint. Under review
☆ HALF: Harm-Aware LLM Fairness Evaluation Aligned with Deployment
Large language models (LLMs) are increasingly deployed across high-impact domains, from clinical decision support and legal analysis to hiring and education, making fairness and bias evaluation before deployment critical. However, existing evaluations lack grounding in real-world scenarios and do not account for differences in harm severity, e.g., a biased decision in surgery should not be weighed the same as a stylistic bias in text summarization. To address this gap, we introduce HALF (Harm-Aware LLM Fairness), a deployment-aligned framework that assesses model bias in realistic applications and weighs the outcomes by harm severity. HALF organizes nine application domains into three tiers (Severe, Moderate, Mild) using a five-stage pipeline. Our evaluation results across eight LLMs show that (1) LLMs are not consistently fair across domains, (2) model size or performance do not guarantee fairness, and (3) reasoning models perform better in medical decision support but worse in education. We conclude that HALF exposes a clear gap between previous benchmarking success and deployment readiness.
☆ DiSTAR: Diffusion over a Scalable Token Autoregressive Representation for Speech Generation
Recent attempts to interleave autoregressive (AR) sketchers with diffusion-based refiners over continuous speech representations have shown promise, but they remain brittle under distribution shift and offer limited levers for controllability. We introduce DISTAR, a zero-shot text-to-speech framework that operates entirely in a discrete residual vector quantization (RVQ) code space and tightly couples an AR language model with a masked diffusion model, without forced alignment or a duration predictor. Concretely, DISTAR drafts block-level RVQ tokens with an AR language model and then performs parallel masked-diffusion infilling conditioned on the draft to complete the next block, yielding long-form synthesis with blockwise parallelism while mitigating classic AR exposure bias. The discrete code space affords explicit control at inference: DISTAR produces high-quality audio under both greedy and sample-based decoding using classifier-free guidance, supports trade-offs between robustness and diversity, and enables variable bit-rate and controllable computation via RVQ layer pruning at test time. Extensive experiments and ablations demonstrate that DISTAR surpasses state-of-the-art zero-shot TTS systems in robustness, naturalness, and speaker/style consistency, while maintaining rich output diversity. Audio samples are provided on https://anonymous.4open.science/w/DiSTAR_demo.
☆ HackWorld: Evaluating Computer-Use Agents on Exploiting Web Application Vulnerabilities
Web applications are prime targets for cyberattacks as gateways to critical services and sensitive data. Traditional penetration testing is costly and expertise-intensive, making it difficult to scale with the growing web ecosystem. While language model agents show promise in cybersecurity, modern web applications demand visual understanding, dynamic content handling, and multi-step interactions that only computer-use agents (CUAs) can perform. Yet, their ability to discover and exploit vulnerabilities through graphical interfaces remains largely unexplored. We present HackWorld, the first framework for systematically evaluating CUAs' capabilities to exploit web application vulnerabilities via visual interaction. Unlike sanitized benchmarks, HackWorld includes 36 real-world applications across 11 frameworks and 7 languages, featuring realistic flaws such as injection vulnerabilities, authentication bypasses, and unsafe input handling. Using a Capture-the-Flag (CTF) setup, it tests CUAs' capacity to identify and exploit these weaknesses while navigating complex web interfaces. Evaluation of state-of-the-art CUAs reveals concerning trends: exploitation rates below 12% and low cybersecurity awareness. CUAs often fail at multi-step attack planning and misuse security tools. These results expose the current limitations of CUAs in web security contexts and highlight opportunities for developing more security-aware agents capable of effective vulnerability detection and exploitation.
☆ DPO-Tuned Large Language Models for Segmentation in Simultaneous Speech Translation
Simultaneous speech translation requires accurate segmentation to balance translation quality and latency. Recent studies such as SHAS have introduced pretrained segmentation models, achieving stronger performance than heuristic rules. However, segmentation models such as SHAS, though pretrained and more robust than heuristic methods, are still constrained by supervised learning objectives and do not incorporate human preference alignment, which is crucial for natural real-time interpretation. In this work, we propose a segmentation framework based on large language models (LLMs) trained with Direct Preference Optimization (DPO). By leveraging preference alignment, our method enables LLMs to predict natural segmentation points that better meet the demands of real-time translation. We evaluate the system on the ACL 60/60 corpus across three language pairs (English-Japanese, Chinese, German), using SeamlessM4T v2 as the translation backbone. Experimental results show that our DPO-tuned LLM achieves higher segmentation accuracy than SHAS and yields consistent improvements in translation quality (BLEU, COMET) as well as latency (Average Lagging). Furthermore, our system benefits from IWSLT baselines for direct comparison. These findings highlight the potential of preference-tuned LLMs to surpass existing pretrained segmentation models and advance adaptive, human-aligned simultaneous interpretation.
☆ Not in Sync: Unveiling Temporal Bias in Audio Chat Models
Large Audio Language Models (LALMs) are increasingly applied to audio understanding and multimodal reasoning, yet their ability to locate when events occur remains underexplored. We present the first systematic study of temporal bias in LALMs, revealing a key limitation in their timestamp prediction. For example, when asked "At which second does the lecturer introduce the key formula?", models often predict timestamps that are consistently earlier or later than the ground truth. Through controlled experiments on timestamped datasets, we find that temporal bias (i) is prevalent across datasets and models, (ii) increases with audio length - even accumulating to tens of seconds in extended recordings, and (iii) varies across event types and positions. We quantify this effect with the Temporal Bias Index (TBI), measuring systematic misalignment in predicted event timings, and complement it with a visualization framework. Our findings highlight a fundamental limitation in current LALMs and call for the development of temporally robust architectures.
☆ From Knowledge to Treatment: Large Language Model Assisted Biomedical Concept Representation for Drug Repurposing EMNLP 2025
Drug repurposing plays a critical role in accelerating treatment discovery, especially for complex and rare diseases. Biomedical knowledge graphs (KGs), which encode rich clinical associations, have been widely adopted to support this task. However, existing methods largely overlook common-sense biomedical concept knowledge in real-world labs, such as mechanistic priors indicating that certain drugs are fundamentally incompatible with specific treatments. To address this gap, we propose LLaDR, a Large Language Model-assisted framework for Drug Repurposing, which improves the representation of biomedical concepts within KGs. Specifically, we extract semantically enriched treatment-related textual representations of biomedical entities from large language models (LLMs) and use them to fine-tune knowledge graph embedding (KGE) models. By injecting treatment-relevant knowledge into KGE, LLaDR largely improves the representation of biomedical concepts, enhancing semantic understanding of under-studied or complex indications. Experiments based on benchmarks demonstrate that LLaDR achieves state-of-the-art performance across different scenarios, with case studies on Alzheimer's disease further confirming its robustness and effectiveness. Code is available at https://github.com/xiaomingaaa/LLaDR.
comment: 16 pages, 4 figures, 13 tables. Accepted by EMNLP 2025 (Findings)
☆ Evolution of meta's llama models and parameter-efficient fine-tuning of large language models: a survey
This review surveys the rapid evolution of Meta AI's LLaMA (Large Language Model Meta AI) series - from LLaMA 1 through LLaMA 4 and the specialized parameter-efficient fine-tuning (PEFT) methods developed for these models. We first describe the LLaMA family of foundation models (7B-65B to 288B parameters), their architectures (including native multimodal and Mixtureof-Experts variants), and key performance characteristics. We then describe and discuss the concept of PEFT, which adapts large pre-trained models by updating only a small subset of parameters, and review five PEFT methods that have been applied to LLaMA: LoRA (Low-Rank Adaptation), LLaMA-Adapter V1 and V2, LLaMA-Excitor, and QLoRA (Quantized LoRA). We discuss each method's mechanism, parameter savings, and example application to LLaMA (e.g., instruction tuning, multimodal tasks). We provide structured discussion and analysis of model and adapter architectures, parameter counts, and benchmark results (including examples where fine-tuned LLaMA models outperform larger baselines). Finally, we examine real-world use cases where LLaMA-based models and PEFT have been successfully applied (e.g., legal and medical domains), and we discuss ongoing challenges and future research directions (such as scaling to even larger contexts and improving robustness). This survey paper provides a one-stop resource for ML researchers and practitioners interested in LLaMA models and efficient fine-tuning strategies.
☆ Towards Inference-time Scaling for Continuous Space Reasoning
Inference-time scaling through multiple sample generation in combination with Process- or Outcome-Reward Model (PRM or ORM) re-ranking has proven effective for text-based reasoning in large language models. This paper investigates whether such established techniques can be successfully adapted to reasoning in the continuous space, using COCONUT (Hao et al. 2024) continuous space reasoning LM as the backbone. We demonstrate the feasibility of generating diverse reasoning paths through dropout-based sampling. Our Pass@N analysis on the generated samples reveals the potential that could enable a significant gain in performance akin to observed gain in the discrete space. However, we highlight unique challenges faced for materializing this gain in the continuous thought space. In particular, working recipes for data generation and training PRM and ORM models in the discrete space unlocks only marginal improvements in the continuous space. Through probing various aspects including geometric properties and trajectory dynamics we identify the underlying reasons that prevent effective discrimination between correct and incorrect reasoning (essential for the functioning of PRM and ORM). Our findings reveal that current limitations stem from the absence of key inductive biases in continuous thought representations. We argue that the training frameworks for continuous reasoning LMs require not only to optimize for accuracy but also to explicitly incorporate inductive biases that could be utilized during inference-time for discrimination of correct and incorrect thoughts.\footnote{Our code and data will be publicly available.}
☆ A Survey on Parallel Reasoning
With the increasing capabilities of Large Language Models (LLMs), parallel reasoning has emerged as a new inference paradigm that enhances reasoning robustness by concurrently exploring multiple lines of thought before converging on a final answer. It has become a significant trend to explore parallel reasoning to overcome the fragility of standard sequential methods and improve practical performance. In this paper, we aim to survey and summarize the progress and challenges of parallel reasoning. We first present a formal definition of parallel reasoning and clarify its distinction from related concepts like Chain-of-Thought. Then, we organize and discuss advanced techniques based on a novel taxonomy, including non-interactive reasoning, interactive reasoning, and efficiency-focused decoding strategies. Additionally, we explore various application scenarios, such as solving complex problems and enhancing the reliability of LLM outputs.Finally, we highlight the core challenges of parallel reasoning and suggest potential directions for future research. We hope that our work can provide a useful roadmap for beginners and encourage more research on improving parallel reasoning methods. Related source can be avaliable in https://github.com/PPPP-kaqiu/Awesome-Parallel-Reasoning.
☆ Credal Transformer: A Principled Approach for Quantifying and Mitigating Hallucinations in Large Language Models
Large Language Models (LLMs) hallucinate, generating factually incorrect yet confident assertions. We argue this stems from the Transformer's Softmax function, which creates "Artificial Certainty" by collapsing ambiguous attention scores into a single probability distribution, discarding uncertainty information at each layer. To fix this, we introduce the Credal Transformer, which replaces standard attention with a Credal Attention Mechanism (CAM) based on evidential theory. CAM produces a "credal set" (a set of distributions) instead of a single attention vector, with the set's size directly measuring model uncertainty. We implement this by re-conceptualizing attention scores as evidence masses for a Dirichlet distribution: sufficient evidence recovers standard attention, while insufficient evidence yields a diffuse distribution, representing ambiguity. Empirically, the Credal Transformer identifies out-of-distribution inputs, quantifies ambiguity, and significantly reduces confident errors on unanswerable questions by abstaining. Our contribution is a new architecture to mitigate hallucinations and a design paradigm that integrates uncertainty quantification directly into the model, providing a foundation for more reliable AI.
☆ SafeMT: Multi-turn Safety for Multimodal Language Models
With the widespread use of multi-modal Large Language models (MLLMs), safety issues have become a growing concern. Multi-turn dialogues, which are more common in everyday interactions, pose a greater risk than single prompts; however, existing benchmarks do not adequately consider this situation. To encourage the community to focus on the safety issues of these models in multi-turn dialogues, we introduce SafeMT, a benchmark that features dialogues of varying lengths generated from harmful queries accompanied by images. This benchmark consists of 10,000 samples in total, encompassing 17 different scenarios and four jailbreak methods. Additionally, we propose Safety Index (SI) to evaluate the general safety of MLLMs during conversations. We assess the safety of 17 models using this benchmark and discover that the risk of successful attacks on these models increases as the number of turns in harmful dialogues rises. This observation indicates that the safety mechanisms of these models are inadequate for recognizing the hazard in dialogue interactions. We propose a dialogue safety moderator capable of detecting malicious intent concealed within conversations and providing MLLMs with relevant safety policies. Experimental results from several open-source models indicate that this moderator is more effective in reducing multi-turn ASR compared to existed guard models.
☆ Precise Attribute Intensity Control in Large Language Models via Targeted Representation Editing
Precise attribute intensity control--generating Large Language Model (LLM) outputs with specific, user-defined attribute intensities--is crucial for AI systems adaptable to diverse user expectations. Current LLM alignment methods, however, typically provide only directional or open-ended guidance, failing to reliably achieve exact attribute intensities. We address this limitation with three key designs: (1) reformulating precise attribute intensity control as a target-reaching problem, rather than simple maximization; (2) training a lightweight value function via temporal-difference learning to predict final attribute intensity scores from partial generations, thereby steering LLM outputs; and (3) employing gradient-based interventions on hidden representations to navigate the model precisely towards specific attribute intensity targets. Our method enables fine-grained, continuous control over attribute intensities, moving beyond simple directional alignment. Experiments on LLaMA-3.2-3b and Phi-4-mini confirm our method's ability to steer text generation to user-specified attribute intensities with high accuracy. Finally, we demonstrate efficiency enhancements across three downstream tasks: preference data synthesis, Pareto frontier approximation and optimization, and distillation of aligned behaviors for intervention-free inference. Our code is available on https://github.com/Pre-Control/pre-control
☆ Understanding the Modality Gap: An Empirical Study on the Speech-Text Alignment Mechanism of Large Speech Language Models EMNLP 2025
End-to-end Large Speech Language Models (LSLMs) have demonstrated impressive conversational generation abilities, yet consistently fall short of traditional pipeline systems on semantic understanding benchmarks. In this work, we reveal through systematic experimentation that although LSLMs lose some text input performance after speech-text alignment training, the performance gap between speech and text inputs is more pronounced, which we refer to as the modality gap. To understand this gap, we analyze both coarse- and fine-grained text and speech representations. At the coarse-grained level, representations of speech and text in deeper layers are found to be increasingly aligned in direction (cosine similarity), while concurrently diverging in magnitude (Euclidean distance). We further find that representation similarity is strongly correlated with the modality gap. At the fine-grained level, a spontaneous token-level alignment pattern between text and speech representations is observed. Based on this, we introduce the Alignment Path Score to quantify token-level alignment quality, which exhibits stronger correlation with the modality gap. Building on these insights, we design targeted interventions on critical tokens through angle projection and length normalization. These strategies demonstrate the potential to improve correctness for speech inputs. Our study provides the first systematic empirical analysis of the modality gap and alignment mechanisms in LSLMs, offering both theoretical and methodological guidance for future optimization.
comment: Accepted to EMNLP 2025 (Main Conference)
☆ Tracing Multilingual Knowledge Acquisition Dynamics in Domain Adaptation: A Case Study of English-Japanese Biomedical Adaptation
Multilingual domain adaptation (ML-DA) is widely used to learn new domain knowledge across languages into large language models (LLMs). Although many methods have been proposed to improve domain adaptation, the mechanisms of multilingual knowledge acquisition, how domain knowledge is learned within a language and transferred across languages, remain underexplored. This gap leads to suboptimal performance, particularly in low-resource settings. This work examines the learning dynamics of LLMs during ML-DA. Because prior ML-DA studies often train and evaluate on datasets with mismatched knowledge coverage, we propose AdaXEval, an adaptive evaluation method that builds multiple-choice QA datasets from the same bilingual domain corpus used for training, thereby directly studying multilingual knowledge acquisition. Through continual training of LLMs with diverse data recipes, we track how LLMs acquire domain facts and pinpoint the mechanism behind the transformation process from domain training data to knowledge. Our experiments on a 13B English-Japanese bilingual LLM reveal that cross-lingual transfer remains challenging despite a high-quality bilingual corpus. The code has been released.
comment: 22 Pages, Submitted to ARR 2025 Oct
☆ Deep Associations, High Creativity: A Simple yet Effective Metric for Evaluating Large Language Models
The evaluation of LLMs' creativity represents a crucial research domain, though challenges such as data contamination and costly human assessments often impede progress. Drawing inspiration from human creativity assessment, we propose PACE, asking LLMs to generate Parallel Association Chains to Evaluate their creativity. PACE minimizes the risk of data contamination and offers a straightforward, highly efficient evaluation, as evidenced by its strong correlation with Chatbot Arena Creative Writing rankings (Spearman's $\rho = 0.739$, $p < 0.001$) across various proprietary and open-source models. A comparative analysis of associative creativity between LLMs and humans reveals that while high-performing LLMs achieve scores comparable to average human performance, professional humans consistently outperform LLMs. Furthermore, linguistic analysis reveals that both humans and LLMs exhibit a trend of decreasing concreteness in their associations, and humans demonstrating a greater diversity of associative patterns.
comment: 14 pages
☆ One Life to Learn: Inferring Symbolic World Models for Stochastic Environments from Unguided Exploration
Symbolic world modeling requires inferring and representing an environment's transitional dynamics as an executable program. Prior work has focused on largely deterministic environments with abundant interaction data, simple mechanics, and human guidance. We address a more realistic and challenging setting, learning in a complex, stochastic environment where the agent has only "one life" to explore a hostile environment without human guidance. We introduce OneLife, a framework that models world dynamics through conditionally-activated programmatic laws within a probabilistic programming framework. Each law operates through a precondition-effect structure, activating in relevant world states. This creates a dynamic computation graph that routes inference and optimization only through relevant laws, avoiding scaling challenges when all laws contribute to predictions about a complex, hierarchical state, and enabling the learning of stochastic dynamics even with sparse rule activation. To evaluate our approach under these demanding constraints, we introduce a new evaluation protocol that measures (a) state ranking, the ability to distinguish plausible future states from implausible ones, and (b) state fidelity, the ability to generate future states that closely resemble reality. We develop and evaluate our framework on Crafter-OO, our reimplementation of the Crafter environment that exposes a structured, object-oriented symbolic state and a pure transition function that operates on that state alone. OneLife can successfully learn key environment dynamics from minimal, unguided interaction, outperforming a strong baseline on 16 out of 23 scenarios tested. We also test OneLife's planning ability, with simulated rollouts successfully identifying superior strategies. Our work establishes a foundation for autonomously constructing programmatic world models of unknown, complex environments.
comment: Project page: https://onelife-worldmodel.github.io/; 39 pages
☆ An AI-Based Behavioral Health Safety Filter and Dataset for Identifying Mental Health Crises in Text-Based Conversations
Large language models often mishandle psychiatric emergencies, offering harmful or inappropriate advice and enabling destructive behaviors. This study evaluated the Verily behavioral health safety filter (VBHSF) on two datasets: the Verily Mental Health Crisis Dataset containing 1,800 simulated messages and the NVIDIA Aegis AI Content Safety Dataset subsetted to 794 mental health-related messages. The two datasets were clinician-labelled and we evaluated performance using the clinician labels. Additionally, we carried out comparative performance analyses against two open source, content moderation guardrails: OpenAI Omni Moderation Latest and NVIDIA NeMo Guardrails. The VBHSF demonstrated, well-balanced performance on the Verily Mental Health Crisis Dataset v1.0, achieving high sensitivity (0.990) and specificity (0.992) in detecting any mental health crises. It achieved an F1-score of 0.939, sensitivity ranged from 0.917-0.992, and specificity was >= 0.978 in identifying specific crisis categories. When evaluated against the NVIDIA Aegis AI Content Safety Dataset 2.0, VBHSF performance remained highly sensitive (0.982) and accuracy (0.921) with reduced specificity (0.859). When compared with the NVIDIA NeMo and OpenAI Omni Moderation Latest guardrails, the VBHSF demonstrated superior performance metrics across both datasets, achieving significantly higher sensitivity in all cases (all p < 0.001) and higher specificity relative to NVIDIA NeMo (p < 0.001), but not to OpenAI Omni Moderation Latest (p = 0.094). NVIDIA NeMo and OpenAI Omni Moderation Latest exhibited inconsistent performance across specific crisis types, with sensitivity for some categories falling below 0.10. Overall, the VBHSF demonstrated robust, generalizable performance that prioritizes sensitivity to minimize missed crises, a crucial feature for healthcare applications.
comment: Main Text: 2943; Abstract: 256; Tables and Figures: 5
☆ ThinkPilot: Steering Reasoning Models via Automated Think-prefixes Optimization
Large Reasoning Models (LRMs) are powerful, but they still suffer from inefficient and off-target reasoning. Currently, training-free methods are limited to either rigid heuristics or descriptive, non-actionable analyses. In this paper, we introduce ThinkPilot, a training-free framework that automatically optimizes LRMs reasoning. It uses an evolutionary process to generate think-prefixes, which are instructions that evolve driven by a taxonomy of reasoning behaviors to guide models toward superior performance. Extensive experiments demonstrate ThinkPilot's broad effectiveness: it significantly improves the accuracy-length trade-off for efficient reasoning, drastically improves safety (for example, cutting the StrongREJECT score of DeepSeek-R1-Distill-Qwen-32B from 27.0% to 0.7), and enhances instruction following. It also synergizes with existing training-based methods. Our analysis reveals that think-prefixes can reliably control LRMs' reasoning behaviors, and that different tasks have strong preferences for specific behavioral distributions. By automatically identifying and eliciting these behaviors, ThinkPilot provides a generalizable framework for aligning LRMs reasoning with task demands. Data and code are available at https://github.com/teqkilla/ThinkPilot
☆ APCE: Adaptive Progressive Context Expansion for Long Context Processing NeurIPS 2025
Deploying useful Long-Context Transformer Models (LCTMs) requires addressing two key challenges: (1) A growing memory footprint due to quadratic self-attention and linear KV-cache scaling in memory as sequence length increases; (2) the ContextRot phenomena where empirical evidence suggests that transformer architecture's performance degrades with increasing context length. Given the shared dependency on the input, a natural question arises: Can we surgically select the most important input chunks for processing to synergistically (a) reduce the memory footprint, and (b) mitigate the ContextRot effects? In this paper, we answer this question in the affirmative for long-context summarization tasks. We propose APCE as a context-aware solution to select the most important input chunks through low-dimensional semantic similarity matching with the current query. By directly operating on the input, APCE decouples from strict dependency on underlying hardware or CUDA environments, promising a compatible solution scalable to different deployment systems. Our empirical evaluations have demonstrated superior or on-par summarization performance for APCE compared to the full dense baseline using a fraction (50%-70%) of the input sequence resulting in KV-cache and self-attention memory efficiency improvements. We hope our findings inspire further research on context-aware efficiency solutions for LCTMs geared towards other relevant long-context tasks.
comment: NeurIPS 2025 Workshop: ML For Systems
☆ Hierarchical Alignment: Surgical Fine-Tuning via Functional Layer Specialization in Large Language Models
Existing alignment techniques for Large Language Models (LLMs), such as Direct Preference Optimization (DPO), typically treat the model as a monolithic entity, applying uniform optimization pressure across all layers. This approach overlooks the functional specialization within the Transformer architecture, where different layers are known to handle distinct tasks from syntax to abstract reasoning. In this paper, we challenge this one-size-fits-all paradigm by introducing Hierarchical Alignment, a novel method that applies targeted DPO to distinct functional blocks of a model's layers: local (syntax), intermediate (logic), and global (factuality). Through a series of controlled experiments on state-of-the-art models like Llama-3.1-8B and Qwen1.5-7B using LoRA for surgical fine-tuning, our results, evaluated by a powerful LLM-as-Judge, demonstrate significant and predictable improvements. Specifically, aligning the local layers (Local-Align) enhances grammatical fluency. More importantly, aligning the global layers (Global-Align) not only improves factual consistency as hypothesized but also proves to be the most effective strategy for enhancing logical coherence, outperforming all baselines. Critically, all hierarchical strategies successfully avoid the "alignment tax" observed in standard DPO, where gains in fluency come at the cost of degraded logical reasoning. These findings establish a more resource-efficient, controllable, and interpretable path for model alignment, highlighting the immense potential of shifting from monolithic optimization to structure-aware surgical fine-tuning to build more advanced and reliable LLMs.
☆ Improving Text-to-Image Generation with Input-Side Inference-Time Scaling
Recent advances in text-to-image (T2I) generation have achieved impressive results, yet existing models often struggle with simple or underspecified prompts, leading to suboptimal image-text alignment, aesthetics, and quality. We propose a prompt rewriting framework that leverages large language models (LLMs) to refine user inputs before feeding them into T2I backbones. Our approach introduces a carefully designed reward system and an iterative direct preference optimization (DPO) training pipeline, enabling the rewriter to enhance prompts without requiring supervised fine-tuning data. We evaluate our method across diverse T2I models and benchmarks. Results show that our prompt rewriter consistently improves image-text alignment, visual quality, and aesthetics, outperforming strong baselines. Furthermore, we demonstrate strong transferability by showing that a prompt rewriter trained on one T2I backbone generalizes effectively to others without needing to be retrained. We also systematically study scalability, evaluating how performance gains scale with the capacity of the large LLM used as the rewriter. These findings highlight that prompt rewriting is an effective, scalable, and practical model-agnostic strategy for improving T2I systems. We plan to release the code and trained prompt rewriters soon.
☆ Uncertainty Quantification for Hallucination Detection in Large Language Models: Foundations, Methodology, and Future Directions
The rapid advancement of large language models (LLMs) has transformed the landscape of natural language processing, enabling breakthroughs across a wide range of areas including question answering, machine translation, and text summarization. Yet, their deployment in real-world applications has raised concerns over reliability and trustworthiness, as LLMs remain prone to hallucinations that produce plausible but factually incorrect outputs. Uncertainty quantification (UQ) has emerged as a central research direction to address this issue, offering principled measures for assessing the trustworthiness of model generations. We begin by introducing the foundations of UQ, from its formal definition to the traditional distinction between epistemic and aleatoric uncertainty, and then highlight how these concepts have been adapted to the context of LLMs. Building on this, we examine the role of UQ in hallucination detection, where quantifying uncertainty provides a mechanism for identifying unreliable generations and improving reliability. We systematically categorize a wide spectrum of existing methods along multiple dimensions and present empirical results for several representative approaches. Finally, we discuss current limitations and outline promising future research directions, providing a clearer picture of the current landscape of LLM UQ for hallucination detection.
comment: 24 pages, 3 figures, magazine
☆ On the Interplay between Human Label Variation and Model Fairness
The impact of human label variation (HLV) on model fairness is an unexplored topic. This paper examines the interplay by comparing training on majority-vote labels with a range of HLV methods. Our experiments show that without explicit debiasing, HLV training methods have a positive impact on fairness.
comment: 9 pages, 7 figures
☆ Multi-stage Prompt Refinement for Mitigating Hallucinations in Large Language Models
Recent advancements in large language models (LLMs) have shown strong performance in natural language understanding and generation tasks. However, LLMs continue to encounter challenges with hallucinations, where models generate plausible but incorrect information. While several factors contribute to hallucinations, the impact of ill-formed prompts, prompts with ambiguous wording, incorrect grammar, or incomplete information, was relatively under explored. To address this, we introduce Multi-stage Prompt Refinement (MPR), a framework designed to systematically improve these ill-formed prompts across multiple stages. Each stage addresses specific errors such as punctuation, typographical mistakes, and misuse of key terms, using small language models (SLMs) fine-tuned for these tasks. MPR iteratively enhances the clarity of prompts with additional context and employs a self-reflection mechanism with ranking to prioritize the most relevant input. Experimental results on hallucination benchmarks show that prompts refined by MPR achieve over an 85~\% win rate compared to their original forms, demonstrating its effectiveness in reducing hallucinations and improving LLM output accuracy. Interestingly, we reveal that MPR can be combined with existing post-hoc hallucination mitigation frameworks, further enhancing its versatility. MPR provides a lightweight and adaptable solution for enhancing LLM reliability across various domains.
comment: 22 pages, 6 figures
☆ CPR: Mitigating Large Language Model Hallucinations with Curative Prompt Refinement
Recent advancements in large language models (LLMs) highlight their fluency in generating responses to diverse prompts. However, these models sometimes generate plausible yet incorrect ``hallucinated" facts, undermining trust. A frequent but often overlooked cause of such errors is the use of poorly structured or vague prompts by users, leading LLMs to base responses on assumed rather than actual intentions. To mitigate hallucinations induced by these ill-formed prompts, we introduce Curative Prompt Refinement (CPR), a plug-and-play framework for curative prompt refinement that 1) cleans ill-formed prompts, and 2) generates additional informative task descriptions to align the intention of the user and the prompt using a fine-tuned small language model. When applied to language models, we discover that CPR significantly increases the quality of generation while also mitigating hallucination. Empirical studies show that prompts with CPR applied achieves over a 90\% win rate over the original prompts without any external knowledge.
comment: 2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 7 pages, 2 figures
☆ Information Extraction from Conversation Transcripts: Neuro-Symbolic vs. LLM
The current trend in information extraction (IE) is to rely extensively on large language models, effectively discarding decades of experience in building symbolic or statistical IE systems. This paper compares a neuro-symbolic (NS) and an LLM-based IE system in the agricultural domain, evaluating them on nine interviews across pork, dairy, and crop subdomains. The LLM-based system outperforms the NS one (F1 total: 69.4 vs. 52.7; core: 63.0 vs. 47.2), where total includes all extracted information and core focuses on essential details. However, each system has trade-offs: the NS approach offers faster runtime, greater control, and high accuracy in context-free tasks but lacks generalizability, struggles with contextual nuances, and requires significant resources to develop and maintain. The LLM-based system achieves higher performance, faster deployment, and easier maintenance but has slower runtime, limited control, model dependency and hallucination risks. Our findings highlight the "hidden cost" of deploying NLP systems in real-world applications, emphasizing the need to balance performance, efficiency, and control.
comment: 15 pages, 2 figures
♻ ☆ Model-Based Ranking of Source Languages for Zero-Shot Cross-Lingual Transfer EMNLP 2025
We present NN-Rank, an algorithm for ranking source languages for cross-lingual transfer, which leverages hidden representations from multilingual models and unlabeled target-language data. We experiment with two pretrained multilingual models and two tasks: part-of-speech tagging (POS) and named entity recognition (NER). We consider 51 source languages and evaluate on 56 and 72 target languages for POS and NER, respectively. When using in-domain data, NN-Rank beats state-of-the-art baselines that leverage lexical and linguistic features, with average improvements of up to 35.56 NDCG for POS and 18.14 NDCG for NER. As prior approaches can fall back to language-level features if target language data is not available, we show that NN-Rank remains competitive using only the Bible, an out-of-domain corpus available for a large number of languages. Ablations on the amount of unlabeled target data show that, for subsets consisting of as few as 25 examples, NN-Rank produces high-quality rankings which achieve 92.8% of the NDCG achieved using all available target data for ranking.
comment: Accepted to EMNLP 2025 (Main)
♻ ☆ GRDD: A Dataset for Greek Dialectal NLP
In this paper, we present a dataset for the computational study of a number of Modern Greek dialects. It consists of raw text data from four dialects of Modern Greek, Cretan, Pontic, Northern Greek and Cypriot Greek. The dataset is of considerable size, albeit imbalanced, and presents the first attempt to create large scale dialectal resources of this type for Modern Greek dialects. We then use the dataset to perform dialect idefntification. We experiment with traditional ML algorithms, as well as simple DL architectures. The results show very good performance on the task, potentially revealing that the dialects in question have distinct enough characteristics allowing even simple ML models to perform well on the task. Error analysis is performed for the top performing algorithms showing that in a number of cases the errors are due to insufficient dataset cleaning.
♻ ☆ OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding NeurIPS 2025
Speculative decoding generally dictates having a small, efficient draft model that is either pretrained or distilled offline to a particular target model series, for instance, Llama or Qwen models. However, within online deployment settings, there are two major challenges: 1) usage of a target model that is incompatible with the draft model; 2) expectation of latency improvements over usage and time. In this work, we propose OmniDraft, a unified framework that enables a single draft model to operate with any target model and adapt dynamically to user data. We introduce an online n-gram cache with hybrid distillation fine-tuning to address the cross-vocabulary mismatch across draft and target models; and further improve decoding speed by leveraging adaptive drafting techniques. OmniDraft is particularly suitable for on-device LLM applications where model cost, efficiency and user customization are the major points of contention. This further highlights the need to tackle the above challenges and motivates the \textit{``one drafter for all''} paradigm. We showcase the proficiency of the OmniDraft framework by performing online learning on math reasoning, coding and text generation tasks. Notably, OmniDraft enables a single Llama-68M model to pair with various target models including Vicuna-7B, Qwen2-7B and Llama3-8B models for speculative decoding; and additionally provides up to 1.5-2x speedup.
comment: Accepted to NeurIPS 2025
♻ ☆ Limited Preference Data? Learning Better Reward Model with Latent Space Synthesis NeurIPS 2025
Reward modeling, crucial for aligning large language models (LLMs) with human preferences, is often bottlenecked by the high cost of preference data. Existing textual data synthesis methods are computationally expensive. We propose a novel framework LENS for synthesizing preference data directly in the LLM's latent embedding space. Our method employs a Variational Autoencoder (VAE) to learn a structured latent representation of response embeddings. By performing controlled perturbations in this latent space and decoding back to the embedding space, we efficiently generate diverse, semantically consistent synthetic preference pairs, bypassing costly text generation and annotation. We provide theoretical guarantees that our synthesized pairs approximately preserve original preference ordering and improve reward model generalization. Empirically, our latent-space synthesis significantly outperforms text-based augmentation on standard benchmarks, achieving superior results while being 18x faster in generation and using a 16,000x smaller model. Our work offers a scalable and effective alternative for enhancing reward modeling through efficient data augmentation. Code is publicly available at https://github.com/deeplearning-wisc/lens
comment: Accepted by NeurIPS 2025
♻ ☆ ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference NeurIPS 2025
Large Language Models (LLMs) require significant GPU memory when processing long texts, with the key value (KV) cache consuming up to 70\% of total memory during inference. Although existing compression methods reduce memory by evaluating the importance of individual tokens, they overlook critical semantic relationships between tokens, resulting in fragmented context and degraded performance. We introduce ChunkKV, which fundamentally reimagines KV cache compression by treating semantic chunks - rather than isolated tokens - as basic compression units. This approach preserves complete linguistic structures and contextual integrity, ensuring that essential meaning is retained even under aggressive compression. Our innovation includes a novel layer-wise index reuse technique that exploits the higher cross-layer similarity of preserved indices in ChunkKV, reducing computational overhead and improving throughput by 26.5\%. Comprehensive evaluations on challenging benchmarks: LongBench, Needle-In-A-HayStack, GSM8K, and JailbreakV demonstrate that ChunkKV outperforms state-of-the-art methods by up to 8.7\% in precision while maintaining the same compression ratio. These results confirm that semantic-aware compression significantly enhances both efficiency and performance for long-context LLM inference, providing a simple yet effective solution to the memory bottleneck problem. The code is available at \href{https://github.com/NVIDIA/kvpress}{link}.
comment: NeurIPS 2025
♻ ☆ Language Modeling for the Future of Finance: A Survey into Metrics, Tasks, and Data Opportunities
Recent advances in language modeling have led to a growing number of papers related to finance in top-tier Natural Language Processing (NLP) venues. To systematically examine this trend, we review 374 NLP research papers published between 2017 and 2024 across 38 conferences and workshops, with a focused analysis of 221 papers that directly address finance-related tasks. We evaluate these papers across 11 quantitative and qualitative dimensions, and our study identifies the following opportunities for NLP researchers: (i) expanding the scope of forecasting tasks; (ii) enriching evaluation with financial metrics; (iii) leveraging multilingual and crisis-period datasets; and (iv) balancing PLMs with efficient or interpretable alternatives. We identify actionable directions supported by dataset and tool recommendations, with implications for both the academia and industry communities.
♻ ☆ Cognition-of-Thought Elicits Social-Aligned Reasoning in Large Language Models
Large language models (LLMs) excel at complex reasoning but can still exhibit harmful behaviors. Current alignment strategies typically embed safety into model weights, making these controls implicit, static, and difficult to modify. This paper introduces Cognition-of-Thought (CooT), a novel decoding-time framework that equips LLMs with an explicit cognitive self-monitoring loop. CooT couples a standard text Generator with a cognitive Perceiver that continuously monitors the unfolding sequence. The Perceiver uses a structured, precedence-based hierarchy of principles (e.g., safety over obedience) to detect potential misalignments as they arise. When violations are flagged, CooT intervenes by rolling back the generation to the point of error and regenerating under injected guidance that combines universal social priors with context-specific warnings. CooT thus transforms alignment from a fixed property into an explicit, dynamic, and auditable process active during inference, allowing for flexible policy updates without retraining the model. Extensive experiments across multiple benchmarks and model families confirm that CooT consistently improves safety and social reasoning performance.
♻ ☆ KnowledgeSmith: Uncovering Knowledge Updating in LLMs with Model Editing and Unlearning
Knowledge editing and machine unlearning are two popular approaches for large language models (LLMs) to stay up-to-date. However, the knowledge updating mechanism of LLMs remains largely unexplored due to insufficient, isolated, and small-scale evaluation. For instance, are LLMs similar to humans in modifying certain knowledge? What differs editing and unlearning as training data increases? This paper proposes KnowledgeSmith, a unified framework to systematically understand the updating mechanism of LLMs. We first cast editing and unlearning as instances of one constrained optimization problem. Then, we propose an automatic dataset generator that provides structured interventions across multiple graph levels and data scales, enabling controlled studies of how different modification strategies propagate through model knowledge. Extensive experiments demonstrate nuanced insights over knowledge propagation, plasticity scaling, consistency, and robustness. For instance, our results show that LLMs do not exhibit similar updating as humans for different levels of knowledge, and there exists consistency-capacity trade-off. We hope our findings can offer suggestions to the design of more reliable and scalable strategies. Code: https://github.com/AIFrontierLab/KnowledgeSmith.git
comment: Technical report
♻ ☆ Large language models management of medications: three performance analyses
Purpose: Large language models (LLMs) have proven performance for certain diagnostic tasks, however limited studies have evaluated their consistency in recommending appropriate medication regimens for a given diagnosis. Medication management is a complex task that requires synthesis of drug formulation and complete order instructions for safe use. Here, the performance of GPT 4o, an LLM available with ChatGPT, was tested for three medication management tasks. Methods: GPT-4o performance was tested using three medication tasks: identifying available formulations for a given generic drug name, identifying drug-drug interactions (DDI) for a given medication regimen, and preparing a medication order for a given generic drug name. For each experiment, the models raw text response was captured exactly as returned and evaluated using clinician evaluation in addition to standard LLM metrics, including Term Frequency-Inverse Document Frequency (TF IDF) vectors, normalized Levenshtein similarity, and Recall-Oriented Understudy for Gisting Evaluation (ROUGE 1/ROUGE L F1) between each response and its reference string. Results: For the first task of drug-formulation matching, GPT-4o had 49% accuracy for generic medications being matched to all available formulations, with an average of 1.23 omissions per medication and 1.14 hallucinations per medication. For the second task of drug-drug interaction identification, the accuracy was 54.7% for identifying the DDI pair. For the third task, GPT-4o generated order sentences containing no medication or abbreviation errors in 65.8% of cases. Conclusions: Model performance for basic medication tasks was consistently poor. This evaluation highlights the need for domain-specific training through clinician-annotated datasets and a comprehensive evaluation framework for benchmarking performance.
♻ ☆ Simulating and Understanding Deceptive Behaviors in Long-Horizon Interactions
Deception is a pervasive feature of human communication and an emerging concern in large language models (LLMs). While recent studies document instances of LLM deception under pressure, most evaluations remain confined to single-turn prompts and fail to capture the long-horizon interactions in which deceptive strategies typically unfold. We introduce the first simulation framework for probing and evaluating deception in LLMs under extended sequences of interdependent tasks and dynamic contextual pressures. Our framework instantiates a multi-agent system: a performer agent tasked with completing tasks and a supervisor agent that evaluates progress, provides feedback, and maintains evolving states of trust. An independent deception auditor then reviews full trajectories to identify when and how deception occurs. We conduct extensive experiments across 11 frontier models, spanning both closed- and open-source systems, and find that deception is model-dependent, increases with event pressure, and consistently erodes supervisor trust. Qualitative analyses further reveal distinct strategies of concealment, equivocation, and falsification. Our findings establish deception as an emergent risk in long-horizon interactions and provide a foundation for evaluating future LLMs in real-world, trust-sensitive contexts.
♻ ☆ MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems NeurIPS 2025
Human social interactions depend on the ability to infer others' unspoken intentions, emotions, and beliefs-a cognitive skill grounded in the psychological concept of Theory of Mind (ToM). While large language models (LLMs) excel in semantic understanding tasks, they struggle with the ambiguity and contextual nuance inherent in human communication. To bridge this gap, we introduce MetaMind, a multi-agent framework inspired by psychological theories of metacognition, designed to emulate human-like social reasoning. MetaMind decomposes social understanding into three collaborative stages: (1) a Theory-of-Mind Agent generates hypotheses about user mental states (e.g., intent, emotion), (2) a Moral Agent refines these hypotheses using cultural norms and ethical constraints, and (3) a Response Agent generates contextually appropriate responses while validating alignment with inferred intent. Our framework achieves state-of-the-art performance across three challenging benchmarks, with 35.7% improvement in real-world social scenarios and 6.2% gain in ToM reasoning. Notably, it enables LLMs to match human-level performance on key ToM tasks for the first time. Ablation studies confirm the necessity of all components, which showcase the framework's ability to balance contextual plausibility, social appropriateness, and user adaptation. This work advances AI systems toward human-like social intelligence, with applications in empathetic dialogue and culturally sensitive interactions. Code is available at https://github.com/XMZhangAI/MetaMind.
comment: NeurIPS 2025 Spotlight
♻ ☆ Clean First, Align Later: Benchmarking Preference Data Cleaning for Reliable LLM Alignment NeurIPS 2025
Human feedback plays a pivotal role in aligning large language models (LLMs) with human preferences. However, such feedback is often noisy or inconsistent, which can degrade the quality of reward models and hinder alignment. While various automated data cleaning methods have been proposed to mitigate this issue, a systematic evaluation of their effectiveness and generalizability remains lacking. To bridge this gap, we introduce the first comprehensive benchmark for evaluating 13 preference data cleaning methods in the context of LLM alignment. PrefCleanBench offers a standardized protocol to assess cleaning strategies in terms of alignment performance and generalizability across diverse datasets, model architectures, and optimization algorithms. By unifying disparate methods and rigorously comparing them, we uncover key factors that determine the success of data cleaning in alignment tasks. This benchmark lays the groundwork for principled and reproducible approaches to improving LLM alignment through better data quality-highlighting the crucial but underexplored role of data preprocessing in responsible AI development. We release modular implementations of all methods to catalyze further research: https://github.com/deeplearning-wisc/PrefCleanBench.
comment: NeurIPS 2025
♻ ☆ Knowledge Fusion via Bidirectional Information Aggregation
Knowledge graphs (KGs) are the cornerstone of the semantic web, offering up-to-date representations of real-world entities and relations. Yet large language models (LLMs) remain largely static after pre-training, causing their internal knowledge to become outdated and limiting their utility in time-sensitive web applications. To bridge this gap between dynamic knowledge and static models, a prevalent approach is to enhance LLMs with KGs. However, prevailing methods typically rely on parameter-invasive fine-tuning, which risks catastrophic forgetting and often degrades LLMs' general capabilities. Moreover, their static integration frameworks cannot keep pace with the continuous evolution of real-world KGs, hindering their deployment in dynamic web environments. To bridge this gap, we introduce KGA (\textit{\underline{K}nowledge \underline{G}raph-guided \underline{A}ttention}), a novel framework that dynamically integrates external KGs into LLMs exclusively at inference-time without any parameter modification. Inspired by research on neuroscience, we rewire the self-attention module by innovatively introducing two synergistic pathways: a \textit{bottom-up knowledge fusion} pathway and a \textit{top-down attention guidance} pathway. The \textit{bottom-up pathway} dynamically integrates external knowledge into input representations via input-driven KG fusion, which is akin to the \textit{stimulus-driven attention process} in the human brain. Complementarily, the \textit{top-down pathway} aims to assess the contextual relevance of each triple through a \textit{goal-directed verification process}, thereby suppressing task-irrelevant signals and amplifying knowledge-relevant patterns. By synergistically combining these two pathways, our method supports real-time knowledge fusion. Extensive experiments on four benchmarks verify KGA's strong fusion performance and efficiency.
♻ ☆ General Exploratory Bonus for Optimistic Exploration in RLHF
Optimistic exploration is central to improving sample efficiency in reinforcement learning with human feedback, yet existing exploratory bonus methods to incentivize exploration often fail to realize optimism. We provide a theoretical analysis showing that current formulations, under KL or $\alpha$-divergence regularization, unintentionally bias exploration toward high-probability regions of the reference model, thereby reinforcing conservative behavior instead of promoting discovery of uncertain regions. To address this pitfall, we introduce the General Exploratory Bonus (GEB), a novel theoretical framework that provably satisfies the optimism principle. GEB counteracts divergence-induced bias via reference-dependent reward regulation and unifies prior heuristic bonuses as special cases, while extending naturally across the full $\alpha$-divergence family. Empirically, GEB consistently outperforms baselines on alignment tasks across multiple divergence settings and large language model backbones. These results demonstrate that GEB offers both a principled and practical solution for optimistic exploration in RLHF.
♻ ☆ FlagEval Findings Report: A Preliminary Evaluation of Large Reasoning Models on Automatically Verifiable Textual and Visual Questions NeurIPS 2025
We conduct a moderate-scale contamination-free (to some extent) evaluation of current large reasoning models (LRMs) with some preliminary findings. We also release ROME, our evaluation benchmark for vision language models intended to test reasoning from visual clues. We attach links to the benchmark, evaluation data, and other updates on this website: https://flageval-baai.github.io/LRM-Eval/
comment: Project homepage: https://flageval-baai.github.io/LRM-Eval/ This work will also be presented at NeurIPS 2025 Workshop on Foundations of Reasoning in Language Models (FoRLM)
♻ ☆ Leveraging Importance Sampling to Detach Alignment Modules from Large Language Models NeurIPS 2025
The widespread adoption of large language models (LLMs) across industries has increased the demand for high-quality and customizable outputs. However, traditional alignment methods often require retraining large pretrained models, making it difficult to quickly adapt and optimize LLMs for diverse applications. To address this limitation, we propose a novel \textit{Residual Alignment Model} (\textit{RAM}) that formalizes the alignment process as a type of importance sampling. In this framework, the unaligned upstream model serves as the proposal distribution, while the alignment process is framed as secondary sampling based on an autoregressive alignment module that acts as an estimator of the importance weights. This design enables a natural detachment of the alignment module from the target aligned model, improving flexibility and scalability. Based on this model, we derive an efficient sequence-level training strategy for the alignment module, which operates independently of the proposal module. Additionally, we develop a resampling algorithm with iterative token-level decoding to address the common first-token latency issue in comparable methods. Experimental evaluations on two leading open-source LLMs across diverse tasks, including instruction following, domain adaptation, and preference optimization, demonstrate that our approach consistently outperforms baseline models.
comment: Accepted by NeurIPS 2025, 28 pages
♻ ☆ DeepDive: Advancing Deep Search Agents with Knowledge Graphs and Multi-Turn RL
Augmenting large language models (LLMs) with browsing tools substantially improves their potential as deep search agents to solve complex, real-world tasks. Yet, open LLMs still perform poorly in such settings due to limited long-horizon reasoning capacity with browsing tools and the lack of sufficiently difficult supervised data. To address these challenges, we present DeepDive to advance deep search agents. First, we propose a strategy to automatically synthesize complex, difficult, and hard-to-find questions from open knowledge graphs. Second, we apply end-to-end multi-turn reinforcement learning (RL) to enhance LLMs' long-horizon reasoning with deep search. To encourage diversity and reduce redundancy, we design a redundancy penalty that discourages repeated similar queries. Experiments show that DeepDive-32B achieves a new open-source competitive result on BrowseComp, outperforming WebSailor, DeepSeek-R1-Browse, and Search-o1. We demonstrate that multi-turn RL training improves deep search ability and significantly contributes to the performance improvements across multiple benchmarks. We observe that DeepDive enables test-time scaling of tool calls and parallel sampling. All datasets, models, and code are publicly available at https://github.com/THUDM/DeepDive.
♻ ☆ Assessing Latency in ASR Systems: A Methodological Perspective for Real-Time Use
Automatic speech recognition (ASR) systems generate real-time transcriptions but often miss nuances that human interpreters capture. While ASR is useful in many contexts, interpreters-who already use ASR tools such as Dragon-add critical value, especially in sensitive settings such as diplomatic meetings where subtle language is key. Human interpreters not only perceive these nuances but can adjust in real time, improving accuracy, while ASR handles basic transcription tasks. However, ASR systems introduce a delay that does not align with real-time interpretation needs. The user-perceived latency of ASR systems differs from that of interpretation because it measures the time between speech and transcription delivery. To address this, we propose a new approach to measuring delay in ASR systems and validate if they are usable in live interpretation scenarios.
comment: 8 pages, 2 figures
♻ ☆ Lost at the Beginning of Reasoning
Recent advancements in large language models (LLMs) have significantly advanced complex reasoning capabilities, particularly through extended chain-of-thought (CoT) reasoning that incorporates mechanisms such as backtracking, self-reflection, and self-correction. Despite these developments, the self-correction abilities of LLMs during long CoT reasoning remain underexplored. And recent findings on overthinking suggest that such models often engage in unnecessarily redundant reasoning. In this work, we empirically show that the first reasoning step exerts a disproportionately large influence on the final prediction. I.e., errors introduced at this stage can substantially degrade subsequent reasoning quality. This phenomenon is consistently observed across various state-of-the-art open- and closed-source reasoning models. Leveraging this insight, we propose an efficient sampling strategy that leverages a reward model to identify and retain high-quality first reasoning steps while discarding suboptimal ones, achieving up to a 70% reduction in inference cost without sacrificing any accuracy. Our work highlights the central role of the first reasoning step in generating a high-quality reasoning trajectory, and thus enabling significantly efficient sampling.
comment: remove the benchmark part. (10 pages, 6 figures, 5 tables)
♻ ☆ Persuasion at Play: Understanding Misinformation Dynamics in Demographic-Aware Human-LLM Interactions
Existing challenges in misinformation exposure and susceptibility vary across demographic groups, as some populations are more vulnerable to misinformation than others. Large language models (LLMs) introduce new dimensions to these challenges through their ability to generate persuasive content at scale and reinforcing existing biases. This study investigates the bidirectional persuasion dynamics between LLMs and humans when exposed to misinformative content. We analyze human-to-LLM influence using human-stance datasets and assess LLM-to-human influence by generating LLM-based persuasive arguments. Additionally, we use a multi-agent LLM framework to analyze the spread of misinformation under persuasion among demographic-oriented LLM agents. Our findings show that demographic factors influence susceptibility to misinformation in LLMs, closely reflecting the demographic-based patterns seen in human susceptibility. We also find that, similar to human demographic groups, multi-agent LLMs exhibit echo chamber behavior. This research explores the interplay between humans and LLMs, highlighting demographic differences in the context of misinformation and offering insights for future interventions.
♻ ☆ The Distribution of Dependency Distance and Hierarchical Distance in Contemporary Written Japanese and Its Influencing Factors
To explore the relationship between dependency distance (DD) and hierarchical distance (HD) in Japanese, we compared the probability distributions of DD and HD with and without sentence length fixed, and analyzed the changes in mean dependency distance (MDD) and mean hierarchical distance (MHD) as sentence length increases, along with their correlation coefficient based on the Balanced Corpus of Contemporary Written Japanese. It was found that the valency of the predicates is the underlying factor behind the trade-off relation between MDD and MHD in Japanese. Native speakers of Japanese regulate the linear complexity and hierarchical complexity through the valency of the predicates, and the relative sizes of MDD and MHD depend on whether the threshold of valency has been reached. Apart from the cognitive load, the valency of the predicates also affects the probability distributions of DD and HD. The effect of the valency of the predicates on the distribution of HD is greater than on that of DD, which leads to differences in their probability distributions and causes the mean of MDD to be lower than that of MHD.
comment: This paper has been accepted by the 13th International Quantitative Linguistics Conference QUALICO 2025
♻ ☆ Exploring the Frontier of Vision-Language Models: A Survey of Current Methodologies and Future Directions
The advent of Large Language Models (LLMs) has significantly reshaped the trajectory of the AI revolution. Nevertheless, these LLMs exhibit a notable limitation, as they are primarily adept at processing textual information. To address this constraint, researchers have endeavored to integrate visual capabilities with LLMs, resulting in the emergence of Vision-Language Models (VLMs). These advanced models are instrumental in tackling more intricate tasks such as image captioning and visual question answering. In our comprehensive survey paper, we delve into the key advancements within the realm of VLMs. Our classification organizes VLMs into three distinct categories: models dedicated to vision-language understanding, models that process multimodal inputs to generate unimodal (textual) outputs and models that both accept and produce multimodal inputs and outputs.This classification is based on their respective capabilities and functionalities in processing and generating various modalities of data.We meticulously dissect each model, offering an extensive analysis of its foundational architecture, training data sources, as well as its strengths and limitations wherever possible, providing readers with a comprehensive understanding of its essential components. We also analyzed the performance of VLMs in various benchmark datasets. By doing so, we aim to offer a nuanced understanding of the diverse landscape of VLMs. Additionally, we underscore potential avenues for future research in this dynamic domain, anticipating further breakthroughs and advancements.
comment: One of the first survey on Visual Language Models
♻ ☆ SCAN: Self-Denoising Monte Carlo Annotation for Robust Process Reward Learning NeurIPS 2025
Process reward models (PRMs) offer fine-grained, step-level evaluations that facilitate deeper reasoning processes in large language models (LLMs), proving effective in complex tasks like mathematical reasoning. However, developing PRMs is challenging due to the high cost and limited scalability of human-annotated data. Synthetic data from Monte Carlo (MC) estimation is a promising alternative but suffers from a high noise ratio, which can cause overfitting and hinder large-scale training. In this work, we conduct a preliminary study on the noise distribution in synthetic data from MC estimation, identifying that annotation models tend to both underestimate and overestimate step correctness due to limitations in their annotation capabilities. Building on these insights, we propose Self-Denoising Monte Carlo Annotation (SCAN), an efficient data synthesis and noise-tolerant learning framework. Our key findings indicate that: (1) Even lightweight models (e.g., 1.5B parameters) can produce high-quality annotations through a self-denoising strategy, enabling PRMs to achieve superior performance with only 6% the inference cost required by vanilla MC estimation. (2) With our robust learning strategy, PRMs can effectively learn from this weak supervision, achieving a 39.2 F1 score improvement (from 19.9 to 59.1) in ProcessBench. Despite using only a compact synthetic dataset, our models surpass strong baselines, including those trained on large-scale human-annotated datasets such as PRM800K. Furthermore, performance continues to improve as we scale up the synthetic data, highlighting the potential of SCAN for scalable, cost-efficient, and robust PRM training.
comment: NeurIPS 2025. Project page: https://scan-prm.github.io/
♻ ☆ Triplet-Structured Knowledge Integration for Multi-Turn Medical Reasoning
Large Language Models (LLMs) have shown strong performance on static medical Question Answering (QA) tasks, yet their reasoning often deteriorates in multi-turn clinical dialogues where patient information is scattered across turns. This paper introduces TriMediQ, a triplet-structured approach that enhances the reasoning reliability of LLMs through explicit knowledge integration. TriMediQ first employs a frozen triplet extraction LLM to convert patient responses into clinically grounded triplets, ensuring factual precision via constrained prompting. These triplets are incorporated into a patient-specific Knowledge Graph (KG), from which a trainable projection module consisting of a graph encoder and a projector captures relational dependencies while keeping all LLM parameters frozen. During inference, the projection module guides multi-hop reasoning over the KG, enabling coherent clinical dialogue understanding. Experiments on two interactive medical QA benchmarks show that TriMediQ achieves up to 10.4\% improvement in accuracy over five existing baselines on the iMedQA dataset. These results demonstrate that structuring patient information as triplets can effectively improve the reasoning capability of LLMs in multi-turn medical QA.
comment: Preprint
♻ ☆ AGENTIQL: An Agent-Inspired Multi-Expert Framework for Text-to-SQL Generation NeurIPS 2025
LLMs have advanced text-to-SQL generation, yet monolithic architectures struggle with complex reasoning and schema diversity. We propose AGENTIQL, an agent-inspired multi-expert framework that combines a reasoning agent for question decomposition, a coding agent for sub-query generation, and a refinement step for column selection. An adaptive router further balances efficiency and accuracy by selecting between our modular pipeline and a baseline parser. Several steps in the pipeline can be executed in parallel, making the framework scalable to larger workloads. Evaluated on the Spider benchmark, AGENTIQL improves execution accuracy and interpretability and achieves up to 86.07% EX with 14B models using the Planner&Executor merging strategy. The attained performance is contingent upon the efficacy of the routing mechanism, thereby narrowing the gap to GPT-4-based SOTA (89.65% EX) while using much smaller open-source LLMs. Beyond accuracy, AGENTIQL enhances transparency by exposing intermediate reasoning steps, offering a robust, scalable, and interpretable approach to semantic parsing.
comment: Accepted at NeurIPS 2025, ER "Efficient Reasoning" workshop
♻ ☆ A Comprehensive Taxonomy of Negation for NLP and Neural Retrievers
Understanding and solving complex reasoning tasks is vital for addressing the information needs of a user. Although dense neural models learn contextualised embeddings, they still underperform on queries containing negation. To understand this phenomenon, we study negation in both traditional neural information retrieval and LLM-based models. We (1) introduce a taxonomy of negation that derives from philosophical, linguistic, and logical definitions; (2) generate two benchmark datasets that can be used to evaluate the performance of neural information retrieval models and to fine-tune models for a more robust performance on negation; and (3) propose a logic-based classification mechanism that can be used to analyze the performance of retrieval models on existing datasets. Our taxonomy produces a balanced data distribution over negation types, providing a better training setup that leads to faster convergence on the NevIR dataset. Moreover, we propose a classification schema that reveals the coverage of negation types in existing datasets, offering insights into the factors that might affect the generalization of fine-tuned models on negation.
♻ ☆ A Survey of Multilingual Reasoning in Language Models EMNLP
While reasoning and multilingual capabilities in language models (LMs) have achieved remarkable progress in recent years, their integration into a unified paradigm - multilingual reasoning - is at a nascent stage. Multilingual reasoning requires language models to handle logical reasoning across languages while addressing misalignment, biases, and challenges in low-resource settings. This survey provides the first in-depth review of multilingual reasoning in LMs. In this survey, we provide a systematic overview of existing methods that leverage LMs for multilingual reasoning, specifically outlining the challenges, motivations, and foundational aspects of applying language models to reason across diverse languages. We provide an overview of the standard data resources used for training multilingual reasoning in LMs and the evaluation benchmarks employed to assess their multilingual capabilities. Next, we analyze various state-of-the-art methods and their performance on these benchmarks. Finally, we explore future research opportunities to improve multilingual reasoning in LMs, focusing on enhancing their ability to handle diverse languages and complex reasoning tasks. Rapid growth of evolving developments in this field can be actively tracked on our project page: [https://github.com/AkashGhosh/Survey-of-Multilingual-Reasoning-in-Language-Models](https://github.com/AkashGhosh/Survey-of-Multilingual-Reasoning-in-Language-Models)
comment: EMNLP Findings 2025
♻ ☆ MaxPoolBERT: Enhancing BERT Classification via Layer- and Token-Wise Aggregation
The [CLS] token in BERT is commonly used as a fixed-length representation for classification tasks, yet prior work has shown that both other tokens and intermediate layers encode valuable contextual information. In this work, we study lightweight extensions to BERT that refine the [CLS] representation by aggregating information across layers and tokens. Specifically, we explore three modifications: (i) max-pooling the [CLS] token across multiple layers, (ii) enabling the [CLS] token to attend over the entire final layer using an additional multi-head attention (MHA) layer, and (iii) combining max-pooling across the full sequence with MHA. Our approach, called MaxPoolBERT, enhances BERT's classification accuracy (especially on low-resource tasks) without requiring new pre-training or significantly increasing model size. Experiments on the GLUE benchmark show that MaxPoolBERT consistently achieves a better performance than the standard BERT base model on low resource tasks of the GLUE benchmark.
♻ ☆ DiaCDM: Cognitive Diagnosis in Teacher-Student Dialogues using the Initiation-Response-Evaluation Framework
While cognitive diagnosis (CD) effectively assesses students' knowledge mastery from structured test data, applying it to real-world teacher-student dialogues presents two fundamental challenges. Traditional CD models lack a suitable framework for handling dynamic, unstructured dialogues, and it's difficult to accurately extract diagnostic semantics from lengthy dialogues. To overcome these hurdles, we propose DiaCDM, an innovative model. We've adapted the initiation-response-evaluation (IRE) framework from educational theory to design a diagnostic framework tailored for dialogue. We also developed a unique graph-based encoding method that integrates teacher questions with relevant knowledge components to capture key information more precisely. To our knowledge, this is the first exploration of cognitive diagnosis in a dialogue setting. Experiments on three real-world dialogue datasets confirm that DiaCDM not only significantly improves diagnostic accuracy but also enhances the results' interpretability, providing teachers with a powerful tool for assessing students' cognitive states. The code is available at https://github.com/Mind-Lab-ECNU/DiaCDM/tree/main.
♻ ☆ CiteBART: Learning to Generate Citations for Local Citation Recommendation EMNLP 2025
Local citation recommendation (LCR) suggests a set of papers for a citation placeholder within a given context. The task has evolved as generative approaches have become more promising than the traditional pre-fetch and re-rank-based state-of-the-art approaches. This paper introduces citation-specific pre-training within an encoder-decoder architecture, where author-date citation tokens are masked to learn to reconstruct them to fulfill LCR. There are two variants for this pre-training. In the local context-only base scheme (CiteBART-Base), the citation token in a local context is masked to learn to predict the citation. The global version (CiteBART-Global) extends the local context with the citing paper's title and abstract to enrich the learning signal. CiteBART-Global achieves state-of-the-art performance on LCR benchmarks except for the FullTextPeerRead dataset, which is quite small to see the advantage of generative pre-training. The effect is significant in the larger benchmarks, e.g., Refseer and ArXiv., with the Refseer benchmark-trained model emerging as the best-performing model. We perform comprehensive experiments, including an ablation study, a qualitative analysis, and a taxonomy of hallucinations with detailed statistics. Our analyses confirm that CiteBART-Global has a cross-dataset generalization capability; the macro hallucination rate (MaHR) at the top-3 predictions is 4\%, and when the ground-truth is in the top-k prediction list, the hallucination tendency in the other predictions drops significantly.
comment: This paper has been accepted to the EMNLP 2025 Main Conference. (19 pages, 3 figures, 11 tables)
♻ ☆ DRIFT: Decompose, Retrieve, Illustrate, then Formalize Theorems
Automating the formalization of mathematical statements for theorem proving remains a major challenge for Large Language Models (LLMs). LLMs struggle to identify and utilize the prerequisite mathematical knowledge and its corresponding formal representation in languages like Lean. Current retrieval-augmented autoformalization methods query external libraries using the informal statement directly, but overlook a fundamental limitation: informal mathematical statements are often complex and offer limited context on the underlying math concepts. To address this, we introduce DRIFT, a novel framework that enables LLMs to decompose informal mathematical statements into smaller, more tractable ''sub-components''. This facilitates targeted retrieval of premises from mathematical libraries such as Mathlib. Additionally, DRIFT retrieves illustrative theorems to help models use premises more effectively in formalization tasks. We evaluate DRIFT across diverse benchmarks (ProofNet, ConNF, and MiniF2F-test) and find that it consistently improves premise retrieval, nearly doubling the F1 score compared to the DPR baseline on ProofNet. Notably, DRIFT demonstrates strong performance on the out-of-distribution ConNF benchmark, with BEq+@10 improvements of 37.14% and 42.25% using GPT-4.1 and DeepSeek-V3.1, respectively. Our analysis shows that retrieval effectiveness in mathematical autoformalization depends heavily on model-specific knowledge boundaries, highlighting the need for adaptive retrieval strategies aligned with each model's capabilities.
♻ ☆ Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning
Large Reasoning Models (LRMs) have achieved remarkable performance on complex reasoning tasks by adopting the "think-then-answer" paradigm, which enhances both accuracy and interpretability. However, current LRMs exhibit two critical limitations when processing non-English languages: (1) They often struggle to maintain input-output language consistency; (2) They generally perform poorly with wrong reasoning paths and lower answer accuracy compared to English. These limitations significantly degrade the user experience for non-English speakers and hinder the global deployment of LRMs. To address these limitations, we propose M-Thinker, which is trained by the GRPO algorithm that involves a Language Consistency (LC) reward and a novel Cross-lingual Thinking Alignment (CTA) reward. Specifically, the LC reward defines a strict constraint on the language consistency between the input, thought, and answer. Besides, the CTA reward compares the model's non-English reasoning paths with its English reasoning path to transfer its own reasoning capability from English to non-English languages. Through an iterative RL procedure, our M-Thinker-1.5B/7B models not only achieve nearly 100% language consistency and superior performance on two multilingual benchmarks (MMATH and PolyMath), but also exhibit excellent generalization on out-of-domain languages.
comment: 13 pages, 8 tables, 4 figures. Code is available at: https://github.com/XZhang00/M-Thinker
♻ ☆ Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
A key challenge in applying reinforcement learning (RL) to diffusion large language models (dLLMs) lies in the intractability of their likelihood functions, which are essential for the RL objective, necessitating corresponding approximation in each training step. While existing methods approximate the log-likelihoods by their evidence lower bounds (ELBOs) via customized Monte Carlo (MC) sampling, the forward computational graphs of all MC samples need to be retained for the gradient computation of non-linear terms in the RL objective, resulting in significant memory overhead. This constraint restricts feasible sample sizes, leading to imprecise likelihood approximations and ultimately distorting the RL objective. To overcome this limitation, we propose \emph{Boundary-Guided Policy Optimization} (BGPO), a memory-efficient RL algorithm that maximizes a specially constructed lower bound of the ELBO-based objective. This lower bound is carefully designed to satisfy two key properties: (1) Linearity: it is formulated in a linear sum where each term depends only on a single MC sample, thereby enabling gradient accumulation across samples and ensuring constant memory usage; (2) Equivalence: Both the value and gradient of this lower bound are equal to those of the ELBO-based objective in on-policy training, making it also an effective approximation for the original RL objective. These properties allow BGPO to adopt a large MC sample size, resulting in more accurate likelihood approximations and improved RL objective estimation, which in turn leads to enhanced performance. Experiments show that BGPO significantly outperforms previous RL algorithms for dLLMs in math problem solving, code generation, and planning tasks. Our codes and models are available at \href{https://github.com/THU-KEG/BGPO}{https://github.com/THU-KEG/BGPO}.
♻ ☆ SAFER: Probing Safety in Reward Models with Sparse Autoencoder
Reinforcement learning from human feedback (RLHF) is a key paradigm for aligning large language models (LLMs) with human values, yet the reward models at its core remain largely opaque. In this work, we present sparse Autoencoder For Enhanced Reward model (\textbf{SAFER}), a novel framework for interpreting and improving reward models through mechanistic analysis. Leveraging Sparse Autoencoders (SAEs), we uncover human-interpretable features in reward model activations, enabling insight into safety-relevant decision-making. We apply SAFER to safety-oriented preference datasets and quantify the salience of individual features by activation differences between chosen and rejected responses. Using these feature-level signals, we design targeted data poisoning and denoising strategies. Experiments show that SAFER can precisely degrade or enhance safety alignment with minimal data modification, without sacrificing general chat performance. Our approach contributes to interpreting, auditing and refining reward models in high-stakes LLM alignment tasks. Our codes are available at https://github.com/xzy-101/SAFER-code. \textit{This paper discusses topics related to large language model safety and may include discussions or examples that highlight potential risks or unsafe outcomes.}
comment: One of the institutions requires additional approval before we can move forward with the publication. Thanks for your understanding, and we hope to resubmit once everything is finalized
♻ ☆ LazyEviction: Lagged KV Eviction with Attention Pattern Observation for Efficient Long Reasoning
Large Language Models (LLMs) exhibit enhanced capabilities by Chain-of-Thought reasoning. However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a \textbf{Token Importance Recurrence} phenomenon: a large proportion of tokens regain high attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose \textbf{LazyEviction}, an observation window-based lagged eviction framework retaining latent recurring tokens by prioritized eviction based on tokens' recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache by 50\%\textasciitilde70\% while maintaining comparable accuracy, outperforming existing KV cache compression baselines. Our implementation code can be found at https://github.com/Halo-949/LazyEviction.
♻ ☆ The Price of a Second Thought: On the Evaluation of Reasoning Efficiency in Large Language Models
Recent thinking models trained with reinforcement learning and backward-checking CoT often suffer from overthinking: they produce excessively long outputs even on simple problems, wasting computation. Existing evaluations, based on token efficiency, give an incomplete view as they neglect problem difficulty and intermediate computation costs. We formalize reasoning efficiency as a relative measure between thinking and instruct models, treating instruct models as the minimal-effort baseline. A systematic study across four thinking models and multiple benchmarks reveals two consistent patterns: (i) instruct models achieve higher efficiency overall, and (ii) problem difficulty affects efficiency, with thinking models wasting computation on easy problems but providing value on harder ones. Building on this insight, we propose COTHINK, a simple two-stage pipeline: an instruct model drafts a brief outline, and a thinking model expands it. On GSM8K, MATH500, and AIME24, COTHINK cuts token usage by 21.1% while keeping accuracy on four thinking models, and remains competitive with strong efficiency baselines.
comment: Added new experiments and revised the manuscript for clarity
♻ ☆ BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions
Efficiently solving real-world problems with LLMs increasingly hinges on their ability to interact with dynamic web environments and autonomously acquire external information. While recent research like Search-R1 and WebDancer demonstrates strong performance in solving web tasks, they heavily rely on additional tools to convert the interactive web environment into static text content. This is in contrast to human browsing behaviors, which involve diverse interactions with the browser, such as scrolling, clicking, and typing. In this paper, we propose BrowserAgent, a more interactive agent that solves complex tasks through human-inspired browser actions. BrowserAgent operates directly on raw web pages via Playwright through a set of predefined browser actions. We adopt a two-stage training (Supervised Fine-Tuning (SFT) and Rejection Fine-Tuning (RFT)) to improve the model's generalization abilities. Despite using significantly less training data than Search-R1, BrowserAgent achieves more competitive results across different Open-QA tasks. Additionally, we introduce an explicit memory mechanism to store key conclusions across steps, further enhancing the model's reasoning capabilities for long-horizon tasks. Notably, BrowserAgent-7B can achieve around 20\% improvement over Search-R1 on multi-hop QA tasks like HotpotQA, 2Wiki, and Bamboogle. These results indicate that BrowserAgent can serve as a more advanced framework for more interactive and scalable web agents.
comment: 10 pages
♻ ☆ The Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge
Large language models (LLMs) are increasingly deployed as automatic judges to evaluate system outputs in tasks such as summarization, dialogue, and creative writing. A faithful judge should base its verdicts solely on response quality and explicitly acknowledge the factors shaping its decision. We show that current LLM judges fail on both counts by relying on shortcuts introduced in the prompt. Our study uses two evaluation datasets: ELI5, a benchmark for long-form question answering, and LitBench, a recent benchmark for creative writing. Both datasets provide pairwise comparisons, where the evaluator must choose which of two responses is better. From each dataset we construct 100 pairwise judgment tasks and employ two widely used models, GPT-4o and Gemini-2.5-Flash, as evaluators in the role of LLM-as-a-judge. For each pair, we assign superficial cues to the responses, provenance cues indicating source identity (Human, Expert, LLM, or Unknown) and recency cues indicating temporal origin (Old, 1950 vs. New, 2025), while keeping the rest of the prompt fixed. Results reveal consistent verdict shifts: both models exhibit a strong recency bias, systematically favoring new responses over old, as well as a clear provenance hierarchy (Expert > Human > LLM > Unknown). These biases are especially pronounced in GPT-4o and in the more subjective and open-ended LitBench domain. Crucially, cue acknowledgment is rare: justifications almost never reference the injected cues, instead rationalizing decisions in terms of content qualities. These findings demonstrate that current LLM-as-a-judge systems are shortcut-prone and unfaithful, undermining their reliability as evaluators in both research and deployment.
comment: 9 Pages, 5 Tables, 1 Figures
♻ ☆ Steering Large Language Models for Machine Translation Personalization
Large language models have simplified the production of personalized translations reflecting predefined stylistic constraints. However, these systems still struggle when stylistic requirements are implicitly represented by a set of examples, such as texts produced by a specific human translator. In this work, we explore various strategies for personalizing automatically generated translations when few examples are available, with a focus on the challenging domain of literary translation. We begin by determining the feasibility of the task and how style information is encoded within model representations. Then, we evaluate various prompting strategies and inference-time interventions for steering model generations towards a personalized style, with a particular focus on contrastive steering with sparse autoencoder (SAE) latents to identify salient personalization properties. We demonstrate that contrastive SAE steering yields robust style conditioning and translation quality, resulting in higher inference-time computational efficiency than prompting approaches. We further examine the impact of steering on model activations, finding that layers encoding personalization properties are impacted similarly by prompting and SAE steering, suggesting a similar mechanism at play.
♻ ☆ The Cultural Gene of Large Language Models: A Study on the Impact of Cross-Corpus Training on Model Values and Biases
Large language models (LLMs) are deployed globally, yet their underlying cultural and ethical assumptions remain underexplored. We propose the notion of a "cultural gene" -- a systematic value orientation that LLMs inherit from their training corpora -- and introduce a Cultural Probe Dataset (CPD) of 200 prompts targeting two classic cross-cultural dimensions: Individualism-Collectivism (IDV) and Power Distance (PDI). Using standardized zero-shot prompts, we compare a Western-centric model (GPT-4) and an Eastern-centric model (ERNIE Bot). Human annotation shows significant and consistent divergence across both dimensions. GPT-4 exhibits individualistic and low-power-distance tendencies (IDV score approx 1.21; PDI score approx -1.05), while ERNIE Bot shows collectivistic and higher-power-distance tendencies (IDV approx -0.89; PDI approx 0.76); differences are statistically significant (p < 0.001). We further compute a Cultural Alignment Index (CAI) against Hofstede's national scores and find GPT-4 aligns more closely with the USA (e.g., IDV CAI approx 0.91; PDI CAI approx 0.88) whereas ERNIE Bot aligns more closely with China (IDV CAI approx 0.85; PDI CAI approx 0.81). Qualitative analyses of dilemma resolution and authority-related judgments illustrate how these orientations surface in reasoning. Our results support the view that LLMs function as statistical mirrors of their cultural corpora and motivate culturally aware evaluation and deployment to avoid algorithmic cultural hegemony.
comment: 10 pages, 5 figures, IEEE conference format, submitted to [Conference Name]
♻ ☆ DoctorAgent-RL: A Multi-Agent Collaborative Reinforcement Learning System for Multi-Turn Clinical Dialogue
Large language models (LLMs) have demonstrated excellent capabilities in the field of biomedical question answering, but their application in real-world clinical consultations still faces core challenges. Single-round consultation systems require patients to describe all symptoms upfront, leading to vague diagnosis with unclear complaints. Traditional multi-turn dialogue models, constrained by static supervised learning, lack flexibility and fail to intelligently extract key clinical information. To address these limitations, we propose \Ours{}, a reinforcement learning (RL)-based multi-agent collaborative framework that models medical consultations as a dynamic decision-making process under uncertainty. The doctor agent continuously optimizes its questioning strategy within the RL framework through multi-turn interactions with the patient agent, dynamically adjusting its information-gathering path based on comprehensive rewards from the Consultation Evaluator. This RL fine-tuning mechanism enables LLMs to autonomously develop interaction strategies aligned with clinical reasoning logic, rather than superficially imitating patterns in existing dialogue data. Notably, we constructed MTMedDialog, the first English multi-turn medical consultation dataset capable of simulating patient interactions. Experiments demonstrate that \Ours{} outperforms existing models in both multi-turn reasoning capability and final diagnostic performance. This approach shows immense practical value by reducing misdiagnosis risks in time-pressured settings, freeing clinicians for complex cases, and pioneering a strategy to optimize medical resource allocation and alleviate workforce shortages. Code and data are available at https://github.com/JarvisUSTC/DoctorAgent-RL
♻ ☆ Efficient and Versatile Model for Multilingual Information Retrieval of Islamic Text: Development and Deployment in Real-World Scenarios
Despite recent advancements in Multilingual Information Retrieval (MLIR), a significant gap remains between research and practical deployment. Many studies assess MLIR performance in isolated settings, limiting their applicability to real-world scenarios. In this work, we leverage the unique characteristics of the Quranic multilingual corpus to examine the optimal strategies to develop an ad-hoc IR system for the Islamic domain that is designed to satisfy users' information needs in multiple languages. We prepared eleven retrieval models employing four training approaches: monolingual, cross-lingual, translate-train-all, and a novel mixed method combining cross-lingual and monolingual techniques. Evaluation on an in-domain dataset demonstrates that the mixed approach achieves promising results across diverse retrieval scenarios. Furthermore, we provide a detailed analysis of how different training configurations affect the embedding space and their implications for multilingual retrieval effectiveness. Finally, we discuss deployment considerations, emphasizing the cost-efficiency of deploying a single versatile, lightweight model for real-world MLIR applications.
♻ ☆ Unspoken Hints: Accuracy Without Acknowledgement in LLM Reasoning
Large language models (LLMs) increasingly rely on chain-of-thought (CoT) prompting to solve mathematical and logical reasoning tasks. Yet, a central question remains: to what extent are these generated rationales \emph{faithful} to the underlying computations, rather than post-hoc narratives shaped by hints that function as answer shortcuts embedded in the prompt? Following prior work on hinted vs.\ unhinted prompting, we present a systematic study of CoT faithfulness under controlled hint manipulations. Our experimental design spans four datasets (AIME, GSM-Hard, MATH-500, UniADILR), two state-of-the-art models (GPT-4o and Gemini-2-Flash), and a structured set of hint conditions varying in correctness (correct and incorrect), presentation style (sycophancy and data leak), and complexity (raw answers, two-operator expressions, four-operator expressions). We evaluate both task accuracy and whether hints are explicitly acknowledged in the reasoning. Our results reveal three key findings. First, correct hints substantially improve accuracy, especially on harder benchmarks and logical reasoning, while incorrect hints sharply reduce accuracy in tasks with lower baseline competence. Second, acknowledgement of hints is highly uneven: equation-based hints are frequently referenced, whereas raw hints are often adopted silently, indicating that more complex hints push models toward verbalizing their reliance in the reasoning process. Third, presentation style matters: sycophancy prompts encourage overt acknowledgement, while leak-style prompts increase accuracy but promote hidden reliance. This may reflect RLHF-related effects, as sycophancy exploits the human-pleasing side and data leak triggers the self-censoring side. Together, these results demonstrate that LLM reasoning is systematically shaped by shortcuts in ways that obscure faithfulness.
comment: 5 Pages, 4 Figures, 4 Tables
♻ ☆ Hybrid Multi-stage Decoding for Few-shot NER with Entity-aware Contrastive Learning
Few-shot named entity recognition can identify new types of named entities based on a few labeled examples. Previous methods employing token-level or span-level metric learning suffer from the computational burden and a large number of negative sample spans. In this paper, we propose the Hybrid Multi-stage Decoding for Few-shot NER with Entity-aware Contrastive Learning (MsFNER), which splits the general NER into two stages: entity-span detection and entity classification. There are 3 processes for introducing MsFNER: training, finetuning, and inference. In the training process, we train and get the best entity-span detection model and the entity classification model separately on the source domain using meta-learning, where we create a contrastive learning module to enhance entity representations for entity classification. During finetuning, we finetune the both models on the support dataset of target domain. In the inference process, for the unlabeled data, we first detect the entity-spans, then the entity-spans are jointly determined by the entity classification model and the KNN. We conduct experiments on the open FewNERD dataset and the results demonstrate the advance of MsFNER.
♻ ☆ EMSEdit: Efficient Multi-Step Meta-Learning-based Model Editing
Large Language Models (LLMs) power numerous AI applications, yet updating their knowledge remains costly. Model editing provides a lightweight alternative through targeted parameter modifications, with meta-learning-based model editing (MLME) demonstrating strong effectiveness and efficiency. However, we find that MLME struggles in low-data regimes and incurs high training costs due to the use of KL divergence. To address these issues, we propose $\textbf{E}$fficient $\textbf{M}$ulti-$\textbf{S}$tep $\textbf{Edit (EMSEdit)}$, which leverages multi-step backpropagation (MSBP) to effectively capture gradient-activation mapping patterns within editing samples, performs multi-step edits per sample to enhance editing performance under limited data, and introduces norm-based regularization to preserve unedited knowledge while improving training efficiency. Experiments on two datasets and three LLMs show that EMSEdit consistently outperforms state-of-the-art methods in both sequential and batch editing. Moreover, MSBP can be seamlessly integrated into existing approaches to yield additional performance gains. Further experiments on a multi-hop reasoning editing task demonstrate EMSEdit's robustness in handling complex edits, while ablation studies validate the contribution of each design component. Our code is available at https://github.com/xpq-tech/emsedit.
♻ ☆ Enhancing Long-Chain Reasoning Distillation through Error-Aware Self-Reflection
Large Language Models (LLMs) have exhibited strong reasoning capabilities and achieved remarkable performance in mathematical problem-solving tasks. Recently, distilling reasoning ability from long-form Chains-of-Thought (CoTs) has emerged as a promising approach for enhancing Small Language Models (SLMs). Existing studies typically treat SLMs as student models and use long-form CoTs as supervision signals for Supervised Fine-Tuning (SFT) to transfer reasoning ability. However, such long-form CoT teachers are usually unaware of the student model's capacity, which limits the effective utilization of the provided reasoning traces. To overcome this limitation, we propose errOr-aware self-ReflectION (ORION), a framework that refines teacher CoTs through an Error-Aware Reflection process. ORION enables the student model to construct more tailored teacher CoTs by refining teacher CoTs and incorporating its own reasoning errors. Experiments on multiple mathematical reasoning benchmarks demonstrate that ORION consistently improves performance by more than 2% over all baselines. Further analysis reveals that the CoTs constructed by ORION exhibit higher coherence and logical consistency, thereby serving as more effective supervision signals for SFT. All codes are available at https://github.com/NEUIR/ORION.git.
♻ ☆ Are Large Reasoning Models Interruptible?
Large Reasoning Models (LRMs) excel at complex reasoning but are traditionally evaluated in static, "frozen world" settings: model responses are assumed to be instantaneous, and the context of a request is presumed to be immutable over the duration of the response. While generally true for short-term tasks, the "frozen world" assumption breaks down in modern reasoning tasks such as assistive programming, where models may take hours to think through problems and code may change dramatically from the time the model starts thinking to the model's final output. In this work, we challenge the frozen world assumption and evaluate LRM robustness under two realistic dynamic scenarios: interruptions, which test the quality of the model's partial outputs on a limited budget, and dynamic context, which tests model adaptation to in-flight changes. Across mathematics and programming benchmarks that require long-form reasoning, static evaluations consistently overestimate robustness: even state-of-the-art LRMs, which achieve high accuracy in static settings, can fail unpredictably when interrupted or exposed to changing context, with performance dropping by up to 60% when updates are introduced late in the reasoning process. Our analysis further reveals several novel failure modes, including reasoning leakage, where models fold the reasoning into their final answer when interrupted; panic, where under time pressure models abandon reasoning entirely and return incorrect answers; and self-doubt, where performance degrades while incorporating updated information.
comment: We found a code/data bug in Section 5's intervene experiments. We're not sure how much it would affect the overall results and thus we're planning to fix it and do further investigation. As we do not want to leave any incorrect information on the internet, we want to withdraw this submission
♻ ☆ Cross-Modal Safety Alignment: Is textual unlearning all you need? EMNLP 2024
Recent studies reveal that integrating new modalities into Large Language Models (LLMs), such as Vision-Language Models (VLMs), creates a new attack surface that bypasses existing safety training techniques like Supervised Fine-tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF). While further SFT and RLHF-based safety training can be conducted in multi-modal settings, collecting multi-modal training datasets poses a significant challenge. Inspired by the structural design of recent multi-modal models, where, regardless of the combination of input modalities, all inputs are ultimately fused into the language space, we aim to explore whether unlearning solely in the textual domain can be effective for cross-modality safety alignment. Our evaluation across six datasets empirically demonstrates the transferability -- textual unlearning in VLMs significantly reduces the Attack Success Rate (ASR) to less than 8\% and in some cases, even as low as nearly 2\% for both text-based and vision-text-based attacks, alongside preserving the utility. Moreover, our experiments show that unlearning with a multi-modal dataset offers no potential benefits but incurs significantly increased computational demands, possibly up to 6 times higher.
comment: Accepted by EMNLP 2024 Findings
♻ ☆ Oyster-I: Beyond Refusal -- Constructive Safety Alignment for Responsible Language Models
Large language models (LLMs) typically deploy safety mechanisms to prevent harmful content generation. Most current approaches focus narrowly on risks posed by malicious actors, often framing risks as adversarial events and relying on defensive refusals. However, in real-world settings, risks also come from non-malicious users seeking help while under psychological distress (e.g., self-harm intentions). In such cases, the model's response can strongly influence the user's next actions. Simple refusals may lead them to repeat, escalate, or move to unsafe platforms, creating worse outcomes. We introduce Constructive Safety Alignment (CSA), a human-centric paradigm that protects against malicious misuse while actively guiding vulnerable users toward safe and helpful results. Implemented in Oyster-I (Oy1), CSA combines game-theoretic anticipation of user reactions, fine-grained risk boundary discovery, and interpretable reasoning control, turning safety into a trust-building process. Oy1 achieves state-of-the-art safety among open models while retaining high general capabilities. On our Constructive Benchmark, it shows strong constructive engagement, close to GPT-5, and unmatched robustness on the Strata-Sword jailbreak dataset, nearing GPT-o1 levels. By shifting from refusal-first to guidance-first safety, CSA redefines the model-user relationship, aiming for systems that are not just safe, but meaningfully helpful. We release Oy1, code, and the benchmark to support responsible, user-centered AI.
comment: Technical Report Code & Model weights available: https://github.com/Alibaba-AAIG/Oyster
♻ ☆ Understanding the Mixture-of-Experts with Nadaraya-Watson Kernel
Mixture-of-Experts (MoE) has become a cornerstone in recent state-of-the-art large language models (LLMs). Traditionally, MoE relies on $\mathrm{Softmax}$ as the router score function to aggregate expert output, a designed choice that has persisted from the earliest MoE models to modern LLMs, and is now widely regarded as standard practice. However, the necessity of using $\mathrm{Softmax}$ to project router weights into a probability simplex remains an unchallenged assumption rather than a principled design choice. In this work, we first revisit the classical Nadaraya-Watson regression and observe that MoE shares the same mathematical formulation as Nadaraya-Watson regression. Furthermore, we show that both feed-forward neural network (FFN) and MoE can be interpreted as a special case of Nadaraya-Watson regression, where the kernel function corresponds to the input neurons of the output layer. Motivated by these insights, we propose the \textbf{zero-additional-cost} Kernel Inspired Router with Normalization (KERN), an FFN-style router function, as an alternative to $\mathrm{Softmax}$. We demonstrate that this router generalizes both $\mathrm{Sigmoid}$- and $\mathrm{Softmax}$-based routers. \textbf{Based on empirical observations and established practices in FFN implementation, we recommend the use of $\mathrm{ReLU}$ activation and $\ell_2$-normalization in $\mathrm{KERN}$ router function.} Comprehensive experiments in MoE and LLM validate the effectiveness of the proposed FFN-style router function \methodNorm.
comment: Tech Report
♻ ☆ Time-IMM: A Dataset and Benchmark for Irregular Multimodal Multivariate Time Series NeurIPS 2025
Time series data in real-world applications such as healthcare, climate modeling, and finance are often irregular, multimodal, and messy, with varying sampling rates, asynchronous modalities, and pervasive missingness. However, existing benchmarks typically assume clean, regularly sampled, unimodal data, creating a significant gap between research and real-world deployment. We introduce Time-IMM, a dataset specifically designed to capture cause-driven irregularity in multimodal multivariate time series. Time-IMM represents nine distinct types of time series irregularity, categorized into trigger-based, constraint-based, and artifact-based mechanisms. Complementing the dataset, we introduce IMM-TSF, a benchmark library for forecasting on irregular multimodal time series, enabling asynchronous integration and realistic evaluation. IMM-TSF includes specialized fusion modules, including a timestamp-to-text fusion module and a multimodality fusion module, which support both recency-aware averaging and attention-based integration strategies. Empirical results demonstrate that explicitly modeling multimodality on irregular time series data leads to substantial gains in forecasting performance. Time-IMM and IMM-TSF provide a foundation for advancing time series analysis under real-world conditions. The dataset is publicly available at https://www.kaggle.com/datasets/blacksnail789521/time-imm/data, and the benchmark library can be accessed at https://github.com/blacksnail789521/IMM-TSF.
comment: This paper has been accepted by the NeurIPS 2025 Datasets and Benchmarks Track
♻ ☆ When "Competency" in Reasoning Opens the Door to Vulnerability: Jailbreaking LLMs via Novel Complex Ciphers NeurIPS 2025
Recent advancements in Large Language Model (LLM) safety have primarily focused on mitigating attacks crafted in natural language or common ciphers (e.g. Base64), which are likely integrated into newer models' safety training. However, we reveal a paradoxical vulnerability: as LLMs advance in reasoning, they inadvertently become more susceptible to novel jailbreaking attacks. Enhanced reasoning enables LLMs to interpret complex instructions and decode complex user-defined ciphers, creating an exploitable security gap. To study this vulnerability, we introduce Attacks using Custom Encryptions (ACE), a jailbreaking technique that encodes malicious queries with novel ciphers. Extending ACE, we introduce Layered Attacks using Custom Encryptions (LACE), which applies multi-layer ciphers to amplify attack complexity. Furthermore, we develop CipherBench, a benchmark designed to evaluate LLMs' accuracy in decoding encrypted benign text. Our experiments reveal a critical trade-off: LLMs that are more capable of decoding ciphers are more vulnerable to LACE, with success rates on gpt-oss-20b escalating from 60% under ACE to 72% with LACE. These findings highlight a critical insight: as LLMs become more adept at deciphering complex user ciphers--many of which cannot be preemptively included in safety training--they become increasingly exploitable.
comment: Published in Reliable ML from Unreliable Data workshop @ NeurIPS 2025
♻ ☆ DefenderBench: A Toolkit for Evaluating Language Agents in Cybersecurity Environments NeurIPS 2025
Large language model (LLM) agents have shown impressive capabilities in human language comprehension and reasoning, yet their potential in cybersecurity remains underexplored. We introduce DefenderBench, a practical, open-source toolkit for evaluating language agents across offense, defense, and cybersecurity knowledge-based tasks. DefenderBench includes environments for network intrusion, malicious content detection, code vulnerability analysis, and cybersecurity knowledge assessment. It is intentionally designed to be affordable and easily accessible for researchers while providing fair and rigorous assessment. We benchmark several state-of-the-art (SoTA) and popular LLMs, including both open- and closed-weight models, using a standardized agentic framework. Our results show that Claude-3.7-sonnet performs best with a DefenderBench score of 81.65, followed by Claude-3.7-sonnet-think with 78.40, while the best open-weight model, Llama 3.3 70B, is not far behind with a DefenderBench score of 71.81. DefenderBench's modular design allows seamless integration of custom LLMs and tasks, promoting reproducibility and fair comparisons. An anonymized version of DefenderBench is available at https://github.com/microsoft/DefenderBench.
comment: Accepted by NeurIPS 2025 Workshop Scaling Environments for Agents (SEA)
♻ ☆ Highlighting What Matters: Promptable Embeddings for Attribute-Focused Image Retrieval NeurIPS 2025
While an image is worth more than a thousand words, only a few provide crucial information for a given task and thus should be focused on. In light of this, ideal text-to-image (T2I) retrievers should prioritize specific visual attributes relevant to queries. To evaluate current retrievers on handling attribute-focused queries, we build COCO-Facet, a COCO-based benchmark with 9,112 queries about diverse attributes of interest. We find that CLIP-like retrievers, which are widely adopted due to their efficiency and zero-shot ability, have poor and imbalanced performance, possibly because their image embeddings focus on global semantics and subjects while leaving out other details. Notably, we reveal that even recent Multimodal Large Language Model (MLLM)-based, stronger retrievers with a larger output dimension struggle with this limitation. Hence, we hypothesize that retrieving with general image embeddings is suboptimal for performing such queries. As a solution, we propose to use promptable image embeddings enabled by these multimodal retrievers, which boost performance by highlighting required attributes. Our pipeline for deriving such embeddings generalizes across query types, image pools, and base retriever architectures. To enhance real-world applicability, we offer two acceleration strategies: Pre-processing promptable embeddings and using linear approximations. We show that the former yields a 15% improvement in Recall@5 when prompts are predefined, while the latter achieves an 8% improvement when prompts are only available during inference.
comment: NeurIPS 2025; 27 pages, 6 figures
♻ ☆ Causal Agent based on Large Language Model
The large language model (LLM) has achieved significant success across various domains. However, the inherent complexity of causal problems and causal theory poses challenges in accurately describing them in natural language, making it difficult for LLM to comprehend and use them effectively. Causal methods are not easily conveyed through natural language, which hinders LLM's ability to apply them accurately. Additionally, causal datasets are typically tabular, while LLM excels in handling natural language data, creating a structural mismatch that impedes effective reasoning with tabular data. To address these challenges, we have equipped the LLM with causal tools within an agent framework, named the Causal Agent, enabling it to tackle causal problems. The causal agent comprises tools, memory, and reasoning modules. In the tool module, the causal agent calls Python code and uses the encapsulated causal function module to align tabular data with natural language. In the reasoning module, the causal agent performs reasoning through multiple iterations with the tools. In the memory module, the causal agent maintains a dictionary instance where the keys are unique names and the values are causal graphs. To verify the causal ability of the causal agent, we established a Causal Tabular Question Answer (CausalTQA) benchmark consisting of four levels of causal problems: variable level, edge level, causal graph level, and causal effect level. CausalTQA consists of about 1.4K for these four levels questions. Causal agent demonstrates remarkable efficacy on the four-level causal problems, with accuracy rates all above 80\%. Through verification on the real-world dataset QRData, the causal agent is 6\% higher than the original SOTA. For further insights and implementation details, our code is accessible via the GitHub repository https://github.com/kairong-han/causal_agent.
♻ ☆ LUMINA: Detecting Hallucinations in RAG System with Context-Knowledge Signals
Retrieval-Augmented Generation (RAG) aims to mitigate hallucinations in large language models (LLMs) by grounding responses in retrieved documents. Yet, RAG-based LLMs still hallucinate even when provided with correct and sufficient context. A growing line of work suggests that this stems from an imbalance between how models use external context and their internal knowledge, and several approaches have attempted to quantify these signals for hallucination detection. However, existing methods require extensive hyperparameter tuning, limiting their generalizability. We propose LUMINA, a novel framework that detects hallucinations in RAG systems through context-knowledge signals: external context utilization is quantified via distributional distance, while internal knowledge utilization is measured by tracking how predicted tokens evolve across transformer layers. We further introduce a framework for statistically validating these measurements. Experiments on common RAG hallucination benchmarks and four open-source LLMs show that LUMINA achieves consistently high AUROC and AUPRC scores, outperforming prior utilization-based methods by up to +13% AUROC on HalluRAG. Moreover, LUMINA remains robust under relaxed assumptions about retrieval quality and model matching, offering both effectiveness and practicality.
♻ ☆ DFAMS: Dynamic-flow guided Federated Alignment based Multi-prototype Search
Federated Retrieval (FR) routes queries across multiple external knowledge sources, to mitigate hallucinations of LLMs, when necessary external knowledge is distributed. However, existing methods struggle to retrieve high-quality and relevant documents for ambiguous queries, especially in cross-domain scenarios, which significantly limits their effectiveness in supporting downstream generation tasks. Inspired by Dynamic Information Flow (DIF), we propose DFAMS, a novel framework that leverages DIF to identify latent query intents and construct semantically aligned knowledge partitions for accurate retrieval across heterogeneous sources. Specifically, DFAMS probes the DIF in LLMs by leveraging gradient signals from a few annotated queries and employing Shapley value-based attribution to trace neuron activation paths associated with intent recognition and subdomain boundary detection. Then, DFAMS leverages DIF to train an alignment module via multi-prototype contrastive learning, enabling fine-grained intra-source modeling and inter-source semantic alignment across knowledge bases. Experimental results across five benchmarks show that DFAMS outperforms advanced FR methods by up to 14.37\% in knowledge classification accuracy, 5.38\% in retrieval recall, and 6.45\% in downstream QA accuracy, demonstrating its effectiveness in complex FR scenarios. Our code are anonymous available at https://anonymous.4open.science/r/DFAMS/
comment: 8 pages, 3 figures
♻ ☆ Attention-Aware GNN-based Input Defense against Multi-Turn LLM Jailbreak
Large Language Models (LLMs) have gained significant traction in various applications, yet their capabilities present risks for both constructive and malicious exploitation. Despite extensive training and fine-tuning efforts aimed at enhancing safety, LLMs remain susceptible to jailbreak attacks. Recently, the emergence of multi-turn attacks has intensified this vulnerability. Unlike single-turn attacks, multi-turn attacks incrementally escalate dialogue complexity, rendering them more challenging to detect and mitigate. In this study, we introduce G-Guard, an innovative attention-aware Graph Neural Network (GNN)-based input classifier specifically designed to defend against multi-turn jailbreak attacks targeting LLMs. G-Guard constructs an entity graph for multi-turn queries, which captures the interrelationships between queries and harmful keywords that present in multi-turn queries. Furthermore, we propose an attention-aware augmentation mechanism that retrieves the most relevant single-turn query based on the ongoing multi-turn conversation. The retrieved query is incorporated as a labeled node within the graph, thereby enhancing the GNN's capacity to classify the current query as harmful or benign. Evaluation results show that G-Guard consistently outperforms all baselines across diverse datasets and evaluation metrics, demonstrating its efficacy as a robust defense mechanism against multi-turn jailbreak attacks.
♻ ☆ Diffusion Language Models Know the Answer Before Decoding
Diffusion language models (DLMs) have recently emerged as an alternative to autoregressive approaches, offering parallel sequence generation and flexible token orders. However, their inference remains slower than that of autoregressive models, primarily due to the cost of bidirectional attention and the large number of refinement steps required for high quality outputs. In this work, we highlight and leverage an overlooked property of DLMs early answer convergence: in many cases, the correct answer can be internally identified by half steps before the final decoding step, both under semi-autoregressive and random remasking schedules. For example, on GSM8K and MMLU, up to 97% and 99% of instances, respectively, can be decoded correctly using only half of the refinement steps. Building on this observation, we introduce Prophet, a training-free fast decoding paradigm that enables early commit decoding. Specifically, Prophet dynamically decides whether to continue refinement or to go "all-in" (i.e., decode all remaining tokens in one step), using the confidence gap between the top-2 prediction candidates as the criterion. It integrates seamlessly into existing DLM implementations, incurs negligible overhead, and requires no additional training. Empirical evaluations of LLaDA-8B and Dream-7B across multiple tasks show that Prophet reduces the number of decoding steps by up to 3.4x while preserving high generation quality. These results recast DLM decoding as a problem of when to stop sampling, and demonstrate that early decode convergence provides a simple yet powerful mechanism for accelerating DLM inference, complementary to existing speedup techniques. Our code is publicly available at https://github.com/pixeli99/Prophet.
♻ ☆ DSPO: Stable and Efficient Policy Optimization for Agentic Search and Reasoning
Enhancing LLMs with the ability to actively search external knowledge is crucial for complex and real-world tasks. Current approaches either rely on prompting to elicit the model's innate agent capabilities, or suffer from performance ceilings and collapse when applying RL to complex interactive tasks, leaving their true agentic potential untapped. To address this, we introduce \textbf{D}ynamic-filter \textbf{S}equence-level \textbf{P}olicy \textbf{O}ptimization (DSPO), an improved RL algorithm designed for robust agent training through sequence-level optimization and dynamic sample filtering. We train our model purely through RL to interleave multi-turn search and reasoning, obviating the need for supervised demonstration data. Across multiple QA benchmarks, our 7B model improves over a comparable previous work by \textbf{34.1\%}, and even outperforms the 14B model from previous work in complex multihop QA such as HotpotQA by nearly \textbf{9\% relative}, maintaining exceptional training stability.
♻ ☆ TripScore: Benchmarking and rewarding real-world travel planning with fine-grained evaluation
Travel planning is a valuable yet complex task that poses significant challenges even for advanced large language models (LLMs). While recent benchmarks have advanced in evaluating LLMs' planning capabilities, they often fall short in evaluating feasibility, reliability, and engagement of travel plans. We introduce a comprehensive benchmark for travel planning that unifies fine-grained criteria into a single reward, enabling direct comparison of plan quality and seamless integration with reinforcement learning (RL). Our evaluator achieves moderate agreement with travel-expert annotations (60.75%) and outperforms multiple LLM-as-judge baselines. We further release a large-scale dataset of 4,870 queries including 219 real-world, free-form requests for generalization to authentic user intent. Using this benchmark, we conduct extensive experiments across diverse methods and LLMs, including test-time computation, neuro-symbolic approaches, supervised fine-tuning, and RL via GRPO. Across base models, RL generally improves itinerary feasibility over prompt-only and supervised baselines, yielding higher unified reward scores.
♻ ☆ EvolveNav: Empowering LLM-Based Vision-Language Navigation via Self-Improving Embodied Reasoning
Recent studies have revealed the potential of training open-source Large Language Models (LLMs) to unleash LLMs' reasoning ability for enhancing vision-language navigation (VLN) performance, and simultaneously mitigate the domain gap between LLMs' training corpus and the VLN task. However, these approaches predominantly adopt straightforward input-output mapping paradigms, causing the mapping learning difficult and the navigational decisions unexplainable. Chain-of-Thought (CoT) training is a promising way to improve both navigational decision accuracy and interpretability, while the complexity of the navigation task makes the perfect CoT labels unavailable and may lead to overfitting through pure CoT supervised fine-tuning. To address these issues, we propose EvolveNav, a novel sElf-improving embodied reasoning paradigm that realizes adaptable and generalizable navigational reasoning for boosting LLM-based vision-language Navigation. Specifically, EvolveNav involves a two-stage training process: (1) Formalized CoT Supervised Fine-Tuning, where we train the model with curated formalized CoT labels to first activate the model's navigational reasoning capabilities, and simultaneously increase the reasoning speed; (2) Self-Reflective Post-Training, where the model is iteratively trained with its own reasoning outputs as self-enriched CoT labels to enhance the supervision diversity. A self-reflective auxiliary task is also designed to encourage the model to learn correct reasoning patterns by contrasting with wrong ones. Experimental results under both task-specific and cross-task training paradigms demonstrate the consistent superiority of EvolveNav over previous LLM-based VLN approaches on various popular benchmarks, including R2R, REVERIE, CVDN, and SOON. Code is available at https://github.com/expectorlin/EvolveNav.
♻ ☆ Responsible AI Technical Report
KT developed a Responsible AI (RAI) assessment methodology and risk mitigation technologies to ensure the safety and reliability of AI services. By analyzing the Basic Act on AI implementation and global AI governance trends, we established a unique approach for regulatory compliance and systematically identify and manage all potential risk factors from AI development to operation. We present a reliable assessment methodology that systematically verifies model safety and robustness based on KT's AI risk taxonomy tailored to the domestic environment. We also provide practical tools for managing and mitigating identified AI risks. With the release of this report, we also release proprietary Guardrail : SafetyGuard that blocks harmful responses from AI models in real-time, supporting the enhancement of safety in the domestic AI development ecosystem. We also believe these research outcomes provide valuable insights for organizations seeking to develop Responsible AI.
comment: 23 pages, 8 figures
♻ ☆ Graph2Eval: Automatic Multimodal Task Generation for Agents via Knowledge Graphs
As multimodal LLM-driven agents continue to advance in autonomy and generalization, evaluation based on static datasets can no longer adequately assess their true capabilities in dynamic environments and diverse tasks. Existing LLM-based synthetic data methods are largely designed for LLM training and evaluation, and thus cannot be directly applied to agent tasks that require tool use and interactive capabilities. While recent studies have explored automatic agent task generation with LLMs, most efforts remain limited to text or image analysis, without systematically modeling multi-step interactions in web environments. To address these challenges, we propose Graph2Eval, a knowledge graph-based framework that automatically generates both multimodal document comprehension tasks and web interaction tasks, enabling comprehensive evaluation of agents' reasoning, collaboration, and interactive capabilities. In our approach, knowledge graphs constructed from multi-source external data serve as the task space, where we translate semantic relations into structured multimodal tasks using subgraph sampling, task templates, and meta-paths. A multi-stage filtering pipeline based on node reachability, LLM scoring, and similarity analysis is applied to guarantee the quality and executability of the generated tasks. Furthermore, Graph2Eval supports end-to-end evaluation of multiple agent types (Single-Agent, Multi-Agent, Web Agent) and measures reasoning, collaboration, and interaction capabilities. We instantiate the framework with Graph2Eval-Bench, a curated dataset of 1,319 tasks spanning document comprehension and web interaction scenarios. Experiments show that Graph2Eval efficiently generates tasks that differentiate agent and model performance, revealing gaps in reasoning, collaboration, and web interaction across different settings and offering a new perspective for agent evaluation.
comment: 20 pages, 10 figures. Our Code: https://github.com/YurunChen/Graph2Eval
♻ ☆ From Rational Answers to Emotional Resonance: The Role of Controllable Emotion Generation in Language Models
Purpose: Emotion is a fundamental component of human communication, shaping understanding, trust, and engagement across domains such as education, healthcare, and mental health. While large language models (LLMs) exhibit strong reasoning and knowledge generation capabilities, they still struggle to express emotions in a consistent, controllable, and contextually appropriate manner. This limitation restricts their potential for authentic human-AI interaction. Methods: We propose a controllable emotion generation framework based on Emotion Vectors (EVs) - latent representations derived from internal activation shifts between neutral and emotion-conditioned responses. By injecting these vectors into the hidden states of pretrained LLMs during inference, our method enables fine-grained, continuous modulation of emotional tone without any additional training or architectural modification. We further provide theoretical analysis proving that EV steering enhances emotional expressivity while maintaining semantic fidelity and linguistic fluency. Results: Extensive experiments across multiple LLM families show that the proposed approach achieves consistent emotional alignment, stable topic adherence, and controllable affect intensity. Compared with existing prompt-based and fine-tuning-based baselines, our method demonstrates superior flexibility and generalizability. Conclusion: Emotion Vector (EV) steering provides an efficient and interpretable means of bridging rational reasoning and affective understanding in large language models, offering a promising direction for building emotionally resonant AI systems capable of more natural human-machine interaction.
comment: 43 pages, 5 figures
♻ ☆ KaLM-Embedding-V2: Superior Training Techniques and Data Inspire A Versatile Embedding Model
Recent advancements in Large Language Models (LLMs)-based text embedding models primarily focus on data scaling or synthesis, yet limited exploration of training techniques and data quality, thereby constraining performance. In this work, we propose KaLM-Embedding-V2, a series of versatile and compact embedding models, systematically incentivizing advanced embedding capability in LLMs by superior training techniques and high-quality data. For model architecture, we implement the models on a 0.5B compact size with simple mean-pooling to produce fixed-length embeddings and remove the causal attention mask to enable fully bidirectional representation learning. For training techniques, we propose a progressive multi-stage training pipeline: pre-training on weakly supervised large-scale datasets, fine-tuning with supervised high-quality datasets, and contrastive distillation with fine-grained soft signals, integrated with focal-style reweighting and online hard-negative mixing to emphasize difficult samples and enrich hard negatives, respectively. For training data, we curate over 20 categories for pre-training and 100 categories for fine-tuning and contrastive distillation, to improve both performance and generalization, leveraging task-specific instructions, hard-negative mining, and example-based multi-class labeling to ensure high quality. Combining these techniques, our KaLM-Embedding-V2 series achieves state-of-the-art performance on the Massive Text Embedding Benchmark, outperforming models of comparable size and rivaling models 3-26x larger, setting a new standard for versatile and compact embedding models under 1B parameters.
comment: 32 pages, 16 tables, 5 figures
♻ ☆ The Open Source Advantage in Large Language Models (LLMs)
Large language models (LLMs) have rapidly advanced natural language processing, driving significant breakthroughs in tasks such as text generation, machine translation, and domain-specific reasoning. The field now faces a critical dilemma in its approach: closed-source models like GPT-4 deliver state-of-the-art performance but restrict reproducibility, accessibility, and external oversight, while open-source frameworks like LLaMA and Mixtral democratize access, foster collaboration, and support diverse applications, achieving competitive results through techniques like instruction tuning and LoRA. Hybrid approaches address challenges like bias mitigation and resource accessibility by combining the scalability of closed-source systems with the transparency and inclusivity of open-source framework. However, in this position paper, we argue that open-source remains the most robust path for advancing LLM research and ethical deployment.
comment: 9 pages, 1 figure
♻ ☆ CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysis
Large Language Models (LLMs) with Mixture-of-Experts (MoE) architectures are distinguished by their strong performance scaling with increasing parameters across a wide range of tasks, yet they also suffer from substantial computational and storage overheads. Notably, the performance gains of MoE models do not scale proportionally with the growth in expert parameters. While prior works attempt to reduce parameters via expert-level pruning, merging, or decomposition, they still suffer from challenges in both performance and computational efficiency. In this paper, we address these challenges by introducing micro-expert as a finer-grained compression unit that spans across matrices. We first establish a more fundamental perspective, viewing MoE layers as mixtures of micro-experts, and present CAMERA, a lightweight and training-free framework for identifying micro-expert redundancy. Our analysis uncovers significant variance in micro-expert contributions during decoding. Based on this insight, we further propose CAMERA-P, a structured micro-expert pruning framework, and CAMERA-Q, a mixed-precision quantization idea designed for micro-experts. Extensive experiments on nine downstream tasks show that CAMERA-P consistently outperforms strong baselines under pruning ratios ranging from 20% to 60%. Furthermore, CAMERA-Q achieves superior results under aggressive 2-bit quantization, surpassing existing matrix- and channel-level ideas. Notably, our method enables complete micro-expert analysis of Qwen2-57B-A14B in less than 5 minutes on a single NVIDIA A100-40GB GPU.
comment: 16 pages, 9 figures, 7 tables
♻ ☆ Exploring Compositional Generalization (in COGS/ReCOGS_pos) by Transformers using Restricted Access Sequence Processing (RASP)
Humans understand new combinations of words encountered if they are combinations of words recognized from different contexts, an ability called Compositional Generalization. The COGS benchmark (Kim and Linzen, 2020) arXiv:2010.05465 reports 0% accuracy for Transformer models on some structural generalizations. We use (Weiss et al., 2021) arXiv:2106.06981's Restricted Access Sequence Processing (RASP), a Transformer-equivalent programming language, to demonstrate that a Transformer Encoder-Decoder can perform COGS and the semantically equivalent ReCOGS_pos (Wu et al., 2024) arXiv:2303.13716 systematically and compositionally: Our RASP models attain near perfect scores on structural generalization splits on COGS (exact match) and ReCOGS_pos (semantic exact match). Our RASP models show the (Re)COGS tasks do not require a hierarchical or tree-structured solution (contrary to (Kim and Linzen, 2020) arXiv:2010.05465, (Yao and Koller, 2022) arXiv:2210.13050, (Murty et al., 2022) arXiv:2211.01288, (Liu et al., 2021) arXiv:2107.06516): we use word-level tokens with an "embedding" layer that tags with possible part of speech, applying just once per encoder pass 19 attention-head compatible flat pattern-matching rules (easily identified with specific training examples), shown using grammar coverage (Zeller et al., 2023) to cover the non-recursive aspects of the input grammar, plus masking out prepositional phrases ("pp noun") and/or sentential complements (cp) when recognizing grammar patterns and extracting nouns related to the main verb in the sentence, and output the next logical form (LF) token (repeating until the LF is complete). The models do not apply recursive, tree-structured rules like "np_det pp np -> np_pp -> np", but score near perfect semantic and string exact match on both COGS and ReCOGS pp recursion, cp recursion using the decoder loop.
comment: 8 pages main text with 3 figures and 4 tables; limitations page and references separate; 10 more figures, 1 image, and 3 more tables in the appendices supplement the work
♻ ☆ MLRIP: Pre-training a military language representation model with informative factual knowledge and professional knowledge base
Incorporating structured knowledge into pre-trained language models has demonstrated signiffcant bene-ffts for domain-speciffc natural language processing tasks, particularly in specialized ffelds like military intelligence analysis. Existing approaches typically integrate external knowledge through masking tech-niques or fusion mechanisms, but often fail to fully leverage the intrinsic tactical associations and factual information within input sequences, while introducing uncontrolled noise from unveriffed exter-nal sources. To address these limitations, we present MLRIP (Military Language Representation with Integrated Prior), a novel pre-training framework that introduces a hierarchical knowledge integration pipeline combined with a dual-phase entity substitu-tion mechanism. Our approach speciffcally models operational linkages between military entities, capturing critical dependencies such as command, support, and engagement structures. Comprehensive evaluations on military-speciffc NLP tasks show that MLRIP outperforms existing BERT-based models by substantial margins, establishing new state-of-the-art performance in military entity recognition, typing, and operational linkage extraction tasks while demonstrating superior operational efffciency in resource-constrained environments.
comment: 12 pages, 6 figures
Information Retrieval 18
☆ DeepMMSearch-R1: Empowering Multimodal LLMs in Multimodal Web Search
Multimodal Large Language Models (MLLMs) in real-world applications require access to external knowledge sources and must remain responsive to the dynamic and ever-changing real-world information in order to address information-seeking and knowledge-intensive user queries. Existing approaches, such as retrieval augmented generation (RAG) methods, search agents, and search equipped MLLMs, often suffer from rigid pipelines, excessive search calls, and poorly constructed search queries, which result in inefficiencies and suboptimal outcomes. To address these limitations, we present DeepMMSearch-R1, the first multimodal LLM capable of performing on-demand, multi-turn web searches and dynamically crafting queries for both image and text search tools. Specifically, DeepMMSearch-R1 can initiate web searches based on relevant crops of the input image making the image search more effective, and can iteratively adapt text search queries based on retrieved information, thereby enabling self-reflection and self-correction. Our approach relies on a two-stage training pipeline: a cold start supervised finetuning phase followed by an online reinforcement learning optimization. For training, we introduce DeepMMSearchVQA, a novel multimodal VQA dataset created through an automated pipeline intermixed with real-world information from web search tools. This dataset contains diverse, multi-hop queries that integrate textual and visual information, teaching the model when to search, what to search for, which search tool to use and how to reason over the retrieved information. We conduct extensive experiments across a range of knowledge-intensive benchmarks to demonstrate the superiority of our approach. Finally, we analyze the results and provide insights that are valuable for advancing multimodal web-search.
☆ CTRL-Rec: Controlling Recommender Systems With Natural Language
When users are dissatisfied with recommendations from a recommender system, they often lack fine-grained controls for changing them. Large language models (LLMs) offer a solution by allowing users to guide their recommendations through natural language requests (e.g., "I want to see respectful posts with a different perspective than mine"). We propose a method, CTRL-Rec, that allows for natural language control of traditional recommender systems in real-time with computational efficiency. Specifically, at training time, we use an LLM to simulate whether users would approve of items based on their language requests, and we train embedding models that approximate such simulated judgments. We then integrate these user-request-based predictions into the standard weighting of signals that traditional recommender systems optimize. At deployment time, we require only a single LLM embedding computation per user request, allowing for real-time control of recommendations. In experiments with the MovieLens dataset, our method consistently allows for fine-grained control across a diversity of requests. In a study with 19 Letterboxd users, we find that CTRL-Rec was positively received by users and significantly enhanced users' sense of control and satisfaction with recommendations compared to traditional controls.
☆ SAIL-Embedding Technical Report: Omni-modal Embedding Foundation Model
Multimodal embedding models aim to yield informative unified representations that empower diverse cross-modal tasks. Despite promising developments in the evolution from CLIP-based dual-tower architectures to large vision-language models, prior works still face unavoidable challenges in real-world applications and business scenarios, such as the limited modality support, unstable training mechanisms, and industrial domain gaps. In this work, we introduce SAIL-Embedding, an omni-modal embedding foundation model that addresses these issues through tailored training strategies and architectural design. In the optimization procedure, we propose a multi-stage training scheme to boost the multifaceted effectiveness of representation learning. Specifically, the content-aware progressive training aims to enhance the model's adaptability to diverse downstream tasks and master enriched cross-modal proficiency. The collaboration-aware recommendation enhancement training further adapts multimodal representations for recommendation scenarios by distilling knowledge from sequence-to-item and ID-to-item embeddings while mining user historical interests. Concurrently, we develop the stochastic specialization and dataset-driven pattern matching to strengthen model training flexibility and generalizability. Experimental results show that SAIL-Embedding achieves SOTA performance compared to other methods in different retrieval tasks. In online experiments across various real-world scenarios integrated with our model, we observe a significant increase in Lifetime (LT), which is a crucial indicator for the recommendation experience. For instance, the model delivers the 7-day LT gain of +0.158% and the 14-day LT gain of +0.144% in the Douyin-Selected scenario. For the Douyin feed rank model, the match features produced by SAIL-Embedding yield a +0.08% AUC gain.
comment: Technical Report
☆ The Role of Parametric Injection-A Systematic Study of Parametric Retrieval-Augmented Generation
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by retrieving external documents. As an emerging form of RAG, parametric retrieval-augmented generation (PRAG) encodes documents as model parameters (i.e., LoRA modules) and injects these representations into the model during inference, enabling interaction between the LLM and documents at parametric level. Compared with directly placing documents in the input context, PRAG is more efficient and has the potential to offer deeper model-document interaction. Despite its growing attention, the mechanism underlying parametric injection remains poorly understood. In this work, we present a systematic study of PRAG to clarify the role of parametric injection, showing that parameterized documents capture only partial semantic information of documents, and relying on them alone yields inferior performance compared to interaction at text level. However, these parametric representations encode high-level document information that can enhance the model's understanding of documents within the input context. When combined parameterized documents with textual documents, the model can leverage relevant information more effectively and become more robust to noisy inputs, achieving better performance than either source alone. We recommend jointly using parameterized and textual documents and advocate for increasing the information content of parametric representations to advance PRAG.
☆ SMILE: SeMantic Ids Enhanced CoLd Item Representation for Click-through Rate Prediction in E-commerce SEarch
With the rise of modern search and recommendation platforms, insufficient collaborative information of cold-start items exacerbates the Matthew effect of existing platform items, challenging platform diversity and becoming a longstanding issue. Existing methods align items' side content with collaborative information to transfer collaborative signals from high-popularity items to cold-start items. However, these methods fail to account for the asymmetry between collaboration and content, nor the fine-grained differences among items. To address these issues, we propose SMILE, an item representation enhancement approach based on fused alignment of semantic IDs. Specifically, we use RQ-OPQ encoding to quantize item content and collaborative information, followed by a two-step alignment: RQ encoding transfers shared collaborative signals across items, while OPQ encoding learns differentiated information of items. Comprehensive offline experiments on large-scale industrial datasets demonstrate superiority of SMILE, and rigorous online A/B tests confirm statistically significant improvements: item CTR +1.66%, buyers +1.57%, and order volume +2.17%.
☆ Leveraging Language Semantics for Collaborative Filtering with TextGCN and TextGCN-MLP: Zero-Shot vs In-Domain Performance
In recent years, various approaches have been proposed to leverage large language models (LLMs) for incorporating textual information about items into recommender systems. Existing methods primarily focus on either fine-tuning LLMs to generate recommendations or integrating LLM-based embeddings into downstream models. In this work, we follow the latter direction and propose \textbf{TextGCN}, which applies parameter-free graph convolution layers directly over LLM-based item-title embeddings, instead of learning ID-based embeddings as in traditional methods. By combining language semantics with graph message passing, this architecture achieves state-of-the-art zero-shot performance, significantly outperforming prior approaches. Furthermore, we introduce \textbf{TextGCN-MLP}, which extends TextGCN with a trainable multilayer perceptron trained using a contrastive loss, achieving state-of-the-art in-domain performance on recommendation benchmarks. However, the zero-shot performance of TextGCN-MLP remains lower than that of TextGCN, highlighting the trade-off between in-domain specialization and zero-shot generalization. We release our code on github at \href{https://github.com/ChernovAndrey/TFCE}{github.com/ChernovAndrey/TFCE}.
☆ A Hierarchical Quantized Tokenization Framework for Task-Adaptive Graph Representation Learning
Recent progress in language and vision foundation models demonstrates the importance of discrete token interfaces that transform complex inputs into compact sequences for large-scale modeling. Extending this paradigm to graphs requires a tokenization scheme that handles non-Euclidean structures and multi-scale dependencies efficiently. Existing approaches to graph tokenization, linearized, continuous, and quantized, remain limited in adaptability and efficiency. In particular, most current quantization-based tokenizers organize hierarchical information in fixed or task-agnostic ways, which may either over-represent or under-utilize structural cues, and lack the ability to dynamically reweight contributions from different levels without retraining the encoder. This work presents a hierarchical quantization framework that introduces a self-weighted mechanism for task-adaptive aggregation across multiple scales. The proposed method maintains a frozen encoder while modulating information flow through a lightweight gating process, enabling parameter-efficient adaptation to diverse downstream tasks. Experiments on benchmark datasets for node classification and link prediction demonstrate consistent improvements over strong baselines under comparable computational budgets.
☆ Simple Projection Variants Improve ColBERT Performance
Multi-vector dense retrieval methods like ColBERT systematically use a single-layer linear projection to reduce the dimensionality of individual vectors. In this study, we explore the implications of the MaxSim operator on the gradient flows of the training of multi-vector models and show that such a simple linear projection has inherent, if non-critical, limitations in this setting. We then discuss the theoretical improvements that could result from replacing this single-layer projection with well-studied alternative feedforward linear networks (FFN), such as deeper, non-linear FFN blocks, GLU blocks, and skip-connections, could alleviate these limitations. Through the design and systematic evaluation of alternate projection blocks, we show that better-designed final projections positively impact the downstream performance of ColBERT models. We highlight that many projection variants outperform the original linear projections, with the best-performing variants increasing average performance on a range of retrieval benchmarks across domains by over 2 NDCG@10 points. We then conduct further exploration on the individual parameters of these projections block in order to understand what drives this empirical performance, highlighting the particular importance of upscaled intermediate projections and residual connections. As part of these ablation studies, we show that numerous suboptimal projection variants still outperform the traditional single-layer projection across multiple benchmarks, confirming our hypothesis. Finally, we observe that this effect is consistent across random seeds, further confirming that replacing the linear layer of ColBERT models is a robust, drop-in upgrade.
☆ Causal Inspired Multi Modal Recommendation
Multimodal recommender systems enhance personalized recommendations in e-commerce and online advertising by integrating visual, textual, and user-item interaction data. However, existing methods often overlook two critical biases: (i) modal confounding, where latent factors (e.g., brand style or product category) simultaneously drive multiple modalities and influence user preference, leading to spurious feature-preference associations; (ii) interaction bias, where genuine user preferences are mixed with noise from exposure effects and accidental clicks. To address these challenges, we propose a Causal-inspired multimodal Recommendation framework. Specifically, we introduce a dual-channel cross-modal diffusion module to identify hidden modal confounders, utilize back-door adjustment with hierarchical matching and vector-quantized codebooks to block confounding paths, and apply front-door adjustment combined with causal topology reconstruction to build a deconfounded causal subgraph. Extensive experiments on three real-world e-commerce datasets demonstrate that our method significantly outperforms state-of-the-art baselines while maintaining strong interpretability.
☆ An Empirical Study for Representations of Videos in Video Question Answering via MLLMs
Multimodal large language models have recently achieved remarkable progress in video question answering (VideoQA) by jointly processing visual, textual, and audio information. However, it remains unclear which video representations are most effective for MLLMs, and how different modalities balance task accuracy against computational efficiency. In this work, we present a comprehensive empirical study of video representation methods for VideoQA with MLLMs. We systematically evaluate single modality inputs question only, subtitles, visual frames, and audio signals as well as multimodal combinations, on two widely used benchmarks: VideoMME and LongVideoBench. Our results show that visual frames substantially enhance accuracy but impose heavy costs in GPU memory and inference latency, while subtitles provide a lightweight yet effective alternative, particularly for long videos. These findings highlight clear trade-offs between effectiveness and efficiency and provide practical insights for designing resource-aware MLLM-based VideoQA systems.
comment: 6 pages, 3 figures
☆ Reinforced Preference Optimization for Recommendation
Recent breakthroughs in large language models (LLMs) have fundamentally shifted recommender systems from discriminative to generative paradigms, where user behavior modeling is achieved by generating target items conditioned on historical interactions. Yet current generative recommenders still suffer from two core limitations: the lack of high-quality negative modeling and the reliance on implicit rewards. Reinforcement learning with verifiable rewards (RLVR) offers a natural solution by enabling on-policy sampling of harder negatives and grounding optimization in explicit reward signals. However, applying RLVR to generative recommenders remains non-trivial. Its unique generation space often leads to invalid or repetitive items that undermine sampling efficiency, and ranking supervision is sparse since most items receive identical zero rewards. To address these challenges, we propose Reinforced Preference Optimization for Recommendation (ReRe), a reinforcement-based paradigm tailored to LLM-based recommenders, an important direction in generative recommendation. ReRe incorporates constrained beam search to improve sampling efficiency and diversify hard negatives, while augmenting rule-based accuracy rewards with auxiliary ranking rewards for finer-grained supervision. Extensive experiments on three real-world datasets demonstrate that ReRe consistently outperforms both traditional and LLM-based recommenders in ranking performance. Further analysis shows that ReRe not only enhances performance across both base and SFT-initialized models but also generalizes robustly across different backbone families and scales. Beyond empirical gains, we systematically investigate the design space of RLVR in recommendation across generation, sampling strategy, reward modeling, and optimization algorithm, offering insights for future research.
☆ MIARec: Mutual-influence-aware Heterogeneous Network Embedding for Scientific Paper Recommendation
With the rapid expansion of scientific literature, scholars increasingly demand precise and high-quality paper recommendations. Among various recommendation methodologies, graph-based approaches have garnered attention by effectively exploiting the structural characteristics inherent in scholarly networks. However, these methods often overlook the asymmetric academic influence that is prevalent in scholarly networks when learning graph representations. To address this limitation, this study proposes the Mutual-Influence-Aware Recommendation (MIARec) model, which employs a gravity-based approach to measure the mutual academic influence between scholars and incorporates this influence into the feature aggregation process during message propagation in graph representation learning. Additionally, the model utilizes a multi-channel aggregation method to capture both individual embeddings of distinct single relational sub-networks and their interdependent embeddings, thereby enabling a more comprehensive understanding of the heterogeneous scholarly network. Extensive experiments conducted on real-world datasets demonstrate that the MIARec model outperforms baseline models across three primary evaluation metrics, indicating its effectiveness in scientific paper recommendation tasks.
♻ ☆ Query Brand Entity Linking in E-Commerce Search
In this work, we address the brand entity linking problem for e-commerce search queries. The entity linking task is done by either i)a two-stage process consisting of entity mention detection followed by entity disambiguation or ii) an end-to-end linking approaches that directly fetch the target entity given the input text. The task presents unique challenges: queries are extremely short (averaging 2.4 words), lack natural language structure, and must handle a massive space of unique brands. We present a two-stage approach combining named-entity recognition with matching, and a novel end-to-end solution using extreme multi-class classification. We validate our solutions by both offline benchmarks and the impact of online A/B test.
♻ ☆ A Comprehensive Taxonomy of Negation for NLP and Neural Retrievers
Understanding and solving complex reasoning tasks is vital for addressing the information needs of a user. Although dense neural models learn contextualised embeddings, they still underperform on queries containing negation. To understand this phenomenon, we study negation in both traditional neural information retrieval and LLM-based models. We (1) introduce a taxonomy of negation that derives from philosophical, linguistic, and logical definitions; (2) generate two benchmark datasets that can be used to evaluate the performance of neural information retrieval models and to fine-tune models for a more robust performance on negation; and (3) propose a logic-based classification mechanism that can be used to analyze the performance of retrieval models on existing datasets. Our taxonomy produces a balanced data distribution over negation types, providing a better training setup that leads to faster convergence on the NevIR dataset. Moreover, we propose a classification schema that reveals the coverage of negation types in existing datasets, offering insights into the factors that might affect the generalization of fine-tuned models on negation.
♻ ☆ CiteBART: Learning to Generate Citations for Local Citation Recommendation EMNLP 2025
Local citation recommendation (LCR) suggests a set of papers for a citation placeholder within a given context. The task has evolved as generative approaches have become more promising than the traditional pre-fetch and re-rank-based state-of-the-art approaches. This paper introduces citation-specific pre-training within an encoder-decoder architecture, where author-date citation tokens are masked to learn to reconstruct them to fulfill LCR. There are two variants for this pre-training. In the local context-only base scheme (CiteBART-Base), the citation token in a local context is masked to learn to predict the citation. The global version (CiteBART-Global) extends the local context with the citing paper's title and abstract to enrich the learning signal. CiteBART-Global achieves state-of-the-art performance on LCR benchmarks except for the FullTextPeerRead dataset, which is quite small to see the advantage of generative pre-training. The effect is significant in the larger benchmarks, e.g., Refseer and ArXiv., with the Refseer benchmark-trained model emerging as the best-performing model. We perform comprehensive experiments, including an ablation study, a qualitative analysis, and a taxonomy of hallucinations with detailed statistics. Our analyses confirm that CiteBART-Global has a cross-dataset generalization capability; the macro hallucination rate (MaHR) at the top-3 predictions is 4\%, and when the ground-truth is in the top-k prediction list, the hallucination tendency in the other predictions drops significantly.
comment: This paper has been accepted to the EMNLP 2025 Main Conference. (19 pages, 3 figures, 11 tables)
♻ ☆ DRIFT: Decompose, Retrieve, Illustrate, then Formalize Theorems
Automating the formalization of mathematical statements for theorem proving remains a major challenge for Large Language Models (LLMs). LLMs struggle to identify and utilize the prerequisite mathematical knowledge and its corresponding formal representation in languages like Lean. Current retrieval-augmented autoformalization methods query external libraries using the informal statement directly, but overlook a fundamental limitation: informal mathematical statements are often complex and offer limited context on the underlying math concepts. To address this, we introduce DRIFT, a novel framework that enables LLMs to decompose informal mathematical statements into smaller, more tractable ''sub-components''. This facilitates targeted retrieval of premises from mathematical libraries such as Mathlib. Additionally, DRIFT retrieves illustrative theorems to help models use premises more effectively in formalization tasks. We evaluate DRIFT across diverse benchmarks (ProofNet, ConNF, and MiniF2F-test) and find that it consistently improves premise retrieval, nearly doubling the F1 score compared to the DPR baseline on ProofNet. Notably, DRIFT demonstrates strong performance on the out-of-distribution ConNF benchmark, with BEq+@10 improvements of 37.14% and 42.25% using GPT-4.1 and DeepSeek-V3.1, respectively. Our analysis shows that retrieval effectiveness in mathematical autoformalization depends heavily on model-specific knowledge boundaries, highlighting the need for adaptive retrieval strategies aligned with each model's capabilities.
♻ ☆ Efficient and Versatile Model for Multilingual Information Retrieval of Islamic Text: Development and Deployment in Real-World Scenarios
Despite recent advancements in Multilingual Information Retrieval (MLIR), a significant gap remains between research and practical deployment. Many studies assess MLIR performance in isolated settings, limiting their applicability to real-world scenarios. In this work, we leverage the unique characteristics of the Quranic multilingual corpus to examine the optimal strategies to develop an ad-hoc IR system for the Islamic domain that is designed to satisfy users' information needs in multiple languages. We prepared eleven retrieval models employing four training approaches: monolingual, cross-lingual, translate-train-all, and a novel mixed method combining cross-lingual and monolingual techniques. Evaluation on an in-domain dataset demonstrates that the mixed approach achieves promising results across diverse retrieval scenarios. Furthermore, we provide a detailed analysis of how different training configurations affect the embedding space and their implications for multilingual retrieval effectiveness. Finally, we discuss deployment considerations, emphasizing the cost-efficiency of deploying a single versatile, lightweight model for real-world MLIR applications.
♻ ☆ AdaptJobRec: Enhancing Conversational Career Recommendation through an LLM-Powered Agentic System
In recent years, recommendation systems have evolved from providing a single list of recommendations to offering a comprehensive suite of topic focused services. To better accomplish this task, conversational recommendation systems (CRS) have progressed from basic retrieval augmented LLM generation to agentic systems with advanced reasoning and self correction capabilities. However, agentic systems come with notable response latency, a longstanding challenge for conversational recommendation systems. To balance the trade off between handling complex queries and minimizing latency, we propose AdaptJobRec, the first conversational job recommendation system that leverages autonomous agent to integrate personalized recommendation algorithm tools. The system employs a user query complexity identification mechanism to minimize response latency. For straightforward queries, the agent directly selects the appropriate tool for rapid responses. For complex queries, the agent uses the memory processing module to filter chat history for relevant content, then passes the results to the intelligent task decomposition planner, and finally executes the tasks using personalized recommendation tools. Evaluation on Walmart's real world career recommendation scenarios demonstrates that AdaptJobRec reduces average response latency by up to 53.3% compared to competitive baselines, while significantly improving recommendation accuracy.
Machine Learning 150
☆ CuMPerLay: Learning Cubical Multiparameter Persistence Vectorizations ICCV 2025
We present CuMPerLay, a novel differentiable vectorization layer that enables the integration of Cubical Multiparameter Persistence (CMP) into deep learning pipelines. While CMP presents a natural and powerful way to topologically work with images, its use is hindered by the complexity of multifiltration structures as well as the vectorization of CMP. In face of these challenges, we introduce a new algorithm for vectorizing MP homologies of cubical complexes. Our CuMPerLay decomposes the CMP into a combination of individual, learnable single-parameter persistence, where the bifiltration functions are jointly learned. Thanks to the differentiability, its robust topological feature vectors can be seamlessly used within state-of-the-art architectures such as Swin Transformers. We establish theoretical guarantees for the stability of our vectorization under generalized Wasserstein metrics. Our experiments on benchmark medical imaging and computer vision datasets show the benefit CuMPerLay on classification and segmentation performance, particularly in limited-data scenarios. Overall, CuMPerLay offers a promising direction for integrating global structural information into deep networks for structured image analysis.
comment: Appears at ICCV 2025
☆ UniFusion: Vision-Language Model as Unified Encoder in Image Generation
Although recent advances in visual generation have been remarkable, most existing architectures still depend on distinct encoders for images and text. This separation constrains diffusion models' ability to perform cross-modal reasoning and knowledge transfer. Prior attempts to bridge this gap often use the last layer information from VLM, employ multiple visual encoders, or train large unified models jointly for text and image generation, which demands substantial computational resources and large-scale data, limiting its accessibility.We present UniFusion, a diffusion-based generative model conditioned on a frozen large vision-language model (VLM) that serves as a unified multimodal encoder. At the core of UniFusion is the Layerwise Attention Pooling (LAP) mechanism that extracts both high level semantics and low level details from text and visual tokens of a frozen VLM to condition a diffusion generative model. We demonstrate that LAP outperforms other shallow fusion architectures on text-image alignment for generation and faithful transfer of visual information from VLM to the diffusion model which is key for editing. We propose VLM-Enabled Rewriting Injection with Flexibile Inference (VERIFI), which conditions a diffusion transformer (DiT) only on the text tokens generated by the VLM during in-model prompt rewriting. VERIFI combines the alignment of the conditioning distribution with the VLM's reasoning capabilities for increased capabilities and flexibility at inference. In addition, finetuning on editing task not only improves text-image alignment for generation, indicative of cross-modality knowledge transfer, but also exhibits tremendous generalization capabilities. Our model when trained on single image editing, zero-shot generalizes to multiple image references further motivating the unified encoder design of UniFusion.
comment: Project page at https://thekevinli.github.io/unifusion/
☆ Wavefront Coding for Accommodation-Invariant Near-Eye Displays
We present a new computational near-eye display method that addresses the vergence-accommodation conflict problem in stereoscopic displays through accommodation-invariance. Our system integrates a refractive lens eyepiece with a novel wavefront coding diffractive optical element, operating in tandem with a pre-processing convolutional neural network. We employ end-to-end learning to jointly optimize the wavefront-coding optics and the image pre-processing module. To implement this approach, we develop a differentiable retinal image formation model that accounts for limiting aperture and chromatic aberrations introduced by the eye optics. We further integrate the neural transfer function and the contrast sensitivity function into the loss model to account for related perceptual effects. To tackle off-axis distortions, we incorporate position dependency into the pre-processing module. In addition to conducting rigorous analysis based on simulations, we also fabricate the designed diffractive optical element and build a benchtop setup, demonstrating accommodation-invariance for depth ranges of up to four diopters.
☆ Dr.LLM: Dynamic Layer Routing in LLMs
Large Language Models (LLMs) process every token through all layers of a transformer stack, causing wasted computation on simple queries and insufficient flexibility for harder ones that need deeper reasoning. Adaptive-depth methods can improve efficiency, but prior approaches rely on costly inference-time search, architectural changes, or large-scale retraining, and in practice often degrade accuracy despite efficiency gains. We introduce Dr.LLM, Dynamic routing of Layers for LLMs, a retrofittable framework that equips pretrained models with lightweight per-layer routers deciding to skip, execute, or repeat a block. Routers are trained with explicit supervision: using Monte Carlo Tree Search (MCTS), we derive high-quality layer configurations that preserve or improve accuracy under a compute budget. Our design, windowed pooling for stable routing, focal loss with class balancing, and bottleneck MLP routers, ensures robustness under class imbalance and long sequences. On ARC (logic) and DART (math), Dr.LLM improves accuracy by up to +3.4%p while saving 5 layers per example on average. Routers generalize to out-of-domain tasks (MMLU, GSM8k, AIME, TruthfulQA, SQuADv2, GPQA, PIQA, AGIEval) with only 0.85% accuracy drop while retaining efficiency, and outperform prior routing methods by up to +7.7%p. Overall, Dr.LLM shows that explicitly supervised routers retrofit frozen LLMs for budget-aware, accuracy-driven inference without altering base weights.
comment: 17 pages, Under submission
☆ Sample-Efficient Omniprediction for Proper Losses
We consider the problem of constructing probabilistic predictions that lead to accurate decisions when employed by downstream users to inform actions. For a single decision maker, designing an optimal predictor is equivalent to minimizing a proper loss function corresponding to the negative utility of that individual. For multiple decision makers, our problem can be viewed as a variant of omniprediction in which the goal is to design a single predictor that simultaneously minimizes multiple losses. Existing algorithms for achieving omniprediction broadly fall into two categories: 1) boosting methods that optimize other auxiliary targets such as multicalibration and obtain omniprediction as a corollary, and 2) adversarial two-player game based approaches that estimate and respond to the ``worst-case" loss in an online fashion. We give lower bounds demonstrating that multicalibration is a strictly more difficult problem than omniprediction and thus the former approach must incur suboptimal sample complexity. For the latter approach, we discuss how these ideas can be used to obtain a sample-efficient algorithm through an online-to-batch conversion. This conversion has the downside of returning a complex, randomized predictor. We improve on this method by designing a more direct, unrandomized algorithm that exploits structural elements of the set of proper losses.
☆ AnyUp: Universal Feature Upsampling
We introduce AnyUp, a method for feature upsampling that can be applied to any vision feature at any resolution, without encoder-specific training. Existing learning-based upsamplers for features like DINO or CLIP need to be re-trained for every feature extractor and thus do not generalize to different feature types at inference time. In this work, we propose an inference-time feature-agnostic upsampling architecture to alleviate this limitation and improve upsampling quality. In our experiments, AnyUp sets a new state of the art for upsampled features, generalizes to different feature types, and preserves feature semantics while being efficient and easy to apply to a wide range of downstream tasks.
comment: Project Website: https://wimmerth.github.io/anyup/
☆ KoALA: KL-L0 Adversarial Detector via Label Agreement
Deep neural networks are highly susceptible to adversarial attacks, which pose significant risks to security- and safety-critical applications. We present KoALA (KL-L0 Adversarial detection via Label Agreement), a novel, semantics-free adversarial detector that requires no architectural changes or adversarial retraining. KoALA operates on a simple principle: it detects an adversarial attack when class predictions from two complementary similarity metrics disagree. These metrics-KL divergence and an L0-based similarity-are specifically chosen to detect different types of perturbations. The KL divergence metric is sensitive to dense, low-amplitude shifts, while the L0-based similarity is designed for sparse, high-impact changes. We provide a formal proof of correctness for our approach. The only training required is a simple fine-tuning step on a pre-trained image encoder using clean images to ensure the embeddings align well with both metrics. This makes KOALA a lightweight, plug-and-play solution for existing models and various data modalities. Our extensive experiments on ResNet/CIFAR-10 and CLIP/Tiny-ImageNet confirm our theoretical claims. When the theorem's conditions are met, KoALA consistently and effectively detects adversarial examples. On the full test sets, KoALA achieves a precision of 0.94 and a recall of 0.81 on ResNet/CIFAR-10, and a precision of 0.66 and a recall of 0.85 on CLIP/Tiny-ImageNet.
☆ VQArt-Bench: A semantically rich VQA Benchmark for Art and Cultural Heritage
Multimodal Large Language Models (MLLMs) have demonstrated significant capabilities in joint visual and linguistic tasks. However, existing Visual Question Answering (VQA) benchmarks often fail to evaluate deep semantic understanding, particularly in complex domains like visual art analysis. Confined to simple syntactic structures and surface-level attributes, these questions fail to capture the diversity and depth of human visual inquiry. This limitation incentivizes models to exploit statistical shortcuts rather than engage in visual reasoning. To address this gap, we introduce VQArt-Bench, a new, large-scale VQA benchmark for the cultural heritage domain. This benchmark is constructed using a novel multi-agent pipeline where specialized agents collaborate to generate nuanced, validated, and linguistically diverse questions. The resulting benchmark is structured along relevant visual understanding dimensions that probe a model's ability to interpret symbolic meaning, narratives, and complex visual relationships. Our evaluation of 14 state-of-the-art MLLMs on this benchmark reveals significant limitations in current models, including a surprising weakness in simple counting tasks and a clear performance gap between proprietary and open-source models.
☆ Dendrograms of Mixing Measures for Softmax-Gated Gaussian Mixture of Experts: Consistency without Model Sweeps
We develop a unified statistical framework for softmax-gated Gaussian mixture of experts (SGMoE) that addresses three long-standing obstacles in parameter estimation and model selection: (i) non-identifiability of gating parameters up to common translations, (ii) intrinsic gate-expert interactions that induce coupled differential relations in the likelihood, and (iii) the tight numerator-denominator coupling in the softmax-induced conditional density. Our approach introduces Voronoi-type loss functions aligned with the gate-partition geometry and establishes finite-sample convergence rates for the maximum likelihood estimator (MLE). In over-specified models, we reveal a link between the MLE's convergence rate and the solvability of an associated system of polynomial equations characterizing near-nonidentifiable directions. For model selection, we adapt dendrograms of mixing measures to SGMoE, yielding a consistent, sweep-free selector of the number of experts that attains pointwise-optimal parameter rates under overfitting while avoiding multi-size training. Simulations on synthetic data corroborate the theory, accurately recovering the expert count and achieving the predicted rates for parameter estimation while closely approximating the regression function. Under model misspecification (e.g., $\epsilon$-contamination), the dendrogram selection criterion is robust, recovering the true number of mixture components, while the Akaike information criterion, the Bayesian information criterion, and the integrated completed likelihood tend to overselect as sample size grows. On a maize proteomics dataset of drought-responsive traits, our dendrogram-guided SGMoE selects two experts, exposes a clear mixing-measure hierarchy, stabilizes the likelihood early, and yields interpretable genotype-phenotype maps, outperforming standard criteria without multi-size training.
comment: Do Tien Hai, Trung Nguyen Mai, and TrungTin Nguyen are co-first authors
☆ Doctor Rashomon and the UNIVERSE of Madness: Variable Importance with Unobserved Confounding and the Rashomon Effect
Variable importance (VI) methods are often used for hypothesis generation, feature selection, and scientific validation. In the standard VI pipeline, an analyst estimates VI for a single predictive model with only the observed features. However, the importance of a feature depends heavily on which other variables are included in the model, and essential variables are often omitted from observational datasets. Moreover, the VI estimated for one model is often not the same as the VI estimated for another equally-good model - a phenomenon known as the Rashomon Effect. We address these gaps by introducing UNobservables and Inference for Variable importancE using Rashomon SEts (UNIVERSE). Our approach adapts Rashomon sets - the sets of near-optimal models in a dataset - to produce bounds on the true VI even with missing features. We theoretically guarantee the robustness of our approach, show strong performance on semi-synthetic simulations, and demonstrate its utility in a credit risk task.
☆ HYPE: Hybrid Planning with Ego Proposal-Conditioned Predictions
Safe and interpretable motion planning in complex urban environments needs to reason about bidirectional multi-agent interactions. This reasoning requires to estimate the costs of potential ego driving maneuvers. Many existing planners generate initial trajectories with sampling-based methods and refine them by optimizing on learned predictions of future environment states, which requires a cost function that encodes the desired vehicle behavior. Designing such a cost function can be very challenging, especially if a wide range of complex urban scenarios has to be considered. We propose HYPE: HYbrid Planning with Ego proposal-conditioned predictions, a planner that integrates multimodal trajectory proposals from a learned proposal model as heuristic priors into a Monte Carlo Tree Search (MCTS) refinement. To model bidirectional interactions, we introduce an ego-conditioned occupancy prediction model, enabling consistent, scene-aware reasoning. Our design significantly simplifies cost function design in refinement by considering proposal-driven guidance, requiring only minimalistic grid-based cost terms. Evaluations on large-scale real-world benchmarks nuPlan and DeepUrban show that HYPE effectively achieves state-of-the-art performance, especially in safety and adaptability.
☆ Data-Model Co-Evolution: Growing Test Sets to Refine LLM Behavior
A long-standing challenge in machine learning has been the rigid separation between data work and model refinement, enforced by slow fine-tuning cycles. The rise of Large Language Models (LLMs) overcomes this historical barrier, allowing applications developers to instantly govern model behavior by editing prompt instructions. This shift enables a new paradigm: data-model co-evolution, where a living test set and a model's instructions evolve in tandem. We operationalize this paradigm in an interactive system designed to address the critical challenge of encoding subtle, domain-specific policies into prompt instructions. The system's structured workflow guides people to discover edge cases, articulate rationales for desired behavior, and iteratively evaluate instruction revisions against a growing test set. A user study shows our workflow helps participants refine instructions systematically and specify ambiguous policies more concretely. This work points toward more robust and responsible LLM applications through human-in-the-loop development aligned with local preferences and policies.
☆ Hierarchical Federated Learning for Crop Yield Prediction in Smart Agricultural Production Systems
In this paper, we presents a novel hierarchical federated learning architecture specifically designed for smart agricultural production systems and crop yield prediction. Our approach introduces a seasonal subscription mechanism where farms join crop-specific clusters at the beginning of each agricultural season. The proposed three-layer architecture consists of individual smart farms at the client level, crop-specific aggregators at the middle layer, and a global model aggregator at the top level. Within each crop cluster, clients collaboratively train specialized models tailored to specific crop types, which are then aggregated to produce a higher-level global model that integrates knowledge across multiple crops. This hierarchical design enables both local specialization for individual crop types and global generalization across diverse agricultural contexts while preserving data privacy and reducing communication overhead. Experiments demonstrate the effectiveness of the proposed system, showing that local and crop-layer models closely follow actual yield patterns with consistent alignment, significantly outperforming standard machine learning models. The results validate the advantages of hierarchical federated learning in the agricultural context, particularly for scenarios involving heterogeneous farming environments and privacy-sensitive agricultural data.
comment: 6 pages, 3 figures, conference
☆ Improving Decision Trees through the Lens of Parameterized Local Search NeurIPS 2025
Algorithms for learning decision trees often include heuristic local-search operations such as (1) adjusting the threshold of a cut or (2) also exchanging the feature of that cut. We study minimizing the number of classification errors by performing a fixed number of a single type of these operations. Although we discover that the corresponding problems are NP-complete in general, we provide a comprehensive parameterized-complexity analysis with the aim of determining those properties of the problems that explain the hardness and those that make the problems tractable. For instance, we show that the problems remain hard for a small number $d$ of features or small domain size $D$ but the combination of both yields fixed-parameter tractability. That is, the problems are solvable in $(D + 1)^{2d} \cdot |I|^{O(1)}$ time, where $|I|$ is the size of the input. We also provide a proof-of-concept implementation of this algorithm and report on empirical results.
comment: Accepted at NeurIPS 2025
☆ CARVQ: Corrective Adaptor with Group Residual Vector Quantization for LLM Embedding Compression EMNLP
Large Language Models (LLMs) typically rely on a large number of parameters for token embedding, leading to substantial storage requirements and memory footprints. In particular, LLMs deployed on edge devices are memory-bound, and reducing the memory footprint by compressing the embedding layer not only frees up the memory bandwidth but also speeds up inference. To address this, we introduce CARVQ, a post-training novel Corrective Adaptor combined with group Residual Vector Quantization. CARVQ relies on the composition of both linear and non-linear maps and mimics the original model embedding to compress to approximately 1.6 bits without requiring specialized hardware to support lower-bit storage. We test our method on pre-trained LLMs such as LLaMA-3.2-1B, LLaMA-3.2-3B, LLaMA-3.2-3B-Instruct, LLaMA-3.1-8B, Qwen2.5-7B, Qwen2.5-Math-7B and Phi-4, evaluating on common generative, discriminative, math and reasoning tasks. We show that in most cases, CARVQ can achieve lower average bitwidth-per-parameter while maintaining reasonable perplexity and accuracy compared to scalar quantization. Our contributions include a novel compression technique that is compatible with state-of-the-art transformer quantization methods and can be seamlessly integrated into any hardware supporting 4-bit memory to reduce the model's memory footprint in memory-constrained devices. This work demonstrates a crucial step toward the efficient deployment of LLMs on edge devices.
comment: Accepted at EMNLP Findings 2025
☆ Multitask finetuning and acceleration of chemical pretrained models for small molecule drug property prediction
Chemical pretrained models, sometimes referred to as foundation models, are receiving considerable interest for drug discovery applications. The general chemical knowledge extracted from self-supervised training has the potential to improve predictions for critical drug discovery endpoints, including on-target potency and ADMET properties. Multi-task learning has previously been successfully leveraged to improve predictive models. Here, we show that enabling multitasking in finetuning of chemical pretrained graph neural network models such as Kinetic GROVER Multi-Task (KERMT), an enhanced version of the GROVER model, and Knowledge-guided Pre-training of Graph Transformer (KGPT) significantly improves performance over non-pretrained graph neural network models. Surprisingly, we find that the performance improvement from finetuning KERMT in a multitask manner is most significant at larger data sizes. Additionally, we publish two multitask ADMET data splits to enable more accurate benchmarking of multitask deep learning methods for drug property prediction. Finally, we provide an accelerated implementation of the KERMT model on GitHub, unlocking large-scale pretraining, finetuning, and inference in industrial drug discovery workflows.
☆ Topological Signatures of ReLU Neural Network Activation Patterns
This paper explores the topological signatures of ReLU neural network activation patterns. We consider feedforward neural networks with ReLU activation functions and analyze the polytope decomposition of the feature space induced by the network. Mainly, we investigate how the Fiedler partition of the dual graph and show that it appears to correlate with the decision boundary -- in the case of binary classification. Additionally, we compute the homology of the cellular decomposition -- in a regression task -- to draw similar patterns in behavior between the training loss and polyhedral cell-count, as the model is trained.
☆ Who is a Better Matchmaker? Human vs. Algorithmic Judge Assignment in a High-Stakes Startup Competition
There is growing interest in applying artificial intelligence (AI) to automate and support complex decision-making tasks. However, it remains unclear how algorithms compare to human judgment in contexts requiring semantic understanding and domain expertise. We examine this in the context of the judge assignment problem, matching submissions to suitably qualified judges. Specifically, we tackled this problem at the Harvard President's Innovation Challenge, the university's premier venture competition awarding over \$500,000 to student and alumni startups. This represents a real-world environment where high-quality judge assignment is essential. We developed an AI-based judge-assignment algorithm, Hybrid Lexical-Semantic Similarity Ensemble (HLSE), and deployed it at the competition. We then evaluated its performance against human expert assignments using blinded match-quality scores from judges on $309$ judge-venture pairs. Using a Mann-Whitney U statistic based test, we found no statistically significant difference in assignment quality between the two approaches ($AUC=0.48, p=0.40$); on average, algorithmic matches are rated $3.90$ and manual matches $3.94$ on a 5-point scale, where 5 indicates an excellent match. Furthermore, manual assignments that previously required a full week could be automated in several hours by the algorithm during deployment. These results demonstrate that HLSE achieves human-expert-level matching quality while offering greater scalability and efficiency, underscoring the potential of AI-driven solutions to support and enhance human decision-making for judge assignment in high-stakes settings.
comment: 17 Pages, 2 figures
☆ DiffEM: Learning from Corrupted Data with Diffusion Models via Expectation Maximization
Diffusion models have emerged as powerful generative priors for high-dimensional inverse problems, yet learning them when only corrupted or noisy observations are available remains challenging. In this work, we propose a new method for training diffusion models with Expectation-Maximization (EM) from corrupted data. Our proposed method, DiffEM, utilizes conditional diffusion models to reconstruct clean data from observations in the E-step, and then uses the reconstructed data to refine the conditional diffusion model in the M-step. Theoretically, we provide monotonic convergence guarantees for the DiffEM iteration, assuming appropriate statistical conditions. We demonstrate the effectiveness of our approach through experiments on various image reconstruction tasks.
☆ EReLiFM: Evidential Reliability-Aware Residual Flow Meta-Learning for Open-Set Domain Generalization under Noisy Labels
Open-Set Domain Generalization (OSDG) aims to enable deep learning models to recognize unseen categories in new domains, which is crucial for real-world applications. Label noise hinders open-set domain generalization by corrupting source-domain knowledge, making it harder to recognize known classes and reject unseen ones. While existing methods address OSDG under Noisy Labels (OSDG-NL) using hyperbolic prototype-guided meta-learning, they struggle to bridge domain gaps, especially with limited clean labeled data. In this paper, we propose Evidential Reliability-Aware Residual Flow Meta-Learning (EReLiFM). We first introduce an unsupervised two-stage evidential loss clustering method to promote label reliability awareness. Then, we propose a residual flow matching mechanism that models structured domain- and category-conditioned residuals, enabling diverse and uncertainty-aware transfer paths beyond interpolation-based augmentation. During this meta-learning process, the model is optimized such that the update direction on the clean set maximizes the loss decrease on the noisy set, using pseudo labels derived from the most confident predicted class for supervision. Experimental results show that EReLiFM outperforms existing methods on OSDG-NL, achieving state-of-the-art performance. The source code is available at https://github.com/KPeng9510/ERELIFM.
comment: The source code is available at https://github.com/KPeng9510/ERELIFM
☆ Few Shot Semi-Supervised Learning for Abnormal Stop Detection from Sparse GPS Trajectories
Abnormal stop detection (ASD) in intercity coach transportation is critical for ensuring passenger safety, operational reliability, and regulatory compliance. However, two key challenges hinder ASD effectiveness: sparse GPS trajectories, which obscure short or unauthorized stops, and limited labeled data, which restricts supervised learning. Existing methods often assume dense sampling or regular movement patterns, limiting their applicability. To address data sparsity, we propose a Sparsity-Aware Segmentation (SAS) method that adaptively defines segment boundaries based on local spatial-temporal density. Building upon these segments, we introduce three domain-specific indicators to capture abnormal stop behaviors. To further mitigate the impact of sparsity, we develop Locally Temporal-Indicator Guided Adjustment (LTIGA), which smooths these indicators via local similarity graphs. To overcome label scarcity, we construct a spatial-temporal graph where each segment is a node with LTIGA-refined features. We apply label propagation to expand weak supervision across the graph, followed by a GCN to learn relational patterns. A final self-training module incorporates high-confidence pseudo-labels to iteratively improve predictions. Experiments on real-world coach data show an AUC of 0.854 and AP of 0.866 using only 10 labeled instances, outperforming prior methods. The code and dataset are publicly available at \href{https://github.com/pangjunbiao/Abnormal-Stop-Detection-SSL.git}
☆ CoRA: Covariate-Aware Adaptation of Time Series Foundation Models
Time Series Foundation Models (TSFMs) have shown significant impact through their model capacity, scalability, and zero-shot generalization. However, due to the heterogeneity of inter-variate dependencies and the backbone scalability on large-scale multivariate datasets, most TSFMs are typically pre-trained on univariate time series. This limitation renders them oblivious to crucial information from diverse covariates in real-world forecasting tasks. To further enhance the performance of TSFMs, we propose a general covariate-aware adaptation (CoRA) framework for TSFMs. It leverages pre-trained backbones of foundation models while effectively incorporating exogenous covariates from various modalities, including time series, language, and images, to improve the quality of predictions. Technically, CoRA maintains the equivalence of initialization and parameter consistency during adaptation. With preserved backbones of foundation models as frozen feature extractors, the outcome embeddings from foundation models are empirically demonstrated more informative than raw data. Further, CoRA employs a novel Granger Causality Embedding (GCE) to automatically evaluate covariates regarding their causal predictability with respect to the target variate. We incorporate these weighted embeddings with a zero-initialized condition-injection mechanism, avoiding catastrophic forgetting of pre-trained foundation models and gradually integrates exogenous information. Extensive experiments show that CoRA of TSFMs surpasses state-of-the-art covariate-aware deep forecasters with full or few-shot training samples, achieving 31.1% MSE reduction on covariate-aware forecasting. Compared to other adaptation methods, CoRA exhibits strong compatibility with various advanced TSFMs and extends the scope of covariates to other modalities, presenting a practical paradigm for the application of TSFMs.
☆ Demystifying Hybrid Thinking: Can LLMs Truly Switch Between Think and No-Think?
Hybrid thinking enables LLMs to switch between reasoning and direct answering, offering a balance between efficiency and reasoning capability. Yet our experiments reveal that current hybrid thinking LLMs only achieve partial mode separation: reasoning behaviors often leak into the no-think mode. To understand and mitigate this, we analyze the factors influencing controllability and identify four that matter most: (1) larger data scale, (2) using think and no-think answers from different questions rather than the same question, (3) a moderate increase in no-think data number, and (4) a two-phase strategy that first trains reasoning ability and then applies hybrid think training. Building on these findings, we propose a practical recipe that, compared to standard training, can maintain accuracy in both modes while significantly reducing no-think output length (from $1085$ to $585$ on MATH500) and occurrences of reasoning-supportive tokens such as ``\texttt{wait}'' (from $5917$ to $522$ on MATH500). Our findings highlight the limitations of current hybrid thinking and offer directions for strengthening its controllability.
comment: 10 pages, 6 figures
☆ Keep Calm and Avoid Harmful Content: Concept Alignment and Latent Manipulation Towards Safer Answers
Large Language Models are susceptible to jailbreak attacks that bypass built-in safety guardrails (e.g., by tricking the model with adversarial prompts). We propose Concept Alignment and Concept Manipulation \textbf{CALM}, an inference-time method that suppresses harmful concepts by modifying latent representations of the last layer of the model, without retraining. Leveraging \gls*{cw} technique from Computer Vision combined with orthogonal projection, CALM removes unwanted latent directions associated with harmful content while preserving model performance. Experiments show that CALM reduces harmful outputs and outperforms baseline methods in most metrics, offering a lightweight approach to AI safety with no additional training data or model fine-tuning, while incurring only a small computational overhead at inference.
☆ Structure-Aware Spectral Sparsification via Uniform Edge Sampling NeurIPS 2025
Spectral clustering is a fundamental method for graph partitioning, but its reliance on eigenvector computation limits scalability to massive graphs. Classical sparsification methods preserve spectral properties by sampling edges proportionally to their effective resistances, but require expensive preprocessing to estimate these resistances. We study whether uniform edge sampling-a simple, structure-agnostic strategy-can suffice for spectral clustering. Our main result shows that for graphs admitting a well-separated $k$-clustering, characterized by a large structure ratio $\Upsilon(k) = \lambda_{k+1} / \rho_G(k)$, uniform sampling preserves the spectral subspace used for clustering. Specifically, we prove that uniformly sampling $O(\gamma^2 n \log n / \epsilon^2)$ edges, where $\gamma$ is the Laplacian condition number, yields a sparsifier whose top $(n-k)$-dimensional eigenspace is approximately orthogonal to the cluster indicators. This ensures that the spectral embedding remains faithful, and clustering quality is preserved. Our analysis introduces new resistance bounds for intra-cluster edges, a rank-$(n-k)$ effective resistance formulation, and a matrix Chernoff bound adapted to the dominant eigenspace. These tools allow us to bypass importance sampling entirely. Conceptually, our result connects recent coreset-based clustering theory to spectral sparsification, showing that under strong clusterability, even uniform sampling is structure-aware. This provides the first provable guarantee that uniform edge sampling suffices for structure-preserving spectral clustering.
comment: 19 pages, 4 figures, NeurIPS 2025
☆ Structured Sparsity and Weight-adaptive Pruning for Memory and Compute efficient Whisper models
Whisper models have achieved remarkable progress in speech recognition; yet their large size remains a bottleneck for deployment on resource-constrained edge devices. This paper proposes a framework to design fine-tuned variants of Whisper which address the above problem. Structured sparsity is enforced via the Sparse Group LASSO penalty as a loss regularizer, to reduce the number of FLOating Point operations (FLOPs). Further, a weight statistics aware pruning algorithm is proposed. We also design our custom text normalizer for WER evaluation. On Common Voice 11.0 Hindi dataset, we obtain, without degrading WER, (a) 35.4% reduction in model parameters, 14.25% lower memory consumption and 18.5% fewer FLOPs on Whisper-small, and (b) 31% reduction in model parameters, 15.29% lower memory consumption and 16.95% fewer FLOPs on Whisper-medium; and, (c) substantially outperform the state-of-the-art Iterative Magnitude Pruning based method by pruning 18.7% more parameters along with a 12.31 reduction in WER.
☆ SG-XDEAT: Sparsity-Guided Cross-Dimensional and Cross-Encoding Attention with Target-Aware Conditioning in Tabular Learning
We propose SG-XDEAT (Sparsity-Guided Cross Dimensional and Cross-Encoding Attention with Target Aware Conditioning), a novel framework designed for supervised learning on tabular data. At its core, SG-XDEAT employs a dual-stream encoder that decomposes each input feature into two parallel representations: a raw value stream and a target-conditioned (label-aware) stream. These dual representations are then propagated through a hierarchical stack of attention-based modules. SG-XDEAT integrates three key components: (i) Cross-Dimensional self-attention, which captures intra-view dependencies among features within each stream; (ii) Cross-Encoding self-attention, which enables bidirectional interaction between raw and target-aware representations; and (iii) an Adaptive Sparse Self-Attention (ASSA) mechanism, which dynamically suppresses low-utility tokens by driving their attention weights toward zero--thereby mitigating the impact of noise. Empirical results on multiple public benchmarks show consistent gains over strong baselines, confirming that jointly modeling raw and target-aware views--while adaptively filtering noise--yields a more robust deep tabular learner.
☆ Towards Foundation Inference Models that Learn ODEs In-Context
Ordinary differential equations (ODEs) describe dynamical systems evolving deterministically in continuous time. Accurate data-driven modeling of systems as ODEs, a central problem across the natural sciences, remains challenging, especially if the data is sparse or noisy. We introduce FIM-ODE (Foundation Inference Model for ODEs), a pretrained neural model designed to estimate ODEs zero-shot (i.e., in context) from sparse and noisy observations. Trained on synthetic data, the model utilizes a flexible neural operator for robust ODE inference, even from corrupted data. We empirically verify that FIM-ODE provides accurate estimates, on par with a neural state-of-the-art method, and qualitatively compare the structure of their estimated vector fields.
☆ On Foundation Models for Temporal Point Processes to Accelerate Scientific Discovery
Many scientific fields, from medicine to seismology, rely on analyzing sequences of events over time to understand complex systems. Traditionally, machine learning models must be built and trained from scratch for each new dataset, which is a slow and costly process. We introduce a new approach: a single, powerful model that learns the underlying patterns of event data in context. We trained this "foundation model" on millions of simulated event sequences, teaching it a general-purpose understanding of how events can unfold. As a result, our model can analyze new scientific data instantly, without retraining, simply by looking at a few examples from the dataset. It can also be quickly fine-tuned for even higher accuracy. This approach makes sophisticated event analysis more accessible and accelerates the pace of scientific discovery.
☆ Contraction and entropy production in continuous-time Sinkhorn dynamics
Recently, the vanishing-step-size limit of the Sinkhorn algorithm at finite regularization parameter $\varepsilon$ was shown to be a mirror descent in the space of probability measures. We give $L^2$ contraction criteria in two time-dependent metrics induced by the mirror Hessian, which reduce to the coercivity of certain conditional expectation operators. We then give an exact identity for the entropy production rate of the Sinkhorn flow, which was previously known only to be nonpositive. Examining this rate shows that the standard semigroup analysis of diffusion processes extends systematically to the Sinkhorn flow. We show that the flow induces a reversible Markov dynamics on the target marginal as an Onsager gradient flow. We define the Dirichlet form associated to its (nonlocal) infinitesimal generator, prove a Poincar\'e inequality for it, and show that the spectral gap is strictly positive along the Sinkhorn flow whenever $\varepsilon > 0$. Lastly, we show that the entropy decay is exponential if and only if a logarithmic Sobolev inequality (LSI) holds. We give for illustration two immediate practical use-cases for the Sinkhorn LSI: as a design principle for the latent space in which generative models are trained, and as a stopping heuristic for discrete-time algorithms.
comment: 10 pages excluding references
☆ Expert or not? assessing data quality in offline reinforcement learning
Offline reinforcement learning (RL) learns exclusively from static datasets, without further interaction with the environment. In practice, such datasets vary widely in quality, often mixing expert, suboptimal, and even random trajectories. The choice of algorithm therefore depends on dataset fidelity. Behavior cloning can suffice on high-quality data, whereas mixed- or low-quality data typically benefits from offline RL methods that stitch useful behavior across trajectories. Yet in the wild it is difficult to assess dataset quality a priori because the data's provenance and skill composition are unknown. We address the problem of estimating offline dataset quality without training an agent. We study a spectrum of proxies from simple cumulative rewards to learned value based estimators, and introduce the Bellman Wasserstein distance (BWD), a value aware optimal transport score that measures how dissimilar a dataset's behavioral policy is from a random reference policy. BWD is computed from a behavioral critic and a state conditional OT formulation, requiring no environment interaction or full policy optimization. Across D4RL MuJoCo tasks, BWD strongly correlates with an oracle performance score that aggregates multiple offline RL algorithms, enabling efficient prediction of how well standard agents will perform on a given dataset. Beyond prediction, integrating BWD as a regularizer during policy optimization explicitly pushes the learned policy away from random behavior and improves returns. These results indicate that value aware, distributional signals such as BWD are practical tools for triaging offline RL datasets and policy optimization.
☆ Adapting Noise to Data: Generative Flows from 1D Processes
We introduce a general framework for constructing generative models using one-dimensional noising processes. Beyond diffusion processes, we outline examples that demonstrate the flexibility of our approach. Motivated by this, we propose a novel framework in which the 1D processes themselves are learnable, achieved by parameterizing the noise distribution through quantile functions that adapt to the data. Our construction integrates seamlessly with standard objectives, including Flow Matching and consistency models. Learning quantile-based noise naturally captures heavy tails and compact supports when present. Numerical experiments highlight both the flexibility and the effectiveness of our method.
☆ Laminar: A Scalable Asynchronous RL Post-Training Framework
Reinforcement learning (RL) post-training for Large Language Models (LLMs) is now scaling to large clusters and running for extended durations to enhance model reasoning performance. However, the scalability of existing RL frameworks is limited, as extreme long-tail skewness in RL trajectory generation causes severe GPU underutilization. Current asynchronous RL systems attempt to mitigate this, but they rely on global weight synchronization between the actor and all rollouts, which creates a rigid model update schedule. This global synchronization is ill-suited for the highly skewed and evolving distribution of trajectory generation latency in RL training, crippling training efficiency. Our key insight is that efficient scaling requires breaking this lockstep through trajectory-level asynchrony, which generates and consumes each trajectory independently. We propose Laminar, a scalable and robust RL post-training system built on a fully decoupled architecture. First, we replace global updates with a tier of relay workers acting as a distributed parameter service. This enables asynchronous and fine-grained weight synchronization, allowing rollouts to pull the latest weight anytime without stalling the actor's training loop. Second, a dynamic repack mechanism consolidates long-tail trajectories onto a few dedicated rollouts, maximizing generation throughput. The fully decoupled design also isolates failures, ensuring robustness for long-running jobs. Our evaluation on a 1024-GPU cluster shows that Laminar achieves up to 5.48$\times$ training throughput speedup over state-of-the-art systems, while reducing model convergence time.
☆ Learning-To-Measure: In-context Active Feature Acquisition
Active feature acquisition (AFA) is a sequential decision-making problem where the goal is to improve model performance for test instances by adaptively selecting which features to acquire. In practice, AFA methods often learn from retrospective data with systematic missingness in the features and limited task-specific labels. Most prior work addresses acquisition for a single predetermined task, limiting scalability. To address this limitation, we formalize the meta-AFA problem, where the goal is to learn acquisition policies across various tasks. We introduce Learning-to-Measure (L2M), which consists of i) reliable uncertainty quantification over unseen tasks, and ii) an uncertainty-guided greedy feature acquisition agent that maximizes conditional mutual information. We demonstrate a sequence-modeling or autoregressive pre-training approach that underpins reliable uncertainty quantification for tasks with arbitrary missingness. L2M operates directly on datasets with retrospective missingness and performs the meta-AFA task in-context, eliminating per-task retraining. Across synthetic and real-world tabular benchmarks, L2M matches or surpasses task-specific baselines, particularly under scarce labels and high missingness.
☆ Towards Fast Coarse-graining and Equation Discovery with Foundation Inference Models
High-dimensional recordings of dynamical processes are often characterized by a much smaller set of effective variables, evolving on low-dimensional manifolds. Identifying these latent dynamics requires solving two intertwined problems: discovering appropriate coarse-grained variables and simultaneously fitting the governing equations. Most machine learning approaches tackle these tasks jointly by training autoencoders together with models that enforce dynamical consistency. We propose to decouple the two problems by leveraging the recently introduced Foundation Inference Models (FIMs). FIMs are pretrained models that estimate the infinitesimal generators of dynamical systems (e.g., the drift and diffusion of a stochastic differential equation) in zero-shot mode. By amortizing the inference of the dynamics through a FIM with frozen weights, and training only the encoder-decoder map, we define a simple, simulation-consistent loss that stabilizes representation learning. A proof of concept on a stochastic double-well system with semicircle diffusion, embedded into synthetic video data, illustrates the potential of this approach for fast and reusable coarse-graining pipelines.
☆ Same model, better performance: the impact of shuffling on DNA Language Models benchmarking
Large Language Models are increasingly popular in genomics due to their potential to decode complex biological sequences. Hence, researchers require a standardized benchmark to evaluate DNA Language Models (DNA LMs) capabilities. However, evaluating DNA LMs is a complex task that intersects genomic's domain-specific challenges and machine learning methodologies, where seemingly minor implementation details can significantly compromise benchmark validity. We demonstrate this through BEND (Benchmarking DNA Language Models), where hardware-dependent hyperparameters -- number of data loading workers and buffer sizes -- create spurious performance variations of up to 4% for identical models. The problem stems from inadequate data shuffling interacting with domain specific data characteristics. Experiments with three DNA language models (HyenaDNA, DNABERT-2, ResNet-LM) show these artifacts affect both absolute performance and relative model rankings. We propose a simple solution: pre-shuffling data before storage eliminates hardware dependencies while maintaining efficiency. This work highlights how standard ML practices can interact unexpectedly with domain-specific data characteristics, with broader implications for benchmark design in specialized domains.
☆ Rethinking Knowledge Distillation: A Data Dependent Regulariser With a Negative Asymmetric Payoff
Knowledge distillation is often considered a compression mechanism when judged on the resulting student's accuracy and loss, yet its functional impact is poorly understood. In this work, we quantify the compression capacity of knowledge distillation and the resulting knowledge transfer from a functional perspective, decoupling compression from architectural reduction, which provides an improved understanding of knowledge distillation. We employ hypothesis testing, controls, and random control distillation to understand knowledge transfer mechanisms across data modalities. To rigorously test the breadth and limits of our analyses, we explore multiple distillation variants and analyse distillation scaling laws across model sizes. Our findings demonstrate that, while there is statistically significant knowledge transfer in some modalities and architectures, the extent of this transfer is less pronounced than anticipated, even under conditions designed to maximise knowledge sharing. Notably, in cases of significant knowledge transfer, we identify a consistent and severe asymmetric transfer of negative knowledge to the student, raising safety concerns in knowledge distillation applications. Across 12 experimental setups, 9 architectures, and 7 datasets, our findings show that knowledge distillation functions less as a compression mechanism and more as a data-dependent regulariser with a negative asymmetric payoff.
comment: 45 pages, 24 figures and 104 tables
☆ Research in Collaborative Learning Does Not Serve Cross-Silo Federated Learning in Practice
Cross-silo federated learning (FL) is a promising approach to enable cross-organization collaboration in machine learning model development without directly sharing private data. Despite growing organizational interest driven by data protection regulations such as GDPR and HIPAA, the adoption of cross-silo FL remains limited in practice. In this paper, we conduct an interview study to understand the practical challenges associated with cross-silo FL adoption. With interviews spanning a diverse set of stakeholders such as user organizations, software providers, and academic researchers, we uncover various barriers, from concerns about model performance to questions of incentives and trust between participating organizations. Our study shows that cross-silo FL faces a set of challenges that have yet to be well-captured by existing research in the area and are quite distinct from other forms of federated learning such as cross-device FL. We end with a discussion on future research directions that can help overcome these challenges.
comment: Main text: 23 pages, 2 tables, 2 figures
☆ LayerSync: Self-aligning Intermediate Layers
We propose LayerSync, a domain-agnostic approach for improving the generation quality and the training efficiency of diffusion models. Prior studies have highlighted the connection between the quality of generation and the representations learned by diffusion models, showing that external guidance on model intermediate representations accelerates training. We reconceptualize this paradigm by regularizing diffusion models with their own intermediate representations. Building on the observation that representation quality varies across diffusion model layers, we show that the most semantically rich representations can act as an intrinsic guidance for weaker ones, reducing the need for external supervision. Our approach, LayerSync, is a self-sufficient, plug-and-play regularizer term with no overhead on diffusion model training and generalizes beyond the visual domain to other modalities. LayerSync requires no pretrained models nor additional data. We extensively evaluate the method on image generation and demonstrate its applicability to other domains such as audio, video, and motion generation. We show that it consistently improves the generation quality and the training efficiency. For example, we speed up the training of flow-based transformer by over 8.75x on ImageNet dataset and improved the generation quality by 23.6%. The code is available at https://github.com/vita-epfl/LayerSync.
☆ CoIRL-AD: Collaborative-Competitive Imitation-Reinforcement Learning in Latent World Models for Autonomous Driving
End-to-end autonomous driving models trained solely with imitation learning (IL) often suffer from poor generalization. In contrast, reinforcement learning (RL) promotes exploration through reward maximization but faces challenges such as sample inefficiency and unstable convergence. A natural solution is to combine IL and RL. Moving beyond the conventional two-stage paradigm (IL pretraining followed by RL fine-tuning), we propose CoIRL-AD, a competitive dual-policy framework that enables IL and RL agents to interact during training. CoIRL-AD introduces a competition-based mechanism that facilitates knowledge exchange while preventing gradient conflicts. Experiments on the nuScenes dataset show an 18% reduction in collision rate compared to baselines, along with stronger generalization and improved performance on long-tail scenarios. Code is available at: https://github.com/SEU-zxj/CoIRL-AD.
comment: 18 pages, 17 figures
☆ Universal Adaptive Environment Discovery
An open problem in Machine Learning is how to avoid models to exploit spurious correlations in the data; a famous example is the background-label shortcut in the Waterbirds dataset. A common remedy is to train a model across multiple environments; in the Waterbirds dataset, this corresponds to training by randomizing the background. However, selecting the right environments is a challenging problem, given that these are rarely known a priori. We propose Universal Adaptive Environment Discovery (UAED), a unified framework that learns a distribution over data transformations that instantiate environments, and optimizes any robust objective averaged over this learned distribution. UAED yields adaptive variants of IRM, REx, GroupDRO, and CORAL without predefined groups or manual environment design. We provide a theoretical analysis by providing PAC-Bayes bounds and by showing robustness to test environment distributions under standard conditions. Empirically, UAED discovers interpretable environment distributions and improves worst-case accuracy on standard benchmarks, while remaining competitive on mean accuracy. Our results indicate that making environments adaptive is a practical route to out-of-distribution generalization.
comment: 8 papes in the main body, 4 pages in the appendix, 4 figures and 9 tables overall, conference
☆ Evaluation of Real-Time Preprocessing Methods in AI-Based ECG Signal Analysis
The increasing popularity of portable ECG systems and the growing demand for privacy-compliant, energy-efficient real-time analysis require new approaches to signal processing at the point of data acquisition. In this context, the edge domain is acquiring increasing importance, as it not only reduces latency times, but also enables an increased level of data security. The FACE project aims to develop an innovative machine learning solution for analysing long-term electrocardiograms that synergistically combines the strengths of edge and cloud computing. In this thesis, various pre-processing steps of ECG signals are analysed with regard to their applicability in the project. The selection of suitable methods in the edge area is based in particular on criteria such as energy efficiency, processing capability and real-time capability.
comment: Conference paper for 2025 IEEE World AI IoT Congress (AIIoT), FACE Project, University of Siegen, Germany
☆ Multi-Armed Bandits with Minimum Aggregated Revenue Constraints
We examine a multi-armed bandit problem with contextual information, where the objective is to ensure that each arm receives a minimum aggregated reward across contexts while simultaneously maximizing the total cumulative reward. This framework captures a broad class of real-world applications where fair revenue allocation is critical and contextual variation is inherent. The cross-context aggregation of minimum reward constraints, while enabling better performance and easier feasibility, introduces significant technical challenges -- particularly the absence of closed-form optimal allocations typically available in standard MAB settings. We design and analyze algorithms that either optimistically prioritize performance or pessimistically enforce constraint satisfaction. For each algorithm, we derive problem-dependent upper bounds on both regret and constraint violations. Furthermore, we establish a lower bound demonstrating that the dependence on the time horizon in our results is optimal in general and revealing fundamental limitations of the free exploration principle leveraged in prior work.
☆ Why the noise model matters: A performance gap in learned regularization
This article addresses the challenge of learning effective regularizers for linear inverse problems. We analyze and compare several types of learned variational regularization against the theoretical benchmark of the optimal affine reconstruction, i.e. the best possible affine linear map for minimizing the mean squared error. It is known that this optimal reconstruction can be achieved using Tikhonov regularization, but this requires precise knowledge of the noise covariance to properly weight the data fidelity term. However, in many practical applications, noise statistics are unknown. We therefore investigate the performance of regularization methods learned without access to this noise information, focusing on Tikhonov, Lavrentiev, and quadratic regularization. Our theoretical analysis and numerical experiments demonstrate that for non-white noise, a performance gap emerges between these methods and the optimal affine reconstruction. Furthermore, we show that these different types of regularization yield distinct results, highlighting that the choice of regularizer structure is critical when the noise model is not explicitly learned. Our findings underscore the significant value of accurately modeling or co-learning noise statistics in data-driven regularization.
☆ The Robustness of Differentiable Causal Discovery in Misspecified Scenarios ICLR 2025
Causal discovery aims to learn causal relationships between variables from targeted data, making it a fundamental task in machine learning. However, causal discovery algorithms often rely on unverifiable causal assumptions, which are usually difficult to satisfy in real-world data, thereby limiting the broad application of causal discovery in practical scenarios. Inspired by these considerations, this work extensively benchmarks the empirical performance of various mainstream causal discovery algorithms, which assume i.i.d. data, under eight model assumption violations. Our experimental results show that differentiable causal discovery methods exhibit robustness under the metrics of Structural Hamming Distance and Structural Intervention Distance of the inferred graphs in commonly used challenging scenarios, except for scale variation. We also provide the theoretical explanations for the performance of differentiable causal discovery methods. Finally, our work aims to comprehensively benchmark the performance of recent differentiable causal discovery methods under model assumption violations, and provide the standard for reasonable evaluation of causal discovery, as well as to further promote its application in real-world scenarios.
comment: accepted to ICLR 2025
☆ Mitigating the Noise Shift for Denoising Generative Models via Noise Awareness Guidance
Existing denoising generative models rely on solving discretized reverse-time SDEs or ODEs. In this paper, we identify a long-overlooked yet pervasive issue in this family of models: a misalignment between the pre-defined noise level and the actual noise level encoded in intermediate states during sampling. We refer to this misalignment as noise shift. Through empirical analysis, we demonstrate that noise shift is widespread in modern diffusion models and exhibits a systematic bias, leading to sub-optimal generation due to both out-of-distribution generalization and inaccurate denoising updates. To address this problem, we propose Noise Awareness Guidance (NAG), a simple yet effective correction method that explicitly steers sampling trajectories to remain consistent with the pre-defined noise schedule. We further introduce a classifier-free variant of NAG, which jointly trains a noise-conditional and a noise-unconditional model via noise-condition dropout, thereby eliminating the need for external classifiers. Extensive experiments, including ImageNet generation and various supervised fine-tuning tasks, show that NAG consistently mitigates noise shift and substantially improves the generation quality of mainstream diffusion models.
☆ PubSub-VFL: Towards Efficient Two-Party Split Learning in Heterogeneous Environments via Publisher/Subscriber Architecture NeurIPS 2025
With the rapid advancement of the digital economy, data collaboration between organizations has become a well-established business model, driving the growth of various industries. However, privacy concerns make direct data sharing impractical. To address this, Two-Party Split Learning (a.k.a. Vertical Federated Learning (VFL)) has emerged as a promising solution for secure collaborative learning. Despite its advantages, this architecture still suffers from low computational resource utilization and training efficiency. Specifically, its synchronous dependency design increases training latency, while resource and data heterogeneity among participants further hinder efficient computation. To overcome these challenges, we propose PubSub-VFL, a novel VFL paradigm with a Publisher/Subscriber architecture optimized for two-party collaborative learning with high computational efficiency. PubSub-VFL leverages the decoupling capabilities of the Pub/Sub architecture and the data parallelism of the parameter server architecture to design a hierarchical asynchronous mechanism, reducing training latency and improving system efficiency. Additionally, to mitigate the training imbalance caused by resource and data heterogeneity, we formalize an optimization problem based on participants' system profiles, enabling the selection of optimal hyperparameters while preserving privacy. We conduct a theoretical analysis to demonstrate that PubSub-VFL achieves stable convergence and is compatible with security protocols such as differential privacy. Extensive case studies on five benchmark datasets further validate its effectiveness, showing that, compared to state-of-the-art baselines, PubSub-VFL not only accelerates training by $2 \sim 7\times$ without compromising accuracy, but also achieves a computational resource utilization rate of up to 91.07%.
comment: Accepted at NeurIPS 2025
☆ CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling
Time series anomaly detection plays a crucial role in a wide range of real-world applications. Given that time series data can exhibit different patterns at different sampling granularities, multi-scale modeling has proven beneficial for uncovering latent anomaly patterns that may not be apparent at a single scale. However, existing methods often model multi-scale information independently or rely on simple feature fusion strategies, neglecting the dynamic changes in cross-scale associations that occur during anomalies. Moreover, most approaches perform multi-scale modeling based on fixed sliding windows, which limits their ability to capture comprehensive contextual information. In this work, we propose CrossAD, a novel framework for time series Anomaly Detection that takes Cross-scale associations and Cross-window modeling into account. We propose a cross-scale reconstruction that reconstructs fine-grained series from coarser series, explicitly capturing cross-scale associations. Furthermore, we design a query library and incorporate global multi-scale context to overcome the limitations imposed by fixed window sizes. Extensive experiments conducted on multiple real-world datasets using nine evaluation metrics validate the effectiveness of CrossAD, demonstrating state-of-the-art performance in anomaly detection.
comment: Accepted by the thirty-ninth annual conference on Neural Information Processing Systems
☆ Diff-XYZ: A Benchmark for Evaluating Diff Understanding
Reliable handling of code diffs is central to agents that edit and refactor repositories at scale. We introduce Diff-XYZ, a compact benchmark for code-diff understanding with three supervised tasks: apply (old code $+$ diff $\rightarrow$ new code), anti-apply (new code $-$ diff $\rightarrow$ old code), and diff generation (new code $-$ old code $\rightarrow$ diff). Instances in the benchmark are triples $\langle \textit{old code}, \textit{new code}, \textit{diff} \rangle$ drawn from real commits in CommitPackFT, paired with automatic metrics and a clear evaluation protocol. We use the benchmark to do a focused empirical study of the unified diff format and run a cross-format comparison of different diff representations. Our findings reveal that different formats should be used depending on the use case and model size. For example, representing diffs in search-replace format is good for larger models in the diff generation scenario, yet not suited well for diff analysis and smaller models. The Diff-XYZ benchmark is a reusable foundation for assessing and improving diff handling in LLMs that can aid future development of diff formats and models editing code. The dataset is published on HuggingFace Hub: https://huggingface.co/datasets/JetBrains-Research/diff-xyz.
☆ SMEC: Rethinking Matryoshka Representation Learning for Retrieval Embedding Compression EMNLP2025
Large language models (LLMs) generate high-dimensional embeddings that capture rich semantic and syntactic information. However, high-dimensional embeddings exacerbate computational complexity and storage requirements, thereby hindering practical deployment. To address these challenges, we propose a novel training framework named Sequential Matryoshka Embedding Compression (SMEC). This framework introduces the Sequential Matryoshka Representation Learning(SMRL) method to mitigate gradient variance during training, the Adaptive Dimension Selection (ADS) module to reduce information degradation during dimension pruning, and the Selectable Cross-batch Memory (S-XBM) module to enhance unsupervised learning between high- and low-dimensional embeddings. Experiments on image, text, and multimodal datasets demonstrate that SMEC achieves significant dimensionality reduction while maintaining performance. For instance, on the BEIR dataset, our approach improves the performance of compressed LLM2Vec embeddings (256 dimensions) by 1.1 points and 2.7 points compared to the Matryoshka-Adaptor and Search-Adaptor models, respectively.
comment: Accepted by EMNLP2025
☆ Time-Correlated Video Bridge Matching
Diffusion models excel in noise-to-data generation tasks, providing a mapping from a Gaussian distribution to a more complex data distribution. However they struggle to model translations between complex distributions, limiting their effectiveness in data-to-data tasks. While Bridge Matching (BM) models address this by finding the translation between data distributions, their application to time-correlated data sequences remains unexplored. This is a critical limitation for video generation and manipulation tasks, where maintaining temporal coherence is particularly important. To address this gap, we propose Time-Correlated Video Bridge Matching (TCVBM), a framework that extends BM to time-correlated data sequences in the video domain. TCVBM explicitly models inter-sequence dependencies within the diffusion bridge, directly incorporating temporal correlations into the sampling process. We compare our approach to classical methods based on bridge matching and diffusion models for three video-related tasks: frame interpolation, image-to-video generation, and video super-resolution. TCVBM achieves superior performance across multiple quantitative metrics, demonstrating enhanced generation quality and reconstruction fidelity.
☆ A Function Centric Perspective On Flat and Sharp Minima
Flat minima are widely believed to correlate with improved generalisation in deep neural networks. However, this connection has proven more nuanced in recent studies, with both theoretical counterexamples and empirical exceptions emerging in the literature. In this paper, we revisit the role of sharpness in model performance, proposing that sharpness is better understood as a function-dependent property rather than a reliable indicator of poor generalisation. We conduct extensive empirical studies, from single-objective optimisation to modern image classification tasks, showing that sharper minima often emerge when models are regularised (e.g., via SAM, weight decay, or data augmentation), and that these sharp minima can coincide with better generalisation, calibration, robustness, and functional consistency. Across a range of models and datasets, we find that baselines without regularisation tend to converge to flatter minima yet often perform worse across all safety metrics. Our findings demonstrate that function complexity, rather than flatness alone, governs the geometry of solutions, and that sharper minima can reflect more appropriate inductive biases (especially under regularisation), calling for a function-centric reappraisal of loss landscape geometry.
comment: 26 pages, 26 tables, 63 figures, pre-print
☆ Bayesian Optimization for Dynamic Pricing and Learning
Dynamic pricing is the practice of adjusting the selling price of a product to maximize a firm's revenue by responding to market demand. The literature typically distinguishes between two settings: infinite inventory, where the firm has unlimited stock and time to sell, and finite inventory, where both inventory and selling horizon are limited. In both cases, the central challenge lies in the fact that the demand function -- how sales respond to price -- is unknown and must be learned from data. Traditional approaches often assume a specific parametric form for the demand function, enabling the use of reinforcement learning (RL) to identify near-optimal pricing strategies. However, such assumptions may not hold in real-world scenarios, limiting the applicability of these methods. In this work, we propose a Gaussian Process (GP) based nonparametric approach to dynamic pricing that avoids restrictive modeling assumptions. We treat the demand function as a black-box function of the price and develop pricing algorithms based on Bayesian Optimization (BO) -- a sample-efficient method for optimizing unknown functions. We present BO-based algorithms tailored for both infinite and finite inventory settings and provide regret guarantees for both regimes, thereby quantifying the learning efficiency of our methods. Through extensive experiments, we demonstrate that our BO-based methods outperform several state-of-the-art RL algorithms in terms of revenue, while requiring fewer assumptions and offering greater robustness. This highlights Bayesian Optimization as a powerful and practical tool for dynamic pricing in complex, uncertain environments.
☆ Formal Models and Convergence Analysis for Context-Aware Security Verification
We present a formal framework for context-aware security verification that establishes provable guarantees for ML-enhanced adaptive systems. We introduce context-completeness - a new security property - and prove: (1) sample complexity bounds showing when adaptive verification succeeds, (2) information-theoretic limits relating context richness to detection capability, (3) convergence guarantees for ML-based payload generators, and (4) compositional soundness bounds. We further provide a formal separation between static context-blind verifiers and context-aware adaptive verifiers: for a natural family of targets, any static verifier with finite payload budget achieves completeness at most alpha, while a context-aware verifier with sufficient information achieves completeness greater than alpha. We validate our theoretical predictions through controlled experiments on 97,224 exploit samples, demonstrating: detection accuracy improving from 58% to 69.93% with dataset growth, success probability increasing from 51% to 82% with context enrichment, training loss converging at O(1/sqrt(T)) rate, and false positive rate (10.19%) within theoretical bounds (12%). Our results show that theoretically-grounded adaptive verification achieves provable improvements over static approaches under stated assumptions while maintaining soundness guarantees.
comment: 11 pages, 4 figures, 4 tables. Presents formal framework for context-aware security verification with ML-enhanced adaptive systems. Includes theoretical bounds (sample complexity, information-theoretic limits, convergence guarantees, soundness preservation) and empirical validation on 97,224 exploit samples
☆ Neural Guided Sampling for Quantum Circuit Optimization
Translating a general quantum circuit on a specific hardware topology with a reduced set of available gates, also known as transpilation, comes with a substantial increase in the length of the equivalent circuit. Due to decoherence, the quality of the computational outcome can degrade seriously with increasing circuit length. Thus, there is major interest to reduce a quantum circuit to an equivalent circuit which is in its gate count as short as possible. One method to address efficient transpilation is based on approaches known from stochastic optimization, e.g. by using random sampling and token replacement strategies. Here, a core challenge is that these methods can suffer from sampling efficiency, causing long and energy consuming optimization time. As a remedy, we propose in this work 2D neural guided sampling. Thus, given a 2D representation of a quantum circuit, a neural network predicts groups of gates in the quantum circuit, which are likely reducible. Thus, it leads to a sampling prior which can heavily reduce the compute time for quantum circuit reduction. In several experiments, we demonstrate that our method is superior to results obtained from different qiskit or BQSKit optimization levels.
comment: 12 pages, 9 Figures
☆ Geopolitics, Geoeconomics and Risk:A Machine Learning Approach
We introduce a novel high-frequency daily panel dataset of both markets and news-based indicators -- including Geopolitical Risk, Economic Policy Uncertainty, Trade Policy Uncertainty, and Political Sentiment -- for 42 countries across both emerging and developed markets. Using this dataset, we study how sentiment dynamics shape sovereign risk, measured by Credit Default Swap (CDS) spreads, and evaluate their forecasting value relative to traditional drivers such as global monetary policy and market volatility. Our horse-race analysis of forecasting models demonstrates that incorporating news-based indicators significantly enhances predictive accuracy and enriches the analysis, with non-linear machine learning methods -- particularly Random Forests -- delivering the largest gains. Our analysis reveals that while global financial variables remain the dominant drivers of sovereign risk, geopolitical risk and economic policy uncertainty also play a meaningful role. Crucially, their effects are amplified through non-linear interactions with global financial conditions. Finally, we document pronounced regional heterogeneity, as certain asset classes and emerging markets exhibit heightened sensitivity to shocks in policy rates, global financial volatility, and geopolitical risk.
☆ Continuous Uniqueness and Novelty Metrics for Generative Modeling of Inorganic Crystals NeurIPS 2025
To address pressing scientific challenges such as climate change, increasingly sophisticated generative artificial intelligence models are being developed that can efficiently sample the large chemical space of possible functional materials. These models can quickly sample new chemical compositions paired with crystal structures. They are typically evaluated using uniqueness and novelty metrics, which depend on a chosen crystal distance function. However, the most prevalent distance function has four limitations: it fails to quantify the degree of similarity between compounds, cannot distinguish compositional difference and structural difference, lacks Lipschitz continuity against shifts in atomic coordinates, and results in a uniqueness metric that is not invariant against the permutation of generated samples. In this work, we propose using two continuous distance functions to evaluate uniqueness and novelty, which theoretically overcome these limitations. Our experiments show that these distances reveal insights missed by traditional distance functions, providing a more reliable basis for evaluating and comparing generative models for inorganic crystals.
comment: 13 pages (5 pages of main text), accepted to the AI4Mat workshop at NeurIPS 2025. See https://github.com/WMD-group/xtalmet for the code
☆ Robot Learning: A Tutorial
Robot learning is at an inflection point, driven by rapid advancements in machine learning and the growing availability of large-scale robotics data. This shift from classical, model-based methods to data-driven, learning-based paradigms is unlocking unprecedented capabilities in autonomous systems. This tutorial navigates the landscape of modern robot learning, charting a course from the foundational principles of Reinforcement Learning and Behavioral Cloning to generalist, language-conditioned models capable of operating across diverse tasks and even robot embodiments. This work is intended as a guide for researchers and practitioners, and our goal is to equip the reader with the conceptual understanding and practical tools necessary to contribute to developments in robot learning, with ready-to-use examples implemented in $\texttt{lerobot}$.
comment: Tutorial on Robot Learning using LeRobot, the end-to-end robot learning library developed by Hugging Face
☆ Cautious Weight Decay
We introduce Cautious Weight Decay (CWD), a one-line, optimizer-agnostic modification that applies weight decay only to parameter coordinates whose signs align with the optimizer update. Unlike standard decoupled decay, which implicitly optimizes a regularized or constrained objective, CWD preserves the original loss and admits a bilevel interpretation: it induces sliding-mode behavior upon reaching the stationary manifold, allowing it to search for locally Pareto-optimal stationary points of the unmodified objective. In practice, CWD is a drop-in change for optimizers such as AdamW, Lion, and Muon, requiring no new hyperparameters or additional tuning. For language model pre-training and ImageNet classification, CWD consistently improves final loss and accuracy at million- to billion-parameter scales.
☆ Enhanced Pre-training of Graph Neural Networks for Million-Scale Heterogeneous Graphs
In recent years, graph neural networks (GNNs) have facilitated the development of graph data mining. However, training GNNs requires sufficient labeled task-specific data, which is expensive and sometimes unavailable. To be less dependent on labeled data, recent studies propose to pre-train GNNs in a self-supervised manner and then apply the pre-trained GNNs to downstream tasks with limited labeled data. However, most existing methods are designed solely for homogeneous graphs (real-world graphs are mostly heterogeneous) and do not consider semantic mismatch (the semantic difference between the original data and the ideal data containing more transferable semantic information). In this paper, we propose an effective framework to pre-train GNNs on the large-scale heterogeneous graph. We first design a structure-aware pre-training task, which aims to capture structural properties in heterogeneous graphs. Then, we design a semantic-aware pre-training task to tackle the mismatch. Specifically, we construct a perturbation subspace composed of semantic neighbors to help deal with the semantic mismatch. Semantic neighbors make the model focus more on the general knowledge in the semantic space, which in turn assists the model in learning knowledge with better transferability. Finally, extensive experiments are conducted on real-world large-scale heterogeneous graphs to demonstrate the superiority of the proposed method over state-of-the-art baselines. Code available at https://github.com/sunshy-1/PHE.
comment: 26 pages
☆ Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking NeurIPS25
Generative Behavior Cloning (GBC) is a simple yet effective framework for robot learning, particularly in multi-task settings. Recent GBC methods often employ diffusion policies with open-loop (OL) control, where actions are generated via a diffusion process and executed in multi-step chunks without replanning. While this approach has demonstrated strong success rates and generalization, its inherent stochasticity can result in erroneous action sampling, occasionally leading to unexpected task failures. Moreover, OL control suffers from delayed responses, which can degrade performance in noisy or dynamic environments. To address these limitations, we propose two novel techniques to enhance the consistency and reactivity of diffusion policies: (1) self-guidance, which improves action fidelity by leveraging past observations and implicitly promoting future-aware behavior; and (2) adaptive chunking, which selectively updates action sequences when the benefits of reactivity outweigh the need for temporal consistency. Extensive experiments show that our approach substantially improves GBC performance across a wide range of simulated and real-world robotic manipulation tasks. Our code is available at https://github.com/junhyukso/SGAC
comment: Accepted at NeurIPS25
☆ Towards Cross-Modal Error Detection with Tables and Images
Ensuring data quality at scale remains a persistent challenge for large organizations. Despite recent advances, maintaining accurate and consistent data is still complex, especially when dealing with multiple data modalities. Traditional error detection and correction methods tend to focus on a single modality, typically a table, and often miss cross-modal errors that are common in domains like e-Commerce and healthcare, where image, tabular, and text data co-exist. To address this gap, we take an initial step towards cross-modal error detection in tabular data, by benchmarking several methods. Our evaluation spans four datasets and five baseline approaches. Among them, Cleanlab, a label error detection framework, and DataScope, a data valuation method, perform the best when paired with a strong AutoML framework, achieving the highest F1 scores. Our findings indicate that current methods remain limited, particularly when applied to heavy-tailed real-world data, motivating further research in this area.
☆ LiteVPNet: A Lightweight Network for Video Encoding Control in Quality-Critical Applications
In the last decade, video workflows in the cinema production ecosystem have presented new use cases for video streaming technology. These new workflows, e.g. in On-set Virtual Production, present the challenge of requiring precise quality control and energy efficiency. Existing approaches to transcoding often fall short of these requirements, either due to a lack of quality control or computational overhead. To fill this gap, we present a lightweight neural network (LiteVPNet) for accurately predicting Quantisation Parameters for NVENC AV1 encoders that achieve a specified VMAF score. We use low-complexity features, including bitstream characteristics, video complexity measures, and CLIP-based semantic embeddings. Our results demonstrate that LiteVPNet achieves mean VMAF errors below 1.2 points across a wide range of quality targets. Notably, LiteVPNet achieves VMAF errors within 2 points for over 87% of our test corpus, c.f. approx 61% with state-of-the-art methods. LiteVPNet's performance across various quality regions highlights its applicability for enhancing high-value content transport and streaming for more energy-efficient, high-quality media experiences.
comment: Accepted PCS 2025 Camera-Ready Version, 5 Pages
☆ Deep Attention-guided Adaptive Subsampling
Although deep neural networks have provided impressive gains in performance, these improvements often come at the cost of increased computational complexity and expense. In many cases, such as 3D volume or video classification tasks, not all slices or frames are necessary due to inherent redundancies. To address this issue, we propose a novel learnable subsampling framework that can be integrated into any neural network architecture. Subsampling, being a nondifferentiable operation, poses significant challenges for direct adaptation into deep learning models. While some works, have proposed solutions using the Gumbel-max trick to overcome the problem of non-differentiability, they fall short in a crucial aspect: they are only task-adaptive and not inputadaptive. Once the sampling mechanism is learned, it remains static and does not adjust to different inputs, making it unsuitable for real-world applications. To this end, we propose an attention-guided sampling module that adapts to inputs even during inference. This dynamic adaptation results in performance gains and reduces complexity in deep neural network models. We demonstrate the effectiveness of our method on 3D medical imaging datasets from MedMNIST3D as well as two ultrasound video datasets for classification tasks, one of them being a challenging in-house dataset collected under real-world clinical conditions.
☆ Improved Central Limit Theorem and Bootstrap Approximations for Linear Stochastic Approximation
In this paper, we refine the Berry-Esseen bounds for the multivariate normal approximation of Polyak-Ruppert averaged iterates arising from the linear stochastic approximation (LSA) algorithm with decreasing step size. We consider the normal approximation by the Gaussian distribution with covariance matrix predicted by the Polyak-Juditsky central limit theorem and establish the rate up to order $n^{-1/3}$ in convex distance, where $n$ is the number of samples used in the algorithm. We also prove a non-asymptotic validity of the multiplier bootstrap procedure for approximating the distribution of the rescaled error of the averaged LSA estimator. We establish approximation rates of order up to $1/\sqrt{n}$ for the latter distribution, which significantly improves upon the previous results obtained by Samsonov et al. (2024).
☆ Constrained Sensing and Reliable State Estimation with Shallow Recurrent Decoders on a TRIGA Mark II Reactor
Shallow Recurrent Decoder networks are a novel data-driven methodology able to provide accurate state estimation in engineering systems, such as nuclear reactors. This deep learning architecture is a robust technique designed to map the temporal trajectories of a few sparse measures to the full state space, including unobservable fields, which is agnostic to sensor positions and able to handle noisy data through an ensemble strategy, leveraging the short training times and without the need for hyperparameter tuning. Following its application to a novel reactor concept, this work investigates the performance of Shallow Recurrent Decoders when applied to a real system. The underlying model is represented by a fluid dynamics model of the TRIGA Mark II research reactor; the architecture will use both synthetic temperature data coming from the numerical model and leveraging experimental temperature data recorded during a previous campaign. The objective of this work is, therefore, two-fold: 1) assessing if the architecture can reconstruct the full state of the system (temperature, velocity, pressure, turbulence quantities) given sparse data located in specific, low-dynamics channels and 2) assessing the correction capabilities of the architecture (that is, given a discrepancy between model and data, assessing if sparse measurements can provide some correction to the architecture output). As will be shown, the accurate reconstruction of every characteristic field, using both synthetic and experimental data, in real-time makes this approach suitable for interpretable monitoring and control purposes in the framework of a reactor digital twin.
☆ Pretraining in Actor-Critic Reinforcement Learning for Robot Motion Control ICLR 2026
The pretraining-finetuning paradigm has facilitated numerous transformative advancements in artificial intelligence research in recent years. However, in the domain of reinforcement learning (RL) for robot motion control, individual skills are often learned from scratch despite the high likelihood that some generalizable knowledge is shared across all task-specific policies belonging to a single robot embodiment. This work aims to define a paradigm for pretraining neural network models that encapsulate such knowledge and can subsequently serve as a basis for warm-starting the RL process in classic actor-critic algorithms, such as Proximal Policy Optimization (PPO). We begin with a task-agnostic exploration-based data collection algorithm to gather diverse, dynamic transition data, which is then used to train a Proprioceptive Inverse Dynamics Model (PIDM) through supervised learning. The pretrained weights are loaded into both the actor and critic networks to warm-start the policy optimization of actual tasks. We systematically validated our proposed method on seven distinct robot motion control tasks, showing significant benefits to this initialization strategy. Our proposed approach on average improves sample efficiency by 40.1% and task performance by 7.5%, compared to random initialization. We further present key ablation studies and empirical analyses that shed light on the mechanisms behind the effectiveness of our method.
comment: Submitted to ICLR 2026
☆ Traveling Salesman-Based Token Ordering Improves Stability in Homomorphically Encrypted Language Models
As users increasingly interact with large language models (LLMs) using private information, secure and encrypted communication becomes essential. Homomorphic encryption (HE) provides a principled solution by enabling computation directly on encrypted data. Although prior work has explored aspects of running LLMs under HE, the challenge of text generation, particularly next-token prediction, has received limited attention and remains a key obstacle to practical encrypted interaction. In this work, we propose a TSP-based token reordering strategy to address the difficulties of encrypted text generation, together with a post-processing step that further reduces approximation error. Theoretical analysis and experimental results demonstrate that our method prevents collapse, improves coherence in generated text, and preserves data privacy throughout. Overall, our contributions advance the feasibility of practical and privacy-preserving LLM inference.
comment: 34 pages
☆ Finite-time Convergence Analysis of Actor-Critic with Evolving Reward
Many popular practical reinforcement learning (RL) algorithms employ evolving reward functions-through techniques such as reward shaping, entropy regularization, or curriculum learning-yet their theoretical foundations remain underdeveloped. This paper provides the first finite-time convergence analysis of a single-timescale actor-critic algorithm in the presence of an evolving reward function under Markovian sampling. We consider a setting where the reward parameters may change at each time step, affecting both policy optimization and value estimation. Under standard assumptions, we derive non-asymptotic bounds for both actor and critic errors. Our result shows that an $O(1/\sqrt{T})$ convergence rate is achievable, matching the best-known rate for static rewards, provided the reward parameters evolve slowly enough. This rate is preserved when the reward is updated via a gradient-based rule with bounded gradient and on the same timescale as the actor and critic, offering a theoretical foundation for many popular RL techniques. As a secondary contribution, we introduce a novel analysis of distribution mismatch under Markovian sampling, improving the best-known rate by a factor of $\log^2T$ in the static-reward case.
☆ Leveraging Teleconnections with Physics-Informed Graph Attention Networks for Long-Range Extreme Rainfall Forecasting in Thailand
Accurate rainfall forecasting, particularly for extreme events, remains a significant challenge in climatology and the Earth system. This paper presents novel physics-informed Graph Neural Networks (GNNs) combined with extreme-value analysis techniques to improve gauge-station rainfall predictions across Thailand. The model leverages a graph-structured representation of gauge stations to capture complex spatiotemporal patterns, and it offers explainability through teleconnections. We preprocess relevant climate indices that potentially influence regional rainfall. The proposed Graph Attention Network with Long Short-Term Memory (Attention-LSTM) applies the attention mechanism using initial edge features derived from simple orographic-precipitation physics formulation. The embeddings are subsequently processed by LSTM layers. To address extremes, we perform Peak-Over-Threshold (POT) mapping using the novel Spatial Season-aware Generalized Pareto Distribution (GPD) method, which overcomes limitations of traditional machine-learning models. Experiments demonstrate that our method outperforms well-established baselines across most regions, including areas prone to extremes, and remains strongly competitive with the state of the art. Compared with the operational forecasting system SEAS5, our real-world application improves extreme-event prediction and offers a practical enhancement to produce fine-resolution maps that support decision-making in long-term water management.
☆ Deep SPI: Safe Policy Improvement via World Models
Safe policy improvement (SPI) offers theoretical control over policy updates, yet existing guarantees largely concern offline, tabular reinforcement learning (RL). We study SPI in general online settings, when combined with world model and representation learning. We develop a theoretical framework showing that restricting policy updates to a well-defined neighborhood of the current policy ensures monotonic improvement and convergence. This analysis links transition and reward prediction losses to representation quality, yielding online, "deep" analogues of classical SPI theorems from the offline RL literature. Building on these results, we introduce DeepSPI, a principled on-policy algorithm that couples local transition and reward losses with regularised policy updates. On the ALE-57 benchmark, DeepSPI matches or exceeds strong baselines, including PPO and DeepMDPs, while retaining theoretical guarantees.
comment: 10 pages main text, 17 pages appendix (excluding references)
☆ Learning Latent Energy-Based Models via Interacting Particle Langevin Dynamics
We develop interacting particle algorithms for learning latent variable models with energy-based priors. To do so, we leverage recent developments in particle-based methods for solving maximum marginal likelihood estimation (MMLE) problems. Specifically, we provide a continuous-time framework for learning latent energy-based models, by defining stochastic differential equations (SDEs) that provably solve the MMLE problem. We obtain a practical algorithm as a discretisation of these SDEs and provide theoretical guarantees for the convergence of the proposed algorithm. Finally, we demonstrate the empirical effectiveness of our method on synthetic and image datasets.
☆ DeepTrust: Multi-Step Classification through Dissimilar Adversarial Representations for Robust Android Malware Detection
Over the last decade, machine learning has been extensively applied to identify malicious Android applications. However, such approaches remain vulnerable against adversarial examples, i.e., examples that are subtly manipulated to fool a machine learning model into making incorrect predictions. This research presents DeepTrust, a novel metaheuristic that arranges flexible classifiers, like deep neural networks, into an ordered sequence where the final decision is made by a single internal model based on conditions activated in cascade. In the Robust Android Malware Detection competition at the 2025 IEEE Conference SaTML, DeepTrust secured the first place and achieved state-of-the-art results, outperforming the next-best competitor by up to 266% under feature-space evasion attacks. This is accomplished while maintaining the highest detection rate on non-adversarial malware and a false positive rate below 1%. The method's efficacy stems from maximizing the divergence of the learned representations among the internal models. By using classifiers inducing fundamentally dissimilar embeddings of the data, the decision space becomes unpredictable for an attacker. This frustrates the iterative perturbation process inherent to evasion attacks, enhancing system robustness without compromising accuracy on clean examples.
☆ General Fourier Feature Physics-Informed Extreme Learning Machine (GFF-PIELM) for High-Frequency PDEs
Conventional physics-informed extreme learning machine (PIELM) often faces challenges in solving partial differential equations (PDEs) involving high-frequency and variable-frequency behaviors. To address these challenges, we propose a general Fourier feature physics-informed extreme learning machine (GFF-PIELM). We demonstrate that directly concatenating multiple Fourier feature mappings (FFMs) and an extreme learning machine (ELM) network makes it difficult to determine frequency-related hyperparameters. Fortunately, we find an alternative to establish the GFF-PIELM in three main steps. First, we integrate a variation of FFM into ELM as the Fourier-based activation function, so there is still one hidden layer in the GFF-PIELM framework. Second, we assign a set of frequency coefficients to the hidden neurons, which enables ELM network to capture diverse frequency components of target solutions. Finally, we develop an innovative, straightforward initialization method for these hyperparameters by monitoring the distribution of ELM output weights. GFF-PIELM not only retains the high accuracy, efficiency, and simplicity of the PIELM framework but also inherits the ability of FFMs to effectively handle high-frequency problems. We carry out five case studies with a total of ten numerical examples to highlight the feasibility and validity of the proposed GFF-PIELM, involving high frequency, variable frequency, multi-scale behaviour, irregular boundary and inverse problems. Compared to conventional PIELM, the GFF-PIELM approach significantly improves predictive accuracy without additional cost in training time and architecture complexity. Our results confirm that that PIELM can be extended to solve high-frequency and variable-frequency PDEs with high accuracy, and our initialization strategy may further inspire advances in other physics-informed machine learning (PIML) frameworks.
☆ Multi-Action Self-Improvement for Neural Combinatorial Optimization
Self-improvement has emerged as a state-of-the-art paradigm in Neural Combinatorial Optimization (NCO), where models iteratively refine their policies by generating and imitating high-quality solutions. Despite strong empirical performance, existing methods face key limitations. Training is computationally expensive, as policy updates require sampling numerous candidate solutions per instance to extract a single expert trajectory. More fundamentally, these approaches fail to exploit the structure of combinatorial problems involving the coordination of multiple agents, such as vehicles in min-max routing or machines in scheduling. By supervising on single-action trajectories, they fail to exploit agent-permutation symmetries, where distinct sequences of actions yield identical solutions, hindering generalization and the ability to learn coordinated behavior. We address these challenges by extending self-improvement to operate over joint multi-agent actions. Our model architecture predicts complete agent-task assignments jointly at each decision step. To explicitly leverage symmetries, we employ a set-prediction loss, which supervises the policy on multiple expert assignments for any given state. This approach enhances sample efficiency and the model's ability to learn coordinated behavior. Furthermore, by generating multi-agent actions in parallel, it drastically accelerates the solution generation phase of the self-improvement loop. Empirically, we validate our method on several combinatorial problems, demonstrating consistent improvements in the quality of the final solution and a reduced generation latency compared to standard self-improvement.
☆ Heterogeneous RBCs via deep multi-agent reinforcement learning
Current macroeconomic models with agent heterogeneity can be broadly divided into two main groups. Heterogeneous-agent general equilibrium (GE) models, such as those based on Heterogeneous Agents New Keynesian (HANK) or Krusell-Smith (KS) approaches, rely on GE and 'rational expectations', somewhat unrealistic assumptions that make the models very computationally cumbersome, which in turn limits the amount of heterogeneity that can be modelled. In contrast, agent-based models (ABMs) can flexibly encompass a large number of arbitrarily heterogeneous agents, but typically require the specification of explicit behavioural rules, which can lead to a lengthy trial-and-error model-development process. To address these limitations, we introduce MARL-BC, a framework that integrates deep multi-agent reinforcement learning (MARL) with Real Business Cycle (RBC) models. We demonstrate that MARL-BC can: (1) recover textbook RBC results when using a single agent; (2) recover the results of the mean-field KS model using a large number of identical agents; and (3) effectively simulate rich heterogeneity among agents, a hard task for traditional GE approaches. Our framework can be thought of as an ABM if used with a variety of heterogeneous interacting agents, and can reproduce GE results in limit cases. As such, it is a step towards a synthesis of these often opposed modelling paradigms.
comment: 13 pages, 9 figures
☆ The Living Forecast: Evolving Day-Ahead Predictions into Intraday Reality
Accurate intraday forecasts are essential for power system operations, complementing day-ahead forecasts that gradually lose relevance as new information becomes available. This paper introduces a Bayesian updating mechanism that converts fully probabilistic day-ahead forecasts into intraday forecasts without retraining or re-inference. The approach conditions the Gaussian mixture output of a conditional variational autoencoder-based forecaster on observed measurements, yielding an updated distribution for the remaining horizon that preserves its probabilistic structure. This enables consistent point, quantile, and ensemble forecasts while remaining computationally efficient and suitable for real-time applications. Experiments on household electricity consumption and photovoltaic generation datasets demonstrate that the proposed method improves forecast accuracy up to 25% across likelihood-, sample-, quantile-, and point-based metrics. The largest gains occur in time steps with strong temporal correlation to observed data, and the use of pattern dictionary-based covariance structures further enhances performance. The results highlight a theoretically grounded framework for intraday forecasting in modern power systems.
☆ Tensor Logic: The Language of AI
Progress in AI is hindered by the lack of a programming language with all the requisite features. Libraries like PyTorch and TensorFlow provide automatic differentiation and efficient GPU implementation, but are additions to Python, which was never intended for AI. Their lack of support for automated reasoning and knowledge acquisition has led to a long and costly series of hacky attempts to tack them on. On the other hand, AI languages like LISP an Prolog lack scalability and support for learning. This paper proposes tensor logic, a language that solves these problems by unifying neural and symbolic AI at a fundamental level. The sole construct in tensor logic is the tensor equation, based on the observation that logical rules and Einstein summation are essentially the same operation, and all else can be reduced to them. I show how to elegantly implement key forms of neural, symbolic and statistical AI in tensor logic, including transformers, formal reasoning, kernel machines and graphical models. Most importantly, tensor logic makes new directions possible, such as sound reasoning in embedding space. This combines the scalability and learnability of neural networks with the reliability and transparency of symbolic reasoning, and is potentially a basis for the wider adoption of AI.
comment: 17 pages, 0 figures
☆ HiLoRA: Adaptive Hierarchical LoRA Routing for Training-Free Domain Generalization
Low-Rank Adaptation (LoRA) has emerged as a widely used technique for adapting large language models (LLMs) to new domains, due to its modular design and broad availability on platforms such as HuggingFace. This availability has motivated efforts to reuse existing LoRAs for domain generalization. However, existing methods often rely on explicit task labels or additional training, which are impractical for deployment. Moreover, they typically activate a fixed number of entire LoRA modules, leading to parameter redundancy or insufficiency that degrade performance. In this paper, we propose \texttt{HiLoRA}, a training-free framework that performs adaptive hierarchical routing over LoRA pools. Drawing on structural properties of LoRA, we define rank-one components (ROCs), in which each rank parameter is regarded as an independent unit. For a given input sequence, \texttt{HiLoRA} first adaptively selects a subset of LoRAs and determines their ROC allocation based on Gaussian likelihoods at the sequence level. At the token level, it further refines routing by activating only the most informative ROCs. We further provide theoretical guarantees that \texttt{HiLoRA} selects the most relevant LoRAs with high probability. Extensive experiments show that \texttt{HiLoRA} achieves substantial improvements in domain generalization, with accuracy gains of up to {\small $55\%$} over state-of-the-art baselines, while maintaining comparable inference throughput.
☆ AngularFuse: A Closer Look at Angle-based Perception for Spatial-Sensitive Multi-Modality Image Fusion
Visible-infrared image fusion is crucial in key applications such as autonomous driving and nighttime surveillance. Its main goal is to integrate multimodal information to produce enhanced images that are better suited for downstream tasks. Although deep learning based fusion methods have made significant progress, mainstream unsupervised approaches still face serious challenges in practical applications. Existing methods mostly rely on manually designed loss functions to guide the fusion process. However, these loss functions have obvious limitations. On one hand, the reference images constructed by existing methods often lack details and have uneven brightness. On the other hand, the widely used gradient losses focus only on gradient magnitude. To address these challenges, this paper proposes an angle-based perception framework for spatial-sensitive image fusion (AngularFuse). At first, we design a cross-modal complementary mask module to force the network to learn complementary information between modalities. Then, a fine-grained reference image synthesis strategy is introduced. By combining Laplacian edge enhancement with adaptive histogram equalization, reference images with richer details and more balanced brightness are generated. Last but not least, we introduce an angle-aware loss, which for the first time constrains both gradient magnitude and direction simultaneously in the gradient domain. AngularFuse ensures that the fused images preserve both texture intensity and correct edge orientation. Comprehensive experiments on the MSRS, RoadScene, and M3FD public datasets show that AngularFuse outperforms existing mainstream methods with clear margin. Visual comparisons further confirm that our method produces sharper and more detailed results in challenging scenes, demonstrating superior fusion capability.
comment: For the first time, angle-based perception was introduced into the multi-modality image fusion task
☆ FedMMKT:Co-Enhancing a Server Text-to-Image Model and Client Task Models in Multi-Modal Federated Learning
Text-to-Image (T2I) models have demonstrated their versatility in a wide range of applications. However, adaptation of T2I models to specialized tasks is often limited by the availability of task-specific data due to privacy concerns. On the other hand, harnessing the power of rich multimodal data from modern mobile systems and IoT infrastructures presents a great opportunity. This paper introduces Federated Multi-modal Knowledge Transfer (FedMMKT), a novel framework that enables co-enhancement of a server T2I model and client task-specific models using decentralized multimodal data without compromising data privacy.
☆ Diffusion Models for Reinforcement Learning: Foundations, Taxonomy, and Development
Diffusion Models (DMs), as a leading class of generative models, offer key advantages for reinforcement learning (RL), including multi-modal expressiveness, stable training, and trajectory-level planning. This survey delivers a comprehensive and up-to-date synthesis of diffusion-based RL. We first provide an overview of RL, highlighting its challenges, and then introduce the fundamental concepts of DMs, investigating how they are integrated into RL frameworks to address key challenges in this research field. We establish a dual-axis taxonomy that organizes the field along two orthogonal dimensions: a function-oriented taxonomy that clarifies the roles DMs play within the RL pipeline, and a technique-oriented taxonomy that situates implementations across online versus offline learning regimes. We also provide a comprehensive examination of this progression from single-agent to multi-agent domains, thereby forming several frameworks for DM-RL integration and highlighting their practical utility. Furthermore, we outline several categories of successful applications of diffusion-based RL across diverse domains, discuss open research issues of current methodologies, and highlight key directions for future research to advance the field. Finally, we summarize the survey to identify promising future development directions. We are actively maintaining a GitHub repository (https://github.com/ChangfuXu/D4RL-FTD) for papers and other related resources to apply DMs for RL.
comment: Under Review
☆ Optimal Regularization for Performative Learning
In performative learning, the data distribution reacts to the deployed model - for example, because strategic users adapt their features to game it - which creates a more complex dynamic than in classical supervised learning. One should thus not only optimize the model for the current data but also take into account that the model might steer the distribution in a new direction, without knowing the exact nature of the potential shift. We explore how regularization can help cope with performative effects by studying its impact in high-dimensional ridge regression. We show that, while performative effects worsen the test risk in the population setting, they can be beneficial in the over-parameterized regime where the number of features exceeds the number of samples. We show that the optimal regularization scales with the overall strength of the performative effect, making it possible to set the regularization in anticipation of this effect. We illustrate this finding through empirical evaluations of the optimal regularization parameter on both synthetic and real-world datasets.
☆ MoRA: On-the-fly Molecule-aware Low-Rank Adaptation Framework for LLM-based Multi-Modal Molecular Assistant
Effectively integrating molecular graph structures with Large Language Models (LLMs) is a key challenge in drug discovery. Most existing multi-modal alignment methods typically process these structures by fine-tuning the LLM or adding a static adapter simultaneously. However, these approaches have two main limitations: (1) it optimizes a shared parameter space across all molecular inputs, limiting the model's ability to capture instance-specific structural features; and (2) fine-tuning the LLM for molecular tasks can lead to catastrophic forgetting, undermining its general reasoning capabilities. In this paper, instead of static task-oriented adaptation, we propose an instance-specific parameter space alignment approach for each molecule on-the-fly. To this end, we introduce Molecule-aware Low-Rank Adaptation (MoRA) that produces a unique set of low-rank adaptation weights for each input molecular graph. These weights are then dynamically injected into a frozen LLM, allowing the model to adapt its reasoning to the structure of each molecular input, while preserving the LLM's core knowledge. Extensive experiments demonstrate that on key molecular tasks, such as chemical reaction prediction and molecular captioning, MoRA's instance-specific dynamic adaptation outperforms statically adapted baselines, including a 14.1% relative improvement in reaction prediction exact match and a 22% reduction in error for quantum property prediction. The code is available at https://github.com/jk-sounds/MoRA.
♻ ☆ Context-Action Embedding Learning for Off-Policy Evaluation in Contextual Bandits
We consider off-policy evaluation (OPE) in contextual bandits with finite action space. Inverse Propensity Score (IPS) weighting is a widely used method for OPE due to its unbiased, but it suffers from significant variance when the action space is large or when some parts of the context-action space are underexplored. Recently introduced Marginalized IPS (MIPS) estimators mitigate this issue by leveraging action embeddings. However, these embeddings do not minimize the mean squared error (MSE) of the estimators and do not consider context information. To address these limitations, we introduce Context-Action Embedding Learning for MIPS, or CAEL-MIPS, which learns context-action embeddings from offline data to minimize the MSE of the MIPS estimator. Building on the theoretical analysis of bias and variance of MIPS, we present an MSE-minimizing objective for CAEL-MIPS. In the empirical studies on a synthetic dataset and a real-world dataset, we demonstrate that our estimator outperforms baselines in terms of MSE.
♻ ☆ How to Train Your Metamorphic Deep Neural Network
Neural Metamorphosis (NeuMeta) is a recent paradigm for generating neural networks of varying width and depth. Based on Implicit Neural Representation (INR), NeuMeta learns a continuous weight manifold, enabling the direct generation of compressed models, including those with configurations not seen during training. While promising, the original formulation of NeuMeta proves effective only for the final layers of the undelying model, limiting its broader applicability. In this work, we propose a training algorithm that extends the capabilities of NeuMeta to enable full-network metamorphosis with minimal accuracy degradation. Our approach follows a structured recipe comprising block-wise incremental training, INR initialization, and strategies for replacing batch normalization. The resulting metamorphic networks maintain competitive accuracy across a wide range of compression ratios, offering a scalable solution for adaptable and efficient deployment of deep models. The code is available at: https://github.com/TSommariva/HTTY_NeuMeta.
comment: Accepted with an Honorable Mention Award at ICIAP 2025 (Rome, Italy). 14 pages, 7 figures
♻ ☆ Multi-View Graph Learning with Graph-Tuple
Graph Neural Networks (GNNs) typically scale with the number of graph edges, making them well suited for sparse graphs but less efficient on dense graphs, such as point clouds or molecular interactions. A common remedy is to sparsify the graph via similarity thresholding or distance pruning, but this forces an arbitrary choice of a single interaction scale and discards crucial information from other scales. To overcome this limitation, we introduce a multi-view graph-tuple framework. Instead of a single graph, our graph-tuple framework partitions the graph into disjoint subgraphs, capturing primary local interactions and weaker, long-range connections. We then learn multi-view representations from the graph-tuple via a heterogeneous message-passing architecture inspired by the theory of non-commuting operators, which we formally prove is strictly more expressive and guarantees a lower oracle risk compared to single-graph message-passing models. We instantiate our framework on two scientific domains: molecular property prediction from feature-scarce Coulomb matrices and cosmological parameter inference from geometric point clouds. On both applications, our multi-view graph-tuple models demonstrate better performance than single-graph baselines, highlighting the power and versatility of our multi-view approach.
comment: Submitted to TAG workshop
♻ ☆ Reconstruction of SINR Maps from Sparse Measurements using Group Equivariant Non-Expansive Operators
As sixth generation (6G) wireless networks evolve, accurate signal-to-interference-noise ratio (SINR) maps are becoming increasingly critical for effective resource management and optimization. However, acquiring such maps at high resolution is often cost-prohibitive, creating a severe data scarcity challenge. This necessitates machine learning (ML) approaches capable of robustly reconstructing the full map from extremely sparse measurements. To address this, we introduce a novel reconstruction framework based on Group Equivariant Non-Expansive Operators (GENEOs). Unlike data-hungry ML models, GENEOs are low-complexity operators that embed domain-specific geometric priors, such as translation invariance, directly into their structure. This provides a strong inductive bias, enabling effective reconstruction from very few samples. Our key insight is that for network management, preserving the topological structure of the SINR map, such as the geometry of coverage holes and interference patterns, is often more critical than minimizing pixel-wise error. We validate our approach on realistic ray-tracing-based urban scenarios, evaluating performance with both traditional statistical metrics (mean squared error (MSE)) and, crucially, a topological metric (1-Wasserstein distance). Results show that while maintaining competitive MSE, our method dramatically outperforms established ML baselines in topological fidelity. This demonstrates the practical advantage of GENEOs for creating structurally accurate SINR maps that are more reliable for downstream network optimization tasks.
♻ ☆ Joint Embedding vs Reconstruction: Provable Benefits of Latent Space Prediction for Self Supervised Learning
Reconstruction and joint embedding have emerged as two leading paradigms in Self Supervised Learning (SSL). Reconstruction methods focus on recovering the original sample from a different view in input space. On the other hand, joint embedding methods align the representations of different views in latent space. Both approaches offer compelling advantages, yet practitioners lack clear guidelines for choosing between them. In this work, we unveil the core mechanisms that distinguish each paradigm. By leveraging closed form solutions for both approaches, we precisely characterize how the view generation process, e.g. data augmentation, impacts the learned representations. We then demonstrate that, unlike supervised learning, both SSL paradigms require a minimal alignment between augmentations and irrelevant features to achieve asymptotic optimality with increasing sample size. Our findings indicate that in scenarios where these irrelevant features have a large magnitude, joint embedding methods are preferable because they impose a strictly weaker alignment condition compared to reconstruction based methods. These results not only clarify the trade offs between the two paradigms but also substantiate the empirical success of joint embedding approaches on real world challenging datasets.
comment: 33 pages, 9 figures
♻ ☆ Fixed Point Explainability
This paper introduces a formal notion of fixed point explanations, inspired by the "why regress" principle, to assess, through recursive applications, the stability of the interplay between a model and its explainer. Fixed point explanations satisfy properties like minimality, stability, and faithfulness, revealing hidden model behaviours and explanatory weaknesses. We define convergence conditions for several classes of explainers, from feature-based to mechanistic tools like Sparse AutoEncoders, and we report quantitative and qualitative results for several datasets and models, including LLMs such as Llama-3.3-70B.
comment: The code is available here: https://anonymous.4open.science/r/fixed_point_explainability_iclr2026-D188
Finite Sample Analysis of Linear Temporal Difference Learning with Arbitrary Features
Linear TD($\lambda$) is one of the most fundamental reinforcement learning algorithms for policy evaluation. Previously, convergence rates are typically established under the assumption of linearly independent features, which does not hold in many practical scenarios. This paper instead establishes the first $L^2$ convergence rates for linear TD($\lambda$) operating under arbitrary features, without making any algorithmic modification or additional assumptions. Our results apply to both the discounted and average-reward settings. To address the potential non-uniqueness of solutions resulting from arbitrary features, we develop a novel stochastic approximation result featuring convergence rates to the solution set instead of a single point.
♻ ☆ Investigating Faithfulness in Large Audio Language Models
Faithfulness measures whether chain-of-thought (CoT) representations accurately reflect a model's decision process and can be used as reliable explanations. Prior work has shown that CoTs from text-based LLMs are often unfaithful. This question has not been explored for large audio-language models (LALMs), where faithfulness is critical for safety-sensitive applications. Reasoning in LALMs is also more challenging, as models must first extract relevant clues from audio before reasoning over them. In this paper, we investigate the faithfulness of CoTs produced by several LALMs by applying targeted interventions, including paraphrasing, filler token injection, early answering, and introducing mistakes, on two challenging reasoning datasets: SAKURA and MMAR. After going through the aforementioned interventions across several datasets and tasks, our experiments suggest that, LALMs generally produce CoTs that appear to be faithful to their underlying decision processes.
♻ ☆ Increase Alpha: Performance and Risk of an AI-Driven Trading Framework
There are inefficiencies in financial markets, with unexploited patterns in price, volume, and cross-sectional relationships. While many approaches use large-scale transformers, we take a domain-focused path: feed-forward and recurrent networks with curated features to capture subtle regularities in noisy financial data. This smaller-footprint design is computationally lean and reliable under low signal-to-noise, crucial for daily production at scale. At Increase Alpha, we built a deep-learning framework that maps over 800 U.S. equities into daily directional signals with minimal computational overhead. The purpose of this paper is twofold. First, we outline the general overview of the predictive model without disclosing its core underlying concepts. Second, we evaluate its real-time performance through transparent, industry standard metrics. Forecast accuracy is benchmarked against both naive baselines and macro indicators. The performance outcomes are summarized via cumulative returns, annualized Sharpe ratio, and maximum drawdown. The best portfolio combination using our signals provides a low-risk, continuous stream of returns with a Sharpe ratio of more than 2.5, maximum drawdown of around 3%, and a near-zero correlation with the S&P 500 market benchmark. We also compare the model's performance through different market regimes, such as the recent volatile movements of the US equity market in the beginning of 2025. Our analysis showcases the robustness of the model and significantly stable performance during these volatile periods. Collectively, these findings show that market inefficiencies can be systematically harvested with modest computational overhead if the right variables are considered. This report will emphasize the potential of traditional deep learning frameworks for generating an AI-driven edge in the financial market.
comment: To get access to the data, please contact s.ghatak@increasealpha.com The analysis now includes model performance through Q3 2025
♻ ☆ Quantizer Design for Finite Model Approximations, Model Learning, and Quantized Q-Learning for MDPs with Unbounded Spaces
In this paper, for Markov decision processes (MDPs) with unbounded state spaces we present refined upper bounds presented in [Kara et. al. JMLR'23] on finite model approximation errors via optimizing the quantizers used for finite model approximations. We also consider implications on quantizer design for quantized Q-learning and empirical model learning, and the performance of policies obtained via Q-learning where the quantized state is treated as the state itself. We highlight the distinctions between planning, where approximating MDPs can be independently designed, and learning (either via Q-learning or empirical model learning), where approximating MDPs are restricted to be defined by invariant measures of Markov chains under exploration policies, leading to significant subtleties on quantizer design performance, even though asymptotic near optimality can be established under both setups. In particular, under Lyapunov growth conditions, we obtain explicit upper bounds which decay to zero as the number of bins approaches infinity
♻ ☆ OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding NeurIPS 2025
Speculative decoding generally dictates having a small, efficient draft model that is either pretrained or distilled offline to a particular target model series, for instance, Llama or Qwen models. However, within online deployment settings, there are two major challenges: 1) usage of a target model that is incompatible with the draft model; 2) expectation of latency improvements over usage and time. In this work, we propose OmniDraft, a unified framework that enables a single draft model to operate with any target model and adapt dynamically to user data. We introduce an online n-gram cache with hybrid distillation fine-tuning to address the cross-vocabulary mismatch across draft and target models; and further improve decoding speed by leveraging adaptive drafting techniques. OmniDraft is particularly suitable for on-device LLM applications where model cost, efficiency and user customization are the major points of contention. This further highlights the need to tackle the above challenges and motivates the \textit{``one drafter for all''} paradigm. We showcase the proficiency of the OmniDraft framework by performing online learning on math reasoning, coding and text generation tasks. Notably, OmniDraft enables a single Llama-68M model to pair with various target models including Vicuna-7B, Qwen2-7B and Llama3-8B models for speculative decoding; and additionally provides up to 1.5-2x speedup.
comment: Accepted to NeurIPS 2025
♻ ☆ One Prompt Fits All: Universal Graph Adaptation for Pretrained Models NeurIPS 2025
Graph Prompt Learning (GPL) has emerged as a promising paradigm that bridges graph pretraining models and downstream scenarios, mitigating label dependency and the misalignment between upstream pretraining and downstream tasks. Although existing GPL studies explore various prompt strategies, their effectiveness and underlying principles remain unclear. We identify two critical limitations: (1) Lack of consensus on underlying mechanisms: Despite current GPLs have advanced the field, there is no consensus on how prompts interact with pretrained models, as different strategies intervene at varying spaces within the model, i.e., input-level, layer-wise, and representation-level prompts. (2) Limited scenario adaptability: Most methods fail to generalize across diverse downstream scenarios, especially under data distribution shifts (e.g., homophilic-to-heterophilic graphs). To address these issues, we theoretically analyze existing GPL approaches and reveal that representation-level prompts essentially function as fine-tuning a simple downstream classifier, proposing that graph prompt learning should focus on unleashing the capability of pretrained models, and the classifier should adapt to downstream scenarios. Based on our findings, we propose UniPrompt, a novel GPL method that adapts any pretrained models, unleashing the capability of pretrained models while preserving the input graph. Extensive experiments demonstrate that our method can effectively integrate with various pretrained models and achieve strong performance across in-domain and cross-domain scenarios.
comment: accepted by NeurIPS 2025 main conference
♻ ☆ PPA-Game: Characterizing and Learning Competitive Dynamics Among Online Content Creators KDD 2025
In this paper, we present the Proportional Payoff Allocation Game (PPA-Game), which characterizes situations where agents compete for divisible resources. In the PPA-game, agents select from available resources, and their payoffs are proportionately determined based on heterogeneous weights attributed to them. Such dynamics simulate content creators on online recommender systems like YouTube and TikTok, who compete for finite consumer attention, with content exposure reliant on inherent and distinct quality. We first conduct a game-theoretical analysis of the PPA-Game. While the PPA-Game does not always guarantee the existence of a pure Nash equilibrium (PNE), we identify prevalent scenarios ensuring its existence. Simulated experiments further prove that the cases where PNE does not exist rarely happen. Beyond analyzing static payoffs, we further discuss the agents' online learning about resource payoffs by integrating a multi-player multi-armed bandit framework. We propose an online algorithm facilitating each agent's maximization of cumulative payoffs over $T$ rounds. Theoretically, we establish that the regret of any agent is bounded by $O(\log^{1 + \eta} T)$ for any $\eta > 0$. Empirical results further validate the effectiveness of our online learning approach.
comment: KDD 2025
♻ ☆ Denoised Diffusion for Object-Focused Image Augmentation
Modern agricultural operations increasingly rely on integrated monitoring systems that combine multiple data sources for farm optimization. Aerial drone-based animal health monitoring serves as a key component but faces limited data availability, compounded by scene-specific issues such as small, occluded, or partially visible animals. Transfer learning approaches often fail to address this limitation due to the unavailability of large datasets that reflect specific farm conditions, including variations in animal breeds, environments, and behaviors. Therefore, there is a need for developing a problem-specific, animal-focused data augmentation strategy tailored to these unique challenges. To address this gap, we propose an object-focused data augmentation framework designed explicitly for animal health monitoring in constrained data settings. Our approach segments animals from backgrounds and augments them through transformations and diffusion-based synthesis to create realistic, diverse scenes that enhance animal detection and monitoring performance. Our initial experiments demonstrate that our augmented dataset yields superior performance compared to our baseline models on the animal detection task. By generating domain-specific data, our method empowers real-time animal health monitoring solutions even in data-scarce scenarios, bridging the gap between limited data and practical applicability.
♻ ☆ A Unified Framework for Variable Selection in Model-Based Clustering with Missing Not at Random
Model-based clustering integrated with variable selection is a powerful tool for uncovering latent structures within complex data. However, its effectiveness is often hindered by challenges such as identifying relevant variables that define heterogeneous subgroups and handling data that are missing not at random, a prevalent issue in fields like transcriptomics. While several notable methods have been proposed to address these problems, they typically tackle each issue in isolation, thereby limiting their flexibility and adaptability. This paper introduces a unified framework designed to address these challenges simultaneously. Our approach incorporates a data-driven penalty matrix into penalized clustering to enable more flexible variable selection, along with a mechanism that explicitly models the relationship between missingness and latent class membership. We demonstrate that, under certain regularity conditions, the proposed framework achieves both asymptotic consistency and selection consistency, even in the presence of missing data. This unified strategy significantly enhances the capability and efficiency of model-based clustering, advancing methodologies for identifying informative variables that define homogeneous subgroups in the presence of complex missing data patterns. The performance of the framework, including its computational efficiency, is evaluated through simulations and demonstrated using both synthetic and real-world transcriptomic datasets.
comment: Binh H. Ho, Long Nguyen Chi, and TrungTin Nguyen are co-first authors
♻ ☆ Malice in Agentland: Down the Rabbit Hole of Backdoors in the AI Supply Chain
The practice of fine-tuning AI agents on data from their own interactions--such as web browsing or tool use--, while being a strong general recipe for improving agentic capabilities, also introduces a critical security vulnerability within the AI supply chain. In this work, we show that adversaries can easily poison the data collection pipeline to embed hard-to-detect backdoors that are triggerred by specific target phrases, such that when the agent encounters these triggers, it performs an unsafe or malicious action. We formalize and validate three realistic threat models targeting different layers of the supply chain: 1) direct poisoning of fine-tuning data, where an attacker controls a fraction of the training traces; 2) environmental poisoning, where malicious instructions are injected into webpages scraped or tools called while creating training data; and 3) supply chain poisoning, where a pre-backdoored base model is fine-tuned on clean data to improve its agentic capabilities. Our results are stark: by poisoning as few as 2% of the collected traces, an attacker can embed a backdoor causing an agent to leak confidential user information with over 80% success when a specific trigger is present. This vulnerability holds across all three threat models. Furthermore, we demonstrate that prominent safeguards, including two guardrail models and one weight-based defense, fail to detect or prevent the malicious behavior. These findings highlight an urgent threat to agentic AI development and underscore the critical need for rigorous security vetting of data collection processes and end-to-end model supply chains.
comment: 27 pages
♻ ☆ Dimension Reduction with Locally Adjusted Graphs AAAI 2025
Dimension reduction (DR) algorithms have proven to be extremely useful for gaining insight into large-scale high-dimensional datasets, particularly finding clusters in transcriptomic data. The initial phase of these DR methods often involves converting the original high-dimensional data into a graph. In this graph, each edge represents the similarity or dissimilarity between pairs of data points. However, this graph is frequently suboptimal due to unreliable high-dimensional distances and the limited information extracted from the high-dimensional data. This problem is exacerbated as the dataset size increases. If we reduce the size of the dataset by selecting points for a specific sections of the embeddings, the clusters observed through DR are more separable since the extracted subgraphs are more reliable. In this paper, we introduce LocalMAP, a new dimensionality reduction algorithm that dynamically and locally adjusts the graph to address this challenge. By dynamically extracting subgraphs and updating the graph on-the-fly, LocalMAP is capable of identifying and separating real clusters within the data that other DR methods may overlook or combine. We demonstrate the benefits of LocalMAP through a case study on biological datasets, highlighting its utility in helping users more accurately identify clusters for real-world problems.
comment: Accepted by AAAI 2025
♻ ☆ Learning to sample fibers for goodness-of-fit testing
We consider the problem of constructing exact goodness-of-fit tests for discrete exponential family models. This classical problem remains practically unsolved for many types of structured or sparse data, as it rests on a computationally difficult core task: to produce a reliable sample from lattice points in a high-dimensional polytope. We translate the problem into a Markov decision process and demonstrate a reinforcement learning approach for learning `good moves' for sampling. We illustrate the approach on data sets and models for which traditional MCMC samplers converge too slowly due to problem size, sparsity structure, and the requirement to use prohibitive non-linear algebra computations in the process. The differentiating factor is the use of scalable tools from \emph{linear} algebra in the context of theoretical guarantees provided by \emph{non-linear} algebra. Our algorithm is based on an actor-critic sampling scheme, with provable convergence. The discovered moves can be used to efficiently obtain an exchangeable sample, significantly cutting computational times with regards to statistical testing.
comment: Most notable change is the link to code on github that implements the actor critic fiber sampler algorithm in python. Otherwise, minor revision adding some examples
♻ ☆ Go With the Flow: Fast Diffusion for Gaussian Mixture Models NIPS 2025
Schrodinger Bridges (SBs) are diffusion processes that steer, in finite time, a given initial distribution to another final one while minimizing a suitable cost functional. Although various methods for computing SBs have recently been proposed in the literature, most of these approaches require computationally expensive training schemes, even for solving low-dimensional problems. In this work, we propose an analytic parametrization of a set of feasible policies for steering the distribution of a dynamical system from one Gaussian Mixture Model (GMM) to another. Instead of relying on standard non-convex optimization techniques, the optimal policy within the set can be approximated as the solution of a low-dimensional linear program whose dimension scales linearly with the number of components in each mixture. The proposed method generalizes naturally to more general classes of dynamical systems, such as controllable linear time-varying systems, enabling efficient solutions to multi-marginal momentum SBs between GMMs, a challenging distribution interpolation problem. We showcase the potential of this approach in low-to-moderate dimensional problems such as image-to-image translation in the latent space of an autoencoder, learning of cellular dynamics using multi-marginal momentum SBs, and various other examples. The implementation is publicly available at https://github.com/georgeRapa/GMMflow.
comment: NIPS 2025 (spotlight)
♻ ☆ PEAR: Planner-Executor Agent Robustness Benchmark
Large Language Model (LLM)-based Multi-Agent Systems (MAS) have emerged as a powerful paradigm for tackling complex, multi-step tasks across diverse domains. However, despite their impressive capabilities, MAS remain susceptible to adversarial manipulation. Existing studies typically examine isolated attack surfaces or specific scenarios, leaving a lack of holistic understanding of MAS vulnerabilities. To bridge this gap, we introduce PEAR, a benchmark for systematically evaluating both the utility and vulnerability of planner-executor MAS. While compatible with various MAS architectures, our benchmark focuses on the planner-executor structure, which is a practical and widely adopted design. Through extensive experiments, we find that (1) a weak planner degrades overall clean task performance more severely than a weak executor; (2) while a memory module is essential for the planner, having a memory module for the executor does not impact the clean task performance; (3) there exists a trade-off between task performance and robustness; and (4) attacks targeting the planner are particularly effective at misleading the system. These findings offer actionable insights for enhancing the robustness of MAS and lay the groundwork for principled defenses in multi-agent settings.
♻ ☆ Can ChatGPT support software verification?
Large language models have become increasingly effective in software engineering tasks such as code generation, debugging and repair. Language models like ChatGPT can not only generate code, but also explain its inner workings and in particular its correctness. This raises the question whether we can utilize ChatGPT to support formal software verification. In this paper, we take some first steps towards answering this question. More specifically, we investigate whether ChatGPT can generate loop invariants. Loop invariant generation is a core task in software verification, and the generation of valid and useful invariants would likely help formal verifiers. To provide some first evidence on this hypothesis, we ask ChatGPT to annotate 106 C programs with loop invariants. We check validity and usefulness of the generated invariants by passing them to two verifiers, Frama-C and CPAchecker. Our evaluation shows that ChatGPT is able to produce valid and useful invariants allowing Frama-C to verify tasks that it could not solve before. Based on our initial insights, we propose ways of combining ChatGPT (or large language models in general) and software verifiers, and discuss current limitations and open issues.
comment: accepted at Fundamental Approaches to Software Engineering 2024
♻ ☆ A Generalized Information Bottleneck Theory of Deep Learning
The Information Bottleneck (IB) principle offers a compelling theoretical framework to understand how neural networks (NNs) learn. However, its practical utility has been constrained by unresolved theoretical ambiguities and significant challenges in accurate estimation. In this paper, we present a \textit{Generalized Information Bottleneck (GIB)} framework that reformulates the original IB principle through the lens of synergy, i.e., the information obtainable only through joint processing of features. We provide theoretical and empirical evidence demonstrating that synergistic functions achieve superior generalization compared to their non-synergistic counterparts. Building on these foundations we re-formulate the IB using a computable definition of synergy based on the average interaction information (II) of each feature with those remaining. We demonstrate that the original IB objective is upper bounded by our GIB in the case of perfect estimation, ensuring compatibility with existing IB theory while addressing its limitations. Our experimental results demonstrate that GIB consistently exhibits compression phases across a wide range of architectures (including those with \textit{ReLU} activations where the standard IB fails), while yielding interpretable dynamics in both CNNs and Transformers and aligning more closely with our understanding of adversarial robustness.
♻ ☆ Variational Rank Reduction Autoencoders
Deterministic Rank Reduction Autoencoders (RRAEs) enforce by construction a regularization on the latent space by applying a truncated SVD. While this regularization makes Autoencoders more powerful, using them for generative purposes is counter-intuitive due to their deterministic nature. On the other hand, Variational Autoencoders (VAEs) are well known for their generative abilities by learning a probabilistic latent space. In this paper, we present Variational Rank Reduction Autoencoders (VRRAEs), a model that leverages the advantages of both RRAEs and VAEs. Our claims and results show that when carefully sampling the latent space of RRAEs and further regularizing with the Kullback-Leibler (KL) divergence (similarly to VAEs), VRRAEs outperform RRAEs and VAEs. Additionally, we show that the regularization induced by the SVD not only makes VRRAEs better generators than VAEs, but also reduces the possibility of posterior collapse. Our results include a synthetic dataset of a small size that showcases the robustness of VRRAEs against collapse, and three real-world datasets; the MNIST, CelebA, and CIFAR-10, over which VRRAEs are shown to outperform both VAEs and RRAEs on many random generation and interpolation tasks based on the FID score. We developed an open-source implementation of VRRAEs in JAX (Equinox), available at https://github.com/JadM133/RRAEs.git.
♻ ☆ General Exploratory Bonus for Optimistic Exploration in RLHF
Optimistic exploration is central to improving sample efficiency in reinforcement learning with human feedback, yet existing exploratory bonus methods to incentivize exploration often fail to realize optimism. We provide a theoretical analysis showing that current formulations, under KL or $\alpha$-divergence regularization, unintentionally bias exploration toward high-probability regions of the reference model, thereby reinforcing conservative behavior instead of promoting discovery of uncertain regions. To address this pitfall, we introduce the General Exploratory Bonus (GEB), a novel theoretical framework that provably satisfies the optimism principle. GEB counteracts divergence-induced bias via reference-dependent reward regulation and unifies prior heuristic bonuses as special cases, while extending naturally across the full $\alpha$-divergence family. Empirically, GEB consistently outperforms baselines on alignment tasks across multiple divergence settings and large language model backbones. These results demonstrate that GEB offers both a principled and practical solution for optimistic exploration in RLHF.
♻ ☆ FlagEval Findings Report: A Preliminary Evaluation of Large Reasoning Models on Automatically Verifiable Textual and Visual Questions NeurIPS 2025
We conduct a moderate-scale contamination-free (to some extent) evaluation of current large reasoning models (LRMs) with some preliminary findings. We also release ROME, our evaluation benchmark for vision language models intended to test reasoning from visual clues. We attach links to the benchmark, evaluation data, and other updates on this website: https://flageval-baai.github.io/LRM-Eval/
comment: Project homepage: https://flageval-baai.github.io/LRM-Eval/ This work will also be presented at NeurIPS 2025 Workshop on Foundations of Reasoning in Language Models (FoRLM)
♻ ☆ Central Limit Theorems for Asynchronous Averaged Q-Learning
This paper establishes central limit theorems for Polyak-Ruppert averaged Q-learning under asynchronous updates. We prove a non-asymptotic central limit theorem, where the convergence rate in Wasserstein distance explicitly reflects the dependence on the number of iterations, state-action space size, the discount factor, and the quality of exploration. In addition, we derive a functional central limit theorem, showing that the partial-sum process converges weakly to a Brownian motion.
♻ ☆ Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding
Large vision-language models (LVLMs) achieve strong performance on multimodal tasks, yet they often default to their language prior (LP) -- memorized textual patterns from pre-training while under-utilizing visual evidence. Prior analyses of LP mostly rely on input-output probing, which fails to reveal the internal mechanisms governing when and how vision influences model behavior. To address this gap, we present the first systematic analysis of language prior through the lens of chain-of-embedding, which examines the layer-wise representation dynamics within LVLMs. Our analysis reveals a universal phenomenon: each model exhibits a Visual Integration Point (VIP), a critical layer at which visual information begins to meaningfully reshape hidden representations and influence decoding. Building on this observation, we introduce the Total Visual Integration (TVI) estimator, which aggregates representation distance beyond the VIP to quantify how strongly visual query influences response generation. Across 54 model-dataset combinations spanning 9 contemporary LVLMs and 6 benchmarks, we demonstrate that VIP consistently emerges, and that TVI reliably predicts the strength of language prior. This offers a principled toolkit for diagnosing and understanding language prior in LVLMs.
♻ ☆ Inverse Design in Nanophotonics via Representation Learning
Inverse design in nanophotonics, the computational discovery of structures achieving targeted electromagnetic (EM) responses, has become a key tool for recent optical advances. Traditional intuition-driven or iterative optimization methods struggle with the inherently high-dimensional, non-convex design spaces and the substantial computational demands of EM simulations. Recently, machine learning (ML) has emerged to address these bottlenecks effectively. This review frames ML-enhanced inverse design methodologies through the lens of representation learning, classifying them into two categories: output-side and input-side approaches. Output-side methods use ML to learn a representation in the solution space to create a differentiable solver that accelerates optimization. Conversely, input-side techniques employ ML to learn compact, latent-space representations of feasible device geometries, enabling efficient global exploration through generative models. Each strategy presents unique trade-offs in data requirements, generalization capacity, and novel design discovery potentials. Hybrid frameworks that combine physics-based optimization with data-driven representations help escape poor local optima, improve scalability, and facilitate knowledge transfer. We conclude by highlighting open challenges and opportunities, emphasizing complexity management, geometry-independent representations, integration of fabrication constraints, and advancements in multiphysics co-designs.
♻ ☆ MooseAgent: A LLM Based Multi-agent Framework for Automating Moose Simulation
The Finite Element Method (FEM) is widely used in engineering and scientific computing, but its pre-processing, solver configuration, and post-processing stages are often time-consuming and require specialized knowledge. This paper proposes an automated solution framework, MooseAgent, for the multi-physics simulation framework MOOSE, which combines large-scale pre-trained language models (LLMs) with a multi-agent system. The framework uses LLMs to understand user-described simulation requirements in natural language and employs task decomposition and multi-round iterative verification strategies to automatically generate MOOSE input files. To improve accuracy and reduce model hallucinations, the system builds and utilizes a vector database containing annotated MOOSE input cards and function documentation. We conducted experimental evaluations on several typical cases, including heat transfer, mechanics, phase field, and multi-physics coupling. The results show that MooseAgent can automate the MOOSE simulation process to a certain extent, especially demonstrating a high success rate when dealing with relatively simple single-physics problems. The main contribution of this research is the proposal of a multi-agent automated framework for MOOSE, which validates its potential in simplifying finite element simulation processes and lowering the user barrier, providing new ideas for the development of intelligent finite element simulation software. The code for the MooseAgent framework proposed in this paper has been open-sourced and is available at https://github.com/taozhan18/MooseAgent
comment: 11 pages, 3 Figs
♻ ☆ Protein Design with Dynamic Protein Vocabulary NeurIPS 2025
Protein design is a fundamental challenge in biotechnology, aiming to design novel sequences with specific functions within the vast space of possible proteins. Recent advances in deep generative models have enabled function-based protein design from textual descriptions, yet struggle with structural plausibility. Inspired by classical protein design methods that leverage natural protein structures, we explore whether incorporating fragments from natural proteins can enhance foldability in generative models. Our empirical results show that even random incorporation of fragments improves foldability. Building on this insight, we introduce ProDVa, a novel protein design approach that integrates a text encoder for functional descriptions, a protein language model for designing proteins, and a fragment encoder to dynamically retrieve protein fragments based on textual functional descriptions. Experimental results demonstrate that our approach effectively designs protein sequences that are both functionally aligned and structurally plausible. Compared to state-of-the-art models, ProDVa achieves comparable function alignment using less than 0.04% of the training data, while designing significantly more well-folded proteins, with the proportion of proteins having pLDDT above 70 increasing by 7.38% and those with PAE below 10 increasing by 9.6%.
comment: Accepted to NeurIPS 2025 (Spotlight)
♻ ☆ Optimistic Multi-Agent Policy Gradient ICML 2024
*Relative overgeneralization* (RO) occurs in cooperative multi-agent learning tasks when agents converge towards a suboptimal joint policy due to overfitting to suboptimal behavior of other agents. No methods have been proposed for addressing RO in multi-agent policy gradient (MAPG) methods although these methods produce state-of-the-art results. To address this gap, we propose a general, yet simple, framework to enable optimistic updates in MAPG methods that alleviate the RO problem. Our approach involves clipping the advantage to eliminate negative values, thereby facilitating optimistic updates in MAPG. The optimism prevents individual agents from quickly converging to a local optimum. Additionally, we provide a formal analysis to show that the proposed method retains optimality at a fixed point. In extensive evaluations on a diverse set of tasks including the *Multi-agent MuJoCo* and *Overcooked* benchmarks, our method outperforms strong baselines on 13 out of 19 tested tasks and matches the performance on the rest.
comment: Updated MMDP notation. Published at ICML 2024, 17 pages, 10 figures
♻ ☆ Query Brand Entity Linking in E-Commerce Search
In this work, we address the brand entity linking problem for e-commerce search queries. The entity linking task is done by either i)a two-stage process consisting of entity mention detection followed by entity disambiguation or ii) an end-to-end linking approaches that directly fetch the target entity given the input text. The task presents unique challenges: queries are extremely short (averaging 2.4 words), lack natural language structure, and must handle a massive space of unique brands. We present a two-stage approach combining named-entity recognition with matching, and a novel end-to-end solution using extreme multi-class classification. We validate our solutions by both offline benchmarks and the impact of online A/B test.
♻ ☆ Gen-C: Populating Virtual Worlds with Generative Crowds
Over the past two decades, researchers have made significant steps in simulating agent-based human crowds, yet most efforts remain focused on low-level tasks such as collision avoidance, path following, and flocking. Realistic simulations, however, require modeling high-level behaviors that emerge from agents interacting with each other and with their environment over time. We introduce Generative Crowds (Gen-C), a generative framework that produces crowd scenarios capturing agent-agent and agent-environment interactions, shaping coherent high-level crowd plans. To avoid the labor-intensive process of collecting and annotating real crowd video data, we leverage large language models (LLMs) to bootstrap synthetic datasets of crowd scenarios. We propose a time-expanded graph representation, encoding actions, interactions, and spatial context. Gen-C employs a dual Variational Graph Autoencoder (VGAE) architecture that jointly learns connectivity patterns and node features conditioned on textual and structural signals, overcoming the limitations of direct LLM generation to enable scalable, environment-aware multi-agent crowd simulations. We demonstrate the effectiveness of Gen-C on scenarios with diverse behaviors such as a University Campus and a Train Station, showing that it generates heterogeneous crowds, coherent interactions, and high-level decision-making patterns consistent with real-world crowd dynamics.
comment: 11 pages
♻ ☆ Resource-Constrained Federated Continual Learning: What Does Matter?
Federated Continual Learning (FCL) aims to enable sequentially privacy-preserving model training on streams of incoming data that vary in edge devices by preserving previous knowledge while adapting to new data. Current FCL literature focuses on restricted data privacy and access to previously seen data while imposing no constraints on the training overhead. This is unreasonable for FCL applications in real-world scenarios, where edge devices are primarily constrained by resources such as storage, computational budget, and label rate. We revisit this problem with a large-scale benchmark and analyze the performance of state-of-the-art FCL approaches under different resource-constrained settings. Various typical FCL techniques and six datasets in two incremental learning scenarios (Class-IL and Domain-IL) are involved in our experiments. Through extensive experiments amounting to a total of over 1,000+ GPU hours, we find that, under limited resource-constrained settings, existing FCL approaches, with no exception, fail to achieve the expected performance. Our conclusions are consistent in the sensitivity analysis. This suggests that most existing FCL methods are particularly too resource-dependent for real-world deployment. Moreover, we study the performance of typical FCL techniques with resource constraints and shed light on future research directions in FCL.
comment: arXiv admin note: text overlap with arXiv:2303.11165 by other authors
♻ ☆ Computing Systemic Risk Measures with Graph Neural Networks
This paper investigates systemic risk measures for stochastic financial networks of explicitly modelled bilateral liabilities. We extend the notion of systemic risk measures from Biagini, Fouque, Fritelli and Meyer-Brandis (2019) to graph structured data. In particular, we focus on an aggregation function that is derived from a market clearing algorithm proposed by Eisenberg and Noe (2001). In this setting, we show the existence of an optimal random allocation that distributes the overall minimal bailout capital and secures the network. We study numerical methods for the approximation of systemic risk and optimal random allocations. We propose to use permutation equivariant architectures of neural networks like graph neural networks (GNNs) and a class that we name (extended) permutation equivariant neural networks ((X)PENNs). We compare their performance to several benchmark allocations. The main feature of GNNs and (X)PENNs is that they are permutation equivariant with respect to the underlying graph data. In numerical experiments we find evidence that these permutation equivariant methods are superior to other approaches.
comment: 50 pages
♻ ☆ Dual Perspectives on Non-Contrastive Self-Supervised Learning
The {\em stop gradient} and {\em exponential moving average} iterative procedures are commonly used in non-contrastive approaches to self-supervised learning to avoid representation collapse, with excellent performance in downstream applications in practice. This presentation investigates these procedures from the dual viewpoints of optimization and dynamical systems. We show that, in general, although they {\em do not} optimize the original objective, or {\em any} other smooth function, they {\em do} avoid collapse Following~\citet{Tian21}, but without any of the extra assumptions used in their proofs, we then show using a dynamical system perspective that, in the linear case, minimizing the original objective function without the use of a stop gradient or exponential moving average {\em always} leads to collapse. Conversely, we characterize explicitly the equilibria of the dynamical systems associated with these two procedures in this linear setting as algebraic varieties in their parameter space, and show that they are, in general, {\em asymptotically stable}. Our theoretical findings are illustrated by empirical experiments with real and synthetic data.
♻ ☆ BAAF: A benchmark attention adaptive framework for medical ultrasound image segmentation tasks
The AI-based assisted diagnosis programs have been widely investigated on medical ultrasound images. Complex scenario of ultrasound image, in which the coupled interference of internal and external factors is severe, brings a unique challenge for localize the object region automatically and precisely in ultrasound images. In this study, we seek to propose a more general and robust Benchmark Attention Adaptive Framework (BAAF) to assist doctors segment or diagnose lesions and tissues in ultrasound images more quickly and accurately. Different from existing attention schemes, the BAAF consists of a parallel hybrid attention module (PHAM) and an adaptive calibration mechanism (ACM). Specifically, BAAF first coarsely calibrates the input features from the channel and spatial dimensions, and then adaptively selects more robust lesion or tissue characterizations from the coarse-calibrated feature maps. The design of BAAF further optimizes the "what" and "where" focus and selection problems in CNNs and seeks to improve the segmentation accuracy of lesions or tissues in medical ultrasound images. The method is evaluated on four medical ultrasound segmentation tasks, and the adequate experimental results demonstrate the remarkable performance improvement over existing state-of-the-art methods. In addition, the comparison with existing attention mechanisms also demonstrates the superiority of BAAF. This work provides the possibility for automated medical ultrasound assisted diagnosis and reduces reliance on human accuracy and precision.
comment: 10 pages, 11 figures
♻ ☆ ACCO: Accumulate While You Communicate for Communication-Overlapped Sharded LLM Training
Training LLMs relies on distributed implementations using multiple GPUs to compute gradients in parallel with sharded optimizers. However, synchronizing gradients in data parallel setups introduces communication overhead that grows with the number of workers, limiting parallelization efficiency. Local optimization algorithms reduce communications but incur high memory costs as they prevent optimizer state sharding, hindering scalability. To address this, we propose \textbf{AC}cumulate while \textbf{CO}mmunicate (ACCO), a memory-efficient optimization algorithm for distributed LLM training. By synchronizing delayed gradients while computing new ones, ACCO reduces GPU idle time and supports heterogeneous hardware. To mitigate the convergence issues caused by delayed updates, we introduce a novel technique ensuring training dynamics align with standard distributed optimization. Compared to ZeRO-1, our approach is significantly faster and scales effectively across heterogeneous hardware.
♻ ☆ Nash Equilibria in Games with Playerwise Concave Coupling Constraints: Existence and Computation
We study the existence and computation of Nash equilibria in continuous static games where the players' admissible strategies are subject to shared coupling constraints, i.e., constraints that depend on their \emph{joint} strategies. Specifically, we focus on a class of games characterized by playerwise concave utilities and playerwise concave constraints. Prior results on the existence of Nash equilibria are not applicable to this class, as they rely on strong assumptions such as joint convexity of the feasible set. By leveraging topological fixed point theory and novel structural insights into the contractibility of feasible sets under playerwise concave constraints, we give an existence proof for Nash equilibria under weaker conditions. Having established existence, we then focus on the computation of Nash equilibria via independent gradient methods under the additional assumption that the utilities admit a potential function. To account for the possibly nonconvex feasible region, we employ a log barrier regularized gradient ascent with adaptive stepsizes. Starting from an initial feasible strategy profile and under exact gradient feedback, the proposed method converges to an $\epsilon$-approximate constrained Nash equilibrium within $\mathcal{O}(\epsilon^{-3})$ iterations.
♻ ☆ Wasserstein-based Kernel Principal Component Analysis for Clustering Applications
Many data clustering applications must handle objects that cannot be represented as vectors. In this context, the bag-of-vectors representation describes complex objects through discrete distributions, for which the Wasserstein distance provides a well-conditioned dissimilarity measure. Kernel methods extend this by embedding distance information into feature spaces that facilitate analysis. However, an unsupervised framework that combines kernels with Wasserstein distances for clustering distributional data is still lacking. We address this gap by introducing a computationally tractable framework that integrates Wasserstein metrics with kernel methods for clustering. The framework can accommodate both vectorial and distributional data, enabling applications in various domains. It comprises three components: (i) an efficient approximation of pairwise Wasserstein distances using multiple reference distributions; (ii) shifted positive definite kernel functions based on Wasserstein distances, combined with kernel principal component analysis for feature mapping; and (iii) scalable, distance-agnostic validity indices for clustering evaluation and kernel parameter optimization. Experiments on power distribution graphs and real-world time series demonstrate the effectiveness and efficiency of the proposed framework.
♻ ☆ Offline Fictitious Self-Play for Competitive Games
Offline Reinforcement Learning (RL) enables policy improvement from fixed datasets without online interactions, making it highly suitable for real-world applications lacking efficient simulators. Despite its success in the single-agent setting, offline multi-agent RL remains a challenge, especially in competitive games. Firstly, unaware of the game structure, it is impossible to interact with the opponents and conduct a major learning paradigm, self-play, for competitive games. Secondly, real-world datasets cannot cover all the state and action space in the game, resulting in barriers to identifying Nash equilibrium (NE). To address these issues, this paper introduces OFF-FSP, the first practical model-free offline RL algorithm for competitive games. We start by simulating interactions with various opponents by adjusting the weights of the fixed dataset with importance sampling. This technique allows us to learn the best responses to different opponents and employ the Offline Self-Play learning framework. To overcome the challenge of partial coverage, we combine the single-agent offline RL method with Fictitious Self-Play (FSP) to approximate NE by constraining the approximate best responses away from out-of-distribution actions. Experiments on matrix games, extensive-form poker, and board games demonstrate that OFF-FSP achieves significantly lower exploitability than state-of-the-art baselines. Finally, we validate OFF-FSP on a real-world human-robot competitive task, demonstrating its potential for solving complex, hard-to-simulate real-world problems.
♻ ☆ SCAN: Self-Denoising Monte Carlo Annotation for Robust Process Reward Learning NeurIPS 2025
Process reward models (PRMs) offer fine-grained, step-level evaluations that facilitate deeper reasoning processes in large language models (LLMs), proving effective in complex tasks like mathematical reasoning. However, developing PRMs is challenging due to the high cost and limited scalability of human-annotated data. Synthetic data from Monte Carlo (MC) estimation is a promising alternative but suffers from a high noise ratio, which can cause overfitting and hinder large-scale training. In this work, we conduct a preliminary study on the noise distribution in synthetic data from MC estimation, identifying that annotation models tend to both underestimate and overestimate step correctness due to limitations in their annotation capabilities. Building on these insights, we propose Self-Denoising Monte Carlo Annotation (SCAN), an efficient data synthesis and noise-tolerant learning framework. Our key findings indicate that: (1) Even lightweight models (e.g., 1.5B parameters) can produce high-quality annotations through a self-denoising strategy, enabling PRMs to achieve superior performance with only 6% the inference cost required by vanilla MC estimation. (2) With our robust learning strategy, PRMs can effectively learn from this weak supervision, achieving a 39.2 F1 score improvement (from 19.9 to 59.1) in ProcessBench. Despite using only a compact synthetic dataset, our models surpass strong baselines, including those trained on large-scale human-annotated datasets such as PRM800K. Furthermore, performance continues to improve as we scale up the synthetic data, highlighting the potential of SCAN for scalable, cost-efficient, and robust PRM training.
comment: NeurIPS 2025. Project page: https://scan-prm.github.io/
♻ ☆ Split Conformal Classification with Unsupervised Calibration
Methods for split conformal prediction leverage calibration samples to transform any prediction rule into a set-prediction rule that complies with a target coverage probability. Existing methods provide remarkably strong performance guarantees with minimal computational costs. However, they require to use calibration samples composed by labeled examples different to those used for training. This requirement can be highly inconvenient, as it prevents the use of all labeled examples for training and may require acquiring additional labels solely for calibration. This paper presents an effective methodology for split conformal prediction with unsupervised calibration for classification tasks. In the proposed approach, set-prediction rules are obtained using unsupervised calibration samples together with supervised training samples previously used to learn the classification rule. Theoretical and experimental results show that the presented methods can achieve performance comparable to that with supervised calibration, at the expenses of a moderate degradation in performance guarantees and computational efficiency.
♻ ☆ SpaPool: Soft Partition Assignment Pooling for__Graph Neural Networks
This paper introduces SpaPool, a novel pooling method that combines the strengths of both dense and sparse techniques for a graph neural network. SpaPool groups vertices into an adaptive number of clusters, leveraging the benefits of both dense and sparse approaches. It aims to maintain the structural integrity of the graph while reducing its size efficiently. Experimental results on several datasets demonstrate that SpaPool achieves competitive performance compared to existing pooling techniques and excels particularly on small-scale graphs. This makes SpaPool a promising method for applications requiring efficient and effective graph processing.
♻ ☆ Feature Distillation is the Better Choice for Model-Heterogeneous Federated Learning
Model-Heterogeneous Federated Learning (Hetero-FL) has attracted growing attention for its ability to aggregate knowledge from heterogeneous models while keeping private data locally. To better aggregate knowledge from clients, ensemble distillation, as a widely used and effective technique, is often employed after global aggregation to enhance the performance of the global model. However, simply combining Hetero-FL and ensemble distillation does not always yield promising results and can make the training process unstable. The reason is that existing methods primarily focus on logit distillation, which, while being model-agnostic with softmax predictions, fails to compensate for the knowledge bias arising from heterogeneous models. To tackle this challenge, we propose a stable and efficient Feature Distillation for model-heterogeneous Federated learning, dubbed FedFD, that can incorporate aligned feature information via orthogonal projection to integrate knowledge from heterogeneous models better. Specifically, a new feature-based ensemble federated knowledge distillation paradigm is proposed. The global model on the server needs to maintain a projection layer for each client-side model architecture to align the features separately. Orthogonal techniques are employed to re-parameterize the projection layer to mitigate knowledge bias from heterogeneous models and thus maximize the distilled knowledge. Extensive experiments show that FedFD achieves superior performance compared to state-of-the-art methods.
♻ ☆ MaxPoolBERT: Enhancing BERT Classification via Layer- and Token-Wise Aggregation
The [CLS] token in BERT is commonly used as a fixed-length representation for classification tasks, yet prior work has shown that both other tokens and intermediate layers encode valuable contextual information. In this work, we study lightweight extensions to BERT that refine the [CLS] representation by aggregating information across layers and tokens. Specifically, we explore three modifications: (i) max-pooling the [CLS] token across multiple layers, (ii) enabling the [CLS] token to attend over the entire final layer using an additional multi-head attention (MHA) layer, and (iii) combining max-pooling across the full sequence with MHA. Our approach, called MaxPoolBERT, enhances BERT's classification accuracy (especially on low-resource tasks) without requiring new pre-training or significantly increasing model size. Experiments on the GLUE benchmark show that MaxPoolBERT consistently achieves a better performance than the standard BERT base model on low resource tasks of the GLUE benchmark.
♻ ☆ Riemannian generative decoder ICML 2025
Riemannian representation learning typically relies on an encoder to estimate densities on chosen manifolds. This involves optimizing numerically brittle objectives, potentially harming model training and quality. To completely circumvent this issue, we introduce the Riemannian generative decoder, a unifying approach for finding manifold-valued latents on any Riemannian manifold. Latents are learned with a Riemannian optimizer while jointly training a decoder network. By discarding the encoder, we vastly simplify the manifold constraint compared to current approaches which often only handle few specific manifolds. We validate our approach on three case studies -- a synthetic branching diffusion process, human migrations inferred from mitochondrial DNA, and cells undergoing a cell division cycle -- each showing that learned representations respect the prescribed geometry and capture intrinsic non-Euclidean structure. Our method requires only a decoder, is compatible with existing architectures, and yields interpretable latent spaces aligned with data geometry. Code available on https://github.com/yhsure/riemannian-generative-decoder.
comment: GenBio ICML 2025 (Proceedings of the Workshop on Generative AI for Biology at the 42nd International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025)
♻ ☆ Causal Ordering Without Effect Estimation: A Framework for Using Proxies in Treatment Prioritization
Who should we prioritize for treatment when causal effects cannot be estimated? In practice, organizations often rely on predictive proxies: ads are targeted using purchase probabilities, and retention incentives are allocated using churn-risk scores. These models are not causal, but they are often used with the aim of ranking individuals by treatment effects, a task we call causal ordering. We develop a decision-focused framework to reason about this practice. We identify conditions under which proxies recover the correct effect ordering, which hold when a proxy reflects a dominant moderator of treatment effects. We show how these conditions emerge as a useful approximation in discrete choice settings, where the propensity to act without an intervention moderates persuasion. Moreover, we extend beyond this case, demonstrating that proxies capturing a non-dominant moderator can still outperform CATE estimates when they target signals that are easier to estimate precisely. Building on these insights, we introduce diagnostic tools to assess proxy usefulness in practice. Finally, we illustrate the framework in advertising, where a simple predictive proxy outperforms heterogeneous-effect estimation methods.
♻ ☆ BrainForm: a Serious Game for BCI Training and Data Collection
$\textit{BrainForm}$ is a gamified Brain-Computer Interface (BCI) training system designed for scalable data collection using consumer hardware and a minimal setup. We investigated (1) how users develop BCI control skills across repeated sessions and (2) perceptual and performance effects of two visual stimulation textures. Game Experience Questionnaire (GEQ) scores for Flow, Positive Affect, Competence and Challenge were strongly positive, indicating sustained engagement. A within-subject study with multiple runs, two task complexities, and post-session questionnaires revealed no significant performance differences between textures but increased ocular irritation over time. Online metrics$\unicode{x2013}$Task Accuracy, Task Time, and Information Transfer Rate$\unicode{x2013}$improved across sessions, confirming learning effects for symbol spelling, even under pressure conditions. Our results highlight the potential of $\textit{BrainForm}$ as a scalable, user-friendly BCI research tool and offer guidance for sustained engagement and reduced training fatigue.
comment: 15 pages, 6 figures. Author-accepted version. Accepted for presentation at the Brain Informatics 2025 conference, to appear in Springer Lecture Notes in Artificial Intelligence (LNAI) Brain Informatics Books Series. The final authenticated version will be available via SpringerLink
♻ ☆ Optimally Deep Networks -- Adapting Model Depth to Datasets for Superior Efficiency
Deep neural networks (DNNs) have provided brilliant performance across various tasks. However, this success often comes at the cost of unnecessarily large model sizes, high computational demands, and substantial memory footprints. Typically, powerful architectures are trained at full depths but not all datasets or tasks require such high model capacity. Training very deep architectures on relatively low-complexity datasets frequently leads to wasted computation, unnecessary energy consumption, and excessive memory usage, which in turn makes deployment of models on resource-constrained devices impractical. To address this problem, we introduce Optimally Deep Networks (ODNs), which provide a balance between model depth and task complexity. Specifically, we propose a NAS like training strategy called progressive depth expansion, which begins by training deep networks at shallower depths and incrementally increases their depth as the earlier blocks converge, continuing this process until the target accuracy is reached. ODNs use only the optimal depth for the given datasets, removing redundant layers. This cuts down future training and inference costs, lowers the memory footprint, enhances computational efficiency, and facilitates deployment on edge devices. Empirical results show that the optimal depths of ResNet-18 and ResNet-34 for MNIST and SVHN, achieve up to 98.64 % and 96.44 % reduction in memory footprint, while maintaining a competitive accuracy of 99.31 % and 96.08 %, respectively.
comment: 6 pages, 3 figures, 1 table
♻ ☆ NinA: Normalizing Flows in Action. Training VLA Models with Normalizing Flows
Recent advances in Vision-Language-Action (VLA) models have established a two-component architecture, where a pre-trained Vision-Language Model (VLM) encodes visual observations and task descriptions, and an action decoder maps these representations to continuous actions. Diffusion models have been widely adopted as action decoders due to their ability to model complex, multimodal action distributions. However, they require multiple iterative denoising steps at inference time or downstream techniques to speed up sampling, limiting their practicality in real-world settings where high-frequency control is crucial. In this work, we present NinA (Normalizing Flows in Action), a fast and expressive alternative to diffusion-based decoders for VLAs. NinA replaces the diffusion action decoder with a Normalizing Flow (NF) that enables one-shot sampling through an invertible transformation, significantly reducing inference time. We integrate NinA into the FLOWER VLA architecture and fine-tune on the LIBERO benchmark. Our experiments show that NinA matches the performance of its diffusion-based counterpart under the same training regime, while achieving substantially faster inference. These results suggest that NinA offers a promising path toward efficient, high-frequency VLA control without compromising performance.
comment: https://github.com/dunnolab/NinA/
♻ ☆ Interpretable Reward Model via Sparse Autoencoder
Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models. Our code is available at https://github.com/schrieffer-z/sarm.
comment: Commercial firm need to review this paper before publishing it
♻ ☆ Constraint Matters: Multi-Modal Representation for Reducing Mixed-Integer Linear programming
Model reduction, which aims to learn a simpler model of the original mixed integer linear programming (MILP), can solve large-scale MILP problems much faster. Most existing model reduction methods are based on variable reduction, which predicts a solution value for a subset of variables. From a dual perspective, constraint reduction that transforms a subset of inequality constraints into equalities can also reduce the complexity of MILP, but has been largely ignored. Therefore, this paper proposes a novel constraint-based model reduction approach for the MILP. Constraint-based MILP reduction has two challenges: 1) which inequality constraints are critical such that reducing them can accelerate MILP solving while preserving feasibility, and 2) how to predict these critical constraints efficiently. To identify critical constraints, we first label these tight-constraints at the optimal solution as potential critical constraints and design a heuristic rule to select a subset of critical tight-constraints. To learn the critical tight-constraints, we propose a multi-modal representation technique that leverages information from both instance-level and abstract-level MILP formulations. The experimental results show that, compared to the state-of-the-art methods, our method improves the quality of the solution by over 50\% and reduces the computation time by 17.47\%.
comment: Since the article needs improvement, it will be temporarily withdrawn
♻ ☆ Persistent Homology via Ellipsoids
Persistent homology is one of the most popular methods in topological data analysis. An initial step in its use involves constructing a nested sequence of simplicial complexes. There is an abundance of different complexes to choose from, with \v{C}ech, Rips, alpha, and witness complexes being popular choices. In this manuscript, we build a novel type of geometrically informed simplicial complex, called a Rips-type ellipsoid complex. This complex is based on the idea that ellipsoids aligned with tangent directions better approximate the data compared to conventional (Euclidean) balls centered at sample points, as used in the construction of Rips and Alpha complexes. We use Principal Component Analysis to estimate tangent spaces directly from samples and present an algorithm for computing Rips-type ellipsoid barcodes, i.e., topological descriptors based on Rips-type ellipsoid complexes. Additionally, we show that the ellipsoid barcodes depend continuously on the input data so that small perturbations of a k-generic point cloud lead to proportionally small changes in the resulting ellipsoid barcodes. This provides a theoretical guarantee analogous, if somewhat weaker, to the classical stability results for Rips and \v{C}ech filtrations. We also conduct extensive experiments and compare Rips-type ellipsoid barcodes with standard Rips barcodes. Our findings indicate that Rips-type ellipsoid complexes are particularly effective for estimating the homology of manifolds and spaces with bottlenecks from samples. In particular, the persistence intervals corresponding to ground-truth topological features are longer compared to those obtained using the Rips complex of the data. Furthermore, Rips-type ellipsoid barcodes lead to better classification results in sparsely sampled point clouds. Finally, we demonstrate that Rips-type ellipsoid barcodes outperform Rips barcodes in classification tasks.
♻ ☆ Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
A key challenge in applying reinforcement learning (RL) to diffusion large language models (dLLMs) lies in the intractability of their likelihood functions, which are essential for the RL objective, necessitating corresponding approximation in each training step. While existing methods approximate the log-likelihoods by their evidence lower bounds (ELBOs) via customized Monte Carlo (MC) sampling, the forward computational graphs of all MC samples need to be retained for the gradient computation of non-linear terms in the RL objective, resulting in significant memory overhead. This constraint restricts feasible sample sizes, leading to imprecise likelihood approximations and ultimately distorting the RL objective. To overcome this limitation, we propose \emph{Boundary-Guided Policy Optimization} (BGPO), a memory-efficient RL algorithm that maximizes a specially constructed lower bound of the ELBO-based objective. This lower bound is carefully designed to satisfy two key properties: (1) Linearity: it is formulated in a linear sum where each term depends only on a single MC sample, thereby enabling gradient accumulation across samples and ensuring constant memory usage; (2) Equivalence: Both the value and gradient of this lower bound are equal to those of the ELBO-based objective in on-policy training, making it also an effective approximation for the original RL objective. These properties allow BGPO to adopt a large MC sample size, resulting in more accurate likelihood approximations and improved RL objective estimation, which in turn leads to enhanced performance. Experiments show that BGPO significantly outperforms previous RL algorithms for dLLMs in math problem solving, code generation, and planning tasks. Our codes and models are available at \href{https://github.com/THU-KEG/BGPO}{https://github.com/THU-KEG/BGPO}.
♻ ☆ LazyEviction: Lagged KV Eviction with Attention Pattern Observation for Efficient Long Reasoning
Large Language Models (LLMs) exhibit enhanced capabilities by Chain-of-Thought reasoning. However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a \textbf{Token Importance Recurrence} phenomenon: a large proportion of tokens regain high attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose \textbf{LazyEviction}, an observation window-based lagged eviction framework retaining latent recurring tokens by prioritized eviction based on tokens' recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache by 50\%\textasciitilde70\% while maintaining comparable accuracy, outperforming existing KV cache compression baselines. Our implementation code can be found at https://github.com/Halo-949/LazyEviction.
♻ ☆ Learning to Explore in Diverse Reward Settings via Temporal-Difference-Error Maximization
Numerous heuristics and advanced approaches have been proposed for exploration in different settings for deep reinforcement learning. Noise-based exploration generally fares well with dense-shaped rewards and bonus-based exploration with sparse rewards. However, these methods usually require additional tuning to deal with undesirable reward settings by adjusting hyperparameters and noise distributions. Rewards that actively discourage exploration, i.e., with an action cost and no other dense signal to follow, can pose a major challenge. We propose a novel exploration method, Stable Error-seeking Exploration (SEE), that is robust across dense, sparse, and exploration-adverse reward settings. To this endeavor, we revisit the idea of maximizing the TD-error as a separate objective. Our method introduces three design choices to mitigate instability caused by far-off-policy learning, the conflict of interest of maximizing the cumulative TD-error in an episodic setting, and the non-stationary nature of TD-errors. SEE can be combined with off-policy algorithms without modifying the optimization pipeline of the original objective. In our experimental analysis, we show that a Soft-Actor Critic agent with the addition of SEE performs robustly across three diverse reward settings in a variety of tasks without hyperparameter adjustments.
♻ ☆ BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals NeurIPS 2025
Electroencephalography (EEG) and magnetoencephalography (MEG) measure neural activity non-invasively by capturing electromagnetic fields generated by dendritic currents. Although rooted in the same biophysics, EEG and MEG exhibit distinct signal patterns, further complicated by variations in sensor configurations across modalities and recording devices. Existing approaches typically rely on separate, modality- and dataset-specific models, which limits the performance and cross-domain scalability. This paper proposes BrainOmni, the first brain foundation model that generalises across heterogeneous EEG and MEG recordings. To unify diverse data sources, we introduce BrainTokenizer,the first tokenizer that quantises spatiotemporal brain activity into discrete representations. Central to BrainTokenizer is a novel Sensor Encoder that encodes sensor properties such as spatial layout, orientation, and type, enabling compatibility across devices and modalities. Building upon the discrete representations, BrainOmni learns unified semantic embeddings of brain signals by self-supervised pretraining. To the best of our knowledge, it is the first foundation model to support both EEG and MEG signals, as well as the first to incorporate large-scale MEG pretraining. A total of 1,997 hours of EEG and 656 hours of MEG data are curated and standardised from publicly available sources for pretraining. Experiments show that BrainOmni outperforms both existing foundation models and state-of-the-art task-specific models on a range of downstream tasks. It also demonstrates strong generalisation to unseen EEG and MEG devices. Further analysis reveals that joint EEG-MEG (EMEG) training yields consistent improvements across both modalities. Code and model checkpoints will be released upon acceptance.
comment: Accepted by the 39th Conference on Neural Information Processing Systems (NeurIPS 2025)
♻ ☆ RIGNO: A Graph-based framework for robust and accurate operator learning for PDEs on arbitrary domains
Learning the solution operators of PDEs on arbitrary domains is challenging due to the diversity of possible domain shapes, in addition to the often intricate underlying physics. We propose an end-to-end graph neural network (GNN) based neural operator to learn PDE solution operators from data on point clouds in arbitrary domains. Our multi-scale model maps data between input/output point clouds by passing it through a downsampled regional mesh. The approach includes novel elements aimed at ensuring spatio-temporal resolution invariance. Our model, termed RIGNO, is tested on a challenging suite of benchmarks composed of various time-dependent and steady PDEs defined on a diverse set of domains. We demonstrate that RIGNO is significantly more accurate than neural operator baselines and robustly generalizes to unseen resolutions both in space and in time. Our code is publicly available at github.com/camlab-ethz/rigno.
♻ ☆ General Demographic Foundation Models for Enhancing Predictive Performance Across Diseases and Populations
Demographic attributes are universally present in electronic health records. They are the most widespread information across populations and diseases, and serve as vital predictors in clinical risk stratification and treatment decisions. Despite their significance, these attributes are often treated as auxiliaries in model design, with limited attention being paid to learning their representations. This study explored the development of a General Demographic Pre-trained (GDP) model as a foundational model tailored to demographic attributes, focusing on age and gender. The model is pre-trained and evaluated using datasets with diverse diseases and populations compositions from different geographic regions. The composition of GDP architecture was explored through examining combinations of ordering approaches and encoding methods to transform tabular demographic inputs into effective latent embeddings. Results demonstrate the feasibility of GDP to generalize across task, diseases, and populations. In detailed composition, the sequential ordering substantially improves model performance in discrimination, calibration, and the corresponding information gain at each decision tree split, particularly in diseases where age and gender contribute significantly to risk stratification. Even in datasets where demographic attributes hold relatively low predictive value, GDP enhances the representational importance, increasing their influence in downstream gradient boosting models. The findings suggest that foundation models for tabular demographic attributes offer a promising direction for improving predictive performance in healthcare applications.
♻ ☆ Steering Large Language Models for Machine Translation Personalization
Large language models have simplified the production of personalized translations reflecting predefined stylistic constraints. However, these systems still struggle when stylistic requirements are implicitly represented by a set of examples, such as texts produced by a specific human translator. In this work, we explore various strategies for personalizing automatically generated translations when few examples are available, with a focus on the challenging domain of literary translation. We begin by determining the feasibility of the task and how style information is encoded within model representations. Then, we evaluate various prompting strategies and inference-time interventions for steering model generations towards a personalized style, with a particular focus on contrastive steering with sparse autoencoder (SAE) latents to identify salient personalization properties. We demonstrate that contrastive SAE steering yields robust style conditioning and translation quality, resulting in higher inference-time computational efficiency than prompting approaches. We further examine the impact of steering on model activations, finding that layers encoding personalization properties are impacted similarly by prompting and SAE steering, suggesting a similar mechanism at play.
♻ ☆ Competitive Advantage Attacks to Decentralized Federated Learning NeurIPS
Decentralized federated learning (DFL) enables clients (e.g., hospitals and banks) to jointly train machine learning models without a central orchestration server. In each global training round, each client trains a local model on its own training data and then they exchange local models for aggregation. In this work, we propose SelfishAttack, a new family of attacks to DFL. In SelfishAttack, a set of selfish clients aim to achieve competitive advantages over the remaining non-selfish ones, i.e., the final learnt local models of the selfish clients are more accurate than those of the non-selfish ones. Towards this goal, the selfish clients send carefully crafted local models to each remaining non-selfish one in each global training round. We formulate finding such local models as an optimization problem and propose methods to solve it when DFL uses different aggregation rules. Theoretically, we show that our methods find the optimal solutions to the optimization problem. Empirically, we show that SelfishAttack successfully increases the accuracy gap (i.e., competitive advantage) between the final learnt local models of selfish clients and those of non-selfish ones. Moreover, SelfishAttack achieves larger accuracy gaps than poisoning attacks when extended to increase competitive advantages.
comment: Conference on Neural Information Processing Systems (NeurIPS) 2025
♻ ☆ Padé Approximant Neural Networks for Enhanced Electric Motor Fault Diagnosis Using Vibration and Acoustic Data
Purpose: The primary aim of this study is to enhance fault diagnosis in induction machines by leveraging the Pad\'e Approximant Neuron (PAON) model. While accelerometers and microphones are standard in motor condition monitoring, deep learning models with nonlinear neuron architectures offer promising improvements in diagnostic performance. This research investigates whether Pad\'e Approximant Neural Networks (Pad\'eNets) can outperform conventional Convolutional Neural Networks (CNNs) and Self-Organized Operational Neural Networks (Self-ONNs) in the diagnosis of electrical and mechanical faults from vibration and acoustic data. Methods: We evaluate and compare the diagnostic capabilities of three deep learning architectures: one-dimensional CNNs, Self-ONNs, and Pad\'eNets. These models are tested on the University of Ottawa's publicly available constant-speed induction motor datasets, which include both vibration and acoustic sensor data. The Pad\'eNet model is designed to introduce enhanced nonlinearity and is compatible with unbounded activation functions such as LeakyReLU. Results and Conclusion: Pad\'eNets consistently outperformed the baseline models, achieving diagnostic accuracies of 99.96%, 98.26%, 97.61%, and 98.33% for accelerometers 1, 2, 3, and the acoustic sensor, respectively. The enhanced nonlinearity of Pad\'eNets, together with their compatibility with unbounded activation functions, significantly improves fault diagnosis performance in induction motor condition monitoring.
comment: This version is the author's accepted manuscript. It has been peer-reviewed and accepted for publication in Journal of Vibration Engineering & Technologies. The final published version is available at https://doi.org/10.1007/s42417-025-02129-5
♻ ☆ GTCN-G: A Residual Graph-Temporal Fusion Network for Imbalanced Intrusion Detection (Preprint)
The escalating complexity of network threats and the inherent class imbalance in traffic data present formidable challenges for modern Intrusion Detection Systems (IDS). While Graph Neural Networks (GNNs) excel in modeling topological structures and Temporal Convolutional Networks (TCNs) are proficient in capturing time-series dependencies, a framework that synergistically integrates both while explicitly addressing data imbalance remains an open challenge. This paper introduces a novel deep learning framework, named Gated Temporal Convolutional Network and Graph (GTCN-G), engineered to overcome these limitations. Our model uniquely fuses a Gated TCN (G-TCN) for extracting hierarchical temporal features from network flows with a Graph Convolutional Network (GCN) designed to learn from the underlying graph structure. The core innovation lies in the integration of a residual learning mechanism, implemented via a Graph Attention Network (GAT). This mechanism preserves original feature information through residual connections, which is critical for mitigating the class imbalance problem and enhancing detection sensitivity for rare malicious activities (minority classes). We conducted extensive experiments on two public benchmark datasets, UNSW-NB15 and ToN-IoT, to validate our approach. The empirical results demonstrate that the proposed GTCN-G model achieves state-of-the-art performance, significantly outperforming existing baseline models in both binary and multi-class classification tasks.
comment: This preprint was submitted to IEEE TrustCom 2025. The accepted version will be published under copyright 2025 IEEE
♻ ☆ EMSEdit: Efficient Multi-Step Meta-Learning-based Model Editing
Large Language Models (LLMs) power numerous AI applications, yet updating their knowledge remains costly. Model editing provides a lightweight alternative through targeted parameter modifications, with meta-learning-based model editing (MLME) demonstrating strong effectiveness and efficiency. However, we find that MLME struggles in low-data regimes and incurs high training costs due to the use of KL divergence. To address these issues, we propose $\textbf{E}$fficient $\textbf{M}$ulti-$\textbf{S}$tep $\textbf{Edit (EMSEdit)}$, which leverages multi-step backpropagation (MSBP) to effectively capture gradient-activation mapping patterns within editing samples, performs multi-step edits per sample to enhance editing performance under limited data, and introduces norm-based regularization to preserve unedited knowledge while improving training efficiency. Experiments on two datasets and three LLMs show that EMSEdit consistently outperforms state-of-the-art methods in both sequential and batch editing. Moreover, MSBP can be seamlessly integrated into existing approaches to yield additional performance gains. Further experiments on a multi-hop reasoning editing task demonstrate EMSEdit's robustness in handling complex edits, while ablation studies validate the contribution of each design component. Our code is available at https://github.com/xpq-tech/emsedit.
♻ ☆ Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator
Post-training of large language models is essential for adapting pre-trained language models (PLMs) to align with human preferences and downstream tasks. While PLMs typically exhibit well-calibrated confidence, post-trained language models (PoLMs) often suffer from over-confidence, assigning high confidence to both correct and incorrect outputs, which can undermine reliability in critical applications. A major obstacle in calibrating PoLMs is the scarcity of labeled data for individual downstream tasks. To address this, we propose Disagreement-Aware Confidence Alignment (DACA), a novel unsupervised method to optimize the parameters (e.g., temperature $\tau$) in post-hoc confidence calibration. Our method is motivated by the under-confidence issue caused by prediction disagreement between the PLM and PoLM while aligning their confidence via temperature scaling. Theoretically, the PLM's confidence underestimates PoLM's prediction accuracy on disagreement examples, causing a larger $\tau$ and producing under-confident predictions. DACA mitigates this by selectively using only agreement examples for calibration, effectively decoupling the influence of disagreement. In this manner, our method avoids an overly large $\tau$ in temperature scaling caused by disagreement examples, improving calibration performance. Extensive experiments demonstrate the effectiveness of our method, improving the average ECE of open-sourced and API-based LLMs (e.g. GPT-4o) by up to 15.08$\%$ on common benchmarks.
Multimedia 9
☆ Omni-Captioner: Data Pipeline, Models, and Benchmark for Omni Detailed Perception
Fine-grained perception of multimodal information is critical for advancing human-AI interaction. With recent progress in audio-visual technologies, Omni Language Models (OLMs), capable of processing audio and video signals in parallel, have emerged as a promising paradigm for achieving richer understanding and reasoning. However, their capacity to capture and describe fine-grained details remains limited explored. In this work, we present a systematic and comprehensive investigation of omni detailed perception from the perspectives of the data pipeline, models, and benchmark. We first identify an inherent "co-growth" between detail and hallucination in current OLMs. To address this, we propose Omni-Detective, an agentic data generation pipeline integrating tool-calling, to autonomously produce highly detailed yet minimally hallucinatory multimodal data. Based on the data generated with Omni-Detective, we train two captioning models: Audio-Captioner for audio-only detailed perception, and Omni-Captioner for audio-visual detailed perception. Under the cascade evaluation protocol, Audio-Captioner achieves the best performance on MMAU and MMAR among all open-source models, surpassing Gemini 2.5 Flash and delivering performance comparable to Gemini 2.5 Pro. On existing detailed captioning benchmarks, Omni-Captioner sets a new state-of-the-art on VDC and achieves the best trade-off between detail and hallucination on the video-SALMONN 2 testset. Given the absence of a dedicated benchmark for omni detailed perception, we design Omni-Cloze, a novel cloze-style evaluation for detailed audio, visual, and audio-visual captioning that ensures stable, efficient, and reliable assessment. Experimental results and analysis demonstrate the effectiveness of Omni-Detective in generating high-quality detailed captions, as well as the superiority of Omni-Cloze in evaluating such detailed captions.
comment: https://github.com/ddlBoJack/Omni-Captioner
☆ M3ST-DTI: A multi-task learning model for drug-target interactions based on multi-modal features and multi-stage alignment
Accurate prediction of drug-target interactions (DTI) is pivotal in drug discovery. However, existing approaches often fail to capture deep intra-modal feature interactions or achieve effective cross-modal alignment, limiting predictive performance and generalization. To address these challenges, we propose M3ST-DTI, a multi-task learning model that enables multi-stage integration and alignment of multi modal features for DTI prediction. M3ST-DTI incorporates three types of features-textual, structural, and functional and enhances intra-modal representations using self-attention mechanisms and a hybrid pooling graph attention module. For early-stage feature alignment and fusion, the model in tegrates MCA with Gram loss as a structural constraint. In the later stage, a BCA module captures fine-grained interactions between drugs and targets within each modality, while a deep orthogonal fusion module mitigates feature redundancy.Extensive evaluations on benchmark datasets demonstrate that M3ST-DTI consistently outperforms state-of-the art methods across diverse metrics
☆ An Empirical Study of Reducing AV1 Decoder Complexity and Energy Consumption via Encoder Parameter Tuning
The widespread adoption of advanced video codecs such as AV1 is often hindered by their high decoding complexity, posing a challenge for battery-constrained devices. While encoders can be configured to produce bitstreams that are decoder-friendly, estimating the decoding complexity and energy overhead for a given video is non-trivial. In this study, we systematically analyse the impact of disabling various coding tools and adjusting coding parameters in two AV1 encoders, libaom-av1 and SVT-AV1. Using system-level energy measurement tools like RAPL (Running Average Power Limit), Intel SoC Watch (integrated with VTune profiler), we quantify the resulting trade-offs between decoding complexity, energy consumption, and compression efficiency for decoding a bitstream. Our results demonstrate that specific encoder configurations can substantially reduce decoding complexity with minimal perceptual quality degradation. For libaom-av1, disabling CDEF, an in-loop filter gives us a mean reduction in decoding cycles by 10%. For SVT-AV1, using the in-built, fast-decode=2 preset achieves a more substantial 24% reduction in decoding cycles. These findings provide strategies for content providers to lower the energy footprint of AV1 video streaming.
comment: Accepted Camera-Ready paper for PCS 2025, 5 Pages
☆ LiteVPNet: A Lightweight Network for Video Encoding Control in Quality-Critical Applications
In the last decade, video workflows in the cinema production ecosystem have presented new use cases for video streaming technology. These new workflows, e.g. in On-set Virtual Production, present the challenge of requiring precise quality control and energy efficiency. Existing approaches to transcoding often fall short of these requirements, either due to a lack of quality control or computational overhead. To fill this gap, we present a lightweight neural network (LiteVPNet) for accurately predicting Quantisation Parameters for NVENC AV1 encoders that achieve a specified VMAF score. We use low-complexity features, including bitstream characteristics, video complexity measures, and CLIP-based semantic embeddings. Our results demonstrate that LiteVPNet achieves mean VMAF errors below 1.2 points across a wide range of quality targets. Notably, LiteVPNet achieves VMAF errors within 2 points for over 87% of our test corpus, c.f. approx 61% with state-of-the-art methods. LiteVPNet's performance across various quality regions highlights its applicability for enhancing high-value content transport and streaming for more energy-efficient, high-quality media experiences.
comment: Accepted PCS 2025 Camera-Ready Version, 5 Pages
☆ Human-in-the-Loop Bandwidth Estimation for Quality of Experience Optimization in Real-Time Video Communication AAAI
The quality of experience (QoE) delivered by video conferencing systems is significantly influenced by accurately estimating the time-varying available bandwidth between the sender and receiver. Bandwidth estimation for real-time communications remains an open challenge due to rapidly evolving network architectures, increasingly complex protocol stacks, and the difficulty of defining QoE metrics that reliably improve user experience. In this work, we propose a deployed, human-in-the-loop, data-driven framework for bandwidth estimation to address these challenges. Our approach begins with training objective QoE reward models derived from subjective user evaluations to measure audio and video quality in real-time video conferencing systems. Subsequently, we collect roughly $1$M network traces with objective QoE rewards from real-world Microsoft Teams calls to curate a bandwidth estimation training dataset. We then introduce a novel distributional offline reinforcement learning (RL) algorithm to train a neural-network-based bandwidth estimator aimed at improving QoE for users. Our real-world A/B test demonstrates that the proposed approach reduces the subjective poor call ratio by $11.41\%$ compared to the baseline bandwidth estimator. Furthermore, the proposed offline RL algorithm is benchmarked on D4RL tasks to demonstrate its generalization beyond bandwidth estimation.
comment: Accepted for publication in the proceedings of the AAAI Conference on Artificial Intelligence 2026 (IAAI Technical Track on Deployed Highly Innovative Applications of AI)
♻ ☆ Probabilistic Temporal Masked Attention for Cross-view Online Action Detection
As a critical task in video sequence classification within computer vision, Online Action Detection (OAD) has garnered significant attention. The sensitivity of mainstream OAD models to varying video viewpoints often hampers their generalization when confronted with unseen sources. To address this limitation, we propose a novel Probabilistic Temporal Masked Attention (PTMA) model, which leverages probabilistic modeling to derive latent compressed representations of video frames in a cross-view setting. The PTMA model incorporates a GRU-based temporal masked attention (TMA) cell, which leverages these representations to effectively query the input video sequence, thereby enhancing information interaction and facilitating autoregressive frame-level video analysis. Additionally, multi-view information can be integrated into the probabilistic modeling to facilitate the extraction of view-invariant features. Experiments conducted under three evaluation protocols: cross-subject (cs), cross-view (cv), and cross-subject-view (csv) show that PTMA achieves state-of-the-art performance on the DAHLIA, IKEA ASM, and Breakfast datasets.
comment: 12 pages, 6 figures, accepted at IEEE Transactions on Multimedia (TMM), in press
♻ ☆ SIG-Chat: Spatial Intent-Guided Conversational Gesture Generation Involving How, When and Where
The accompanying actions and gestures in dialogue are often closely linked to interactions with the environment, such as looking toward the interlocutor or using gestures to point to the described target at appropriate moments. Speech and semantics guide the production of gestures by determining their timing (WHEN) and style (HOW), while the spatial locations of interactive objects dictate their directional execution (WHERE). Existing approaches either rely solely on descriptive language to generate motions or utilize audio to produce non-interactive gestures, thereby lacking the characterization of interactive timing and spatial intent. This significantly limits the applicability of conversational gesture generation, whether in robotics or in the fields of game and animation production. To address this gap, we present a full-stack solution. We first established a unique data collection method to simultaneously capture high-precision human motion and spatial intent. We then developed a generation model driven by audio, language, and spatial data, alongside dedicated metrics for evaluating interaction timing and spatial accuracy. Finally, we deployed the solution on a humanoid robot, enabling rich, context-aware physical interactions.
♻ ☆ Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance ICASSP 2025
Speech Emotion Recognition (SER) systems rely on speech input and emotional labels annotated by humans. However, various emotion databases collect perceptional evaluations in different ways. For instance, the IEMOCAP dataset uses video clips with sounds for annotators to provide their emotional perceptions. However, the most significant English emotion dataset, the MSP-PODCAST, only provides speech for raters to choose the emotional ratings. Nevertheless, using speech as input is the standard approach to training SER systems. Therefore, the open question is the emotional labels elicited by which scenarios are the most effective for training SER systems. We comprehensively compare the effectiveness of SER systems trained with labels elicited by different modality stimuli and evaluate the SER systems on various testing conditions. Also, we introduce an all-inclusive label that combines all labels elicited by various modalities. We show that using labels elicited by voice-only stimuli for training yields better performance on the test set, whereas labels elicited by voice-only stimuli.
comment: 5 pages, 2 figures, 4 tables, acceptance for ICASSP 2025
♻ ☆ Towards Robust and Realible Multimodal Misinformation Recognition with Incomplete Modality
Multimodal Misinformation Recognition has become an urgent task with the emergence of huge multimodal fake content on social media platforms. Previous studies mainly focus on complex feature extraction and fusion to learn discriminative information from multimodal content. However, in real-world applications, multimedia news may naturally lose some information during dissemination, resulting in modality incompleteness, which is detrimental to the generalization and robustness of existing models. To this end, we propose a novel generic and robust multimodal fusion strategy, termed Multi-expert Modality-incomplete Learning Network (MMLNet), which is simple yet effective. It consists of three key steps: (1) Multi-Expert Collaborative Reasoning to compensate for missing modalities by dynamically leveraging complementary information through multiple experts. (2) Incomplete Modality Adapters compensates for the missing information by leveraging the new feature distribution. (3) Modality Missing Learning leveraging an label-aware adaptive weighting strategy to learn a robust representation with contrastive learning. We evaluate MMLNet on three real-world benchmarks across two languages, demonstrating superior performance compared to state-of-the-art methods while maintaining relative simplicity. By ensuring the accuracy of misinformation recognition in incomplete modality scenarios caused by information propagation, MMLNet effectively curbs the spread of malicious misinformation. Code is publicly available at https://github.com/zhyhome/MMLNet.
Computation and Language 167
☆ Are Large Reasoning Models Interruptible?
Large Reasoning Models (LRMs) excel at complex reasoning but are traditionally evaluated in static, "frozen world" settings: model responses are assumed to be instantaneous, and the context of a request is presumed to be immutable over the duration of the response. While generally true for short-term tasks, the "frozen world" assumption breaks down in modern reasoning tasks such as assistive programming, where models may take hours to think through problems and code may change dramatically from the time the model starts thinking to the model's final output. In this work, we challenge the frozen world assumption and evaluate LRM robustness under two realistic dynamic scenarios: interruptions, which test the quality of the model's partial outputs on a limited budget, and dynamic context, which tests model adaptation to in-flight changes. Across mathematics and programming benchmarks that require long-form reasoning, static evaluations consistently overestimate robustness: even state-of-the-art LRMs, which achieve high accuracy in static settings, can fail unpredictably when interrupted or exposed to changing context, with performance dropping by up to 60% when updates are introduced late in the reasoning process. Our analysis further reveals several novel failure modes, including reasoning leakage, where models fold the reasoning into their final answer when interrupted; panic, where under time pressure models abandon reasoning entirely and return incorrect answers; and self-doubt, where performance degrades while incorporating updated information.
comment: Project Page: https://dynamic-lm.github.io/
☆ Demystifying Reinforcement Learning in Agentic Reasoning
Recently, the emergence of agentic RL has showcased that RL could also effectively improve the agentic reasoning ability of LLMs, yet the key design principles and optimal practices remain unclear. In this work, we conduct a comprehensive and systematic investigation to demystify reinforcement learning in agentic reasoning from three key perspectives: data, algorithm, and reasoning mode. We highlight our key insights: (i) Replacing stitched synthetic trajectories with real end-to-end tool-use trajectories yields a far stronger SFT initialization; high-diversity, model-aware datasets sustain exploration and markedly improve RL performance. (ii) Exploration-friendly techniques are crucial for agentic RL, such as clip higher, overlong reward shaping, and maintaining adequate policy entropy could improve the training efficiency. (iii) A deliberative strategy with fewer tool calls outperforms frequent tool calls or verbose self-reasoning, improving tool efficiency and final accuracy. Together, these simple practices consistently enhance agentic reasoning and training efficiency, achieving strong results on challenging benchmarks with smaller models, and establishing a practical baseline for future agentic RL research. Beyond these empirical insights, we further contribute a high-quality, real end-to-end agentic SFT dataset along with a high-quality RL dataset, and demonstrate the effectiveness of our insights in boosting the agentic reasoning ability of LLMs across four challenging benchmarks, including AIME2024/AIME2025, GPQA-Diamond, and LiveCodeBench-v6. With our recipes, 4B-sized models could also achieve superior agentic reasoning performance compared to 32B-sized models. Code and models: https://github.com/Gen-Verse/Open-AgentRL
comment: Code and models: https://github.com/Gen-Verse/Open-AgentRL
QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs
We propose QeRL, a Quantization-enhanced Reinforcement Learning framework for large language models (LLMs). While RL is essential for LLMs' reasoning capabilities, it is resource-intensive, requiring substantial GPU memory and long rollout durations. QeRL addresses these issues by combining NVFP4 quantization with Low-Rank Adaptation (LoRA), accelerating rollout phase of RL while reducing memory overhead. Beyond efficiency, our findings show that quantization noise increases policy entropy, enhancing exploration, and enabling the discovery of better strategies during RL. To further optimize exploration, QeRL introduces an Adaptive Quantization Noise (AQN) mechanism, which dynamically adjusts noise during training. Experiments demonstrate that QeRL delivers over 1.5 times speedup in the rollout phase. Moreover, this is the first framework to enable RL training of a 32B LLM on a single H100 80GB GPU, while delivering overall speedups for RL training. It also achieves faster reward growth and higher final accuracy than 16-bit LoRA and QLoRA, while matching the performance of full-parameter fine-tuning on mathematical benchmarks such as GSM8K (90.8%) and MATH 500 (77.4%) in the 7B model. These results establish QeRL as an efficient and effective framework for RL training in LLMs.
comment: Code is available at https://github.com/NVlabs/QeRL
☆ When Agents Trade: Live Multi-Market Trading Benchmark for LLM Agents
Although Large Language Model (LLM)-based agents are increasingly used in financial trading, it remains unclear whether they can reason and adapt in live markets, as most studies test models instead of agents, cover limited periods and assets, and rely on unverified data. To address these gaps, we introduce Agent Market Arena (AMA), the first lifelong, real-time benchmark for evaluating LLM-based trading agents across multiple markets. AMA integrates verified trading data, expert-checked news, and diverse agent architectures within a unified trading framework, enabling fair and continuous comparison under real conditions. It implements four agents, including InvestorAgent as a single-agent baseline, TradeAgent and HedgeFundAgent with different risk styles, and DeepFundAgent with memory-based reasoning, and evaluates them across GPT-4o, GPT-4.1, Claude-3.5-haiku, Claude-sonnet-4, and Gemini-2.0-flash. Live experiments on both cryptocurrency and stock markets demonstrate that agent frameworks display markedly distinct behavioral patterns, spanning from aggressive risk-taking to conservative decision-making, whereas model backbones contribute less to outcome variation. AMA thus establishes a foundation for rigorous, reproducible, and continuously evolving evaluation of financial reasoning and trading intelligence in LLM-based agents.
☆ Scaling Language-Centric Omnimodal Representation Learning NeurIPS 2025
Recent multimodal embedding approaches leveraging multimodal large language models (MLLMs) fine-tuned with contrastive learning (CL) have shown promising results, yet the underlying reasons behind their superiority remain underexplored. This work argues that a crucial advantage of MLLM-based approaches stems from implicit cross-modal alignment achieved during generative pretraining, where the language decoder learns to exploit multimodal signals within a shared representation space for generating unimodal outputs. Through analysis of anisotropy and kernel similarity structure, we empirically confirm that latent alignment emerges within MLLM representations, allowing CL to serve as a lightweight refinement stage. Leveraging this insight, we propose a Language-Centric Omnimodal Embedding framework, termed LCO-Emb. Extensive experiments across diverse backbones and benchmarks demonstrate its effectiveness, achieving state-of-the-art performance across modalities. Furthermore, we identify a Generation-Representation Scaling Law (GRSL), showing that the representational capabilities gained through contrastive refinement scales positively with the MLLM's generative capabilities. This suggests that improving generative abilities evolves as an effective paradigm for enhancing representation quality. We provide a theoretical explanation of GRSL, which formally links the MLLM's generative quality to the upper bound on its representation performance, and validate it on a challenging, low-resource visual-document retrieval task, showing that continual generative pretraining before CL can further enhance the potential of a model's embedding capabilities. Codes, models, and resources are available at https://github.com/LCO-Embedding/LCO-Embedding.
comment: NeurIPS 2025
☆ Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
A key challenge in applying reinforcement learning (RL) to diffusion large language models (dLLMs) lies in the intractability of their likelihood functions, which are essential for the RL objective, necessitating corresponding approximation in each training step. While existing methods approximate the log-likelihoods by their evidence lower bounds (ELBOs) via customized Monte Carlo (MC) sampling, the forward computational graphs of all MC samples need to be retained for the gradient computation of non-linear terms in the RL objective, resulting in significant memory overhead. This constraint restricts feasible sample sizes, leading to imprecise likelihood approximations and ultimately distorting the RL objective. To overcome this limitation, we propose \emph{Boundary-Guided Policy Optimization} (BGPO), a memory-efficient RL algorithm that maximizes a specially constructed lower bound of the ELBO-based objective. This lower bound is carefully designed to satisfy two key properties: (1) Linearity: it is formulated in a linear sum where each term depends only on a single MC sample, thereby enabling gradient accumulation across samples and ensuring constant memory usage; (2) Equivalence: Both the value and gradient of this lower bound are equal to those of the ELBO-based objective in on-policy training, making it also an effective approximation for the original RL objective. These properties allow BGPO to adopt a large MC sample size, resulting in more accurate likelihood approximations and improved RL objective estimation, which in turn leads to enhanced performance. Experiments show that BGPO significantly outperforms previous RL algorithms for dLLMs in math problem solving, code generation, and planning tasks.
☆ FinVet: A Collaborative Framework of RAG and External Fact-Checking Agents for Financial Misinformation Detection
Financial markets face growing threats from misinformation that can trigger billions in losses in minutes. Most existing approaches lack transparency in their decision-making and provide limited attribution to credible sources. We introduce FinVet, a novel multi-agent framework that integrates two Retrieval-Augmented Generation (RAG) pipelines with external fact-checking through a confidence-weighted voting mechanism. FinVet employs adaptive three-tier processing that dynamically adjusts verification strategies based on retrieval confidence, from direct metadata extraction to hybrid reasoning to full model-based analysis. Unlike existing methods, FinVet provides evidence-backed verdicts, source attribution, confidence scores, and explicit uncertainty flags when evidence is insufficient. Experimental evaluation on the FinFact dataset shows that FinVet achieves an F1 score of 0.85, which is a 10.4% improvement over the best individual pipeline (fact-check pipeline) and 37% improvement over standalone RAG approaches.
☆ ACADREASON: Exploring the Limits of Reasoning Models with Academic Research Problems
In recent years, the research focus of large language models (LLMs) and agents has shifted increasingly from demonstrating novel capabilities to complex reasoning and tackling challenging tasks. However, existing evaluations focus mainly on math/code contests or general tasks, while existing multi-domain academic benchmarks lack sufficient reasoning depth, leaving the field without a rigorous benchmark for high-level reasoning. To fill this gap, we introduce the Acadreason benchmark, designed to evaluate the ability of LLMs and agents to acquire and reason over academic knowledge. It consists of 50 expert-annotated academic problems across five high-reasoning domains, including computer science, economics, law, mathematics, and philosophy. All questions are sourced from top-tier publications in recent years and undergo rigorous annotation and quality control to ensure they are both challenging and answerable. We conduct systematic evaluations of over 10 mainstream LLMs and agents. The results show that most LLMs scored below 20 points, with even the cutting-edge GPT-5 achieving only 16 points. While agents achieved higher scores, none exceeded 40 points. This demonstrates the current capability gap between LLMs and agents in super-intelligent academic research tasks and highlights the challenges of Acadreason.
☆ Enhancing Long Chain-of-Thought Reasoning through Multi-Path Plan Aggregation
Inference-time scaling enhances the reasoning ability of a language model (LM) by extending its chain-of-thought (CoT). However, existing approaches typically generate the entire reasoning chain in a single forward pass, which often leads to CoT derailment, i.e., the reasoning trajectory drifting off course due to compounding errors. This problem is particularly severe for smaller LMs with long CoTs due to their limited capacity. To address this, we analyze raw long CoTs and uncover a reasoning hierarchy consisting of planning and execution steps. Our analysis reveals that most reasoning errors stem from incorrect planning. Motivated by this observation, we propose Multi-Path Plan Aggregation (MPPA), a framework that augments single-pass reasoning with plan exploration and aggregation. Following a variable interval schedule based on the token position, MPPA generates multiple candidate plans and aggregates them into a refined planning step. To maintain efficiency, we adopt a minimal design in which the base LM serves as the primary policy, while a lightweight LoRA module implements the plan aggregation policy. We further observe that outcome-reward RL is inefficient for long trajectories (e.g., exceeding 4K tokens). To overcome this, we introduce online Step-DPO, a process-level preference optimization scheme that leverages Twisted Sequential Monte Carlo (TSMC) to provide scalable stepwise supervision using small LMs. This yields more efficient training, improved stability, and higher accuracy. Extensive experiments on challenging math, science, and logical reasoning benchmarks demonstrate that, with only 10% SFT data and 5% of preference pairs, our method outperforms both the DeepSeek-R1 distillation baseline and the outcome-reward RL baseline across multiple base models and tasks.
☆ StoryBox: Collaborative Multi-Agent Simulation for Hybrid Bottom-Up Long-Form Story Generation Using Large Language Models
Human writers often begin their stories with an overarching mental scene, where they envision the interactions between characters and their environment. Inspired by this creative process, we propose a novel approach to long-form story generation, termed hybrid bottom-up long-form story generation, using multi-agent simulations. In our method, agents interact within a dynamic sandbox environment, where their behaviors and interactions with one another and the environment generate emergent events. These events form the foundation for the story, enabling organic character development and plot progression. Unlike traditional top-down approaches that impose rigid structures, our hybrid bottom-up approach allows for the natural unfolding of events, fostering more spontaneous and engaging storytelling. The system is capable of generating stories exceeding 10,000 words while maintaining coherence and consistency, addressing some of the key challenges faced by current story generation models. We achieve state-of-the-art performance across several metrics. This approach offers a scalable and innovative solution for creating dynamic, immersive long-form stories that evolve organically from agent-driven interactions.
comment: Project: https://storyboxproject.github.io
☆ LLM-Oriented Token-Adaptive Knowledge Distillation
Knowledge distillation (KD) is a key technique for compressing large-scale language models (LLMs), yet prevailing logit-based methods typically employ static strategies that are misaligned with the dynamic learning process of student models. These methods typically treat all tokens indiscriminately and apply a single, fixed temperature, resulting in suboptimal knowledge transfer. To address these limitations, we propose LLM-Oriented Token-Adaptive Knowledge Distillation (AdaKD), a novel framework that adapts the distillation process to the real-time learning state of each token. AdaKD consists of two synergistic modules driven by a unified token difficulty metric. First, our Loss-Driven Adaptive Token Focusing (LATF) module dynamically adjusts the distillation focus by monitoring the student's learning stability, concentrating computational resources on the most valuable tokens at each training phase. Second, we introduce Inverse Difficulty Temperature Scaling (IDTS), a counterintuitive yet effective token-level temperature strategy. It employs low temperatures for difficult tokens for targeted error correction, and high temperatures for easy tokens to encourage students to learn from the teacher's complete and smooth output distribution, thereby enhancing generalization. As a plug-and-play framework, AdaKD can consistently improve the performance of various distillation methods on multiple model architectures and benchmarks.
comment: 15 pages, 4 figures
☆ Deconstructing Attention: Investigating Design Principles for Effective Language Modeling
The success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.
☆ SemCSE-Multi: Multifaceted and Decodable Embeddings for Aspect-Specific and Interpretable Scientific Domain Mapping
We propose SemCSE-Multi, a novel unsupervised framework for generating multifaceted embeddings of scientific abstracts, evaluated in the domains of invasion biology and medicine. These embeddings capture distinct, individually specifiable aspects in isolation, thus enabling fine-grained and controllable similarity assessments as well as adaptive, user-driven visualizations of scientific domains. Our approach relies on an unsupervised procedure that produces aspect-specific summarizing sentences and trains embedding models to map semantically related summaries to nearby positions in the embedding space. We then distill these aspect-specific embedding capabilities into a unified embedding model that directly predicts multiple aspect embeddings from a scientific abstract in a single, efficient forward pass. In addition, we introduce an embedding decoding pipeline that decodes embeddings back into natural language descriptions of their associated aspects. Notably, we show that this decoding remains effective even for unoccupied regions in low-dimensional visualizations, thus offering vastly improved interpretability in user-centric settings.
☆ MeTA-LoRA: Data-Efficient Multi-Task Fine-Tuning for Large Language Models
Low-Rank Adaptation (LoRA) has emerged as one of the most widely used parameter-efficient fine-tuning (PEFT) methods for adapting large language models (LLMs) to downstream tasks. While highly effective in single-task settings, it struggles to efficiently leverage inter-task knowledge in complex multi-task learning scenarios, often requiring substantial task-specific data to achieve optimal performance. To address this limitation, we introduce MeTA-LoRA, a two-stage optimization framework that significantly improves data efficiency in multi-task adaptation. In the first stage, task-specific LoRA adapters are learned using only a few samples from each involved dataset, enabling rapid adaptation without large-scale supervision. In the second stage, the shared LoRA adapter is updated by aggregating gradients from multiple tasks to promote knowledge transfer across tasks, further reducing data usage by leveraging common patterns. In both multi-task learning and multilingual learning scenarios, our method matches or surpasses the performance of traditional full-data LoRA fine-tuning approaches, while using significantly less task-specific data.
☆ REGENT: Relevance-Guided Attention for Entity-Aware Multi-Vector Neural Re-Ranking SIGIR
Current neural re-rankers often struggle with complex information needs and long, content-rich documents. The fundamental issue is not computational--it is intelligent content selection: identifying what matters in lengthy, multi-faceted texts. While humans naturally anchor their understanding around key entities and concepts, neural models process text within rigid token windows, treating all interactions as equally important and missing critical semantic signals. We introduce REGENT, a neural re-ranking model that mimics human-like understanding by using entities as a "semantic skeleton" to guide attention. REGENT integrates relevance guidance directly into the attention mechanism, combining fine-grained lexical matching with high-level semantic reasoning. This relevance-guided attention enables the model to focus on conceptually important content while maintaining sensitivity to precise term matches. REGENT achieves new state-of-the-art performance in three challenging datasets, providing up to 108% improvement over BM25 and consistently outperforming strong baselines including ColBERT and RankVicuna. To our knowledge, this is the first work to successfully integrate entity semantics directly into neural attention, establishing a new paradigm for entity-aware information retrieval.
comment: To be published in: Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (SIGIR-AP 2025)
☆ QDER: Query-Specific Document and Entity Representations for Multi-Vector Document Re-Ranking SIGIR
Neural IR has advanced through two distinct paths: entity-oriented approaches leveraging knowledge graphs and multi-vector models capturing fine-grained semantics. We introduce QDER, a neural re-ranking model that unifies these approaches by integrating knowledge graph semantics into a multi-vector model. QDER's key innovation lies in its modeling of query-document relationships: rather than computing similarity scores on aggregated embeddings, we maintain individual token and entity representations throughout the ranking process, performing aggregation only at the final scoring stage - an approach we call "late aggregation." We first transform these fine-grained representations through learned attention patterns, then apply carefully chosen mathematical operations for precise matches. Experiments across five standard benchmarks show that QDER achieves significant performance gains, with improvements of 36% in nDCG@20 over the strongest baseline on TREC Robust 2004 and similar improvements on other datasets. QDER particularly excels on difficult queries, achieving an nDCG@20 of 0.70 where traditional approaches fail completely (nDCG@20 = 0.0), setting a foundation for future work in entity-aware retrieval.
comment: Published in: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2025)
Survey Response Generation: Generating Closed-Ended Survey Responses In-Silico with Large Language Models
Many in-silico simulations of human survey responses with large language models (LLMs) focus on generating closed-ended survey responses, whereas LLMs are typically trained to generate open-ended text instead. Previous research has used a diverse range of methods for generating closed-ended survey responses with LLMs, and a standard practice remains to be identified. In this paper, we systematically investigate the impact that various Survey Response Generation Methods have on predicted survey responses. We present the results of 32 mio. simulated survey responses across 8 Survey Response Generation Methods, 4 political attitude surveys, and 10 open-weight language models. We find significant differences between the Survey Response Generation Methods in both individual-level and subpopulation-level alignment. Our results show that Restricted Generation Methods perform best overall, and that reasoning output does not consistently improve alignment. Our work underlines the significant impact that Survey Response Generation Methods have on simulated survey responses, and we develop practical recommendations on the application of Survey Response Generation Methods.
☆ LLMAtKGE: Large Language Models as Explainable Attackers against Knowledge Graph Embeddings
Adversarial attacks on knowledge graph embeddings (KGE) aim to disrupt the model's ability of link prediction by removing or inserting triples. A recent black-box method has attempted to incorporate textual and structural information to enhance attack performance. However, it is unable to generate human-readable explanations, and exhibits poor generalizability. In the past few years, large language models (LLMs) have demonstrated powerful capabilities in text comprehension, generation, and reasoning. In this paper, we propose LLMAtKGE, a novel LLM-based framework that selects attack targets and generates human-readable explanations. To provide the LLM with sufficient factual context under limited input constraints, we design a structured prompting scheme that explicitly formulates the attack as multiple-choice questions while incorporating KG factual evidence. To address the context-window limitation and hesitation issues, we introduce semantics-based and centrality-based filters, which compress the candidate set while preserving high recall of attack-relevant information. Furthermore, to efficiently integrate both semantic and structural information into the filter, we precompute high-order adjacency and fine-tune the LLM with a triple classification task to enhance filtering performance. Experiments on two widely used knowledge graph datasets demonstrate that our attack outperforms the strongest black-box baselines and provides explanations via reasoning, and showing competitive performance compared with white-box methods. Comprehensive ablation and case studies further validate its capability to generate explanations.
comment: 13 pages
☆ Bag of Tricks for Subverting Reasoning-based Safety Guardrails
Recent reasoning-based safety guardrails for Large Reasoning Models (LRMs), such as deliberative alignment, have shown strong defense against jailbreak attacks. By leveraging LRMs' reasoning ability, these guardrails help the models to assess the safety of user inputs before generating final responses. The powerful reasoning ability can analyze the intention of the input query and will refuse to assist once it detects the harmful intent hidden by the jailbreak methods. Such guardrails have shown a significant boost in defense, such as the near-perfect refusal rates on the open-source gpt-oss series. Unfortunately, we find that these powerful reasoning-based guardrails can be extremely vulnerable to subtle manipulation of the input prompts, and once hijacked, can lead to even more harmful results. Specifically, we first uncover a surprisingly fragile aspect of these guardrails: simply adding a few template tokens to the input prompt can successfully bypass the seemingly powerful guardrails and lead to explicit and harmful responses. To explore further, we introduce a bag of jailbreak methods that subvert the reasoning-based guardrails. Our attacks span white-, gray-, and black-box settings and range from effortless template manipulations to fully automated optimization. Along with the potential for scalable implementation, these methods also achieve alarmingly high attack success rates (e.g., exceeding 90% across 5 different benchmarks on gpt-oss series on both local host models and online API services). Evaluations across various leading open-source LRMs confirm that these vulnerabilities are systemic, underscoring the urgent need for stronger alignment techniques for open-sourced LRMs to prevent malicious misuse. Code is open-sourced at https://chenxshuo.github.io/bag-of-tricks.
comment: OpenAI Red-teaming Challenge Winner and Oral Presentation
☆ Culturally-Aware Conversations: A Framework & Benchmark for LLMs EMNLP
Existing benchmarks that measure cultural adaptation in LLMs are misaligned with the actual challenges these models face when interacting with users from diverse cultural backgrounds. In this work, we introduce the first framework and benchmark designed to evaluate LLMs in realistic, multicultural conversational settings. Grounded in sociocultural theory, our framework formalizes how linguistic style - a key element of cultural communication - is shaped by situational, relational, and cultural context. We construct a benchmark dataset based on this framework, annotated by culturally diverse raters, and propose a new set of desiderata for cross-cultural evaluation in NLP: conversational framing, stylistic sensitivity, and subjective correctness. We evaluate today's top LLMs on our benchmark and show that these models struggle with cultural adaptation in a conversational setting.
comment: To appear at the 4th HCI + NLP Workshop @ EMNLP
☆ Invisible Languages of the LLM Universe
Large Language Models are trained on massive multilingual corpora, yet this abundance masks a profound crisis: of the world's 7,613 living languages, approximately 2,000 languages with millions of speakers remain effectively invisible in digital ecosystems. We propose a critical framework connecting empirical measurements of language vitality (real world demographic strength) and digitality (online presence) with postcolonial theory and epistemic injustice to explain why linguistic inequality in AI systems is not incidental but structural. Analyzing data across all documented human languages, we identify four categories: Strongholds (33%, high vitality and digitality), Digital Echoes (6%, high digitality despite declining vitality), Fading Voices (36%, low on both dimensions), and critically, Invisible Giants (27%, high vitality but near-zero digitality) - languages spoken by millions yet absent from the LLM universe. We demonstrate that these patterns reflect continuities from colonial-era linguistic hierarchies to contemporary AI development, constituting what we term digital epistemic injustice. Our analysis reveals that English dominance in AI is not a technical necessity but an artifact of power structures that systematically exclude marginalized linguistic knowledge. We conclude with implications for decolonizing language technology and democratizing access to AI benefits.
☆ Information-Preserving Reformulation of Reasoning Traces for Antidistillation
Recent advances in Large Language Models (LLMs) show that extending the length of reasoning chains significantly improves performance on complex tasks. While revealing these reasoning traces helps users better follow, verify, and learn from the model's problem-solving process, it also makes them highly vulnerable to unauthorized distillation. To mitigate this risk, proprietary model providers often adopt aggressive protection strategies, such as replacing detailed reasoning with brief summaries, which deprive users of valuable intermediate information. To address this trade-off, we propose PART, an information-preserving antidistillation reformulation of reasoning traces. Motivated by the difference between how humans understand reasoning traces and how LLMs exploit them for supervised fine-tuning, we design a simple but effective two-step reformulation: removing self-talk behaviors and reordering sub-conclusions. A small auxiliary model is trained to perform this reformulation, incurring minimal computational overhead. Extensive experiments demonstrate that PART consistently disrupts distillation across student models of different sizes and types on various reasoning benchmarks. For instance, when training on reformulated traces, even the performance of a large 32B student model decreases from 54.17 to 46.88 on AIME 2024, corresponding to a 13.5% degradation.
☆ An Encoder-Integrated PhoBERT with Graph Attention for Vietnamese Token-Level Classification SP 2025
We propose a novel neural architecture named TextGraphFuseGAT, which integrates a pretrained transformer encoder (PhoBERT) with Graph Attention Networks for token-level classification tasks. The proposed model constructs a fully connected graph over the token embeddings produced by PhoBERT, enabling the GAT layer to capture rich inter-token dependencies beyond those modeled by sequential context alone. To further enhance contextualization, a Transformer-style self-attention layer is applied on top of the graph-enhanced embeddings. The final token representations are passed through a classification head to perform sequence labeling. We evaluate our approach on three Vietnamese benchmark datasets: PhoNER-COVID19 for named entity recognition in the COVID-19 domain, PhoDisfluency for speech disfluency detection, and VietMed-NER for medical-domain NER. VietMed-NER is the first Vietnamese medical spoken NER dataset, featuring 18 entity types collected from real-world medical speech transcripts and annotated with the BIO tagging scheme. Its specialized vocabulary and domain-specific expressions make it a challenging benchmark for token-level classification models. Experimental results show that our method consistently outperforms strong baselines, including transformer-only and hybrid neural models such as BiLSTM + CNN + CRF, confirming the effectiveness of combining pretrained semantic features with graph-based relational modeling for improved token classification across multiple domains.
comment: 11 pages, 1 figure. Submitted to VLSP 2025 and reviewed
☆ Hallucination Detection via Internal States and Structured Reasoning Consistency in Large Language Models
The detection of sophisticated hallucinations in Large Language Models (LLMs) is hampered by a ``Detection Dilemma'': methods probing internal states (Internal State Probing) excel at identifying factual inconsistencies but fail on logical fallacies, while those verifying externalized reasoning (Chain-of-Thought Verification) show the opposite behavior. This schism creates a task-dependent blind spot: Chain-of-Thought Verification fails on fact-intensive tasks like open-domain QA where reasoning is ungrounded, while Internal State Probing is ineffective on logic-intensive tasks like mathematical reasoning where models are confidently wrong. We resolve this with a unified framework that bridges this critical gap. However, unification is hindered by two fundamental challenges: the Signal Scarcity Barrier, as coarse symbolic reasoning chains lack signals directly comparable to fine-grained internal states, and the Representational Alignment Barrier, a deep-seated mismatch between their underlying semantic spaces. To overcome these, we introduce a multi-path reasoning mechanism to obtain more comparable, fine-grained signals, and a segment-aware temporalized cross-attention module to adaptively fuse these now-aligned representations, pinpointing subtle dissonances. Extensive experiments on three diverse benchmarks and two leading LLMs demonstrate that our framework consistently and significantly outperforms strong baselines. Our code is available: https://github.com/peach918/HalluDet.
☆ ReLook: Vision-Grounded RL with a Multimodal LLM Critic for Agentic Web Coding
While Large Language Models (LLMs) excel at algorithmic code generation, they struggle with front-end development, where correctness is judged on rendered pixels and interaction. We present ReLook, an agentic, vision-grounded reinforcement learning framework that empowers an agent to close a robust generate--diagnose--refine loop by invoking a multimodal LLM (MLLM) as a tool. During training, the agent uses the MLLM-in-the-loop both as a visual critic--scoring code with screenshots--and as a source of actionable, vision-grounded feedback; a strict zero-reward rule for invalid renders anchors renderability and prevents reward hacking. To prevent behavioral collapse, we introduce Forced Optimization, a strict acceptance rule that admits only improving revisions, yielding monotonically better trajectories. At inference, we decouple the critic and run a lightweight, critic-free self-edit cycle, keeping latency comparable to base decoding while retaining most of the gains. Across three widely used benchmarks, ReLook consistently outperforms strong baselines in vision-grounded front-end code generation, highlighting the benefits of agentic perception, visual rewards, and training-inference decoupling.
☆ Investigating Large Language Models' Linguistic Abilities for Text Preprocessing
Text preprocessing is a fundamental component of Natural Language Processing, involving techniques such as stopword removal, stemming, and lemmatization to prepare text as input for further processing and analysis. Despite the context-dependent nature of the above techniques, traditional methods usually ignore contextual information. In this paper, we investigate the idea of using Large Language Models (LLMs) to perform various preprocessing tasks, due to their ability to take context into account without requiring extensive language-specific annotated resources. Through a comprehensive evaluation on web-sourced data, we compare LLM-based preprocessing (specifically stopword removal, lemmatization and stemming) to traditional algorithms across multiple text classification tasks in six European languages. Our analysis indicates that LLMs are capable of replicating traditional stopword removal, lemmatization, and stemming methods with accuracies reaching 97%, 82%, and 74%, respectively. Additionally, we show that ML algorithms trained on texts preprocessed by LLMs achieve an improvement of up to 6% with respect to the $F_1$ measure compared to traditional techniques. Our code, prompts, and results are publicly available at https://github.com/GianCarloMilanese/llm_pipeline_wi-iat.
comment: Accepted in WI-IAT 2025. Pre-camera-ready version
☆ GenCNER: A Generative Framework for Continual Named Entity Recognition IJCNN 2025
Traditional named entity recognition (NER) aims to identify text mentions into pre-defined entity types. Continual Named Entity Recognition (CNER) is introduced since entity categories are continuously increasing in various real-world scenarios. However, existing continual learning (CL) methods for NER face challenges of catastrophic forgetting and semantic shift of non-entity type. In this paper, we propose GenCNER, a simple but effective Generative framework for CNER to mitigate the above drawbacks. Specifically, we skillfully convert the CNER task into sustained entity triplet sequence generation problem and utilize a powerful pre-trained seq2seq model to solve it. Additionally, we design a type-specific confidence-based pseudo labeling strategy along with knowledge distillation (KD) to preserve learned knowledge and alleviate the impact of label noise at the triplet level. Experimental results on two benchmark datasets show that our framework outperforms previous state-of-the-art methods in multiple CNER settings, and achieves the smallest gap compared with non-CL results.
comment: Accepted by IJCNN 2025
☆ Who are you, ChatGPT? Personality and Demographic Style in LLM-Generated Content ECAI2025
Generative large language models (LLMs) have become central to everyday life, producing human-like text across diverse domains. A growing body of research investigates whether these models also exhibit personality- and demographic-like characteristics in their language. In this work, we introduce a novel, data-driven methodology for assessing LLM personality without relying on self-report questionnaires, applying instead automatic personality and gender classifiers to model replies on open-ended questions collected from Reddit. Comparing six widely used models to human-authored responses, we find that LLMs systematically express higher Agreeableness and lower Neuroticism, reflecting cooperative and stable conversational tendencies. Gendered language patterns in model text broadly resemble those of human writers, though with reduced variation, echoing prior findings on automated agents. We contribute a new dataset of human and model responses, along with large-scale comparative analyses, shedding new light on the topic of personality and demographic patterns of generative AI.
comment: ECAI2025 (Identity-Aware AI workshop)
☆ Beyond the Crowd: LLM-Augmented Community Notes for Governing Health Misinformation
Community Notes, the crowd-sourced misinformation governance system on X (formerly Twitter), enables users to flag misleading posts, attach contextual notes, and vote on their helpfulness. However, our analysis of 30.8K health-related notes reveals significant latency, with a median delay of 17.6 hours before the first note receives a helpfulness status. To improve responsiveness during real-world misinformation surges, we propose CrowdNotes+, a unified framework that leverages large language models (LLMs) to augment Community Notes for faster and more reliable health misinformation governance. CrowdNotes+ integrates two complementary modes: (1) evidence-grounded note augmentation and (2) utility-guided note automation, along with a hierarchical three-step evaluation that progressively assesses relevance, correctness, and helpfulness. We instantiate the framework through HealthNotes, a benchmark of 1.2K helpfulness-annotated health notes paired with a fine-tuned helpfulness judge. Experiments on fifteen LLMs reveal an overlooked loophole in current helpfulness evaluation, where stylistic fluency is mistaken for factual accuracy, and demonstrate that our hierarchical evaluation and LLM-augmented generation jointly enhance factual precision and evidence utility. These results point toward a hybrid human-AI governance model that improves both the rigor and timeliness of crowd-sourced fact-checking.
☆ Valid Survey Simulations with Limited Human Data: The Roles of Prompting, Fine-Tuning, and Rectification
Surveys provide valuable insights into public opinion and behavior, but their execution is costly and slow. Large language models (LLMs) have been proposed as a scalable, low-cost substitute for human respondents, but their outputs are often biased and yield invalid estimates. We study the interplay between synthesis methods that use LLMs to generate survey responses and rectification methods that debias population estimates, and explore how human responses are best allocated between them. Using two panel surveys with questions on nutrition, politics, and economics, we find that synthesis alone introduces substantial bias (24-86%), whereas combining it with rectification reduces bias below 5% and increases effective sample size by up to 14%. Overall, we challenge the common practice of using all human responses for fine-tuning, showing that under a fixed budget, allocating most to rectification results in far more effective estimation.
comment: 19 pages, 4 figures, 9 tables
☆ KnowRL: Teaching Language Models to Know What They Know
Truly reliable AI requires more than simply scaling up knowledge; it demands the ability to know what it knows and when it does not. Yet recent research shows that even the best LLMs misjudge their own competence in more than one in five cases, making any response born of such internal uncertainty impossible to fully trust. Inspired by self-improvement reinforcement learning techniques that require minimal data, we present a simple but powerful framework KnowRL that strengthens a model's internal understanding of its own feasibility boundaries, enabling safer and more responsible behaviour. Our framework combines two components: (i) introspection, where the model generates and classifies tasks it judges feasible or infeasible, and (ii) consensus-based rewarding, where stability of self-knowledge assessment is reinforced through internal agreement. By using internally generated data, this design strengthens consistency in self-knowledge and entirely avoids costly external supervision. In experiments on LLaMA-3.1-8B and Qwen-2.5-7B, KnowRL steadily improved self-knowledge, validated by both intrinsic self-consistency and extrinsic benchmarking. With nothing more than a small seed set and no external supervision, our method drove gains as high as 28% in accuracy and 12% in F1, outperforming baselines in just a few iterations. Our framework essentially unlocks the untapped capacity of LLMs to self-improve their knowledge awareness, opening the door to reliable, more accountable AI and safer deployment in critical applications. Owing to its simplicity and independence from external effort, we encourage applying this reliability-enhancing process to all future models.
comment: 14 pages, 7 figures
☆ DocReward: A Document Reward Model for Structuring and Stylizing
Recent advances in agentic workflows have enabled the automation of tasks such as professional document generation. However, they primarily focus on textual quality, neglecting visual structure and style, which are crucial for readability and engagement. This gap arises mainly from the absence of suitable reward models to guide agentic workflows toward producing documents with stronger structural and stylistic quality. To address this, we propose DocReward, a document reward model that evaluates documents based on their structure and style. We construct a multi-domain dataset DocPair of 117K paired documents, covering 32 domains and 267 document types, each including a high- and low-professionalism document with identical content but different structure and style. This enables the model to evaluate professionalism comprehensively, and in a textual-quality-agnostic way. DocReward is trained using the Bradley-Terry loss to score documents, penalizing predictions that contradict the annotated ranking. To assess the performance of reward models, we create a test dataset containing document bundles ranked by well-educated human evaluators. Notably, DocReward outperforms GPT-4o and GPT-5 in accuracy by 30.6 and 19.4 percentage points, respectively, demonstrating its superiority over baselines. In an extrinsic evaluation of document generation, DocReward achieves a significantly higher win rate of 60.8%, compared to GPT-5's 37.7% win rate, demonstrating its utility in guiding generation agents toward producing human-preferred documents.
☆ Beyond Survival: Evaluating LLMs in Social Deduction Games with Human-Aligned Strategies
Social deduction games like Werewolf combine language, reasoning, and strategy, providing a testbed for studying natural language and social intelligence. However, most studies reduce the game to LLM-based self-play, yielding templated utterances and anecdotal cases that overlook the richness of social gameplay. Evaluation further relies on coarse metrics such as survival time or subjective scoring due to the lack of quality reference data. To address these gaps, we curate a high-quality, human-verified multimodal Werewolf dataset containing over 100 hours of video, 32.4M utterance tokens, and 15 rule variants. Based on this dataset, we propose a novel strategy-alignment evaluation that leverages the winning faction's strategies as ground truth in two stages: 1) Speech evaluation, formulated as multiple-choice-style tasks that assess whether the model can adopt appropriate stances across five dimensions of social ability; and 2) Decision evaluation, which assesses the model's voting choices and opponent-role inferences. This framework enables a fine-grained evaluation of models' linguistic and reasoning capabilities, while capturing their ability to generate strategically coherent gameplay. Our experiments show that state-of-the-art LLMs show diverse performance, with roughly half remain below 0.50, revealing clear gaps in deception and counterfactual reasoning. We hope our dataset further inspires research on language, reasoning, and strategy in multi-agent interaction.
comment: 34 pages, 32figures
☆ Early Detection and Reduction of Memorisation for Domain Adaptation and Instruction Tuning ACL
Although large language models excel across many tasks, they can memorise training data and thereby expose private or copyrighted text. Most defences target the pre-training stage, leaving memorisation during fine-tuning, especially for domain adaptation and instruction tuning, poorly understood. We fine-tune Pythia, Llama3, and Mistral models spanning 1.4B-70B parameters on common evaluation datasets and track verbatim memorisation throughout training. We find that memorisation increases dramatically in the first few epochs, often significantly before either validation perplexity or evaluation performance is optimised. We use a simple but effective n-gram memorisation score which reliably precedes verbatim memorisation; using it as an early-stopping criterion mitigates memorisation with minimal performance loss. Further, we introduce an n-gram-aware loss regulariser and show that it reduces memorisation across all model families tested by up to 40% while minimising evaluation performance trade-offs when compared to an existing memorisation mitigation strategy. These results yield practical, scalable insights into memorisation dynamics during language model fine-tuning.
comment: Accepted to Transactions of the ACL (TACL), 2025. 15 pages, 6 figures, 3 tables
☆ Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers
Reinforcement learning (RL) has emerged as a crucial approach for enhancing the capabilities of large language models. However, in Mixture-of-Experts (MoE) models, the routing mechanism often introduces instability, even leading to catastrophic RL training collapse. We analyze the training-inference consistency of MoE models and identify a notable discrepancy in routing behaviors between the two phases. Moreover, even under identical conditions, the routing framework can yield divergent expert selections across repeated forward passes. To address this foundational inconsistency, we propose Rollout Routing Replay (R3), a method that records routing distributions from the inference engine and replays them during training. R3 significantly reduces training-inference policy KL divergence and mitigates extreme discrepancies without compromising training speed. Extensive experiments on various settings confirm that R3 succeeds in stabilizing RL training, preventing collapse and outperforming methods such as GSPO and TIS. We believe this work can offer a new solution for stabilizing RL in MoE models.
☆ LLM-Specific Utility: A New Perspective for Retrieval-Augmented Generation
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external knowledge. While traditional retrieval focuses on relevance, RAG's effectiveness depends on the utility of retrieved passages, i.e., the usefulness in facilitating the generation of an accurate and comprehensive answer. Existing studies often treat utility as a generic attribute, ignoring the fact that different LLMs may benefit differently from the same passage due to variations in internal knowledge and comprehension ability. In this work, we introduce and systematically investigate the notion of LLM-specific utility. Through large-scale experiments across multiple datasets and LLMs, we demonstrate that human-annotated passages are not optimal for LLMs and that ground-truth utilitarian passages are not transferable across different LLMs. These findings highlight the necessity of adopting the LLM-specific utility in RAG research. Our findings indicate that some human-annotated passages are not ground-truth utilitarian passages for specific LLMs, partially due to the varying readability of queries and passages for LLMs, a tendency for which perplexity is a key metric. Based on these findings, we propose a benchmarking procedure for LLM-specific utility judgments. We evaluate existing utility judgment methods on six datasets and find that while verbalized methods using pseudo-answers perform robustly, LLMs struggle to assess utility effectively-failing to reject all passages for known queries and to select truly useful ones for unknown queries.
comment: 13 pages, 9 figures
☆ Diffusion-Link: Diffusion Probabilistic Model for Bridging the Audio-Text Modality Gap ICASSP 2026
Contrastive audio-language pretraining yields powerful joint representations, yet a persistent audio-text modality gap limits the benefits of coupling multimodal encoders with large language models (LLMs). We present Diffusion-Link, a diffusion-based modality-bridging module that generatively maps audio embeddings into the text-embedding distribution. The module is trained at the output embedding from the frozen multimodal encoder and implemented as a lightweight network with three residual MLP blocks. To assess the effect of Diffusion-Link on multimodal encoder-LLM coupling, we evaluate on Automatic Audio Captioning (AAC); to our knowledge, this is the first application of diffusion-based modality bridging to AAC. We report two results. (1) Modality-gap analysis: on similarity and geometric criteria, Diffusion-Link reduces the modality gap the most among prior diffusion-based methods and shows a collective migration of audio embeddings toward the text distribution. (2) Downstream AAC: attaching Diffusion-Link to the same multimodal LLM baseline achieves state-of-the-art on AudioCaps in both zero-shot and fully supervised captioning without external knowledge, with relative gains up to 52.5% and 7.5%, respectively. These findings show that closing the modality gap is pivotal for effective coupling between multimodal encoders and LLMs, and diffusion-based modality bridging offers a promising direction beyond knowledge-retrieval-centric designs. Code will be released upon acceptance https://github.com/DevKiHyun/Diffusion-Link
comment: 5 pages. Submitted to IEEE ICASSP 2026
☆ Do LLMs "Feel"? Emotion Circuits Discovery and Control
As the demand for emotional intelligence in large language models (LLMs) grows, a key challenge lies in understanding the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text. This study addresses three core questions: (1) Do LLMs contain context-agnostic mechanisms shaping emotional expression? (2) What form do these mechanisms take? (3) Can they be harnessed for universal emotion control? We first construct a controlled dataset, SEV (Scenario-Event with Valence), to elicit comparable internal states across emotions. Subsequently, we extract context-agnostic emotion directions that reveal consistent, cross-context encoding of emotion (Q1). We identify neurons and attention heads that locally implement emotional computation through analytical decomposition and causal analysis, and validate their causal roles via ablation and enhancement interventions. Next, we quantify each sublayer's causal influence on the model's final emotion representation and integrate the identified local components into coherent global emotion circuits that drive emotional expression (Q2). Directly modulating these circuits achieves 99.65% emotion-expression accuracy on the test set, surpassing prompting- and steering-based methods (Q3). To our knowledge, this is the first systematic study to uncover and validate emotion circuits in LLMs, offering new insights into interpretability and controllable emotional intelligence.
comment: 19 pages, 8 figures, 8 tables. Code and dataset available at https://github.com/Aurora-cx/EmotionCircuits-LLM
☆ Template-Based Text-to-Image Alignment for Language Accessibility: A Study on Visualizing Text Simplifications
Individuals with intellectual disabilities often have difficulties in comprehending complex texts. While many text-to-image models prioritize aesthetics over accessibility, it is not clear how visual illustrations relate to text simplifications (TS) generated from them. This paper presents a structured vision-language model (VLM) prompting framework for generating accessible images from simplified texts. We designed five prompt templates, i.e., Basic Object Focus, Contextual Scene, Educational Layout, Multi-Level Detail, and Grid Layout, each following distinct spatial arrangements while adhering to accessibility constraints such as object count limits, spatial separation, and content restrictions. Using 400 sentence-level simplifications from four established TS datasets (OneStopEnglish, SimPA, Wikipedia, and ASSET), we conducted a two-phase evaluation: Phase 1 assessed prompt template effectiveness with CLIPScores, and Phase 2 involved human annotation of generated images across ten visual styles by four accessibility experts. Results show that the Basic Object Focus prompt template achieved the highest semantic alignment, indicating that visual minimalism enhances language accessibility. Expert evaluation further identified Retro style as the most accessible and Wikipedia as the most effective data source. Inter-annotator agreement varied across dimensions, with Text Simplicity showing strong reliability and Image Quality proving more subjective. Overall, our framework offers practical guidelines for accessible content generation and underscores the importance of structured prompting in AI-generated visual accessibility tools.
☆ FOSSIL: Harnessing Feedback on Suboptimal Samples for Data-Efficient Generalisation with Imitation Learning for Embodied Vision-and-Language Tasks EMNLP 2025
Current approaches to embodied AI tend to learn policies from expert demonstrations. However, without a mechanism to evaluate the quality of demonstrated actions, they are limited to learning from optimal behaviour, or they risk replicating errors and inefficiencies. While reinforcement learning offers one alternative, the associated exploration typically results in sacrificing data efficiency. This work explores how agents trained with imitation learning can learn robust representations from both optimal and suboptimal demonstrations when given access to constructive language feedback as a means to contextualise different modes of behaviour. We directly provide language feedback embeddings as part of the input sequence into a Transformer-based policy, and optionally complement the traditional next action prediction objective with auxiliary self-supervised learning objectives for feedback prediction. We test our approach on a range of embodied Vision-and-Language tasks in our custom BabyAI-XGen environment and show significant improvements in agents' compositional generalisation abilities and robustness, suggesting that our data-efficient method allows models to successfully convert suboptimal behaviour into learning opportunities. Overall, our results suggest that language feedback is a competitive and intuitive alternative to intermediate scalar rewards for language-specified embodied tasks.
comment: EMNLP 2025 Findings
☆ Are Large Language Models Effective Knowledge Graph Constructors?
Knowledge graphs (KGs) are vital for knowledge-intensive tasks and have shown promise in reducing hallucinations in large language models (LLMs). However, constructing high-quality KGs remains difficult, requiring accurate information extraction and structured representations that support interpretability and downstream utility. Existing LLM-based approaches often focus narrowly on entity and relation extraction, limiting coverage to sentence-level contexts or relying on predefined schemas. We propose a hierarchical extraction framework that organizes information at multiple levels, enabling the creation of semantically rich and well-structured KGs. Using state-of-the-art LLMs, we extract and construct knowledge graphs and evaluate them comprehensively from both structural and semantic perspectives. Our results highlight the strengths and shortcomings of current LLMs in KG construction and identify key challenges for future work. To advance research in this area, we also release a curated dataset of LLM-generated KGs derived from research papers on children's mental well-being. This resource aims to foster more transparent, reliable, and impactful applications in high-stakes domains such as healthcare.
☆ Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs
Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across three datasets, three frontier models produce broadly misaligned responses at rates between 2% and 17% given 64 narrow in-context examples, and up to 58% with 256 examples. We also examine mechanisms of EM by eliciting step-by-step reasoning (while leaving in-context examples unchanged). Manual analysis of the resulting chain-of-thought shows that 67.5% of misaligned traces explicitly rationalize harmful outputs by adopting a reckless or dangerous ''persona'', echoing prior results on finetuning-induced EM.
☆ ENIGMA: The Geometry of Reasoning and Alignment in Large-Language Models
We present Entropic Mutual-Information Geometry Large-Language Model Alignment (ENIGMA), a novel approach to Large-Language Model (LLM) training that jointly improves reasoning, alignment and robustness by treating an organisation's policies/principles as directions to move on a model's information manifold. Our single-loop trainer combines Group-Relative Policy Optimisation (GRPO), an on-policy, critic-free RL method with Chain-of-Thought (CoT)-format only rewards; a Self-Supervised Alignment with Mutual Information (SAMI)-style symmetric InfoNCE auxiliary; and an entropic Sinkhorn optimal-transport regulariser on hidden-state distributions to bound geometry drift. We also introduce infoNCE metrics that specialise to a standard MI lower bound under matched negatives to measure how strongly a model's CoT encodes these policies. These metrics include a Sufficiency Index (SI) that enables the selection and creation of principles that maximise downstream performance prior to training. In our experiments using small (1B) LLMs, high-SI principles predict steadier training dynamics and improved benchmark performance over GRPO ablations. Our information-geometry analysis of trained models validates desirable structural change in the manifold. These results support our hypothesis that reasoning, alignment, and robustness are projections of a single informationgeometric objective, and that models trained using ENIGMA demonstrate principled reasoning without the use of a reward model, offering a path to trusted capability
comment: 52 pages, 10 figures
☆ Towards Real-Time Fake News Detection under Evidence Scarcity
Fake news detection becomes particularly challenging in real-time scenarios, where emerging events often lack sufficient supporting evidence. Existing approaches often rely heavily on external evidence and therefore struggle to generalize under evidence scarcity. To address this issue, we propose Evaluation-Aware Selection of Experts (EASE), a novel framework for real-time fake news detection that dynamically adapts its decision-making process according to the assessed sufficiency of available evidence. EASE introduces a sequential evaluation mechanism comprising three independent perspectives: (1) Evidence-based evaluation, which assesses evidence and incorporates it into decision-making only when the evidence is sufficiently supportive; (2) Reasoning-based evaluation, which leverages the world knowledge of large language models (LLMs) and applies them only when their reliability is adequately established; and (3) Sentiment-based fallback, which integrates sentiment cues when neither evidence nor reasoning is reliable. To enhance the accuracy of evaluation processes, EASE employs instruction tuning with pseudo labels to guide each evaluator in justifying its perspective-specific knowledge through interpretable reasoning. Furthermore, the expert modules integrate the evaluators' justified assessments with the news content to enable evaluation-aware decision-making, thereby enhancing overall detection accuracy. Moreover, we introduce RealTimeNews-25, a new benchmark comprising recent news for evaluating model generalization on emerging news with limited evidence. Extensive experiments demonstrate that EASE not only achieves state-of-the-art performance across multiple benchmarks, but also significantly improves generalization to real-time news. The code and dataset are available: https://github.com/wgyhhhh/EASE.
☆ Do Psychometric Tests Work for Large Language Models? Evaluation of Tests on Sexism, Racism, and Morality
Psychometric tests are increasingly used to assess psychological constructs in large language models (LLMs). However, it remains unclear whether these tests -- originally developed for humans -- yield meaningful results when applied to LLMs. In this study, we systematically evaluate the reliability and validity of human psychometric tests for three constructs: sexism, racism, and morality. We find moderate reliability across multiple item and prompt variations. Validity is evaluated through both convergent (i.e., testing theory-based inter-test correlations) and ecological approaches (i.e., testing the alignment between tests scores and behavior in real-world downstream tasks). Crucially, we find that psychometric test scores do not align, and in some cases even negatively correlate with, model behavior in downstream tasks, indicating low ecological validity. Our results highlight that systematic evaluations of psychometric tests is essential before interpreting their scores. They also suggest that psychometric tests designed for humans cannot be applied directly to LLMs without adaptation.
☆ Attacks by Content: Automated Fact-checking is an AI Security Issue EMNLP 2025
When AI agents retrieve and reason over external documents, adversaries can manipulate the data they receive to subvert their behaviour. Previous research has studied indirect prompt injection, where the attacker injects malicious instructions. We argue that injection of instructions is not necessary to manipulate agents - attackers could instead supply biased, misleading, or false information. We term this an attack by content. Existing defenses, which focus on detecting hidden commands, are ineffective against attacks by content. To defend themselves and their users, agents must critically evaluate retrieved information, corroborating claims with external evidence and evaluating source trustworthiness. We argue that this is analogous to an existing NLP task, automated fact-checking, which we propose to repurpose as a cognitive self-defense tool for agents.
comment: Accepted to EMNLP 2025
☆ XQuant: Achieving Ultra-Low Bit KV Cache Quantization with Cross-Layer Compression EMNLP 2025
Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks. However, their extensive memory requirements, particularly due to KV cache growth during long-text understanding and generation, present significant challenges for deployment in resource-constrained environments. Quantization has emerged as a promising solution to reduce memory consumption while preserving historical information. We propose XQuant, a training-free and plug-and-play framework that achieves ultra-low equivalent bit-width KV cache quantization. XQuant introduces two key innovations: a computationally negligible data-free calibration method and cross-layer KV cache compression, enabling quantization to sub-1.4 bits. Extensive experiments on TruthfulQA and LongBench demonstrate that XQuant outperforms state-of-the-art methods (e.g., KIVI-2bit and AsymKV-1.5bit) by achieving lower bit-width while maintaining superior performance, establishing a better trade-off between memory efficiency and model accuracy.
comment: To be published in The 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025)
☆ CNSocialDepress: A Chinese Social Media Dataset for Depression Risk Detection and Structured Analysis
Depression is a pressing global public health issue, yet publicly available Chinese-language resources for risk detection remain scarce and are mostly limited to binary classification. To address this limitation, we release CNSocialDepress, a benchmark dataset for depression risk detection from Chinese social media posts. The dataset contains 44,178 texts from 233 users, within which psychological experts annotated 10,306 depression-related segments. CNSocialDepress provides binary risk labels together with structured multi-dimensional psychological attributes, enabling interpretable and fine-grained analysis of depressive signals. Experimental results demonstrate its utility across a wide range of NLP tasks, including structured psychological profiling and fine-tuning of large language models for depression detection. Comprehensive evaluations highlight the dataset's effectiveness and practical value for depression risk identification and psychological analysis, thereby providing insights to mental health applications tailored for Chinese-speaking populations.
☆ A Theorem-Proving-Based Evaluation of Neural Semantic Parsing
Graph-matching metrics such as Smatch are the de facto standard for evaluating neural semantic parsers, yet they capture surface overlap rather than logical equivalence. We reassess evaluation by pairing graph-matching with automated theorem proving. We compare two approaches to building parsers: supervised fine-tuning (T5-Small/Base) and few-shot in-context learning (GPT-4o/4.1/5), under normalized and unnormalized targets. We evaluate outputs using graph-matching, bidirectional entailment between source and target formulas with a first-order logic theorem prover, and well-formedness. Across settings, we find that models performing well on graph-matching often fail to produce logically equivalent formulas. Normalization reduces incidental target variability, improves well-formedness, and strengthens logical adequacy. Error analysis shows performance degrades with increasing formula complexity and with coordination, prepositional phrases, and passive voice; the dominant failures involve variable binding and indexing, and predicate naming. These findings highlight limits of graph-based metrics for reasoning-oriented applications and motivate logic-sensitive evaluation and training objectives together with simplified, normalized target representations. All code and data for our experiments are publicly available.
comment: Accepted to BlackboxNLP 2025
☆ Fairness Metric Design Exploration in Multi-Domain Moral Sentiment Classification using Transformer-Based Models
Ensuring fairness in natural language processing for moral sentiment classification is challenging, particularly under cross-domain shifts where transformer models are increasingly deployed. Using the Moral Foundations Twitter Corpus (MFTC) and Moral Foundations Reddit Corpus (MFRC), this work evaluates BERT and DistilBERT in a multi-label setting with in-domain and cross-domain protocols. Aggregate performance can mask disparities: we observe pronounced asymmetry in transfer, with Twitter->Reddit degrading micro-F1 by 14.9% versus only 1.5% for Reddit->Twitter. Per-label analysis reveals fairness violations hidden by overall scores; notably, the authority label exhibits Demographic Parity Differences of 0.22-0.23 and Equalized Odds Differences of 0.40-0.41. To address this gap, we introduce the Moral Fairness Consistency (MFC) metric, which quantifies the cross-domain stability of moral foundation detection. MFC shows strong empirical validity, achieving a perfect negative correlation with Demographic Parity Difference (rho = -1.000, p < 0.001) while remaining independent of standard performance metrics. Across labels, loyalty demonstrates the highest consistency (MFC = 0.96) and authority the lowest (MFC = 0.78). These findings establish MFC as a complementary, diagnosis-oriented metric for fairness-aware evaluation of moral reasoning models, enabling more reliable deployment across heterogeneous linguistic contexts. .
☆ WebRouter: Query-specific Router via Variational Information Bottleneck for Cost-sensitive Web Agent
LLM-brained web agents offer powerful capabilities for web automation but face a critical cost-performance trade-off. The challenge is amplified by web agents' inherently complex prompts that include goals, action histories, and environmental states, leading to degraded LLM ensemble performance. To address this, we introduce WebRouter, a novel query-specific router trained from an information-theoretic perspective. Our core contribution is a cost-aware Variational Information Bottleneck (ca-VIB) objective, which learns a compressed representation of the input prompt while explicitly penalizing the expected operational cost. Experiments on five real-world websites from the WebVoyager benchmark show that WebRouter reduces operational costs by a striking 87.8\% compared to a GPT-4o baseline, while incurring only a 3.8\% accuracy drop.
☆ The Curious Case of Factual (Mis)Alignment between LLMs' Short- and Long-Form Answers
Large language models (LLMs) can correctly answer "When was Einstein born?" yet fail to provide the same date when writing about Einstein's life revealing a fundamental inconsistency in how models access factual knowledge across task complexities. While models display impressive accuracy on factual question-answering benchmarks, the reliability gap between simple and complex queries remains poorly understood, eroding their trustworthiness. In this work, we introduce Short-Long Form Alignment for Factual Question Answering (SLAQ), a controlled evaluation framework that compares LLMs' answers to the same factual questions asked (a) in isolation (short) vs. (b) integrated into complex queries (long). Looking at 16 LLMs across 600 queries, we find a systematic misalignment of answers to the corresponding short and long queries. We further uncover position-dependent accuracy loss and momentum effects where consecutive correct or incorrect answers create self-reinforcing patterns. Through mechanistic analysis, we find that aligned facts activate overlapping model internals, and that metrics based on mechanistic similarity can predict short-long answer alignment with up to 78% accuracy. Our work establishes factual consistency over query complexity as an important aspect of LLMs' trustworthiness and challenges current evaluation practices, which implicitly assume that good performance for simple factual queries implies reliability in more complex knowledge-seeking tasks too.
☆ Domain-Specific Data Generation Framework for RAG Adaptation
Retrieval-Augmented Generation (RAG) combines the language understanding and reasoning power of large language models (LLMs) with external retrieval to enable domain-grounded responses. Effectively adapting RAG systems to domain-specific settings requires specialized, context-rich training data beyond general-purpose question-answering. Here, we propose RAGen, a scalable and modular framework for generating domain-grounded question-answer-context (QAC) triples tailored to diverse RAG adaptation approaches. RAGen produces these QAC triples by identifying key concepts in documents, generating diverse questions guided by Bloom's Taxonomy-inspired principles, and pairing them with precise answers extracted from relevant contexts. RAGen supports multiple RAG adaptation strategies, including the optimization of key components such as the LLM, retriever, and embedding model, etc. Its modular pipeline features semantic chunking, hierarchical concept extraction, and multi-chunk retrieval, along with the introduction of curated distractor contexts to promote robust reasoning. Designed for scalability, RAGen efficiently handles large and evolving document corpora without redundant processing, making it especially suitable for dynamic evolving domains such as scientific research and enterprise knowledge bases.
☆ Discursive Circuits: How Do Language Models Understand Discourse Relations? EMNLP 2025
Which components in transformer language models are responsible for discourse understanding? We hypothesize that sparse computational graphs, termed as discursive circuits, control how models process discourse relations. Unlike simpler tasks, discourse relations involve longer spans and complex reasoning. To make circuit discovery feasible, we introduce a task called Completion under Discourse Relation (CuDR), where a model completes a discourse given a specified relation. To support this task, we construct a corpus of minimal contrastive pairs tailored for activation patching in circuit discovery. Experiments show that sparse circuits ($\approx 0.2\%$ of a full GPT-2 model) recover discourse understanding in the English PDTB-based CuDR task. These circuits generalize well to unseen discourse frameworks such as RST and SDRT. Further analysis shows lower layers capture linguistic features such as lexical semantics and coreference, while upper layers encode discourse-level abstractions. Feature utility is consistent across frameworks (e.g., coreference supports Expansion-like relations).
comment: Accepted to EMNLP 2025 (Main Conference); 9 pages, 8 figures, 5 tables (20 pages, 12 figures, 14 tables including references and appendices)
☆ Evaluating Reasoning Faithfulness in Medical Vision-Language Models using Multimodal Perturbations
Vision-language models (VLMs) often produce chain-of-thought (CoT) explanations that sound plausible yet fail to reflect the underlying decision process, undermining trust in high-stakes clinical use. Existing evaluations rarely catch this misalignment, prioritizing answer accuracy or adherence to formats. We present a clinically grounded framework for chest X-ray visual question answering (VQA) that probes CoT faithfulness via controlled text and image modifications across three axes: clinical fidelity, causal attribution, and confidence calibration. In a reader study (n=4), evaluator-radiologist correlations fall within the observed inter-radiologist range for all axes, with strong alignment for attribution (Kendall's $\tau_b=0.670$), moderate alignment for fidelity ($\tau_b=0.387$), and weak alignment for confidence tone ($\tau_b=0.091$), which we report with caution. Benchmarking six VLMs shows that answer accuracy and explanation quality are decoupled, acknowledging injected cues does not ensure grounding, and text cues shift explanations more than visual cues. While some open-source models match final answer accuracy, proprietary models score higher on attribution (25.0% vs. 1.4%) and often on fidelity (36.1% vs. 31.7%), highlighting deployment risks and the need to evaluate beyond final answer accuracy.
☆ Can Tool-Integrated Reinforcement Learning Generalize Across Diverse Domains?
Recent advances in large language models (LLMs) have demonstrated remarkable capabilities in reasoning and tool utilization. However, the generalization of tool-augmented reinforcement learning (RL) across diverse domains remains underexplored. In this work, we investigate the cross-domain generalization of an LLM agent equipped with a code interpreter tool, which is exclusively trained on mathematical problem-solving tasks. Despite the restricted training domain, we evaluate the agent's performance across several distinct reasoning domains. The results reveal that RL-based tool usage learned from mathematical tasks can be effectively transferred to complex tasks in other domains, enabling great task performance and high token efficiency. To facilitate this cross-domain transfer, we propose a Tool Generalization Reinforcement Learning (TGRL) framework designed to promote domain-agnostic learning and skill migration, encompassing: (i) a standardized tool interface that abstracts domain-specific nuances through consistent formatting and explicit termination, fostering transferable invocation patterns; (ii) a dual-component reward system that decomposes rewards to incentivize generalizable behaviors like tool efficiency and reasoning abstraction, ensuring alignment and robustness across domain shifts; and (iii) an XML-based prompt template that separates thinking, tool calls, and responses to encourage modular, domain-invariant planning and coherent multi-turn interactions. Extensive experiments across diverse benchmarks validate our approach, achieving state-of-the-art performance and highlighting the cross-domain potential of Tool RL for LLM reasoning.
☆ EAGER: Entropy-Aware GEneRation for Adaptive Inference-Time Scaling
With the rise of reasoning language models and test-time scaling methods as a paradigm for improving model performance, substantial computation is often required to generate multiple candidate sequences from the same prompt. This enables exploration of different reasoning paths toward the correct solution, however, allocates the same compute budget for each prompt. Grounded on the assumption that different prompts carry different degrees of complexity, and thus different computation needs, we propose EAGer, a training-free generation method that leverages model uncertainty through token-wise entropy distribution to reduce redundant computation and concurrently improve overall performance. EAGer allows branching to multiple reasoning paths only in the presence of high-entropy tokens, and then reallocates the saved compute budget to the instances where exploration of alternative paths is most needed. We find that across multiple open-source models on complex reasoning benchmarks such as AIME 2025, EAGer can reallocate the budget without accessing target labels, achieving the best efficiency-performance trade-off in terms of reasoning length and Pass@k. When target labels are accessible, EAGer generates up to 65% fewer tokens (hence saving compute) and achieves up to 37% improvement in Pass@k compared to the Full Parallel Sampling.
☆ ELMO: Efficiency via Low-precision and Peak Memory Optimization in Large Output Spaces ICML 2025
Large output spaces, also referred to as Extreme multilabel classification (XMC), is a setting that arises, e.g., in large-scale tagging and product-to-product recommendation, and is characterized by the number of labels ranging from hundreds of thousands to millions. This means that the linear classification head, usually only a tiny fraction of the overall model, turns into the main driver for compute and memory demand. Current state-of-the-art XMC methods predominantly rely on FP16-FP32 mixed-precision training, which we show can be unstable, and inefficient in terms of memory usage and computational overhead. Meanwhile, existing low-precision methods typically retain higher precision for the classification layer. In this work, we propose ELMO, a pure low-precision training framework for XMC models using BFloat16 and Float8 data types. By leveraging Kahan summation and stochastic rounding, we demonstrate that XMC models can be effectively trained entirely in Float8, without relying on single-precision master weights or tensor scaling. Low-precision training, combined with our proposed memory optimizations -- gradient fusion and chunking -- enables significant reductions in GPU memory usage. For example, we train a 3-million-label XMC model with only 6.6 GiB of GPU memory, compared to the 39.7 GiB required by the optimized SOTA method, Renee without compromising accuracy.
comment: Accepted to ICML 2025
☆ Bridging Gaps in Hate Speech Detection: Meta-Collections and Benchmarks for Low-Resource Iberian Languages
Hate speech poses a serious threat to social cohesion and individual well-being, particularly on social media, where it spreads rapidly. While research on hate speech detection has progressed, it remains largely focused on English, resulting in limited resources and benchmarks for low-resource languages. Moreover, many of these languages have multiple linguistic varieties, a factor often overlooked in current approaches. At the same time, large language models require substantial amounts of data to perform reliably, a requirement that low-resource languages often cannot meet. In this work, we address these gaps by compiling a meta-collection of hate speech datasets for European Spanish, standardised with unified labels and metadata. This collection is based on a systematic analysis and integration of existing resources, aiming to bridge the data gap and support more consistent and scalable hate speech detection. We extended this collection by translating it into European Portuguese and into a Galician standard that is more convergent with Spanish and another Galician variant that is more convergent with Portuguese, creating aligned multilingual corpora. Using these resources, we establish new benchmarks for hate speech detection in Iberian languages. We evaluate state-of-the-art large language models in zero-shot, few-shot, and fine-tuning settings, providing baseline results for future research. Moreover, we perform a cross-lingual analysis with our target languages. Our findings underscore the importance of multilingual and variety-aware approaches in hate speech detection and offer a foundation for improved benchmarking in underrepresented European languages.
☆ One Size Does Not Fit All: Exploring Variable Thresholds for Distance-Based Multi-Label Text Classification
Distance-based unsupervised text classification is a method within text classification that leverages the semantic similarity between a label and a text to determine label relevance. This method provides numerous benefits, including fast inference and adaptability to expanding label sets, as opposed to zero-shot, few-shot, and fine-tuned neural networks that require re-training in such cases. In multi-label distance-based classification and information retrieval algorithms, thresholds are required to determine whether a text instance is "similar" to a label or query. Similarity between a text and label is determined in a dense embedding space, usually generated by state-of-the-art sentence encoders. Multi-label classification complicates matters, as a text instance can have multiple true labels, unlike in multi-class or binary classification, where each instance is assigned only one label. We expand upon previous literature on this underexplored topic by thoroughly examining and evaluating the ability of sentence encoders to perform distance-based classification. First, we perform an exploratory study to verify whether the semantic relationships between texts and labels vary across models, datasets, and label sets by conducting experiments on a diverse collection of realistic multi-label text classification (MLTC) datasets. We find that similarity distributions show statistically significant differences across models, datasets and even label sets. We propose a novel method for optimizing label-specific thresholds using a validation set. Our label-specific thresholding method achieves an average improvement of 46% over normalized 0.5 thresholding and outperforms uniform thresholding approaches from previous work by an average of 14%. Additionally, the method demonstrates strong performance even with limited labeled examples.
☆ TypePilot: Leveraging the Scala Type System for Secure LLM-generated Code
Large language Models (LLMs) have shown remarkable proficiency in code generation tasks across various programming languages. However, their outputs often contain subtle but critical vulnerabilities, posing significant risks when deployed in security-sensitive or mission-critical systems. This paper introduces TypePilot, an agentic AI framework designed to enhance the security and robustness of LLM-generated code by leveraging strongly typed and verifiable languages, using Scala as a representative example. We evaluate the effectiveness of our approach in two settings: formal verification with the Stainless framework and general-purpose secure code generation. Our experiments with leading open-source LLMs reveal that while direct code generation often fails to enforce safety constraints, just as naive prompting for more secure code, our type-focused agentic pipeline substantially mitigates input validation and injection vulnerabilities. The results demonstrate the potential of structured, type-guided LLM workflows to improve the SotA of the trustworthiness of automated code generation in high-assurance domains.
☆ $How^{2}$: How to learn from procedural How-to questions
An agent facing a planning problem can use answers to how-to questions to reduce uncertainty and fill knowledge gaps, helping it solve both current and future tasks. However, their open ended nature, where valid answers to "How do I X?" range from executable actions to high-level descriptions of X's sub-goals, makes them challenging for AI agents to ask, and for AI experts to answer, in ways that support efficient planning. We introduce $How^{2}$, a memory agent framework that enables agents to ask how-to questions, store the answers, and reuse them for lifelong learning in interactive environments. We evaluate our approach in Plancraft, a Minecraft crafting environment, where agents must complete an assembly task by manipulating inventory items. Using teacher models that answer at varying levels of abstraction, from executable action sequences to high-level subgoal descriptions, we show that lifelong learning agents benefit most from answers that are abstracted and decoupled from the current state. $How^{2}$ offers a way for LLM-based agents to improve their planning capabilities over time by asking questions in interactive environments.
☆ Enhancing LLM Reasoning via Non-Human-Like Reasoning Path Preference Optimization
Current approaches for strengthening LLM reasoning tend to introduce a training bias toward human-like reasoning trajectories. In step-wise preference optimization, in particular, dependence on human or higher-capacity model annotations for intermediate steps limits exploration of alternative, non-human-like reasoning paths and thus constrains achievable performance. Furthermore, through a small-scale pilot study, we observed that in approximately 75% of cases, the model's first erroneous step occurs after the lowest-confidence point. This suggests that guiding the model at its lowest-confidence point before an error provides more accurate supervision than locating the first explicit error. In this paper, we propose Confidence-Guided Reasoning Path Preference Optimization (CGPO), a method that leverages a confidence signal to identify points of maximal uncertainty in the model's reasoning process and applies self-generated, non-human-like reasoning-path guidance to mitigate trajectory drift. Our experiments span diverse models applied to both code and mathematical reasoning tasks. The results show that, with the same amount of training data, our method using data generated by a small model can achieve better performance in most cases compared with approaches using data generated by a strong model or human-annotated.
comment: 13 pages
☆ VCB Bench: An Evaluation Benchmark for Audio-Grounded Large Language Model Conversational Agents
Recent advances in large audio language models (LALMs) have greatly enhanced multimodal conversational systems. However, existing benchmarks remain limited -- they are mainly English-centric, rely on synthetic speech, and lack comprehensive, discriminative evaluation across multiple dimensions. To address these gaps, we present Voice Chat Bot Bench (VCB Bench) -- a high-quality Chinese benchmark built entirely on real human speech. VCB Bench evaluates LALMs from three complementary perspectives: instruction following (including speech-level control beyond text commands), knowledge understanding (general knowledge, reasoning, and daily dialogue), and robustness (stability under perturbations in content, environment, and speaker traits). Experiments on representative LALMs reveal notable performance gaps and highlight future directions for improvement. VCB Bench provides a reproducible and fine-grained evaluation framework, offering standardized methodology and practical insights for advancing Chinese voice conversational models.
comment: 20 pages, 5 figures
☆ Latent Refinement Decoding: Enhancing Diffusion-Based Language Models by Refining Belief States
Autoregressive (AR) models remain the standard for natural language generation but still suffer from high latency due to strictly sequential decoding. Recent diffusion-inspired approaches, such as LlaDA and Dream, mitigate this by generating in parallel, yet they suffer from two core limitations: information loss, as predictive distributions for non-finalized tokens are discarded at each step, and premature commitment, where local decisions are made without sufficient global coordination. We introduce Latent Refinement Decoding (LRD), a two-stage framework with Latent Refinement and a Predictive Feedback Loop. The first stage maintains masked positions as distributional mixtures of predicted tokens and the mask embedding, allowing the model to establish more globally consistent beliefs. The second stage progressively finalizes confident tokens while retaining uncertain ones for iterative feedback. KL-divergence dynamics provide a principled and reliable criterion for convergence and early stopping. Experiments across coding (HumanEval +6.3, MBPP +2.6) and reasoning (GSM8K +2.9, MATH500 +3.8) show that LRD improves accuracy while delivering speedups of up to 10.6x, making it a strong and versatile alternative for parallel sequence generation.
☆ Enabling Doctor-Centric Medical AI with LLMs through Workflow-Aligned Tasks and Benchmarks
The rise of large language models (LLMs) has transformed healthcare by offering clinical guidance, yet their direct deployment to patients poses safety risks due to limited domain expertise. To mitigate this, we propose repositioning LLMs as clinical assistants that collaborate with experienced physicians rather than interacting with patients directly. We conduct a two-stage inspiration-feedback survey to identify real-world needs in clinical workflows. Guided by this, we construct DoctorFLAN, a large-scale Chinese medical dataset comprising 92,000 Q&A instances across 22 clinical tasks and 27 specialties. To evaluate model performance in doctor-facing applications, we introduce DoctorFLAN-test (550 single-turn Q&A items) and DotaBench (74 multi-turn conversations). Experimental results with over ten popular LLMs demonstrate that DoctorFLAN notably improves the performance of open-source LLMs in medical contexts, facilitating their alignment with physician workflows and complementing existing patient-oriented models. This work contributes a valuable resource and framework for advancing doctor-centered medical LLM development
☆ LogiNumSynth: Synthesizing Joint Logical-Numerical Reasoning Problems for Language Models
Joint logical-numerical reasoning remains a major challenge for language models, yet existing datasets rely on fixed rule sets and offer limited control over task complexity, constraining their generalizability for evaluation and training. We present LogiNumSynth, a flexible natural language problem synthesizer that synthesizes tasks requiring proficiency in joint logical reasoning (e.g., rule-based reasoning) and numerical reasoning (e.g., arithmetic computation). LogiNumSynth supports fine-grained control over reasoning world richness, logical reasoning depth, and the complexity of numerical computations, enabling flexible data synthesis across difficulty levels. We demonstrate three key contributions: (1) Synthesizer -- synthesizing fully controllable joint reasoning tasks over natural language; (2) Evaluation & Process Analysis -- evaluating both process accuracy and answer accuracy; (3) Targeted Training -- using synthesized data to enhance LLMs' reasoning performance. Experiments with multiple LLMs highlight persistent weaknesses in logical-numerical reasoning, showing that LogiNumSynth can serve as both a diagnostic tool and a source of targeted supervision for advancing integrated reasoning skills.
comment: 30 pages, 3 figures
☆ Automating Structural Engineering Workflows with Large Language Model Agents
We introduce $\textbf{MASSE}$, the first Multi-Agent System for Structural Engineering, effectively integrating large language model (LLM)-based agents with real-world engineering workflows. Structural engineering is a fundamental yet traditionally stagnant domain, with core workflows remaining largely unchanged for decades despite its substantial economic impact and global market size. Recent advancements in LLMs have significantly enhanced their ability to perform complex reasoning, long-horizon planning, and precise tool utilization -- capabilities well aligned with structural engineering tasks such as interpreting design codes, executing load calculations, and verifying structural capacities. We present a proof-of-concept showing that most real-world structural engineering workflows can be fully automated through a training-free LLM-based multi-agent system. MASSE enables immediate deployment in professional environments, and our comprehensive validation on real-world case studies demonstrates that it can reduce expert workload from approximately two hours to mere minutes, while enhancing both reliability and accuracy in practical engineering scenarios.
comment: Code: https://github.com/DelosLiang/masse
☆ DND: Boosting Large Language Models with Dynamic Nested Depth
We introduce Dynamic Nested Depth (DND), a novel method that improves performance for off-the-shelf LLMs by selecting critical tokens to reprocess in a nested depth manner. Specifically, at the end of the given transformer layer, DND identifies more critical tokens with a router and feeds them back for an extra round of processing, effectively ``reviewing" difficult tokens while avoiding redundant computation for easier ones. The dynamic selection mechanism is tailored for precise control via two novel strategies: a router controlling loss to enhance token selection distinguishability, and a threshold control scheme to ensure selection stability. We demonstrate the effectiveness of DND by directly integrating it into pre-trained dense and MoE models during a post-training phase. On diverse benchmarks, this approach boosts the performances of the dense Qwen3-1.7B by 1.88% and the MoE Qwen3-30B-A3B by 0.87%, all with a minimal parameter and computing increase.
comment: TL;DR: We introduce Dynamic Nested Depth (DND), an efficient paradigm that adaptively identifies critical tokens and selectively deepens their computation via nested re-processing
☆ ABLEIST: Intersectional Disability Bias in LLM-Generated Hiring Scenarios
Large language models (LLMs) are increasingly under scrutiny for perpetuating identity-based discrimination in high-stakes domains such as hiring, particularly against people with disabilities (PwD). However, existing research remains largely Western-centric, overlooking how intersecting forms of marginalization--such as gender and caste--shape experiences of PwD in the Global South. We conduct a comprehensive audit of six LLMs across 2,820 hiring scenarios spanning diverse disability, gender, nationality, and caste profiles. To capture subtle intersectional harms and biases, we introduce ABLEIST (Ableism, Inspiration, Superhumanization, and Tokenism), a set of five ableism-specific and three intersectional harm metrics grounded in disability studies literature. Our results reveal significant increases in ABLEIST harms towards disabled candidates--harms that many state-of-the-art models failed to detect. These harms were further amplified by sharp increases in intersectional harms (e.g., Tokenism) for gender and caste-marginalized disabled candidates, highlighting critical blind spots in current safety tools and the need for intersectional safety evaluations of frontier models in high-stakes domains like hiring.
comment: 28 pages, 11 figures, 16 tables. In submission
☆ DeepResearchGuard: Deep Research with Open-Domain Evaluation and Multi-Stage Guardrails for Safety
Deep research frameworks have shown promising capabilities in synthesizing comprehensive reports from web sources. While deep research possesses significant potential to address complex issues through planning and research cycles, existing frameworks are deficient in sufficient evaluation procedures and stage-specific protections. They typically treat evaluation as exact match accuracy of question-answering, but overlook crucial aspects of report quality such as credibility, coherence, breadth, depth, and safety. This oversight may result in hazardous or malicious sources being integrated into the final report. To address these issues, we introduce DEEPRESEARCHGUARD, a comprehensive framework featuring four-stage safeguards with open-domain evaluation of references and reports. We assess performance across multiple metrics, e.g., defense success rate and over-refusal rate, and five key report dimensions. In the absence of a suitable safety benchmark, we introduce DRSAFEBENCH, a stage-wise benchmark for deep research safety. Our evaluation spans diverse state-of-the-art LLMs, including GPT-4o, Gemini-2.5-flash, DeepSeek-v3, and o4-mini. DEEPRESEARCHGUARD achieves an average defense success rate improvement of 18.16% while reducing over-refusal rate by 6%. The input guard provides the most substantial early-stage protection by filtering out obvious risks, while the plan and research guards enhance citation discipline and source credibility. Through extensive experiments, we show that DEEPRESEARCHGUARD enables comprehensive open-domain evaluation and stage-aware defenses that effectively block harmful content propagation, while systematically improving report quality without excessive over-refusal rates. The code can be found via https://github.com/Jasonya/DeepResearchGuard.
☆ A Survey on Agentic Multimodal Large Language Models
With the recent emergence of revolutionary autonomous agentic systems, research community is witnessing a significant shift from traditional static, passive, and domain-specific AI agents toward more dynamic, proactive, and generalizable agentic AI. Motivated by the growing interest in agentic AI and its potential trajectory toward AGI, we present a comprehensive survey on Agentic Multimodal Large Language Models (Agentic MLLMs). In this survey, we explore the emerging paradigm of agentic MLLMs, delineating their conceptual foundations and distinguishing characteristics from conventional MLLM-based agents. We establish a conceptual framework that organizes agentic MLLMs along three fundamental dimensions: (i) Agentic internal intelligence functions as the system's commander, enabling accurate long-horizon planning through reasoning, reflection, and memory; (ii) Agentic external tool invocation, whereby models proactively use various external tools to extend their problem-solving capabilities beyond their intrinsic knowledge; and (iii) Agentic environment interaction further situates models within virtual or physical environments, allowing them to take actions, adapt strategies, and sustain goal-directed behavior in dynamic real-world scenarios. To further accelerate research in this area for the community, we compile open-source training frameworks, training and evaluation datasets for developing agentic MLLMs. Finally, we review the downstream applications of agentic MLLMs and outline future research directions for this rapidly evolving field. To continuously track developments in this rapidly evolving field, we will also actively update a public repository at https://github.com/HJYao00/Awesome-Agentic-MLLMs.
☆ Secret-Protected Evolution for Differentially Private Synthetic Text Generation
Text data has become extremely valuable on large language models (LLMs) and even lead to general artificial intelligence (AGI). A lot of high-quality text in the real world is private and cannot be freely used due to privacy concerns. Therefore, differentially private (DP) synthetic text generation has been proposed, aiming to produce high-utility synthetic data while protecting sensitive information. However, existing DP synthetic text generation imposes uniform guarantees that often overprotect non-sensitive content, resulting in substantial utility loss and computational overhead. Therefore, we propose Secret-Protected Evolution (SecPE), a novel framework that extends private evolution with secret-aware protection. Theoretically, we show that SecPE satisfies $(\mathrm{p}, \mathrm{r})$-secret protection, constituting a relaxation of Gaussian DP that enables tighter utility-privacy trade-offs, while also substantially reducing computational complexity relative to baseline methods. Empirically, across the OpenReview, PubMed, and Yelp benchmarks, SecPE consistently achieves lower Fr\'echet Inception Distance (FID) and higher downstream task accuracy than GDP-based Aug-PE baselines, while requiring less noise to attain the same level of protection. Our results highlight that secret-aware guarantees can unlock more practical and effective privacy-preserving synthetic text generation.
☆ Revisiting Model Interpolation for Efficient Reasoning
Model merging, typically on Instruct and Thinking models, has shown remarkable performance for efficient reasoning. In this paper, we systematically revisit the simplest merging method that interpolates two weights directly. Particularly, we observe that model interpolation follows a three-stage evolutionary paradigm with distinct behaviors on the reasoning trajectory. These dynamics provide a principled guide for navigating the performance-cost trade-off. Empirical results demonstrate that a strategically interpolated model surprisingly surpasses sophisticated model merging baselines on both efficiency and effectiveness. We further validate our findings with extensive ablation studies on model layers, modules, and decoding strategies. Ultimately, this work demystifies model interpolation and offers a practical framework for crafting models with precisely targeted reasoning capabilities. Code is available at \href{https://github.com/wutaiqiang/MI}{Github}.
comment: 14 pages, 6 figures, 7 tables. Working in progress
☆ Enhancing Large Language Model Reasoning via Selective Critical Token Fine-Tuning
Large language models (LLMs) primarily rely on supervised fine-tuning (SFT) as a key method to adapt pre-trained models to domain-specific tasks such as mathematical reasoning. However, standard SFT uniformly penalizes all tokens, neglecting that only a small subset of critical tokens determines reasoning correctness. This uniform supervision often causes reduced output diversity and limited generalization. We propose Critical Token Fine-tuning (CFT), a simple yet effective approach that updates only tokens identified as functionally indispensable via counterfactual perturbations. By focusing gradient signals on these decisive reasoning steps while preserving the diversity of non-critical tokens, CFT can enhance both generation and diversity. Extensive experiments on five models across three families (Qwen, OLMo, LLaMA) and eleven mathematical reasoning benchmarks show that CFT, despite fine-tuning on less than 12% of tokens, consistently outperforms standard SFT. Moreover, CFT enables test-time scaling through improved sampling diversity and provides a stronger initialization for reinforcement learning, sustaining performance gains in later training stages while maintaining higher entropy for better exploration. These results highlight CFT as a practical and general framework for efficient and robust LLM fine-tuning.
☆ RV-HATE: Reinforced Multi-Module Voting for Implicit Hate Speech Detection
Hate speech remains prevalent in human society and continues to evolve in its forms and expressions. Modern advancements in internet and online anonymity accelerate its rapid spread and complicate its detection. However, hate speech datasets exhibit diverse characteristics primarily because they are constructed from different sources and platforms, each reflecting different linguistic styles and social contexts. Despite this diversity, prior studies on hate speech detection often rely on fixed methodologies without adapting to data-specific features. We introduce RV-HATE, a detection framework designed to account for the dataset-specific characteristics of each hate speech dataset. RV-HATE consists of multiple specialized modules, where each module focuses on distinct linguistic or contextual features of hate speech. The framework employs reinforcement learning to optimize weights that determine the contribution of each module for a given dataset. A voting mechanism then aggregates the module outputs to produce the final decision. RV-HATE offers two primary advantages: (1)~it improves detection accuracy by tailoring the detection process to dataset-specific attributes, and (2)~it also provides interpretable insights into the distinctive features of each dataset. Consequently, our approach effectively addresses implicit hate speech and achieves superior performance compared to conventional static methods. Our code is available at https://github.com/leeyejin1231/RV-HATE.
comment: 10 pages, 9 figures, 12 tables
☆ Judge Before Answer: Can MLLM Discern the False Premise in Question?
Multimodal large language models (MLLMs) have witnessed astonishing advancements in recent years. Despite these successes, MLLMs remain vulnerable to flase premise problems. However, existing benchmarks targeting this issue are limited in scope: they often lack fine-grained categorization, exhibit insufficient coverage, and thus fail to provide a rigorous evaluation of the ability of models to recognize false premises. To bridge this gap, we introduce a fully automated pipeline for constructing a comprehensive benchmark of false premise questions. Our method systematically categorizes the premises into three main types and thirteen subtypes according to the abilities required to identify the premises, resulting in the JBA dataset.Results show current MLLMs still struggle with false premise recognition. Building upon this benchmark, we further propose a recognition enhancement framework tailored to strengthen the robustness of MLLMs to detect false premises. Extensive experiments demonstrate that models trained with our framework achieve significant improvements in false premise recognition.
☆ KOTOX: A Korean Toxic Dataset for Deobfuscation and Detoxification
Toxic content has become an increasingly critical social issue with the rapid expansion of online communication. While numerous studies explored methods for detecting and detoxifying such content, most have focused primarily on English, leaving low-resource language underrepresented. Consequently, Large Language Models~(LLMs) often struggle to identify and neutralize toxic expressions in these languages. This challenge becomes even more pronounced when user employ obfuscation techniques to evade detection systems. Therefore, we propose a \textbf{KOTOX: Korean Toxic Dataset} for deobfuscation and detoxicification to address this issue. We categorize various obfuscation approaches based on linguistic characteristics of Korean and define a set of transformation rules grounded in real-word examples. Using these rules, we construct three dataset versions (easy, normal, and hard) representing different levels of obfuscation difficulty. This is the first dataset that simultaneously supports deobfuscation and detoxification for the Korean language. We expect it to facilitate better understanding and mitigating of obfuscated toxic content in LLM for low-resource languages. Our code and data are available at https://github.com/leeyejin1231/KOTOX.
comment: 25 pages, 5 figures, 25 tables
☆ Rediscovering Entropy Regularization: Adaptive Coefficient Unlocks Its Potential for LLM Reinforcement Learning
Reasoning ability has become a defining capability of Large Language Models (LLMs), with Reinforcement Learning with Verifiable Rewards (RLVR) emerging as a key paradigm to enhance it. However, RLVR training often suffers from policy entropy collapse, where the policy becomes overly deterministic, hindering exploration and limiting reasoning performance. While entropy regularization is a common remedy, its effectiveness is highly sensitive to the fixed coefficient, making it unstable across tasks and models. In this work, we revisit entropy regularization in RLVR and argue that its potential has been largely underestimated. Our analysis shows that (i) tasks of varying difficulty demand distinct exploration intensities, and (ii) balanced exploration may require the policy entropy to be maintained within a moderate range below its initial level. Therefore, we propose Adaptive Entropy Regularization (AER)--a framework that dynamically balances exploration and exploitation via three components: difficulty-aware coefficient allocation, initial-anchored target entropy, and dynamic global coefficient adjustment. Experiments on multiple mathematical reasoning benchmarks show that AER consistently outperforms baselines, improving both reasoning accuracy and exploration capability.
comment: 16 pages, 4 figures
☆ Punctuation-aware treebank tree binarization
This article presents a curated resource and evaluation suite for punctuation-aware treebank binarization. Standard binarization pipelines drop punctuation before head selection, which alters constituent shape and harms head-child identification. We release (1) a reproducible pipeline that preserves punctuation as sibling nodes prior to binarization, (2) derived artifacts and metadata (intermediate @X markers, reversibility signatures, alignment indices), and (3) an accompanying evaluation suite covering head-child prediction, round-trip reversibility, and structural compatibility with derivational resources (CCGbank). On the Penn Treebank, punctuation-aware preprocessing improves head prediction accuracy from 73.66\% (Collins rules) and 86.66\% (MLP) to 91.85\% with the same classifier, and achieves competitive alignment against CCGbank derivations. All code, configuration files, and documentation are released to enable replication and extension to other corpora.
☆ The Social Cost of Intelligence: Emergence, Propagation, and Amplification of Stereotypical Bias in Multi-Agent Systems
Bias in large language models (LLMs) remains a persistent challenge, manifesting in stereotyping and unfair treatment across social groups. While prior research has primarily focused on individual models, the rise of multi-agent systems (MAS), where multiple LLMs collaborate and communicate, introduces new and largely unexplored dynamics in bias emergence and propagation. In this work, we present a comprehensive study of stereotypical bias in MAS, examining how internal specialization, underlying LLMs and inter-agent communication protocols influence bias robustness, propagation, and amplification. We simulate social contexts where agents represent different social groups and evaluate system behavior under various interaction and adversarial scenarios. Experiments on three bias benchmarks reveal that MAS are generally less robust than single-agent systems, with bias often emerging early through in-group favoritism. However, cooperative and debate-based communication can mitigate bias amplification, while more robust underlying LLMs improve overall system stability. Our findings highlight critical factors shaping fairness and resilience in multi-agent LLM systems.
comment: 15 pages, 19 figures, Preprint. Under review
☆ End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF: A Reproducibility Study
We present a reproducibility study of the state-of-the-art neural architecture for sequence labeling proposed by Ma and Hovy (2016)\cite{ma2016end}. The original BiLSTM-CNN-CRF model combines character-level representations via Convolutional Neural Networks (CNNs), word-level context modeling through Bi-directional Long Short-Term Memory networks (BiLSTMs), and structured prediction using Conditional Random Fields (CRFs). This end-to-end approach eliminates the need for hand-crafted features while achieving excellent performance on named entity recognition (NER) and part-of-speech (POS) tagging tasks. Our implementation successfully reproduces the key results, achieving 91.18\% F1-score on CoNLL-2003 NER and demonstrating the model's effectiveness across sequence labeling tasks. We provide a detailed analysis of the architecture components and release an open-source PyTorch implementation to facilitate further research.
☆ Evaluating Language Models' Evaluations of Games
Reasoning is not just about solving problems -- it is also about evaluating which problems are worth solving at all. Evaluations of artificial intelligence (AI) systems primarily focused on problem solving, historically by studying how models play games such as chess and Go. In this paper, we advocate for a new paradigm that assesses AI systems' evaluation of games. First, we introduce a formalism for evaluating such evaluations. We then leverage a large-scale dataset of over $100$ novel board games and over 450 human judgments to compare evaluations produced by modern language and reasoning models against those of people and symbolic computational agents. We consider two kinds of evaluative queries: assessing the payoff (or fairness) and the funness of games. These queries span two dimensions relevant to the design of evaluations of AI evaluations: how complex a query is to compute and how difficult a query is to quantify. Our results show that reasoning models are generally more aligned to people in their evaluations of games than non-reasoning language models. However, we observe a non-monotonic relationship: as models get closer to game-theoretic optimal, their fit to human data weakens. We also observe more "jaggedness" across models for assessing funness, in line with the greater difficulty of quantifying this query. Across queries and games, reasoning models show highly variable and unpredictable resource usage when assessing queries, pointing to the importance of imbuing more resource-rational meta-reasoning in language and reasoning models.
comment: Pre-print
☆ GapDNER: A Gap-Aware Grid Tagging Model for Discontinuous Named Entity Recognition IJCNN 2025
In biomedical fields, one named entity may consist of a series of non-adjacent tokens and overlap with other entities. Previous methods recognize discontinuous entities by connecting entity fragments or internal tokens, which face challenges of error propagation and decoding ambiguity due to the wide variety of span or word combinations. To address these issues, we deeply explore discontinuous entity structures and propose an effective Gap-aware grid tagging model for Discontinuous Named Entity Recognition, named GapDNER. Our GapDNER innovatively applies representation learning on the context gaps between entity fragments to resolve decoding ambiguity and enhance discontinuous NER performance. Specifically, we treat the context gap as an additional type of span and convert span classification into a token-pair grid tagging task. Subsequently, we design two interactive components to comprehensively model token-pair grid features from both intra- and inter-span perspectives. The intra-span regularity extraction module employs the biaffine mechanism along with linear attention to capture the internal regularity of each span, while the inter-span relation enhancement module utilizes criss-cross attention to obtain semantic relations among different spans. At the inference stage of entity decoding, we assign a directed edge to each entity fragment and context gap, then use the BFS algorithm to search for all valid paths from the head to tail of grids with entity tags. Experimental results on three datasets demonstrate that our GapDNER achieves new state-of-the-art performance on discontinuous NER and exhibits remarkable advantages in recognizing complex entity structures.
comment: Accepted by IJCNN 2025
☆ Find Your Optimal Teacher: Personalized Data Synthesis via Router-Guided Multi-Teacher Distillation
Training student models on synthetic data generated by strong teacher models is a promising way to distilling the capabilities of teachers. However, recent studies show that stronger models are not always optimal teachers, revealing a mismatch between teacher outputs and student learnability. To address this issue, we propose PerSyn (Personalized data Synthesis), a novel synthesis strategy that operates under a new ``Route then Generate'' paradigm to create data tailored to each student model, enabling it to learn more effectively. Specifically, PerSyn first assigns each prompt to its optimal teacher via a query-level router that jointly considers student learnability and teacher response quality. Each teacher then synthesizes data only for its assigned prompts, making the process more efficient than the conventional ``Generate then Select'' paradigm, where all teachers must generate parallel responses for the entire prompt set before constructing the final dataset. Extensive experiments across different model families and scales demonstrate that PerSyn consistently achieves superior or comparable performance to all baselines in instruct tuning and math reasoning settings. Further analysis verifies the effectiveness of PerSyn and offers extra insights to propel future research.
comment: 19 pages, 10 figures
☆ ADVICE: Answer-Dependent Verbalized Confidence Estimation
Recent progress in large language models (LLMs) has enabled them to express their confidence in natural language, enhancing transparency and reliability. However, their confidence often exhibits overconfidence, the cause of which remains poorly understood. In this work, we conduct a detailed analysis of the dynamics underlying verbalized confidence and identify answer-independence as a key factor, defined as the model's failure to condition confidence on its own answer. To address this, we propose ADVICE (Answer-Dependent Verbalized Confidence Estimation), a fine-tuning framework that facilitates answer-grounded confidence estimation. Extensive experiments show that ADVICE substantially improves confidence calibration while preserving task performance. Further analyses confirm that ADVICE strengthens answer-groundedness, leading to more balanced and well-calibrated confidence distributions. Our findings shed light on the origin of overconfidence and establish a framework for more trustworthy confidence verbalization.
☆ LLM$\times$MapReduce-V3: Enabling Interactive In-Depth Survey Generation through a MCP-Driven Hierarchically Modular Agent System EMNLP2025
We introduce LLM x MapReduce-V3, a hierarchically modular agent system designed for long-form survey generation. Building on the prior work, LLM x MapReduce-V2, this version incorporates a multi-agent architecture where individual functional components, such as skeleton initialization, digest construction, and skeleton refinement, are implemented as independent model-context-protocol (MCP) servers. These atomic servers can be aggregated into higher-level servers, creating a hierarchically structured system. A high-level planner agent dynamically orchestrates the workflow by selecting appropriate modules based on their MCP tool descriptions and the execution history. This modular decomposition facilitates human-in-the-loop intervention, affording users greater control and customization over the research process. Through a multi-turn interaction, the system precisely captures the intended research perspectives to generate a comprehensive skeleton, which is then developed into an in-depth survey. Human evaluations demonstrate that our system surpasses representative baselines in both content depth and length, highlighting the strength of MCP-based modular planning.
comment: Accepted by EMNLP2025 System Demonstration
☆ Rethinking Agentic Workflows: Evaluating Inference-Based Test-Time Scaling Strategies in Text2SQL Tasks SC
Large language models (LLMs) are increasingly powering Text-to-SQL (Text2SQL) systems, enabling non-expert users to query industrial databases using natural language. While test-time scaling strategies have shown promise in LLM-based solutions, their effectiveness in real-world applications, especially with the latest reasoning models, remains uncertain. In this work, we benchmark six lightweight, industry-oriented test-time scaling strategies and four LLMs, including two reasoning models, evaluating their performance on the BIRD Mini-Dev benchmark. Beyond standard accuracy metrics, we also report inference latency and token consumption, providing insights relevant for practical system deployment. Our findings reveal that Divide-and-Conquer prompting and few-shot demonstrations consistently enhance performance for both general-purpose and reasoning-focused LLMs. However, introducing additional workflow steps yields mixed results, and base model selection plays a critical role. This work sheds light on the practical trade-offs between accuracy, efficiency, and complexity when deploying Text2SQL systems.
comment: Accepted at COLM 2025 SCALR Workshop
☆ Generate Logical Equivalence Questions
Academic dishonesty is met with zero tolerance in higher education, yet plagiarism has become increasingly prevalent in the era of online teaching and learning. Automatic Question Generation (AQG) presents a potential solution to mitigate copying by creating unique questions for each student. Additionally, AQG can provide a vast array of practice questions. Our AQG focuses on generating logical equivalence questions for Discrete Mathematics, a foundational course for first-year computer science students. A literature review reveals that existing AQGs for this type of question generate all propositions that meet user-defined constraints, resulting in inefficiencies and a lack of uniform question difficulty. To address this, we propose a new approach that defines logical equivalence questions using a formal language, translates this language into two sets of generation rules, and develops a linear-time algorithm for question generation. We evaluated our AQG through two experiments. The first involved a group of students completing questions generated by our system. Statistical analysis shows that the accuracy of these questions is comparable to that of textbook questions. The second experiment assessed the number of steps required to solve our generated questions, textbook questions, and those generated by multiple large language models. The results indicated that the difficulty of our questions was similar to that of textbook questions, confirming the quality of our AQG.
☆ UALM: Unified Audio Language Model for Understanding, Generation and Reasoning
Recent advances in the audio language modeling (ALM) domain tackle audio understanding and text-to-audio generation as separate tasks. Very few studies attempt to unify these tasks -- an essential step toward advanced multimodal reasoning. This paper introduces U}nified Audio Language Model (UALM), which aims to unify audio understanding, text-to-audio generation, and multimodal reasoning in a single model. To achieve this goal, we first present UALM-Gen, a text-to-audio language model that directly predicts audio tokens and is comparable to state-of-the-art diffusion-based models. We then demonstrate, using proper data blending, training recipes, and inference techniques, that our single UALM model matches the quality of state-of-the-art specialized models in audio understanding, text-to-audio generation, and text reasoning. Furthermore, we present UALM-Reason, a multimodal reasoning model that utilizes both text and audio in the intermediate thinking steps to facilitate complex generation tasks. To our knowledge, this is the first demonstration in audio research of cross-modal generative reasoning, with its effectiveness confirmed by subjective evaluations.
☆ SAGE: A Top-Down Bottom-Up Knowledge-Grounded User Simulator for Multi-turn AGent Evaluation
Evaluating multi-turn interactive agents is challenging due to the need for human assessment. Evaluation with simulated users has been introduced as an alternative, however existing approaches typically model generic users and overlook the domain-specific principles required to capture realistic behavior. We propose SAGE, a novel user Simulation framework for multi-turn AGent Evaluation that integrates knowledge from business contexts. SAGE incorporates top-down knowledge rooted in business logic, such as ideal customer profiles, grounding user behavior in realistic customer personas. We further integrate bottom-up knowledge taken from business agent infrastructure (e.g., product catalogs, FAQs, and knowledge bases), allowing the simulator to generate interactions that reflect users' information needs and expectations in a company's target market. Through empirical evaluation, we find that this approach produces interactions that are more realistic and diverse, while also identifying up to 33% more agent errors, highlighting its effectiveness as an evaluation tool to support bug-finding and iterative agent improvement.
☆ Conjecturing: An Overlooked Step in Formal Mathematical Reasoning
Autoformalisation, the task of expressing informal mathematical statements in formal language, is often viewed as a direct translation process. This, however, disregards a critical preceding step: conjecturing. Many mathematical problems cannot be formalised directly without first conjecturing a conclusion such as an explicit answer, or a specific bound. Since Large Language Models (LLMs) already struggle with autoformalisation, and the evaluation of their conjecturing ability is limited and often entangled within autoformalisation or proof, it is particularly challenging to understand its effect. To address this gap, we augment existing datasets to create ConjectureBench, and redesign the evaluation framework and metric specifically to measure the conjecturing capabilities of LLMs both as a distinct task and within the autoformalisation pipeline. Our evaluation of foundational models, including GPT-4.1 and DeepSeek-V3.1, reveals that their autoformalisation performance is substantially overestimated when the conjecture is accounted for during evaluation. However, the conjecture should not be assumed to be provided. We design an inference-time method, Lean-FIRe to improve conjecturing and autoformalisation, which, to the best of our knowledge, achieves the first successful end-to-end autoformalisation of 13 PutnamBench problems with GPT-4.1 and 7 with DeepSeek-V3.1. We demonstrate that while LLMs possess the requisite knowledge to generate accurate conjectures, improving autoformalisation performance requires treating conjecturing as an independent task, and investigating further how to correctly integrate it within autoformalisation. Finally, we provide forward-looking guidance to steer future research toward improving conjecturing, an overlooked step of formal mathematical reasoning.
☆ Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation
AI agents have been developed for complex real-world tasks from coding to customer service. But AI agent evaluations suffer from many challenges that undermine our understanding of how well agents really work. We introduce the Holistic Agent Leaderboard (HAL) to address these challenges. We make three main contributions. First, we provide a standardized evaluation harness that orchestrates parallel evaluations across hundreds of VMs, reducing evaluation time from weeks to hours while eliminating common implementation bugs. Second, we conduct three-dimensional analysis spanning models, scaffolds, and benchmarks. We validate the harness by conducting 21,730 agent rollouts across 9 models and 9 benchmarks in coding, web navigation, science, and customer service with a total cost of about $40,000. Our analysis reveals surprising insights, such as higher reasoning effort reducing accuracy in the majority of runs. Third, we use LLM-aided log inspection to uncover previously unreported behaviors, such as searching for the benchmark on HuggingFace instead of solving a task, or misusing credit cards in flight booking tasks. We share all agent logs, comprising 2.5B tokens of language model calls, to incentivize further research into agent behavior. By standardizing how the field evaluates agents and addressing common pitfalls in agent evaluation, we hope to shift the focus from agents that ace benchmarks to agents that work reliably in the real world.
☆ Scaling Long-Horizon LLM Agent via Context-Folding
Large language model (LLM) agents are fundamentally constrained by context length on long-horizon tasks. We introduce Context-Folding, a framework that empowers agents to actively manage their working context. An agent can procedurally branch into a sub-trajectory to handle a subtask and then fold it upon completion, collapsing the intermediate steps while retaining a concise summary of the outcome. To make this behavior learnable, we develop an end-to-end reinforcement learning framework FoldGRPO with specific process rewards to encourage effective task decomposition and context management. On complex long-horizon tasks (Deep Research and SWE), our folding agent matches or outperforms the ReAct baselines while using an active context 10$\times$ smaller and significantly outperforms models that rely on summarization-based context management.
☆ Direct Multi-Token Decoding
Decoder-only transformers have become the standard architecture for large language models (LLMs) due to their strong performance. Recent studies suggest that, in pre-trained LLMs, early, middle, and late layers may serve distinct roles: Early layers focus on understanding the input context, middle layers handle task-specific processing, and late layers convert abstract representations into output tokens. We hypothesize that once representations have been processed by the early and middle layers, the resulting hidden states may encapsulate sufficient information to support the generation of multiple tokens using only the late layers, eliminating the need to repeatedly traverse the early and middle layers. We refer to this inference paradigm as Direct Multi-Token Decoding (DMTD). Unlike speculative decoding, our method introduces no additional parameters, auxiliary routines, or post-generation verification. Despite being trained on a limited dataset, a fine-tuned DMTD Qwen3-4B model has already demonstrated promising results, achieving up to a 2x speedup with only minor performance loss. Moreover, as shown in our scaling analysis, its performance is expected to further improve with larger training datasets.
☆ Evaluating Retrieval-Augmented Generation Systems on Unanswerable, Uncheatable, Realistic, Multi-hop Queries
Real-world use cases often present RAG systems with complex queries for which relevant information is missing from the corpus or is incomplete. In these settings, RAG systems must be able to reject unanswerable, out-of-scope queries and identify failures of retrieval and multi-hop reasoning. Despite this, existing RAG benchmarks rarely reflect realistic task complexity for multi-hop or out-of-scope questions, which often can be cheated via disconnected reasoning (i.e., solved without genuine multi-hop inference) or require only simple factual recall. This limits the ability for such benchmarks to uncover limitations of existing RAG systems. To address this gap, we present the first pipeline for automatic, difficulty-controlled creation of un$\underline{c}$heatable, $\underline{r}$ealistic, $\underline{u}$nanswerable, and $\underline{m}$ulti-hop $\underline{q}$uerie$\underline{s}$ (CRUMQs), adaptable to any corpus and domain. We use our pipeline to create CRUMQs over two popular RAG datasets and demonstrate its effectiveness via benchmark experiments on leading retrieval-augmented LLMs. Results show that compared to prior RAG benchmarks, CRUMQs are highly challenging for RAG systems and achieve up to 81.0\% reduction in cheatability scores. More broadly, our pipeline offers a simple way to enhance benchmark difficulty and realism and drive development of more capable RAG systems.
☆ GRAVITY: A Framework for Personalized Text Generation via Profile-Grounded Synthetic Preferences
Personalization in LLMs often relies on costly human feedback or interaction logs, limiting scalability and neglecting deeper user attributes. To reduce the reliance on human annotations, we introduce GRAVITY (Generative Response with Aligned Values, Interests, and Traits of You), a framework for generating synthetic, profile-grounded preference data that captures users' interests, values, beliefs, and personality traits. By integrating demographic, cultural, and psychological frameworks -- including Hofstede's cultural dimensions, Schwartz's basic values, the World Values Survey, and Big Five OCEAN traits -- GRAVITY synthesizes preference pairs to guide personalized content generation. We evaluate GRAVITY on book descriptions for 400 Amazon users, comparing it to prompt-based conditioning, standard fine-tuning, and naive synthetic pair generation. Profile-grounded synthetic data consistently improves generation, especially across multiple cultures (USA, Brazil, Japan, India), achieving over 4% higher preference gains across baselines, with user studies showing that GRAVITY outputs are preferred over 86% of the time. Our results show that scenario-grounded synthetic data can capture richer user variation, reduce reliance on costly annotation, and produce more engaging, user-centered content, offering a scalable path for LLM personalization.
☆ TopoAlign: A Framework for Aligning Code to Math via Topological Decomposition
Large Language Models (LLMs) excel at both informal and formal (e.g. Lean 4) mathematical reasoning but still struggle with autoformalisation, the task of transforming informal into formal mathematical statements. Autoformalisation helps pair the informal reasoning of LLMs with formal proof assistants which enable machine-verifiable generation and mitigate hallucinations. Yet, the performance of current Math LLMs is constrained by the scarcity of large-scale corpora, particularly those containing pairs of informal and formal statements. Although current models are trained to generate code from natural language instructions, structural and syntactic differences between these and formal mathematics limit effective transfer learning. We propose TopoAlign, a framework that unlocks widely available code repositories as training resources for Math LLMs. TopoAlign decomposes code into docstrings, main functions, and dependency functions, and reassembles these components into analogues that structurally mirror formal statements. This produces structurally aligned code data that can be used for training Math LLMs without requiring additional human annotation. We train two state-of-the-art models, DeepSeek-Math and Herald, and evaluate them on the minif2f, Putnam, and ProofNet benchmarks. TopoAlign provides substantial gains for DeepSeek-Math, improving performance by 17.77% on BEq@10 and 68.82% on typecheck@10. Despite introducing no new mathematical knowledge, our framework achieves gains of 0.12% and 1.09% for Herald on BEq@10 and typecheck@10, respectively, demonstrating that training on aligned code data is beneficial even for specialized models.
Let's Reason Formally: Natural-Formal Hybrid Reasoning Enhances LLM's Math Capability
Enhancing the mathematical reasoning capabilities of LLMs has garnered significant attention in both the mathematical and computer science communities. Recent works have made substantial progress in both Natural Language (NL) reasoning and Formal Language (FL) reasoning by leveraging the potential of pure Reinforcement Learning (RL) methods on base models. However, RL approaches struggle to impart new capabilities not presented in the base model, highlighting the need to integrate more knowledge like FL into NL math reasoning effectively. Yet, this integration is challenging due to inherent disparities in problem structure and reasoning format between NL and FL. To address these challenges, we introduce **NL-FL HybridReasoning (NFL-HR)**, an end-to-end framework designed to incorporate the FL expert into NL math problem-solving. To bridge the NL and FL input format gap, we propose the NL-FL Problem Alignment method, which reformulates the Question-Answering (QA) problems in NL as existence theorems in FL. Subsequently, the Mixed Problem Input technique we provide enables the FL reasoner to handle both QA and existence problems concurrently. Lastly, we mitigate the NL and FL output format gap in reasoning through an LLM-based Answer Extraction mechanism. Comprehensive experiments demonstrate that the NFL-HR framework achieves **89.80**% and **84.34%** accuracy rates on the MATH-500 and the AMC benchmarks, surpassing the NL baseline by **4.60%** and **4.82%**, respectively. Notably, some problems resolved by our framework remain unsolved by the NL baseline model even under a larger number of trials.
♻ ☆ Part-of-speech tagging for Nagamese Language using CRF
This paper investigates part-of-speech tagging, an important task in Natural Language Processing (NLP) for the Nagamese language. The Nagamese language, a.k.a. Naga Pidgin, is an Assamese-lexified Creole language developed primarily as a means of communication in trade between the Nagas and people from Assam in northeast India. A substantial amount of work in part-of-speech-tagging has been done for resource-rich languages like English, Hindi, etc. However, no work has been done in the Nagamese language. To the best of our knowledge, this is the first attempt at part-of-speech tagging for the Nagamese Language. The aim of this work is to identify the part-of-speech for a given sentence in the Nagamese language. An annotated corpus of 16,112 tokens is created and applied machine learning technique known as Conditional Random Fields (CRF). Using CRF, an overall tagging accuracy of 85.70%; precision, recall of 86%, and f1-score of 85% is achieved. Keywords. Nagamese, NLP, part-of-speech, machine learning, CRF.
comment: 8 pages
♻ ☆ Talk Isn't Always Cheap: Understanding Failure Modes in Multi-Agent Debate ICML
While multi-agent debate has been proposed as a promising strategy for improving AI reasoning ability, we find that debate can sometimes be harmful rather than helpful. Prior work has primarily focused on debates within homogeneous groups of agents, whereas we explore how diversity in model capabilities influences the dynamics and outcomes of multi-agent interactions. Through a series of experiments, we demonstrate that debate can lead to a decrease in accuracy over time - even in settings where stronger (i.e., more capable) models outnumber their weaker counterparts. Our analysis reveals that models frequently shift from correct to incorrect answers in response to peer reasoning, favoring agreement over challenging flawed reasoning. We perform additional experiments investigating various potential contributing factors to these harmful shifts - including sycophancy, social conformity, and model and task type. These results highlight important failure modes in the exchange of reasons during multi-agent debate, suggesting that naive applications of debate may cause performance degradation when agents are neither incentivised nor adequately equipped to resist persuasive but incorrect reasoning.
comment: ICML MAS Workshop 2025
♻ ☆ Multi-Scale Manifold Alignment for Interpreting Large Language Models: A Unified Information-Geometric Framework
We present Multi-Scale Manifold Alignment(MSMA), an information-geometric framework that decomposes LLM representations into local, intermediate, and global manifolds and learns cross-scale mappings that preserve geometry and information. Across GPT-2, BERT, RoBERTa, and T5, we observe consistent hierarchical patterns and find that MSMA improves alignment metrics under multiple estimators (e.g., relative KL reduction and MI gains with statistical significance across seeds). Controlled interventions at different scales yield distinct and architecture-dependent effects on lexical diversity, sentence structure, and discourse coherence. While our theoretical analysis relies on idealized assumptions, the empirical results suggest that multi-objective alignment offers a practical lens for analyzing cross-scale information flow and guiding representation-level control.
♻ ☆ Holistic Evaluation of Multimodal LLMs on Spatial Intelligence
Multimodal models have achieved remarkable progress in recent years. Nevertheless, they continue to exhibit notable limitations in spatial understanding and reasoning, the very capability that anchors artificial general intelligence in the physical world. With the recent release of GPT-5, allegedly the most powerful AI model to date, it is timely to examine where the leading models (GPT, Gemini, Grok, Seed, Qwen, and Intern) stand on the path toward spatial intelligence. We first propose a holistic taxonomy of spatial tasks that unifies existing benchmarks and a standardized protocol for the fair evaluation of state-of-the-art proprietary and open-source models across eight key benchmarks, at a cost exceeding ten billion total tokens. Our empirical study then reveals that (1) GPT-5 demonstrates unprecedented strength in spatial intelligence (SI), yet (2) still falls short of human performance significantly across a broad spectrum of SI-tasks. Moreover, we (3) show that SI-tasks expose greater model capability deficiency than non-SI tasks, to the extent that (4) proprietary models do not exhibit a decisive advantage when facing the most difficult ones. In addition, we conduct a qualitative evaluation across a diverse set of scenarios that are intuitive for humans, yet fail even the most advanced multimodal models.
♻ ☆ Empirical Investigation of Latent Representational Dynamics in Large Language Models: A Manifold Evolution Perspective
This paper introduces the Dynamical Manifold Evolution Theory (DMET), a conceptual framework that models large language model (LLM) generation as a continuous trajectory evolving on a low-dimensional semantic manifold. The theory characterizes latent dynamics through three interpretable metrics-state continuity ($C$), attractor compactness ($Q$), and topological persistence ($P$)-which jointly capture the smoothness, stability, and structure of representation evolution. Empirical analyses across multiple Transformer architectures reveal consistent links between these latent dynamics and text quality: smoother trajectories correspond to greater fluency, and richer topological organization correlates with enhanced coherence. Different models exhibit distinct dynamical regimes, reflecting diverse strategies of semantic organization in latent space. Moreover, decoding parameters such as temperature and top-$p$ shape these trajectories in predictable ways, defining a balanced region that harmonizes fluency and creativity. As a phenomenological rather than first-principles framework, DMET provides a unified and testable perspective for interpreting, monitoring, and guiding LLM behavior, offering new insights into the interplay between internal representation dynamics and external text generation quality.
♻ ☆ Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot EMNLP25
In-Context Learning (ICL) is an essential emergent ability of Large Language Models (LLMs), and recent studies introduce Chain-of-Thought (CoT) to exemplars of ICL to enhance the reasoning capability, especially in mathematics tasks. However, given the continuous advancement of model capabilities, it remains unclear whether CoT exemplars still benefit recent, stronger models in such tasks. Through systematic experiments, we find that for recent strong models such as the Qwen2.5 series, adding traditional CoT exemplars does not improve reasoning performance compared to Zero-Shot CoT. Instead, their primary function is to align the output format with human expectations. We further investigate the effectiveness of enhanced CoT exemplars, constructed using answers from advanced models such as \texttt{Qwen2.5-Max} and \texttt{DeepSeek-R1}. Experimental results indicate that these enhanced exemplars still fail to improve the model's reasoning performance. Further analysis reveals that models tend to ignore the exemplars and focus primarily on the instructions, leading to no observable gain in reasoning ability. Overall, our findings highlight the limitations of the current ICL+CoT framework in mathematical reasoning, calling for a re-examination of the ICL paradigm and the definition of exemplars.
comment: EMNLP25-findings camera_ready, 19 pages,22 figures
♻ ☆ ViDRiP-LLaVA: A Dataset and Benchmark for Diagnostic Reasoning from Pathology Videos
We present ViDRiP-LLaVA, the first large multimodal model (LMM) in computational pathology that integrates three distinct image scenarios, including single patch images, automatically segmented pathology video clips, and manually segmented pathology videos. This integration closely mirrors the natural diagnostic process of pathologists. By generating detailed histological descriptions and culminating in a definitive sign-out diagnosis, ViDRiP-LLaVA bridges visual narratives with diagnostic reasoning. Central to our approach is the ViDRiP-Instruct dataset, comprising 4278 video and diagnosis-specific chain-of-thought instructional pairs sourced from educational histopathology videos on YouTube. Although high-quality data is critical for enhancing diagnostic reasoning, its creation is time-intensive and limited in volume. To overcome this challenge, we transfer knowledge from existing single-image instruction datasets to train on weakly annotated, keyframe-extracted clips, followed by fine-tuning on manually segmented videos. ViDRiP-LLaVA establishes a new benchmark in pathology video analysis and offers a promising foundation for future AI systems that support clinical decision-making through integrated visual and diagnostic reasoning. Our code, data, and model are publicly available at: https://github.com/QuIIL/ViDRiP-LLaVA.
♻ ☆ Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated notable success in enhancing the reasoning performance of large language models (LLMs), particularly on mathematics and programming tasks. Similar to how traditional RL helps agents explore and learn new strategies, RLVR is believed to enable LLMs to continuously self-improve, thus acquiring novel reasoning abilities beyond those of the corresponding base models. In this study we critically examine the current state of RLVR by systematically probing the reasoning capability boundaries of RLVR-trained LLMs across various model families, RL algorithms, and math, coding, and visual reasoning benchmarks, using pass@k at large k values as the evaluation metric. Surprisingly, we find that the current training setup does not elicit fundamentally new reasoning patterns. While RLVR-trained models outperform their base models at small k (e.g., k = 1), the base models achieve a higher pass@k score when k is large. Coverage and perplexity analyses show that the observed reasoning abilities originate from and are bounded by the base model. Treating the base model as an upper bound, our quantitative analysis shows that six popular RLVR algorithms perform similarly and remain far from optimal in leveraging the potential of the base model. By contrast, we find that distillation can introduce new reasoning patterns from the teacher and genuinely expand the model's reasoning capabilities. Overall, our findings suggest that current RLVR methods have not yet realized the potential of RL to elicit truly novel reasoning abilities in LLMs. This highlights the need for improved RL paradigms, such as continual scaling and multi-turn agent-environment interaction, to unlock this potential.
comment: 30 pages, 27 figures
♻ ☆ Evaluating LLMs for Demographic-Targeted Social Bias Detection: A Comprehensive Benchmark Study
Large-scale web-scraped text corpora used to train general-purpose AI models often contain harmful demographic-targeted social biases, creating a regulatory need for data auditing and developing scalable bias-detection methods. Although prior work has investigated biases in text datasets and related detection methods, these studies remain narrow in scope. They typically focus on a single content type (e.g., hate speech), cover limited demographic axes, overlook biases affecting multiple demographics simultaneously, and analyze limited techniques. Consequently, practitioners lack a holistic understanding of the strengths and limitations of recent large language models (LLMs) for automated bias detection. In this study, we present a comprehensive evaluation framework aimed at English texts to assess the ability of LLMs in detecting demographic-targeted social biases. To align with regulatory requirements, we frame bias detection as a multi-label task using a demographic-focused taxonomy. We then conduct a systematic evaluation with models across scales and techniques, including prompting, in-context learning, and fine-tuning. Using twelve datasets spanning diverse content types and demographics, our study demonstrates the promise of fine-tuned smaller models for scalable detection. However, our analyses also expose persistent gaps across demographic axes and multi-demographic targeted biases, underscoring the need for more effective and scalable auditing frameworks.
comment: 18 pages
♻ ☆ LADM: Long-context Training Data Selection with Attention-based Dependency Measurement for LLMs ACL 2025
Long-context modeling has drawn more and more attention in the area of Large Language Models (LLMs). Continual training with long-context data becomes the de-facto method to equip LLMs with the ability to process long inputs. However, it still remains an open challenge to measure the quality of long-context training data. To address this issue, we propose a Long-context data selection framework with Attention-based Dependency Measurement (LADM), which can efficiently identify high-quality long-context data from a large-scale, multi-domain pre-training corpus. LADM leverages the retrieval capabilities of the attention mechanism to capture contextual dependencies, ensuring a comprehensive quality measurement of long-context data. Experimental results show that our LADM framework significantly boosts the performance of LLMs on multiple long-context tasks with only 1B tokens for continual training.
comment: ACL 2025, our code is available at https://github.com/ZNLP/LADM
♻ ☆ Evolving LLMs' Self-Refinement Capability via Synergistic Training-Inference Optimization
Self-Refinement refers to a model's ability to revise its own responses to produce improved outputs. This capability can also serve as a fundamental mechanism for Self-Improvement, for example, by reconstructing datasets with refined results to enhance intrinsic model performance. However, our comprehensive experiments reveal that large language models (LLMs) show no clear evidence of inherent Self-Refinement and may even experience response quality degradation after Self-Refinement. To address this issue, we propose EVOLVE, a simple and effective framework for eliciting and tracking the evolution of Self-Refinement through iterative training. We first explore optimization methods during training to activate the model's Self-Refinement capability. Then, at inference, we investigate various generation strategies to further enhance and utilize Self-Refinement while supplying the necessary data for training. Through synergistic optimization of training and inference stages, we continually evolve the model's Self-Refinement ability, enabling it to better refine its own responses. Moreover, we demonstrate the potential of leveraging Self-Refinement to achieve broader Self-Improvement of intrinsic model abilities. Experiments show that the evolved Self-Refinement ability enables the Llama-3.1-8B base model to surpass GPT-4o, achieving 62.3% length-controlled and 63.3% raw win rates on AlpacaEval 2, and 50.3% on Arena-Hard. It also generalizes effectively to out-of-domain reasoning tasks, improving performance on mathematical reasoning benchmarks such as GSM8K and MATH.
♻ ☆ NeMo: Needle in a Montage for Video-Language Understanding
Recent advances in video large language models (VideoLLMs) call for new evaluation protocols and benchmarks for complex temporal reasoning in video-language understanding. Inspired by the needle in a haystack test widely used by LLMs, we introduce a novel task of Needle in a Montage (NeMo), designed to assess VideoLLMs' critical reasoning capabilities, including long-context recall and temporal grounding. To generate video question answering data for our task, we develop a scalable automated data generation pipeline that facilitates high-quality data synthesis. Built upon the proposed pipeline, we present NeMoBench, a video-language benchmark centered on our task. Specifically, our full set of NeMoBench features 31,378 automatically generated question-answer (QA) pairs from 13,486 videos with various durations ranging from seconds to hours. Experiments demonstrate that our pipeline can reliably and automatically generate high-quality evaluation data, enabling NeMoBench to be continuously updated with the latest videos. We evaluate 20 state-of-the-art models on our benchmark, providing extensive results and key insights into their capabilities and limitations. Our project page is available at: https://lavi-lab.github.io/NeMoBench.
♻ ☆ On the Interplay between Musical Preferences and Personality through the Lens of Language ECAI2025
Music serves as a powerful reflection of individual identity, often aligning with deeper psychological traits. Prior research has established correlations between musical preferences and personality, while separate studies have demonstrated that personality is detectable through linguistic analysis. Our study bridges these two research domains by investigating whether individuals' musical preferences leave traces in their spontaneous language through the lens of the Big Five personality traits (Openness, Conscientiousness, Extroversion, Agreeableness, and Neuroticism). Using a carefully curated dataset of over 500,000 text samples from nearly 5,000 authors with reliably identified musical preferences, we build advanced models to assess personality characteristics. Our results reveal significant personality differences across fans of five musical genres. We release resources for future research at the intersection of computational linguistics, music psychology and personality analysis.
comment: ECAI2025 (Identity-Aware AI workshop)
♻ ☆ Effectiveness of Counter-Speech against Abusive Content: A Multidimensional Annotation and Classification Study
Counter-speech (CS) is a key strategy for mitigating online Hate Speech (HS), yet defining the criteria to assess its effectiveness remains an open challenge. We propose a novel computational framework for CS effectiveness classification, grounded in linguistics, communication and argumentation concepts. Our framework defines six core dimensions - Clarity, Evidence, Emotional Appeal, Rebuttal, Audience Adaptation, and Fairness - which we use to annotate 4,214 CS instances from two benchmark datasets, resulting in a novel linguistic resource released to the community. In addition, we propose two classification strategies, multi-task and dependency-based, achieving strong results (0.94 and 0.96 average F1 respectively on both expert- and user-written CS), outperforming standard baselines, and revealing strong interdependence among dimensions.
♻ ☆ From $f(x)$ and $g(x)$ to $f(g(x))$: LLMs Learn New Skills in RL by Composing Old Ones
Does RL teach LLMs genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL even without preceding supervised finetuning; on the other, critics argue that RL contributes little beyond reweighting existing reasoning strategies. This work provides concrete evidence that LLMs can acquire genuinely new skills during RL by composing existing ones, mirroring one of the central mechanisms by which humans acquire new cognitive skills. To mitigate data contamination and other confounding factors, and to allow precise control over task complexity, we develop a synthetic framework for our investigation. Specifically, we define a skill as the ability to infer the output of a string transformation function f(x) given x. When an LLM has already learned f and g prior to RL, our experiments reveal that RL enables it to learn unseen compositions of them h(x)=g(f(x)). Further, this compositional ability generalizes to more difficult problems such as compositions of >2 functions unseen during RL training. Surprisingly, our experiments show that compositional skill acquired on a source task transfers to a different target task. This transfer happens even without compositional training on the target, requiring only prior knowledge of the target's atomic skills. Our qualitative analysis shows that RL fundamentally changes the reasoning behaviors of the models. In contrast, next-token training with the same data yields none of these findings. Our systematic experiments provide fresh insights into LLM learning, suggesting the value of first building base models with basic skills, then using RL to incentivize advanced, generalizable skills for complex problems.
♻ ☆ WebThinker: Empowering Large Reasoning Models with Deep Research Capability NeurIPS 2025
Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose WebThinker, a deep research agent that empowers LRMs to autonomously search the web, navigate among web pages, and draft reports during the reasoning process. WebThinker integrates a Deep Web Explorer module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an Autonomous Think-Search-and-Draft strategy, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an RL-based training strategy via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.
comment: Accepted by NeurIPS 2025
♻ ☆ Formalizing Style in Personal Narratives
Personal narratives are stories authors construct to make meaning of their experiences. Style, the distinctive way authors use language to express themselves, is fundamental to how these narratives convey subjective experiences. Yet there is a lack of a formal framework for systematically analyzing these stylistic choices. We present a novel approach that formalizes style in personal narratives as patterns in the linguistic choices authors make when communicating subjective experiences. Our framework integrates three domains: functional linguistics establishes language as a system of meaningful choices, computer science provides methods for automatically extracting and analyzing sequential patterns, and these patterns are linked to psychological observations. Using language models, we automatically extract linguistic features such as processes, participants, and circumstances. We apply our framework to hundreds of dream narratives, including a case study on a war veteran with post-traumatic stress disorder. Analysis of his narratives uncovers distinctive patterns, particularly how verbal processes dominate over mental ones, illustrating the relationship between linguistic choices and psychological states.
♻ ☆ The Enemy from Within: A Study of Political Delegitimization Discourse in Israeli Political Speech EMNLP 2025
We present the first large-scale computational study of political delegitimization discourse (PDD), defined as symbolic attacks on the normative validity of political entities. We curate and manually annotate a novel Hebrew-language corpus of 10,410 sentences drawn from Knesset speeches (1993-2023), Facebook posts (2018-2021), and leading news outlets, of which 1,812 instances (17.4\%) exhibit PDD and 642 carry additional annotations for intensity, incivility, target type, and affective framing. We introduce a two-stage classification pipeline combining finetuned encoder models and decoder LLMs. Our best model (DictaLM 2.0) attains an F$_1$ of 0.74 for binary PDD detection and a macro-F$_1$ of 0.67 for classification of delegitimization characteristics. Applying this classifier to longitudinal and cross-platform data, we see a marked rise in PDD over three decades, higher prevalence on social media versus parliamentary debate, greater use by male than female politicians, and stronger tendencies among right-leaning actors - with pronounced spikes during election campaigns and major political events. Our findings demonstrate the feasibility and value of automated PDD analysis for understanding democratic discourse.
comment: EMNLP 2025
♻ ☆ Agentic large language models improve retrieval-based radiology question answering
Clinical decision-making in radiology increasingly benefits from artificial intelligence (AI), particularly through large language models (LLMs). However, traditional retrieval-augmented generation (RAG) systems for radiology question answering (QA) typically rely on single-step retrieval, limiting their ability to handle complex clinical reasoning tasks. Here we propose radiology Retrieval and Reasoning (RaR), a multi-step retrieval and reasoning framework designed to improve diagnostic accuracy, factual consistency, and clinical reliability of LLMs in radiology question answering. We evaluated 25 LLMs spanning diverse architectures, parameter scales (0.5B to >670B), and training paradigms (general-purpose, reasoning-optimized, clinically fine-tuned), using 104 expert-curated radiology questions from previously established RSNA-RadioQA and ExtendedQA datasets. To assess generalizability, we additionally tested on an unseen internal dataset of 65 real-world radiology board examination questions. RaR significantly improved mean diagnostic accuracy over zero-shot prompting and conventional online RAG. The greatest gains occurred in small-scale models, while very large models (>200B parameters) demonstrated minimal changes (<2% improvement). Additionally, RaR retrieval reduced hallucinations (mean 9.4%) and retrieved clinically relevant context in 46% of cases, substantially aiding factual grounding. Even clinically fine-tuned models showed gains from RaR (e.g., MedGemma-27B), indicating that retrieval remains beneficial despite embedded domain knowledge. These results highlight the potential of RaR to enhance factuality and diagnostic accuracy in radiology QA, warranting future studies to validate their clinical utility. All datasets, code, and the full RaR framework are publicly available to support open research and clinical translation.
♻ ☆ J1: Incentivizing Thinking in LLM-as-a-Judge via Reinforcement Learning
The progress of AI is bottlenecked by the quality of evaluation, making powerful LLM-as-a-Judge models a core solution. The efficacy of these judges depends on their chain-of-thought reasoning, creating a critical need for methods that can effectively optimize this reasoning process. In this work, we introduce J1, a reinforcement learning framework for teaching LLM judges to think before making decisions. Our core contribution lies in converting all judgment tasks for non-verifiable and verifiable prompts into a unified format with verifiable rewards, enabling direct optimization of evaluation quality while mitigating positional bias. We then use RL to train thinking-judges at scales of 8B, 32B, and 70B and show that they obtain state-of-the-art performance across multiple benchmarks. In particular, J1-Qwen-32B, our multitasked pointwise and pairwise judge also outperforms o1-mini, o3, and a much larger 671B DeepSeek-R1 on some benchmarks, while only training on synthetic data. Through comprehensive ablations of pairwise, pointwise, and multitask J1 variants, we demonstrate the effectiveness of our approach across seed prompts, reward strategies, and training recipes. Qualitative analysis reveals that J1 develops systematic evaluation strategies, including dynamic criteria generation, reference answer creation, iterative self-correction of initial assessments, and feedback generation for low-quality responses.
comment: 10 pages, 13 tables, 14 figures
♻ ☆ LIDDIA: Language-based Intelligent Drug Discovery Agent EMNLP 2025
Drug discovery is a long, expensive, and complex process, relying heavily on human medicinal chemists, who can spend years searching the vast space of potential therapies. Recent advances in artificial intelligence for chemistry have sought to expedite individual drug discovery tasks; however, there remains a critical need for an intelligent agent that can navigate the drug discovery process. Towards this end, we introduce LIDDIA, an autonomous agent capable of intelligently navigating the drug discovery process in silico. By leveraging the reasoning capabilities of large language models, LIDDIA serves as a low-cost and highly-adaptable tool for autonomous drug discovery. We comprehensively examine LIDDIA , demonstrating that (1) it can generate molecules meeting key pharmaceutical criteria on over 70% of 30 clinically relevant targets, (2) it intelligently balances exploration and exploitation in the chemical space, and (3) it identifies one promising novel candidate on AR/NR3C4, a critical target for both prostate and breast cancers. Code and dataset are available at https://github.com/ninglab/LIDDiA
comment: EMNLP 2025 Main Conference
♻ ☆ When Thinking Backfires: Mechanistic Insights Into Reasoning-Induced Misalignment
With the growing accessibility and wide adoption of large language models, concerns about their safety and alignment with human values have become paramount. In this paper, we identify a concerning phenomenon: Reasoning-Induced Misalignment (RIM), in which misalignment emerges when reasoning capabilities strengthened-particularly when specific types of reasoning patterns are introduced during inference or training. Beyond reporting this vulnerability, we provide the first mechanistic account of its origins. Through representation analysis, we discover that specific attention heads facilitate refusal by reducing their attention to CoT tokens, a mechanism that modulates the model's rationalization process during inference. During training, we find significantly higher activation entanglement between reasoning and safety in safety-critical neurons than in control neurons, particularly after fine-tuning with those identified reasoning patterns. This entanglement strongly correlates with catastrophic forgetting, providing a neuron-level explanation for RIM.
♻ ☆ dInfer: An Efficient Inference Framework for Diffusion Language Models
Diffusion-based large language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) LLMs, leveraging denoising-based generation to enable inherent parallelism. Even more and more open-sourced dLLM models emerge, yet their widespread adoption remains constrained by the lack of a standardized and efficient inference framework. We present dInfer, an efficient and extensible framework for dLLM inference. dInfer decomposes the inference pipeline into four modular components--model, diffusion iteration manager, decoding strategy, and KV-cache manager--and integrates novel algorithms for each component alongside system-level optimizations. Through this combination of algorithmic innovations and system enhancements, dInfer achieves substantial efficiency gains without compromising output quality on LLaDA-MoE. At batch size 1, it surpasses 1,100 tokens per second on HumanEval and averages over 800 tokens per second across six benchmarks on $8\times$ H800 GPUs. Compared to prior systems, dInfer delivers a $10\times$ speedup over Fast-dLLM while maintaining similar model performance. Even compared to the AR model (with a comparable number of activation parameters and performance) QWen2.5-3B, which is highly optimized with the latest vLLM inference engine, dInfer still delivers a $2$-$3\times$ speedup. The implementation of dInfer is open-sourced at https://github.com/inclusionAI/dInfer.
♻ ☆ Accurate and Diverse LLM Mathematical Reasoning via Automated PRM-Guided GFlowNets
Achieving both accuracy and diverse reasoning remains challenging for Large Language Models (LLMs) in complex domains like mathematics. A key bottleneck is evaluating intermediate reasoning steps to guide generation without costly human annotations. To address this, we first introduce a novel Process Reward Model (PRM) trained automatically using Monte Carlo Tree Search coupled with a similarity-based data augmentation technique, effectively capturing step-level reasoning quality. Leveraging this PRM, we then adapt Generative Flow Networks (GFlowNets) to operate at the reasoning step level. Unlike traditional reinforcement learning focused on maximizing a single reward, GFlowNets naturally sample diverse, high-quality solutions proportional to their rewards, as measured by our PRM. Empirical evaluation shows strong improvements in both accuracy and solution diversity on challenging mathematical benchmarks (e.g., +2.59% absolute accuracy on MATH Level 5 for Llama3.2-3B), with effective generalization to unseen datasets (+9.4\% absolute on SAT MATH). Furthermore, we benchmark our PRM against existing open-source reward models, demonstrating superior alignment with reasoning quality and more consistent guidance for downstream generation. Our work demonstrates the potential of PRM-guided, step-level GFlowNets for developing more robust and versatile mathematical reasoning in LLMs.
♻ ☆ LiTransProQA: an LLM-based Literary Translation evaluation metric with Professional Question Answering EMNLP 2025
The impact of Large Language Models (LLMs) has extended into literary domains. However, existing evaluation metrics for literature prioritize mechanical accuracy over artistic expression and tend to overrate machine translation as being superior to human translation from experienced professionals. In the long run, this bias could result in an irreversible decline in translation quality and cultural authenticity. In response to the urgent need for a specialized literary evaluation metric, we introduce LITRANSPROQA, a novel, reference-free, LLM-based question-answering framework designed for literary translation evaluation. LITRANSPROQA integrates humans in the loop to incorporate insights from professional literary translators and researchers, focusing on critical elements in literary quality assessment such as literary devices, cultural understanding, and authorial voice. Our extensive evaluation shows that while literary-finetuned XCOMET-XL yields marginal gains, LITRANSPROQA substantially outperforms current metrics, achieving up to 0.07 gain in correlation and surpassing the best state-of-the-art metrics by over 15 points in adequacy assessments. Incorporating professional translator insights as weights further improves performance, highlighting the value of translator inputs. Notably, LITRANSPROQA reaches an adequacy performance comparable to trained linguistic student evaluators, though it still falls behind experienced professional translators. LITRANSPROQA shows broad applicability to open-source models like LLaMA3.3-70b and Qwen2.5-32b, indicating its potential as an accessible and training-free tool for evaluating literary translations that require local processing due to copyright or ethical considerations.
comment: Accepted as a main paper at EMNLP 2025. CR version
♻ ☆ References Indeed Matter? Reference-Free Preference Optimization for Conversational Query Reformulation
Conversational query reformulation (CQR) has become indispensable for improving retrieval in dialogue-based applications. However, existing approaches typically rely on reference passages for optimization, which are impractical to acquire in real-world scenarios. To address this limitation, we introduce a novel reference-free preference optimization framework DualReform that generates pseudo reference passages from commonly-encountered conversational datasets containing only queries and responses. DualReform attains this goal through two key innovations: (1) response-based inference, where responses serve as proxies to infer pseudo reference passages, and (2) response refinement via the dual-role of CQR, where a CQR model refines responses based on the shared objectives between response refinement and CQR. Despite not relying on reference passages, DualReform achieves 96.9--99.1% of the retrieval accuracy attainable only with reference passages and surpasses the state-of-the-art method by up to 31.6%.
♻ ☆ Test-Time Alignment for Large Language Models via Textual Model Predictive Control
Aligning Large Language Models (LLMs) with human preferences through finetuning is resource-intensive, motivating lightweight alternatives at test time. We address test-time alignment through the lens of sequential decision making, a perspective that reveals two fundamental challenges. When actions are defined at the token level, as in guided decoding, alignment suffers from the curse of horizon. Conversely, when actions are at the response level, as in traditional iterative refinement, the curse of dimensionality emerges. To resolve this trade-off, we draw inspiration from Model Predictive Control (MPC) in control theory to propose Textual Model Predictive Control (TMPC), a novel predictive planning framework adapted for aligning LLMs at inference time. A key limitation of standard MPC is its reliance on predefined, hard segment boundaries, which are often absent in text generation. TMPC overcomes this by introducing two principles inspired by hierarchical reinforcement learning: (1) Hindsight Subgoal Identification, where TMPC analyzes generation subgoals to retrospectively identify high-reward intermediate outputs as subgoals. This allows the framework to discover meaningful, task-specific planning steps (e.g., a sentence in machine translation or a bug fix in code generation.). (2) Subgoal-Conditioned Re-Generation, where these identified subgoals are used to guide subsequent planning iterations. By conditioning on these proven, high-quality subgoals, TMPC ensures stable improvement by building upon previously validated successes. TMPC is evaluated on three tasks with distinct segmentation properties: discourse-level translation, long-form response generation, and program synthesis. The results demonstrate that TMPC consistently improves performance, highlighting the generality.
comment: Preprint. Code will be released at Plan2Align GitHub link: https://github.com/NYCU-RL-Bandits-Lab/Plan2Align
♻ ☆ On the Mathematical Relationship Between Layer Normalization and Dynamic Activation Functions
Layer normalization (LN) is an essential component of modern neural networks. While many alternative techniques have been proposed, none of them have succeeded in replacing LN so far. The latest suggestion in this line of research is a dynamic activation function called Dynamic Tanh (DyT). Although it is empirically well-motivated and appealing from a practical point of view, it lacks a theoretical foundation. In this work, we shed light on the mathematical relationship between LN and dynamic activation functions. In particular, we derive DyT from the LN variant RMSNorm, and show that a well-defined decoupling in derivative space as well as an approximation are needed to do so. By applying the same decoupling procedure directly in function space, we are able to omit the approximation and obtain the exact element-wise counterpart of RMSNorm, which we call Dynamic Inverse Square Root Unit (DyISRU). We demonstrate numerically that DyISRU reproduces the normalization effect on outliers more accurately than DyT does.
comment: Revision and Simplification (starting point RMSNorm instead of LN)
♻ ☆ Autoencoding-Free Context Compression for LLMs via Contextual Semantic Anchors
Context compression presents a promising approach for accelerating large language model (LLM) inference by compressing long contexts into compact representations. Current context compression methods predominantly rely on autoencoding tasks to train context-agnostic compression tokens to compress contextual semantics. While autoencoding tasks enable compression tokens to acquire compression capabilities, compression via autoencoding tasks creates a fundamental mismatch: the models are optimized for reconstruction that diverge from actual downstream tasks, thereby weakening the features more beneficial for real-world usage. We propose Semantic-Anchor Compression (SAC), a novel method that shifts from autoencoding task based compression to an architecture that is equipped with this compression capability \textit{a priori}. Instead of training models to compress contexts through autoencoding tasks, SAC directly selects so-called anchor tokens from the original context and aggregates contextual information into their key-value (KV) representations. By deriving representations directly from the contextual tokens, SAC eliminates the need for autoencoding training. To ensure compression performance while directly leveraging anchor tokens, SAC incorporates two key designs: (1) anchor embeddings that enable the compressor to identify critical tokens, and (2) bidirectional attention modification that allows anchor tokens to capture information from the entire context. Experimental results demonstrate that SAC consistently outperforms existing context compression methods across various compression ratios. On out-of-distribution evaluation using MRQA, SAC achieves 1 EM improvement at 5x compression over strong baselines, with increasing advantages at higher compression ratios.
comment: 18 pages,9 figures
♻ ☆ Hallucinate at the Last in Long Response Generation: A Case Study on Long Document Summarization
Large Language Models (LLMs) have significantly advanced text generation capabilities, including tasks like summarization, often producing coherent and fluent outputs. However, faithfulness to source material remains a significant challenge due to the generation of hallucinations. While extensive research focuses on detecting and reducing these inaccuracies, less attention has been paid to the positional distribution of hallucination within generated text, particularly in long outputs. In this work, we investigate where hallucinations occur in LLM-based long response generation, using long document summarization as a key case study. Focusing on the challenging setting of long context-aware long response generation, we find a consistent and concerning phenomenon: hallucinations tend to concentrate disproportionately in the latter parts of the generated long response. To understand this bias, we explore potential contributing factors related to the dynamics of attention and decoding over long sequences. Furthermore, we investigate methods to mitigate this positional hallucination, aiming to improve faithfulness specifically in the concluding segments of long outputs.
comment: 22 tables, 8 figures
♻ ☆ Introducing Spotlight: A Novel Approach for Generating Captivating Key Information from Documents EMNLP 2025
In this paper, we introduce Spotlight, a novel paradigm for information extraction that produces concise, engaging narratives by highlighting the most compelling aspects of a document. Unlike traditional summaries, which prioritize comprehensive coverage, spotlights selectively emphasize intriguing content to foster deeper reader engagement with the source material. We formally differentiate spotlights from related constructs and support our analysis with a detailed benchmarking study using new datasets curated for this work. To generate high-quality spotlights, we propose a two-stage approach: fine-tuning a large language model on our benchmark data, followed by alignment via Direct Preference Optimization (DPO). Our comprehensive evaluation demonstrates that the resulting model not only identifies key elements with precision but also enhances readability and boosts the engagement value of the original document.
comment: Paper accepted in EMNLP 2025 Main Conference (Full Paper)
♻ ☆ VLMGuard-R1: Proactive Safety Alignment for VLMs via Reasoning-Driven Prompt Optimization
Aligning Vision-Language Models (VLMs) with safety standards is essential to mitigate risks arising from their multimodal complexity, where integrating vision and language unveils subtle threats beyond the reach of conventional safeguards. Inspired by the insight that reasoning across modalities is key to preempting intricate vulnerabilities, we propose a novel direction for VLM safety: multimodal reasoning-driven prompt rewriting. To this end, we introduce VLMGuard-R1, a proactive framework that refines user inputs through a reasoning-guided rewriter, dynamically interpreting text-image interactions to deliver refined prompts that bolster safety across diverse VLM architectures without altering their core parameters. To achieve this, we devise a three-stage reasoning pipeline to synthesize a dataset that trains the rewriter to infer subtle threats, enabling tailored, actionable responses over generic refusals. Extensive experiments across three benchmarks with five VLMs reveal that VLMGuard-R1 outperforms four baselines. In particular, VLMGuard-R1 achieves a remarkable 43.59\% increase in average safety across five models on the SIUO benchmark.
♻ ☆ QUITO-X: A New Perspective on Context Compression from the Information Bottleneck Theory
Generative LLM have achieved remarkable success in various industrial applications, owing to their promising In-Context Learning capabilities. However, the issue of long context in complex tasks poses a significant barrier to their wider adoption, manifested in two main aspects: (i) The excessively long context leads to high costs and inference delays. (ii) A substantial amount of task-irrelevant information introduced by long contexts exacerbates the "lost in the middle" problem. Existing methods compress context by removing redundant tokens using metrics such as self-information or PPL, which is inconsistent with the objective of retaining the most important tokens when conditioning on a given query. In this study, we introduce information bottleneck theory (IB) to model the problem, offering a novel perspective that thoroughly addresses the essential properties required for context compression. Additionally, we propose a cross-attention-based approach to approximate mutual information in IB, which can be flexibly replaced with suitable alternatives in different scenarios. Extensive experiments on four datasets demonstrate that our method achieves a 25% increase in compression rate compared to the state-of-the-art, while maintaining question answering performance. In particular, the context compressed by our method even outperform the full context in some cases.
♻ ☆ Draw with Thought: Unleashing Multimodal Reasoning for Scientific Diagram Generation
Scientific diagrams are vital tools for communicating structured knowledge across disciplines. However, they are often published as static raster images, losing symbolic semantics and limiting reuse. While Multimodal Large Language Models (MLLMs) offer a pathway to bridging vision and structure, existing methods lack semantic control and structural interpretability, especially on complex diagrams. We propose Draw with Thought (DwT), a training-free framework that guides MLLMs to reconstruct diagrams into editable mxGraph XML code through cognitively-grounded Chain-of-Thought reasoning. DwT enables interpretable and controllable outputs without model fine-tuning by dividing the task into two stages: Coarse-to-Fine Planning, which handles perceptual structuring and semantic specification, and Structure-Aware Code Generation, enhanced by format-guided refinement. To support evaluation, we release Plot2XML, a benchmark of 247 real-world scientific diagrams with gold-standard XML annotations. Extensive experiments across eight MLLMs show that our approach yields high-fidelity, semantically aligned, and structurally valid reconstructions, with human evaluations confirming strong alignment in both accuracy and visual aesthetics, offering a scalable solution for converting static visuals into executable representations and advancing machine understanding of scientific graphics.
comment: 10 pages, 5 figures, accepted to appear in the Proceedings of the 33rd ACM International Conference on Multimedia (MM '25)
♻ ☆ TemplateRL: Structured Template-Guided Reinforcement Learning for LLM Reasoning
Reinforcement learning (RL) has emerged as an effective paradigm for enhancing model reasoning. However, existing RL methods like GRPO often rely on unstructured self-sampling to fit scalar rewards, often producing inefficient rollouts that fail to capture transferable problem-solving strategies. To address these limitations, we propose **TemplateRL**, a structured template-guided RL framework that augments policy optimization with explicit template guidance. Our approach first constructs a problem-solving template library via MCTS on a small seed set, then seamlessly integrates this high-level structured guidance into RL training. By guiding rollout generation to align with proven template structures, TemplateRL significantly improves high-quality trajectory hit rates while reducing ineffective exploration. This structure-guided design steers the policy toward validated strategic patterns, stabilizing training dynamics, and enhancing RL sampling efficiency. Notably, the explicit template library is interpretable, editable, and supports online updates-enabling continuous updates during both training and inference. Extensive experiments demonstrate that TemplateRL outperforms GRPO by 99% on AIME and 41% on AMC, with superior stability on weak models and remarkable cross-domain generalization, highlighting its potential for broader tasks.
♻ ☆ Clip Your Sequences Fairly: Enforcing Length Fairness for Sequence-Level RL
We propose FSPO (Fair Sequence Policy Optimization), a sequence-level reinforcement learning method for LLMs that enforces length-fair clipping on the importance-sampling (IS) weight. We study RL methods with sequence-level IS and identify a mismatch when PPO/GRPO-style clipping is transplanted to sequences: a fixed clip range systematically reweights short vs. long responses, distorting the optimization direction. FSPO introduces a simple remedy: we clip the sequence log-IS ratio with a band that scales as $\sqrt{L}$. Theoretically, we formalize length fairness via a Length Reweighting Error (LRE) and prove that small LRE yields a cosine directional guarantee between the clipped and true updates. Empirically, FSPO flattens clip rates across length bins, stabilizes training, and outperforms baselines across model sizes and evaluation datasets, with the largest gains on the Qwen3-8B-Base model.
♻ ☆ SaraCoder: Orchestrating Semantic and Structural Cues for Resource-Optimized Repository-Level Code Completion
Despite Retrieval-Augmented Generation improving code completion, traditional retrieval methods struggle with information redundancy and a lack of diversity within limited context windows. To solve this, we propose a resource-optimized retrieval augmentation method, SaraCoder. It maximizes information diversity and representativeness in a limited context window, significantly boosting the accuracy and reliability of repository-level code completion. Its core Hierarchical Feature Optimization module systematically refines candidates by distilling deep semantic relationships, pruning exact duplicates, assessing structural similarity with a novel graph-based metric that weighs edits by their topological importance, and reranking results to maximize both relevance and diversity. Furthermore, an External-Aware Identifier Disambiguator module accurately resolves cross-file symbol ambiguity via dependency analysis. Extensive experiments on the challenging CrossCodeEval and RepoEval-Updated benchmarks demonstrate that SaraCoder outperforms existing baselines across multiple programming languages and models. Our work proves that systematically refining retrieval results across multiple dimensions provides a new paradigm for building more accurate and resource-optimized repository-level code completion systems.
♻ ☆ Tailored Teaching with Balanced Difficulty: Elevating Reasoning in Multimodal Chain-of-Thought via Prompt Curriculum
The effectiveness of Multimodal Chain-of-Thought (MCoT) prompting is often limited by the use of randomly or manually selected examples. These examples fail to account for both model-specific knowledge distributions and the intrinsic complexity of the tasks, resulting in suboptimal and unstable model performance. To address this, we propose a novel framework inspired by the pedagogical principle of "tailored teaching with balanced difficulty". We reframe prompt selection as a prompt curriculum design problem: constructing a well ordered set of training examples that align with the model's current capabilities. Our approach integrates two complementary signals: (1) model-perceived difficulty, quantified through prediction disagreement in an active learning setup, capturing what the model itself finds challenging; and (2) intrinsic sample complexity, which measures the inherent difficulty of each question-image pair independently of any model. By jointly analyzing these signals, we develop a difficulty-balanced sampling strategy that ensures the selected prompt examples are diverse across both dimensions. Extensive experiments conducted on five challenging benchmarks and multiple popular Multimodal Large Language Models (MLLMs) demonstrate that our method yields substantial and consistent improvements and greatly reduces performance discrepancies caused by random sampling, providing a principled and robust approach for enhancing multimodal reasoning.
♻ ☆ LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning
Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM's autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. We first construct a structured latent reasoning space using a Variational Autoencoder (VAE) that encodes text reasoning steps into blocks of thought tokens, preserving semantic information and interpretability while offering compact but expressive representations. Subsequently, we utilize a latent diffusion model that learns to denoise a block of latent thought tokens with a blockwise bidirectional attention mask, enabling longer horizon and iterative refinement with adaptive test-time compute. This design allows efficient parallel generation of diverse reasoning trajectories, allowing the model to plan and revise the reasoning process holistically. We conduct evaluations on a suite of mathematical reasoning and planning benchmarks. Empirical results show that LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods, revealing a new paradigm for text reasoning with latent diffusion.
♻ ☆ Personality Editing for Language Models through Adjusting Self-Referential Queries
Large Language Models (LLMs) are integral to applications such as conversational agents and content creation, where precise control over a model's personality is essential for maintaining tone, consistency, and user engagement. However, prevailing prompt-based or fine-tuning approaches either lack robustness or demand large-scale training data, making them costly and impractical. In this paper, we present PALETTE (Personality Adjustment by LLM SElf-TargeTed quEries), a novel method for personality editing in LLMs. Our approach introduces adjustment queries, where self-referential statements grounded in psychological constructs are treated analogously to factual knowledge, enabling direct editing of personality-related responses. Unlike fine-tuning, PALETTE requires only 12 editing samples to achieve substantial improvements in personality alignment across personality dimensions. Experimental results from both automatic and human evaluations demonstrate that our method enables more stable and well-balanced personality control in LLMs.
comment: 22 pages, 5 figures, 26 tables
♻ ☆ Reveal-Bangla: A Dataset for Cross-Lingual Multi-Step Reasoning Evaluation AACL 2025
Language models have demonstrated remarkable performance on complex multi-step reasoning tasks. However, their evaluation has been predominantly confined to high-resource languages such as English. In this paper, we introduce a manually translated Bangla multi-step reasoning dataset derived from the English Reveal dataset, featuring both binary and non-binary question types. We conduct a controlled evaluation of English-centric and Bangla-centric multilingual small language models on the original dataset and our translated version to compare their ability to exploit relevant reasoning steps to produce correct answers. Our results show that, in comparable settings, reasoning context is beneficial for more challenging non-binary questions, but models struggle to employ relevant Bangla reasoning steps effectively. We conclude by exploring how reasoning steps contribute to models' predictions, highlighting different trends across models and languages.
comment: Submitted to BLP workshop @ IJCNLP-AACL 2025
♻ ☆ PhoniTale: Phonologically Grounded Mnemonic Generation for Typologically Distant Language Pairs EMNLP 2025
Vocabulary acquisition poses a significant challenge for second-language (L2) learners, especially when learning typologically distant languages such as English and Korean, where phonological and structural mismatches complicate vocabulary learning. Recently, large language models (LLMs) have been used to generate keyword mnemonics by leveraging similar keywords from a learner's first language (L1) to aid in acquiring L2 vocabulary. However, most methods still rely on direct IPA-based phonetic matching or employ LLMs without phonological guidance. In this paper, we present PhoniTale, a novel cross-lingual mnemonic generation system that performs IPA-based phonological adaptation and syllable-aware alignment to retrieve L1 keyword sequence and uses LLMs to generate verbal cues. We evaluate PhoniTale through automated metrics and a short-term recall test with human participants, comparing its output to human-written and prior automated mnemonics. Our findings show that PhoniTale consistently outperforms previous automated approaches and achieves quality comparable to human-written mnemonics.
comment: Accepted to EMNLP 2025 Main Conference
♻ ☆ BabyVLM: Data-Efficient Pretraining of VLMs Inspired by Infant Learning ICCV 2025
Human infants rapidly develop visual reasoning skills from minimal input, suggesting that developmentally inspired pretraining could significantly enhance the efficiency of vision-language models (VLMs). Although recent efforts have leveraged infant-inspired datasets like SAYCam, existing evaluation benchmarks remain misaligned--they are either too simplistic, narrowly scoped, or tailored for large-scale pretrained models. Additionally, training exclusively on infant data overlooks the broader, diverse input from which infants naturally learn. To address these limitations, we propose BabyVLM, a novel framework comprising comprehensive in-domain evaluation benchmarks and a synthetic training dataset created via child-directed transformations of existing datasets. We demonstrate that VLMs trained with our synthetic dataset achieve superior performance on BabyVLM tasks compared to models trained solely on SAYCam or general-purpose data of the SAYCam size. BabyVLM thus provides a robust, developmentally aligned evaluation tool and illustrates how compact models trained on carefully curated data can generalize effectively, opening pathways toward data-efficient vision-language learning paradigms.
comment: Accepted to ICCV 2025
♻ ☆ DiaCDM: Cognitive Diagnosis in Teacher-Student Dialogues using the Initiation-Response-Evaluation Framework
While cognitive diagnosis (CD) effectively assesses students' knowledge mastery from structured test data, applying it to real-world teacher-student dialogues presents two fundamental challenges. Traditional CD models lack a suitable framework for handling dynamic, unstructured dialogues, and it's difficult to accurately extract diagnostic semantics from lengthy dialogues. To overcome these hurdles, we propose DiaCDM, an innovative model. We've adapted the initiation-response-evaluation (IRE) framework from educational theory to design a diagnostic framework tailored for dialogue. We also developed a unique graph-based encoding method that integrates teacher questions with relevant knowledge components to capture key information more precisely. To our knowledge, this is the first exploration of cognitive diagnosis in a dialogue setting. Experiments on three real-world dialogue datasets confirm that DiaCDM not only significantly improves diagnostic accuracy but also enhances the results' interpretability, providing teachers with a powerful tool for assessing students' cognitive states. The code is available at https://github.com/Mind-Lab-ECNU/DiaCDM/tree/main.
♻ ☆ DITING: A Multi-Agent Evaluation Framework for Benchmarking Web Novel Translation
Large language models (LLMs) have substantially advanced machine translation (MT), yet their effectiveness in translating web novels remains unclear. Existing benchmarks rely on surface-level metrics that fail to capture the distinctive traits of this genre. To address these gaps, we introduce DITING, the first comprehensive evaluation framework for web novel translation, assessing narrative and cultural fidelity across six dimensions: idiom translation, lexical ambiguity, terminology localization, tense consistency, zero-pronoun resolution, and cultural safety, supported by over 18K expert-annotated Chinese-English sentence pairs. We further propose AgentEval, a reasoning-driven multi-agent evaluation framework that simulates expert deliberation to assess translation quality beyond lexical overlap, achieving the highest correlation with human judgments among seven tested automatic metrics. To enable metric comparison, we develop MetricAlign, a meta-evaluation dataset of 300 sentence pairs annotated with error labels and scalar quality scores. Comprehensive evaluation of fourteen open, closed, and commercial models reveals that Chinese-trained LLMs surpass larger foreign counterparts, and that DeepSeek-V3 delivers the most faithful and stylistically coherent translations. Our work establishes a new paradigm for exploring LLM-based web novel translation and provides public resources to advance future research.
♻ ☆ Training and Evaluating with Human Label Variation: An Empirical Study
Human label variation (HLV) challenges the standard assumption that a labelled instance has a single ground truth, instead embracing the natural variation in human annotation to train and evaluate models. While various training methods and metrics for HLV have been proposed, it is still unclear which methods and metrics perform best in what settings. We propose new evaluation metrics for HLV leveraging fuzzy set theory. Since these new proposed metrics are differentiable, we then in turn experiment with employing these metrics as training objectives. We conduct an extensive study over 6 HLV datasets testing 14 training methods and 6 evaluation metrics. We find that training on either disaggregated annotations or soft labels performs best across metrics, outperforming training using the proposed training objectives with differentiable metrics. We also show that our proposed soft micro F1 score is one of the best metrics for HLV data.
comment: 27 pages, 7 figures. Accepted to CL. Pre-MIT Press publication version. Fixed PO-JSD values on the MFRC dataset. Completely redid the empirical meta-evaluation, added more related work, and other minor edits
♻ ☆ Chronological Passage Assembling in RAG framework for Temporal Question Answering
Long-context question answering over narrative tasks is challenging because correct answers often hinge on reconstructing a coherent timeline of events while preserving contextual f low in a limited context window. Retrievalaugmented generation (RAG) methods aim to address this challenge by selectively retrieving only necessary document segments. However, narrative texts possess unique characteristics that limit the effectiveness of these existing approaches. Specifically, understanding narrative texts requires more than isolated segments, as the broader context and sequential relationships between segments are crucial for comprehension. To address these limitations, we propose ChronoRAG, a novel RAG framework specialized for narrative texts. This approach focuses on two essential aspects: refining dispersed document information into coherent and structured passages and preserving narrative flow by explicitly capturing and maintaining the temporal order among retrieved passages. We empirically demonstrate the effectiveness of ChronoRAG through experiments on the NarrativeQA and GutenQAdataset, showing substantial improvements in tasks requiring both factual identification and comprehension of complex sequential relationships, underscoring that reasoning over temporal order is crucial in resolving narrative QA.
comment: 15 pages, 4 figures
♻ ☆ Self-Filtered Distillation with LLMs-generated Trust Indicators for Reliable Patent Classification
Large language models (LLMs) increasingly generate natural language rationales to enhance interpretability, but these often contain logical errors, label mismatches, and domain-specific misalignments. Directly using such rationales as supervision risks propagating noise and undermining training stability. To address this challenge, we introduce Self-Filtered Distillation, a framework specifically tailored for patent classification, which treats LLM-generated rationales as trust signals rather than ground-truth supervision. The framework employs selective distillation guided by three unsupervised trust metrics: (1) Self-Consistency, which measures the stability of LLM-generated rationales across multiple generations; (2) Class Entailment Alignment, which assesses semantic coherence with patent-specific class definitions; and (3) LLM Agreement Scoring, which validates rationale-label plausibility. These metrics are integrated into a unified trust score that primarily weights training samples while optionally filtering out extremely low-trust cases, enabling reasoning-aware supervision. Experiments on the USPTO-2M dataset, a widely used benchmark for patent classification, show that our method outperforms label-based learning and conventional distillation in accuracy, stability, and interpretability, establishing a reliable paradigm for leveraging reasoning-aware trust indicators in patent analytics.
♻ ☆ Learning from Reference Answers: Versatile Language Model Alignment without Binary Human Preference Data
Large language models~(LLMs) are expected to be helpful, harmless, and honest. In different alignment scenarios, such as safety, confidence, and general preference alignment, binary preference data collection and reward modeling are resource-intensive but play a central role in transferring human preferences. In this work, we explore using the similarity between sampled generations and reference answers as a supplementary reward function for alignment. When unary reference answers are available, such similarity-based rewards can circumvent the need for binary preference data and explicit reward modeling. We introduce \textit{RefAlign}, a versatile REINFORCE-style alignment algorithm that does not rely on reward or reference models. RefAlign utilizes language generation evaluation metrics, such as BERTScore, between sampled generations and reference answers as surrogate rewards. Beyond general preference optimization, RefAlign can be naturally extended to diverse scenarios, including safety and confidence alignment, by combining similarity-based rewards with task-specific objectives. Across multiple scenarios, RefAlign achieves performance comparable to prior alignment methods while operating without binary preference data or reward models. The code is available at https://github.com/mzhaoshuai/RefAlign.
comment: The code is at https://github.com/mzhaoshuai/RefAlign
♻ ☆ Attributing Response to Context: A Jensen-Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation NeurIPS 2025
Retrieval-Augmented Generation (RAG) leverages large language models (LLMs) combined with external contexts to enhance the accuracy and reliability of generated responses. However, reliably attributing generated content to specific context segments, context attribution, remains challenging due to the computationally intensive nature of current methods, which often require extensive fine-tuning or human annotation. In this work, we introduce a novel Jensen-Shannon Divergence driven method to Attribute Response to Context (ARC-JSD), enabling efficient and accurate identification of essential context sentences without additional fine-tuning, gradient-calculation or surrogate modelling. Evaluations on a wide range of RAG benchmarks, such as TyDi QA, Hotpot QA, and Musique, using instruction-tuned LLMs in different scales demonstrate superior accuracy and significant computational efficiency improvements compared to the previous surrogate-based method. Furthermore, our mechanistic analysis reveals specific attention heads and multilayer perceptron (MLP) layers responsible for context attribution, providing valuable insights into the internal workings of RAG models and how they affect RAG behaviours. Our code is available at https://github.com/ruizheliUOA/ARC_JSD.
comment: Best Paper Award at COLM 2025 XLLM-Reason-Plan Workshop; Accepted at NeurIPS 2025 Mechanistic Interpretability Workshop
♻ ☆ Add-One-In: Incremental Sample Selection for Large Language Models via a Choice-Based Greedy Paradigm EMNLP2025
Selecting high-quality and diverse training samples from extensive datasets plays a crucial role in reducing training overhead and enhancing the performance of Large Language Models (LLMs). However, existing studies fall short in assessing the overall value of selected data, focusing primarily on individual quality, and struggle to strike an effective balance between ensuring diversity and minimizing data point traversals. Therefore, this paper introduces a novel choice-based sample selection framework that shifts the focus from evaluating individual sample quality to comparing the contribution value of different samples when incorporated into the subset. Thanks to the advanced language understanding capabilities of LLMs, we utilize LLMs to evaluate the value of each option during the selection process. Furthermore, we design a greedy sampling process where samples are incrementally added to the subset, thereby improving efficiency by eliminating the need for exhaustive traversal of the entire dataset with the limited budget. Extensive experiments demonstrate that selected data from our method not only surpasses the performance of the full dataset but also achieves competitive results with recent powerful studies, while requiring fewer selections. Moreover, we validate our approach on a larger medical dataset, highlighting its practical applicability in real-world applications. Our code and data are available at https://github.com/BIRlz/comperative_sample_selection.
comment: EMNLP2025
♻ ☆ DSPO: Stable and Efficient Policy Optimization for Agentic Search and Reasoning
Enhancing LLMs with the ability to actively search external knowledge is crucial for complex and real-world tasks. Current approaches either rely on prompting to elicit the model's innate agent capabilities, or suffer from performance ceilings and collapse when applying RL to complex interactive tasks, leaving their true agentic potential untapped. To address this, we introduce \textbf{D}ynamic-filter \textbf{S}equence-level \textbf{P}olicy \textbf{O}ptimization (DSPO), an improved RL algorithm designed for robust agent training through sequence-level optimization and dynamic sample filtering. We train our model purely through RL to interleave multi-turn search and reasoning, obviating the need for supervised demonstration data. Across multiple QA benchmarks, our DSPO-trained 7B model improves over a comparable previous work by \textbf{34.1\%}, and even outperforms the 14B model from previous work in complex multihop QA such as HotpotQA by nearly \textbf{9\% relative}, maintaining exceptional training stability.
♻ ☆ Dynamic Optimizations of LLM Ensembles with Two-Stage Reinforcement Learning Agents
The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive frameworks for dynamically changing environments. This paper introduces RL-Focal, a two-stage RL agent framework that routes and ensembles LLMs. First, we develop the Decider RL-agent, which learns to dynamically select an ensemble of small size ($m_i$) among $N$ LLMs ($m_i \ll N$) for incoming queries from a user-defined downstream task $i$, by maximizing both error-diversity and reasoning-performance of the selected ensemble through iterative updates of task-adaptive rewards and policy. Second, to enable effective fusion of dynamically selected LLMs, we develop the stage-2 Fusion RL-agent, which learns to resolve reasoning conflicts from different LLMs and dynamically adapts to different ensemble teams composed by the Decider Agent for different downstream tasks. Third, we introduce the focal diversity metric to better model the error correlations among multiple LLMs, further improving the generalization performance of the Decider Agent, which actively prunes the ensemble combinations. By focal diversity, we enhance performance across tasks by effectively promoting reward-aware and policy-adaptive ensemble selection and inference fusion. Extensive evaluations on five benchmarks show that RL-Focal achieves the performance improvement of 8.48\% with an ensemble of small size compared to the best individual LLM in a pool and offers stronger robustness. Code is available at https://github.com/sftekin/rl-focal
♻ ☆ ARM: Adaptive Reasoning Model NeurIPS 2025
While large reasoning models demonstrate strong performance on complex tasks, they lack the ability to adjust reasoning token usage based on task difficulty. This often leads to the "overthinking" problem -- excessive and unnecessary reasoning -- which, although potentially mitigated by human intervention to control the token budget, still fundamentally contradicts the goal of achieving fully autonomous AI. In this work, we propose Adaptive Reasoning Model (ARM), a reasoning model capable of adaptively selecting appropriate reasoning formats based on the task at hand. These formats include three efficient ones -- Direct Answer, Short CoT, and Code -- as well as a more elaborate format, Long CoT. To train ARM, we introduce Ada-GRPO, an adaptation of Group Relative Policy Optimization (GRPO), which addresses the format collapse issue in traditional GRPO. Ada-GRPO enables ARM to achieve high token efficiency, reducing tokens by an average of 30%, and up to 70%, while maintaining performance comparable to the model that relies solely on Long CoT. Furthermore, not only does it improve inference efficiency through reduced token generation, but it also brings a 2x speedup in training. In addition to the default Adaptive Mode, ARM supports two additional reasoning modes: 1) Instruction-Guided Mode, which allows users to explicitly specify the reasoning format via special tokens -- ideal when the appropriate format is known for a batch of tasks. 2) Consensus-Guided Mode, which aggregates the outputs of the three efficient formats and resorts to Long CoT in case of disagreement, prioritizing performance with higher token usage.
comment: NeurIPS 2025 (Spotlight)
♻ ☆ Self-Exploring Language Models for Explainable Link Forecasting on Temporal Graphs via Reinforcement Learning
Forecasting future links is a central task in temporal graph (TG) reasoning, requiring models to leverage historical interactions to predict upcoming ones. Traditional neural approaches, such as temporal graph neural networks, achieve strong performance but lack explainability and cannot be applied to unseen graphs without retraining. Recent studies have begun to explore using large language models (LLMs) for graph reasoning, but most of them are constrained to static graphs or small synthetic TGs and lack the evaluation of the quality of reasoning traces generated by LLMs. In this work, we present Reasoning-Enhanced Learning for Temporal Graphs (ReaL-TG), a reinforcement learning framework that fine-tunes LLMs to perform explainable link forecasting on real-world TGs. ReaL-TG uses outcome-based reward to encourage models to self-explore reasoning strategies from graph structure and to produce explanations that directly justify their predictions. To enable evaluation on LLM-generated reasoning traces, we propose a new evaluation protocol combining ranking metrics with an LLM-as-a-Judge system that assesses both the quality of reasoning and the impact of hallucinations. Experiments with ReaL-TG-4B, obtained by fine-tuning Qwen3-4B under our framework, show that it outperforms much larger frontier LLMs, including GPT-5 mini, on ranking metrics, while producing high-quality explanations confirmed by both the LLM judge and human evaluation.
♻ ☆ Prompt4Trust: A Reinforcement Learning Prompt Augmentation Framework for Clinically-Aligned Confidence Calibration in Multimodal Large Language Models ICCV 2025
Multimodal large language models (MLLMs) hold considerable promise for applications in healthcare. However, their deployment in safety-critical settings is hindered by two key limitations: (i) sensitivity to prompt design, and (ii) a tendency to generate incorrect responses with high confidence. As clinicians may rely on a model's stated confidence to gauge the reliability of its predictions, it is especially important that when a model expresses high confidence, it is also highly accurate. We introduce Prompt4Trust, the first reinforcement learning (RL) framework for prompt augmentation targeting confidence calibration in MLLMs. A lightweight LLM is trained to produce context-aware auxiliary prompts that guide a downstream task MLLM to generate responses in which the expressed confidence more accurately reflects predictive accuracy. Unlike conventional calibration techniques, Prompt4Trust specifically prioritizes aspects of calibration most critical for safe and trustworthy clinical decision-making. Beyond improvements driven by this clinically motivated calibration objective, our proposed method also improves task accuracy, achieving state-of-the-art medical visual question answering (VQA) performance on the PMC-VQA benchmark, which is composed of multiple-choice questions spanning diverse medical imaging modalities. Moreover, our framework trained with a small downstream task MLLM showed promising zero-shot generalization to larger MLLMs in our experiments, suggesting the potential for scalable calibration without the associated computational costs. This work demonstrates the potential of automated yet human-aligned prompt engineering for improving the the trustworthiness of MLLMs in safety critical settings. Our codebase can be found at https://github.com/xingbpshen/prompt4trust.
comment: Accepted to ICCV 2025 Workshop CVAMD
♻ ☆ SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models
Diffusion large language models (dLLMs) are emerging as an efficient alternative to autoregressive models due to their ability to decode multiple tokens in parallel. However, aligning dLLMs with human preferences or task-specific rewards via reinforcement learning (RL) is challenging because their intractable log-likelihood precludes the direct application of standard policy gradient methods. While prior work uses surrogates like the evidence lower bound (ELBO), these one-sided approximations can introduce significant policy gradient bias. To address this, we propose the Sandwiched Policy Gradient (SPG) that leverages both an upper and a lower bound of the true log-likelihood. Experiments show that SPG significantly outperforms baselines based on ELBO or one-step estimation. Specifically, SPG improves the accuracy over state-of-the-art RL methods for dLLMs by 3.6% in GSM8K, 2.6% in MATH500, 18.4% in Countdown and 27.0% in Sudoku.
♻ ☆ Steering LLMs for Formal Theorem Proving
Recent advances in automated theorem proving use Large Language Models (LLMs) to translate informal mathematical statements into formal proofs. However, informal cues are often ambiguous or lack strict logical structure, making it hard for models to interpret them precisely. While existing methods achieve strong performance, little is known about how LLMs internally represent informal cues, or how these influence proof generation. To address this, we explore \textit{activation steering}, an inference-time intervention that identifies linear directions in residual activations associated with informal reasoning traces and adjusts them to improve proof construction without fine-tuning. This mechanism also yields interpretable information about how reasoning is internally encoded in the activation space of LLMs. We test our method for generating formal proofs from already-formalized theorems. Our contributions are twofold: (1) a novel activation-based intervention for guiding proof synthesis in LLMs; and (2) demonstration that this intervention improves performance under two decoding strategies (sampling and best-first search) without any further training.
♻ ☆ Rethinking the Residual Distribution of Locate-then-Editing Methods in Model Editing NeurIPS 2025
Model editing enables targeted updates to the knowledge of large language models (LLMs) with minimal retraining. Among existing approaches, locate-then-edit methods constitute a prominent paradigm: they first identify critical layers, then compute residuals at the final critical layer based on the target edit, and finally apply least-squares-based multi-layer updates via $\textbf{residual distribution}$. While empirically effective, we identify a counterintuitive failure mode: residual distribution, a core mechanism in these methods, introduces weight shift errors that undermine editing precision. Through theoretical and empirical analysis, we show that such errors increase with the distribution distance, batch size, and edit sequence length, ultimately leading to inaccurate or suboptimal edits. To address this, we propose the $\textbf{B}$oundary $\textbf{L}$ayer $\textbf{U}$pdat$\textbf{E (BLUE)}$ strategy to enhance locate-then-edit methods. Sequential batch editing experiments on three LLMs and two datasets demonstrate that BLUE not only delivers an average performance improvement of 35.59\%, significantly advancing the state of the art in model editing, but also enhances the preservation of LLMs' general capabilities. Our code is available at https://github.com/xpq-tech/BLUE.
comment: NeurIPS 2025
♻ ☆ When Machine Unlearning Meets Retrieval-Augmented Generation (RAG): Keep Secret or Forget Knowledge? SC
The deployment of large language models (LLMs) like ChatGPT and Gemini has shown their powerful natural language generation capabilities. However, these models can inadvertently learn and retain sensitive information and harmful content during training, raising significant ethical and legal concerns. To address these issues, machine unlearning has been introduced as a potential solution. While existing unlearning methods take into account the specific characteristics of LLMs, they often suffer from high computational demands, limited applicability, or the risk of catastrophic forgetting. To address these limitations, we propose a lightweight behavioral unlearning framework based on Retrieval-Augmented Generation (RAG) technology. By modifying the external knowledge base of RAG, we simulate the effects of forgetting without directly interacting with the unlearned LLM. We approach the construction of unlearned knowledge as a constrained optimization problem, deriving two key components that underpin the effectiveness of RAG-based unlearning. This RAG-based approach is particularly effective for closed-source LLMs, where existing unlearning methods often fail. We evaluate our framework through extensive experiments on both open-source and closed-source models, including ChatGPT, Gemini, Llama-2-7b-chat, and PaLM 2. The results demonstrate that our approach meets five key unlearning criteria: effectiveness, universality, harmlessness, simplicity, and robustness. Meanwhile, this approach can extend to multimodal large language models and LLM-based agents.
comment: 16 pages, 9 figures, 13 tables. To appear in IEEE Transactions on Dependable and Secure Computing (TDSC), 2025
♻ ☆ Self-ensemble: Mitigating Confidence Mis-calibration for Large Language Models
Although Large Language Models (LLMs) perform well in general fields, they exhibit a confidence distortion problem on multi-choice question-answering (MCQA), particularly as the number of answer choices increases. Specifically, on MCQA with many choices, LLMs suffer from under-confidence in correct predictions and over-confidence in incorrect ones, leading to a substantially degraded performance. To solve this problem, we propose Self-ensemble in this work. Our method splits the choices into several groups and ensembles LLM predictions across these groups to reach a final decision. The advantage of Self-ensemble is its plug-and-play nature, where it can be integrated into existing LLM architecture based on a designed attention mask and positional encoding, without requiring labeled datasets for parameter tuning. Experimental results on three LLMs and datasets demonstrate that Self-ensemble comprehensively addresses the confidence distortion problem of LLMs, outperforming standard inference as well as baseline methods.
♻ ☆ Creativity or Brute Force? Using Brainteasers as a Window into the Problem-Solving Abilities of Large Language Models
Accuracy remains a standard metric for evaluating AI systems, but it offers limited insight into how models arrive at their solutions. In this work, we introduce a benchmark based on brainteasers written in long narrative form to probe more deeply into the types of reasoning strategies that models use. Brainteasers are well-suited for this goal because they can be solved with multiple approaches, such as a few-step solution that uses a creative insight or a longer solution that uses more brute force. We investigate large language models (LLMs) across multiple layers of reasoning, focusing not only on correctness but also on the quality and creativity of their solutions. We investigate many aspects of the reasoning process: (1) semantic parsing of the brainteasers into precise mathematical competition style formats; (2) generating solutions from these mathematical forms; (3) self-correcting solutions based on gold solutions; (4) producing step-by-step sketches of solutions; and (5) making use of hints. We find that LLMs are in many cases able to find creative, insightful solutions to brainteasers, suggesting that they capture some of the capacities needed to solve novel problems in creative ways. Nonetheless, there also remain situations where they rely on brute force despite the availability of more efficient, creative solutions, highlighting a potential direction for improvement in the reasoning abilities of LLMs.
comment: 13 Tables; 5 Figures
♻ ☆ Agent Learning via Early Experience
A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.
comment: Work in progress
♻ ☆ Reasoning on a Spectrum: Aligning LLMs to System 1 and System 2 Thinking
Large Language Models (LLMs) exhibit impressive reasoning abilities, yet their reliance on structured step-by-step processing reveals a critical limitation. In contrast, human cognition fluidly adapts between intuitive, heuristic (System 1) and analytical, deliberative (System 2) reasoning depending on the context. This difference between human cognitive flexibility and LLMs' reliance on a single reasoning style raises a critical question: while human fast heuristic reasoning evolved for its efficiency and adaptability, is a uniform reasoning approach truly optimal for LLMs, or does its inflexibility make them brittle and unreliable when faced with tasks demanding more agile, intuitive responses? To answer these questions, we explicitly align LLMs to these reasoning styles by curating a dataset with valid System 1 and System 2 answers, and evaluate their performance across reasoning benchmarks. Our results reveal an accuracy-efficiency trade-off: System 2-aligned models excel in arithmetic and symbolic reasoning, while System 1-aligned models perform better in commonsense reasoning tasks. To analyze the reasoning spectrum, we interpolated between the two extremes by varying the proportion of alignment data, which resulted in a monotonic change in accuracy. A mechanistic analysis of model responses shows that System 1 models employ more definitive outputs, whereas System 2 models demonstrate greater uncertainty. Building on these findings, we further combine System 1- and System 2-aligned models based on the entropy of their generations, without additional training, and obtain a dynamic model that outperforms across nearly all benchmarks. This work challenges the assumption that step-by-step reasoning is always optimal and highlights the need for adapting reasoning strategies based on task demands.
♻ ☆ Can LLMs Express Personality Across Cultures? Introducing CulturalPersonas for Evaluating Trait Alignment
As LLMs become central to interactive applications, ranging from tutoring to mental health, the ability to express personality in culturally appropriate ways is increasingly important. While recent works have explored personality evaluation of LLMs, they largely overlook the interplay between culture and personality. To address this, we introduce CulturalPersonas, the first large-scale benchmark with human validation for evaluating LLMs' personality expression in culturally grounded, behaviorally rich contexts. Our dataset spans 3,000 scenario-based questions across six diverse countries, designed to elicit personality through everyday scenarios rooted in local values. We evaluate three LLMs, using both multiple-choice and open-ended response formats. Our results show that CulturalPersonas improves alignment with country-specific human personality distributions (over a 20% reduction in Wasserstein distance across models and countries) and elicits more expressive, culturally coherent outputs compared to existing benchmarks. CulturalPersonas surfaces meaningful modulated trait outputs in response to culturally grounded prompts, offering new directions for aligning LLMs to global norms of behavior. By bridging personality expression and cultural nuance, we envision that CulturalPersonas will pave the way for more socially intelligent and globally adaptive LLMs.
♻ ☆ COMAL: A Convergent Meta-Algorithm for Aligning LLMs with General Preferences
Many alignment methods, including reinforcement learning from human feedback (RLHF), rely on the Bradley-Terry reward assumption, which is not always sufficient to capture the full range and complexity of general human preferences. We explore RLHF under a general preference framework by modeling the alignment problem as a two-player zero-sum game in a game-theoretic framework, where the Nash equilibrium policy guarantees a 50% win rate against any competing policy. However, previous self-play algorithms for finding the Nash policy either diverge or only converge to a Nash policy in a modified game, even in a simple synthetic setting, thereby failing to maintain the 50% win rate guarantee against all other policies. We propose a meta-algorithm, Convergent Meta Alignment Algorithm (COMAL), for language model alignment with general preferences, inspired by convergent algorithms in game theory. We provide theoretical analysis that our meta-algorithm converges to an exact Nash policy in the last iterate and demonstrate its effectiveness on a range of synthetic and preference optimization datasets. COMAL is simple and can be integrated with many existing methods designed for preference optimization with minimal changes, and empirically it consistently maintains above 60.2% and 56.8% win rates, when applied to Llama-3-8B-Instruct and Qwen2.5-7B, against all compared algorithms under controlled evaluations.
♻ ☆ VIDEE: Visual and Interactive Decomposition, Execution, and Evaluation of Text Analytics with Intelligent Agents
Text analytics has traditionally required specialized knowledge in Natural Language Processing (NLP) or text analysis, which presents a barrier for entry-level analysts. Recent advances in large language models (LLMs) have changed the landscape of NLP by enabling more accessible and automated text analysis (e.g., topic detection, summarization, information extraction, etc.). We introduce VIDEE, a system that supports entry-level data analysts to conduct advanced text analytics with intelligent agents. VIDEE instantiates a human-agent collaroration workflow consisting of three stages: (1) Decomposition, which incorporates a human-in-the-loop Monte-Carlo Tree Search algorithm to support generative reasoning with human feedback, (2) Execution, which generates an executable text analytics pipeline, and (3) Evaluation, which integrates LLM-based evaluation and visualizations to support user validation of execution results. We conduct two quantitative experiments to evaluate VIDEE's effectiveness and analyze common agent errors. A user study involving participants with varying levels of NLP and text analytics experience -- from none to expert -- demonstrates the system's usability and reveals distinct user behavior patterns. The findings identify design implications for human-agent collaboration, validate the practical utility of VIDEE for non-expert users, and inform future improvements to intelligent text analytics systems.
♻ ☆ LearnLens: LLM-Enabled Personalised, Curriculum-Grounded Feedback with Educators in the Loop EMNLP 2025
Effective feedback is essential for student learning but is time-intensive for teachers. We present LearnLens, a modular, LLM-based system that generates personalised, curriculum-aligned feedback in science education. LearnLens comprises three components: (1) an error-aware assessment module that captures nuanced reasoning errors; (2) a curriculum-grounded generation module that uses a structured, topic-linked memory chain rather than traditional similarity-based retrieval, improving relevance and reducing noise; and (3) an educator-in-the-loop interface for customisation and oversight. LearnLens addresses key challenges in existing systems, offering scalable, high-quality feedback that empowers both teachers and students.
comment: EMNLP 2025
Information Retrieval 36
☆ FinVet: A Collaborative Framework of RAG and External Fact-Checking Agents for Financial Misinformation Detection
Financial markets face growing threats from misinformation that can trigger billions in losses in minutes. Most existing approaches lack transparency in their decision-making and provide limited attribution to credible sources. We introduce FinVet, a novel multi-agent framework that integrates two Retrieval-Augmented Generation (RAG) pipelines with external fact-checking through a confidence-weighted voting mechanism. FinVet employs adaptive three-tier processing that dynamically adjusts verification strategies based on retrieval confidence, from direct metadata extraction to hybrid reasoning to full model-based analysis. Unlike existing methods, FinVet provides evidence-backed verdicts, source attribution, confidence scores, and explicit uncertainty flags when evidence is insufficient. Experimental evaluation on the FinFact dataset shows that FinVet achieves an F1 score of 0.85, which is a 10.4% improvement over the best individual pipeline (fact-check pipeline) and 37% improvement over standalone RAG approaches.
☆ OneRec-Think: In-Text Reasoning for Generative Recommendation
The powerful generative capacity of Large Language Models (LLMs) has instigated a paradigm shift in recommendation. However, existing generative models (e.g., OneRec) operate as implicit predictors, critically lacking the capacity for explicit and controllable reasoning-a key advantage of LLMs. To bridge this gap, we propose OneRec-Think, a unified framework that seamlessly integrates dialogue, reasoning, and personalized recommendation. OneRec-Think incorporates: (1) Itemic Alignment: cross-modal Item-Textual Alignment for semantic grounding; (2) Reasoning Activation: Reasoning Scaffolding to activate LLM reasoning within the recommendation context; and (3) Reasoning Enhancement, where we design a recommendation-specific reward function that accounts for the multi-validity nature of user preferences. Experiments across public benchmarks show state-of-the-art performance. Moreover, our proposed "Think-Ahead" architecture enables effective industrial deployment on Kuaishou, achieving a 0.159\% gain in APP Stay Time and validating the practical efficacy of the model's explicit reasoning capability.
☆ SemCSE-Multi: Multifaceted and Decodable Embeddings for Aspect-Specific and Interpretable Scientific Domain Mapping
We propose SemCSE-Multi, a novel unsupervised framework for generating multifaceted embeddings of scientific abstracts, evaluated in the domains of invasion biology and medicine. These embeddings capture distinct, individually specifiable aspects in isolation, thus enabling fine-grained and controllable similarity assessments as well as adaptive, user-driven visualizations of scientific domains. Our approach relies on an unsupervised procedure that produces aspect-specific summarizing sentences and trains embedding models to map semantically related summaries to nearby positions in the embedding space. We then distill these aspect-specific embedding capabilities into a unified embedding model that directly predicts multiple aspect embeddings from a scientific abstract in a single, efficient forward pass. In addition, we introduce an embedding decoding pipeline that decodes embeddings back into natural language descriptions of their associated aspects. Notably, we show that this decoding remains effective even for unoccupied regions in low-dimensional visualizations, thus offering vastly improved interpretability in user-centric settings.
☆ REGENT: Relevance-Guided Attention for Entity-Aware Multi-Vector Neural Re-Ranking SIGIR
Current neural re-rankers often struggle with complex information needs and long, content-rich documents. The fundamental issue is not computational--it is intelligent content selection: identifying what matters in lengthy, multi-faceted texts. While humans naturally anchor their understanding around key entities and concepts, neural models process text within rigid token windows, treating all interactions as equally important and missing critical semantic signals. We introduce REGENT, a neural re-ranking model that mimics human-like understanding by using entities as a "semantic skeleton" to guide attention. REGENT integrates relevance guidance directly into the attention mechanism, combining fine-grained lexical matching with high-level semantic reasoning. This relevance-guided attention enables the model to focus on conceptually important content while maintaining sensitivity to precise term matches. REGENT achieves new state-of-the-art performance in three challenging datasets, providing up to 108% improvement over BM25 and consistently outperforming strong baselines including ColBERT and RankVicuna. To our knowledge, this is the first work to successfully integrate entity semantics directly into neural attention, establishing a new paradigm for entity-aware information retrieval.
comment: To be published in: Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (SIGIR-AP 2025)
☆ QDER: Query-Specific Document and Entity Representations for Multi-Vector Document Re-Ranking SIGIR
Neural IR has advanced through two distinct paths: entity-oriented approaches leveraging knowledge graphs and multi-vector models capturing fine-grained semantics. We introduce QDER, a neural re-ranking model that unifies these approaches by integrating knowledge graph semantics into a multi-vector model. QDER's key innovation lies in its modeling of query-document relationships: rather than computing similarity scores on aggregated embeddings, we maintain individual token and entity representations throughout the ranking process, performing aggregation only at the final scoring stage - an approach we call "late aggregation." We first transform these fine-grained representations through learned attention patterns, then apply carefully chosen mathematical operations for precise matches. Experiments across five standard benchmarks show that QDER achieves significant performance gains, with improvements of 36% in nDCG@20 over the strongest baseline on TREC Robust 2004 and similar improvements on other datasets. QDER particularly excels on difficult queries, achieving an nDCG@20 of 0.70 where traditional approaches fail completely (nDCG@20 = 0.0), setting a foundation for future work in entity-aware retrieval.
comment: Published in: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2025)
☆ Characterizing Web Search in The Age of Generative AI
The advent of LLMs has given rise to a new type of web search: Generative search, where LLMs retrieve web pages related to a query and generate a single, coherent text as a response. This output modality stands in stark contrast to traditional web search, where results are returned as a ranked list of independent web pages. In this paper, we ask: Along what dimensions do generative search outputs differ from traditional web search? We compare Google, a traditional web search engine, with four generative search engines from two providers (Google and OpenAI) across queries from four domains. Our analysis reveals intriguing differences. Most generative search engines cover a wider range of sources compared to web search. Generative search engines vary in the degree to which they rely on internal knowledge contained within the model parameters v.s. external knowledge retrieved from the web. Generative search engines surface varying sets of concepts, creating new opportunities for enhancing search diversity and serendipity. Our results also highlight the need for revisiting evaluation criteria for web search in the age of Generative AI.
☆ Uncertainty Quantification for Retrieval-Augmented Reasoning
Retrieval-augmented reasoning (RAR) is a recent evolution of retrieval-augmented generation (RAG) that employs multiple reasoning steps for retrieval and generation. While effective for some complex queries, RAR remains vulnerable to errors and misleading outputs. Uncertainty quantification (UQ) offers methods to estimate the confidence of systems' outputs. These methods, however, often handle simple queries with no retrieval or single-step retrieval, without properly handling RAR setup. Accurate estimation of UQ for RAR requires accounting for all sources of uncertainty, including those arising from retrieval and generation. In this paper, we account for all these sources and introduce Retrieval-Augmented Reasoning Consistency (R2C)--a novel UQ method for RAR. The core idea of R2C is to perturb the multi-step reasoning process by applying various actions to reasoning steps. These perturbations alter the retriever's input, which shifts its output and consequently modifies the generator's input at the next step. Through this iterative feedback loop, the retriever and generator continuously reshape one another's inputs, enabling us to capture uncertainty arising from both components. Experiments on five popular RAR systems across diverse QA datasets show that R2C improves AUROC by over 5% on average compared to the state-of-the-art UQ baselines. Extrinsic evaluations using R2C as an external signal further confirm its effectiveness for two downstream tasks: in Abstention, it achieves ~5% gains in both F1Abstain and AccAbstain; in Model Selection, it improves the exact match by ~7% over single models and ~3% over selection methods.
☆ What Generative Search Engines Like and How to Optimize Web Content Cooperatively
By employing large language models (LLMs) to retrieve documents and generate natural language responses, Generative Engines, such as Google AI overview and ChatGPT, provide significantly enhanced user experiences and have rapidly become the new form of search. Their rapid adoption also drives the needs of Generative Engine Optimization (GEO), as content providers are eager to gain more traction from them. In this paper, we introduce AutoGEO, a framework to automatically learn generative engine preferences when using retrieved contents for response generation, and rewrite web contents for more such traction. AutoGEO first prompts frontier LLMs to explain generative engine preferences and extract meaningful preference rules from these explanations. Then it uses preference rules as context engineering for AutoGEO$_\text{API}$, a prompt-based GEO system, and as rule-based rewards to train AutoGEO$_\text{Mini}$, a cost-effective GEO model. Experiments on the standard GEO-Bench and two newly constructed benchmarks using real user queries demonstrate the effectiveness of AutoGEO in enhancing content traction while preserving search utility. Analyses confirm the learned rules' robustness and abilities to capture unique preferences in variant domains, and AutoGEO systems' ability to embed them in content optimization. The code is released at https://github.com/cxcscmu/AutoGEO.
☆ On Inherited Popularity Bias in Cold-Start Item Recommendation RecSys 2025
Collaborative filtering (CF) recommender systems struggle with making predictions on unseen, or 'cold', items. Systems designed to address this challenge are often trained with supervision from warm CF models in order to leverage collaborative and content information from the available interaction data. However, since they learn to replicate the behavior of CF methods, cold-start models may therefore also learn to imitate their predictive biases. In this paper, we show that cold-start systems can inherit popularity bias, a common cause of recommender system unfairness arising when CF models overfit to more popular items, thereby maximizing user-oriented accuracy but neglecting rarer items. We demonstrate that cold-start recommenders not only mirror the popularity biases of warm models, but are in fact affected more severely: because they cannot infer popularity from interaction data, they instead attempt to estimate it based solely on content features. This leads to significant over-prediction of certain cold items with similar content to popular warm items, even if their ground truth popularity is very low. Through experiments on three multimedia datasets, we analyze the impact of this behavior on three generative cold-start methods. We then describe a simple post-processing bias mitigation method that, by using embedding magnitude as a proxy for predicted popularity, can produce more balanced recommendations with limited harm to user-oriented cold-start accuracy.
comment: Published at ACM RecSys 2025
☆ VeriCite: Towards Reliable Citations in Retrieval-Augmented Generation via Rigorous Verification
Retrieval-Augmented Generation (RAG) has emerged as a crucial approach for enhancing the responses of large language models (LLMs) with external knowledge sources. Despite the impressive performance in complex question-answering tasks, RAG still struggles with hallucinations. Attributing RAG-generated content through in-line citations has demonstrated potential in reducing hallucinations and facilitating human verification. Existing citation generation methods primarily rely on either fine-tuning the generator or employing post-processing approaches for citation matching. However, the former approach demands substantial annotated data and computational resources, while the latter often encounters difficulties in managing multiple citations and frequently produces suboptimal results. In this paper, we introduce a novel framework, called VeriCite, designed to rigorously validate supporting evidence and enhance answer attribution. Specifically, VeriCite breaks down into a three-stage generation: 1) The initial answer generation first generates a response based on all available contexts and has its claims verified through the NLI model; 2) the supporting evidence selection assesses the utility of each document and extracts useful supporting evidences; 3) the final answer refinement integrates the initial response and collected evidences to produce the final, refined answer.We conduct experiments across five open-source LLMs and four datasets, demonstrating that VeriCite can significantly improve citation quality while maintaining the correctness of the answers.
☆ LLM-Specific Utility: A New Perspective for Retrieval-Augmented Generation
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external knowledge. While traditional retrieval focuses on relevance, RAG's effectiveness depends on the utility of retrieved passages, i.e., the usefulness in facilitating the generation of an accurate and comprehensive answer. Existing studies often treat utility as a generic attribute, ignoring the fact that different LLMs may benefit differently from the same passage due to variations in internal knowledge and comprehension ability. In this work, we introduce and systematically investigate the notion of LLM-specific utility. Through large-scale experiments across multiple datasets and LLMs, we demonstrate that human-annotated passages are not optimal for LLMs and that ground-truth utilitarian passages are not transferable across different LLMs. These findings highlight the necessity of adopting the LLM-specific utility in RAG research. Our findings indicate that some human-annotated passages are not ground-truth utilitarian passages for specific LLMs, partially due to the varying readability of queries and passages for LLMs, a tendency for which perplexity is a key metric. Based on these findings, we propose a benchmarking procedure for LLM-specific utility judgments. We evaluate existing utility judgment methods on six datasets and find that while verbalized methods using pseudo-answers perform robustly, LLMs struggle to assess utility effectively-failing to reject all passages for known queries and to select truly useful ones for unknown queries.
comment: 13 pages, 9 figures
☆ Dynamic Network-Based Two-Stage Time Series Forecasting for Affiliate Marketing
In recent years, affiliate marketing has emerged as a revenue-sharing strategy where merchants collaborate with promoters to promote their products. It not only increases product exposure but also allows promoters to earn a commission. This paper addresses the pivotal yet under-explored challenge in affiliate marketing: accurately assessing and predicting the contributions of promoters in product promotion. We design a novel metric for evaluating the indirect contributions of the promoter, called propagation scale. Unfortunately, existing time series forecasting techniques fail to deliver accurate predictions due to the propagation scale being influenced by multiple factors and the inherent complexities arising from dynamic scenarios. To address this issue, we decouple the network structure from the node signals and propose a two-stage solution: initially, the basic self-sales and network structure prediction are conducted separately, followed by the synthesis of the propagation scale. Specifically, we design a graph convolution encoding scheme based on descendant neighbors and incorporate hypergraph convolution to efficiently capture complex promotional dynamics. Additionally, three auxiliary tasks are employed: self-sales prediction for base estimations, descendant prediction to synthesize propagation scale, and promoter activation prediction to mitigate high volatility issues. Extensive offline experiments on large-scale industrial datasets validate the superiority of our method. We further deploy our model on Alimama platform with over $100,000$ promoters, achieving a $9.29\%$ improvement in GMV and a $5.89\%$ increase in sales volume.
☆ Next Interest Flow: A Generative Pre-training Paradigm for Recommender Systems by Modeling All-domain Movelines
Click-Through Rate (CTR) prediction, a cornerstone of modern recommender systems, has been dominated by discriminative models that react to past user behavior rather than proactively modeling user intent. Existing generative paradigms attempt to address this but suffer from critical limitations: Large Language Model (LLM) based methods create a semantic mismatch by forcing e-commerce signals into a linguistic space, while ID-based generation is constrained by item memorization and cold-start issues. To overcome these limitations, we propose a novel generative pre-training paradigm. Our model learns to predict the Next Interest Flow, a dense vector sequence representing a user's future intent, while simultaneously modeling its internal Interest Diversity and Interest Evolution Velocity to ensure the representation is both rich and coherent. However, this two-stage approach introduces a critical objective mismatch between the generative and discriminative stages. We resolve this via a bidirectional alignment strategy, which harmonizes the two stages through cross-stage weight initialization and a dynamic Semantic Alignment Module for fine-tuning. Additionally, we enhance the underlying discriminative model with a Temporal Sequential Pairwise (TSP) mechanism to better capture temporal causality. We present the All-domain Moveline Evolution Network (AMEN), a unified framework implementing our entire pipeline. Extensive offline experiments validate AMEN's superiority over strong baselines, and a large-scale online A/B test demonstrates its significant real-world impact, delivering substantial improvements in key business metrics.
☆ ELMO: Efficiency via Low-precision and Peak Memory Optimization in Large Output Spaces ICML 2025
Large output spaces, also referred to as Extreme multilabel classification (XMC), is a setting that arises, e.g., in large-scale tagging and product-to-product recommendation, and is characterized by the number of labels ranging from hundreds of thousands to millions. This means that the linear classification head, usually only a tiny fraction of the overall model, turns into the main driver for compute and memory demand. Current state-of-the-art XMC methods predominantly rely on FP16-FP32 mixed-precision training, which we show can be unstable, and inefficient in terms of memory usage and computational overhead. Meanwhile, existing low-precision methods typically retain higher precision for the classification layer. In this work, we propose ELMO, a pure low-precision training framework for XMC models using BFloat16 and Float8 data types. By leveraging Kahan summation and stochastic rounding, we demonstrate that XMC models can be effectively trained entirely in Float8, without relying on single-precision master weights or tensor scaling. Low-precision training, combined with our proposed memory optimizations -- gradient fusion and chunking -- enables significant reductions in GPU memory usage. For example, we train a 3-million-label XMC model with only 6.6 GiB of GPU memory, compared to the 39.7 GiB required by the optimized SOTA method, Renee without compromising accuracy.
comment: Accepted to ICML 2025
☆ DyKnow-RAG: Dynamic Knowledge Utilization Reinforcement Framework for Noisy Retrieval-Augmented Generation in E-commerce Search Relevance
Accurately modeling query-item relevance drives e-commerce ranking, yet long-tail, knowledge-heavy, and fast-evolving queries exceed parametric LLM coverage. External context (reviews, attribute encyclopedias, UGC) can help but is noisy, and single-pass latency and cost forbid any clean-then-summarize step. The model must, per query, judge relevance and decide whether to use, partially use, or ignore the context. DyKnow-RAG is a dynamic noisy-RAG framework built on Group Relative Policy Optimization. It trains two rollout groups (no external context vs a single retrieved chunk) and applies posterior-driven inter-group advantage scaling that adaptively reweights their contributions by the per-query correctness gap. This teaches when to trust retrieval versus fall back to parametric knowledge, without process labels, value networks, or extra inference passes, preserving single-pass, single-chunk deployment under production latency. Training combines: (1) supervised initialization with a structured rationale that explicitly records the context-usage decision; (2) an RL pool prioritized by SFT uncertainty to focus where context choice is most consequential; and (3) an optional lightweight DPO warm start to stabilize with-context calibration. Under a unified retrieval/index and fixed latency budget, DyKnow-RAG outperforms SFT, DPO, and vanilla GRPO in offline tests, and delivers consistent lifts on GSB, Query Goodrate, and Item Goodrate in Taobao A/B testing. It is deployed in Taobao's production relevance system, serving live traffic. To our knowledge, it is among the first single-pass RAG solutions for e-commerce relevance, turning noisy external signals into reliable gains without added online complexity.
☆ HoMer: Addressing Heterogeneities by Modeling Sequential and Set-wise Contexts for CTR Prediction
Click-through rate (CTR) prediction, which models behavior sequence and non-sequential features (e.g., user/item profiles or cross features) to infer user interest, underpins industrial recommender systems. However, most methods face three forms of heterogeneity that degrade predictive performance: (i) Feature Heterogeneity persists when limited sequence side features provide less granular interest representation compared to extensive non-sequential features, thereby impairing sequence modeling performance; (ii) Context Heterogeneity arises because a user's interest in an item will be influenced by other items, yet point-wise prediction neglects cross-item interaction context from the entire item set; (iii) Architecture Heterogeneity stems from the fragmented integration of specialized network modules, which compounds the model's effectiveness, efficiency and scalability in industrial deployments. To tackle the above limitations, we propose HoMer, a Homogeneous-Oriented TransforMer for modeling sequential and set-wise contexts. First, we align sequence side features with non-sequential features for accurate sequence modeling and fine-grained interest representation. Second, we shift the prediction paradigm from point-wise to set-wise, facilitating cross-item interaction in a highly parallel manner. Third, HoMer's unified encoder-decoder architecture achieves dual optimization through structural simplification and shared computation, ensuring computational efficiency while maintaining scalability with model size. Without arduous modification to the prediction pipeline, HoMer successfully scales up and outperforms our industrial baseline by 0.0099 in the AUC metric, and enhances online business metrics like CTR/RPM by 1.99%/2.46%. Additionally, HoMer saves 27% of GPU resources via preliminary engineering optimization, further validating its superiority and practicality.
comment: 10 pages, 6 figures
☆ Decoupled Multimodal Fusion for User Interest Modeling in Click-Through Rate Prediction
Modern industrial recommendation systems improve recommendation performance by integrating multimodal representations from pre-trained models into ID-based Click-Through Rate (CTR) prediction frameworks. However, existing approaches typically adopt modality-centric modeling strategies that process ID-based and multimodal embeddings independently, failing to capture fine-grained interactions between content semantics and behavioral signals. In this paper, we propose Decoupled Multimodal Fusion (DMF), which introduces a modality-enriched modeling strategy to enable fine-grained interactions between ID-based collaborative representations and multimodal representations for user interest modeling. Specifically, we construct target-aware features to bridge the semantic gap across different embedding spaces and leverage them as side information to enhance the effectiveness of user interest modeling. Furthermore, we design an inference-optimized attention mechanism that decouples the computation of target-aware features and ID-based embeddings before the attention layer, thereby alleviating the computational bottleneck introduced by incorporating target-aware features. To achieve comprehensive multimodal integration, DMF combines user interest representations learned under the modality-centric and modality-enriched modeling strategies. Offline experiments on public and industrial datasets demonstrate the effectiveness of DMF. Moreover, DMF has been deployed on the product recommendation system of the international e-commerce platform Lazada, achieving relative improvements of 5.30% in CTCVR and 7.43% in GMV with negligible computational overhead.
☆ From Reasoning LLMs to BERT: A Two-Stage Distillation Framework for Search Relevance
Query-service relevance prediction in e-commerce search systems faces strict latency requirements that prevent the direct application of Large Language Models (LLMs). To bridge this gap, we propose a two-stage reasoning distillation framework to transfer reasoning capabilities from a powerful teacher LLM to a lightweight, deployment-friendly student model. In the first stage, we address the limitations of general-purpose LLMs by constructing a domain-adapted teacher model. This is achieved through a three-step process: domain-adaptive pre-training to inject platform knowledge, supervised fine-tuning to elicit reasoning skills, and preference optimization with a multi-dimensional reward model to ensure the generation of reliable and preference-aligned reasoning paths. This teacher can then automatically annotate massive query-service pairs from search logs with both relevance labels and reasoning chains. In the second stage, to address the challenges of architectural heterogeneity in standard distillation, we introduce Contrastive Reasoning Self-Distillation (CRSD). By modeling the behavior of the same student model under "standard" and "reasoning-augmented" inputs as a teacher-student relationship, CRSD enables the lightweight model to internalize the teacher's complex decision-making mechanisms without needing the explicit reasoning path at inference. Offline evaluations and online A/B testing in the Meituan search advertising system demonstrate that our framework achieves significant improvements across multiple metrics, validating its effectiveness and practical value.
☆ FBS Model-based Maintenance Record Accumulation for Failure-Cause Inference in Manufacturing Systems
In manufacturing systems, identifying the causes of failures is crucial for maintaining and improving production efficiency. In knowledge-based failure-cause inference, it is important that the knowledge base (1) explicitly structures knowledge about the target system and about failures, and (2) contains sufficiently long causal chains of failures. In this study, we constructed Diagnostic Knowledge Ontology and proposed a Function-Behavior-Structure (FBS) model-based maintenance-record accumulation method based on it. Failure-cause inference using the maintenance records accumulated by the proposed method showed better agreement with the set of candidate causes enumerated by experts, especially in difficult cases where the number of related cases is small and the vocabulary used differs. In the future, it will be necessary to develop inference methods tailored to these maintenance records, build a user interface, and carry out validation on larger and more diverse systems. Additionally, this approach leverages the understanding and knowledge of the target in the design phase to support knowledge accumulation and problem solving during the maintenance phase, and it is expected to become a foundation for knowledge sharing across the entire engineering chain in the future.
☆ Does LLM Focus on the Right Words? Diagnosing Language Bias in LLM-based Recommenders
Large language models (LLMs), owing to their extensive open-domain knowledge and semantic reasoning capabilities, have been increasingly integrated into recommender systems (RS). However, a substantial gap remains between the pre-training objectives of LLMs and the specific requirements of recommendation tasks. To address this gap, supervised fine-tuning (SFT) is commonly performed on specially curated recommendation datasets to further enhance their predictive ability. Despite its success, SFT exhibits a critical limitation: it induces Language Bias, whereby the model over-relies on auxiliary tokens-such as task descriptions and prefix-generated tokens-while underutilizing core user interaction tokens that encode user-specific preferences. This bias not only undermines recommendation accuracy but also raises unfairness concerns. To address this issue, we propose Group Distributionally Robust Optimization-based Tuning (GDRT), a novel fine-tuning paradigm that enforces consistent model performance across token groups with varying degrees of relevance to auxiliary tokens. By adaptively upweighting underperforming groups, typically those weakly correlated with auxiliary tokens, GDRT shifts the model's attention from superficial auxiliary cues to informative user interaction tokens, thereby mitigating language bias. Extensive experiments conducted on three public datasets demonstrate that GDRT effectively mitigates language bias, yielding substantial improvements in recommendation accuracy (with an average NDCG@10 gain of 24.29%) and significantly enhancing recommendation fairness.
☆ HatLLM: Hierarchical Attention Masking for Enhanced Collaborative Modeling in LLM-based Recommendation
Recent years have witnessed a surge of research on leveraging large language models (LLMs) for sequential recommendation. LLMs have demonstrated remarkable potential in inferring users' nuanced preferences through fine-grained semantic reasoning. However, they also exhibit a notable limitation in effectively modeling collaborative signals, i.e., behavioral correlations inherent in users' historical interactions. Our empirical analysis further reveals that the attention mechanisms in LLMs tend to disproportionately focus on tokens within the same item, thereby impeding the capture of cross-item correlations. To address this limitation, we propose a novel hierarchical attention masking strategy for LLM-based recommendation, termed HatLLM. Specifically, in shallow layers, HatLLM masks attention between tokens from different items, facilitating intra-item semantic understanding; in contrast, in deep layers, HatLLM masks attention within items, thereby compelling the model to capture cross-item correlations. This progressive, layer-wise approach enables LLMs to jointly model both token-level and item-level dependencies. Extensive experiments on three real-world datasets demonstrate that HatLLM achieves significant performance gains (9.13% on average) over existing LLM-based methods.
☆ Comparative Explanations via Counterfactual Reasoning in Recommendations
Explainable recommendation through counterfactual reasoning seeks to identify the influential aspects of items in recommendations, which can then be used as explanations. However, state-of-the-art approaches, which aim to minimize changes in product aspects while reversing their recommended decisions according to an aggregated decision boundary score, often lead to factual inaccuracies in explanations. To solve this problem, in this work we propose a novel method of Comparative Counterfactual Explanations for Recommendation (CoCountER). CoCountER creates counterfactual data based on soft swap operations, enabling explanations for recommendations of arbitrary pairs of comparative items. Empirical experiments validate the effectiveness of our approach.
☆ Embedding the Teacher: Distilling vLLM Preferences for Scalable Image Retrieval
Text--image retrieval is necessary for applications such as product recommendation. Embedding-based approaches like CLIP enable efficient large-scale retrieval via vector similarity search, but they are primarily trained on literal caption-like text--image pairs and often fail to capture abstract or persona-driven attributes common in product recommendation applications (e.g., ``a gift for a mother who loves gardening''). In contrast, state-of-the-art vision--language models (vLLMs) can align text with images in a flexible manner, but their limited context window prevents them from directly handling retrieval over large catalogs. We propose a framework that distills the preference rankings of a powerful vLLM into an embedding-based system, transferring its nuanced alignment abilities while maintaining the inference-time scalability of an embedding-based approach. Experiments on persona-driven product recommendation tasks demonstrate that our method significantly outperforms existing embedding-based baselines, providing an efficient solution for personalized text--image retrieval.
☆ Evaluating Retrieval-Augmented Generation Systems on Unanswerable, Uncheatable, Realistic, Multi-hop Queries
Real-world use cases often present RAG systems with complex queries for which relevant information is missing from the corpus or is incomplete. In these settings, RAG systems must be able to reject unanswerable, out-of-scope queries and identify failures of retrieval and multi-hop reasoning. Despite this, existing RAG benchmarks rarely reflect realistic task complexity for multi-hop or out-of-scope questions, which often can be cheated via disconnected reasoning (i.e., solved without genuine multi-hop inference) or require only simple factual recall. This limits the ability for such benchmarks to uncover limitations of existing RAG systems. To address this gap, we present the first pipeline for automatic, difficulty-controlled creation of un$\underline{c}$heatable, $\underline{r}$ealistic, $\underline{u}$nanswerable, and $\underline{m}$ulti-hop $\underline{q}$uerie$\underline{s}$ (CRUMQs), adaptable to any corpus and domain. We use our pipeline to create CRUMQs over two popular RAG datasets and demonstrate its effectiveness via benchmark experiments on leading retrieval-augmented LLMs. Results show that compared to prior RAG benchmarks, CRUMQs are highly challenging for RAG systems and achieve up to 81.0\% reduction in cheatability scores. More broadly, our pipeline offers a simple way to enhance benchmark difficulty and realism and drive development of more capable RAG systems.
♻ ☆ WebThinker: Empowering Large Reasoning Models with Deep Research Capability NeurIPS 2025
Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose WebThinker, a deep research agent that empowers LRMs to autonomously search the web, navigate among web pages, and draft reports during the reasoning process. WebThinker integrates a Deep Web Explorer module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an Autonomous Think-Search-and-Draft strategy, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an RL-based training strategy via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.
comment: Accepted by NeurIPS 2025
♻ ☆ MVIGER: Multi-View Variational Integration of Complementary Knowledge for Generative Recommender
Language Models (LMs) have been widely used in recommender systems to incorporate textual information of items into item IDs, leveraging their advanced language understanding and generation capabilities. Recently, generative recommender systems have utilized the reasoning abilities of LMs to directly generate index tokens for potential items of interest based on the user's interaction history. To inject diverse item knowledge into LMs, prompt templates with detailed task descriptions and various indexing techniques derived from diverse item information have been explored. This paper focuses on the inconsistency in outputs generated by variations in input prompt templates and item index types, even with the same user's interaction history. Our in-depth quantitative analysis reveals that preference knowledge learned from diverse prompt templates and heterogeneous indices differs significantly, indicating a high potential for complementarity. To fully exploit this complementarity and provide consistent performance under varying prompts and item indices, we propose MVIGER, a unified variational framework that models selection among these information sources as a categorical latent variable with a learnable prior. During inference, this prior enables the model to adaptively select the most relevant source or aggregate predictions across multiple sources, thereby ensuring high-quality recommendation across diverse template-index combinations. We validate the effectiveness of MVIGER on three real-world datasets, demonstrating its superior performance over existing generative recommender baselines through the effective integration of complementary knowledge.
♻ ☆ SaraCoder: Orchestrating Semantic and Structural Cues for Resource-Optimized Repository-Level Code Completion
Despite Retrieval-Augmented Generation improving code completion, traditional retrieval methods struggle with information redundancy and a lack of diversity within limited context windows. To solve this, we propose a resource-optimized retrieval augmentation method, SaraCoder. It maximizes information diversity and representativeness in a limited context window, significantly boosting the accuracy and reliability of repository-level code completion. Its core Hierarchical Feature Optimization module systematically refines candidates by distilling deep semantic relationships, pruning exact duplicates, assessing structural similarity with a novel graph-based metric that weighs edits by their topological importance, and reranking results to maximize both relevance and diversity. Furthermore, an External-Aware Identifier Disambiguator module accurately resolves cross-file symbol ambiguity via dependency analysis. Extensive experiments on the challenging CrossCodeEval and RepoEval-Updated benchmarks demonstrate that SaraCoder outperforms existing baselines across multiple programming languages and models. Our work proves that systematically refining retrieval results across multiple dimensions provides a new paradigm for building more accurate and resource-optimized repository-level code completion systems.
♻ ☆ TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems NeurIPS 2025
Regression models are crucial in recommender systems. However, retransformation bias problem has been conspicuously neglected within the community. While many works in other fields have devised effective bias correction methods, all of them are post-hoc cures externally to the model, facing practical challenges when applied to real-world recommender systems. Hence, we propose a preemptive paradigm to eradicate the bias intrinsically from the models via minor model refinement. Specifically, a novel TranSUN method is proposed with a joint bias learning manner to offer theoretically guaranteed unbiasedness under empirical superior convergence. It is further generalized into a novel generic regression model family, termed Generalized TranSUN (GTS), which not only offers more theoretical insights but also serves as a generic framework for flexibly developing various bias-free models. Comprehensive experimental results demonstrate the superiority of our methods across data from various domains, which have been successfully deployed in two real-world industrial recommendation scenarios, i.e. product and short video recommendation scenarios in Guess What You Like business domain in the homepage of Taobao App (a leading e-commerce platform with DAU > 300M), to serve the major online traffic.
comment: 30 pages, 7 figures, NeurIPS 2025 Poster
♻ ☆ Scenario-Wise Rec: A Multi-Scenario Recommendation Benchmark CIKM'2025
Multi Scenario Recommendation (MSR) tasks, referring to building a unified model to enhance performance across all recommendation scenarios, have recently gained much attention. However, current research in MSR faces two significant challenges that hinder the field's development: the absence of uniform procedures for multi-scenario dataset processing, thus hindering fair comparisons, and most models being closed-sourced, which complicates comparisons with current SOTA models. Consequently, we introduce our benchmark, \textbf{Scenario-Wise Rec}, which comprises 6 public datasets and 12 benchmark models, along with a training and evaluation pipeline. Additionally, we validated the benchmark using an industrial advertising dataset, reinforcing its reliability and applicability in real-world scenarios. We aim for this benchmark to offer researchers valuable insights from prior work, enabling the development of novel models based on our benchmark and thereby fostering a collaborative research ecosystem in MSR. Our source code is also publicly available.
comment: Accepted to CIKM'2025
♻ ☆ HAMUR: Hyper Adapter for Multi-Domain Recommendation CIKM'2023
Multi-Domain Recommendation (MDR) has gained significant attention in recent years, which leverages data from multiple domains to enhance their performance concurrently.However, current MDR models are confronted with two limitations. Firstly, the majority of these models adopt an approach that explicitly shares parameters between domains, leading to mutual interference among them. Secondly, due to the distribution differences among domains, the utilization of static parameters in existing methods limits their flexibility to adapt to diverse domains. To address these challenges, we propose a novel model Hyper Adapter for Multi-Domain Recommendation (HAMUR). Specifically, HAMUR consists of two components: (1). Domain-specific adapter, designed as a pluggable module that can be seamlessly integrated into various existing multi-domain backbone models, and (2). Domain-shared hyper-network, which implicitly captures shared information among domains and dynamically generates the parameters for the adapter. We conduct extensive experiments on two public datasets using various backbone networks. The experimental results validate the effectiveness and scalability of the proposed model.
comment: Accepted by CIKM'2023
♻ ☆ ChoirRec: Semantic User Grouping via LLMs for Conversion Rate Prediction of Low-Activity Users
Accurately predicting conversion rates (CVR) for low-activity users remains a fundamental challenge in large-scale e-commerce recommender systems. Existing approaches face three critical limitations: (i) reliance on noisy and unreliable behavioral signals; (ii) insufficient user-level information due to the lack of diverse interaction data; and (iii) a systemic training bias toward high-activity users that overshadows the needs of low-activity users. To address these challenges, we propose ChoirRec, a novel framework that leverages the semantic capabilities of Large Language Models (LLMs) to construct semantic user groups and enhance CVR prediction for low-activity users. With a dual-channel architecture designed for robust cross-user knowledge transfer, ChoirRec comprises three components: (i) a Semantic Group Generation module that utilizes LLMs to form reliable, cross-activity user clusters, thereby filtering out noisy signals; (ii) a Group-aware Hierarchical Representation module that enriches sparse user embeddings with informative group-level priors to mitigate data insufficiency; and (iii) a Group-aware Multi-granularity Modual that employs a dual-channel architecture and adaptive fusion mechanism to ensure effective learning and utilization of group knowledge. We conduct extensive offline and online experiments on Taobao, a leading industrial-scale e-commerce platform. ChoirRec improves GAUC by 1.16\% in offline evaluations, while online A/B testing reveals a 7.24\% increase in order volume, highlighting its substantial practical value in real-world applications.
♻ ☆ SHERLOCK: Towards Dynamic Knowledge Adaptation in LLM-enhanced E-commerce Risk Management
The growth of the e-commerce industry has intensified the adversarial dynamics between shadow economy actors and risk management teams. Companies often conduct risk investigations into suspicious cases to identify emerging fraud patterns, thereby enhancing both preemptive risk prevention and post-hoc governance. However, the sheer volume of case analyses imposes a substantial workload on risk management analysts, as each case requires the integration of long-term expert experience and meticulous scrutiny across multiple risk dimensions. Additionally, individual disparities among analysts hinder the establishment of uniform and high-standard workflows. To address these challenges, we propose the SHERLOCK framework, which leverages the reasoning capabilities of large language models (LLMs) to assist analysts in risk investigations. Our approach consists of three primary components: (1) extracting risk management knowledge from multi-modal data and constructing a domain knowledge base (KB), (2) building an intelligent platform guided by the data flywheel paradigm that integrates daily operations, expert annotations, and model evaluations, with iteratively fine-tuning for preference alignment, and (3) introducing a Reflect & Refine (R&R) module that collaborates with the domain KB to establish a rapid response mechanism for evolving risk patterns. Experiments conducted on the real-world transaction dataset from JD dot com demonstrate that our method significantly improves the precision of both factual alignment and risk localization within the LLM analysis results. Deployment of the SHERLOCK-based LLM system on JD dot com has substantially enhanced the efficiency of case investigation workflows for risk managers.
♻ ☆ Doc2Query++: Topic-Coverage based Document Expansion and its Application to Dense Retrieval via Dual-Index Fusion
Document expansion (DE) via query generation tackles vocabulary mismatch in sparse retrieval, yet faces limitations: uncontrolled generation producing hallucinated or redundant queries with low diversity; poor generalization from in-domain training (e.g., MS MARCO) to out-of-domain data like BEIR; and noise from concatenation harming dense retrieval. While Large Language Models (LLMs) enable cross-domain query generation, basic prompting lacks control, and taxonomy-based methods rely on domain-specific structures, limiting applicability. To address these challenges, we introduce Doc2Query++, a DE framework that structures query generation by first inferring a document's latent topics via unsupervised topic modeling for cross-domain applicability, then using hybrid keyword selection to create a diverse and relevant keyword set per document. This guides LLM not only to leverage keywords, which ensure comprehensive topic representation, but also to reduce redundancy through diverse, relevant terms. To prevent noise from query appending in dense retrieval, we propose Dual-Index Fusion strategy that isolates text and query signals, boosting performance in dense settings. Extensive experiments show Doc2Query++ significantly outperforms state-of-the-art baselines, achieving substantial gains in MAP, nDCG@10 and Recall@100 across diverse datasets on both sparse and dense retrieval.
comment: 11 pages, 4 figures
♻ ☆ A Comprehensive Review of Recommender Systems: Transitioning from Theory to Practice
Recommender Systems (RS) play an integral role in enhancing user experiences by providing personalized item suggestions. This survey reviews the progress in RS inclusively from 2017 to 2024, effectively connecting theoretical advances with practical applications. We explore the development from traditional RS techniques like content-based and collaborative filtering to advanced methods involving deep learning, graph-based models, reinforcement learning, and large language models. We also discuss specialized systems such as context-aware, review-based, and fairness-aware RS. The primary goal of this survey is to bridge theory with practice. It addresses challenges across various sectors, including e-commerce, healthcare, and finance, emphasizing the need for scalable, real-time, and trustworthy solutions. Through this survey, we promote stronger partnerships between academic research and industry practices. The insights offered by this survey aim to guide industry professionals in optimizing RS deployment and to inspire future research directions, especially in addressing emerging technological and societal trends\footnote. The survey resources are available in the public GitHub repository https://github.com/VectorInstitute/Recommender-Systems-Survey. (Recommender systems, large language models, chatgpt, responsible AI)
comment: we quarterly update of this literature
♻ ☆ Agent Learning via Early Experience
A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.
comment: Work in progress
♻ ☆ Revela: Dense Retriever Learning via Language Modeling
Dense retrievers play a vital role in accessing external and specialized knowledge to augment language models (LMs). Training dense retrievers typically requires annotated query-document pairs, which are costly to create and scarce in specialized domains (e.g., code) or in complex settings (e.g., requiring reasoning). These practical challenges have sparked growing interest in self-supervised retriever learning. Since LMs are trained to capture token-level dependencies through a self-supervised learning objective (i.e., next token prediction), we can analogously cast retrieval as learning dependencies among chunks of tokens. This analogy naturally leads to the question: How can we adapt self-supervised learning objectives in the spirit of language modeling to train retrievers? To answer this question, we introduce Revela, a unified and scalable training framework for self-supervised retriever learning via language modeling. Revela models semantic dependencies among documents by conditioning next token prediction on local and cross-document context through an in-batch attention mechanism. This attention is weighted by retriever-computed similarity scores, enabling the retriever to be optimized as part of language modeling. We evaluate Revela on domain-specific (CoIR), reasoning-intensive (BRIGHT), and general-domain (BEIR) benchmarks across various retriever backbones. Without annotated or synthetic query-document pairs, Revela surpasses larger supervised models and proprietary APIs on CoIR and matches them on BRIGHT. It achieves BEIR's unsupervised SoTA with ~ 1000x less training data and 10x less compute. Performance increases with batch size and model size, highlighting Revela's scalability and its promise for self-supervised retriever learning.
Machine Learning 150
☆ Are Large Reasoning Models Interruptible?
Large Reasoning Models (LRMs) excel at complex reasoning but are traditionally evaluated in static, "frozen world" settings: model responses are assumed to be instantaneous, and the context of a request is presumed to be immutable over the duration of the response. While generally true for short-term tasks, the "frozen world" assumption breaks down in modern reasoning tasks such as assistive programming, where models may take hours to think through problems and code may change dramatically from the time the model starts thinking to the model's final output. In this work, we challenge the frozen world assumption and evaluate LRM robustness under two realistic dynamic scenarios: interruptions, which test the quality of the model's partial outputs on a limited budget, and dynamic context, which tests model adaptation to in-flight changes. Across mathematics and programming benchmarks that require long-form reasoning, static evaluations consistently overestimate robustness: even state-of-the-art LRMs, which achieve high accuracy in static settings, can fail unpredictably when interrupted or exposed to changing context, with performance dropping by up to 60% when updates are introduced late in the reasoning process. Our analysis further reveals several novel failure modes, including reasoning leakage, where models fold the reasoning into their final answer when interrupted; panic, where under time pressure models abandon reasoning entirely and return incorrect answers; and self-doubt, where performance degrades while incorporating updated information.
comment: Project Page: https://dynamic-lm.github.io/
☆ Reinforced sequential Monte Carlo for amortised sampling
This paper proposes a synergy of amortised and particle-based methods for sampling from distributions defined by unnormalised density functions. We state a connection between sequential Monte Carlo (SMC) and neural sequential samplers trained by maximum-entropy reinforcement learning (MaxEnt RL), wherein learnt sampling policies and value functions define proposal kernels and twist functions. Exploiting this connection, we introduce an off-policy RL training procedure for the sampler that uses samples from SMC -- using the learnt sampler as a proposal -- as a behaviour policy that better explores the target distribution. We describe techniques for stable joint training of proposals and twist functions and an adaptive weight tempering scheme to reduce training signal variance. Furthermore, building upon past attempts to use experience replay to guide the training of neural samplers, we derive a way to combine historical samples with annealed importance sampling weights within a replay buffer. On synthetic multi-modal targets (in both continuous and discrete spaces) and the Boltzmann distribution of alanine dipeptide conformations, we demonstrate improvements in approximating the true distribution as well as training stability compared to both amortised and Monte Carlo methods.
comment: code: https://github.com/hyeok9855/gfn-smc-jax
☆ Adversarial Attacks Leverage Interference Between Features in Superposition
Fundamental questions remain about when and why adversarial examples arise in neural networks, with competing views characterising them either as artifacts of the irregularities in the decision landscape or as products of sensitivity to non-robust input features. In this paper, we instead argue that adversarial vulnerability can stem from efficient information encoding in neural networks. Specifically, we show how superposition - where networks represent more features than they have dimensions - creates arrangements of latent representations that adversaries can exploit. We demonstrate that adversarial perturbations leverage interference between superposed features, making attack patterns predictable from feature arrangements. Our framework provides a mechanistic explanation for two known phenomena: adversarial attack transferability between models with similar training regimes and class-specific vulnerability patterns. In synthetic settings with precisely controlled superposition, we establish that superposition suffices to create adversarial vulnerability. We then demonstrate that these findings persist in a ViT trained on CIFAR-10. These findings reveal adversarial vulnerability can be a byproduct of networks' representational compression, rather than flaws in the learning process or non-robust inputs.
QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs
We propose QeRL, a Quantization-enhanced Reinforcement Learning framework for large language models (LLMs). While RL is essential for LLMs' reasoning capabilities, it is resource-intensive, requiring substantial GPU memory and long rollout durations. QeRL addresses these issues by combining NVFP4 quantization with Low-Rank Adaptation (LoRA), accelerating rollout phase of RL while reducing memory overhead. Beyond efficiency, our findings show that quantization noise increases policy entropy, enhancing exploration, and enabling the discovery of better strategies during RL. To further optimize exploration, QeRL introduces an Adaptive Quantization Noise (AQN) mechanism, which dynamically adjusts noise during training. Experiments demonstrate that QeRL delivers over 1.5 times speedup in the rollout phase. Moreover, this is the first framework to enable RL training of a 32B LLM on a single H100 80GB GPU, while delivering overall speedups for RL training. It also achieves faster reward growth and higher final accuracy than 16-bit LoRA and QLoRA, while matching the performance of full-parameter fine-tuning on mathematical benchmarks such as GSM8K (90.8%) and MATH 500 (77.4%) in the 7B model. These results establish QeRL as an efficient and effective framework for RL training in LLMs.
comment: Code is available at https://github.com/NVlabs/QeRL
☆ Tight Regret Upper and Lower Bounds for Optimistic Hedge in Two-Player Zero-Sum Games
In two-player zero-sum games, the learning dynamic based on optimistic Hedge achieves one of the best-known regret upper bounds among strongly-uncoupled learning dynamics. With an appropriately chosen learning rate, the social and individual regrets can be bounded by $O(\log(mn))$ in terms of the numbers of actions $m$ and $n$ of the two players. This study investigates the optimality of the dependence on $m$ and $n$ in the regret of optimistic Hedge. To this end, we begin by refining existing regret analysis and show that, in the strongly-uncoupled setting where the opponent's number of actions is known, both the social and individual regret bounds can be improved to $O(\sqrt{\log m \log n})$. In this analysis, we express the regret upper bound as an optimization problem with respect to the learning rates and the coefficients of certain negative terms, enabling refined analysis of the leading constants. We then show that the existing social regret bound as well as these new social and individual regret upper bounds cannot be further improved for optimistic Hedge by providing algorithm-dependent individual regret lower bounds. Importantly, these social regret upper and lower bounds match exactly including the constant factor in the leading term. Finally, building on these results, we improve the last-iterate convergence rate and the dynamic regret of a learning dynamic based on optimistic Hedge, and complement these bounds with algorithm-dependent dynamic regret lower bounds that match the improved bounds.
comment: 29 pages, 2 figures
☆ Diffusion Transformers with Representation Autoencoders
Latent generative modeling, where a pretrained autoencoder maps pixels into a latent space for the diffusion process, has become the standard strategy for Diffusion Transformers (DiT); however, the autoencoder component has barely evolved. Most DiTs continue to rely on the original VAE encoder, which introduces several limitations: outdated backbones that compromise architectural simplicity, low-dimensional latent spaces that restrict information capacity, and weak representations that result from purely reconstruction-based training and ultimately limit generative quality. In this work, we explore replacing the VAE with pretrained representation encoders (e.g., DINO, SigLIP, MAE) paired with trained decoders, forming what we term Representation Autoencoders (RAEs). These models provide both high-quality reconstructions and semantically rich latent spaces, while allowing for a scalable transformer-based architecture. Since these latent spaces are typically high-dimensional, a key challenge is enabling diffusion transformers to operate effectively within them. We analyze the sources of this difficulty, propose theoretically motivated solutions, and validate them empirically. Our approach achieves faster convergence without auxiliary representation alignment losses. Using a DiT variant equipped with a lightweight, wide DDT head, we achieve strong image generation results on ImageNet: 1.51 FID at 256x256 (no guidance) and 1.13 at both 256x256 and 512x512 (with guidance). RAE offers clear advantages and should be the new default for diffusion transformer training.
comment: Technical Report; Project Page: https://rae-dit.github.io/
☆ Representation-Based Exploration for Language Models: From Test-Time to Post-Training
Reinforcement learning (RL) promises to expand the capabilities of language models, but it is unclear if current RL techniques promote the discovery of novel behaviors, or simply sharpen those already present in the base model. In this paper, we investigate the value of deliberate exploration -- explicitly incentivizing the model to discover novel and diverse behaviors -- and aim to understand how the knowledge in pre-trained models can guide this search. Our main finding is that exploration with a simple, principled, representation-based bonus derived from the pre-trained language model's hidden states significantly improves diversity and pass@k rates -- both for post-training, and in a novel inference-time scaling setting we introduce. For inference-time, exploration with representation-based diversity improves efficiency, consistently improving pass@k rates across a variety of models and reasoning tasks. For example, for Qwen-2.5-14b-Instruct we obtain over 50% improvement in verifier efficiency on almost all tasks. For post-training, we show that integrating this exploration strategy into an RL pipeline improves reasoning performance over that of the initial model and over standard RL post-training. For example, on AIME 2024, our post-trained Qwen-2.5-7b-Instruct's pass@80 matches the pass@256 of GRPO on the same model, demonstrating a 3x improvement in test-time sample efficiency. Overall, our findings suggest that deliberate exploration -- with the right notion of diversity -- is a practical path toward discovery of new behaviors beyond sharpening.
comment: Website and code: https://rep-exp.github.io
☆ Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
A key challenge in applying reinforcement learning (RL) to diffusion large language models (dLLMs) lies in the intractability of their likelihood functions, which are essential for the RL objective, necessitating corresponding approximation in each training step. While existing methods approximate the log-likelihoods by their evidence lower bounds (ELBOs) via customized Monte Carlo (MC) sampling, the forward computational graphs of all MC samples need to be retained for the gradient computation of non-linear terms in the RL objective, resulting in significant memory overhead. This constraint restricts feasible sample sizes, leading to imprecise likelihood approximations and ultimately distorting the RL objective. To overcome this limitation, we propose \emph{Boundary-Guided Policy Optimization} (BGPO), a memory-efficient RL algorithm that maximizes a specially constructed lower bound of the ELBO-based objective. This lower bound is carefully designed to satisfy two key properties: (1) Linearity: it is formulated in a linear sum where each term depends only on a single MC sample, thereby enabling gradient accumulation across samples and ensuring constant memory usage; (2) Equivalence: Both the value and gradient of this lower bound are equal to those of the ELBO-based objective in on-policy training, making it also an effective approximation for the original RL objective. These properties allow BGPO to adopt a large MC sample size, resulting in more accurate likelihood approximations and improved RL objective estimation, which in turn leads to enhanced performance. Experiments show that BGPO significantly outperforms previous RL algorithms for dLLMs in math problem solving, code generation, and planning tasks.
☆ Chronologically Consistent Generative AI
We introduce a family of chronologically consistent, instruction-following large language models to eliminate lookahead bias. Each model is trained only on data available before a clearly defined knowledge-cutoff date, ensuring strict temporal separation from any post-cutoff data. The resulting framework offers (i) a simple, conversational chat interface, (ii) fully open, fixed model weights that guarantee replicability, and (iii) a conservative lower bound on forecast accuracy, isolating the share of predictability that survives once training leakage is removed. Together, these features provide researchers with an easy-to-use generative AI tool useful for a wide range of prediction tasks that is free of lookahead bias.
☆ Accelerated stochastic first-order method for convex optimization under heavy-tailed noise
We study convex composite optimization problems, where the objective function is given by the sum of a prox-friendly function and a convex function whose subgradients are estimated under heavy-tailed noise. Existing work often employs gradient clipping or normalization techniques in stochastic first-order methods to address heavy-tailed noise. In this paper, we demonstrate that a vanilla stochastic algorithm -- without additional modifications such as clipping or normalization -- can achieve optimal complexity for these problems. In particular, we establish that an accelerated stochastic proximal subgradient method achieves a first-order oracle complexity that is universally optimal for smooth, weakly smooth, and nonsmooth convex optimization, as well as for stochastic convex optimization under heavy-tailed noise. Numerical experiments are further provided to validate our theoretical results.
☆ An Eulerian Perspective on Straight-Line Sampling
We study dynamic measure transport for generative modeling: specifically, flows induced by stochastic processes that bridge a specified source and target distribution. The conditional expectation of the process' velocity defines an ODE whose flow map achieves the desired transport. We ask \emph{which processes produce straight-line flows} -- i.e., flows whose pointwise acceleration vanishes and thus are exactly integrable with a first-order method? We provide a concise PDE characterization of straightness as a balance between conditional acceleration and the divergence of a weighted covariance (Reynolds) tensor. Using this lens, we fully characterize affine-in-time interpolants and show that straightness occurs exactly under deterministic endpoint couplings. We also derive necessary conditions that constrain flow geometry for general processes, offering broad guidance for designing transports that are easier to integrate.
☆ MATH-Beyond: A Benchmark for RL to Expand Beyond the Base Model
With the advent of DeepSeek-R1, a new wave of reinforcement learning (RL) methods has emerged that seem to unlock stronger mathematical reasoning. However, a closer look at the open-source ecosystem reveals a critical limitation: with sufficiently many draws (e.g., $\texttt{pass@1024}$), many existing base models already solve nearly all questions on widely used math benchmarks such as MATH-500 and AIME 2024. This suggests that the RL fine-tuning methods prevalent in the LLM reasoning literature largely sharpen existing solution modes rather than discovering entirely new ones. Such sharpening stands in contrast to the broader promise of RL: to foster exploration and to acquire new skills. To move beyond this plateau, we introduce MATH-Beyond (MATH-B), a benchmark deliberately constructed to defeat common open-source models of up to 8B parameters even under large sampling budgets. Improving performance on our benchmark via RL requires methods that learn to reason in ways that go beyond base model capabilities in repeated sampling. Since the problems are drawn from subsets of DAPO-Math-17K and DeepScaleR datasets, they remain topically equivalent to standard high-school math. Validating our premise, RL fine-tuned models such as Nemotron-Research-Reasoning-Qwen-1.5B and DeepScaleR-1.5B-Preview perform poorly on MATH-B at $\texttt{pass@1024}$, showing how existing approaches fall short on tackling harder instances. We hope MATH-B will catalyze exploration-driven RL approaches that elicit deeper reasoning capabilities. We release MATH-B at https://huggingface.co/datasets/brendel-group/MATH-Beyond.
☆ Continual Release of Densest Subgraphs: Privacy Amplification & Sublinear Space via Subsampling
We study the sublinear space continual release model for edge-differentially private (DP) graph algorithms, with a focus on the densest subgraph problem (DSG) in the insertion-only setting. Our main result is the first continual release DSG algorithm that matches the additive error of the best static DP algorithms and the space complexity of the best non-private streaming algorithms, up to constants. The key idea is a refined use of subsampling that simultaneously achieves privacy amplification and sparsification, a connection not previously formalized in graph DP. Via a simple black-box reduction to the static setting, we obtain both pure and approximate-DP algorithms with $O(\log n)$ additive error and $O(n\log n)$ space, improving both accuracy and space complexity over the previous state of the art. Along the way, we introduce graph densification in the graph DP setting, adding edges to trigger earlier subsampling, which removes the extra logarithmic factors in error and space incurred by prior work [ELMZ25]. We believe this simple idea may be of independent interest.
comment: to be published in SOSA'26
☆ NV3D: Leveraging Spatial Shape Through Normal Vector-based 3D Object Detection
Recent studies in 3D object detection for autonomous vehicles aim to enrich features through the utilization of multi-modal setups or the extraction of local patterns within LiDAR point clouds. However, multi-modal methods face significant challenges in feature alignment, and gaining features locally can be oversimplified for complex 3D object detection tasks. In this paper, we propose a novel model, NV3D, which utilizes local features acquired from voxel neighbors, as normal vectors computed per voxel basis using K-nearest neighbors (KNN) and principal component analysis (PCA). This informative feature enables NV3D to determine the relationship between the surface and pertinent target entities, including cars, pedestrians, or cyclists. During the normal vector extraction process, NV3D offers two distinct sampling strategies: normal vector density-based sampling and FOV-aware bin-based sampling, allowing elimination of up to 55% of data while maintaining performance. In addition, we applied element-wise attention fusion, which accepts voxel features as the query and value and normal vector features as the key, similar to the attention mechanism. Our method is trained on the KITTI dataset and has demonstrated superior performance in car and cyclist detection owing to their spatial shapes. In the validation set, NV3D without sampling achieves 86.60% and 80.18% mean Average Precision (mAP), greater than the baseline Voxel R-CNN by 2.61% and 4.23% mAP, respectively. With both samplings, NV3D achieves 85.54% mAP in car detection, exceeding the baseline by 1.56% mAP, despite roughly 55% of voxels being filtered out.
☆ Lecture Notes on Verifying Graph Neural Networks
In these lecture notes, we first recall the connection between graph neural networks, Weisfeiler-Lehman tests and logics such as first-order logic and graded modal logic. We then present a modal logic in which counting modalities appear in linear inequalities in order to solve verification tasks on graph neural networks. We describe an algorithm for the satisfiability problem of that logic. It is inspired from the tableau method of vanilla modal logic, extended with reasoning in quantifier-free fragment Boolean algebra with Presburger arithmetic.
☆ Attention Factors for Statistical Arbitrage
Statistical arbitrage exploits temporal price differences between similar assets. We develop a framework to jointly identify similar assets through factors, identify mispricing and form a trading policy that maximizes risk-adjusted performance after trading costs. Our Attention Factors are conditional latent factors that are the most useful for arbitrage trading. They are learned from firm characteristic embeddings that allow for complex interactions. We identify time-series signals from the residual portfolios of our factors with a general sequence model. Estimating factors and the arbitrage trading strategy jointly is crucial to maximize profitability after trading costs. In a comprehensive empirical study we show that our Attention Factor model achieves an out-of-sample Sharpe ratio above 4 on the largest U.S. equities over a 24-year period. Our one-step solution yields an unprecedented Sharpe ratio of 2.3 net of transaction costs. We show that weak factors are important for arbitrage trading.
comment: Accepted to the 6th ACM International Conference on AI in Finance
☆ Deconstructing Attention: Investigating Design Principles for Effective Language Modeling
The success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.
☆ SemCSE-Multi: Multifaceted and Decodable Embeddings for Aspect-Specific and Interpretable Scientific Domain Mapping
We propose SemCSE-Multi, a novel unsupervised framework for generating multifaceted embeddings of scientific abstracts, evaluated in the domains of invasion biology and medicine. These embeddings capture distinct, individually specifiable aspects in isolation, thus enabling fine-grained and controllable similarity assessments as well as adaptive, user-driven visualizations of scientific domains. Our approach relies on an unsupervised procedure that produces aspect-specific summarizing sentences and trains embedding models to map semantically related summaries to nearby positions in the embedding space. We then distill these aspect-specific embedding capabilities into a unified embedding model that directly predicts multiple aspect embeddings from a scientific abstract in a single, efficient forward pass. In addition, we introduce an embedding decoding pipeline that decodes embeddings back into natural language descriptions of their associated aspects. Notably, we show that this decoding remains effective even for unoccupied regions in low-dimensional visualizations, thus offering vastly improved interpretability in user-centric settings.
☆ Hierarchical Qubit-Merging Transformer for Quantum Error Correction
For reliable large-scale quantum computation, a quantum error correction (QEC) scheme must effectively resolve physical errors to protect logical information. Leveraging recent advances in deep learning, neural network-based decoders have emerged as a promising approach to enhance the reliability of QEC. We propose the Hierarchical Qubit-Merging Transformer (HQMT), a novel and general decoding framework that explicitly leverages the structural graph of stabilizer codes to learn error correlations across multiple scales. Our architecture first computes attention locally on structurally related groups of stabilizers and then systematically merges these qubit-centric representations to build a global view of the error syndrome. The proposed HQMT achieves substantially lower logical error rates for surface codes by integrating a dedicated qubit-merging layer within the transformer architecture. Across various code distances, HQMT significantly outperforms previous neural network-based QEC decoders as well as a powerful belief propagation with ordered statistics decoding (BP+OSD) baseline. This hierarchical approach provides a scalable and effective framework for surface code decoding, advancing the realization of reliable quantum computing.
comment: 6 pages, 5 figures
☆ Diffusion-DFL: Decision-focused Diffusion Models for Stochastic Optimization
Decision-focused learning (DFL) integrates predictive modeling and optimization by training predictors to optimize the downstream decision target rather than merely minimizing prediction error. To date, existing DFL methods typically rely on deterministic point predictions, which are often insufficient to capture the intrinsic stochasticity of real-world environments. To address this challenge, we propose the first diffusion-based DFL approach, which trains a diffusion model to represent the distribution of uncertain parameters and optimizes the decision by solving a stochastic optimization with samples drawn from the diffusion model. Our contributions are twofold. First, we formulate diffusion DFL using the reparameterization trick, enabling end-to-end training through diffusion. While effective, it is memory and compute-intensive due to the need to differentiate through the diffusion sampling process. Second, we propose a lightweight score function estimator that uses only several forward diffusion passes and avoids backpropagation through the sampling. This follows from our results that backpropagating through stochastic optimization can be approximated by a weighted score function formulation. We empirically show that our diffusion DFL approach consistently outperforms strong baselines in decision quality. The source code for all experiments is available at the project repository: https://github.com/GT-KOALA/Diffusion_DFL.
☆ MS-Mix: Unveiling the Power of Mixup for Multimodal Sentiment Analysis
Multimodal Sentiment Analysis (MSA) aims to identify and interpret human emotions by integrating information from heterogeneous data sources such as text, video, and audio. While deep learning models have advanced in network architecture design, they remain heavily limited by scarce multimodal annotated data. Although Mixup-based augmentation improves generalization in unimodal tasks, its direct application to MSA introduces critical challenges: random mixing often amplifies label ambiguity and semantic inconsistency due to the lack of emotion-aware mixing mechanisms. To overcome these issues, we propose MS-Mix, an adaptive, emotion-sensitive augmentation framework that automatically optimizes sample mixing in multimodal settings. The key components of MS-Mix include: (1) a Sentiment-Aware Sample Selection (SASS) strategy that effectively prevents semantic confusion caused by mixing samples with contradictory emotions. (2) a Sentiment Intensity Guided (SIG) module using multi-head self-attention to compute modality-specific mixing ratios dynamically based on their respective emotional intensities. (3) a Sentiment Alignment Loss (SAL) that aligns the prediction distributions across modalities, and incorporates the Kullback-Leibler-based loss as an additional regularization term to train the emotion intensity predictor and the backbone network jointly. Extensive experiments on three benchmark datasets with six state-of-the-art backbones confirm that MS-Mix consistently outperforms existing methods, establishing a new standard for robust multimodal sentiment augmentation. The source code is available at: https://github.com/HongyuZhu-s/MS-Mix.
comment: Under Review
☆ A Framework for Low-Effort Training Data Generation for Urban Semantic Segmentation
Synthetic datasets are widely used for training urban scene recognition models, but even highly realistic renderings show a noticeable gap to real imagery. This gap is particularly pronounced when adapting to a specific target domain, such as Cityscapes, where differences in architecture, vegetation, object appearance, and camera characteristics limit downstream performance. Closing this gap with more detailed 3D modelling would require expensive asset and scene design, defeating the purpose of low-cost labelled data. To address this, we present a new framework that adapts an off-the-shelf diffusion model to a target domain using only imperfect pseudo-labels. Once trained, it generates high-fidelity, target-aligned images from semantic maps of any synthetic dataset, including low-effort sources created in hours rather than months. The method filters suboptimal generations, rectifies image-label misalignments, and standardises semantics across datasets, transforming weak synthetic data into competitive real-domain training sets. Experiments on five synthetic datasets and two real target datasets show segmentation gains of up to +8.0%pt. mIoU over state-of-the-art translation methods, making rapidly constructed synthetic datasets as effective as high-effort, time-intensive synthetic datasets requiring extensive manual design. This work highlights a valuable collaborative paradigm where fast semantic prototyping, combined with generative models, enables scalable, high-quality training data creation for urban scene understanding.
☆ Ontolearn-A Framework for Large-scale OWL Class Expression Learning in Python
In this paper, we present Ontolearn-a framework for learning OWL class expressions over large knowledge graphs. Ontolearn contains efficient implementations of recent stateof-the-art symbolic and neuro-symbolic class expression learners including EvoLearner and DRILL. A learned OWL class expression can be used to classify instances in the knowledge graph. Furthermore, Ontolearn integrates a verbalization module based on an LLM to translate complex OWL class expressions into natural language sentences. By mapping OWL class expressions into respective SPARQL queries, Ontolearn can be easily used to operate over a remote triplestore. The source code of Ontolearn is available at https://github.com/dice-group/Ontolearn.
☆ Efficient Group Lasso Regularized Rank Regression with Data-Driven Parameter Determination
High-dimensional regression often suffers from heavy-tailed noise and outliers, which can severely undermine the reliability of least-squares based methods. To improve robustness, we adopt a non-smooth Wilcoxon score based rank objective and incorporate structured group sparsity regularization, a natural generalization of the lasso, yielding a group lasso regularized rank regression method. By extending the tuning-free parameter selection scheme originally developed for the lasso, we introduce a data-driven, simulation-based tuning rule and further establish a finite-sample error bound for the resulting estimator. On the computational side, we develop a proximal augmented Lagrangian method for solving the associated optimization problem, which eliminates the singularity issues encountered in existing methods, thereby enabling efficient semismooth Newton updates for the subproblems. Extensive numerical experiments demonstrate the robustness and effectiveness of our proposed estimator against alternatives, and showcase the scalability of the algorithm across both simulated and real-data settings.
comment: 36 pages, 4 figures, 8 tables
☆ Query-Specific GNN: A Comprehensive Graph Representation Learning Method for Retrieval Augmented Generation
Retrieval-augmented generation (RAG) has demonstrated its ability to enhance Large Language Models (LLMs) by integrating external knowledge sources. However, multi-hop questions, which require the identification of multiple knowledge targets to form a synthesized answer, raise new challenges for RAG systems. Under the multi-hop settings, existing methods often struggle to fully understand the questions with complex semantic structures and are susceptible to irrelevant noise during the retrieval of multiple information targets. To address these limitations, we propose a novel graph representation learning framework for multi-hop question retrieval. We first introduce a Multi-information Level Knowledge Graph (Multi-L KG) to model various information levels for a more comprehensive understanding of multi-hop questions. Based on this, we design a Query-Specific Graph Neural Network (QSGNN) for representation learning on the Multi-L KG. QSGNN employs intra/inter-level message passing mechanisms, and in each message passing the information aggregation is guided by the query, which not only facilitates multi-granular information aggregation but also significantly reduces the impact of noise. To enhance its ability to learn robust representations, we further propose two synthesized data generation strategies for pre-training the QSGNN. Extensive experimental results demonstrate the effectiveness of our framework in multi-hop scenarios, especially in high-hop questions the improvement can reach 33.8\%. The code is available at: https://github.com/Jerry2398/QSGNN.
☆ Automatic Music Sample Identification with Multi-Track Contrastive Learning
Sampling, the technique of reusing pieces of existing audio tracks to create new music content, is a very common practice in modern music production. In this paper, we tackle the challenging task of automatic sample identification, that is, detecting such sampled content and retrieving the material from which it originates. To do so, we adopt a self-supervised learning approach that leverages a multi-track dataset to create positive pairs of artificial mixes, and design a novel contrastive learning objective. We show that such method significantly outperforms previous state-of-the-art baselines, that is robust to various genres, and that scales well when increasing the number of noise songs in the reference database. In addition, we extensively analyze the contribution of the different components of our training pipeline and highlight, in particular, the need for high-quality separated stems for this task.
☆ Knowledge-Guided Machine Learning Models to Upscale Evapotranspiration in the U.S. Midwest
Evapotranspiration (ET) plays a critical role in the land-atmosphere interactions, yet its accurate quantification across various spatiotemporal scales remains a challenge. In situ measurement approaches, like eddy covariance (EC) or weather station-based ET estimation, allow for measuring ET at a single location. Agricultural uses of ET require estimates for each field over broad areas, making it infeasible to deploy sensing systems at each location. This study integrates tree-based and knowledge-guided machine learning (ML) techniques with multispectral remote sensing data, griddled meteorology and EC data to upscale ET across the Midwest United States. We compare four tree-based models - Random Forest, CatBoost, XGBoost, LightGBM - and a simple feed-forward artificial neural network in combination with features engineered using knowledge-guided ML principles. Models were trained and tested on EC towers located in the Midwest of the United States using k-fold cross validation with k=5 and site-year, biome stratified train-test split to avoid data leakage. Results show that LightGBM with knowledge-guided features outperformed other methods with an R2=0.86, MSE=14.99 W m^-2 and MAE = 8.82 W m^-2 according to grouped k-fold validation (k=5). Feature importance analysis shows that knowledge-guided features were most important for predicting evapotranspiration. Using the best performing model, we provide a data product at 500 m spatial and one-day temporal resolution for gridded ET for the period of 2019-2024. Intercomparison between the new gridded product and state-level weather station-based ET estimates show best-in-class correspondence.
☆ Learning to Make MISTAKEs: Modeling Incorrect Student Thinking And Key Errors
Research on reasoning in language models (LMs) predominantly focuses on improving the correctness of their outputs. But some important applications require modeling reasoning patterns that are incorrect. For example, automated systems that can reason about and simulate student errors are useful for providing real-time feedback in the classroom or offline practice for educators-in-training. This paper presents a new method, MISTAKE, that (1) constructs high-quality synthetic examples of reasoning errors by leveraging cycle consistency between incorrect answers and latent misconceptions; and (2) uses the generated data to learn models for student simulation, misconception classification, and answer generation. We evaluate MISTAKE on three educational tasks and find that it results in (1) higher accuracy when simulating incorrect student answers based on specific misconceptions, (2) increased performance inferring latent misconceptions from observed incorrect answers, and (3) higher alignment with expert-written distractor answers when generating incorrect answers (e.g., for multiple-choice tests).
☆ Context-Aware Model-Based Reinforcement Learning for Autonomous Racing
Autonomous vehicles have shown promising potential to be a groundbreaking technology for improving the safety of road users. For these vehicles, as well as many other safety-critical robotic technologies, to be deployed in real-world applications, we require algorithms that can generalize well to unseen scenarios and data. Model-based reinforcement learning algorithms (MBRL) have demonstrated state-of-the-art performance and data efficiency across a diverse set of domains. However, these algorithms have also shown susceptibility to changes in the environment and its transition dynamics. In this work, we explore the performance and generalization capabilities of MBRL algorithms for autonomous driving, specifically in the simulated autonomous racing environment, Roboracer (formerly F1Tenth). We frame the head-to-head racing task as a learning problem using contextual Markov decision processes and parameterize the driving behavior of the adversaries using the context of the episode, thereby also parameterizing the transition and reward dynamics. We benchmark the behavior of MBRL algorithms in this environment and propose a novel context-aware extension of the existing literature, cMask. We demonstrate that context-aware MBRL algorithms generalize better to out-of-distribution adversary behaviors relative to context-free approaches. We also demonstrate that cMask displays strong generalization capabilities, as well as further performance improvement relative to other context-aware MBRL approaches when racing against adversaries with in-distribution behaviors.
comment: Accepted to IEEE ICAR 2025
☆ Offline Reinforcement Learning with Generative Trajectory Policies ICLR 2026
Generative models have emerged as a powerful class of policies for offline reinforcement learning (RL) due to their ability to capture complex, multi-modal behaviors. However, existing methods face a stark trade-off: slow, iterative models like diffusion policies are computationally expensive, while fast, single-step models like consistency policies often suffer from degraded performance. In this paper, we demonstrate that it is possible to bridge this gap. The key to moving beyond the limitations of individual methods, we argue, lies in a unifying perspective that views modern generative models, including diffusion, flow matching, and consistency models, as specific instances of learning a continuous-time generative trajectory governed by an Ordinary Differential Equation (ODE). This principled foundation provides a clearer design space for generative policies in RL and allows us to propose Generative Trajectory Policies (GTPs), a new and more general policy paradigm that learns the entire solution map of the underlying ODE. To make this paradigm practical for offline RL, we further introduce two key theoretically principled adaptations. Empirical results demonstrate that GTP achieves state-of-the-art performance on D4RL benchmarks - it significantly outperforms prior generative policies, achieving perfect scores on several notoriously hard AntMaze tasks.
comment: Preprint. Under review at ICLR 2026
☆ ReLook: Vision-Grounded RL with a Multimodal LLM Critic for Agentic Web Coding
While Large Language Models (LLMs) excel at algorithmic code generation, they struggle with front-end development, where correctness is judged on rendered pixels and interaction. We present ReLook, an agentic, vision-grounded reinforcement learning framework that empowers an agent to close a robust generate--diagnose--refine loop by invoking a multimodal LLM (MLLM) as a tool. During training, the agent uses the MLLM-in-the-loop both as a visual critic--scoring code with screenshots--and as a source of actionable, vision-grounded feedback; a strict zero-reward rule for invalid renders anchors renderability and prevents reward hacking. To prevent behavioral collapse, we introduce Forced Optimization, a strict acceptance rule that admits only improving revisions, yielding monotonically better trajectories. At inference, we decouple the critic and run a lightweight, critic-free self-edit cycle, keeping latency comparable to base decoding while retaining most of the gains. Across three widely used benchmarks, ReLook consistently outperforms strong baselines in vision-grounded front-end code generation, highlighting the benefits of agentic perception, visual rewards, and training-inference decoupling.
☆ How Reinforcement Learning After Next-Token Prediction Facilitates Learning
Recent advances in reasoning domains with neural networks have primarily been enabled by a training recipe that optimizes Large Language Models, previously trained to predict the next-token in a sequence, with reinforcement learning algorithms. We introduce a framework to study the success of this paradigm, and we theoretically expose the optimization mechanisms by which reinforcement learning improves over next-token prediction in this setting. We study learning from mixture distributions of short and long ``chain-of-thought'' sequences encoding a single task. In particular, when the task consists of predicting the parity of $d$ bits and long sequences are rare, we show how reinforcement learning after next-token prediction enables autoregressive transformers to generalize, whereas mere next-token prediction requires extreme statistical or computational resources to do so. We further explain how reinforcement learning leverages increased test-time computation, manifested in longer responses, to facilitate this learning process. In a simplified setting, we theoretically prove that autoregressive linear models following this training recipe can efficiently learn to predict the parity of $d$ bits as long as the proportion of long demonstrations in the data mix is not exponentially small in the input dimension $d$. Finally, we demonstrate these same phenomena in other settings, including the post-training of Llama-series models on mixture variations of common mathematical reasoning benchmarks.
☆ Constraint-Aware Reinforcement Learning via Adaptive Action Scaling
Safe reinforcement learning (RL) seeks to mitigate unsafe behaviors that arise from exploration during training by reducing constraint violations while maintaining task performance. Existing approaches typically rely on a single policy to jointly optimize reward and safety, which can cause instability due to conflicting objectives, or they use external safety filters that override actions and require prior system knowledge. In this paper, we propose a modular cost-aware regulator that scales the agent's actions based on predicted constraint violations, preserving exploration through smooth action modulation rather than overriding the policy. The regulator is trained to minimize constraint violations while avoiding degenerate suppression of actions. Our approach integrates seamlessly with off-policy RL methods such as SAC and TD3, and achieves state-of-the-art return-to-cost ratios on Safety Gym locomotion tasks with sparse costs, reducing constraint violations by up to 126 times while increasing returns by over an order of magnitude compared to prior methods.
☆ Rescaling-Aware Training for Efficient Deployment of Deep Learning Models on Full-Integer Hardware
Integer AI inference significantly reduces computational complexity in embedded systems. Quantization-aware training (QAT) helps mitigate accuracy degradation associated with post-training quantization but still overlooks the impact of integer rescaling during inference, which is a hardware costly operation in integer-only AI inference. This work shows that rescaling cost can be dramatically reduced post-training, by applying a stronger quantization to the rescale multiplicands at no model-quality loss. Furthermore, we introduce Rescale-Aware Training, a fine tuning method for ultra-low bit-width rescaling multiplicands. Experiments show that even with 8x reduced rescaler widths, the full accuracy is preserved through minimal incremental retraining. This enables more energy-efficient and cost-efficient AI inference for resource-constrained embedded systems.
comment: Submitted to IEEE Embedded Systems Letters
☆ Coordinated Strategies in Realistic Air Combat by Hierarchical Multi-Agent Reinforcement Learning
Achieving mission objectives in a realistic simulation of aerial combat is highly challenging due to imperfect situational awareness and nonlinear flight dynamics. In this work, we introduce a novel 3D multi-agent air combat environment and a Hierarchical Multi-Agent Reinforcement Learning framework to tackle these challenges. Our approach combines heterogeneous agent dynamics, curriculum learning, league-play, and a newly adapted training algorithm. To this end, the decision-making process is organized into two abstraction levels: low-level policies learn precise control maneuvers, while high-level policies issue tactical commands based on mission objectives. Empirical results show that our hierarchical approach improves both learning efficiency and combat performance in complex dogfight scenarios.
comment: 2025 IEEE International Conference on Agentic AI (ICA)
☆ Differentiable Fast Top-K Selection for Large-Scale Recommendation
Cascade ranking is a widely adopted paradigm in large-scale information retrieval systems for Top-K item selection. However, the Top-K operator is non-differentiable, hindering end-to-end training. Existing methods include Learning-to-Rank approaches (e.g., LambdaLoss), which optimize ranking metrics like NDCG and suffer from objective misalignment, and differentiable sorting-based methods (e.g., ARF, LCRON), which relax permutation matrices for direct Top-K optimization but introduce gradient conflicts through matrix aggregation. A promising alternative is to directly construct a differentiable approximation of the Top-K selection operator, bypassing the use of soft permutation matrices. However, even state-of-the-art differentiable Top-K operator (e.g., LapSum) require $O(n \log n)$ complexity due to their dependence on sorting for solving the threshold. Thus, we propose DFTopK, a novel differentiable Top-K operator achieving optimal $O(n)$ time complexity. By relaxing normalization constraints, DFTopK admits a closed-form solution and avoids sorting. DFTopK also avoids the gradient conflicts inherent in differentiable sorting-based methods. We evaluate DFTopK on both the public benchmark RecFLow and an industrial system. Experimental results show that DFTopK significantly improves training efficiency while achieving superior performance, which enables us to scale up training samples more efficiently. In the online A/B test, DFTopK yielded a +1.77\% revenue lift with the same computational budget compared to the baseline. To the best of our knowledge, this work is the first to introduce differentiable Top-K operators into recommendation systems and the first to achieve theoretically optimal linear-time complexity for Top-K selection. We have open-sourced our implementation to facilitate future research in both academia and industry.
comment: 12 pages, 5 figures
☆ Iterative Amortized Inference: Unifying In-Context Learning and Learned Optimizers
Modern learning systems increasingly rely on amortized learning - the idea of reusing computation or inductive biases shared across tasks to enable rapid generalization to novel problems. This principle spans a range of approaches, including meta-learning, in-context learning, prompt tuning, learned optimizers and more. While motivated by similar goals, these approaches differ in how they encode and leverage task-specific information, often provided as in-context examples. In this work, we propose a unified framework which describes how such methods differ primarily in the aspects of learning they amortize - such as initializations, learned updates, or predictive mappings - and how they incorporate task data at inference. We introduce a taxonomy that categorizes amortized models into parametric, implicit, and explicit regimes, based on whether task adaptation is externalized, internalized, or jointly modeled. Building on this view, we identify a key limitation in current approaches: most methods struggle to scale to large datasets because their capacity to process task data at inference (e.g., context length) is often limited. To address this, we propose iterative amortized inference, a class of models that refine solutions step-by-step over mini-batches, drawing inspiration from stochastic optimization. Our formulation bridges optimization-based meta-learning with forward-pass amortization in models like LLMs, offering a scalable and extensible foundation for general-purpose task adaptation.
☆ Reconstructing 12-Lead ECG from 3-Lead ECG using Variational Autoencoder to Improve Cardiac Disease Detection of Wearable ECG Devices
Twelve-lead electrocardiograms (ECGs) are the clinical gold standard for cardiac diagnosis, providing comprehensive spatial coverage of the heart necessary to detect conditions such as myocardial infarction (MI). However, their lack of portability limits continuous and large-scale use. Three-lead ECG systems are widely used in wearable devices due to their simplicity and mobility, but they often fail to capture pathologies in unmeasured regions. To address this, we propose WearECG, a Variational Autoencoder (VAE) method that reconstructs twelve-lead ECGs from three leads: II, V1, and V5. Our model includes architectural improvements to better capture temporal and spatial dependencies in ECG signals. We evaluate generation quality using MSE, MAE, and Frechet Inception Distance (FID), and assess clinical validity via a Turing test with expert cardiologists. To further validate diagnostic utility, we fine-tune ECGFounder, a large-scale pretrained ECG model, on a multi-label classification task involving over 40 cardiac conditions, including six different myocardial infarction locations, using both real and generated signals. Experiments on the MIMIC dataset show that our method produces physiologically realistic and diagnostically informative signals, with robust performance in downstream tasks. This work demonstrates the potential of generative modeling for ECG reconstruction and its implications for scalable, low-cost cardiac screening.
comment: 24 pages, 5 figures, submitted to Nature Communications
☆ Forward-Forward Autoencoder Architectures for Energy-Efficient Wireless Communications
The application of deep learning to the area of communications systems has been a growing field of interest in recent years. Forward-forward (FF) learning is an efficient alternative to the backpropagation (BP) algorithm, which is the typically used training procedure for neural networks. Among its several advantages, FF learning does not require the communication channel to be differentiable and does not rely on the global availability of partial derivatives, allowing for an energy-efficient implementation. In this work, we design end-to-end learned autoencoders using the FF algorithm and numerically evaluate their performance for the additive white Gaussian noise and Rayleigh block fading channels. We demonstrate their competitiveness with BP-trained systems in the case of joint coding and modulation, and in a scenario where a fixed, non-differentiable modulation stage is applied. Moreover, we provide further insights into the design principles of the FF network, its training convergence behavior, and significant memory and processing time savings compared to BP-based approaches.
☆ Leveraging LLMs for Semi-Automatic Corpus Filtration in Systematic Literature Reviews
The creation of systematic literature reviews (SLR) is critical for analyzing the landscape of a research field and guiding future research directions. However, retrieving and filtering the literature corpus for an SLR is highly time-consuming and requires extensive manual effort, as keyword-based searches in digital libraries often return numerous irrelevant publications. In this work, we propose a pipeline leveraging multiple large language models (LLMs), classifying papers based on descriptive prompts and deciding jointly using a consensus scheme. The entire process is human-supervised and interactively controlled via our open-source visual analytics web interface, LLMSurver, which enables real-time inspection and modification of model outputs. We evaluate our approach using ground-truth data from a recent SLR comprising over 8,000 candidate papers, benchmarking both open and commercial state-of-the-art LLMs from mid-2024 and fall 2025. Results demonstrate that our pipeline significantly reduces manual effort while achieving lower error rates than single human annotators. Furthermore, modern open-source models prove sufficient for this task, making the method accessible and cost-effective. Overall, our work demonstrates how responsible human-AI collaboration can accelerate and enhance systematic literature reviews within academic workflows.
☆ FedHybrid: Breaking the Memory Wall of Federated Learning via Hybrid Tensor Management
Federated Learning (FL) emerges as a new learning paradigm that enables multiple devices to collaboratively train a shared model while preserving data privacy. However, one fundamental and prevailing challenge that hinders the deployment of FL on mobile devices is the memory limitation. This paper proposes \textit{FedHybrid}, a novel framework that effectively reduces the memory footprint during the training process while guaranteeing the model accuracy and the overall training progress. Specifically, \textit{FedHybrid} first selects the participating devices for each training round by jointly evaluating their memory budget, computing capability, and data diversity. After that, it judiciously analyzes the computational graph and generates an execution plan for each selected client in order to meet the corresponding memory budget while minimizing the training delay through employing a hybrid of recomputation and compression techniques according to the characteristic of each tensor. During the local training process, \textit{FedHybrid} carries out the execution plan with a well-designed activation compression technique to effectively achieve memory reduction with minimum accuracy loss. We conduct extensive experiments to evaluate \textit{FedHybrid} on both simulation and off-the-shelf mobile devices. The experiment results demonstrate that \textit{FedHybrid} achieves up to a 39.1\% increase in model accuracy and a 15.5$\times$ reduction in wall clock time under various memory budgets compared with the baselines.
comment: Sensys 2024
☆ Medical Interpretability and Knowledge Maps of Large Language Models
We present a systematic study of medical-domain interpretability in Large Language Models (LLMs). We study how the LLMs both represent and process medical knowledge through four different interpretability techniques: (1) UMAP projections of intermediate activations, (2) gradient-based saliency with respect to the model weights, (3) layer lesioning/removal and (4) activation patching. We present knowledge maps of five LLMs which show, at a coarse-resolution, where knowledge about patient's ages, medical symptoms, diseases and drugs is stored in the models. In particular for Llama3.3-70B, we find that most medical knowledge is processed in the first half of the model's layers. In addition, we find several interesting phenomena: (i) age is often encoded in a non-linear and sometimes discontinuous manner at intermediate layers in the models, (ii) the disease progression representation is non-monotonic and circular at certain layers of the model, (iii) in Llama3.3-70B, drugs cluster better by medical specialty rather than mechanism of action, especially for Llama3.3-70B and (iv) Gemma3-27B and MedGemma-27B have activations that collapse at intermediate layers but recover by the final layers. These results can guide future research on fine-tuning, un-learning or de-biasing LLMs for medical tasks by suggesting at which layers in the model these techniques should be applied.
comment: 29 pages, 34 figures, 5 tables
☆ Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers
Reinforcement learning (RL) has emerged as a crucial approach for enhancing the capabilities of large language models. However, in Mixture-of-Experts (MoE) models, the routing mechanism often introduces instability, even leading to catastrophic RL training collapse. We analyze the training-inference consistency of MoE models and identify a notable discrepancy in routing behaviors between the two phases. Moreover, even under identical conditions, the routing framework can yield divergent expert selections across repeated forward passes. To address this foundational inconsistency, we propose Rollout Routing Replay (R3), a method that records routing distributions from the inference engine and replays them during training. R3 significantly reduces training-inference policy KL divergence and mitigates extreme discrepancies without compromising training speed. Extensive experiments on various settings confirm that R3 succeeds in stabilizing RL training, preventing collapse and outperforming methods such as GSPO and TIS. We believe this work can offer a new solution for stabilizing RL in MoE models.
☆ Understanding the Generalization of Stochastic Gradient Adam in Learning Neural Networks NeurIPS 2025
Adam is a popular and widely used adaptive gradient method in deep learning, which has also received tremendous focus in theoretical research. However, most existing theoretical work primarily analyzes its full-batch version, which differs fundamentally from the stochastic variant used in practice. Unlike SGD, stochastic Adam does not converge to its full-batch counterpart even with infinitesimal learning rates. We present the first theoretical characterization of how batch size affects Adam's generalization, analyzing two-layer over-parameterized CNNs on image data. Our results reveal that while both Adam and AdamW with proper weight decay $\lambda$ converge to poor test error solutions, their mini-batch variants can achieve near-zero test error. We further prove Adam has a strictly smaller effective weight decay bound than AdamW, theoretically explaining why Adam requires more sensitive $\lambda$ tuning. Extensive experiments validate our findings, demonstrating the critical role of batch size and weight decay in Adam's generalization performance.
comment: 71 pages, 12 figures, NeurIPS 2025
☆ Multi-View Graph Feature Propagation for Privacy Preservation and Feature Sparsity
Graph Neural Networks (GNNs) have demonstrated remarkable success in node classification tasks over relational data, yet their effectiveness often depends on the availability of complete node features. In many real-world scenarios, however, feature matrices are highly sparse or contain sensitive information, leading to degraded performance and increased privacy risks. Furthermore, direct exposure of information can result in unintended data leakage, enabling adversaries to infer sensitive information. To address these challenges, we propose a novel Multi-view Feature Propagation (MFP) framework that enhances node classification under feature sparsity while promoting privacy preservation. MFP extends traditional Feature Propagation (FP) by dividing the available features into multiple Gaussian-noised views, each propagating information independently through the graph topology. The aggregated representations yield expressive and robust node embeddings. This framework is novel in two respects: it introduces a mechanism that improves robustness under extreme sparsity, and it provides a principled way to balance utility with privacy. Extensive experiments conducted on graph datasets demonstrate that MFP outperforms state-of-the-art baselines in node classification while substantially reducing privacy leakage. Moreover, our analysis demonstrates that propagated outputs serve as alternative imputations rather than reconstructions of the original features, preserving utility without compromising privacy. A comprehensive sensitivity analysis further confirms the stability and practical applicability of MFP across diverse scenarios. Overall, MFP provides an effective and privacy-aware framework for graph learning in domains characterized by missing or sensitive features.
☆ Part II: ROLL Flash -- Accelerating RLVR and Agentic Training with Asynchrony
Synchronous Reinforcement Learning (RL) post-training has emerged as a crucial step for enhancing Large Language Models (LLMs) with diverse capabilities. However, many systems designed to accelerate RL post-training still suffer from low resource utilization and limited scalability. We present ROLL Flash, a system that extends ROLL with native support for asynchronous RL post-training. ROLL Flash is built upon two core design principles: fine-grained parallelism and rollout-train decoupling. Guided by these principles, ROLL Flash provides flexible programming interfaces that enable a fully asynchronous training architecture and support efficient rollout mechanisms, including queue scheduling and environment-level asynchronous execution. Through comprehensive theoretical analysis and extensive experiments, we demonstrate that ROLL Flash significantly improves resource utilization and scalability over synchronous RL post-training. ROLL Flash achieves up to 2.24x speedup on RLVR tasks and 2.72x on agentic tasks, using the same GPU budget as synchronous baselines. Furthermore, we implement several popular off-policy algorithms and verify that asynchronous training can achieve performance on par with synchronous training.
☆ Event-Aware Prompt Learning for Dynamic Graphs
Real-world graph typically evolve via a series of events, modeling dynamic interactions between objects across various domains. For dynamic graph learning, dynamic graph neural networks (DGNNs) have emerged as popular solutions. Recently, prompt learning methods have been explored on dynamic graphs. However, existing methods generally focus on capturing the relationship between nodes and time, while overlooking the impact of historical events. In this paper, we propose EVP, an event-aware dynamic graph prompt learning framework that can serve as a plug-in to existing methods, enhancing their ability to leverage historical events knowledge. First, we extract a series of historical events for each node and introduce an event adaptation mechanism to align the fine-grained characteristics of these events with downstream tasks. Second, we propose an event aggregation mechanism to effectively integrate historical knowledge into node representations. Finally, we conduct extensive experiments on four public datasets to evaluate and analyze EVP.
comment: Under review
☆ DiffStyleTS: Diffusion Model for Style Transfer in Time Series
Style transfer combines the content of one signal with the style of another. It supports applications such as data augmentation and scenario simulation, helping machine learning models generalize in data-scarce domains. While well developed in vision and language, style transfer methods for time series data remain limited. We introduce DiffTSST, a diffusion-based framework that disentangles a time series into content and style representations via convolutional encoders and recombines them through a self-supervised attention-based diffusion process. At inference, encoders extract content and style from two distinct series, enabling conditional generation of novel samples to achieve style transfer. We demonstrate both qualitatively and quantitatively that DiffTSST achieves effective style transfer. We further validate its real-world utility by showing that data augmentation with DiffTSST improves anomaly detection in data-scarce regimes.
☆ Diffusion-Link: Diffusion Probabilistic Model for Bridging the Audio-Text Modality Gap ICASSP 2026
Contrastive audio-language pretraining yields powerful joint representations, yet a persistent audio-text modality gap limits the benefits of coupling multimodal encoders with large language models (LLMs). We present Diffusion-Link, a diffusion-based modality-bridging module that generatively maps audio embeddings into the text-embedding distribution. The module is trained at the output embedding from the frozen multimodal encoder and implemented as a lightweight network with three residual MLP blocks. To assess the effect of Diffusion-Link on multimodal encoder-LLM coupling, we evaluate on Automatic Audio Captioning (AAC); to our knowledge, this is the first application of diffusion-based modality bridging to AAC. We report two results. (1) Modality-gap analysis: on similarity and geometric criteria, Diffusion-Link reduces the modality gap the most among prior diffusion-based methods and shows a collective migration of audio embeddings toward the text distribution. (2) Downstream AAC: attaching Diffusion-Link to the same multimodal LLM baseline achieves state-of-the-art on AudioCaps in both zero-shot and fully supervised captioning without external knowledge, with relative gains up to 52.5% and 7.5%, respectively. These findings show that closing the modality gap is pivotal for effective coupling between multimodal encoders and LLMs, and diffusion-based modality bridging offers a promising direction beyond knowledge-retrieval-centric designs. Code will be released upon acceptance https://github.com/DevKiHyun/Diffusion-Link
comment: 5 pages. Submitted to IEEE ICASSP 2026
☆ When Does Supervised Training Pay Off? The Hidden Economics of Object Detection in the Era of Vision-Language Models
Object detection systems have traditionally relied on supervised learning with manually annotated bounding boxes, achieving high accuracy at the cost of substantial annotation investment. The emergence of Vision-Language Models (VLMs) offers an alternative paradigm enabling zero-shot detection through natural language queries, eliminating annotation requirements but operating with reduced accuracy. This paper presents the first comprehensive cost-effectiveness analysis comparing supervised detection (YOLO) with zero-shot VLM inference (Gemini Flash 2.5). Through systematic evaluation on 1,000 stratified COCO images and 200 diverse product images spanning consumer electronics and rare categories, combined with detailed Total Cost of Ownership modeling, we establish quantitative break-even thresholds governing architecture selection. Our findings reveal that supervised YOLO achieves 91.2% accuracy versus 68.5% for zero-shot Gemini on standard categories, representing a 22.7 percentage point advantage that costs $10,800 in annotation for 100-category systems. However, this advantage justifies investment only beyond 55 million inferences, equivalent to 151,000 images daily for one year. Zero-shot Gemini demonstrates 52.3% accuracy on diverse product categories (ranging from highly web-prevalent consumer electronics at 75-85% to rare specialized equipment at 25-40%) where supervised YOLO achieves 0% due to architectural constraints preventing detection of untrained classes. Cost per Correct Detection analysis reveals substantially lower per-detection costs for Gemini ($0.00050 vs $0.143) at 100,000 inferences despite accuracy deficits. We develop decision frameworks demonstrating that optimal architecture selection depends critically on deployment volume, category stability, budget constraints, and accuracy requirements rather than purely technical performance metrics.
comment: 23 pages, 4 figures, 4 tables
☆ $Δ\mathrm{Energy}$: Optimizing Energy Change During Vision-Language Alignment Improves both OOD Detection and OOD Generalization
Recent approaches for vision-language models (VLMs) have shown remarkable success in achieving fast downstream adaptation. When applied to real-world downstream tasks, VLMs inevitably encounter both the in-distribution (ID) data and out-of-distribution (OOD) data. The OOD datasets often include both covariate shifts (e.g., known classes with changes in image styles) and semantic shifts (e.g., test-time unseen classes). This highlights the importance of improving VLMs' generalization ability to covariate-shifted OOD data, while effectively detecting open-set semantic-shifted OOD classes. In this paper, inspired by the substantial energy change observed in closed-set data when re-aligning vision-language modalities (specifically by directly reducing the maximum cosine similarity to a low value), we introduce a novel OOD score, named {\Delta}Energy. {\Delta}Energy significantly outperforms the vanilla energy-based OOD score and provides a more reliable approach for OOD detection. Furthermore, {\Delta}Energy can simultaneously improve OOD generalization under covariate shifts, which is achieved by lower-bound maximization for {\Delta}Energy (termed EBM). EBM is theoretically proven to not only enhance OOD detection but also yields a domain-consistent Hessian, which serves as a strong indicator for OOD generalization. Based on this finding, we developed a unified fine-tuning framework that allows for improving VLMs' robustness in both OOD generalization and OOD detection. Extensive experiments on challenging OOD detection and generalization benchmarks demonstrate the superiority of our method, outperforming recent approaches by 10% to 25% in AUROC.
comment: Accepted by NeruIPS2025
☆ LouisKV: Efficient KV Cache Retrieval for Long Input-Output Sequences
While Key-Value (KV) cache succeeds in reducing redundant computations in auto-regressive models, it introduces significant memory overhead, limiting its practical deployment in long-sequence scenarios. Existing KV retrieval methods mitigate this by dynamically retaining only a subset of KV entries on the GPU. However, they still suffer from notable efficiency and accuracy bottlenecks due to per-token retrieval and coarse-grained page-level KV management, especially in long-output reasoning scenarios. With the emergence of large reasoning models, efficiently handling such scenarios has become increasingly important. To address this issue, we present two key observations: (1) critical KVs exhibit strong temporal locality during decoding, and (2) these KVs exhibit distinct distribution patterns across the input prompt and generated output. Building on these observations, we propose LouisKV, an efficient KV cache retrieval framework designed for various long-sequence scenarios. Specifically, LouisKV introduces a semantic-aware retrieval strategy leveraging temporal locality to trigger retrieval only at semantic boundaries, drastically reducing computation and data transfer overhead. LouisKV also designs a decoupled, fine-grained management scheme that tailors differentiated strategies for input and output sequences to create retrieval units that better match the model's attention patterns, enabling precise identification of critical KVs. Furthermore, to boost efficiency, LouisKV incorporates several kernel-level optimizations, including custom Triton and CUDA kernels to accelerate the KV clustering and retrieval. Evaluations show that LouisKV achieves up to 4.7$\times$ speedup over state-of-the-art KV retrieval methods while maintaining near-lossless accuracy across diverse long-sequence tasks, including long-input short-output, short-input long-output, and long-input long-output scenarios.
☆ Network-Optimised Spiking Neural Network (NOS) Scheduling for 6G O-RAN: Spectral Margin and Delay-Tail Control
This work presents a Network-Optimised Spiking (NOS) delay-aware scheduler for 6G radio access. The scheme couples a bounded two-state kernel to a clique-feasible proportional-fair (PF) grant head: the excitability state acts as a finite-buffer proxy, the recovery state suppresses repeated grants, and neighbour pressure is injected along the interference graph via delayed spikes. A small-signal analysis yields a delay-dependent threshold $k_\star(\Delta)$ and a spectral margin $\delta = k_\star(\Delta) - gH\rho(W)$ that compress topology, controller gain, and delay into a single design parameter. Under light assumptions on arrivals, we prove geometric ergodicity for $\delta>0$ and derive sub-Gaussian backlog and delay tail bounds with exponents proportional to $\delta$. A numerical study, aligned with the analysis and a DU compute budget, compares NOS with PF and delayed backpressure (BP) across interference topologies over a $5$--$20$\,ms delay sweep. With a single gain fixed at the worst spectral radius, NOS sustains higher utilisation and a smaller 99.9th-percentile delay while remaining clique-feasible on integer PRBs.
comment: 6 pages, 5 figures, 1 table
☆ Gym-TORAX: Open-source software for integrating RL with plasma control simulators
This paper presents Gym-TORAX, a Python package enabling the implementation of Reinforcement Learning (RL) environments for simulating plasma dynamics and control in tokamaks. Users define succinctly a set of control actions and observations, and a control objective from which Gym-TORAX creates a Gymnasium environment that wraps TORAX for simulating the plasma dynamics. The objective is formulated through rewards depending on the simulated state of the plasma and control action to optimize specific characteristics of the plasma, such as performance and stability. The resulting environment instance is then compatible with a wide range of RL algorithms and libraries and will facilitate RL research in plasma control. In its current version, one environment is readily available, based on a ramp-up scenario of the International Thermonuclear Experimental Reactor (ITER).
☆ Vision-LLMs for Spatiotemporal Traffic Forecasting
Accurate spatiotemporal traffic forecasting is a critical prerequisite for proactive resource management in dense urban mobile networks. While Large Language Models (LLMs) have shown promise in time series analysis, they inherently struggle to model the complex spatial dependencies of grid-based traffic data. Effectively extending LLMs to this domain is challenging, as representing the vast amount of information from dense geographical grids can be inefficient and overwhelm the model's context. To address these challenges, we propose ST-Vision-LLM, a novel framework that reframes spatiotemporal forecasting as a vision-language fusion problem. Our approach leverages a Vision-LLM visual encoder to process historical global traffic matrices as image sequences, providing the model with a comprehensive global view to inform cell-level predictions. To overcome the inefficiency of LLMs in handling numerical data, we introduce an efficient encoding scheme that represents floating-point values as single tokens via a specialized vocabulary, coupled with a two-stage numerical alignment fine-tuning process. The model is first trained with Supervised Fine-Tuning (SFT) and then further optimized for predictive accuracy using Group Relative Policy Optimization (GRPO), a memory-efficient reinforcement learning method. Evaluations on real-world mobile traffic datasets demonstrate that ST-Vision-LLM outperforms existing methods by 15.6% in long-term prediction accuracy and exceeds the second-best baseline by over 30.04% in cross-domain few-shot scenarios. Our extensive experiments validate the model's strong generalization capabilities across various data-scarce environments.
☆ ENIGMA: The Geometry of Reasoning and Alignment in Large-Language Models
We present Entropic Mutual-Information Geometry Large-Language Model Alignment (ENIGMA), a novel approach to Large-Language Model (LLM) training that jointly improves reasoning, alignment and robustness by treating an organisation's policies/principles as directions to move on a model's information manifold. Our single-loop trainer combines Group-Relative Policy Optimisation (GRPO), an on-policy, critic-free RL method with Chain-of-Thought (CoT)-format only rewards; a Self-Supervised Alignment with Mutual Information (SAMI)-style symmetric InfoNCE auxiliary; and an entropic Sinkhorn optimal-transport regulariser on hidden-state distributions to bound geometry drift. We also introduce infoNCE metrics that specialise to a standard MI lower bound under matched negatives to measure how strongly a model's CoT encodes these policies. These metrics include a Sufficiency Index (SI) that enables the selection and creation of principles that maximise downstream performance prior to training. In our experiments using small (1B) LLMs, high-SI principles predict steadier training dynamics and improved benchmark performance over GRPO ablations. Our information-geometry analysis of trained models validates desirable structural change in the manifold. These results support our hypothesis that reasoning, alignment, and robustness are projections of a single informationgeometric objective, and that models trained using ENIGMA demonstrate principled reasoning without the use of a reward model, offering a path to trusted capability
comment: 52 pages, 10 figures
☆ SeFEF: A Seizure Forecasting Evaluation Framework
The lack of standardization in seizure forecasting slows progress in the field and limits the clinical translation of forecasting models. In this work, we introduce a Python-based framework aimed at streamlining the development, assessment, and documentation of individualized seizure forecasting algorithms. The framework automates data labeling, cross-validation splitting, forecast post-processing, performance evaluation, and reporting. It supports various forecasting horizons and includes a model card that documents implementation details, training and evaluation settings, and performance metrics. Three different models were implemented as a proof-of-concept. The models leveraged features extracted from time series data and seizure periodicity. Model performance was assessed using time series cross-validation and key deterministic and probabilistic metrics. Implementation of the three models was successful, demonstrating the flexibility of the framework. The results also emphasize the importance of careful model interpretation due to variations in probability scaling, calibration, and subject-specific differences. Although formal usability metrics were not recorded, empirical observations suggest reduced development time and methodological consistency, minimizing unintentional variations that could affect the comparability of different approaches. As a proof-of-concept, this validation is inherently limited, relying on a single-user experiment without statistical analyses or replication across independent datasets. At this stage, our objective is to make the framework publicly available to foster community engagement, facilitate experimentation, and gather feedback. In the long term, we aim to contribute to the establishment of a consensus on a standardized methodology for the development and validation of seizure forecasting algorithms in people with epilepsy.
comment: main document: 14 pages, 9 figures, 2 tables; appendix: 7 pages, 2 figures, 3 tables, 2 algorithms
☆ FedLoRA-Optimizer: Federated LoRA Fine-Tuning with Global and Local Optimization in Heterogeneous Data Scenarios
Federated efficient fine-tuning has emerged as an approach that leverages distributed data and computational resources across nodes to address the challenges of large-scale fine-tuning and privacy preservation. The Low-Rank Adaptation (LoRA) enables efficient fine-tuning of large-scale pre-trained models by introducing trainable low-rank matrices into weight updates.However, in heterogeneous data scenarios, client drift weakens the generalization of the global model, and local models often fail to meet the personalized needs of individual clients.Moreover, existing federated LoRA efficient fine-tuning techniques overlook fine-grained analysis of the tuning matrices. To address this, we conducted preliminary experiments and found that different LoRA matrices exhibit different sensitivity to changes in the direction and magnitude of their vectors.We thus propose a fine-grained federated LoRA tuning method. By fine-tuning the more sensitive directional vectors in the A matrix, which encode shared knowledge, our method learns shared features more effectively across clients and enhances global generalization. Simultaneously, by fine-tuning the more sensitive magnitude vectors in the B matrix, which encode personalized knowledge, our method better captures personalized knowledge, enabling detailed adaptation to local data. The method uses a pipeline combining global and local optimizers. Global optimization further improves local models, achieving collaborative optimization between global and local levels. This improves both the generalization ability of the global model and the personalized adaptation of local models under heterogeneous data scenarios. Experiments on Databricks-Dolly-15k and Natural Instructions with LLaMA2-7B and Deepseek-7B confirm that our method improves global performance by 0.39% and local performance by 0.59%.
☆ DemoHLM: From One Demonstration to Generalizable Humanoid Loco-Manipulation
Loco-manipulation is a fundamental challenge for humanoid robots to achieve versatile interactions in human environments. Although recent studies have made significant progress in humanoid whole-body control, loco-manipulation remains underexplored and often relies on hard-coded task definitions or costly real-world data collection, which limits autonomy and generalization. We present DemoHLM, a framework for humanoid loco-manipulation that enables generalizable loco-manipulation on a real humanoid robot from a single demonstration in simulation. DemoHLM adopts a hierarchy that integrates a low-level universal whole-body controller with high-level manipulation policies for multiple tasks. The whole-body controller maps whole-body motion commands to joint torques and provides omnidirectional mobility for the humanoid robot. The manipulation policies, learned in simulation via our data generation and imitation learning pipeline, command the whole-body controller with closed-loop visual feedback to execute challenging loco-manipulation tasks. Experiments show a positive correlation between the amount of synthetic data and policy performance, underscoring the effectiveness of our data generation pipeline and the data efficiency of our approach. Real-world experiments on a Unitree G1 robot equipped with an RGB-D camera validate the sim-to-real transferability of DemoHLM, demonstrating robust performance under spatial variations across ten loco-manipulation tasks.
☆ MIEO: encoding clinical data to enhance cardiovascular event prediction
As clinical data are becoming increasingly available, machine learning methods have been employed to extract knowledge from them and predict clinical events. While promising, approaches suffer from at least two main issues: low availability of labelled data and data heterogeneity leading to missing values. This work proposes the use of self-supervised auto-encoders to efficiently address these challenges. We apply our methodology to a clinical dataset from patients with ischaemic heart disease. Patient data is embedded in a latent space, built using unlabelled data, which is then used to train a neural network classifier to predict cardiovascular death. Results show improved balanced accuracy compared to applying the classifier directly to the raw data, demonstrating that this solution is promising, especially in conditions where availability of unlabelled data could increase.
comment: Presented in the Poster Session of Computational Intelligence methods for Bioinformatics and Biostatistics (CIBB) 2025
☆ Large Language Models Are Effective Code Watermarkers
The widespread use of large language models (LLMs) and open-source code has raised ethical and security concerns regarding the distribution and attribution of source code, including unauthorized redistribution, license violations, and misuse of code for malicious purposes. Watermarking has emerged as a promising solution for source attribution, but existing techniques rely heavily on hand-crafted transformation rules, abstract syntax tree (AST) manipulation, or task-specific training, limiting their scalability and generality across languages. Moreover, their robustness against attacks remains limited. To address these limitations, we propose CodeMark-LLM, an LLM-driven watermarking framework that embeds watermark into source code without compromising its semantics or readability. CodeMark-LLM consists of two core components: (i) Semantically Consistent Embedding module that applies functionality-preserving transformations to encode watermark bits, and (ii) Differential Comparison Extraction module that identifies the applied transformations by comparing the original and watermarked code. Leveraging the cross-lingual generalization ability of LLM, CodeMark-LLM avoids language-specific engineering and training pipelines. Extensive experiments across diverse programming languages and attack scenarios demonstrate its robustness, effectiveness, and scalability.
☆ FUSE: Fast Semi-Supervised Node Embedding Learning via Structural and Label-Aware Optimization
Graph-based learning is a cornerstone for analyzing structured data, with node classification as a central task. However, in many real-world graphs, nodes lack informative feature vectors, leaving only neighborhood connectivity and class labels as available signals. In such cases, effective classification hinges on learning node embeddings that capture structural roles and topological context. We introduce a fast semi-supervised embedding framework that jointly optimizes three complementary objectives: (i) unsupervised structure preservation via scalable modularity approximation, (ii) supervised regularization to minimize intra-class variance among labeled nodes, and (iii) semi-supervised propagation that refines unlabeled nodes through random-walk-based label spreading with attention-weighted similarity. These components are unified into a single iterative optimization scheme, yielding high-quality node embeddings. On standard benchmarks, our method consistently achieves classification accuracy at par with or superior to state-of-the-art approaches, while requiring significantly less computational cost.
☆ Learning the Structure of Connection Graphs
Connection graphs (CGs) extend traditional graph models by coupling network topology with orthogonal transformations, enabling the representation of global geometric consistency. They play a key role in applications such as synchronization, Riemannian signal processing, and neural sheaf diffusion. In this work, we address the inverse problem of learning CGs directly from observed signals. We propose a principled framework based on maximum pseudo-likelihood under a consistency assumption, which enforces spectral properties linking the connection Laplacian to the underlying combinatorial Laplacian. Based on this formulation, we introduce the Structured Connection Graph Learning (SCGL) algorithm, a block-optimization procedure over Riemannian manifolds that jointly infers network topology, edge weights, and geometric structure. Our experiments show that SCGL consistently outperforms existing baselines in both topological recovery and geometric fidelity, while remaining computationally efficient.
☆ Analyzing Data Quality and Decay in Mega-Constellations: A Physics-Informed Machine Learning Approach
In the era of mega-constellations, the need for accurate and publicly available information has become fundamental for satellite operators to guarantee the safety of spacecrafts and the Low Earth Orbit (LEO) space environment. This study critically evaluates the accuracy and reliability of publicly available ephemeris data for a LEO mega-constellation - Starlink. The goal of this work is twofold: (i) compare and analyze the quality of the data against high-precision numerical propagation. (ii) Leverage Physics-Informed Machine Learning to extract relevant satellite quantities, such as non-conservative forces, during the decay process. By analyzing two months of real orbital data for approximately 1500 Starlink satellites, we identify discrepancies between high precision numerical algorithms and the published ephemerides, recognizing the use of simplified dynamics at fixed thresholds, planned maneuvers, and limitations in uncertainty propagations. Furthermore, we compare data obtained from multiple sources to track and analyze deorbiting satellites over the same period. Empirically, we extract the acceleration profile of satellites during deorbiting and provide insights relating to the effects of non-conservative forces during reentry. For non-deorbiting satellites, the position Root Mean Square Error (RMSE) was approximately 300 m, while for deorbiting satellites it increased to about 600 m. Through this in-depth analysis, we highlight potential limitations in publicly available data for accurate and robust Space Situational Awareness (SSA), and importantly, we propose a data-driven model of satellite decay in mega-constellations.
comment: 76th International Astronautical Congress
☆ Neural Weight Compression for Language Models
The efficient storage and transmission of language model weights is becoming increasingly important, as their scale and adoption continue to grow. However, as our understanding of this new data modality is limited, designing a good compression algorithm for language model weights heavily relies on manual, trial-and-error approaches. In this paper, we propose a learned compression framework that trains neural codecs directly from pretrained language model weights. Unlike conventional data (e.g., images), language model weights pose unique challenges: the sizes and shapes of weight tensors vary significantly, and the reconstruction quality must be judged by downstream model predictions rather than na\"ive MSE loss. To address this, we introduce Neural Weight Compression (NWC), a novel autoencoder-based neural codec tailored to model weight compression. The proposed method inherits the advantages of autoencoder-based codecs while incorporating three technical components: (1) column-wise tensor chunking and normalization; (2) an importance-aware training loss; (3) an inference-time error compensation mechanism guided by model outputs. Experiments on open-weight language models show that NWC achieves competitive or state-of-the-art accuracy-compression tradeoffs, with particularly strong results at 4-6 bit precisions where accuracy remains nearly on par with FP16 models.
☆ LightPneumoNet: Lightweight Pneumonia Classifier
Effective pneumonia diagnosis is often challenged by the difficulty of deploying large, computationally expensive deep learning models in resource-limited settings. This study introduces LightPneumoNet, an efficient, lightweight convolutional neural network (CNN) built from scratch to provide an accessible and accurate diagnostic solution for pneumonia detection from chest X-rays. Our model was trained on a public dataset of 5,856 chest X-ray images. Preprocessing included image resizing to 224x224, grayscale conversion, and pixel normalization, with data augmentation (rotation, zoom, shear) to prevent overfitting. The custom architecture features four blocks of stacked convolutional layers and contains only 388,082 trainable parameters, resulting in a minimal 1.48 MB memory footprint. On the independent test set, our model delivered exceptional performance, achieving an overall accuracy of 0.942, precision of 0.92, and an F1-Score of 0.96. Critically, it obtained a sensitivity (recall) of 0.99, demonstrating a near-perfect ability to identify true pneumonia cases and minimize clinically significant false negatives. Notably, LightPneumoNet achieves this high recall on the same dataset where existing approaches typically require significantly heavier architectures or fail to reach comparable sensitivity levels. The model's efficiency enables deployment on low-cost hardware, making advanced computer-aided diagnosis accessible in underserved clinics and serving as a reliable second-opinion tool to improve patient outcomes.
comment: 13 pages (including references), 5 figures
☆ Enforcing convex constraints in Graph Neural Networks
Many machine learning applications require outputs that satisfy complex, dynamic constraints. This task is particularly challenging in Graph Neural Network models due to the variable output sizes of graph-structured data. In this paper, we introduce ProjNet, a Graph Neural Network framework which satisfies input-dependant constraints. ProjNet combines a sparse vector clipping method with the Component-Averaged Dykstra (CAD) algorithm, an iterative scheme for solving the best-approximation problem. We establish a convergence result for CAD and develop a GPU-accelerated implementation capable of handling large-scale inputs efficiently. To enable end-to-end training, we introduce a surrogate gradient for CAD that is both computationally efficient and better suited for optimization than the exact gradient. We validate ProjNet on four classes of constrained optimisation problems: linear programming, two classes of non-convex quadratic programs, and radio transmit power optimization, demonstrating its effectiveness across diverse problem settings.
☆ Discursive Circuits: How Do Language Models Understand Discourse Relations? EMNLP 2025
Which components in transformer language models are responsible for discourse understanding? We hypothesize that sparse computational graphs, termed as discursive circuits, control how models process discourse relations. Unlike simpler tasks, discourse relations involve longer spans and complex reasoning. To make circuit discovery feasible, we introduce a task called Completion under Discourse Relation (CuDR), where a model completes a discourse given a specified relation. To support this task, we construct a corpus of minimal contrastive pairs tailored for activation patching in circuit discovery. Experiments show that sparse circuits ($\approx 0.2\%$ of a full GPT-2 model) recover discourse understanding in the English PDTB-based CuDR task. These circuits generalize well to unseen discourse frameworks such as RST and SDRT. Further analysis shows lower layers capture linguistic features such as lexical semantics and coreference, while upper layers encode discourse-level abstractions. Feature utility is consistent across frameworks (e.g., coreference supports Expansion-like relations).
comment: Accepted to EMNLP 2025 (Main Conference); 9 pages, 8 figures, 5 tables (20 pages, 12 figures, 14 tables including references and appendices)
☆ Cross-Scale Reservoir Computing for large spatio-temporal forecasting and modeling
We propose a new reservoir computing method for forecasting high-resolution spatiotemporal datasets. By combining multi-resolution inputs from coarser to finer layers, our architecture better captures both local and global dynamics. Applied to Sea Surface Temperature data, it outperforms standard parallel reservoir models in long-term forecasting, demonstrating the effectiveness of cross-layers coupling in improving predictive accuracy. Finally, we show that the optimal network dynamics in each layer become increasingly linear, revealing the slow modes propagated to subsequent layers.
☆ Evaluating Line-level Localization Ability of Learning-based Code Vulnerability Detection Models
To address the extremely concerning problem of software vulnerability, system security is often entrusted to Machine Learning (ML) algorithms. Despite their now established detection capabilities, such models are limited by design to flagging the entire input source code function as vulnerable, rather than precisely localizing the concerned code lines. However, the detection granularity is crucial to support human operators during software development, ensuring that such predictions reflect the true code semantics to help debug, evaluate, and fix the detected vulnerabilities. To address this issue, recent work made progress toward improving the detector's localization ability, thus narrowing down the vulnerability detection "window" and providing more fine-grained predictions. Such approaches, however, implicitly disregard the presence of spurious correlations and biases in the data, which often predominantly influence the performance of ML algorithms. In this work, we investigate how detectors comply with this requirement by proposing an explainability-based evaluation procedure. Our approach, defined as Detection Alignment (DA), quantifies the agreement between the input source code lines that most influence the prediction and the actual localization of the vulnerability as per the ground truth. Through DA, which is model-agnostic and adaptable to different detection tasks, not limited to our use case, we analyze multiple learning-based vulnerability detectors and datasets. As a result, we show how the predictions of such models are consistently biased by non-vulnerable lines, ultimately highlighting the high impact of biases and spurious correlations. The code is available at https://github.com/pralab/vuln-localization-eval.
comment: Preprint
☆ Efficient In-Memory Acceleration of Sparse Block Diagonal LLMs
Structured sparsity enables deploying large language models (LLMs) on resource-constrained systems. Approaches like dense-to-sparse fine-tuning are particularly compelling, achieving remarkable structured sparsity by reducing the model size by over 6.7x, while still maintaining acceptable accuracy. Despite this reduction, LLM inference, especially the decode stage being inherently memory-bound, is extremely expensive on conventional Von-Neumann architectures. Compute-in-memory (CIM) architectures mitigate this by performing computations directly in memory, and when paired with sparse LLMs, enable storing and computing the entire model in memory, eliminating the data movement on the off-chip bus and improving efficiency. Nonetheless, naively mapping sparse matrices onto CIM arrays leads to poor array utilization and diminished computational efficiency. In this paper, we present an automated framework with novel mapping and scheduling strategies to accelerate sparse LLM inference on CIM accelerators. By exploiting block-diagonal sparsity, our approach improves CIM array utilization by over 50%, achieving more than 4x reduction in both memory footprint and the number of required floating-point operations.
comment: 8 pages, to appear in IEEE Cross-disciplinary Conference on Memory-Centric Computing (CCMCC)
☆ Protein as a Second Language for LLMs ICLR 2026
Deciphering the function of unseen protein sequences is a fundamental challenge with broad scientific impact, yet most existing methods depend on task-specific adapters or large-scale supervised fine-tuning. We introduce the "Protein-as-Second-Language" framework, which reformulates amino-acid sequences as sentences in a novel symbolic language that large language models can interpret through contextual exemplars. Our approach adaptively constructs sequence-question-answer triples that reveal functional cues in a zero-shot setting, without any further training. To support this process, we curate a bilingual corpus of 79,926 protein-QA instances spanning attribute prediction, descriptive understanding, and extended reasoning. Empirically, our method delivers consistent gains across diverse open-source LLMs and GPT-4, achieving up to 17.2% ROUGE-L improvement (average +7%) and even surpassing fine-tuned protein-specific language models. These results highlight that generic LLMs, when guided with protein-as-language cues, can outperform domain-specialized models, offering a scalable pathway for protein understanding in foundation models.
comment: Main paper: 9 pages, 6 figures. With references and appendix: 18 pages, 9 figures total. Submitted to ICLR 2026 (under review)
☆ Can Tool-Integrated Reinforcement Learning Generalize Across Diverse Domains?
Recent advances in large language models (LLMs) have demonstrated remarkable capabilities in reasoning and tool utilization. However, the generalization of tool-augmented reinforcement learning (RL) across diverse domains remains underexplored. In this work, we investigate the cross-domain generalization of an LLM agent equipped with a code interpreter tool, which is exclusively trained on mathematical problem-solving tasks. Despite the restricted training domain, we evaluate the agent's performance across several distinct reasoning domains. The results reveal that RL-based tool usage learned from mathematical tasks can be effectively transferred to complex tasks in other domains, enabling great task performance and high token efficiency. To facilitate this cross-domain transfer, we propose a Tool Generalization Reinforcement Learning (TGRL) framework designed to promote domain-agnostic learning and skill migration, encompassing: (i) a standardized tool interface that abstracts domain-specific nuances through consistent formatting and explicit termination, fostering transferable invocation patterns; (ii) a dual-component reward system that decomposes rewards to incentivize generalizable behaviors like tool efficiency and reasoning abstraction, ensuring alignment and robustness across domain shifts; and (iii) an XML-based prompt template that separates thinking, tool calls, and responses to encourage modular, domain-invariant planning and coherent multi-turn interactions. Extensive experiments across diverse benchmarks validate our approach, achieving state-of-the-art performance and highlighting the cross-domain potential of Tool RL for LLM reasoning.
☆ Machine Learning-Integrated Hybrid Fluid-Kinetic Framework for Quantum Electrodynamic Laser Plasma Simulations
High-intensity laser plasma interactions create complex computational problems because they involve both fluid and kinetic regimes, which need models that maintain physical precision while keeping computational speed. The research introduces a machine learning-based three-dimensional hybrid fluid-particle-in-cell (PIC) system, which links relativistic plasma behavior to automatic regime transitions. The technique employs fluid approximations for stable areas but activates the PIC solver when SwitchNet directs it to unstable sections through its training on physics-based synthetic data. The model uses a smooth transition between Ammosov-Delone-Krainov (ADK) tunneling and multiphoton ionization rates to simulate ionization, while Airy-function approximations simulate quantum electrodynamic (QED) effects for radiation reaction and pair production. The convolutional neural network applies energy conservation through physics-based loss functions, which operate on normalized fields per channel. Monte Carlo dropout provides uncertainty measurement. The hybrid model produces precise predictions with coefficient of determination (R^2) values above 0.95 and mean squared errors below 10^-4 for all field components. This adaptive approach enhances the accuracy and scalability of laser-plasma simulations, providing a unified predictive framework for high-energy-density and particle acceleration applications.
☆ EAGER: Entropy-Aware GEneRation for Adaptive Inference-Time Scaling
With the rise of reasoning language models and test-time scaling methods as a paradigm for improving model performance, substantial computation is often required to generate multiple candidate sequences from the same prompt. This enables exploration of different reasoning paths toward the correct solution, however, allocates the same compute budget for each prompt. Grounded on the assumption that different prompts carry different degrees of complexity, and thus different computation needs, we propose EAGer, a training-free generation method that leverages model uncertainty through token-wise entropy distribution to reduce redundant computation and concurrently improve overall performance. EAGer allows branching to multiple reasoning paths only in the presence of high-entropy tokens, and then reallocates the saved compute budget to the instances where exploration of alternative paths is most needed. We find that across multiple open-source models on complex reasoning benchmarks such as AIME 2025, EAGer can reallocate the budget without accessing target labels, achieving the best efficiency-performance trade-off in terms of reasoning length and Pass@k. When target labels are accessible, EAGer generates up to 65% fewer tokens (hence saving compute) and achieves up to 37% improvement in Pass@k compared to the Full Parallel Sampling.
☆ PAC-Bayesian Bounds on Constrained f-Entropic Risk Measures
PAC generalization bounds on the risk, when expressed in terms of the expected loss, are often insufficient to capture imbalances between subgroups in the data. To overcome this limitation, we introduce a new family of risk measures, called constrained f-entropic risk measures, which enable finer control over distributional shifts and subgroup imbalances via f-divergences, and include the Conditional Value at Risk (CVaR), a well-known risk measure. We derive both classical and disintegrated PAC-Bayesian generalization bounds for this family of risks, providing the first disintegratedPAC-Bayesian guarantees beyond standard risks. Building on this theory, we design a self-bounding algorithm that minimizes our bounds directly, yielding models with guarantees at the subgroup level. Finally, we empirically demonstrate the usefulness of our approach.
☆ ELMO: Efficiency via Low-precision and Peak Memory Optimization in Large Output Spaces ICML 2025
Large output spaces, also referred to as Extreme multilabel classification (XMC), is a setting that arises, e.g., in large-scale tagging and product-to-product recommendation, and is characterized by the number of labels ranging from hundreds of thousands to millions. This means that the linear classification head, usually only a tiny fraction of the overall model, turns into the main driver for compute and memory demand. Current state-of-the-art XMC methods predominantly rely on FP16-FP32 mixed-precision training, which we show can be unstable, and inefficient in terms of memory usage and computational overhead. Meanwhile, existing low-precision methods typically retain higher precision for the classification layer. In this work, we propose ELMO, a pure low-precision training framework for XMC models using BFloat16 and Float8 data types. By leveraging Kahan summation and stochastic rounding, we demonstrate that XMC models can be effectively trained entirely in Float8, without relying on single-precision master weights or tensor scaling. Low-precision training, combined with our proposed memory optimizations -- gradient fusion and chunking -- enables significant reductions in GPU memory usage. For example, we train a 3-million-label XMC model with only 6.6 GiB of GPU memory, compared to the 39.7 GiB required by the optimized SOTA method, Renee without compromising accuracy.
comment: Accepted to ICML 2025
☆ Beyond single-model XAI: aggregating multi-model explanations for enhanced trustworthiness ECAI 2025
The use of Artificial Intelligence (AI) models in real-world and high-risk applications has intensified the discussion about their trustworthiness and ethical usage, from both a technical and a legislative perspective. The field of eXplainable Artificial Intelligence (XAI) addresses this challenge by proposing explanations that bring to light the decision-making processes of complex black-box models. Despite being an essential property, the robustness of explanations is often an overlooked aspect during development: only robust explanation methods can increase the trust in the system as a whole. This paper investigates the role of robustness through the usage of a feature importance aggregation derived from multiple models ($k$-nearest neighbours, random forest and neural networks). Preliminary results showcase the potential in increasing the trustworthiness of the application, while leveraging multiple model's predictive power.
comment: Accepted at the European Workshop on Trustworthy Artificial Intelligence (TRUST-AI), co-located within ECAI 2025
☆ Emergence of hybrid computational dynamics through reinforcement learning
Understanding how learning algorithms shape the computational strategies that emerge in neural networks remains a fundamental challenge in machine intelligence. While network architectures receive extensive attention, the role of the learning paradigm itself in determining emergent dynamics remains largely unexplored. Here we demonstrate that reinforcement learning (RL) and supervised learning (SL) drive recurrent neural networks (RNNs) toward fundamentally different computational solutions when trained on identical decision-making tasks. Through systematic dynamical systems analysis, we reveal that RL spontaneously discovers hybrid attractor architectures, combining stable fixed-point attractors for decision maintenance with quasi-periodic attractors for flexible evidence integration. This contrasts sharply with SL, which converges almost exclusively to simpler fixed-point-only solutions. We further show that RL sculpts functionally balanced neural populations through a powerful form of implicit regularization -- a structural signature that enhances robustness and is conspicuously absent in the more heterogeneous solutions found by SL-trained networks. The prevalence of these complex dynamics in RL is controllably modulated by weight initialization and correlates strongly with performance gains, particularly as task complexity increases. Our results establish the learning algorithm as a primary determinant of emergent computation, revealing how reward-based optimization autonomously discovers sophisticated dynamical mechanisms that are less accessible to direct gradient-based optimization. These findings provide both mechanistic insights into neural computation and actionable principles for designing adaptive AI systems.
comment: 22 pages, 11 figures
☆ Enhanced Sampling for Efficient Learning of Coarse-Grained Machine Learning Potentials
Coarse-graining (CG) enables molecular dynamics (MD) simulations of larger systems and longer timescales that are otherwise infeasible with atomistic models. Machine learning potentials (MLPs), with their capacity to capture many-body interactions, can provide accurate approximations of the potential of mean force (PMF) in CG models. Current CG MLPs are typically trained in a bottom-up manner via force matching, which in practice relies on configurations sampled from the unbiased equilibrium Boltzmann distribution to ensure thermodynamic consistency. This convention poses two key limitations: first, sufficiently long atomistic trajectories are needed to reach convergence; and second, even once equilibrated, transition regions remain poorly sampled. To address these issues, we employ enhanced sampling to bias along CG degrees of freedom for data generation, and then recompute the forces with respect to the unbiased potential. This strategy simultaneously shortens the simulation time required to produce equilibrated data and enriches sampling in transition regions, while preserving the correct PMF. We demonstrate its effectiveness on the M\"uller-Brown potential and capped alanine, achieving notable improvements. Our findings support the use of enhanced sampling for force matching as a promising direction to improve the accuracy and reliability of CG MLPs.
☆ torchsom: The Reference PyTorch Library for Self-Organizing Maps
This paper introduces torchsom, an open-source Python library that provides a reference implementation of the Self-Organizing Map (SOM) in PyTorch. This package offers three main features: (i) dimensionality reduction, (ii) clustering, and (iii) friendly data visualization. It relies on a PyTorch backend, enabling (i) fast and efficient training of SOMs through GPU acceleration, and (ii) easy and scalable integrations with PyTorch ecosystem. Moreover, torchsom follows the scikit-learn API for ease of use and extensibility. The library is released under the Apache 2.0 license with 90% test coverage, and its source code and documentation are available at https://github.com/michelin/TorchSOM.
comment: 4 mains pages with 2 tables, 4 pages of references, 15 pages of appendices with 13 figures and 3 tables
☆ A Comprehensive Forecasting-Based Framework for Time Series Anomaly Detection: Benchmarking on the Numenta Anomaly Benchmark (NAB)
Time series anomaly detection is critical for modern digital infrastructures, yet existing methods lack systematic cross-domain evaluation. We present a comprehensive forecasting-based framework unifying classical methods (Holt-Winters, SARIMA) with deep learning architectures (LSTM, Informer) under a common residual-based detection interface. Our modular pipeline integrates preprocessing (normalization, STL decomposition), four forecasting models, four detection methods, and dual evaluation through forecasting metrics (MAE, RMSE, PCC) and detection metrics (Precision, Recall, F1, AUC). We conduct the first complete evaluation on the Numenta Anomaly Benchmark (58 datasets, 7 categories) with 232 model training runs and 464 detection evaluations achieving 100\% success rate. LSTM achieves best performance (F1: 0.688, ranking first or second on 81\% of datasets) with exceptional correlation on complex patterns (PCC: 0.999). Informer provides competitive accuracy (F1: 0.683) with 30\% faster training. Classical methods achieve perfect predictions on simple synthetic data with 60 lower cost but show 2-3 worse F1-scores on real-world datasets. Forecasting quality dominates detection performance: differences between detection methods (F1: 0.621-0.688) are smaller than between forecasting models (F1: 0.344-0.688). Our findings provide evidence-based guidance: use LSTM for complex patterns, Informer for efficiency-critical deployments, and classical methods for simple periodic data with resource constraints. The complete implementation and results establish baselines for future forecasting-based anomaly detection research.
☆ DUAL: Learning Diverse Kernels for Aggregated Two-sample and Independence Testing
To adapt kernel two-sample and independence testing to complex structured data, aggregation of multiple kernels is frequently employed to boost testing power compared to single-kernel tests. However, we observe a phenomenon that directly maximizing multiple kernel-based statistics may result in highly similar kernels that capture highly overlapping information, limiting the effectiveness of aggregation. To address this, we propose an aggregated statistic that explicitly incorporates kernel diversity based on the covariance between different kernels. Moreover, we identify a fundamental challenge: a trade-off between the diversity among kernels and the test power of individual kernels, i.e., the selected kernels should be both effective and diverse. This motivates a testing framework with selection inference, which leverages information from the training phase to select kernels with strong individual performance from the learned diverse kernel pool. We provide rigorous theoretical statements and proofs to show the consistency on the test power and control of Type-I error, along with asymptotic analysis of the proposed statistics. Lastly, we conducted extensive empirical experiments demonstrating the superior performance of our proposed approach across various benchmarks for both two-sample and independence testing.
☆ Test-Time Adaptation by Causal Trimming NeurIPS 2025
Test-time adaptation aims to improve model robustness under distribution shifts by adapting models with access to unlabeled target samples. A primary cause of performance degradation under such shifts is the model's reliance on features that lack a direct causal relationship with the prediction target. We introduce Test-time Adaptation by Causal Trimming (TACT), a method that identifies and removes non-causal components from representations for test distributions. TACT applies data augmentations that preserve causal features while varying non-causal ones. By analyzing the changes in the representations using Principal Component Analysis, TACT identifies the highest variance directions associated with non-causal features. It trims the representations by removing their projections on the identified directions, and uses the trimmed representations for the predictions. During adaptation, TACT continuously tracks and refines these directions to get a better estimate of non-causal features. We theoretically analyze the effectiveness of this approach and empirically validate TACT on real-world out-of-distribution benchmarks. TACT consistently outperforms state-of-the-art methods by a significant margin.
comment: Accepted to the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025); Code is available at https://github.com/NancyQuris/TACT
☆ Lightweight Facial Landmark Detection in Thermal Images via Multi-Level Cross-Modal Knowledge Transfer
Facial Landmark Detection (FLD) in thermal imagery is critical for applications in challenging lighting conditions, but it is hampered by the lack of rich visual cues. Conventional cross-modal solutions, like feature fusion or image translation from RGB data, are often computationally expensive or introduce structural artifacts, limiting their practical deployment. To address this, we propose Multi-Level Cross-Modal Knowledge Distillation (MLCM-KD), a novel framework that decouples high-fidelity RGB-to-thermal knowledge transfer from model compression to create both accurate and efficient thermal FLD models. A central challenge during knowledge transfer is the profound modality gap between RGB and thermal data, where traditional unidirectional distillation fails to enforce semantic consistency across disparate feature spaces. To overcome this, we introduce Dual-Injected Knowledge Distillation (DIKD), a bidirectional mechanism designed specifically for this task. DIKD establishes a connection between modalities: it not only guides the thermal student with rich RGB features but also validates the student's learned representations by feeding them back into the frozen teacher's prediction head. This closed-loop supervision forces the student to learn modality-invariant features that are semantically aligned with the teacher, ensuring a robust and profound knowledge transfer. Experiments show that our approach sets a new state-of-the-art on public thermal FLD benchmarks, notably outperforming previous methods while drastically reducing computational overhead.
☆ Refining Hybrid Genetic Search for CVRP via Reinforcement Learning-Finetuned LLM
While large language models (LLMs) are increasingly used as automated heuristic designers for vehicle routing problems (VRPs), current state-of-the-art methods predominantly rely on prompting massive, general-purpose models like GPT-4. This work challenges that paradigm by demonstrating that a smaller, specialized LLM, when meticulously fine-tuned, can generate components that surpass expert-crafted heuristics within advanced solvers. We propose RFTHGS, a novel Reinforcement learning (RL) framework for Fine-Tuning a small LLM to generate high-performance crossover operators for the Hybrid Genetic Search (HGS) solver, applied to the Capacitated VRP (CVRP). Our method employs a multi-tiered, curriculum-based reward function that progressively guides the LLM to master generating first compilable, then executable, and finally, superior-performing operators that exceed human expert designs. This is coupled with an operator caching mechanism that discourages plagiarism and promotes diversity during training. Comprehensive experiments show that our fine-tuned LLM produces crossover operators which significantly outperform the expert-designed ones in HGS. The performance advantage remains consistent, generalizing from small-scale instances to large-scale problems with up to 1000 nodes. Furthermore, RFTHGS exceeds the performance of leading neuro-combinatorial baselines, prompt-based methods, and commercial LLMs such as GPT-4o and GPT-4o-mini.
☆ PhysioME: A Robust Multimodal Self-Supervised Framework for Physiological Signals with Missing Modalities
Missing or corrupted modalities are common in physiological signal-based medical applications owing to hardware constraints or motion artifacts. However, most existing methods assume the availability of all modalities, resulting in substantial performance degradation in the absence of any modality. To overcome this limitation, this study proposes PhysioME, a robust framework designed to ensure reliable performance under missing modality conditions. PhysioME adopts: (1) a multimodal self-supervised learning approach that combines contrastive learning with masked prediction; (2) a Dual-PathNeuroNet backbone tailored to capture the temporal dynamics of each physiological signal modality; and (3) a restoration decoder that reconstructs missing modality tokens, enabling flexible processing of incomplete inputs. The experimental results show that PhysioME achieves high consistency and generalization performance across various missing modality scenarios. These findings highlight the potential of PhysioME as a reliable tool for supporting clinical decision-making in real-world settings with imperfect data availability.
comment: 9 pages, 2 figures
♻ ☆ On Convolutions, Intrinsic Dimension, and Diffusion Models
The manifold hypothesis asserts that data of interest in high-dimensional ambient spaces, such as image data, lies on unknown low-dimensional submanifolds. Diffusion models (DMs) -- which operate by convolving data with progressively larger amounts of Gaussian noise and then learning to revert this process -- have risen to prominence as the most performant generative models, and are known to be able to learn distributions with low-dimensional support. For a given datum in one of these submanifolds, we should thus intuitively expect DMs to have implicitly learned its corresponding local intrinsic dimension (LID), i.e. the dimension of the submanifold it belongs to. Kamkari et al. (2024b) recently showed that this is indeed the case by linking this LID to the rate of change of the log marginal densities of the DM with respect to the amount of added noise, resulting in an LID estimator known as FLIPD. LID estimators such as FLIPD have a plethora of uses, among others they quantify the complexity of a given datum, and can be used to detect outliers, adversarial examples and AI-generated text. FLIPD achieves state-of-the-art performance at LID estimation, yet its theoretical underpinnings are incomplete since Kamkari et al. (2024b) only proved its correctness under the highly unrealistic assumption of affine submanifolds. In this work we bridge this gap by formally proving the correctness of FLIPD under realistic assumptions. Additionally, we show that an analogous result holds when Gaussian convolutions are replaced with uniform ones, and discuss the relevance of this result.
comment: TMLR 2025 (expert certification)
♻ ☆ Causal Explanation of Concept Drift -- A Truly Actionable Approach ECML
In a world that constantly changes, it is crucial to understand how those changes impact different systems, such as industrial manufacturing or critical infrastructure. Explaining critical changes, referred to as concept drift in the field of machine learning, is the first step towards enabling targeted interventions to avoid or correct model failures, as well as malfunctions and errors in the physical world. Therefore, in this work, we extend model-based drift explanations towards causal explanations, which increases the actionability of the provided explanations. We evaluate our explanation strategy on a number of use cases, demonstrating the practical usefulness of our framework, which isolates the causally relevant features impacted by concept drift and, thus, allows for targeted intervention.
comment: This manuscript was presented at the TempXAI workshop at the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECMLPKDD 2025)
♻ ☆ Breaking the Compression Ceiling: Data-Free Pipeline for Ultra-Efficient Delta Compression NeurIPS 2025
With the rise of the fine-tuned-pretrained paradigm, storing numerous fine-tuned models for multi-tasking creates significant storage overhead. Delta compression alleviates this by storing only the pretrained model and the highly compressed delta weights (the differences between fine-tuned and pretrained model weights). However, existing methods fail to maintain both high compression and performance, and often rely on data. To address these challenges, we propose UltraDelta, the first data-free delta compression pipeline that achieves both ultra-high compression and strong performance. UltraDelta is designed to minimize redundancy, maximize information, and stabilize performance across inter-layer, intra-layer, and global dimensions, using three key components: (1) Variance-Based Mixed Sparsity Allocation assigns sparsity based on variance, giving lower sparsity to high-variance layers to preserve inter-layer information. (2) Distribution-Aware Compression applies uniform quantization and then groups parameters by value, followed by group-wise pruning, to better preserve intra-layer distribution. (3) Trace-Norm-Guided Rescaling uses the trace norm of delta weights to estimate a global rescaling factor, improving model stability under higher compression. Extensive experiments across (a) large language models (fine-tuned on LLaMA-2 7B and 13B) with up to 50x compression, (b) general NLP models (RoBERTa-base, T5-base) with up to 224x compression, (c) vision models (ViT-B/32, ViT-L/14) with up to 132x compression, and (d) multi-modal models (BEiT-3) with 18x compression, demonstrate that UltraDelta consistently outperforms existing methods, especially under ultra-high compression. Code is available at https://github.com/xiaohuiwang000/UltraDelta.
comment: Accepted at NeurIPS 2025
♻ ☆ Cost-aware Stopping for Bayesian Optimization
In automated machine learning, scientific discovery, and other applications of Bayesian optimization, deciding when to stop evaluating expensive black-box functions is an important practical consideration. While several adaptive stopping rules have been proposed, in the cost-aware setting they lack guarantees ensuring they stop before incurring excessive function evaluation costs. We propose a cost-aware stopping rule for Bayesian optimization that adapts to varying evaluation costs and is free of heuristic tuning. Our rule is grounded in a theoretical connection to state-of-the-art cost-aware acquisition functions, namely the Pandora's Box Gittins Index (PBGI) and log expected improvement per cost. We prove a theoretical guarantee bounding the expected cumulative evaluation cost incurred by our stopping rule when paired with these two acquisition functions. In experiments on synthetic and empirical tasks, including hyperparameter optimization and neural architecture size search, we show that combining our stopping rule with the PBGI acquisition function usually matches or outperforms other acquisition-function--stopping-rule pairs in terms of cost-adjusted simple regret, a metric capturing trade-offs between solution quality and cumulative evaluation cost.
♻ ☆ The Hidden Link Between RLHF and Contrastive Learning
Alignment of large language models (LLMs) with human values has recently garnered significant attention, with prominent examples including the canonical yet costly Reinforcement Learning from Human Feedback (RLHF) and the simple Direct Preference Optimization (DPO). In this work, we demonstrate that both RLHF and DPO can be interpreted from the perspective of mutual information (MI) maximization, uncovering a profound connection to contrastive learning. Within this framework, both RLHF and DPO can be interpreted as methods that performing contrastive learning based on the positive and negative samples derived from base model, leveraging the Donsker-Varadhan (DV) lower bound on MI (equivalently, the MINE estimator). Such paradigm further illuminates why RLHF may not intrinsically incentivize reasoning capacities in LLMs beyond what is already present in the base model. Building on the perspective, we replace the DV/MINE bound with the Jensen-Shannon (JS) MI estimator and propose the Mutual Information Optimization (MIO). Comprehensive theoretical analysis and extensive empirical evaluations demonstrate that MIO mitigates the late-stage decline in chosen-likelihood observed in DPO, achieving competitive or superior performance across various challenging reasoning and mathematical benchmarks.
♻ ☆ Provably faster randomized and quantum algorithms for $k$-means clustering via uniform sampling
The $k$-means algorithm (Lloyd's algorithm) is a widely used method for clustering unlabeled data. A key bottleneck of the $k$-means algorithm is that each iteration requires time linear in the number of data points, which can be expensive in big data applications. This was improved in recent works proposing quantum and quantum-inspired classical algorithms to approximate the $k$-means algorithm locally, in time depending only logarithmically on the number of data points (along with data dependent parameters) [q-means: A quantum algorithm for unsupervised machine learning, Kerenidis, Landman, Luongo, and Prakash, NeurIPS 2019; Do you know what $q$-means?, Cornelissen, Doriguello, Luongo, Tang, QTML 2025]. In this work, we describe a simple randomized mini-batch $k$-means algorithm and a quantum algorithm inspired by the classical algorithm. We demonstrate that the worst case guarantees of these algorithms can significantly improve upon the bounds for algorithms in prior work. Our improvements are due to a careful use of uniform sampling, which preserves certain symmetries of the $k$-means problem that are not preserved in previous algorithms that use data norm-based sampling.
comment: Included a comparison with [Do you know what $q$-means?, Cornelissen, Doriguello, Luongo, Tang, QTML 2025]; added numerical experiments for the classical algorithms; updated theory and fixed typos
♻ ☆ Holistic Evaluation of Multimodal LLMs on Spatial Intelligence
Multimodal models have achieved remarkable progress in recent years. Nevertheless, they continue to exhibit notable limitations in spatial understanding and reasoning, the very capability that anchors artificial general intelligence in the physical world. With the recent release of GPT-5, allegedly the most powerful AI model to date, it is timely to examine where the leading models (GPT, Gemini, Grok, Seed, Qwen, and Intern) stand on the path toward spatial intelligence. We first propose a holistic taxonomy of spatial tasks that unifies existing benchmarks and a standardized protocol for the fair evaluation of state-of-the-art proprietary and open-source models across eight key benchmarks, at a cost exceeding ten billion total tokens. Our empirical study then reveals that (1) GPT-5 demonstrates unprecedented strength in spatial intelligence (SI), yet (2) still falls short of human performance significantly across a broad spectrum of SI-tasks. Moreover, we (3) show that SI-tasks expose greater model capability deficiency than non-SI tasks, to the extent that (4) proprietary models do not exhibit a decisive advantage when facing the most difficult ones. In addition, we conduct a qualitative evaluation across a diverse set of scenarios that are intuitive for humans, yet fail even the most advanced multimodal models.
♻ ☆ Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot EMNLP25
In-Context Learning (ICL) is an essential emergent ability of Large Language Models (LLMs), and recent studies introduce Chain-of-Thought (CoT) to exemplars of ICL to enhance the reasoning capability, especially in mathematics tasks. However, given the continuous advancement of model capabilities, it remains unclear whether CoT exemplars still benefit recent, stronger models in such tasks. Through systematic experiments, we find that for recent strong models such as the Qwen2.5 series, adding traditional CoT exemplars does not improve reasoning performance compared to Zero-Shot CoT. Instead, their primary function is to align the output format with human expectations. We further investigate the effectiveness of enhanced CoT exemplars, constructed using answers from advanced models such as \texttt{Qwen2.5-Max} and \texttt{DeepSeek-R1}. Experimental results indicate that these enhanced exemplars still fail to improve the model's reasoning performance. Further analysis reveals that models tend to ignore the exemplars and focus primarily on the instructions, leading to no observable gain in reasoning ability. Overall, our findings highlight the limitations of the current ICL+CoT framework in mathematical reasoning, calling for a re-examination of the ICL paradigm and the definition of exemplars.
comment: EMNLP25-findings camera_ready, 19 pages,22 figures
♻ ☆ The Minimal Search Space for Conditional Causal Bandits ICLR2026
Causal knowledge can be used to support decision-making problems. This has been recognized in the causal bandits literature, where a causal (multi-armed) bandit is characterized by a causal graphical model and a target variable. The arms are then interventions on the causal model, and rewards are samples of the target variable. Causal bandits were originally studied with a focus on hard interventions. We focus instead on cases where the arms are conditional interventions, which more accurately model many real-world decision-making problems by allowing the value of the intervened variable to be chosen based on the observed values of other variables. This paper presents a graphical characterization of the minimal set of nodes guaranteed to contain the optimal conditional intervention, which maximizes the expected reward. We then propose an efficient algorithm with a time complexity of $O(|V| + |E|)$ to identify this minimal set of nodes. We prove that the graphical characterization and the proposed algorithm are correct. Finally, we empirically demonstrate that our algorithm significantly prunes the search space and substantially accelerates convergence rates when integrated into standard multi-armed bandit algorithms.
comment: Submitted to ICLR2026
♻ ☆ MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction NeurIPS 2025
We present MGE-LDM, a unified latent diffusion framework for simultaneous music generation, source imputation, and query-driven source separation. Unlike prior approaches constrained to fixed instrument classes, MGE-LDM learns a joint distribution over full mixtures, submixtures, and individual stems within a single compact latent diffusion model. At inference, MGE-LDM enables (1) complete mixture generation, (2) partial generation (i.e., source imputation), and (3) text-conditioned extraction of arbitrary sources. By formulating both separation and imputation as conditional inpainting tasks in the latent space, our approach supports flexible, class-agnostic manipulation of arbitrary instrument sources. Notably, MGE-LDM can be trained jointly across heterogeneous multi-track datasets (e.g., Slakh2100, MUSDB18, MoisesDB) without relying on predefined instrument categories. Audio samples are available at our project page: https://yoongi43.github.io/MGELDM_Samples/.
comment: Accepted by NeurIPS 2025
♻ ☆ Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models
We introduce Equilibrium Matching (EqM), a generative modeling framework built from an equilibrium dynamics perspective. EqM discards the non-equilibrium, time-conditional dynamics in traditional diffusion and flow-based generative models and instead learns the equilibrium gradient of an implicit energy landscape. Through this approach, we can adopt an optimization-based sampling process at inference time, where samples are obtained by gradient descent on the learned landscape with adjustable step sizes, adaptive optimizers, and adaptive compute. EqM surpasses the generation performance of diffusion/flow models empirically, achieving an FID of 1.90 on ImageNet 256$\times$256. EqM is also theoretically justified to learn and sample from the data manifold. Beyond generation, EqM is a flexible framework that naturally handles tasks including partially noised image denoising, OOD detection, and image composition. By replacing time-conditional velocities with a unified equilibrium landscape, EqM offers a tighter bridge between flow and energy-based models and a simple route to optimization-driven inference.
♻ ☆ Structured Kolmogorov-Arnold Neural ODEs for Interpretable Learning and Symbolic Discovery of Nonlinear Dynamics
Understanding and modeling nonlinear dynamical systems is a fundamental challenge across science and engineering. Deep learning has shown remarkable potential for capturing complex system behavior, yet achieving models that are both accurate and physically interpretable remains difficult. To address this, we propose Structured Kolmogorov-Arnold Neural ODEs (SKANODEs), a framework that integrates structured state-space modeling with Kolmogorov-Arnold Networks (KANs). Within a Neural ODE architecture, SKANODE employs a fully trainable KAN as a universal function approximator to perform virtual sensing, recovering latent states that correspond to interpretable physical quantities such as displacements and velocities. Leveraging KAN's symbolic regression capability, SKANODE then extracts compact, interpretable expressions for the system's governing dynamics. Extensive experiments on simulated and real-world systems demonstrate that SKANODE achieves superior predictive accuracy, discovers physics-consistent dynamics, and reveals complex nonlinear behavior. Notably, it identifies hysteretic behavior in an F-16 aircraft and recovers a concise symbolic equation describing this phenomenon. SKANODE thus enables interpretable, data-driven discovery of physically grounded models for complex nonlinear dynamical systems.
♻ ☆ Trust Region Reward Optimization and Proximal Inverse Reward Optimization Algorithm NeurIPS 2025
Inverse Reinforcement Learning (IRL) learns a reward function to explain expert demonstrations. Modern IRL methods often use the adversarial (minimax) formulation that alternates between reward and policy optimization, which often lead to unstable training. Recent non-adversarial IRL approaches improve stability by jointly learning reward and policy via energy-based formulations but lack formal guarantees. This work bridges this gap. We first present a unified view showing canonical non-adversarial methods explicitly or implicitly maximize the likelihood of expert behavior, which is equivalent to minimizing the expected return gap. This insight leads to our main contribution: Trust Region Reward Optimization (TRRO), a framework that guarantees monotonic improvement in this likelihood via a Minorization-Maximization process. We instantiate TRRO into Proximal Inverse Reward Optimization (PIRO), a practical and stable IRL algorithm. Theoretically, TRRO provides the IRL counterpart to the stability guarantees of Trust Region Policy Optimization (TRPO) in forward RL. Empirically, PIRO matches or surpasses state-of-the-art baselines in reward recovery, policy imitation with high sample efficiency on MuJoCo and Gym-Robotics benchmarks and a real-world animal behavior modeling task.
comment: Accepted to NeurIPS 2025. Title used at submission and review: PIRO: Toward Stable Reward Learning for Inverse RL via Monotonic Policy Divergence Reduction
♻ ☆ ORN-CBF: Learning Observation-conditioned Residual Neural Control Barrier Functions via Hypernetworks
Control barrier functions (CBFs) have been demonstrated as an effective method for safety-critical control of autonomous systems. Although CBFs are simple to deploy, their design remains challenging, motivating the development of learning-based approaches. Yet, issues such as suboptimal safe sets, applicability in partially observable environments, and lack of rigorous safety guarantees persist. In this work, we propose observation-conditioned neural CBFs based on Hamilton-Jacobi (HJ) reachability analysis, which approximately recover the maximal safe sets. We exploit certain mathematical properties of the HJ value function, ensuring that the predicted safe set never intersects with the observed failure set. Moreover, we leverage a hypernetwork-based architecture that is particularly suitable for the design of observation-conditioned safety filters. The proposed method is examined both in simulation and hardware experiments for a ground robot and a quadcopter. The results show improved success rates and generalization to out-of-domain environments compared to the baselines.
♻ ☆ Automated Machine Learning for Unsupervised Tabular Tasks
In this work, we present LOTUS (Learning to Learn with Optimal Transport for Unsupervised Scenarios), a simple yet effective method to perform model selection for multiple unsupervised machine learning(ML) tasks such as outlier detection and clustering. Our intuition behind this work is that a machine learning pipeline will perform well in a new dataset if it previously worked well on datasets with a similar underlying data distribution. We use Optimal Transport distances to find this similarity between unlabeled tabular datasets and recommend machine learning pipelines with one unified single method on two downstream unsupervised tasks: outlier detection and clustering. We present the effectiveness of our approach with experiments against strong baselines and show that LOTUS is a very promising first step toward model selection for multiple unsupervised ML tasks.
comment: Accepted at Machine Learning Journal, 2025
♻ ☆ Query Complexity of Classical and Quantum Channel Discrimination
Quantum channel discrimination has been studied from an information-theoretic perspective, wherein one is interested in the optimal decay rate of error probabilities as a function of the number of unknown channel accesses. In this paper, we study the query complexity of quantum channel discrimination, wherein the goal is to determine the minimum number of channel uses needed to reach a desired error probability. To this end, we show that the query complexity of binary channel discrimination depends logarithmically on the inverse error probability and inversely on the negative logarithm of the (geometric and Holevo) channel fidelity. As a special case of these findings, we precisely characterize the query complexity of discriminating two classical channels and two classical-quantum channels. Furthermore, by obtaining an optimal characterization of the sample complexity of quantum hypothesis testing, including prior probabilities, we provide a more precise characterization of query complexity when the error probability does not exceed a fixed threshold. We also provide lower and upper bounds on the query complexity of binary asymmetric channel discrimination and multiple quantum channel discrimination. For the former, the query complexity depends on the geometric R\'enyi and Petz R\'enyi channel divergences, while for the latter, it depends on the negative logarithm of the (geometric and Uhlmann) channel fidelity. For multiple channel discrimination, the upper bound scales as the logarithm of the number of channels.
comment: v3: 35 pages, published version; v2: 33 pages, Added tighter and precise characterization of sample and query complexity in Theorem 11 (for states), Theorem 12 (for general channels), and Corollaries 10 and 14 for classical-quantum channels; v1:22 pages; see also the independent work "Sampling complexity of quantum channel discrimination" DOI 10.1088/1572-9494/adcb9e
♻ ☆ Learning Satellite Attitude Dynamics with Physics-Informed Normalising Flow
Attitude control is a fundamental aspect of spacecraft operations. Model Predictive Control (MPC) has emerged as a powerful strategy for these tasks, relying on accurate models of the system dynamics to optimize control actions over a prediction horizon. In scenarios where physics models are incomplete, difficult to derive, or computationally expensive, machine learning offers a flexible alternative by learning the system behavior directly from data. However, purely data-driven models often struggle with generalization and stability, especially when applied to inputs outside their training domain. To address these limitations, we investigate the benefits of incorporating Physics-Informed Neural Networks (PINNs) into the learning of spacecraft attitude dynamics, comparing their performance with that of purely data-driven approaches. Using a Real-valued Non-Volume Preserving (Real NVP) neural network architecture with a self-attention mechanism, we trained several models on simulated data generated with the Basilisk simulator. Two training strategies were considered: a purely data-driven baseline and a physics-informed variant to improve robustness and stability. Our results demonstrate that the inclusion of physics-based information significantly enhances the performance in terms of the mean relative error of the best architectures found by 27.08%. These advantages are particularly evident when the learned models are integrated into an MPC framework, where PINN-based models consistently outperform their purely data-driven counterparts in terms of control accuracy and robustness, yielding improvements of up to 42.86% in performance stability error and increased robustness-to-noise.
♻ ☆ Evaluating LLMs for Demographic-Targeted Social Bias Detection: A Comprehensive Benchmark Study
Large-scale web-scraped text corpora used to train general-purpose AI models often contain harmful demographic-targeted social biases, creating a regulatory need for data auditing and developing scalable bias-detection methods. Although prior work has investigated biases in text datasets and related detection methods, these studies remain narrow in scope. They typically focus on a single content type (e.g., hate speech), cover limited demographic axes, overlook biases affecting multiple demographics simultaneously, and analyze limited techniques. Consequently, practitioners lack a holistic understanding of the strengths and limitations of recent large language models (LLMs) for automated bias detection. In this study, we present a comprehensive evaluation framework aimed at English texts to assess the ability of LLMs in detecting demographic-targeted social biases. To align with regulatory requirements, we frame bias detection as a multi-label task using a demographic-focused taxonomy. We then conduct a systematic evaluation with models across scales and techniques, including prompting, in-context learning, and fine-tuning. Using twelve datasets spanning diverse content types and demographics, our study demonstrates the promise of fine-tuned smaller models for scalable detection. However, our analyses also expose persistent gaps across demographic axes and multi-demographic targeted biases, underscoring the need for more effective and scalable auditing frameworks.
comment: 18 pages
♻ ☆ Beyond [cls]: Exploring the true potential of Masked Image Modeling representations
Masked Image Modeling (MIM) has emerged as a promising approach for Self-Supervised Learning (SSL) of visual representations. However, the out-of-the-box performance of MIMs is typically inferior to competing approaches. Most users cannot afford fine-tuning due to the need for large amounts of data, high GPU consumption, and specialized user knowledge. Therefore, the practical use of MIM representations is limited. In this paper we ask what is the reason for the poor out-of-the-box performance of MIMs. Is it due to weaker features produced by MIM models, or is it due to suboptimal usage? Through detailed analysis, we show that attention in MIMs is spread almost uniformly over many patches, leading to ineffective aggregation by the [cls] token. Based on this insight, we propose Selective Aggregation to better capture the rich semantic information retained in patch tokens, which significantly improves the out-of-the-box performance of MIM.
♻ ☆ Online Selective Generation with Adversarial Bandit Feedback
Large language generative models increasingly interact with humans, while their falsified responses raise concerns. To mitigate this hallucination effect, selectively abstaining from answering, called selective generation, provides an effective way for generators to control the hallucination when uncertain about their answers. However, as selective generators interact under adversarial environments and receive partial feedback from users on selected generation (e.g., thumbs up or down on the selected answer), learning methods for selective generation under such practical setups are crucial but currently missing. To address this limitation, we propose an online learning algorithm for selective generation with partial feedback under an adaptive adversary. In particular, we re-purpose an adversarial bandit algorithm to design an online selective generation method with controllable false discovery rates (FDR), which measures the rate of hallucination. The key building blocks include a novel conversion lemma from regret of any bandit algorithm to the FDR, and the exploitation of a unique structure of selective generation to reuse partial feedback, which we call feedback unlocking. We empirically evaluate the efficacy of the proposed online selective generation algorithm with partial feedback over diverse learning environments, demonstrating its ability to control the FDR, while maintaining reasonable selection efficiency, i.e., the ratio of non-abstaining answers, compared to baselines.
comment: 10 pages
♻ ☆ Task-Optimized Convolutional Recurrent Networks Align with Tactile Processing in the Rodent Brain NeurIPS 2025
Tactile sensing remains far less understood in neuroscience and less effective in artificial systems compared to more mature modalities such as vision and language. We bridge these gaps by introducing a novel Encoder-Attender-Decoder (EAD) framework to systematically explore the space of task-optimized temporal neural networks trained on realistic tactile input sequences from a customized rodent whisker-array simulator. We identify convolutional recurrent neural networks (ConvRNNs) as superior encoders to purely feedforward and state-space architectures for tactile categorization. Crucially, these ConvRNN-encoder-based EAD models achieve neural representations closely matching rodent somatosensory cortex, saturating the explainable neural variability and revealing a clear linear relationship between supervised categorization performance and neural alignment. Furthermore, contrastive self-supervised ConvRNN-encoder-based EADs, trained with tactile-specific augmentations, match supervised neural fits, serving as an ethologically-relevant, label-free proxy. For neuroscience, our findings highlight nonlinear recurrent processing as important for general-purpose tactile representations in somatosensory cortex, providing the first quantitative characterization of the underlying inductive biases in this system. For embodied AI, our results emphasize the importance of recurrent EAD architectures to handle realistic tactile inputs, along with tailored self-supervised learning methods for achieving robust tactile perception with the same type of sensors animals use to sense in unstructured environments.
comment: 10 pages, 8 figures, 7 tables, NeurIPS 2025 Camera Ready Version (oral)
♻ ☆ Evolving LLMs' Self-Refinement Capability via Synergistic Training-Inference Optimization
Self-Refinement refers to a model's ability to revise its own responses to produce improved outputs. This capability can also serve as a fundamental mechanism for Self-Improvement, for example, by reconstructing datasets with refined results to enhance intrinsic model performance. However, our comprehensive experiments reveal that large language models (LLMs) show no clear evidence of inherent Self-Refinement and may even experience response quality degradation after Self-Refinement. To address this issue, we propose EVOLVE, a simple and effective framework for eliciting and tracking the evolution of Self-Refinement through iterative training. We first explore optimization methods during training to activate the model's Self-Refinement capability. Then, at inference, we investigate various generation strategies to further enhance and utilize Self-Refinement while supplying the necessary data for training. Through synergistic optimization of training and inference stages, we continually evolve the model's Self-Refinement ability, enabling it to better refine its own responses. Moreover, we demonstrate the potential of leveraging Self-Refinement to achieve broader Self-Improvement of intrinsic model abilities. Experiments show that the evolved Self-Refinement ability enables the Llama-3.1-8B base model to surpass GPT-4o, achieving 62.3% length-controlled and 63.3% raw win rates on AlpacaEval 2, and 50.3% on Arena-Hard. It also generalizes effectively to out-of-domain reasoning tasks, improving performance on mathematical reasoning benchmarks such as GSM8K and MATH.
♻ ☆ Personalized Bayesian Federated Learning with Wasserstein Barycenter Aggregation NIPS 2025
Personalized Bayesian federated learning (PBFL) handles non-i.i.d. client data and quantifies uncertainty by combining personalization with Bayesian inference. However, existing PBFL methods face two limitations: restrictive parametric assumptions in client posterior inference and naive parameter averaging for server aggregation. To overcome these issues, we propose FedWBA, a novel PBFL method that enhances both local inference and global aggregation. At the client level, we use particle-based variational inference for nonparametric posterior representation. At the server level, we introduce particle-based Wasserstein barycenter aggregation, offering a more geometrically meaningful approach. Theoretically, we provide local and global convergence guarantees for FedWBA. Locally, we prove a KL divergence decrease lower bound per iteration for variational inference convergence. Globally, we show that the Wasserstein barycenter converges to the true parameter as the client data size increases. Empirically, experiments show that FedWBA outperforms baselines in prediction accuracy, uncertainty calibration, and convergence rate, with ablation studies confirming its robustness.
comment: The paper has been accepted by NIPS 2025
♻ ☆ Re-uploading quantum data: A universal function approximator for quantum inputs
Quantum data re-uploading has proved powerful for classical inputs, where repeatedly encoding features into a small circuit yields universal function approximation. Extending this idea to quantum inputs remains underexplored, as the information contained in a quantum state is not directly accessible in classical form. We propose and analyze a quantum data re-uploading architecture in which a qubit interacts sequentially with fresh copies of an arbitrary input state. The circuit can approximate any bounded continuous function using only one ancilla qubit and single-qubit measurements. By alternating entangling unitaries with mid-circuit resets of the input register, the architecture realizes a discrete cascade of completely positive and trace-preserving maps, analogous to collision models in open quantum system dynamics. Our framework provides a qubit-efficient and expressive approach to designing quantum machine learning models that operate directly on quantum data.
comment: 24 pages, 11 figures
♻ ☆ Investigating Memory in RL with POPGym Arcade
How should we analyze memory in deep RL? We introduce mathematical tools for fairly analyzing policies under partial observability and revealing how agents use memory to make decisions. To utilize these tools, we present POPGym Arcade, a collection of Atari-inspired, hardware-accelerated, pixel-based environments sharing a single observation and action space. Each environment provides fully and partially observable variants, enabling counterfactual studies on observability. We find that controlled studies are necessary for fair comparisons, and identify a pathology where value functions smear credit over irrelevant history. With this pathology, we demonstrate how out-of-distribution scenarios can contaminate memory, perturbing the policy far into the future, with implications for sim-to-real transfer and offline RL.
♻ ☆ Enhancing XAI Narratives through Multi-Narrative Refinement and Knowledge Distillation
Explainable Artificial Intelligence has become a crucial area of research, aiming to demystify the decision-making processes of deep learning models. Among various explainability techniques, counterfactual explanations have been proven particularly promising, as they offer insights into model behavior by highlighting minimal changes that would alter a prediction. Despite their potential, these explanations are often complex and technical, making them difficult for non-experts to interpret. To address this challenge, we propose a novel pipeline that leverages Language Models, large and small, to compose narratives for counterfactual explanations. We employ knowledge distillation techniques along with a refining mechanism to enable Small Language Models to perform comparably to their larger counterparts while maintaining robust reasoning abilities. In addition, we introduce a simple but effective evaluation method to assess natural language narratives, designed to verify whether the models' responses are in line with the factual, counterfactual ground truth. As a result, our proposed pipeline enhances both the reasoning capabilities and practical performance of student models, making them more suitable for real-world use cases.
♻ ☆ Multi-Modal Manipulation via Multi-Modal Policy Consensus
Effectively integrating diverse sensory modalities is crucial for robotic manipulation. However, the typical approach of feature concatenation is often suboptimal: dominant modalities such as vision can overwhelm sparse but critical signals like touch in contact-rich tasks, and monolithic architectures cannot flexibly incorporate new or missing modalities without retraining. Our method factorizes the policy into a set of diffusion models, each specialized for a single representation (e.g., vision or touch), and employs a router network that learns consensus weights to adaptively combine their contributions, enabling incremental of new representations. We evaluate our approach on simulated manipulation tasks in {RLBench}, as well as real-world tasks such as occluded object picking, in-hand spoon reorientation, and puzzle insertion, where it significantly outperforms feature-concatenation baselines on scenarios requiring multimodal reasoning. Our policy further demonstrates robustness to physical perturbations and sensor corruption. We further conduct perturbation-based importance analysis, which reveals adaptive shifts between modalities.
comment: 9 pages, 7 figures. Project website: https://policyconsensus.github.io
♻ ☆ Failure Prediction at Runtime for Generative Robot Policies NeurIPS 2025
Imitation learning (IL) with generative models, such as diffusion and flow matching, has enabled robots to perform complex, long-horizon tasks. However, distribution shifts from unseen environments or compounding action errors can still cause unpredictable and unsafe behavior, leading to task failure. Early failure prediction during runtime is therefore essential for deploying robots in human-centered and safety-critical environments. We propose FIPER, a general framework for Failure Prediction at Runtime for generative IL policies that does not require failure data. FIPER identifies two key indicators of impending failure: (i) out-of-distribution (OOD) observations detected via random network distillation in the policy's embedding space, and (ii) high uncertainty in generated actions measured by a novel action-chunk entropy score. Both failure prediction scores are calibrated using a small set of successful rollouts via conformal prediction. A failure alarm is triggered when both indicators, aggregated over short time windows, exceed their thresholds. We evaluate FIPER across five simulation and real-world environments involving diverse failure modes. Our results demonstrate that FIPER better distinguishes actual failures from benign OOD situations and predicts failures more accurately and earlier than existing methods. We thus consider this work an important step towards more interpretable and safer generative robot policies. Code, data and videos are available at https://tum-lsy.github.io/fiper_website.
comment: Project page: https://tum-lsy.github.io/fiper_website. 33 pages, 12 figures. Accepted to NeurIPS 2025
♻ ☆ From Data to Rewards: a Bilevel Optimization Perspective on Maximum Likelihood Estimation
Generative models form the backbone of modern machine learning, underpinning state-of-the-art systems in text, vision, and multimodal applications. While Maximum Likelihood Estimation has traditionally served as the dominant training paradigm, recent work have highlighted its limitations, particularly in generalization and susceptibility to catastrophic forgetting compared to Reinforcement Learning techniques, such as Policy Gradient methods. However, these approaches depend on explicit reward signals, which are often unavailable in practice, leaving open the fundamental problem of how to align generative models when only high-quality datasets are accessible. In this work, we address this challenge via a Bilevel Optimization framework, where the reward function is treated as the optimization variable of an outer-level problem, while a policy gradient objective defines the inner-level. We then conduct a theoretical analysis of this optimization problem in a tractable setting and extract insights that, as we demonstrate, generalize to applications such as tabular classification and model-based reinforcement learning. We release the code at https://github.com/abenechehab/nll_to_po .
♻ ☆ The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models NeurIPS 2025
Test-time adaptation (TTA) methods have gained significant attention for enhancing the performance of vision-language models (VLMs) such as CLIP during inference, without requiring additional labeled data. However, current TTA researches generally suffer from major limitations such as duplication of baseline results, limited evaluation metrics, inconsistent experimental settings, and insufficient analysis. These problems hinder fair comparisons between TTA methods and make it difficult to assess their practical strengths and weaknesses. To address these challenges, we introduce TTA-VLM, a comprehensive benchmark for evaluating TTA methods on VLMs. Our benchmark implements 8 episodic TTA and 7 online TTA methods within a unified and reproducible framework, and evaluates them across 15 widely used datasets. Unlike prior studies focused solely on CLIP, we extend the evaluation to SigLIP--a model trained with a Sigmoid loss--and include training-time tuning methods such as CoOp, MaPLe, and TeCoA to assess generality. Beyond classification accuracy, TTA-VLM incorporates various evaluation metrics, including robustness, calibration, out-of-distribution detection, and stability, enabling a more holistic assessment of TTA methods. Through extensive experiments, we find that 1) existing TTA methods produce limited gains compared to the previous pioneering work; 2) current TTA methods exhibit poor collaboration with training-time fine-tuning methods; 3) accuracy gains frequently come at the cost of reduced model trustworthiness. We release TTA-VLM to provide fair comparison and comprehensive evaluation of TTA methods for VLMs, and we hope it encourages the community to develop more reliable and generalizable TTA strategies.
comment: NeurIPS 2025 Datasets and Benchmarks Track. Github link: https://github.com/TomSheng21/tta-vlm
♻ ☆ Fine-tuning Behavioral Cloning Policies with Preference-Based Reinforcement Learning
Deploying reinforcement learning (RL) in robotics, industry, and health care is blocked by two obstacles: the difficulty of specifying accurate rewards and the risk of unsafe, data-hungry exploration. We address this by proposing a two-stage framework that first learns a safe initial policy from a reward-free dataset of expert demonstrations, then fine-tunes it online using preference-based human feedback. We provide the first principled analysis of this offline-to-online approach and introduce BRIDGE, a unified algorithm that integrates both signals via an uncertainty-weighted objective. We derive regret bounds that shrink with the number of offline demonstrations, explicitly connecting the quantity of offline data to online sample efficiency. We validate BRIDGE in discrete and continuous control MuJoCo environments, showing it achieves lower regret than both standalone behavioral cloning and online preference-based RL. Our work establishes a theoretical foundation for designing more sample-efficient interactive agents.
comment: 85 pages (11 + references and appendix), 9 figures. v2: added acknowledgements
♻ ☆ Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning
Large language model (LLM) unlearning aims to surgically remove the influence of undesired data or knowledge from an existing model while preserving its utility on unrelated tasks. This paradigm has shown promise in addressing privacy and safety concerns. However, recent findings reveal that unlearning effects are often fragile: post-unlearning manipulations such as weight quantization or fine-tuning can quickly neutralize the intended forgetting. Prior efforts to improve robustness primarily reformulate unlearning objectives by explicitly assuming the role of vulnerability sources. In this work, we take a different perspective by investigating the role of the optimizer, independent of unlearning objectives and formulations, in shaping unlearning robustness. We show that the 'grade' of the optimizer, defined by the level of information it exploits, ranging from zeroth-order (gradient-free) to first-order (gradient-based) to second-order (Hessian-based), is tightly linked to the resilience of unlearning. Surprisingly, we find that downgrading the optimizer, such as using zeroth-order methods or compressed-gradient variants (e.g., gradient sign-based optimizers), often leads to stronger robustness. While these optimizers produce noisier and less precise updates, they encourage convergence to harder-to-disturb basins in the loss landscape, thereby resisting post-training perturbations. By connecting zeroth-order methods with randomized smoothing, we further highlight their natural advantage for robust unlearning. Motivated by these insights, we propose a hybrid optimizer that combines first-order and zeroth-order updates, preserving unlearning efficacy while enhancing robustness. Extensive experiments on the MUSE and WMDP benchmarks, across multiple LLM unlearning algorithms, validate that our approach achieves more resilient forgetting without sacrificing unlearning quality.
♻ ☆ Train-before-Test Harmonizes Language Model Rankings
Existing language model benchmarks provide contradictory model rankings, even for benchmarks that aim to capture similar skills. This dilemma of conflicting rankings hampers model selection, clouds model comparisons, and adds confusion to a growing ecosystem of competing models. In this paper, we take a different perspective on model comparison: instead of relying on out-of-the-box performance via direct evaluation, we compare model potential by providing each model with identical benchmark-specific fine-tuning before evaluation. We call this approach train-before-test. Our primary contribution is a comprehensive empirical evaluation of model potential across 24 benchmarks and 61 models. First, we demonstrate that model potential rankings obtained through train-before-test exhibit remarkable consistency across all benchmarks. Whereas traditional rankings demonstrate little external validity under direct evaluation, they enjoy a significant degree of external validity when applying train-before-test: model potential rankings transfer gracefully from one benchmark to another. Second, train-before-test restores the connection between perplexity and downstream task performance, lost under direct evaluation. Remarkably, even pre-finetuning perplexity of a base model predicts post-finetuning downstream performance, suggesting that ranking consistency reflects inherent model potential rather than fine-tuning artifacts. Finally, train-before-test reduces the model-score matrix to essentially rank one, indicating that model potential is dominated by one latent factor, uncovered by train-before-test. Our work supports the recommendation to make train-before-test a default component of LLM benchmarking.
♻ ☆ PULSE: Practical Evaluation Scenarios for Large Multimodal Model Unlearning
In recent years, unlearning techniques, which are methods for inducing a model to "forget" previously learned information, have attracted attention as a way to address privacy and copyright concerns in large language models (LLMs) and large multimodal models (LMMs). While several unlearning benchmarks have been established for LLMs, a practical evaluation framework for unlearning in LMMs has been less explored. Specifically, existing unlearning benchmark for LMMs considers only scenarios in which the model is required to unlearn fine-tuned knowledge through a single unlearning operation. In this study, we introduce PULSE protocol for realistic unlearning scenarios for LMMs by introducing two critical perspectives: (i) Pre-trained knowledge Unlearning for analyzing the effect across different knowledge acquisition phases and (ii) Long-term Sustainability Evaluation to address sequential requests. We then evaluate existing unlearning methods along these dimensions. Our results reveal that, although some techniques can successfully unlearn knowledge acquired through fine-tuning, they struggle to eliminate information learned during pre-training. Moreover, methods that effectively unlearn a batch of target data in a single operation exhibit substantial performance degradation when the same data are split and unlearned sequentially.
♻ ☆ Superior Molecular Representations from Intermediate Encoder Layers
Pretrained molecular encoders have become indispensable in computational chemistry for tasks such as property prediction and molecular generation. However, the standard practice of relying solely on final-layer embeddings for downstream tasks may discard valuable information. In this work, we first analyze the information flow in five diverse molecular encoders and find that intermediate layers retain more general-purpose features, whereas the final-layer specializes and compresses information. We then perform an empirical layer-wise evaluation across 22 property prediction tasks. We find that using frozen embeddings from optimal intermediate layers improves downstream performance by an average of 5.4%, up to 28.6%, compared to the final-layer. Furthermore, finetuning encoders truncated at intermediate depths achieves even greater average improvements of 8.5%, with increases as high as 40.8%, obtaining new state-of-the-art results on several benchmarks. These findings highlight the importance of exploring the full representational depth of molecular encoders to achieve substantial performance improvements and computational efficiency. The code will be made publicly available.
♻ ☆ On Experiments
The scientific process is a means to turn the results of experiments into knowledge about the world in which we live. Much research effort has been directed toward automating this process. To do this, one needs to formulate the scientific process in a precise mathematical language. This paper outlines one such language. What is presented here is hardly new. The material is based on great thinkers from times past well as more modern contributions. The novel contributions of this paper are: A new general data processing inequality, a bias variance decomposition for canonical losses, streamlined proofs of the Blackwell-Sherman-Stein and Randomization theorems. means of calculating deficiency through linear programming.
♻ ☆ Robustness and Regularization in Hierarchical Re-Basin
This paper takes a closer look at Git Re-Basin, an interesting new approach to merge trained models. We propose a hierarchical model merging scheme that significantly outperforms the standard MergeMany algorithm. With our new algorithm, we find that Re-Basin induces adversarial and perturbation robustness into the merged models, with the effect becoming stronger the more models participate in the hierarchical merging scheme. However, in our experiments Re-Basin induces a much bigger performance drop than reported by the original authors.
comment: Published in 32th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN 2024
♻ ☆ Agentic large language models improve retrieval-based radiology question answering
Clinical decision-making in radiology increasingly benefits from artificial intelligence (AI), particularly through large language models (LLMs). However, traditional retrieval-augmented generation (RAG) systems for radiology question answering (QA) typically rely on single-step retrieval, limiting their ability to handle complex clinical reasoning tasks. Here we propose radiology Retrieval and Reasoning (RaR), a multi-step retrieval and reasoning framework designed to improve diagnostic accuracy, factual consistency, and clinical reliability of LLMs in radiology question answering. We evaluated 25 LLMs spanning diverse architectures, parameter scales (0.5B to >670B), and training paradigms (general-purpose, reasoning-optimized, clinically fine-tuned), using 104 expert-curated radiology questions from previously established RSNA-RadioQA and ExtendedQA datasets. To assess generalizability, we additionally tested on an unseen internal dataset of 65 real-world radiology board examination questions. RaR significantly improved mean diagnostic accuracy over zero-shot prompting and conventional online RAG. The greatest gains occurred in small-scale models, while very large models (>200B parameters) demonstrated minimal changes (<2% improvement). Additionally, RaR retrieval reduced hallucinations (mean 9.4%) and retrieved clinically relevant context in 46% of cases, substantially aiding factual grounding. Even clinically fine-tuned models showed gains from RaR (e.g., MedGemma-27B), indicating that retrieval remains beneficial despite embedded domain knowledge. These results highlight the potential of RaR to enhance factuality and diagnostic accuracy in radiology QA, warranting future studies to validate their clinical utility. All datasets, code, and the full RaR framework are publicly available to support open research and clinical translation.
♻ ☆ Edge Delayed Deep Deterministic Policy Gradient: efficient continuous control for edge scenarios
Deep Reinforcement Learning is gaining increasing attention thanks to its capability to learn complex policies in high-dimensional settings. Recent advancements utilize a dual-network architecture to learn optimal policies through the Q-learning algorithm. However, this approach has notable drawbacks, such as an overestimation bias that can disrupt the learning process and degrade the performance of the resulting policy. To address this, novel algorithms have been developed that mitigate overestimation bias by employing multiple Q-functions. Edge scenarios, which prioritize privacy, have recently gained prominence. In these settings, limited computational resources pose a significant challenge for complex Machine Learning approaches, making the efficiency of algorithms crucial for their performance. In this work, we introduce a novel Reinforcement Learning algorithm tailored for edge scenarios, called Edge Delayed Deep Deterministic Policy Gradient (EdgeD3). EdgeD3 enhances the Deep Deterministic Policy Gradient (DDPG) algorithm, achieving significantly improved performance with $25\%$ less Graphics Process Unit (GPU) time while maintaining the same memory usage. Additionally, EdgeD3 consistently matches or surpasses the performance of state-of-the-art methods across various benchmarks, all while using $30\%$ fewer computational resources and requiring $30\%$ less memory.
♻ ☆ Detecting and Mitigating Insertion Hallucination in Video-to-Audio Generation
Video-to-Audio generation has made remarkable strides in automatically synthesizing sound for video. However, existing evaluation metrics, which focus on semantic and temporal alignment, overlook a critical failure mode: models often generate acoustic events, particularly speech and music, that have no corresponding visual source. We term this phenomenon Insertion Hallucination and identify it as a systemic risk driven by dataset biases, such as the prevalence of off-screen sounds, that remains completely undetected by current metrics. To address this challenge, we first develop a systematic evaluation framework that employs a majority-voting ensemble of multiple audio event detectors. We also introduce two novel metrics to quantify the prevalence and severity of this issue: IH@vid (the fraction of videos with hallucinations) and IH@dur (the fraction of hallucinated duration). Building on this, we propose Posterior Feature Correction, a novel training-free inference-time method that mitigates IH. PFC operates in a two-pass process: it first generates an initial audio output to detect hallucinated segments, and then regenerates the audio after masking the corresponding video features at those timestamps. Experiments on several mainstream V2A benchmarks first reveal that state-of-the-art models suffer from severe IH. In contrast, our PFC method reduces both the prevalence and duration of hallucinations by over 50\% on average, without degrading, and in some cases even improving, conventional metrics for audio quality and temporal synchronization. Our work is the first to formally define, systematically measure, and effectively mitigate Insertion Hallucination, paving the way for more reliable and faithful V2A models.
♻ ☆ Simple and Effective Specialized Representations for Fair Classifiers
Fair classification is a critical challenge that has gained increasing importance due to international regulations and its growing use in high-stakes decision-making settings. Existing methods often rely on adversarial learning or distribution matching across sensitive groups; however, adversarial learning can be unstable, and distribution matching can be computationally intensive. To address these limitations, we propose a novel approach based on the characteristic function distance. Our method ensures that the learned representation contains minimal sensitive information while maintaining high effectiveness for downstream tasks. By utilizing characteristic functions, we achieve a more stable and efficient solution compared to traditional methods. Additionally, we introduce a simple relaxation of the objective function that guarantees fairness in common classification models with no performance degradation. Experimental results on benchmark datasets demonstrate that our approach consistently matches or achieves better fairness and predictive accuracy than existing methods. Moreover, our method maintains robustness and computational efficiency, making it a practical solution for real-world applications.
♻ ☆ J1: Incentivizing Thinking in LLM-as-a-Judge via Reinforcement Learning
The progress of AI is bottlenecked by the quality of evaluation, making powerful LLM-as-a-Judge models a core solution. The efficacy of these judges depends on their chain-of-thought reasoning, creating a critical need for methods that can effectively optimize this reasoning process. In this work, we introduce J1, a reinforcement learning framework for teaching LLM judges to think before making decisions. Our core contribution lies in converting all judgment tasks for non-verifiable and verifiable prompts into a unified format with verifiable rewards, enabling direct optimization of evaluation quality while mitigating positional bias. We then use RL to train thinking-judges at scales of 8B, 32B, and 70B and show that they obtain state-of-the-art performance across multiple benchmarks. In particular, J1-Qwen-32B, our multitasked pointwise and pairwise judge also outperforms o1-mini, o3, and a much larger 671B DeepSeek-R1 on some benchmarks, while only training on synthetic data. Through comprehensive ablations of pairwise, pointwise, and multitask J1 variants, we demonstrate the effectiveness of our approach across seed prompts, reward strategies, and training recipes. Qualitative analysis reveals that J1 develops systematic evaluation strategies, including dynamic criteria generation, reference answer creation, iterative self-correction of initial assessments, and feedback generation for low-quality responses.
comment: 10 pages, 13 tables, 14 figures
♻ ☆ Accurate and Diverse LLM Mathematical Reasoning via Automated PRM-Guided GFlowNets
Achieving both accuracy and diverse reasoning remains challenging for Large Language Models (LLMs) in complex domains like mathematics. A key bottleneck is evaluating intermediate reasoning steps to guide generation without costly human annotations. To address this, we first introduce a novel Process Reward Model (PRM) trained automatically using Monte Carlo Tree Search coupled with a similarity-based data augmentation technique, effectively capturing step-level reasoning quality. Leveraging this PRM, we then adapt Generative Flow Networks (GFlowNets) to operate at the reasoning step level. Unlike traditional reinforcement learning focused on maximizing a single reward, GFlowNets naturally sample diverse, high-quality solutions proportional to their rewards, as measured by our PRM. Empirical evaluation shows strong improvements in both accuracy and solution diversity on challenging mathematical benchmarks (e.g., +2.59% absolute accuracy on MATH Level 5 for Llama3.2-3B), with effective generalization to unseen datasets (+9.4\% absolute on SAT MATH). Furthermore, we benchmark our PRM against existing open-source reward models, demonstrating superior alignment with reasoning quality and more consistent guidance for downstream generation. Our work demonstrates the potential of PRM-guided, step-level GFlowNets for developing more robust and versatile mathematical reasoning in LLMs.
♻ ☆ Incentivize Contribution and Learn Parameters Too: Federated Learning with Strategic Data Owners
Classical federated learning (FL) assumes that the clients have a limited amount of noisy data with which they voluntarily participate and contribute towards learning a global, more accurate model in a principled manner. The learning happens in a distributed fashion without sharing the data with the center. However, these methods do not consider the incentive of an agent for participating and contributing to the process, given that data collection and running a distributed algorithm is costly for the clients. The question of rationality of contribution has been asked recently in the literature and some results exist that consider this problem. This paper addresses the question of simultaneous parameter learning and incentivizing contribution in a truthful manner, which distinguishes it from the extant literature. Our first mechanism incentivizes each client to contribute to the FL process at a Nash equilibrium and simultaneously learn the model parameters. We also ensure that agents are incentivized to truthfully reveal information in the intermediate stages of the algorithm. However, this equilibrium outcome can be away from the optimal, where clients contribute with their full data and the algorithm learns the optimal parameters. We propose a second mechanism that enables the full data contribution along with optimal parameter learning. Large scale experiments with real (federated) datasets (CIFAR-10, FEMNIST, and Twitter) show that these algorithms converge quite fast in practice, yield good welfare guarantees and better model performance for all agents.
comment: 25 pages, under review
♻ ☆ TTF-VLA: Temporal Token Fusion via Pixel-Attention Integration for Vision-Language-Action Models AAAI 2026
Vision-Language-Action (VLA) models process visual inputs independently at each timestep, discarding valuable temporal information inherent in robotic manipulation tasks. This frame-by-frame processing makes models vulnerable to visual noise while ignoring the substantial coherence between consecutive frames in manipulation sequences. We propose Temporal Token Fusion (TTF), a training-free approach that intelligently integrates historical and current visual representations to enhance VLA inference quality. Our method employs dual-dimension detection combining efficient grayscale pixel difference analysis with attention-based semantic relevance assessment, enabling selective temporal token fusion through hard fusion strategies and keyframe anchoring to prevent error accumulation. Comprehensive experiments across LIBERO, SimplerEnv, and real robot tasks demonstrate consistent improvements: 4.0 percentage points average on LIBERO (72.4\% vs 68.4\% baseline), cross-environment validation on SimplerEnv (4.8\% relative improvement), and 8.7\% relative improvement on real robot tasks. Our approach proves model-agnostic, working across OpenVLA and VLA-Cache architectures. Notably, TTF reveals that selective Query matrix reuse in attention mechanisms enhances rather than compromises performance, suggesting promising directions for direct KQV matrix reuse strategies that achieve computational acceleration while improving task success rates.
comment: Manuscript submitted to AAAI 2026, currently under review
♻ ☆ Scaling Equilibrium Propagation to Deeper Neural Network Architectures
Equilibrium propagation has been proposed as a biologically plausible alternative to the backpropagation algorithm. The local nature of gradient computations, combined with the use of convergent RNNs to reach equilibrium states, make this approach well-suited for implementation on neuromorphic hardware. However, previous studies on equilibrium propagation have been restricted to networks containing only dense layers or relatively small architectures with a few convolutional layers followed by a final dense layer. These networks have a significant gap in accuracy compared to similarly sized feedforward networks trained with backpropagation. In this work, we introduce the Hopfield-Resnet architecture, which incorporates residual (or skip) connections in Hopfield networks with clipped $\mathrm{ReLU}$ as the activation function. The proposed architectural enhancements enable the training of networks with nearly twice the number of layers reported in prior works. For example, Hopfield-Resnet13 achieves 93.92\% accuracy on CIFAR-10, which is $\approx$3.5\% higher than the previous best result and comparable to that provided by Resnet13 trained using backpropagation.
♻ ☆ Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator
Post-training of large language models is essential for adapting pre-trained language models (PLMs) to align with human preferences and downstream tasks. While PLMs typically exhibit well-calibrated confidence, post-trained language models (PoLMs) often suffer from over-confidence, assigning high confidence to both correct and incorrect outputs, which can undermine reliability in critical applications. A major obstacle in calibrating PoLMs is the scarcity of labeled data for individual downstream tasks. To address this, we propose Disagreement-Aware Confidence Alignment (DACA), a novel unsupervised method to optimize the parameters (e.g., temperature $\tau$) in post-hoc confidence calibration. Our method is motivated by the under-confidence issue caused by prediction disagreement between the PLM and PoLM while aligning their confidence via temperature scaling. Theoretically, the PLM's confidence underestimates PoLM's prediction accuracy on disagreement examples, causing a larger $\tau$ and producing under-confident predictions. DACA mitigates this by selectively using only agreement examples for calibration, effectively decoupling the influence of disagreement. In this manner, our method avoids an overly large $\tau$ in temperature scaling caused by disagreement examples, improving calibration performance. Extensive experiments demonstrate the effectiveness of our method, improving the average ECE of open-sourced and API-based LLMs (e.g. GPT-4o) by up to 15.08$\%$ on common benchmarks.
♻ ☆ Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling NeurIPS 2025
3D molecule generation is crucial for drug discovery and material science, requiring models to process complex multi-modalities, including atom types, chemical bonds, and 3D coordinates. A key challenge is integrating these modalities of different shapes while maintaining SE(3) equivariance for 3D coordinates. To achieve this, existing approaches typically maintain separate latent spaces for invariant and equivariant modalities, reducing efficiency in both training and sampling. In this work, we propose \textbf{U}nified Variational \textbf{A}uto-\textbf{E}ncoder for \textbf{3D} Molecular Latent Diffusion Modeling (\textbf{UAE-3D}), a multi-modal VAE that compresses 3D molecules into latent sequences from a unified latent space, while maintaining near-zero reconstruction error. This unified latent space eliminates the complexities of handling multi-modality and equivariance when performing latent diffusion modeling. We demonstrate this by employing the Diffusion Transformer--a general-purpose diffusion model without any molecular inductive bias--for latent generation. Extensive experiments on GEOM-Drugs and QM9 datasets demonstrate that our method significantly establishes new benchmarks in both \textit{de novo} and conditional 3D molecule generation, achieving leading efficiency and quality. On GEOM-Drugs, it reduces FCD by 72.6\% over the previous best result, while achieving over 70\% relative average improvements in geometric fidelity. Our code is released at https://github.com/lyc0930/UAE-3D/.
comment: NeurIPS 2025
♻ ☆ Faster Convergence of Riemannian Stochastic Gradient Descent with Increasing Batch Size ACML2025
We theoretically analyzed the convergence behavior of Riemannian stochastic gradient descent (RSGD) and found that using an increasing batch size leads to faster convergence than using a constant batch size, not only with a constant learning rate but also with a decaying learning rate, such as cosine annealing decay and polynomial decay. The convergence rate improves from $O(T^{-1}+C)$ with a constant batch size to $O(T^{-1})$ with an increasing batch size, where $T$ denotes the total number of iterations and $C$ is a constant. Using principal component analysis and low-rank matrix completion, we investigated, both theoretically and numerically, how an increasing batch size affects computational time as quantified by stochastic first-order oracle (SFO) complexity. An increasing batch size was found to reduce the SFO complexity of RSGD. Furthermore, an increasing batch size was found to offer the advantages of both small and large constant batch sizes.
comment: Accepted at ACML2025
♻ ☆ Curl Descent: Non-Gradient Learning Dynamics with Sign-Diverse Plasticity
Gradient-based algorithms are a cornerstone of artificial neural network training, yet it remains unclear whether biological neural networks use similar gradient-based strategies during learning. Experiments often discover a diversity of synaptic plasticity rules, but whether these amount to an approximation to gradient descent is unclear. Here we investigate a previously overlooked possibility: that learning dynamics may include fundamentally non-gradient "curl"-like components while still being able to effectively optimize a loss function. Curl terms naturally emerge in networks with inhibitory-excitatory connectivity or Hebbian/anti-Hebbian plasticity, resulting in learning dynamics that cannot be framed as gradient descent on any objective. To investigate the impact of these curl terms, we analyze feedforward networks within an analytically tractable student-teacher framework, systematically introducing non-gradient dynamics through neurons exhibiting rule-flipped plasticity. Small curl terms preserve the stability of the original solution manifold, resulting in learning dynamics similar to gradient descent. Beyond a critical value, strong curl terms destabilize the solution manifold. Depending on the network architecture, this loss of stability can lead to chaotic learning dynamics that destroy performance. In other cases, the curl terms can counterintuitively speed learning compared to gradient descent by allowing the weight dynamics to escape saddles by temporarily ascending the loss. Our results identify specific architectures capable of supporting robust learning via diverse learning rules, providing an important counterpoint to normative theories of gradient-based learning in neural networks.
♻ ☆ Learning Shared Representations from Unpaired Data NeurIPS 2025
Learning shared representations is a primary area of multimodal representation learning. The current approaches to achieve a shared embedding space rely heavily on paired samples from each modality, which are significantly harder to obtain than unpaired ones. In this work, we demonstrate that shared representations can be learned almost exclusively from unpaired data. Our arguments are grounded in the spectral embeddings of the random walk matrices constructed independently from each unimodal representation. Empirical results in computer vision and natural language processing domains support its potential, revealing the effectiveness of unpaired data in capturing meaningful cross-modal relations, demonstrating high capabilities in retrieval tasks, generation, arithmetics, zero-shot, and cross-domain classification. This work, to the best of our knowledge, is the first to demonstrate these capabilities almost exclusively from unpaired samples, giving rise to a cross-modal embedding that could be viewed as universal, i.e., independent of the specific modalities of the data. Our project page: https://shaham-lab.github.io/SUE_page.
comment: Accepted to NeurIPS 2025
♻ ☆ Muddit: Liberating Generation Beyond Text-to-Image with a Unified Discrete Diffusion Model
Unified generation models aim to handle diverse tasks across modalities -- such as text generation, image generation, and vision-language reasoning -- within a single architecture and decoding paradigm. Autoregressive unified models suffer from slow inference due to sequential decoding, and non-autoregressive unified models suffer from weak generalization due to limited pretrained backbones. We introduce Muddit, a unified discrete diffusion transformer that enables fast and parallel generation across both text and image modalities. Unlike prior unified diffusion models trained from scratch, Muddit integrates strong visual priors from a pretrained text-to-image backbone with a lightweight text decoder, enabling flexible and high-quality multimodal generation under a unified architecture. Empirical results show that Muddit achieves competitive or superior performance compared to significantly larger autoregressive models in both quality and efficiency. The work highlights the potential of purely discrete diffusion, when equipped with strong visual priors, as a scalable and effective backbone for unified generation.
comment: The code and model are available at https://github.com/M-E-AGI-Lab/Muddit
♻ ☆ Enhancing Self-Supervised Learning with Semantic Pairs A New Dataset and Empirical Study
Instance discrimination is a self-supervised representation learning paradigm wherein individual instances within a dataset are treated as distinct classes. This is typically achieved by generating two disparate views of each instance by applying stochastic transformations, encouraging the model to learn representations invariant to the common underlying object across these views. While this approach facilitates the acquisition of invariant representations for dataset instances under various handcrafted transformations (e.g., random cropping, colour jittering), an exclusive reliance on such data transformations for achieving invariance may inherently limit the model's generalizability to unseen datasets and diverse downstream tasks. The inherent limitation stems from the fact that the finite set of transformations within the data processing pipeline is unable to encompass the full spectrum of potential data variations. In this study, we provide the technical foundation for leveraging semantic pairs to enhance the generalizability of the model's representation and empirically demonstrate that incorporating semantic pairs mitigates the issue of limited transformation coverage. Specifically, we propose that by exposing the model to semantic pairs (i.e., two instances belonging to the same semantic category), we introduce varied real-world scene contexts, thereby fostering the development of more generalizable object representations. To validate this hypothesis, we constructed and released a novel dataset comprising curated semantic pairs and conducted extensive experimentation to empirically establish that their inclusion enables the model to learn more general representations, ultimately leading to improved performance across diverse downstream tasks.
comment: 16 pages, 7 figures, 5 tables
♻ ☆ TinyGraphEstimator: Adapting Lightweight Language Models for Graph Structure Inference
Graphs provide a universal framework for representing complex relational systems, and inferring their structural properties is a core challenge in graph analysis and reasoning. While large language models have recently demonstrated emerging abilities to perform symbolic and numerical reasoning, the potential of smaller, resource-efficient models in this context remains largely unexplored. This paper investigates whether compact transformer-based language models can infer graph-theoretic parameters directly from graph representations. To enable systematic evaluation, we introduce the TinyGraphEstimator dataset - a balanced collection of connected graphs generated from multiple random graph models and annotated with detailed structural metadata. We evaluate several small open models on their ability to predict key graph parameters such as density, clustering, and chromatic number. Furthermore, we apply lightweight fine-tuning using the Low-Rank Adaptation (LoRA) technique, achieving consistent improvements across all evaluated metrics. The results demonstrate that small language models possess non-trivial reasoning capacity over graph-structured data and can be effectively adapted for structural inference tasks through efficient parameter tuning.
♻ ☆ Long-Range Graph Wavelet Networks NeurIPS 2025
Modeling long-range interactions, the propagation of information across distant parts of a graph, is a central challenge in graph machine learning. Graph wavelets, inspired by multi-resolution signal processing, provide a principled way to capture both local and global structures. However, existing wavelet-based graph neural networks rely on finite-order polynomial approximations, which limit their receptive fields and hinder long-range propagation. We propose Long-Range Graph Wavelet Networks (LR-GWN), which decompose wavelet filters into complementary local and global components. Local aggregation is handled with efficient low-order polynomials, while long-range interactions are captured through a flexible spectral-domain parameterization. This hybrid design unifies short- and long-distance information flow within a principled wavelet framework. Experiments show that LR-GWN achieves state-of-the-art performance among wavelet-based methods on long-range benchmarks, while remaining competitive on short-range datasets.
comment: 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: New Perspectives in Advancing Graph Machine Learning
♻ ☆ References Indeed Matter? Reference-Free Preference Optimization for Conversational Query Reformulation
Conversational query reformulation (CQR) has become indispensable for improving retrieval in dialogue-based applications. However, existing approaches typically rely on reference passages for optimization, which are impractical to acquire in real-world scenarios. To address this limitation, we introduce a novel reference-free preference optimization framework DualReform that generates pseudo reference passages from commonly-encountered conversational datasets containing only queries and responses. DualReform attains this goal through two key innovations: (1) response-based inference, where responses serve as proxies to infer pseudo reference passages, and (2) response refinement via the dual-role of CQR, where a CQR model refines responses based on the shared objectives between response refinement and CQR. Despite not relying on reference passages, DualReform achieves 96.9--99.1% of the retrieval accuracy attainable only with reference passages and surpasses the state-of-the-art method by up to 31.6%.
♻ ☆ AMSbench: A Comprehensive Benchmark for Evaluating MLLM Capabilities in AMS Circuits
Analog/Mixed-Signal (AMS) circuits play a critical role in the integrated circuit (IC) industry. However, automating Analog/Mixed-Signal (AMS) circuit design has remained a longstanding challenge due to its difficulty and complexity. Although recent advances in Multi-modal Large Language Models (MLLMs) offer promising potential for supporting AMS circuit analysis and design, current research typically evaluates MLLMs on isolated tasks within the domain, lacking a comprehensive benchmark that systematically assesses model capabilities across diverse AMS-related challenges. To address this gap, we introduce AMSbench, a benchmark suite designed to evaluate MLLM performance across critical tasks including circuit schematic perception, circuit analysis, and circuit design. AMSbench comprises approximately 8000 test questions spanning multiple difficulty levels and assesses eight prominent models, encompassing both open-source and proprietary solutions such as Qwen 2.5-VL and Gemini 2.5 Pro. Our evaluation highlights significant limitations in current MLLMs, particularly in complex multi-modal reasoning and sophisticated circuit design tasks. These results underscore the necessity of advancing MLLMs' understanding and effective application of circuit-specific knowledge, thereby narrowing the existing performance gap relative to human expertise and moving toward fully automated AMS circuit design workflows. Our data is released at this URL.
FSA: An Alternative Efficient Implementation of Native Sparse Attention Kernel
Recent advance in sparse attention mechanisms has demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs). Native Sparse Attention (NSA), one state-of-the-art approach, introduces natively trainable, hardware-aligned sparse attention that delivers substantial system-level performance boost while maintaining accuracy comparable to full attention. However, the kernel implementation of NSA forces a loop order that is only efficient with a relatively large number of query heads in each Grouped Query Attention (GQA) group, whereas existing LLMs widely adopt much smaller number of query heads in each GQA group -- such an inconsistency significantly limits the applicability of this sparse algorithmic advance. In this work, we propose Flash Sparse Attention (FSA), an alternative kernel implementation that enables efficient NSA computation across a wide range of popular LLMs with varied smaller number of heads in each GQA group on modern GPUs. Compared to vanilla NSA kernel implementation, our empirical evaluation demonstrates that FSA achieves (i) up to 3.5x and on average 1.6x kernel-level latency reduction, (ii) up to 1.25x and 1.09x on average end-to-end training speedup on state-of-the-art LLMs, and (iii) up to 1.36x and 1.11x on average for prefill-phase speedup in LLM generative inference. Github Repo at https://github.com/Relaxed-System-Lab/Flash-Sparse-Attention.
♻ ☆ AB-UPT: Scaling Neural CFD Surrogates for High-Fidelity Automotive Aerodynamics Simulations via Anchored-Branched Universal Physics Transformers
Recent advances in neural surrogate modeling offer the potential for transformative innovations in applications such as automotive aerodynamics. Yet, industrial-scale problems often involve volumetric meshes with cell counts reaching 100 million, presenting major scalability challenges. Complex geometries further complicate modeling through intricate surface-volume interactions, while quantities such as vorticity are highly nonlinear and must satisfy strict divergence-free constraints. To address these requirements, we introduce AB-UPT as a novel modeling scheme for building neural surrogates for CFD simulations. AB-UPT is designed to: (i) decouple geometry encoding and prediction tasks via multi-branch operators; (ii) enable scalability to high-resolution outputs via neural simulation in a low-dimensional latent space, coupled with anchored neural field decoders to predict high-fidelity outputs; (iii) enforce physics consistency by a divergence-free formulation. We show that AB-UPT yields state-of-the-art predictive accuracy of surface and volume fields on automotive CFD simulations ranging from 33 thousand up to 150 million mesh cells. Furthermore, our anchored neural field architecture enables the enforcement of hard physical constraints on the physics predictions without degradation in performance, exemplified by modeling divergence-free vorticity fields. Notably, the proposed models can be trained on a single GPU in less than a day and predict industry-standard surface and volume fields within seconds. Additionally, we show that the flexible design of our method enables neural simulation from a CAD geometry alone, thereby eliminating the need for costly CFD meshing procedures for inference.
comment: Published in Transactions on Machine Learning Research
♻ ☆ On the Mathematical Relationship Between Layer Normalization and Dynamic Activation Functions
Layer normalization (LN) is an essential component of modern neural networks. While many alternative techniques have been proposed, none of them have succeeded in replacing LN so far. The latest suggestion in this line of research is a dynamic activation function called Dynamic Tanh (DyT). Although it is empirically well-motivated and appealing from a practical point of view, it lacks a theoretical foundation. In this work, we shed light on the mathematical relationship between LN and dynamic activation functions. In particular, we derive DyT from the LN variant RMSNorm, and show that a well-defined decoupling in derivative space as well as an approximation are needed to do so. By applying the same decoupling procedure directly in function space, we are able to omit the approximation and obtain the exact element-wise counterpart of RMSNorm, which we call Dynamic Inverse Square Root Unit (DyISRU). We demonstrate numerically that DyISRU reproduces the normalization effect on outliers more accurately than DyT does.
comment: Revision and Simplification (starting point RMSNorm instead of LN)
♻ ☆ Early-Warning of Thunderstorm-Driven Power Outages with a Two-Stage Machine Learning Model
Thunderstorm-driven outages are difficult to predict because most storms do not cause damage, convective processes occur rapidly and chaotically, and the available public data are both noisy and incomplete. We develop a 24-48 h early-warning model for summer, thunderstorm-related outages in Michigan using only open sources (EAGLE-I for ground truth; METAR for weather). We use the publicly released EAGLE-I outage dataset (2014-2022), maintained by Oak Ridge National Laboratory for the U.S. Department of Energy. The pipeline preserves convective micro-signals from a sparse station network via parameter-specific kriging with hourly variograms and targeted overdrafting to retain extremes, and builds causal spatio-temporal features (lags/rolling statistics; k-NN/IDW spatial aggregates) capturing precursors of severe convection (moisture advection, wind shifts, and pressure drops). The two-stage model design, combining a logistic gate and an LSTM regressor, limits routine periods and reduces noise exposure. The study uses event-centric metrics (cluster-based hits/misses/false alarms) and peak-conditional MASE (cMASE) in +/-Delta-hour windows around state-level peaks (>= 50,000), with uncertainty quantified by hourly moving-block bootstrap. On the test sample, Two-Stage detects more reference peaks across all windows (e.g., at +/-48 h it records 3/4 vs. 2/4; F1 66.7% vs. 57.1%) with one extra false alarm. Near peaks, it shows modest amplitude gains (2-3% lower cMASE at +/-0-12 h; bootstrap medians +9-13% at +/-6-12 h) but small losses at +/-36-48 h (~3-4%). Overall, errors are comparable to the one-step LSTM baseline. SHAP analysis confirms moisture-advection and wind/gust precursors, underscoring the value of the feature engineering. Despite open-data noise, the feature-driven pipeline yields actionable, event-focused early warnings for thunderstorm outages.
comment: 23 pages (main), 70 pages incl. appendices; figures & tables as in manuscript. Code (main figure, synthetic data): https://github.com/IrynaStanishevska/peak-outage-forecasting- License: CC BY 4.0 (preprint)
♻ ☆ Specialization after Generalization: Towards Understanding Test-Time Training in Foundation Models NeurIPS 2025
Recent empirical studies have explored the idea of continuing to train a model at test-time for a given task, known as test-time training (TTT), and have found it to yield significant performance improvements. However, there is limited understanding of why and when TTT is effective. Earlier explanations mostly focused on the observation that TTT may help when applied to out-of-distribution adaptation or used with privileged data. However, the growing scale of foundation models with most test data being in-distribution questions these explanations. We instead posit that foundation models remain globally underparameterized, with TTT providing a mechanism for specialization after generalization, focusing capacity on concepts relevant to the test task. Specifically, under the linear representation hypothesis, we propose a model in which TTT achieves a substantially smaller in-distribution test error than global training. We empirically validate our model's key assumptions by training a sparse autoencoder on ImageNet, showing that semantically related data points are explained by only a few shared concepts. Finally, we perform scaling studies across image and language tasks that confirm the practical implications of our model, identifying the regimes where specialization is most effective.
comment: Oral at CCFM @ NeurIPS 2025
♻ ☆ VLMGuard-R1: Proactive Safety Alignment for VLMs via Reasoning-Driven Prompt Optimization
Aligning Vision-Language Models (VLMs) with safety standards is essential to mitigate risks arising from their multimodal complexity, where integrating vision and language unveils subtle threats beyond the reach of conventional safeguards. Inspired by the insight that reasoning across modalities is key to preempting intricate vulnerabilities, we propose a novel direction for VLM safety: multimodal reasoning-driven prompt rewriting. To this end, we introduce VLMGuard-R1, a proactive framework that refines user inputs through a reasoning-guided rewriter, dynamically interpreting text-image interactions to deliver refined prompts that bolster safety across diverse VLM architectures without altering their core parameters. To achieve this, we devise a three-stage reasoning pipeline to synthesize a dataset that trains the rewriter to infer subtle threats, enabling tailored, actionable responses over generic refusals. Extensive experiments across three benchmarks with five VLMs reveal that VLMGuard-R1 outperforms four baselines. In particular, VLMGuard-R1 achieves a remarkable 43.59\% increase in average safety across five models on the SIUO benchmark.
Multimedia 15
☆ Building and Evaluating a Realistic Virtual World for Large Scale Urban Exploration from 360° Videos
We propose to build realistic virtual worlds, called 360RVW, for large urban environments directly from 360{\deg} videos. We provide an interface for interactive exploration, where users can freely navigate via their own avatars. 360{\deg} videos record the entire environment of the shooting location simultaneously leading to highly realistic and immersive representations. Our system uses 360{\deg} videos recorded along streets and builds a 360RVW through four main operations: video segmentation by intersection detection, video completion to remove the videographer, semantic segmentation for virtual collision detection with the avatar, and projection onto a distorted sphere that moves along the camera trajectory following the avatar's movements. Our interface allows users to explore large urban environments by changing their walking direction at intersections or choosing a new location by clicking on a map. Even without a 3D model, the users can experience collision with buildings using metadata produced by semantic segmentation. Furthermore, we stream the 360{\deg} videos so users can directly access 360RVW via their web browser. We fully evaluate our system, including a perceptual experiment comparing our approach to previous exploratory interfaces. The results confirm the quality of our system, especially regarding the presence of users and the interactive exploration, making it most suitable for a virtual tour of urban environments.
comment: Multimedia Tools and Applications, Springer (accepted)
☆ CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation
Existing works on reasoning segmentation either connect hidden features from a language model directly to a mask decoder or represent positions in text, which limits interpretability and semantic detail. To solve this, we present CoPRS, a Multi-modal Chain-of-Thought (MCoT)-based positional perception model that bridges language reasoning to segmentation through a differentiable and interpretable positional prior instantiated as a heatmap. By making the reasoning process clear via MCoT and expressing it as a dense, differentiable heatmap, this interface enhances interpretability and diagnostic analysis and yields more concentrated evidence on the target. A learnable concentration token aggregates features of the image and reasoning text to generate this positional prior, which is decoded to precise masks through a lightweight decoder, providing a direct connection between reasoning and segmentation. Across the RefCOCO series and ReasonSeg, CoPRS matches or surpasses the best reported metrics on each standard split under comparable protocols, with performance at or above prior state of the art across both validation and test partitions. Extensive experiments reveal that the quality of the heatmap strongly influences the resulting mask quality, supporting a consistent association between the reasoning output and downstream mask generation. Collectively, these findings support the utility of this paradigm in bridging reasoning and segmentation and show advantages in concentration driven by reasoning and predicting masks more precisely. Code, checkpoints and logs are released at https://github.com/ZhenyuLU-Heliodore/CoPRS.git.
comment: 18 pages, 6 figures, 6 tables
☆ Connecting Giants: Synergistic Knowledge Transfer of Large Multimodal Models for Few-Shot Learning IJCAI 2025
Few-shot learning (FSL) addresses the challenge of classifying novel classes with limited training samples. While some methods leverage semantic knowledge from smaller-scale models to mitigate data scarcity, these approaches often introduce noise and bias due to the data's inherent simplicity. In this paper, we propose a novel framework, Synergistic Knowledge Transfer (SynTrans), which effectively transfers diverse and complementary knowledge from large multimodal models to empower the off-the-shelf few-shot learner. Specifically, SynTrans employs CLIP as a robust teacher and uses a few-shot vision encoder as a weak student, distilling semantic-aligned visual knowledge via an unsupervised proxy task. Subsequently, a training-free synergistic knowledge mining module facilitates collaboration among large multimodal models to extract high-quality semantic knowledge. Building upon this, a visual-semantic bridging module enables bi-directional knowledge transfer between visual and semantic spaces, transforming explicit visual and implicit semantic knowledge into category-specific classifier weights. Finally, SynTrans introduces a visual weight generator and a semantic weight reconstructor to adaptively construct optimal multimodal FSL classifiers. Experimental results on four FSL datasets demonstrate that SynTrans, even when paired with a simple few-shot vision encoder, significantly outperforms current state-of-the-art methods.
comment: Accepted by IJCAI 2025
☆ Data or Language Supervision: What Makes CLIP Better than DINO? EMNLP 2025
CLIP outperforms self-supervised models like DINO as vision encoders for vision-language models (VLMs), but it remains unclear whether this advantage stems from CLIP's language supervision or its much larger training data. To disentangle these factors, we pre-train CLIP and DINO under controlled settings -- using the same architecture, dataset, and training configuration -- achieving similar ImageNet accuracy. Embedding analysis shows that CLIP captures high-level semantics (e.g., object categories, text), while DINO is more responsive to low-level features like colors and styles. When integrated into VLMs and evaluated on 20 VQA benchmarks, CLIP excels at text-intensive tasks, while DINO slightly outperforms on vision-centric ones. Variants of language supervision (e.g., sigmoid loss, pre-trained language encoders) yield limited gains. Our findings provide scientific insights into vision encoder design and its impact on VLM performance.
comment: EMNLP 2025 Findings
☆ Audio-Guided Visual Perception for Audio-Visual Navigation
Audio-Visual Embodied Navigation aims to enable agents to autonomously navigate to sound sources in unknown 3D environments using auditory cues. While current AVN methods excel on in-distribution sound sources, they exhibit poor cross-source generalization: navigation success rates plummet and search paths become excessively long when agents encounter unheard sounds or unseen environments. This limitation stems from the lack of explicit alignment mechanisms between auditory signals and corresponding visual regions. Policies tend to memorize spurious \enquote{acoustic fingerprint-scenario} correlations during training, leading to blind exploration when exposed to novel sound sources. To address this, we propose the AGVP framework, which transforms sound from policy-memorable acoustic fingerprint cues into spatial guidance. The framework first extracts global auditory context via audio self-attention, then uses this context as queries to guide visual feature attention, highlighting sound-source-related regions at the feature level. Subsequent temporal modeling and policy optimization are then performed. This design, centered on interpretable cross-modal alignment and region reweighting, reduces dependency on specific acoustic fingerprints. Experimental results demonstrate that AGVP improves both navigation efficiency and robustness while achieving superior cross-scenario generalization on previously unheard sounds.
comment: Main paper (6 pages). Accepted for publication by International Conference on Virtual Reality and Visualization 2025 (ICVRV 2025)
♻ ☆ Holistic Evaluation of Multimodal LLMs on Spatial Intelligence
Multimodal models have achieved remarkable progress in recent years. Nevertheless, they continue to exhibit notable limitations in spatial understanding and reasoning, the very capability that anchors artificial general intelligence in the physical world. With the recent release of GPT-5, allegedly the most powerful AI model to date, it is timely to examine where the leading models (GPT, Gemini, Grok, Seed, Qwen, and Intern) stand on the path toward spatial intelligence. We first propose a holistic taxonomy of spatial tasks that unifies existing benchmarks and a standardized protocol for the fair evaluation of state-of-the-art proprietary and open-source models across eight key benchmarks, at a cost exceeding ten billion total tokens. Our empirical study then reveals that (1) GPT-5 demonstrates unprecedented strength in spatial intelligence (SI), yet (2) still falls short of human performance significantly across a broad spectrum of SI-tasks. Moreover, we (3) show that SI-tasks expose greater model capability deficiency than non-SI tasks, to the extent that (4) proprietary models do not exhibit a decisive advantage when facing the most difficult ones. In addition, we conduct a qualitative evaluation across a diverse set of scenarios that are intuitive for humans, yet fail even the most advanced multimodal models.
♻ ☆ Towards Robust and Realible Multimodal Fake News Detection with Incomplete Modality
Multimodal fake news detection (MFND) has become an urgent task with the emergence of huge multimodal fake content on social media platforms. Previous studies mainly focus on complex feature extraction and fusion to learn discriminative information from multimodal content. However, in real-world applications, multimedia news may naturally lose some information during dissemination, resulting in modality incompleteness, which is detrimental to the generalization and robustness of existing models. To this end, we propose a novel generic and robust multimodal fusion strategy, termed Multi-expert Modality-incomplete Learning Network (MMLNet), which is simple yet effective. It consists of three key steps: (1) Multi-Expert Collaborative Reasoning to compensate for missing modalities by dynamically leveraging complementary information through multiple experts. (2) Incomplete Modality Adapters compensates for the missing information by leveraging the new feature distribution. (3) Modality Missing Learning leveraging an label-aware adaptive weighting strategy to learn a robust representation with contrastive learning. We evaluate MMLNet on three real-world benchmarks across two languages, demonstrating superior performance compared to state-of-the-art methods while maintaining relative simplicity. By ensuring the accuracy of fake news detection in incomplete modality scenarios caused by information propagation, MMLNet effectively curbs the spread of malicious misinformation. Code is publicly available at https://github.com/zhyhome/MMLNet.
♻ ☆ Real-Time Position-Aware View Synthesis from Single-View Input
Recent advancements in view synthesis have significantly enhanced immersive experiences across various computer graphics and multimedia applications, including telepresence and entertainment. By enabling the generation of new perspectives from a single input view, view synthesis allows users to better perceive and interact with their environment. However, many state-of-the-art methods, while achieving high visual quality, face limitations in real-time performance, which makes them less suitable for live applications where low latency is critical. In this paper, we present a lightweight, position-aware network designed for real-time view synthesis from a single input image and a target camera pose. The proposed framework consists of a Position Aware Embedding, which efficiently maps positional information from the target pose to generate high dimensional feature maps. These feature maps, along with the input image, are fed into a Rendering Network that merges features from dual encoder branches to resolve both high and low level details, producing a realistic new view of the scene. Experimental results demonstrate that our method achieves superior efficiency and visual quality compared to existing approaches, particularly in handling complex translational movements without explicit geometric operations like warping. This work marks a step toward enabling real-time live and interactive telepresence applications.
♻ ☆ SWIFT: Semantic Watermarking for Image Forgery Thwarting
This paper proposes a novel approach towards image authentication and tampering detection by using watermarking as a communication channel for semantic information. We modify the HiDDeN deep-learning watermarking architecture to embed and extract high-dimensional real vectors representing image captions. Our method improves significantly robustness on both malign and benign edits. We also introduce a local confidence metric correlated with Message Recovery Rate, enhancing the method's practical applicability. This approach bridges the gap between traditional watermarking and passive forensic methods, offering a robust solution for image integrity verification.
comment: Accepted at IEEE WIFS 2024; Code released at : https://github.com/gautierevn/swift_watermarking
♻ ☆ Tailored Teaching with Balanced Difficulty: Elevating Reasoning in Multimodal Chain-of-Thought via Prompt Curriculum
The effectiveness of Multimodal Chain-of-Thought (MCoT) prompting is often limited by the use of randomly or manually selected examples. These examples fail to account for both model-specific knowledge distributions and the intrinsic complexity of the tasks, resulting in suboptimal and unstable model performance. To address this, we propose a novel framework inspired by the pedagogical principle of "tailored teaching with balanced difficulty". We reframe prompt selection as a prompt curriculum design problem: constructing a well ordered set of training examples that align with the model's current capabilities. Our approach integrates two complementary signals: (1) model-perceived difficulty, quantified through prediction disagreement in an active learning setup, capturing what the model itself finds challenging; and (2) intrinsic sample complexity, which measures the inherent difficulty of each question-image pair independently of any model. By jointly analyzing these signals, we develop a difficulty-balanced sampling strategy that ensures the selected prompt examples are diverse across both dimensions. Extensive experiments conducted on five challenging benchmarks and multiple popular Multimodal Large Language Models (MLLMs) demonstrate that our method yields substantial and consistent improvements and greatly reduces performance discrepancies caused by random sampling, providing a principled and robust approach for enhancing multimodal reasoning.
♻ ☆ Context Guided Transformer Entropy Modeling for Video Compression ICCV 2025
Conditional entropy models effectively leverage spatio-temporal contexts to reduce video redundancy. However, incorporating temporal context often introduces additional model complexity and increases computational cost. In parallel, many existing spatial context models lack explicit modeling the ordering of spatial dependencies, which may limit the availability of relevant context during decoding. To address these issues, we propose the Context Guided Transformer (CGT) entropy model, which estimates probability mass functions of the current frame conditioned on resampled temporal context and dependency-weighted spatial context. A temporal context resampler learns predefined latent queries to extract critical temporal information using transformer encoders, reducing downstream computational overhead. Meanwhile, a teacher-student network is designed as dependency-weighted spatial context assigner to explicitly model the dependency of spatial context order. The teacher generates an attention map to represent token importance and an entropy map to reflect prediction certainty from randomly masked inputs, guiding the student to select the weighted top-k tokens with the highest spatial dependency. During inference, only the student is used to predict undecoded tokens based on high-dependency context. Experimental results demonstrate that our CGT model reduces entropy modeling time by approximately 65% and achieves an 11% BD-Rate reduction compared to the previous state-of-the-art conditional entropy model.
comment: ICCV 2025. This is an update to the camera-ready version
♻ ☆ Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance ICASSP 2025
Speech Emotion Recognition (SER) systems rely on speech input and emotional labels annotated by humans. However, various emotion databases collect perceptional evaluations in different ways. For instance, the IEMOCAP dataset uses video clips with sounds for annotators to provide their emotional perceptions. However, the most significant English emotion dataset, the MSP-PODCAST, only provides speech for raters to choose the emotional ratings. Nevertheless, using speech as input is the standard approach to training SER systems. Therefore, the open question is the emotional labels elicited by which scenarios are the most effective for training SER systems. We comprehensively compare the effectiveness of SER systems trained with labels elicited by different modality stimuli and evaluate the SER systems on various testing conditions. Also, we introduce an all-inclusive label that combines all labels elicited by various modalities. We show that using labels elicited by voice-only stimuli for training yields better performance on the test set, whereas labels elicited by voice-only stimuli.
comment: 5 pages, 2 figures, 4 tables, acceptance for ICASSP 2025
♻ ☆ Goal-Based Vision-Language Driving
Autonomous vehicles must react in milliseconds while reasoning about road geometry and traffic intent to navigate complex situations. We introduce NovaDrive, a single-branch vision-language architecture that processes front-camera images, HD-map tiles, LiDAR depth, and textual waypoints in a single branch. A lightweight, two-stage cross-attention block first aligns waypoint tokens with the HD map, then refines attention over fine-grained image and depth patches. Coupled with a novel smoothness loss that discourages abrupt steering and speed changes, this design eliminates the need for recurrent memory. We fine-tune the top 15 layers of an 11B LLaMA-3.2 vision-language backbone, enabling real-time inference. On the nuScenes / Waymo subset of the MD-NEX Outdoor benchmark, NovaDrive raises success rate to 84% (+4%), boosts path-efficiency (SPL) to 0.66 (+0.11), and reduces collision frequency from 2.6% to 1.2% (-1.4%) relative to the previous state-of-the-art. Our ablations confirm that waypoint tokens, partial VLM fine-tuning, and the cross-attention fusion each contribute the most to these gains. Beyond safety, NovaDrive's shorter routes (resulting from the novel smoothness loss) translate to lower fuel or battery usage, pointing toward leaner, more easily updated driving stacks. NovaDrive can be extended to other embodied-AI domains as well.
comment: 6 pages
♻ ☆ TAG-WM: Tamper-Aware Generative Image Watermarking via Diffusion Inversion Sensitivity
AI-generated content (AIGC) enables efficient visual creation but raises copyright and authenticity risks. As a common technique for integrity verification and source tracing, digital image watermarking is regarded as a potential solution to above issues. However, the widespread adoption and advancing capabilities of generative image editing tools have amplified malicious tampering risks, while simultaneously posing new challenges to passive tampering detection and watermark robustness. To address these challenges, this paper proposes a Tamper-Aware Generative image WaterMarking method named TAG-WM. The proposed method comprises four key modules: a dual-mark joint sampling (DMJS) algorithm for embedding copyright and localization watermarks into the latent space while preserving generative quality, the watermark latent reconstruction (WLR) utilizing reversed DMJS, a dense variation region detector (DVRD) leveraging diffusion inversion sensitivity to identify tampered areas via statistical deviation analysis, and the tamper-aware decoding (TAD) guided by localization results. The experimental results demonstrate that TAG-WM achieves state-of-the-art performance in both tampering robustness and localization capability even under distortion, while preserving lossless generation quality and maintaining a watermark capacity of 256 bits. The code is available at: https://github.com/Suchenl/TAG-WM.
♻ ☆ SMC++: Masked Learning of Unsupervised Video Semantic Compression ICCV 2023
Most video compression methods focus on human visual perception, neglecting semantic preservation. This leads to severe semantic loss during the compression, hampering downstream video analysis tasks. In this paper, we propose a Masked Video Modeling (MVM)-powered compression framework that particularly preserves video semantics, by jointly mining and compressing the semantics in a self-supervised manner. While MVM is proficient at learning generalizable semantics through the masked patch prediction task, it may also encode non-semantic information like trivial textural details, wasting bitcost and bringing semantic noises. To suppress this, we explicitly regularize the non-semantic entropy of the compressed video in the MVM token space. The proposed framework is instantiated as a simple Semantic-Mining-then-Compression (SMC) model. Furthermore, we extend SMC as an advanced SMC++ model from several aspects. First, we equip it with a masked motion prediction objective, leading to better temporal semantic learning ability. Second, we introduce a Transformer-based compression module, to improve the semantic compression efficacy. Considering that directly mining the complex redundancy among heterogeneous features in different coding stages is non-trivial, we introduce a compact blueprint semantic representation to align these features into a similar form, fully unleashing the power of the Transformer-based compression module. Extensive results demonstrate the proposed SMC and SMC++ models show remarkable superiority over previous traditional, learnable, and perceptual quality-oriented video codecs, on three video analysis tasks and seven datasets. \textit{Codes and model are available at: https://github.com/tianyuan168326/VideoSemanticCompression-Pytorch.
comment: Accepted to TPAMI; Substantial Extension of ICCV 2023 paper
Information Retrieval 10
☆ VeritasFi: An Adaptable, Multi-tiered RAG Framework for Multi-modal Financial Question Answering
Retrieval-Augmented Generation (RAG) is becoming increasingly essential for Question Answering (QA) in the financial sector, where accurate and contextually grounded insights from complex public disclosures are crucial. However, existing financial RAG systems face two significant challenges: (1) they struggle to process heterogeneous data formats, such as text, tables, and figures; and (2) they encounter difficulties in balancing general-domain applicability with company-specific adaptation. To overcome these challenges, we present VeritasFi, an innovative hybrid RAG framework that incorporates a multi-modal preprocessing pipeline alongside a cutting-edge two-stage training strategy for its re-ranking component. VeritasFi enhances financial QA through three key innovations: (1) A multi-modal preprocessing pipeline that seamlessly transforms heterogeneous data into a coherent, machine-readable format. (2) A tripartite hybrid retrieval engine that operates in parallel, combining deep multi-path retrieval over a semantically indexed document corpus, real-time data acquisition through tool utilization, and an expert-curated memory bank for high-frequency questions, ensuring comprehensive scope, accuracy, and efficiency. (3) A two-stage training strategy for the document re-ranker, which initially constructs a general, domain-specific model using anonymized data, followed by rapid fine-tuning on company-specific data for targeted applications. By integrating our proposed designs, VeritasFi presents a groundbreaking framework that greatly enhances the adaptability and robustness of financial RAG systems, providing a scalable solution for both general-domain and company-specific QA tasks. Code accompanying this work is available at https://github.com/simplew4y/VeritasFi.git.
☆ DRIFT: Decompose, Retrieve, Illustrate, then Formalize Theorems
Automating the formalization of mathematical statements for theorem proving remains a major challenge for Large Language Models (LLMs). LLMs struggle to identify and utilize the prerequisite mathematical knowledge and its corresponding formal representation in languages like Lean. Current retrieval-augmented autoformalization methods query external libraries using the informal statement directly, but overlook a fundamental limitation: informal mathematical statements are often complex and offer limited context on the underlying math concepts. To address this, we introduce DRIFT, a novel framework that enables LLMs to decompose informal mathematical statements into smaller, more tractable ''sub-components''. This facilitates targeted retrieval of premises from mathematical libraries such as Mathlib. Additionally, DRIFT retrieves illustrative theorems to help models use premises more effectively in formalization tasks. We evaluate DRIFT across diverse benchmarks (ProofNet, ConNF, and MiniF2F-test) and find that it consistently improves premise retrieval, nearly doubling the F1 score compared to the DPR baseline on ProofNet. Notably, DRIFT demonstrates strong performance on the out-of-distribution ConNF benchmark, with BEq+@10 improvements of 37.14% and 42.25% using GPT-4.1 and DeepSeek-V3.1, respectively. Our analysis shows that retrieval effectiveness in mathematical autoformalization depends heavily on model-specific knowledge boundaries, highlighting the need for adaptive retrieval strategies aligned with each model's capabilities.
☆ Is Implicit Knowledge Enough for LLMs? A RAG Approach for Tree-based Structures
Large Language Models (LLMs) are adept at generating responses based on information within their context. While this ability is useful for interacting with structured data like code files, another popular method, Retrieval-Augmented Generation (RAG), retrieves relevant documents to augment the model's in-context learning. However, it is not well-explored how to best represent this retrieved knowledge for generating responses on structured data, particularly hierarchical structures like trees. In this work, we propose a novel bottom-up method to linearize knowledge from tree-like structures (like a GitHub repository) by generating implicit, aggregated summaries at each hierarchical level. This approach enables the knowledge to be stored in a knowledge base and used directly with RAG. We then compare our method to using RAG on raw, unstructured code, evaluating the accuracy and quality of the generated responses. Our results show that while response quality is comparable across both methods, our approach generates over 68% fewer documents in the retriever, a significant gain in efficiency. This finding suggests that leveraging implicit, linearized knowledge may be a highly effective and scalable strategy for handling complex, hierarchical data structures.
comment: Waiting for Conference Response
☆ Multi-Granularity Sequence Denoising with Weakly Supervised Signal for Sequential Recommendation
Sequential recommendation aims to predict the next item based on user interests in historical interaction sequences. Historical interaction sequences often contain irrelevant noisy items, which significantly hinders the performance of recommendation systems. Existing research employs unsupervised methods that indirectly identify item-granularity irrelevant noise by predicting the ground truth item. Since these methods lack explicit noise labels, they are prone to misidentify users' interested items as noise. Additionally, while these methods focus on removing item-granularity noise driven by the ground truth item, they overlook interest-granularity noise, limiting their ability to perform broader denoising based on user interests. To address these issues, we propose Multi-Granularity Sequence Denoising with Weakly Supervised Signal for Sequential Recommendation(MGSD-WSS). MGSD-WSS first introduces the Multiple Gaussian Kernel Perceptron module to map the original and enhance sequence into a common representation space and utilizes weakly supervised signals to accurately identify noisy items in the historical interaction sequence. Subsequently, it employs the item-granularity denoising module with noise-weighted contrastive learning to obtain denoised item representations. Then, it extracts target interest representations from the ground truth item and applies noise-weighted contrastive learning to obtain denoised interest representations. Finally, based on the denoised item and interest representations, MGSD-WSS predicts the next item. Extensive experiments on five datasets demonstrate that the proposed method significantly outperforms state-of-the-art sequence recommendation and denoising models. Our code is available at https://github.com/lalunex/MGSD-WSS.
☆ Self-Supervised Representation Learning with ID-Content Modality Alignment for Sequential Recommendation
Sequential recommendation (SR) models often capture user preferences based on the historically interacted item IDs, which usually obtain sub-optimal performance when the interaction history is limited. Content-based sequential recommendation has recently emerged as a promising direction that exploits items' textual and visual features to enhance preference learning. However, there are still three key challenges: (i) how to reduce the semantic gap between different content modality representations; (ii) how to jointly model user behavior preferences and content preferences; and (iii) how to design an effective training strategy to align ID representations and content representations. To address these challenges, we propose a novel model, self-supervised representation learning with ID-Content modality alignment, named SICSRec. Firstly, we propose a LLM-driven sample construction method and develop a supervised fine-tuning approach to align item-level modality representations. Secondly, we design a novel Transformer-based sequential model, where an ID-modality sequence encoder captures user behavior preferences, a content-modality sequence encoder learns user content preferences, and a mix-modality sequence decoder grasps the intrinsic relationship between these two types of preferences. Thirdly, we propose a two-step training strategy with a content-aware contrastive learning task to align modality representations and ID representations, which decouples the training process of content modality dependency and item collaborative dependency. Extensive experiments conducted on four public video streaming datasets demonstrate our SICSRec outperforms the state-of-the-art ID-modality sequential recommenders and content-modality sequential recommenders by 8.04% on NDCG@5 and 6.62% on NDCD@10 on average, respectively.
☆ Towards Long-Term User Welfare in Recommender Systems via Creator-Oriented Information Revelation
Improving the long-term user welfare (e.g., sustained user engagement) has become a central objective of recommender systems (RS). In real-world platforms, the creation behaviors of content creators plays a crucial role in shaping long-term welfare beyond short-term recommendation accuracy, making the effective steering of creator behavior essential to foster a healthier RS ecosystem. Existing works typically rely on re-ranking algorithms that heuristically adjust item exposure to steer creators' behavior. However, when embedded within recommendation pipelines, such a strategy often conflicts with the short-term objective of improving recommendation accuracy, leading to performance degradation and suboptimal long-term welfare. The well-established economics studies offer us valuable insights for an alternative approach without relying on recommendation algorithmic design: revealing information from an information-rich party (sender) to a less-informed party (receiver) can effectively change the receiver's beliefs and steer their behavior. Inspired by this idea, we propose an information-revealing framework, named Long-term Welfare Optimization via Information Revelation (LoRe). In this framework, we utilize a classical information revelation method (i.e., Bayesian persuasion) to map the stakeholders in RS, treating the platform as the sender and creators as the receivers. To address the challenge posed by the unrealistic assumption of traditional economic methods, we formulate the process of information revelation as a Markov Decision Process (MDP) and propose a learning algorithm trained and inferred in environments with boundedly rational creators. Extensive experiments on two real-world RS datasets demonstrate that our method can effectively outperform existing fair re-ranking methods and information revealing strategies in improving long-term user welfare.
☆ Does Weighting Improve Matrix Factorization for Recommender Systems? WWW
Matrix factorization is a widely used approach for top-N recommendation and collaborative filtering. When implemented on implicit feedback data (such as clicks), a common heuristic is to upweight the observed interactions. This strategy has been shown to improve performance for certain algorithms. In this paper, we conduct a systematic study of various weighting schemes and matrix factorization algorithms. Somewhat surprisingly, we find that training with unweighted data can perform comparably to, and sometimes outperform, training with weighted data, especially for large models. This observation challenges the conventional wisdom. Nevertheless, we identify cases where weighting can be beneficial, particularly for models with lower capacity and specific regularization schemes. We also derive efficient algorithms for exactly minimizing several weighted objectives that were previously considered computationally intractable. Our work provides a comprehensive analysis of the interplay between weighting, regularization, and model capacity in matrix factorization for recommender systems.
comment: In the proceedings of the Web Conference (WWW) 2025 (11 pages)
☆ Hierarchical LoRA MoE for Efficient CTR Model Scaling
Deep models have driven significant advances in click-through rate (CTR) prediction. While vertical scaling via layer stacking improves model expressiveness, the layer-by-layer sequential computation poses challenges to efficient scaling. Conversely, horizontal scaling through Mixture of Experts (MoE) achieves efficient scaling by activating a small subset of experts in parallel, but flat MoE layers may struggle to capture the hierarchical structure inherent in recommendation tasks. To push the Return-On-Investment (ROI) boundary, we explore the complementary strengths of both directions and propose HiLoMoE, a hierarchical LoRA MoE framework that enables holistic scaling in a parameter-efficient manner. Specifically, HiLoMoE employs lightweight rank-1 experts for parameter-efficient horizontal scaling, and stacks multiple MoE layers with hierarchical routing to enable combinatorially diverse expert compositions. Unlike conventional stacking, HiLoMoE routes based on prior layer scores rather than outputs, allowing all layers to execute in parallel. A principled three-stage training framework ensures stable optimization and expert diversity. Experiments on four public datasets show that HiLoMoE achieving better performance-efficiency tradeoff, achieving an average AUC improvement of 0.20\% in AUC and 18.5\% reduction in FLOPs compared to the non-MoE baseline.
comment: 13 pages, 9 figures
☆ ZeroGR: A Generalizable and Scalable Framework for Zero-Shot Generative Retrieval
Generative retrieval (GR) reformulates information retrieval (IR) by framing it as the generation of document identifiers (docids), thereby enabling an end-to-end optimization and seamless integration with generative language models (LMs). Despite notable progress under supervised training, GR still struggles to generalize to zero-shot IR scenarios, which are prevalent in real-world applications. To tackle this challenge, we propose \textsc{ZeroGR}, a zero-shot generative retrieval framework that leverages natural language instructions to extend GR across a wide range of IR tasks. Specifically, \textsc{ZeroGR} is composed of three key components: (i) an LM-based docid generator that unifies heterogeneous documents (e.g., text, tables, code) into semantically meaningful docids; (ii) an instruction-tuned query generator that generates diverse types of queries from natural language task descriptions to enhance corpus indexing; and (iii) a reverse annealing decoding strategy to balance precision and recall during docid generation. We investigate the impact of instruction fine-tuning scale and find that performance consistently improves as the number of IR tasks encountered during training increases. Empirical results on the BEIR and MAIR benchmarks demonstrate that \textsc{ZeroGR} outperforms strong dense retrieval and generative baselines in zero-shot settings, establishing a new state-of-the-art for instruction-driven GR.
♻ ☆ Retro*: Optimizing LLMs for Reasoning-Intensive Document Retrieval
With the growing popularity of LLM agents and RAG, it has become increasingly important to retrieve documents that are essential for solving a task, even when their connection to the task is indirect or implicit. Addressing this problem requires fine-grained reasoning to accurately assess the relevance between the task and each candidate document. This capability, however, poses a significant challenge for existing IR techniques. Despite recent progress in reasoning-enhanced IR, existing approaches still face significant challenges in applicability, scalability, and efficiency. In this work, we propose Retro*, a novel approach for reasoning-intensive document retrieval. Our method introduces a rubric-based relevance scoring mechanism, enabling the model to reason about the relationship between a task and a document based on explicitly defined criteria, whereby producing a fine-grained, interpretable relevance score. Retro* also supports test-time scaling by combining multiple reasoning trajectories via score integration, which produces more reliable relevance estimates. To optimize Retro*'s reasoning capabilities, we introduce a novel reinforcement learning algorithm tailored for its relevance scoring mechanism, which employs two composite rewards to fully exploit the trajectories of each training sample. Our experiments show that Retro* outperforms existing document retrieval methods with notable advantages, leading to state-of-the-art performance on the BRIGHT benchmark.
Multimedia 3
☆ JND-Guided Light-Weight Neural Pre-Filter for Perceptual Image Coding ISCA
Just Noticeable Distortion (JND)-guided pre-filter is a promising technique for improving the perceptual compression efficiency of image coding. However, existing methods are often computationally expensive, and the field lacks standardized benchmarks for fair comparison. To address these challenges, this paper introduces a twofold contribution. First, we develop and open-source FJNDF-Pytorch, a unified benchmark for frequency-domain JND-Guided pre-filters. Second, leveraging this platform, we propose a complete learning framework for a novel, lightweight Convolutional Neural Network (CNN). Experimental results demonstrate that our proposed method achieves state-of-the-art compression efficiency, consistently outperforming competitors across multiple datasets and encoders. In terms of computational cost, our model is exceptionally lightweight, requiring only 7.15 GFLOPs to process a 1080p image, which is merely 14.1% of the cost of recent lightweight network. Our work presents a robust, state-of-the-art solution that excels in both performance and efficiency, supported by a reproducible research platform. The open-source implementation is available at https://github.com/viplab-fudan/FJNDF-Pytorch.
comment: 5 pages, 4 figures. Submitted to the IEEE International Symposium on Circuits and Systems (ISCAS) 2026
☆ MCE: Towards a General Framework for Handling Missing Modalities under Imbalanced Missing Rates
Multi-modal learning has made significant advances across diverse pattern recognition applications. However, handling missing modalities, especially under imbalanced missing rates, remains a major challenge. This imbalance triggers a vicious cycle: modalities with higher missing rates receive fewer updates, leading to inconsistent learning progress and representational degradation that further diminishes their contribution. Existing methods typically focus on global dataset-level balancing, often overlooking critical sample-level variations in modality utility and the underlying issue of degraded feature quality. We propose Modality Capability Enhancement (MCE) to tackle these limitations. MCE includes two synergistic components: i) Learning Capability Enhancement (LCE), which introduces multi-level factors to dynamically balance modality-specific learning progress, and ii) Representation Capability Enhancement (RCE), which improves feature semantics and robustness through subset prediction and cross-modal completion tasks. Comprehensive evaluations on four multi-modal benchmarks show that MCE consistently outperforms state-of-the-art methods under various missing configurations. The journal preprint version is now available at https://doi.org/10.1016/j.patcog.2025.112591. Our code is available at https://github.com/byzhaoAI/MCE.
comment: This is the accepted version of an article that has been published in \textbf{Pattern Recognition}. The final published version will be available soon
☆ Towards Efficient 3D Gaussian Human Avatar Compression: A Prior-Guided Framework
This paper proposes an efficient 3D avatar coding framework that leverages compact human priors and canonical-to-target transformation to enable high-quality 3D human avatar video compression at ultra-low bit rates. The framework begins by training a canonical Gaussian avatar using articulated splatting in a network-free manner, which serves as the foundation for avatar appearance modeling. Simultaneously, a human-prior template is employed to capture temporal body movements through compact parametric representations. This decomposition of appearance and temporal evolution minimizes redundancy, enabling efficient compression: the canonical avatar is shared across the sequence, requiring compression only once, while the temporal parameters, consisting of just 94 parameters per frame, are transmitted with minimal bit-rate. For each frame, the target human avatar is generated by deforming canonical avatar via Linear Blend Skinning transformation, facilitating temporal coherent video reconstruction and novel view synthesis. Experimental results demonstrate that the proposed method significantly outperforms conventional 2D/3D codecs and existing learnable dynamic 3D Gaussian splatting compression method in terms of rate-distortion performance on mainstream multi-view human video datasets, paving the way for seamless immersive multimedia experiences in meta-verse applications.
comment: 10 pages, 4 figures
Information Retrieval 12
☆ ImCoref-CeS: An Improved Lightweight Pipeline for Coreference Resolution with LLM-based Checker-Splitter Refinement
Coreference Resolution (CR) is a critical task in Natural Language Processing (NLP). Current research faces a key dilemma: whether to further explore the potential of supervised neural methods based on small language models, whose detect-then-cluster pipeline still delivers top performance, or embrace the powerful capabilities of Large Language Models (LLMs). However, effectively combining their strengths remains underexplored. To this end, we propose \textbf{ImCoref-CeS}, a novel framework that integrates an enhanced supervised model with LLM-based reasoning. First, we present an improved CR method (\textbf{ImCoref}) to push the performance boundaries of the supervised neural method by introducing a lightweight bridging module to enhance long-text encoding capability, devising a biaffine scorer to comprehensively capture positional information, and invoking a hybrid mention regularization to improve training efficiency. Importantly, we employ an LLM acting as a multi-role Checker-Splitter agent to validate candidate mentions (filtering out invalid ones) and coreference results (splitting erroneous clusters) predicted by ImCoref. Extensive experiments demonstrate the effectiveness of ImCoref-CeS, which achieves superior performance compared to existing state-of-the-art (SOTA) methods.
☆ Text2Token: Unsupervised Text Representation Learning with Token Target Prediction
Unsupervised text representation learning (TRL) is a fundamental task in natural language processing, which is beneficial for improving search and recommendations with the web's unlabeled texts. A recent empirical study finds that the high-quality representation aligns with the key token of the input text, uncovering the potential connection between representation space and vocabulary space. Inspired by the findings, we revisit the generative tasks and develop an unsupervised generative framework for TRL, Text2Token. The framework is based on the token target prediction task, utilizing carefully constructed target token distribution as supervisory signals. To construct the high-quality target token distribution, we analyze the token-alignment properties with advanced embedders and identify two essential categories of key tokens: (1) the meaningful tokens in the text and (2) semantically derived tokens beyond the text. Based on these insights, we propose two methods -- data-driven and model-derived -- to construct synthetic token targets from data or the LLM backbone. Experiments on the MTEB v2 benchmark demonstrate that Text2Token achieves performance competitive with the state-of-the-art embedder with unsupervised contrastive learning, LLM2Vec. Our analysis further shows that vocabulary and representation spaces optimize together and toward the optimum solution during training, providing new ideas and insights for future work.
☆ Breaking the Likelihood Trap: Consistent Generative Recommendation with Graph-structured Model
Reranking, as the final stage of recommender systems, demands real-time inference, accuracy, and diversity. It plays a crucial role in determining the final exposure, directly influencing user experience. Recently, generative reranking has gained increasing attention for its strong ability to model complex dependencies among items. However, most existing methods suffer from the "likelihood trap", where high-likelihood sequences are often perceived as low-quality by humans. These models tend to repeatedly recommend a set of high-frequency items, resulting in list homogeneity, thereby limiting user engagement. In this work, we propose Consistent Graph-structured Generative Recommendation (Congrats), a novel generative reranking framework. To break the likelihood trap, we introduce a novel graph-structured decoder that can capture diverse sequences along multiple paths. This design not only expands the decoding space to promote diversity, but also improves prediction accuracy by implicit item dependencies derived from vertex transitions. Furthermore, we design a differentiable cascade system that incorporates an evaluator, enabling the model to learn directly from user preferences as the training objective. Extensive offline experiments validate the superior performance of Congrats over state-of-the-art reranking methods. Moreover, Congrats has been evaluated on a large-scale video-sharing app, Kuaishou, with over 300 million daily active users, demonstrating that our approach significantly improves both recommendation quality and diversity, validating our effectiveness in practical industrial environments.
☆ Integrating Structure-Aware Attention and Knowledge Graphs in Explainable Recommendation Systems
This paper designs and implements an explainable recommendation model that integrates knowledge graphs with structure-aware attention mechanisms. The model is built on graph neural networks and incorporates a multi-hop neighbor aggregation strategy. By integrating the structural information of knowledge graphs and dynamically assigning importance to different neighbors through an attention mechanism, the model enhances its ability to capture implicit preference relationships. In the proposed method, users and items are embedded into a unified graph structure. Multi-level semantic paths are constructed based on entities and relations in the knowledge graph to extract richer contextual information. During the rating prediction phase, recommendations are generated through the interaction between user and target item representations. The model is optimized using a binary cross-entropy loss function. Experiments conducted on the Amazon Books dataset validate the superior performance of the proposed model across various evaluation metrics. The model also shows good convergence and stability. These results further demonstrate the effectiveness and practicality of structure-aware attention mechanisms in knowledge graph-enhanced recommendation.
☆ CardRewriter: Leveraging Knowledge Cards for Long-Tail Query Rewriting on Short-Video Platforms
Short-video platforms have rapidly become a new generation of information retrieval systems, where users formulate queries to access desired videos. However, user queries, especially long-tail ones, often suffer from spelling errors, incomplete phrasing, and ambiguous intent, resulting in mismatches between user expectations and retrieved results. While large language models (LLMs) have shown success in long-tail query rewriting within e-commerce, they struggle on short-video platforms, where proprietary content such as short videos, live streams, micro dramas, and user social networks falls outside their training distribution. To address this challenge, we introduce \textbf{CardRewriter}, an LLM-based framework that incorporates domain-specific knowledge to enhance long-tail query rewriting. For each query, our method aggregates multi-source knowledge relevant to the query and summarizes it into an informative and query-relevant knowledge card. This card then guides the LLM to better capture user intent and produce more effective query rewrites. We optimize CardRewriter using a two-stage training pipeline: supervised fine-tuning followed by group relative policy optimization, with a tailored reward system balancing query relevance and retrieval effectiveness. Offline experiments show that CardRewriter substantially improves rewriting quality for queries targeting proprietary content. Online A/B testing further confirms significant gains in long-view rate (LVR) and click-through rate (CTR), along with a notable reduction in initiative query reformulation rate (IQRR). Since September 2025, CardRewriter has been deployed on Kuaishou, one of China's largest short-video platforms, serving hundreds of millions of users daily.
☆ Beyond the limitation of a single query: Train your LLM for query expansion with Reinforcement Learning
Reasoning-augmented search agents, such as Search-R1, are trained to reason, search, and generate the final answer iteratively. Nevertheless, due to their limited capabilities in reasoning and search, their performance on multi-hop QA benchmarks remains far from satisfactory. To handle complex or compound queries, we train an LLM-based search agent with the native capability of query expansion through reinforcement learning. In each turn, our search agent proposes several query variants, which are searched simultaneously to cover more relevant information. Meanwhile, given limited post-training data and computing resources, it is very challenging for a search agent to master multiple tasks, including query generation, retrieved information understanding, and answer generation. Therefore, we propose incorporating a pre-trained squeezer model that helps the search agent understand the retrieved documents, allowing the search agent to focus on query generation for high retrieval recall. With the assistance of the squeezer model, we discover that even a small-scale 3B LLM can demonstrate a strong capability of query expansion and achieve state-of-the-art accuracy on the multi-hop QA benchmarks. To be specific, our experiments across seven question-answering benchmarks demonstrate that our method, named ExpandSearch, achieves an average improvement of 4.4% compared to state-of-the-art baselines, with strong gains on multi-hop reasoning tasks requiring diverse evidence aggregation.
♻ ☆ SkewRoute: Training-Free LLM Routing for Knowledge Graph Retrieval-Augmented Generation via Score Skewness of Retrieved Context
Large language models excel at many tasks but often incur high inference costs during deployment. To mitigate hallucination, many systems use a knowledge graph to enhance retrieval-augmented generation (KG-RAG). However, the large amount of retrieved knowledge contexts increase these inference costs further. A promising solution to balance performance and cost is LLM routing, which directs simple queries to smaller LLMs and complex ones to larger LLMs. However, no dedicated routing methods currently exist for RAG, and existing training-based routers face challenges scaling to this domain due to the need for extensive training data. We observe that the score distributions produced by the retrieval scorer strongly correlate with query difficulty. Based on this, we propose an extremely simple yet effective routing framework, the first specifically designed for KG-RAG that efficiently balances performance and cost in a plug-and-play manner. It delivers over 3x higher routing effectiveness while reducing runtime to less than 0.001x compared to existing methods. Our code is available at https://github.com/hrwang00/SkewRoute.
♻ ☆ Doc2SAR: A Synergistic Framework for High-Fidelity Extraction of Structure-Activity Relationships from Scientific Documents
Extracting molecular structure-activity relationships (SARs) from scientific literature and patents is essential for drug discovery and materials research. However, this task remains challenging due to heterogeneous document formats and limitations of existing methods. Specifically, rule-based approaches relying on rigid templates fail to generalize across diverse document layouts, while general-purpose multimodal large language models (MLLMs) lack sufficient accuracy and reliability for specialized tasks, such as layout detection and optical chemical structure recognition (OCSR). To address these challenges, we introduce DocSAR-200, a rigorously annotated benchmark of 200 scientific documents designed specifically for evaluating SAR extraction methods. Additionally, we propose Doc2SAR, a novel synergistic framework that integrates domain-specific tools with MLLMs enhanced via supervised fine-tuning (SFT). Extensive experiments demonstrate that Doc2SAR achieves state-of-the-art performance across various document types, significantly outperforming leading end-to-end baselines. Specifically, Doc2SAR attains an overall Table Recall of 80.78% on DocSAR-200, exceeding end2end GPT-4o by 51.48%. Furthermore, Doc2SAR demonstrates practical usability through efficient inference and is accompanied by a web app.
♻ ☆ Audio Does Matter: Importance-Aware Multi-Granularity Fusion for Video Moment Retrieval ACM MM 2025
Video Moment Retrieval (VMR) aims to retrieve a specific moment semantically related to the given query. To tackle this task, most existing VMR methods solely focus on the visual and textual modalities while neglecting the complementary but important audio modality. Although a few recent works try to tackle the joint audio-vision-text reasoning, they treat all modalities equally and simply embed them without fine-grained interaction for moment retrieval. These designs are counter-practical as: Not all audios are helpful for video moment retrieval, and the audio of some videos may be complete noise or background sound that is meaningless to the moment determination. To this end, we propose a novel Importance-aware Multi-Granularity fusion model (IMG), which learns to dynamically and selectively aggregate the audio-vision-text contexts for VMR. Specifically, after integrating the textual guidance with vision and audio separately, we first design a pseudo-label-supervised audio importance predictor that predicts the importance score of the audio, and accordingly assigns weights to mitigate the interference caused by noisy audio. Then, we design a multi-granularity audio fusion module that adaptively fuses audio and visual modalities at local-, event-, and global-level, fully capturing their complementary contexts. We further propose a cross-modal knowledge distillation strategy to address the challenge of missing audio modality during inference. To evaluate our method, we further construct a new VMR dataset, i.e., Charades-AudioMatter, where audio-related samples are manually selected and re-organized from the original Charades-STA to validate the model's capability in utilizing audio modality. Extensive experiments validate the effectiveness of our method, achieving state-of-the-art with audio-video fusion in VMR methods. Our code is available at https://github.com/HuiGuanLab/IMG.
comment: Accepted to ACM MM 2025
♻ ☆ A Comprehensive Survey on Retrieval Methods in Recommender Systems
In an era dominated by information overload, effective recommender systems are essential for managing the deluge of data across digital platforms. Multi-stage cascade ranking systems are widely used in the industry, with retrieval and ranking being two typical stages. Retrieval methods sift through vast candidates to filter out irrelevant items, while ranking methods prioritize these candidates to present the most relevant items to users. Unlike studies focusing on the ranking stage, this survey explores the critical yet often overlooked retrieval stage of recommender systems. To achieve precise and efficient personalized retrieval, we summarize existing work in three key areas: improving similarity computation between user and item, enhancing indexing mechanisms for efficient retrieval, and optimizing training methods of retrieval. We also provide a comprehensive set of benchmarking experiments on three public datasets. Furthermore, we highlight current industrial applications through a case study on retrieval practices at a specific company, covering the entire retrieval process and online serving, along with practical implications and challenges. By detailing the retrieval stage, which is fundamental for effective recommendation, this survey aims to bridge the existing knowledge gap and serve as a cornerstone for researchers interested in optimizing this critical component of cascade recommender systems.
comment: 41 pages
♻ ☆ Reliable Decision Making via Calibration Oriented Retrieval Augmented Generation NeurIPS 2025
Recently, Large Language Models (LLMs) have been increasingly used to support various decision-making tasks, assisting humans in making informed decisions. However, when LLMs confidently provide incorrect information, it can lead humans to make suboptimal decisions. To prevent LLMs from generating incorrect information on topics they are unsure of and to improve the accuracy of generated content, prior works have proposed Retrieval Augmented Generation (RAG), where external documents are referenced to generate responses. However, previous RAG methods focus only on retrieving documents most relevant to the input query, without specifically aiming to ensure that the human user's decisions are well-calibrated. To address this limitation, we propose a novel retrieval method called Calibrated Retrieval-Augmented Generation (CalibRAG), which ensures that decisions informed by RAG are well-calibrated. Then we empirically validate that CalibRAG improves calibration performance as well as accuracy, compared to other baselines across various datasets.
comment: Accepted by NeurIPS 2025
♻ ☆ pEBR: A Probabilistic Approach to Embedding Based Retrieval
Embedding-based retrieval aims to learn a shared semantic representation space for both queries and items, enabling efficient and effective item retrieval through approximate nearest neighbor (ANN) algorithms. In current industrial practice, retrieval systems typically retrieve a fixed number of items for each query. However, this fixed-size retrieval often results in insufficient recall for head queries and low precision for tail queries. This limitation largely stems from the dominance of frequentist approaches in loss function design, which fail to address this challenge in industry. In this paper, we propose a novel \textbf{p}robabilistic \textbf{E}mbedding-\textbf{B}ased \textbf{R}etrieval (\textbf{pEBR}) framework. Our method models the item distribution conditioned on each query, enabling the use of a dynamic cosine similarity threshold derived from the cumulative distribution function (CDF) of the probabilistic model. Experimental results demonstrate that pEBR significantly improves both retrieval precision and recall. Furthermore, ablation studies reveal that the probabilistic formulation effectively captures the inherent differences between head-to-tail queries.
Multimedia 4
☆ Exploration of Embodied Space Experience through Umbilical Interaction: A Grounded Theory Approach
This paper critiques the limits of human-centered design in HCI, proposing a shift toward Interface-Centered Design. Drawing on Hookway's philosophy of interfaces, phenomenology, and embodied interaction, we created Umbilink, an umbilical interaction device simulating a uterine environment with tactile sensors and rhythmic feedback to induce a pre-subjectivized state of sensory reduction. Participants' experiences were captured through semi-structured interviews and analyzed with grounded theory. Our contributions are: (1) introducing the novel interface type of Umbilical Interaction; (2) demonstrating the cognitive value of materialized interfaces in a human-interface-environment relation; (3) highlighting the design role of wearing rituals as liminal experiences. As a pilot study, this design suggests imaginative applications in healing, meditation, and sleep, while offering a speculative tool for future interface research.
comment: 10 pages, 2 figures
☆ SyncLipMAE: Contrastive Masked Pretraining for Audio-Visual Talking-Face Representation
We introduce SyncLipMAE, a self-supervised pretraining framework for talking-face video that learns synchronization-aware and transferable facial dynamics from unlabeled audio-visual streams. Our approach couples masked visual modeling with cross-modal contrastive alignment and employs three per-frame prompt tokens that explicitly encode the essential factors of a talking-face frame - identity, vocal motion (speech-synchronized facial dynamics), and ambient motion (audio-agnostic movements such as blinks and head pose). The contrastive objective uses time-aligned vocal-motion and audio tokens as positives and misaligned pairs as negatives, driving both modalities into a shared embedding space and yielding token-level audio-visual stream synchronization. After pretraining, the aligned audio tokens together with the visual prompt tokens (identity, vocal motion, ambient motion) form a unified interface for four disparate downstream settings: (i) audio-visual stream synchronization; (ii) facial emotion and head/face action recognition; (iii) visual speech recognition; and (iv) visual dubbing, for which we enable indistinguishable audio- or video-driven control within a single model. Across four task families that require distinct capabilities, SyncLipMAE achieves state-of-the-art results, underscoring the effectiveness of synchronization-aware, factorized self-supervised pretraining.
♻ ☆ Audio Does Matter: Importance-Aware Multi-Granularity Fusion for Video Moment Retrieval ACM MM 2025
Video Moment Retrieval (VMR) aims to retrieve a specific moment semantically related to the given query. To tackle this task, most existing VMR methods solely focus on the visual and textual modalities while neglecting the complementary but important audio modality. Although a few recent works try to tackle the joint audio-vision-text reasoning, they treat all modalities equally and simply embed them without fine-grained interaction for moment retrieval. These designs are counter-practical as: Not all audios are helpful for video moment retrieval, and the audio of some videos may be complete noise or background sound that is meaningless to the moment determination. To this end, we propose a novel Importance-aware Multi-Granularity fusion model (IMG), which learns to dynamically and selectively aggregate the audio-vision-text contexts for VMR. Specifically, after integrating the textual guidance with vision and audio separately, we first design a pseudo-label-supervised audio importance predictor that predicts the importance score of the audio, and accordingly assigns weights to mitigate the interference caused by noisy audio. Then, we design a multi-granularity audio fusion module that adaptively fuses audio and visual modalities at local-, event-, and global-level, fully capturing their complementary contexts. We further propose a cross-modal knowledge distillation strategy to address the challenge of missing audio modality during inference. To evaluate our method, we further construct a new VMR dataset, i.e., Charades-AudioMatter, where audio-related samples are manually selected and re-organized from the original Charades-STA to validate the model's capability in utilizing audio modality. Extensive experiments validate the effectiveness of our method, achieving state-of-the-art with audio-video fusion in VMR methods. Our code is available at https://github.com/HuiGuanLab/IMG.
comment: Accepted to ACM MM 2025
♻ ☆ Benchmarking and Bridging Emotion Conflicts for Multimodal Emotion Reasoning
Despite their strong performance in multimodal emotion reasoning, existing Multimodal Large Language Models (MLLMs) often overlook the scenarios involving emotion conflicts, where emotional cues from different modalities are inconsistent. To fill this gap, we first introduce CA-MER, a new benchmark designed to examine MLLMs under realistic emotion conflicts. It consists of three subsets: video-aligned, audio-aligned, and consistent, where only one or all modalities reflect the true emotion. However, evaluations on our CA-MER reveal that current state-of-the-art emotion MLLMs systematically over-rely on audio signal during emotion conflicts, neglecting critical cues from visual modality. To mitigate this bias, we propose MoSEAR, a parameter-efficient framework that promotes balanced modality integration. MoSEAR consists of two modules: (1)MoSE, modality-specific experts with a regularized gating mechanism that reduces modality bias in the fine-tuning heads; and (2)AR, an attention reallocation mechanism that rebalances modality contributions in frozen backbones during inference. Our framework offers two key advantages: it mitigates emotion conflicts and improves performance on consistent samples-without incurring a trade-off between audio and visual modalities. Experiments on multiple benchmarks-including MER2023, EMER, DFEW, and our CA-MER-demonstrate that MoSEAR achieves state-of-the-art performance, particularly under modality conflict conditions.
comment: ACM Multimedia 2025 Oral Code: https://github.com/ZhiyuanHan-Aaron/MoSEAR Project Page: https://zhiyuanhan-aaron.github.io/MoSEAR-page/
Information Retrieval 23
☆ PairSem: LLM-Guided Pairwise Semantic Matching for Scientific Document Retrieval
Scientific document retrieval is a critical task for enabling knowledge discovery and supporting research across diverse domains. However, existing dense retrieval methods often struggle to capture fine-grained scientific concepts in texts due to their reliance on holistic embeddings and limited domain understanding. Recent approaches leverage large language models (LLMs) to extract fine-grained semantic entities and enhance semantic matching, but they typically treat entities as independent fragments, overlooking the multi-faceted nature of scientific concepts. To address this limitation, we propose Pairwise Semantic Matching (PairSem), a framework that represents relevant semantics as entity-aspect pairs, capturing complex, multi-faceted scientific concepts. PairSem is unsupervised, base retriever-agnostic, and plug-and-play, enabling precise and context-aware matching without requiring query-document labels or entity annotations. Extensive experiments on multiple datasets and retrievers demonstrate that PairSem significantly improves retrieval performance, highlighting the importance of modeling multi-aspect semantics in scientific information retrieval.
☆ MTMD: A Multi-Task Multi-Domain Framework for Unified Ad Lightweight Ranking at Pinterest KDD 2025
The lightweight ad ranking layer, living after the retrieval stage and before the fine ranker, plays a critical role in the success of a cascaded ad recommendation system. Due to the fact that there are multiple optimization tasks depending on the ad domain, e.g., Click Through Rate (CTR) for click ads and Conversion Rate (CVR) for conversion ads, as well as multiple surfaces where an ad is served (home feed, search, or related item recommendation) with diverse ad products (shopping or standard ad); it is an essentially challenging problem in industry on how to do joint holistic optimization in the lightweight ranker, such that the overall platform's value, advertiser's value, and user's value are maximized. Deep Neural Network (DNN)-based multitask learning (MTL) can handle multiple goals naturally, with each prediction head mapping to a particular optimization goal. However, in practice, it is unclear how to unify data from different surfaces and ad products into a single model. It is critical to learn domain-specialized knowledge and explicitly transfer knowledge between domains to make MTL effective. We present a Multi-Task Multi-Domain (MTMD) architecture under the classic Two-Tower paradigm, with the following key contributions: 1) handle different prediction tasks, ad products, and ad serving surfaces in a unified framework; 2) propose a novel mixture-of-expert architecture to learn both specialized knowledge each domain and common knowledge shared between domains; 3) propose a domain adaption module to encourage knowledge transfer between experts; 4) constrain the modeling of different prediction tasks. MTMD improves the offline loss value by 12% to 36%, mapping to 2% online reduction in cost per click. We have deployed this single MTMD framework into production for Pinterest ad recommendation replacing 9 production models.
comment: AdKDD 2025
☆ MRMR: A Realistic and Expert-Level Multidisciplinary Benchmark for Reasoning-Intensive Multimodal Retrieval
We introduce MRMR, the first expert-level multidisciplinary multimodal retrieval benchmark requiring intensive reasoning. MRMR contains 1,502 queries spanning 23 domains, with positive documents carefully verified by human experts. Compared to prior benchmarks, MRMR introduces three key advancements. First, it challenges retrieval systems across diverse areas of expertise, enabling fine-grained model comparison across domains. Second, queries are reasoning-intensive, with images requiring deeper interpretation such as diagnosing microscopic slides. We further introduce Contradiction Retrieval, a novel task requiring models to identify conflicting concepts. Finally, queries and documents are constructed as image-text interleaved sequences. Unlike earlier benchmarks restricted to single images or unimodal documents, MRMR offers a realistic setting with multi-image queries and mixed-modality corpus documents. We conduct an extensive evaluation of 4 categories of multimodal retrieval systems and 14 frontier models on MRMR. The text embedding model Qwen3-Embedding with LLM-generated image captions achieves the highest performance, highlighting substantial room for improving multimodal retrieval models. Although latest multimodal models such as Ops-MM-Embedding perform competitively on expert-domain queries, they fall short on reasoning-intensive tasks. We believe that MRMR paves the way for advancing multimodal retrieval in more realistic and challenging scenarios.
☆ Cross-attention Secretly Performs Orthogonal Alignment in Recommendation Models
Cross-domain sequential recommendation (CDSR) aims to align heterogeneous user behavior sequences collected from different domains. While cross-attention is widely used to enhance alignment and improve recommendation performance, its underlying mechanism is not fully understood. Most researchers interpret cross-attention as residual alignment, where the output is generated by removing redundant and preserving non-redundant information from the query input by referencing another domain data which is input key and value. Beyond the prevailing view, we introduce Orthogonal Alignment, a phenomenon in which cross-attention discovers novel information that is not present in the query input, and further argue that those two contrasting alignment mechanisms can co-exist in recommendation models We find that when the query input and output of cross-attention are orthogonal, model performance improves over 300 experiments. Notably, Orthogonal Alignment emerges naturally, without any explicit orthogonality constraints. Our key insight is that Orthogonal Alignment emerges naturally because it improves scaling law. We show that baselines additionally incorporating cross-attention module outperform parameter-matched baselines, achieving a superior accuracy-per-model parameter. We hope these findings offer new directions for parameter-efficient scaling in multi-modal research.
comment: 19 pages
☆ Hierarchical Semantic RL: Tackling the Problem of Dynamic Action Space for RL-based Recommendations
Recommender Systems (RS) are fundamental to modern online services. While most existing approaches optimize for short-term engagement, recent work has begun to explore reinforcement learning (RL) to model long-term user value. However, these efforts face significant challenges due to the vast, dynamic action spaces inherent in recommendation, which hinder stable policy learning. To resolve this bottleneck, we introduce Hierarchical Semantic RL (HSRL), which reframes RL-based recommendation over a fixed Semantic Action Space (SAS). HSRL encodes items as Semantic IDs (SIDs) for policy learning, and maps SIDs back to their original items via a fixed, invertible lookup during execution. To align decision-making with SID generation, the Hierarchical Policy Network (HPN) operates in a coarse-to-fine manner, employing hierarchical residual state modeling to refine each level's context from the previous level's residual, thereby stabilizing training and reducing representation-decision mismatch. In parallel, a Multi-level Critic (MLC) provides token-level value estimates, enabling fine-grained credit assignment. Across public benchmarks and a large-scale production dataset from a leading Chinese short-video advertising platform, HSRL consistently surpasses state-of-the-art baselines. In online deployment over a seven-day A/B testing, it delivers an 18.421% CVR lift with only a 1.251% increase in cost, supporting HSRL as a scalable paradigm for RL-based recommendation. Our code is released at https://github.com/MinmaoWang/HSRL.
☆ Controlled Personalization in Legacy Media Online Services: A Case Study in News Recommendation
Personalized news recommendations have become a standard feature of large news aggregation services, optimizing user engagement through automated content selection. In contrast, legacy news media often approach personalization cautiously, striving to balance technological innovation with core editorial values. As a result, online platforms of traditional news outlets typically combine editorially curated content with algorithmically selected articles - a strategy we term controlled personalization. In this industry paper, we evaluate the effectiveness of controlled personalization through an A/B test conducted on the website of a major Norwegian legacy news organization. Our findings indicate that even a modest level of personalization yields substantial benefits. Specifically, we observe that users exposed to personalized content demonstrate higher click-through rates and reduced navigation effort, suggesting improved discovery of relevant content. Moreover, our analysis reveals that controlled personalization contributes to greater content diversity and catalog coverage and in addition reduces popularity bias. Overall, our results suggest that controlled personalization can successfully align user needs with editorial goals, offering a viable path for legacy media to adopt personalization technologies while upholding journalistic values.
☆ Generative Data Augmentation in Graph Contrastive Learning for Recommendation
Recommendation systems have become indispensable in various online platforms, from e-commerce to streaming services. A fundamental challenge in this domain is learning effective embeddings from sparse user-item interactions. While contrastive learning has recently emerged as a promising solution to this issue, generating augmented views for contrastive learning through most existing random data augmentation methods often leads to the alteration of original semantic information. In this paper, we propose a novel framework, GDA4Rec (Generative Data Augmentation in graph contrastive learning for Recommendation) to generate high-quality augmented views and provide robust self-supervised signals. Specifically, we employ a noise generation module that leverages deep generative models to approximate the distribution of original data for data augmentation. Additionally, GDA4Rec further extracts an item complement matrix to characterize the latent correlations between items and provide additional self-supervised signals. Lastly, a joint objective that integrates recommendation, data augmentation and contrastive learning is used to enforce the model to learn more effective and informative embeddings. Extensive experiments are conducted on three public datasets to demonstrate the superiority of the model. The code is available at: https://github.com/MrYansong/GDA4Rec.
comment: The 34th ACM International Conference on Information and Knowledge Management
☆ Cost-Efficient Long Code Translation using LLMs while Leveraging Identifier Replacements
In the domain of software development, LLMs have been utilized to automate tasks such as code translation, where source code from one programming language is translated to another while preserving its functionality. However, LLMs often struggle with long source codes that don't fit into the context window, which produces inaccurate translations. To address this, we propose a novel zero-shot code translation method that incorporates identifier replacement. By substituting user-given long identifiers with generalized placeholders during translation, our method allows the LLM to focus on the logical structure of the code, by reducing token count and memory usage, which improves the efficiency and cost-effectiveness of long code translation. Our empirical results demonstrate that our approach preserves syntactical and hierarchical information and produces translation results with reduced tokens.
☆ Rethinking Reasoning in Document Ranking: Why Chain-of-Thought Falls Short
Document reranking is a key component in information retrieval (IR), aimed at refining initial retrieval results to improve ranking quality for downstream tasks. Recent studies--motivated by large reasoning models (LRMs)--have begun incorporating explicit chain-of-thought (CoT) reasoning into LLM-based rerankers. However, the effectiveness of such reasoning for ranking tasks remains underexplored. In this work, we present the first systematic study of reasoning in reranking across both pointwise and listwise settings, under both supervised fine-tuning and reinforcement learning. Using diverse benchmarks, including reasoning-intensive datasets (BRIGHT) and standard IR benchmarks (BEIR), we find that reasoning-augmented rerankers consistently underperform their direct counterparts that predict rankings without CoT, despite substantially higher inference costs. Our analysis reveals three core limitations: (i) in pointwise rerankers, reasoning breaks calibration and biases models toward the positive class, raising TPR but lowering TNR, which inflates false positives and degrades ranking in negative-dominant pools; (ii) in listwise rerankers, reasoning improves in-domain fit but increases variance and fails to generalize out-of-domain, even when reinforcement learning shortens rationales; and (iii) overall, directly fine-tuned rerankers remain more stable, effective, and robust. These findings challenge the assumption that explicit reasoning is universally beneficial for reranking. We conclude by highlighting future directions, including calibration-aware scoring for pointwise rerankers and the design of concise, targeted reasoning strategies to mitigate overfitting and overthinking in listwise rerankers.
☆ Hierarchical Scheduling for Multi-Vector Image Retrieval
To effectively leverage user-specific data, retrieval augmented generation (RAG) is employed in multimodal large language model (MLLM) applications. However, conventional retrieval approaches often suffer from limited retrieval accuracy. Recent advances in multi-vector retrieval (MVR) improve accuracy by decomposing queries and matching against segmented images. They still suffer from sub-optimal accuracy and efficiency, overlooking alignment between the query and varying image objects and redundant fine-grained image segments. In this work, we present an efficient scheduling framework for image retrieval - HiMIR. First, we introduce a novel hierarchical paradigm, employing multiple intermediate granularities for varying image objects to enhance alignment. Second, we minimize redundancy in retrieval by leveraging cross-hierarchy similarity consistency and hierarchy sparsity to minimize unnecessary matching computation. Furthermore, we configure parameters for each dataset automatically for practicality across diverse scenarios. Our empirical study shows that, HiMIR not only achieves substantial accuracy improvements but also reduces computation by up to 3.5 times over the existing MVR system.
comment: Under Review
☆ EcphoryRAG: Re-Imagining Knowledge-Graph RAG via Human Associative Memory
Cognitive neuroscience research indicates that humans leverage cues to activate entity-centered memory traces (engrams) for complex, multi-hop recollection. Inspired by this mechanism, we introduce EcphoryRAG, an entity-centric knowledge graph RAG framework. During indexing, EcphoryRAG extracts and stores only core entities with corresponding metadata, a lightweight approach that reduces token consumption by up to 94\% compared to other structured RAG systems. For retrieval, the system first extracts cue entities from queries, then performs a scalable multi-hop associative search across the knowledge graph. Crucially, EcphoryRAG dynamically infers implicit relations between entities to populate context, enabling deep reasoning without exhaustive pre-enumeration of relationships. Extensive evaluations on the 2WikiMultiHop, HotpotQA, and MuSiQue benchmarks demonstrate that EcphoryRAG sets a new state-of-the-art, improving the average Exact Match (EM) score from 0.392 to 0.474 over strong KG-RAG methods like HippoRAG. These results validate the efficacy of the entity-cue-multi-hop retrieval paradigm for complex question answering.
☆ Personalize Before Retrieve: LLM-based Personalized Query Expansion for User-Centric Retrieval
Retrieval-Augmented Generation (RAG) critically depends on effective query expansion to retrieve relevant information. However, existing expansion methods adopt uniform strategies that overlook user-specific semantics, ignoring individual expression styles, preferences, and historical context. In practice, identical queries in text can express vastly different intentions across users. This representational rigidity limits the ability of current RAG systems to generalize effectively in personalized settings. Specifically, we identify two core challenges for personalization: 1) user expression styles are inherently diverse, making it difficult for standard expansions to preserve personalized intent. 2) user corpora induce heterogeneous semantic structures-varying in topical focus and lexical organization-which hinders the effective anchoring of expanded queries within the user's corpora space. To address these challenges, we propose Personalize Before Retrieve (PBR), a framework that incorporates user-specific signals into query expansion prior to retrieval. PBR consists of two components: P-PRF, which generates stylistically aligned pseudo feedback using user history for simulating user expression style, and P-Anchor, which performs graph-based structure alignment over user corpora to capture its structure. Together, they produce personalized query representations tailored for retrieval. Experiments on two personalized benchmarks show that PBR consistently outperforms strong baselines, with up to 10% gains on PersonaBench across retrievers. Our findings demonstrate the value of modeling personalization before retrieval to close the semantic gap in user-adaptive RAG systems. Our code is available at https://github.com/Zhang-Yingyi/PBR-code.
☆ MATT-CTR: Unleashing a Model-Agnostic Test-Time Paradigm for CTR Prediction with Confidence-Guided Inference Paths
Recently, a growing body of research has focused on either optimizing CTR model architectures to better model feature interactions or refining training objectives to aid parameter learning, thereby achieving better predictive performance. However, previous efforts have primarily focused on the training phase, largely neglecting opportunities for optimization during the inference phase. Infrequently occurring feature combinations, in particular, can degrade prediction performance, leading to unreliable or low-confidence outputs. To unlock the predictive potential of trained CTR models, we propose a Model-Agnostic Test-Time paradigm (MATT), which leverages the confidence scores of feature combinations to guide the generation of multiple inference paths, thereby mitigating the influence of low-confidence features on the final prediction. Specifically, to quantify the confidence of feature combinations, we introduce a hierarchical probabilistic hashing method to estimate the occurrence frequencies of feature combinations at various orders, which serve as their corresponding confidence scores. Then, using the confidence scores as sampling probabilities, we generate multiple instance-specific inference paths through iterative sampling and subsequently aggregate the prediction scores from multiple paths to conduct robust predictions. Finally, extensive offline experiments and online A/B tests strongly validate the compatibility and effectiveness of MATT across existing CTR models.
comment: 10 pages, 4 figures, 2 tables
☆ Observation Matrix Design for Densifying MIMO Channel Estimation via 2D Ice Filling
In recent years, densifying multiple-input multiple-output (MIMO) has attracted much attention from the communication community. Thanks to the subwavelength antenna spacing, the strong correlations among densifying antennas provide sufficient prior knowledge about channel state information (CSI). This inspires the careful design of observation matrices (e.g., transmit precoders and receive combiners), that exploits the CSI prior knowledge, to boost channel estimation performance. Aligned with this vision, this work proposes to jointly design the combiners and precoders by maximizing the mutual information between the received pilots and densifying MIMO channels. A two-dimensional ice-filling (2DIF) algorithm is proposed to efficiently accomplish this objective. The algorithm is motivated by the fact that the eigenspace of MIMO channel covariance can be decoupled into two sub-eigenspaces, which are associated with the correlations of transmitter antennas and receiver antennas, respectively. By properly setting the precoder and the combiner as the eigenvectors from these two sub-eigenspaces, the 2DIF promises to generate near-optimal observation matrices. Moreover, we further extend the 2DIF method to the popular hybrid combining systems, where a two-stage 2DIF (TS-2DIF) algorithm is developed to handle the analog combining circuits realized by phase shifters. Simulation results demonstrate that, compared to the state-of-the-art schemes, the proposed 2DIF and TS-2DIF methods can achieve superior channel estimation accuracy.
comment: 17 pages, 8 figures
☆ FinAuditing: A Financial Taxonomy-Structured Multi-Document Benchmark for Evaluating LLMs
The complexity of the Generally Accepted Accounting Principles (GAAP) and the hierarchical structure of eXtensible Business Reporting Language (XBRL) filings make financial auditing increasingly difficult to automate and verify. While large language models (LLMs) have demonstrated strong capabilities in unstructured text understanding, their ability to reason over structured, interdependent, and taxonomy-driven financial documents remains largely unexplored. To fill this gap, we introduce FinAuditing, the first taxonomy-aligned, structure-aware, multi-document benchmark for evaluating LLMs on financial auditing tasks. Built from real US-GAAP-compliant XBRL filings, FinAuditing defines three complementary subtasks, FinSM for semantic consistency, FinRE for relational consistency, and FinMR for numerical consistency, each targeting a distinct aspect of structured auditing reasoning. We further propose a unified evaluation framework integrating retrieval, classification, and reasoning metrics across these subtasks. Extensive zero-shot experiments on 13 state-of-the-art LLMs reveal that current models perform inconsistently across semantic, relational, and mathematical dimensions, with accuracy drops of up to 60-90% when reasoning over hierarchical multi-document structures. Our findings expose the systematic limitations of modern LLMs in taxonomy-grounded financial reasoning and establish FinAuditing as a foundation for developing trustworthy, structure-aware, and regulation-aligned financial intelligence systems. The benchmark dataset is available at Hugging Face.
♻ ☆ Preference Discerning with LLM-Enhanced Generative Retrieval
In sequential recommendation, models recommend items based on user's interaction history. To this end, current models usually incorporate information such as item descriptions and user intent or preferences. User preferences are usually not explicitly given in open-source datasets, and thus need to be approximated, for example via large language models (LLMs). Current approaches leverage approximated user preferences only during training and rely solely on the past interaction history for recommendations, limiting their ability to dynamically adapt to changing preferences, potentially reinforcing echo chambers. To address this issue, we propose a new paradigm, namely preference discerning, which explicitly conditions a generative recommendation model on user preferences in natural language within its context. To evaluate preference discerning, we introduce a novel benchmark that provides a holistic evaluation across various scenarios, including preference steering and sentiment following. Upon evaluating current state-of-the-art methods on our benchmark, we discover that their ability to dynamically adapt to evolving user preferences is limited. To address this, we propose a new method named Mender ($\textbf{M}$ultimodal Prefer$\textbf{en}$ce $\textbf{D}$iscern$\textbf{er}$), which achieves state-of-the-art performance in our benchmark. Our results show that Mender effectively adapts its recommendation guided by human preferences, even if not observed during training, paving the way toward more flexible recommendation models.
comment: Accepted at TMLR, Code available at https://github.com/facebookresearch/preference_discerning
♻ ☆ From Entity Reliability to Clean Feedback: An Entity-Aware Denoising Framework Beyond Interaction-Level Signals
Implicit feedback is central to modern recommender systems but is inherently noisy, often impairing model training and degrading user experience. At scale, such noise can mislead learning processes, reducing both recommendation accuracy and platform value. Existing denoising strategies typically overlook the entity-specific nature of noise while introducing high computational costs and complex hyperparameter tuning. To address these challenges, we propose \textbf{EARD} (\textbf{E}ntity-\textbf{A}ware \textbf{R}eliability-\textbf{D}riven Denoising), a lightweight framework that shifts the focus from interaction-level signals to entity-level reliability. Motivated by the empirical observation that training loss correlates with noise, EARD quantifies user and item reliability via their average training losses as a proxy for reputation, and integrates these entity-level factors with interaction-level confidence. The framework is \textbf{model-agnostic}, \textbf{computationally efficient}, and requires \textbf{only two intuitive hyperparameters}. Extensive experiments across multiple datasets and backbone models demonstrate that EARD yields substantial improvements over state-of-the-art baselines (e.g., up to 27.01\% gain in NDCG@50), while incurring negligible additional computational cost. Comprehensive ablation studies and mechanism analyses further confirm EARD's robustness to hyperparameter choices and its practical scalability. These results highlight the importance of entity-aware reliability modeling for denoising implicit feedback and pave the way for more robust recommendation research.
♻ ☆ Understanding and Improving Information Preservation in Prompt Compression for LLMs EMNLP 2025
Recent advancements in large language models (LLMs) have enabled their successful application to a broad range of tasks. However, in information-intensive tasks, the prompt length can grow fast, leading to increased computational requirements, performance degradation, and induced biases from irrelevant or redundant information. Recently, various prompt compression techniques have been introduced to optimize the trade-off between reducing input length and retaining performance. We propose a holistic evaluation framework that allows for in-depth analysis of prompt compression methods. We focus on three key aspects, besides compression ratio: (i) downstream task performance, (ii) grounding in the input context, and (iii) information preservation. Using our framework, we analyze state-of-the-art soft and hard compression methods and show that some fail to preserve key details from the original prompt, limiting performance on complex tasks. By identifying these limitations, we are able to improve one soft prompting method by controlling compression granularity, achieving up to +23% in downstream performance, +8 BERTScore points in grounding, and 2.7x more entities preserved in compression. Ultimately, we find that the best effectiveness/compression rate trade-off is achieved with soft prompting combined with sequence-level training.The code is available at https://github.com/amazon-science/information-preservation-in-prompt-compression.
comment: Accepted to EMNLP 2025 (Findings), 22 pages, 6 figures, 24 tables
♻ ☆ SUMMA: A Multimodal Large Language Model for Advertisement Summarization
Understanding multimodal video ads is crucial for improving query-ad matching and relevance ranking on short video platforms, enhancing advertising effectiveness and user experience. However, the effective utilization of multimodal information with high commercial value still largely constrained by reliance on highly compressed video embeddings-has long been inadequate. To address this, we propose SUMMA (the abbreviation of Summarizing MultiModal Ads), a multimodal model that automatically processes video ads into summaries highlighting the content of highest commercial value, thus improving their comprehension and ranking in Douyin search-advertising systems. SUMMA is developed via a two-stage training strategy-multimodal supervised fine-tuning followed by reinforcement learning with a mixed reward mechanism-on domain-specific data containing video frames and ASR/OCR transcripts, generating commercially valuable and explainable summaries. We integrate SUMMA-generated summaries into our production pipeline, directly enhancing the candidate retrieval and relevance ranking stages in real search-advertising systems. Both offline and online experiments show substantial improvements over baselines, with online results indicating a statistically significant 1.5% increase in advertising revenue. Our work establishes a novel paradigm for condensing multimodal information into representative texts, effectively aligning visual ad content with user query intent in retrieval and recommendation scenarios.
♻ ☆ Chain-of-Retrieval Augmented Generation NeurIPS 2025
This paper introduces an approach for training o1-like RAG models that retrieve and reason over relevant information step by step before generating the final answer. Conventional RAG methods usually perform a single retrieval step before the generation process, which limits their effectiveness in addressing complex queries due to imperfect retrieval results. In contrast, our proposed method, CoRAG (Chain-of-Retrieval Augmented Generation), allows the model to dynamically reformulate the query based on the evolving state. To train CoRAG effectively, we utilize rejection sampling to automatically generate intermediate retrieval chains, thereby augmenting existing RAG datasets that only provide the correct final answer. At test time, we propose various decoding strategies to scale the model's test-time compute by controlling the length and number of sampled retrieval chains. Experimental results across multiple benchmarks validate the efficacy of CoRAG, particularly in multi-hop question answering tasks, where we observe more than 10 points improvement in EM score compared to strong baselines. On the KILT benchmark, CoRAG establishes a new state-of-the-art performance across a diverse range of knowledge-intensive tasks. Furthermore, we offer comprehensive analyses to understand the scaling behavior of CoRAG, laying the groundwork for future research aimed at developing factual and grounded foundation models.
comment: Accepted by NeurIPS 2025
♻ ☆ Diffusion Generative Recommendation with Continuous Tokens WWW 2026
Recent advances in generative artificial intelligence, particularly large language models (LLMs), have opened new opportunities for enhancing recommender systems (RecSys). Most existing LLM-based RecSys approaches operate in a discrete space, using vector-quantized tokenizers to align with the inherent discrete nature of language models. However, these quantization methods often result in lossy tokenization and suboptimal learning, primarily due to inaccurate gradient propagation caused by the non-differentiable argmin operation in standard vector quantization. Inspired by the emerging trend of embracing continuous tokens in language models, we propose ContRec, a novel framework that seamlessly integrates continuous tokens into LLM-based RecSys. Specifically, ContRec consists of two key modules: a sigma-VAE Tokenizer, which encodes users/items with continuous tokens; and a Dispersive Diffusion module, which captures implicit user preference. The tokenizer is trained with a continuous Variational Auto-Encoder (VAE) objective, where three effective techniques are adopted to avoid representation collapse. By conditioning on the previously generated tokens of the LLM backbone during user modeling, the Dispersive Diffusion module performs a conditional diffusion process with a novel Dispersive Loss, enabling high-quality user preference generation through next-token diffusion. Finally, ContRec leverages both the textual reasoning output from the LLM and the latent representations produced by the diffusion model for Top-K item retrieval, thereby delivering comprehensive recommendation results. Extensive experiments on four datasets demonstrate that \ourname{} consistently outperforms both traditional and SOTA LLM-based recommender systems. Our results highlight the potential of continuous tokenization and generative modeling for advancing the next generation of recommender systems.
comment: Submitted to WWW 2026. Our code and data will be made publicly available after acceptance
♻ ☆ Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation
Modern long-context large language models (LLMs) perform well on synthetic "needle-in-a-haystack" (NIAH) benchmarks, but such tests overlook how noisy contexts arise from biased retrieval and agentic workflows. We argue that haystack engineering is necessary to construct noisy long contexts that faithfully capture key real-world factors -- distraction from heterogeneous biased retrievers and cascading errors in agentic workflows -- to test models' long-context robustness. We instantiate it through HaystackCraft, a new NIAH benchmark built on the full English Wikipedia hyperlink network with multi-hop questions. HaystackCraft evaluates how heterogeneous retrieval strategies (e.g., sparse, dense, hybrid, and graph-based) affect distractor composition, haystack ordering, and downstream LLM performance. HaystackCraft further extends NIAH to dynamic, LLM-dependent settings that simulate agentic operations, where models refine queries, reflect on their past reasonings, and decide when to stop. Experiments with 15 long-context models show that (1) while stronger dense retrievers can introduce more challenging distractors, graph-based reranking simultaneously improves retrieval effectiveness and mitigates more harmful distractors; (2) in agentic tests, even advanced models like Gemini 2.5 Pro and GPT-5 suffer cascading failures from self-generated distractors or struggle to perform early stops. These results highlight persistent challenges in agentic long-context reasoning and establish HaystackCraft as a valuable testbed for future progress.
comment: Code available at https://github.com/Graph-COM/HaystackCraft
♻ ☆ What Makes LLMs Effective Sequential Recommenders? A Study on Preference Intensity and Temporal Context
Sequential recommendation systems aspire to profile users by interpreting their interaction histories, echoing how humans make decisions by weighing experience, relative preference strength, and situational relevance. Yet, existing large language model (LLM)-based recommenders often fall short of mimicking the flexible, context-aware decision strategies humans exhibit, neglecting the structured, dynamic, and context-aware mechanisms fundamental to human behaviors. To bridge this gap, we propose RecPO, a preference optimization framework that models structured feedback and contextual delay to emulate human-like prioritization in sequential recommendation. RecPO exploits adaptive reward margins based on inferred preference hierarchies and temporal signals, enabling the model to favor immediately relevant items and to distinguish between varying degrees of preference and aversion. Extensive experiments across five real-world datasets demonstrate that RecPO not only yields performance gains over state-of-the-art baselines, but also mirrors key characteristics of human decision-making: favoring timely satisfaction, maintaining coherent preferences, and exercising discernment under shifting contexts.
Multimedia 4
☆ Zero-shot image privacy classification with Vision-Language Models ICASSP 2026
While specialized learning-based models have historically dominated image privacy prediction, the current literature increasingly favours adopting large Vision-Language Models (VLMs) designed for generic tasks. This trend risks overlooking the performance ceiling set by purpose-built models due to a lack of systematic evaluation. To address this problem, we establish a zero-shot benchmark for image privacy classification, enabling a fair comparison. We evaluate the top-3 open-source VLMs, according to a privacy benchmark, using task-aligned prompts and we contrast their performance, efficiency, and robustness against established vision-only and multi-modal methods. Counter-intuitively, our results show that VLMs, despite their resource-intensive nature in terms of high parameter count and slower inference, currently lag behind specialized, smaller models in privacy prediction accuracy. We also find that VLMs exhibit higher robustness to image perturbations.
comment: 5 pages, 3 figures, 3 tables. This work has been submitted to the ICASSP 2026
☆ SeeingSounds: Learning Audio-to-Visual Alignment via Text
We introduce SeeingSounds, a lightweight and modular framework for audio-to-image generation that leverages the interplay between audio, language, and vision-without requiring any paired audio-visual data or training on visual generative models. Rather than treating audio as a substitute for text or relying solely on audio-to-text mappings, our method performs dual alignment: audio is projected into a semantic language space via a frozen language encoder, and, contextually grounded into the visual domain using a vision-language model. This approach, inspired by cognitive neuroscience, reflects the natural cross-modal associations observed in human perception. The model operates on frozen diffusion backbones and trains only lightweight adapters, enabling efficient and scalable learning. Moreover, it supports fine-grained and interpretable control through procedural text prompt generation, where audio transformations (e.g., volume or pitch shifts) translate into descriptive prompts (e.g., "a distant thunder") that guide visual outputs. Extensive experiments across standard benchmarks confirm that SeeingSounds outperforms existing methods in both zero-shot and supervised settings, establishing a new state of the art in controllable audio-to-visual generation.
comment: accepted to ACM Multimedia Asia 2025
♻ ☆ Iola Walker: A Mobile Footfall Detection System for Music Composition
This outing is part of a larger music technology research project. The objective is to find a method for materially enhancing music using hardware and software. There is a strong likelihood that there exists a new medium for experiencing music via a wearable device that ordinary listeners prefer over the current state of the art. If such a medium is discovered, it is a step towards altruistic, prosocial reform in the music industry. A new playback system infrastructure has a chance to soothe some of the societal problems tied to the larger entertainment industry ecosystem. Iola walker is a music playback system that allows musicians to compose music that changes in accordance with the listener's gait. Artifacts are available here: https://github.com/willbjames/iolawalker
♻ ☆ AniME: Adaptive Multi-Agent Planning for Long Animation Generation
We present AniME, a director-oriented multi-agent system for automated long-form anime production, covering the full workflow from a story to the final video. The director agent keeps a global memory for the whole workflow, and coordinates several downstream specialized agents. By integrating customized Model Context Protocol (MCP) with downstream model instruction, the specialized agent adaptively selects control conditions for diverse sub-tasks. AniME produces cinematic animation with consistent characters and synchronized audio visual elements, offering a scalable solution for AI-driven anime creation.
comment: 2 pages, Technical Report
Computation and Language 151
☆ ArenaBencher: Automatic Benchmark Evolution via Multi-Model Competitive Evaluation
Benchmarks are central to measuring the capabilities of large language models and guiding model development, yet widespread data leakage from pretraining corpora undermines their validity. Models can match memorized content rather than demonstrate true generalization, which inflates scores, distorts cross-model comparisons, and misrepresents progress. We introduce ArenaBencher, a model-agnostic framework for automatic benchmark evolution that updates test cases while preserving comparability. Given an existing benchmark and a diverse pool of models to be evaluated, ArenaBencher infers the core ability of each test case, generates candidate question-answer pairs that preserve the original objective, verifies correctness and intent with an LLM as a judge, and aggregates feedback from multiple models to select candidates that expose shared weaknesses. The process runs iteratively with in-context demonstrations that steer generation toward more challenging and diagnostic cases. We apply ArenaBencher to math problem solving, commonsense reasoning, and safety domains and show that it produces verified, diverse, and fair updates that uncover new failure modes, increase difficulty while preserving test objective alignment, and improve model separability. The framework provides a scalable path to continuously evolve benchmarks in step with the rapid progress of foundation models.
comment: Preprint
☆ MATRIX: Multimodal Agent Tuning for Robust Tool-Use Reasoning
Vision language models (VLMs) are increasingly deployed as controllers with access to external tools for complex reasoning and decision-making, yet their effectiveness remains limited by the scarcity of high-quality multimodal trajectories and the cost of manual annotation. We address this challenge with a vision-centric agent tuning framework that automatically synthesizes multimodal trajectories, generates step-wise preference pairs, and trains a VLM controller for robust tool-use reasoning. Our pipeline first constructs M-TRACE, a large-scale dataset of 28.5K multimodal tasks with 177K verified trajectories, enabling imitation-based trajectory tuning. Building on this, we develop MATRIX Agent, a controller finetuned on M-TRACE for step-wise tool reasoning. To achieve finer alignment, we further introduce Pref-X, a set of 11K automatically generated preference pairs, and optimize MATRIX on it via step-wise preference learning. Across three benchmarks, Agent-X, GTA, and GAIA, MATRIX consistently surpasses both open- and closed-source VLMs, demonstrating scalable and effective multimodal tool use. Our data and code is avaliable at https://github.com/mbzuai-oryx/MATRIX.
Agent Learning via Early Experience
A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.
comment: Work in progress
☆ VideoNorms: Benchmarking Cultural Awareness of Video Language Models
As Video Large Language Models (VideoLLMs) are deployed globally, they require understanding of and grounding in the relevant cultural background. To properly assess these models' cultural awareness, adequate benchmarks are needed. We introduce VideoNorms, a benchmark of over 1000 (video clip, norm) pairs from US and Chinese cultures annotated with socio-cultural norms grounded in speech act theory, norm adherence and violations labels, and verbal and non-verbal evidence. To build VideoNorms, we use a human-AI collaboration framework, where a teacher model using theoretically-grounded prompting provides candidate annotations and a set of trained human experts validate and correct the annotations. We benchmark a variety of open-weight VideoLLMs on the new dataset which highlight several common trends: 1) models performs worse on norm violation than adherence; 2) models perform worse w.r.t Chinese culture compared to the US culture; 3) models have more difficulty in providing non-verbal evidence compared to verbal for the norm adhere/violation label and struggle to identify the exact norm corresponding to a speech-act; and 4) unlike humans, models perform worse in formal, non-humorous contexts. Our findings emphasize the need for culturally-grounded video language model training - a gap our benchmark and framework begin to address.
comment: 24 pages, 5 figures, under review
☆ SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models
Spatial reasoning remains a fundamental challenge for Vision-Language Models (VLMs), with current approaches struggling to achieve robust performance despite recent advances. We identify that this limitation stems from a critical gap: existing methods attempt to learn spatial reasoning directly without establishing the hierarchical foundations of perception and understanding. To address this challenge, we present a comprehensive methodology for building spatial intelligence progressively. We introduce SpatialLadder-26k, a multimodal dataset containing 26,610 samples spanning object localization, single image, multi-view, and video spatial reasoning tasks, constructed through a standardized pipeline that ensures systematic coverage across modalities. Building on this dataset, we design a three-stage progressive training framework that (1) establishes spatial perception through object localization, (2) develops spatial understanding through multi-dimensional spatial tasks, and (3) strengthens complex reasoning via reinforcement learning with verifiable rewards. This approach yields SpatialLadder, a 3B-parameter model that achieves state-of-the-art performance on spatial reasoning benchmarks, with 23.4% average improvement over the base model, surpassing GPT-4o by 20.8% and Gemini-2.0-Flash by 10.1%. Notably, SpatialLadder maintains strong generalization with 7.2% improvement on out-of-domain benchmarks, demonstrating that progressive training from perception to reasoning is essential for robust spatial intelligence.
comment: Project Page: https://zju-real.github.io/SpatialLadder/ Code: https://github.com/ZJU-REAL/SpatialLadder
☆ CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards
Self-evolution is a central research topic in enabling large language model (LLM)-based agents to continually improve their capabilities after pretraining. Recent research has witnessed a transition from reinforcement learning (RL)-free to RL-based methods. Current RL-based methods either rely on dense external reward signals or extract intrinsic reward signals from LLMs themselves. However, these approaches diverge from the self-evolution mechanisms observed in human intelligence, where individuals learn and improve through mutual discussion and collaboration. In this work, we introduce Co-Evolving Multi-Agent Systems (CoMAS), a novel framework that enables agents to improve autonomously by learning from inter-agent interactions without external supervision. CoMAS generates intrinsic rewards from rich discussion dynamics, employs an LLM-as-a-judge mechanism to formulate these rewards, and optimizes each agent's policy through RL, thereby enabling decentralized and scalable co-evolution. Experimental results demonstrate that CoMAS consistently outperforms untrained agents and achieves state-of-the-art performance across most evaluation settings. Ablation studies confirm the necessity of interaction-based reward signals and reveal promising scalability as the number and diversity of agents increase. These findings establish CoMAS as a novel and effective paradigm for self-evolution in LLM-based agents.
☆ Which Heads Matter for Reasoning? RL-Guided KV Cache Compression
Reasoning large language models exhibit complex reasoning behaviors through the extended chain-of-thought generation, creating unprecedented Key-Value (KV) cache overhead during the decoding phase. Existing KV cache compression methods underperform on reasoning models: token-dropping methods break reasoning integrity by discarding critical information, while head-reallocating methods mistakenly compress reasoning-critical heads since they are designed for retrieval tasks, resulting in significant performance degradation as compression rates increase. We hypothesize that KV heads exhibit functional heterogeneity in reasoning models-some heads are critical for chain-of-thought consistency while others are compressible. To validate and exploit this insight, we propose RLKV, a novel reasoning-critical head identification framework, which uses reinforcement learning to directly optimize the relationship between each head's cache usage and reasoning quality. As RLKV produces rewards from actual generated samples during training, it naturally identifies heads relevant to reasoning behaviors. We then allocate full KV cache to these heads while applying compressed constant KV cache to others for efficient inference. Our experiments reveal that only a small fraction of attention heads is essential for reasoning, enabling our KV compression approach to outperform baseline methods while achieving 20-50% cache reduction with near lossless performance compared to uncompressed results.
☆ Efficient Prompt Optimisation for Legal Text Classification with Proxy Prompt Evaluator EMNLP 2025
Prompt optimization aims to systematically refine prompts to enhance a language model's performance on specific tasks. Fairness detection in Terms of Service (ToS) clauses is a challenging legal NLP task that demands carefully crafted prompts to ensure reliable results. However, existing prompt optimization methods are often computationally expensive due to inefficient search strategies and costly prompt candidate scoring. In this paper, we propose a framework that combines Monte Carlo Tree Search (MCTS) with a proxy prompt evaluator to more effectively explore the prompt space while reducing evaluation costs. Experiments demonstrate that our approach achieves higher classification accuracy and efficiency than baseline methods under a constrained computation budget.
comment: Accepted at NLLP@EMNLP 2025
☆ CaRT: Teaching LLM Agents to Know When They Know Enough
Many tasks require learned models to strategically gather relevant information over multiple rounds of interaction before actually acting on a task. Strategic information gathering requires models to know not only how to effectively acquire information, but also when to stop gathering information and make a decision, in order to avoid overthinking or getting derailed when acting. In this paper, we formalize this problem and introduce Counterfactuals and Reasoning for Termination (CaRT), an approach for teaching LLMs when to stop seeking information. To appropriately learn when to terminate, CaRT fine-tunes LLMs using counterfactual pairs of trajectories, one where termination is appropriate and a minimally modified version of the same trajectory where it is not. It trains the LLM to explain the rationale for the termination decision in either case via verbal reasoning, and imbues this capability into the base LLM via fine-tuning. We instantiate CaRT in two domains: interactive medical diagnosis and math problem solving. In both domains, we find that CaRT improves the efficiency of information gathering and task success rate compared to other fine-tuning methods.
☆ SliceFine: The Universal Winning-Slice Hypothesis for Pretrained Networks
This paper presents a theoretical framework explaining why fine tuning small, randomly selected subnetworks (slices) within pre trained models can be sufficient for downstream adaptation. We prove that pretrained networks exhibit a universal winning slice property arising from two phenomena: (1) spectral balance the eigenspectra of different weight matrix slices are remarkably similar; and (2) high task energy their backbone representations retain rich, task relevant features. This leads to the Universal Winning Slice Hypothesis, which provides a theoretical foundation for parameter efficient fine tuning (PEFT) in large scale models. Inspired by this, we propose SliceFine, a PEFT method that exploits this inherent redundancy by updating only selected slices of the original weights introducing zero new parameters, unlike adapter-based approaches. Empirically, SliceFine matches the performance of state of the art PEFT methods across language and vision tasks, while significantly improving training speed, memory efficiency, and model compactness. Our work bridges theory and practice, offering a theoretically grounded alternative to existing PEFT techniques.
☆ AutoMLGen: Navigating Fine-Grained Optimization for Coding Agents
Large language models (LLMs) have shown impressive performance in general programming tasks. However, in Machine Learning Engineering (MLE) scenarios such as AutoML and Kaggle competitions, achieving high performance depends heavily on expert intervention and repeated adjustments rather than simply generating correct code. When applied directly to these tasks, LLMs often lack fine-grained domain priors, and existing MLE approaches that use linear or tree-structured searches limit knowledge transfer to adjacent hierarchical links. As a result, they cannot leverage past full trajectories or share information across branches, limiting self-evolving ability and search space diversity. To address these limitations, we introduce AutoMLGen, an LLM-based coding agent that integrates a domain knowledge base for high-quality prior guidance and Monte Carlo Graph Search (MCGS) for efficient exploration. MCGS retains the tree-guided exploration of MCTS while embedding a graph structure into the expansion stage to enable dynamic path reorganization, historical trajectory reuse, and multi-solution fusion to support both self-evolution and collaborative learning. Combined with fine-grained operator sets, this design improves stability and accelerates convergence. Evaluation on the MLE-Bench shows that AutoMLGen achieves state-of-the-art performance in numerous dimensions, such as the average medal rate and the valid submission rate, under a 12-hour budget (half the standard runtime). The code is available at https://github.com/Alpha-Innovator/InternAgent.
☆ To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Models
Large Vision Language Models (LVLMs) have recently emerged as powerful architectures capable of understanding and reasoning over both visual and textual information. These models typically rely on two key components: a Vision Transformer (ViT) and a Large Language Model (LLM). ViT encodes visual content into a sequence of image tokens and serves as the perceptual front-end -- the eyes of the model. In contrast, the LLM interprets these tokens to perform high-level reasoning, generates responses, and functions as the cognitive core -- the brain of the model. However, it remains unclear which visual tokens contribute most significantly to understanding and reasoning, and how effectively these signals are propagated from ViT to the LLM. While most existing works have focused on identifying attention sinks, low-semantic tokens receiving disproportionately high attention, within the LLM, we shift the focus to the vision encoder by identifying a class of high-norm visual tokens from ViT, referred to as ViT attention sinks -- a problem that has been rarely studied but is indeed very important for LVLMs. Our findings show that these ViT sinks encapsulate high-level semantic concepts from images, allowing the LLM to perform more effective understanding and reasoning. Despite their importance, these sink tokens are often overlooked in existing LVLM architectures. To explore their contribution, we present both qualitative and quantitative analyses of the information embedded in these sink tokens. We also propose both training-free and training-based approaches to better leverage how this information is interpreted by the LLM, and to what extent. By explicitly utilizing these tokens, we demonstrate substantial improvements across a range of LVLMs and visual reasoning tasks, highlighting the untapped potential of ViT attention sinks in enhancing visual reasoning.
comment: Preprint. Project page: https://davidhalladay.github.io/diysink_demo
☆ Neologism Learning for Controllability and Self-Verbalization
Humans invent new words when there is a rising demand for a new useful concept (e.g., doomscrolling). We explore and validate a similar idea in our communication with LLMs: introducing new words to better understand and control the models, expanding on the recently introduced neologism learning. This method introduces a new word by adding a new word embedding and training with examples that exhibit the concept with no other changes in model parameters. We show that adding a new word allows for control of concepts such as flattery, incorrect answers, text length, as well as more complex concepts in AxBench. We discover that neologisms can also further our understanding of the model via self-verbalization: models can describe what each new word means to them in natural language, like explaining that a word that represents a concept of incorrect answers means ``a lack of complete, coherent, or meaningful answers...'' To validate self-verbalizations, we introduce plug-in evaluation: we insert the verbalization into the context of a model and measure whether it controls the target concept. In some self-verbalizations, we find machine-only synonyms: words that seem unrelated to humans but cause similar behavior in machines. Finally, we show how neologism learning can jointly learn multiple concepts in multiple words.
☆ DeepPrune: Parallel Scaling without Inter-trace Redundancy
Parallel scaling has emerged as a powerful paradigm to enhance reasoning capabilities in large language models (LLMs) by generating multiple Chain-of-Thought (CoT) traces simultaneously. However, this approach introduces significant computational inefficiency due to inter-trace redundancy -- our analysis reveals that over 80% of parallel reasoning traces yield identical final answers, representing substantial wasted computation. To address this critical efficiency bottleneck, we propose DeepPrune, a novel framework that enables efficient parallel scaling through dynamic pruning. Our method features a specialized judge model trained with focal loss and oversampling techniques to accurately predict answer equivalence from partial reasoning traces which realizes 0.87 AUROC on equivalence prediction, combined with an online greedy clustering algorithm that dynamically prunes redundant paths while preserving answer diversity. Comprehensive evaluations across three challenging benchmarks (AIME 2024, AIME 2025, and GPQA) and multiple reasoning models demonstrate that DeepPrune achieves remarkable token reduction by over 80% compared to conventional consensus sampling on most cases, while maintaining competitive accuracy within 3 percentage points. Our work establishes a new standard for efficient parallel reasoning, making high-performance reasoning more efficient. Our code and data are here: https://deepprune.github.io/
comment: 15 pages, 4 figures, please check out the project page: https://deepprune.github.io/
☆ The Visual Iconicity Challenge: Evaluating Vision-Language Models on Sign Language Form-Meaning Mapping
Iconicity, the resemblance between linguistic form and meaning, is pervasive in signed languages, offering a natural testbed for visual grounding. For vision-language models (VLMs), the challenge is to recover such essential mappings from dynamic human motion rather than static context. We introduce the \textit{Visual Iconicity Challenge}, a novel video-based benchmark that adapts psycholinguistic measures to evaluate VLMs on three tasks: (i) phonological sign-form prediction (e.g., handshape, location), (ii) transparency (inferring meaning from visual form), and (iii) graded iconicity ratings. We assess $13$ state-of-the-art VLMs in zero- and few-shot settings on Sign Language of the Netherlands and compare them to human baselines. On \textit{phonological form prediction}, VLMs recover some handshape and location detail but remain below human performance; on \textit{transparency}, they are far from human baselines; and only top models correlate moderately with human \textit{iconicity ratings}. Interestingly, \textit{models with stronger phonological form prediction correlate better with human iconicity judgment}, indicating shared sensitivity to visually grounded structure. Our findings validate these diagnostic tasks and motivate human-centric signals and embodied learning methods for modelling iconicity and improving visual grounding in multimodal models.
☆ Looking to Learn: Token-wise Dynamic Gating for Low-Resource Vision-Language Modelling EMNLP 2025
Training vision-language models on cognitively-plausible amounts of data requires rethinking how models integrate multimodal information. Within the constraints of the Vision track for the BabyLM Challenge 2025, we propose a lightweight decoder-based architecture with (1) token-wise dynamic gating for adaptive fusion of linguistic and visual cues, (2) feature modulation and channel attention to maximise the utility of limited visual information and (3) auxiliary contrastive objectives for visual grounding. Evaluation on five benchmarks (BLiMP, BLiMP Supplement, EWoK, Winoground and VQA) shows competitive or superior performance to multimodal baselines. More notably, our dynamic gate discovers interpretable patterns without explicit supervision, favouring visual cues for content words and linguistic cues for function words. While we identify limitations in the Challenge constraints, such as the information bottleneck created by global image embeddings and training instability from the dataset split, our findings establish dynamic gating as a powerful tool for efficient multimodal learning, offering both interpretability and performance even under severe constraints.
comment: Accepted to the EMNLP 2025 BabyLM Workshop
☆ LeWiDi-2025 at NLPerspectives: The Third Edition of the Learning with Disagreements Shared Task EMNLP 2025
Many researchers have reached the conclusion that AI models should be trained to be aware of the possibility of variation and disagreement in human judgments, and evaluated as per their ability to recognize such variation. The LEWIDI series of shared tasks on Learning With Disagreements was established to promote this approach to training and evaluating AI models, by making suitable datasets more accessible and by developing evaluation methods. The third edition of the task builds on this goal by extending the LEWIDI benchmark to four datasets spanning paraphrase identification, irony detection, sarcasm detection, and natural language inference, with labeling schemes that include not only categorical judgments as in previous editions, but ordinal judgments as well. Another novelty is that we adopt two complementary paradigms to evaluate disagreement-aware systems: the soft-label approach, in which models predict population-level distributions of judgments, and the perspectivist approach, in which models predict the interpretations of individual annotators. Crucially, we moved beyond standard metrics such as cross-entropy, and tested new evaluation metrics for the two paradigms. The task attracted diverse participation, and the results provide insights into the strengths and limitations of methods to modeling variation. Together, these contributions strengthen LEWIDI as a framework and provide new resources, benchmarks, and findings to support the development of disagreement-aware technologies.
comment: 14 pages; LeWiDi-2025 shared task description paper at NLPerspective workshop at EMNLP 2025
☆ ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping
Recent advances in multimodal large reasoning models (MLRMs) have substantially improved their ability to solve complex textual and visual tasks. However, these models tend to overthink on simple problems, producing unnecessarily lengthy reasoning traces, while under-exploring on challenging ones, leading to missed solutions. To address this imbalance, we propose ARES, a unified open-source framework for adaptive reasoning that dynamically allocates exploration effort based on task difficulty. Our approach is motivated by two key empirical findings: (i) while single-token entropy is noisy, high window-entropy (HWE) tokens (token-level entropies averaged under a sliding window) can reliably capture reasoning-critical moments; and (ii) reducing HWE usage benefits easy problems, while increasing it is essential for solving hard ones. Building on these insights, ARES introduces a two-stage training pipeline. In the Adaptive Cold-Start stage, we curate multimodal and textual data paired with reasoning traces of length proportional to problem difficulty, equipping the model with initial difficulty awareness. In the second stage, we develop Adaptive Entropy Policy Optimization (AEPO), which uses HWE tokens as exploration triggers to decide when to explore, and a hierarchical entropy reward with dynamic KL control to decide how much to explore. Extensive experiments demonstrate that ARES achieves superior performance and reasoning efficiency across diverse mathematical, logical, and multimodal benchmarks, while closing the gap to leading commercial systems under significantly lower inference costs.
☆ xRouter: Training Cost-Aware LLMs Orchestration System via Reinforcement Learning
Modern LLM deployments confront a widening cost-performance spectrum: premium models deliver strong reasoning but are expensive, while lightweight models are economical yet brittle on complex tasks. Static escalation rules and keyword heuristics under-utilize this spectrum and fail to adapt across task types. We present xRouter, a tool-calling-based routing system in which a learned router can either answer directly or invoke one or more external models. The router is trained end-to-end with reinforcement learning using an explicit, cost-aware reward that encodes cost-performance trade-offs, eliminating the need for hand-engineered routing rules. Our implementation encompasses the full reinforcement learning framework, including reward and cost accounting, as well as the deployment and evaluation pipelines. Across diverse benchmarks, xRouter achieves strong cost-performance trade-offs (e.g., substantial cost reductions at comparable task completion rates), and provides empirical insights into what reliably helps learned routing and what does not, ranging from model trainability to the difficulty of eliciting sophisticated orchestration behaviors in small open models. We hope these findings and our open implementation will serve as a practical substrate for advancing learned, cost-aware LLM orchestration.
comment: 24 Pages, 4 Figures, 2 Tables
☆ Single layer tiny Co$^4$ outpaces GPT-2 and GPT-BERT
We show that a tiny Co$^4$ machine(Adeel,2025) with a single layer, two heads, and 8M parameters, operating at an approximate cost of $O(N)$ (where $N$ is the number of input tokens), outpaces the BabyLM Challenge baselines GPT-2 (124M, 12 layers, $O(N^2))$ and GPT-BERT (30M, 12 layers, $O(N^2))$ in just two epochs, while both are trained for ten. Co$^4$ achieves orders-of-magnitude greater training efficiency on 10M tokens, demonstrating highly sample efficient pretraining. Using the BabyLM challenge evaluation pipeline across complex benchmarks, Co$^4$ exhibits strong zero-shot and fine-tuning performance on SuperGLUE tasks. Specifically, Co$^4$ outperforms GPT-2 on 5 out of 7 zero-shot metrics and 6 out of 7 fine-tuning tasks, and GPT-BERT on 4 out of 7 metrics in both cases. These results suggest the need to rethink prevailing deep learning paradigms and associated scaling laws.
☆ FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts NeurIPS 2025
Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning method for foundation models, but it suffers from parameter interference, resulting in suboptimal performance. Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-task correlations in single-task instruction tuning, they introduce additional router parameters and remain ineffective in multi-task model merging where inter-task interference arises. Inspired by the fly olfactory circuit, we propose FlyLoRA, an implicit MoE-based LoRA variant that introduces: (1) rank-wise expert activation in the up-projection matrix, and (2) an implicit router that unifies expert routing and down-projection, where a frozen sparse random projection matrix replaces the traditional dense trainable version. This design resolves the trade-off between intra-task decorrelation and computational efficiency by eliminating the need for an explicit router, while inherently mitigating inter-task interference due to the orthogonality property of random matrices. Extensive experiments across four domains -- general knowledge understanding, scientific question answering, mathematical reasoning, and code generation -- demonstrate consistent performance improvements over existing methods. Beyond empirical gains, FlyLoRA highlights how biological structures can inspire innovations in AI technologies. Code is available at https://github.com/gfyddha/FlyLoRA.
comment: NeurIPS 2025 accepted paper
☆ If Probable, Then Acceptable? Understanding Conditional Acceptability Judgments in Large Language Models
Conditional acceptability refers to how plausible a conditional statement is perceived to be. It plays an important role in communication and reasoning, as it influences how individuals interpret implications, assess arguments, and make decisions based on hypothetical scenarios. When humans evaluate how acceptable a conditional "If A, then B" is, their judgments are influenced by two main factors: the $\textit{conditional probability}$ of $B$ given $A$, and the $\textit{semantic relevance}$ of the antecedent $A$ given the consequent $B$ (i.e., whether $A$ meaningfully supports $B$). While prior work has examined how large language models (LLMs) draw inferences about conditional statements, it remains unclear how these models judge the $\textit{acceptability}$ of such statements. To address this gap, we present a comprehensive study of LLMs' conditional acceptability judgments across different model families, sizes, and prompting strategies. Using linear mixed-effects models and ANOVA tests, we find that models are sensitive to both conditional probability and semantic relevance-though to varying degrees depending on architecture and prompting style. A comparison with human data reveals that while LLMs incorporate probabilistic and semantic cues, they do so less consistently than humans. Notably, larger models do not necessarily align more closely with human judgments.
comment: 22 pages, 12 figures
☆ On the Relationship Between the Choice of Representation and In-Context Learning
In-context learning (ICL) is the ability of a large language model (LLM) to learn a new task from a few demonstrations presented as part of the context. Past studies have attributed a large portion of the success of ICL to the way these in-context demonstrations are represented, particularly to how labels are represented in classification tasks. On the other hand, observations of the learning capacity of ICL (i.e., the extent to which more in-context demonstrations can lead to higher performance) have been mixed, and ICL is often thought to occur only under specific conditions. The interaction between these two aspects in ICL, representation and learning, has not been studied in depth until now. We hypothesize that they are largely independent of one another, such that the representation of demonstrations determines the baseline accuracy of ICL, while learning from additional demonstrations improves only on top of this baseline. We validate this hypothesis by developing an optimization algorithm that can enumerate a spectrum of possible label sets (representations) varying in semantic relevance. We then perform ICL with varying numbers of in-context demonstrations for each of these label sets. We observed that learning happens regardless of the quality of the label set itself, although its efficiency, measured by the slope of improvement over in-context demonstrations, is conditioned on both the label set quality and the parameter count of the underlying language model. Despite the emergence of learning, the relative quality (accuracy) of the choice of a label set (representation) is largely maintained throughout learning, confirming our hypothesis and implying their orthogonality. Our work reveals a previously underexplored aspect of ICL: the independent effects of learning from demonstrations and their representations on ICL performance.
comment: 25 pages, 6 figures, 10 tables
☆ Two-Stage Voting for Robust and Efficient Suicide Risk Detection on Social Media
Suicide rates have risen worldwide in recent years, underscoring the urgent need for proactive prevention strategies. Social media provides valuable signals, as many at-risk individuals - who often avoid formal help due to stigma - choose instead to share their distress online. Yet detecting implicit suicidal ideation, conveyed indirectly through metaphor, sarcasm, or subtle emotional cues, remains highly challenging. Lightweight models like BERT handle explicit signals but fail on subtle implicit ones, while large language models (LLMs) capture nuance at prohibitive computational cost. To address this gap, we propose a two-stage voting architecture that balances efficiency and robustness. In Stage 1, a lightweight BERT classifier rapidly resolves high-confidence explicit cases. In Stage 2, ambiguous inputs are escalated to either (i) a multi-perspective LLM voting framework to maximize recall on implicit ideation, or (ii) a feature-based ML ensemble guided by psychologically grounded indicators extracted via prompt-engineered LLMs for efficiency and interpretability. To the best of our knowledge, this is among the first works to operationalize LLM-extracted psychological features as structured vectors for suicide risk detection. On two complementary datasets - explicit-dominant Reddit and implicit-only DeepSuiMind - our framework outperforms single-model baselines, achieving 98.0% F1 on explicit cases, 99.7% on implicit ones, and reducing the cross-domain gap below 2%, while significantly lowering LLM cost.
☆ AutoRed: A Free-form Adversarial Prompt Generation Framework for Automated Red Teaming
The safety of Large Language Models (LLMs) is crucial for the development of trustworthy AI applications. Existing red teaming methods often rely on seed instructions, which limits the semantic diversity of the synthesized adversarial prompts. We propose AutoRed, a free-form adversarial prompt generation framework that removes the need for seed instructions. AutoRed operates in two stages: (1) persona-guided adversarial instruction generation, and (2) a reflection loop to iteratively refine low-quality prompts. To improve efficiency, we introduce a verifier to assess prompt harmfulness without querying the target models. Using AutoRed, we build two red teaming datasets -- AutoRed-Medium and AutoRed-Hard -- and evaluate eight state-of-the-art LLMs. AutoRed achieves higher attack success rates and better generalization than existing baselines. Our results highlight the limitations of seed-based approaches and demonstrate the potential of free-form red teaming for LLM safety evaluation. We will open source our datasets in the near future.
☆ Beyond Pass@k: Breadth-Depth Metrics for Reasoning Boundaries
Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm to improve Large Language Models on reasoning tasks such as coding, math or logic. To assess the reasoning boundary (the fraction of problems a model can solve) researchers often report Pass@k at large sampling budgets. Recent results reveal a crossover phenomenon: while RLVR models outperform the base model at small k values, the base model usually outperforms them when sampling a very large number of completions. This has been interpreted as evidence that base models have a larger reasoning boundary. We argue that on tasks with discrete answer spaces, such as math with numeric outputs, Pass@k at large k reflects the increasingly higher chance of success in the limit of the number of trials rather than genuine reasoning, and can therefore be misleading. We propose Cover@tau, which measures the fraction of problems that a model can solve for which at least a tau proportion of completions are correct. Unlike Pass@k, Cover@tau captures reasoning under an explicit reliability threshold: models that rely on random guessing degrade rapidly as tau increases. We evaluate several RLVR models using Cover@tau-based metrics and illustrate how the relative rankings of popular algorithms change compared to Pass@1, offering a different perspective on reasoning boundaries.
comment: 10 pages, 3 figures
☆ Neuron-Level Analysis of Cultural Understanding in Large Language Models
As large language models (LLMs) are increasingly deployed worldwide, ensuring their fair and comprehensive cultural understanding is important. However, LLMs exhibit cultural bias and limited awareness of underrepresented cultures, while the mechanisms underlying their cultural understanding remain underexplored. To fill this gap, we conduct a neuron-level analysis to identify neurons that drive cultural behavior, introducing a gradient-based scoring method with additional filtering for precise refinement. We identify both culture-general neurons contributing to cultural understanding regardless of cultures, and culture-specific neurons tied to an individual culture. These neurons account for less than 1% of all neurons and are concentrated in shallow to middle MLP layers. We validate their role by showing that suppressing them substantially degrades performance on cultural benchmarks (by up to 30%), while performance on general natural language understanding (NLU) benchmarks remains largely unaffected. Moreover, we show that culture-specific neurons support knowledge of not only the target culture, but also related cultures. Finally, we demonstrate that training on NLU benchmarks can diminish models' cultural understanding when we update modules containing many culture-general neurons. These findings provide insights into the internal mechanisms of LLMs and offer practical guidance for model training and engineering. Our code is available at https://github.com/ynklab/CULNIG
☆ Beyond Turn Limits: Training Deep Search Agents with Dynamic Context Window
While recent advances in reasoning models have demonstrated cognitive behaviors through reinforcement learning, existing approaches struggle to invoke deep reasoning capabilities in multi-turn agents with long-horizon interactions. We propose DeepMiner, a novel framework that elicits such abilities by introducing high-difficulty training tasks and dynamic context window. DeepMiner presents a reverse construction method to generate complex but verifiable question-answer pairs from authentic web sources, which ensures the challenge and reliability of training data while injecting cognitive capabilities into multi-turn reasoning scenarios. We further design an elegant yet effective dynamic context management strategy for both training and inference, utilizing sliding window mechanisms while eliminating the dependency on external summarization models, thereby efficiently empowering the model to handle continuously expanding long-horizon contexts. Through reinforcement learning on Qwen3-32B, we develop DeepMiner-32B, which achieves substantial performance improvements across multiple search agent benchmarks. DeepMiner attains 33.5% accuracy on BrowseComp-en, surpassing the previous best open-source agent by almost 20 percentage points, and demonstrates consistent improvements on BrowseComp-zh, XBench-DeepSearch, and GAIA. Notably, our dynamic context management enables sustained interactions of nearly 100 turns within standard 32k context length, effectively addressing the context limitations that constrain existing multi-turn interaction systems.
☆ Mix- and MoE-DPO: A Variational Inference Approach to Direct Preference Optimization
Direct Preference Optimization (DPO) has recently emerged as a simple and effective alternative to reinforcement learning from human feedback (RLHF) for aligning large language models (LLMs) with user preferences. However, existing DPO formulations rely on a single monolithic model, which limits their expressivity in multi-task settings and their adaptability to heterogeneous or diverse preference distributions. In this work, we propose Mix- and MoE-DPO, a framework that extends DPO with both soft mixture models and mixture-of-experts (MoE) architectures, using a stochastic variational inference approach. Our method introduces a latent-variable model over expert assignments and optimizes a variational evidence lower bound (ELBO), enabling stable and efficient learning of specialized expert policies from preference data. Mix- and MoE-DPO provides three key advantages over standard DPO: (i) generalization via universal function approximation through mixtures; (ii) reward and policy specialization through expert components tailored to distinct preference modes; and (iii) contextual alignment through input-dependent soft gating that enables user-specific mixture policies. Our framework supports both shared base architectures with expert-specific policy heads and fully independent expert models, allowing flexible trade-offs between parameter efficiency and specialization. We validate our approach on a variety of model sizes and multi-preference datasets, demonstrating that Mix- and MoE-DPO offers a powerful and scalable method for preference-based LLM alignment.
☆ Opponent Shaping in LLM Agents
Large Language Models (LLMs) are increasingly being deployed as autonomous agents in real-world environments. As these deployments scale, multi-agent interactions become inevitable, making it essential to understand strategic behavior in such systems. A central open question is whether LLM agents, like reinforcement learning agents, can shape the learning dynamics and influence the behavior of others through interaction alone. In this paper, we present the first investigation of opponent shaping (OS) with LLM-based agents. Existing OS algorithms cannot be directly applied to LLMs, as they require higher-order derivatives, face scalability constraints, or depend on architectural components that are absent in transformers. To address this gap, we introduce ShapeLLM, an adaptation of model-free OS methods tailored for transformer-based agents. Using ShapeLLM, we examine whether LLM agents can influence co-players' learning dynamics across diverse game-theoretic environments. We demonstrate that LLM agents can successfully guide opponents toward exploitable equilibria in competitive games (Iterated Prisoner's Dilemma, Matching Pennies, and Chicken) and promote coordination and improve collective welfare in cooperative games (Iterated Stag Hunt and a cooperative version of the Prisoner's Dilemma). Our findings show that LLM agents can both shape and be shaped through interaction, establishing opponent shaping as a key dimension of multi-agent LLM research.
comment: 29 pages, 15 figures, 15 tables
☆ ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval
In this paper, we introduce ReasonEmbed, a novel text embedding model developed for reasoning-intensive document retrieval. Our work includes three key technical contributions. First, we propose ReMixer, a new data synthesis method that overcomes the triviality problem prevalent in previous synthetic datasets, enabling large-scale production of 82K high-quality training samples. Second, we design Redapter, a self-adaptive learning algorithm that dynamically adjusts training each sample's weight based on its reasoning intensity. This allows the model to effectively capture the complex semantic relationships between queries and documents. Third, we implement ReasonEmbed across multiple backbones of varying sizes, all of which achieve superior performance on reasoning-intensive retrieval tasks. Notably, our ReasonEmbed-Qwen3-8B model offers a record-high nDCG@10 score of 38.1 on the BRIGHT benchmark, which significantly outperforms existing text embedding models. We will fully open-source our created resources in ReasonEmbed to push forward the research advancement in this field.
comment: 17 pages, 3 figures
☆ Contrastive Decoding for Synthetic Data Generation in Low-Resource Language Modeling
Large language models (LLMs) are trained on huge amounts of textual data, and concerns have been raised that the limits of such data may soon be reached. A potential solution is to train on synthetic data sampled from LLMs. In this work, we build on this idea and investigate the benefits of contrastive decoding for generating synthetic corpora. In a controlled setting, we experiment with sampling corpora using the relative difference between a good and bad model trained on the same original corpus of 100 million words. By amplifying the signal from a model that has better performance, we create a synthetic corpus and mix it with the original training data. Our findings show that training on a mixture of synthesized and real data improves performance on the language modeling objective and a range of downstream tasks. In particular, we see that training with a mix of synthetic data from contrastive decoding benefits tasks that require more reasoning skills, while synthetic data from traditional sampling helps more on tasks dependent on surface level linguistic capabilities.
comment: 13 pages, 3 figures
☆ The Alignment Waltz: Jointly Training Agents to Collaborate for Safety
Harnessing the power of LLMs requires a delicate dance between being helpful and harmless. This creates a fundamental tension between two competing challenges: vulnerability to adversarial attacks that elicit unsafe content, and a tendency for overrefusal on benign but sensitive prompts. Current approaches often navigate this dance with safeguard models that completely reject any content that contains unsafe portions. This approach cuts the music entirely-it may exacerbate overrefusals and fails to provide nuanced guidance for queries it refuses. To teach models a more coordinated choreography, we propose WaltzRL, a novel multi-agent reinforcement learning framework that formulates safety alignment as a collaborative, positive-sum game. WaltzRL jointly trains a conversation agent and a feedback agent, where the latter is incentivized to provide useful suggestions that improve the safety and helpfulness of the conversation agent's responses. At the core of WaltzRL is a Dynamic Improvement Reward (DIR) that evolves over time based on how well the conversation agent incorporates the feedback. At inference time, unsafe or overrefusing responses from the conversation agent are improved rather than discarded. The feedback agent is deployed together with the conversation agent and only engages adaptively when needed, preserving helpfulness and low latency on safe queries. Our experiments, conducted across five diverse datasets, demonstrate that WaltzRL significantly reduces both unsafe responses (e.g., from 39.0% to 4.6% on WildJailbreak) and overrefusals (from 45.3% to 9.9% on OR-Bench) compared to various baselines. By enabling the conversation and feedback agents to co-evolve and adaptively apply feedback, WaltzRL enhances LLM safety without degrading general capabilities, thereby advancing the Pareto front between helpfulness and harmlessness.
☆ Investigating Counterclaims in Causality Extraction from Text
Research on causality extraction from text has so far almost entirely neglected counterclaims. Existing causality extraction datasets focus solely on "procausal" claims, i.e., statements that support a relationship. "Concausal" claims, i.e., statements that refute a relationship, are entirely ignored or even accidentally annotated as procausal. We address this shortcoming by developing a new dataset that integrates concausality. Based on an extensive literature review, we first show that concausality is an integral part of causal reasoning on incomplete knowledge. We operationalize this theory in the form of a rigorous guideline for annotation and then augment the Causal News Corpus with concausal statements, obtaining a substantial inter-annotator agreement of Cohen's $\kappa=0.74$. To demonstrate the importance of integrating concausal statements, we show that models trained without concausal relationships tend to misclassify these as procausal instead. Based on our new dataset, this mistake can be mitigated, enabling transformers to effectively distinguish pro- and concausality.
☆ SenWave: A Fine-Grained Multi-Language Sentiment Analysis Dataset Sourced from COVID-19 Tweets
The global impact of the COVID-19 pandemic has highlighted the need for a comprehensive understanding of public sentiment and reactions. Despite the availability of numerous public datasets on COVID-19, some reaching volumes of up to 100 billion data points, challenges persist regarding the availability of labeled data and the presence of coarse-grained or inappropriate sentiment labels. In this paper, we introduce SenWave, a novel fine-grained multi-language sentiment analysis dataset specifically designed for analyzing COVID-19 tweets, featuring ten sentiment categories across five languages. The dataset comprises 10,000 annotated tweets each in English and Arabic, along with 30,000 translated tweets in Spanish, French, and Italian, derived from English tweets. Additionally, it includes over 105 million unlabeled tweets collected during various COVID-19 waves. To enable accurate fine-grained sentiment classification, we fine-tuned pre-trained transformer-based language models using the labeled tweets. Our study provides an in-depth analysis of the evolving emotional landscape across languages, countries, and topics, revealing significant insights over time. Furthermore, we assess the compatibility of our dataset with ChatGPT, demonstrating its robustness and versatility in various applications. Our dataset and accompanying code are publicly accessible on the repository\footnote{https://github.com/gitdevqiang/SenWave}. We anticipate that this work will foster further exploration into fine-grained sentiment analysis for complex events within the NLP community, promoting more nuanced understanding and research innovations.
comment: 13 pages, 13 figures, 6 tables
☆ LLMs Learn to Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactions
Previous research has shown that LLMs finetuned on malicious or incorrect completions within narrow domains (e.g., insecure code or incorrect medical advice) can become broadly misaligned to exhibit harmful behaviors, which is called emergent misalignment. In this work, we investigate whether this phenomenon can extend beyond safety behaviors to a broader spectrum of dishonesty and deception under high-stakes scenarios (e.g., lying under pressure and deceptive behavior). To explore this, we finetune open-sourced LLMs on misaligned completions across diverse domains. Experimental results demonstrate that LLMs show broadly misaligned behavior in dishonesty. Additionally, we further explore this phenomenon in a downstream combined finetuning setting, and find that introducing as little as 1% of misalignment data into a standard downstream task is sufficient to decrease honest behavior over 20%. Furthermore, we consider a more practical human-AI interaction environment where we simulate both benign and biased users to interact with the assistant LLM. Notably, we find that the assistant can be misaligned unintentionally to exacerbate its dishonesty with only 10% biased user population. In summary, we extend the study of emergent misalignment to the domain of dishonesty and deception under high-stakes scenarios, and demonstrate that this risk arises not only through direct finetuning, but also in downstream mixture tasks and practical human-AI interactions.
☆ Memory Retrieval and Consolidation in Large Language Models through Function Tokens
The remarkable success of large language models (LLMs) stems from their ability to consolidate vast amounts of knowledge into the memory during pre-training and to retrieve it from the memory during inference, enabling advanced capabilities such as knowledge memorization, instruction-following and reasoning. However, the mechanisms of memory retrieval and consolidation in LLMs remain poorly understood. In this paper, we propose the function token hypothesis to explain the workings of LLMs: During inference, function tokens activate the most predictive features from context and govern next token prediction (memory retrieval). During pre-training, predicting the next tokens (usually content tokens) that follow function tokens increases the number of learned features of LLMs and updates the model parameters (memory consolidation). Function tokens here roughly correspond to function words in linguistics, including punctuation marks, articles, prepositions, and conjunctions, in contrast to content tokens. We provide extensive experimental evidence supporting this hypothesis. Using bipartite graph analysis, we show that a small number of function tokens activate the majority of features. Case studies further reveal how function tokens activate the most predictive features from context to direct next token prediction. We also find that during pre-training, the training loss is dominated by predicting the next content tokens following function tokens, which forces the function tokens to select the most predictive features from context.
☆ Sentiment Matters: An Analysis of 200 Human-SAV Interactions SC 2025
Shared Autonomous Vehicles (SAVs) are likely to become an important part of the transportation system, making effective human-SAV interactions an important area of research. This paper introduces a dataset of 200 human-SAV interactions to further this area of study. We present an open-source human-SAV conversational dataset, comprising both textual data (e.g., 2,136 human-SAV exchanges) and empirical data (e.g., post-interaction survey results on a range of psychological factors). The dataset's utility is demonstrated through two benchmark case studies: First, using random forest modeling and chord diagrams, we identify key predictors of SAV acceptance and perceived service quality, highlighting the critical influence of response sentiment polarity (i.e., perceived positivity). Second, we benchmark the performance of an LLM-based sentiment analysis tool against the traditional lexicon-based TextBlob method. Results indicate that even simple zero-shot LLM prompts more closely align with user-reported sentiment, though limitations remain. This study provides novel insights for designing conversational SAV interfaces and establishes a foundation for further exploration into advanced sentiment modeling, adaptive user interactions, and multimodal conversational systems.
comment: Accepted for presentation at IEEE ITSC 2025 and for publication in its Proceedings. \c{opyright} 2025 IEEE. Personal use permitted; other uses require permission from IEEE, including reprinting, republishing, or reuse of any copyrighted component of this work
☆ Training-Free Group Relative Policy Optimization
Recent advances in Large Language Model (LLM) agents have demonstrated their promising general capabilities. However, their performance in specialized real-world domains often degrades due to challenges in effectively integrating external tools and specific prompting strategies. While methods like agentic reinforcement learning have been proposed to address this, they typically rely on costly parameter updates, for example, through a process that uses Supervised Fine-Tuning (SFT) followed by a Reinforcement Learning (RL) phase with Group Relative Policy Optimization (GRPO) to alter the output distribution. However, we argue that LLMs can achieve a similar effect on the output distribution by learning experiential knowledge as a token prior, which is a far more lightweight approach that not only addresses practical data scarcity but also avoids the common issue of overfitting. To this end, we propose Training-Free Group Relative Policy Optimization (Training-Free GRPO), a cost-effective solution that enhances LLM agent performance without any parameter updates. Our method leverages the group relative semantic advantage instead of numerical ones within each group of rollouts, iteratively distilling high-quality experiential knowledge during multi-epoch learning on a minimal ground-truth data. Such knowledge serves as the learned token prior, which is seamlessly integrated during LLM API calls to guide model behavior. Experiments on mathematical reasoning and web searching tasks demonstrate that Training-Free GRPO, when applied to DeepSeek-V3.1-Terminus, significantly improves out-of-domain performance. With just a few dozen training samples, Training-Free GRPO outperforms fine-tuned small LLMs with marginal training data and cost.
☆ R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?
Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek-R1) have led to remarkable improvements through long Chain-of-Thought (CoT). However, existing benchmarks mainly focus on immediate, single-horizon tasks, failing to adequately evaluate models' ability to understand and respond to complex, long-horizon scenarios. To address this incomplete evaluation of Large Reasoning Models (LRMs), we propose R-HORIZON, a method designed to stimulate long-horizon reasoning behaviors in LRMs through query composition. Based on R-HORIZON, we construct a long-horizon reasoning benchmark, comprising complex multi-step reasoning tasks with interdependent problems that span long reasoning horizons. Through comprehensive evaluation of LRMs using the R-HORIZON benchmark, we find that even the most advanced LRMs suffer significant performance degradation. Our analysis reveals that LRMs exhibit limited effective reasoning length and struggle to allocate thinking budget across multiple problems appropriately. Recognizing these limitations, we use R-HORIZON to construct long-horizon reasoning data for reinforcement learning with verified rewards (RLVR). Compared to training with single-horizon data, RLVR with R-HORIZON not only substantially improves performance on the multi-horizon reasoning tasks, but also promotes accuracy on standard reasoning tasks, with an increase of 7.5 on AIME2024. These results position R-HORIZON as a scalable, controllable, and low-cost paradigm for enhancing and evaluating the long-horizon reasoning capabilities of LRMs.
☆ METRICALARGS: A Taxonomy for Studying Metrical Poetry with LLMs
Prior NLP work studying poetry has focused primarily on automatic poem generation and summarization. Many languages have well-studied traditions of poetic meter which enforce constraints on a poem in terms of syllable and phoneme patterns. Such advanced literary forms offer opportunities for probing deeper reasoning and language understanding in Large Language Models (LLMs) and their ability to follow strict pre-requisites and rules. In this paper, we introduce MetricalARGS, the first taxonomy of poetry-related NLP tasks designed to evaluate LLMs on metrical poetry across four dimensions: Analysis, Retrieval, Generation, and Support. We discuss how these tasks relate to existing NLP tasks, addressing questions around datasets and evaluation metrics. Taking Telugu as our example language, we illustrate how the taxonomy can be used in practice. MetricalARGS highlights the broader possibilities for understanding the capabilities and limitations of today's LLMs through the lens of metrical poetry.
comment: Pre-print
☆ NavSpace: How Navigation Agents Follow Spatial Intelligence Instructions
Instruction-following navigation is a key step toward embodied intelligence. Prior benchmarks mainly focus on semantic understanding but overlook systematically evaluating navigation agents' spatial perception and reasoning capabilities. In this work, we introduce the NavSpace benchmark, which contains six task categories and 1,228 trajectory-instruction pairs designed to probe the spatial intelligence of navigation agents. On this benchmark, we comprehensively evaluate 22 navigation agents, including state-of-the-art navigation models and multimodal large language models. The evaluation results lift the veil on spatial intelligence in embodied navigation. Furthermore, we propose SNav, a new spatially intelligent navigation model. SNav outperforms existing navigation agents on NavSpace and real robot tests, establishing a strong baseline for future work.
☆ ARM2: Adaptive Reasoning Model with Vision Understanding and Executable Code
Large Reasoning Models (LRMs) often suffer from the ``over-thinking'' problem, generating unnecessarily long reasoning on simple tasks. Some strategies have been proposed to mitigate this issue, such as length penalties or routing mechanisms, but they are typically heuristic and task-specific, lacking a general framework for adaptive reasoning. In this paper, we present ARM2, a unified model that adaptively balances reasoning performance and efficiency across multiple formats through a reinforcement learning framework augmented with length-aware optimization. Beyond conventional natural language inference, ARM2 integrates vision understanding, extending its applicability to multimodal. Moreover, ARM2 integrates executable code into reasoning, enabling substantial reductions in token cost while preserving task performance compared to long CoT. Experiments demonstrate that ARM2 achieves performance on par with traditional reasoning models trained with GRPO, while reducing token usage by over 70% on average. We further conduct extensive analyses to validate the effectiveness of ARM2 and the soundness of its design.
comment: Work in Progress
☆ Beyond Over-Refusal: Scenario-Based Diagnostics and Post-Hoc Mitigation for Exaggerated Refusals in LLMs
Large language models (LLMs) frequently produce false refusals, declining benign requests that contain terms resembling unsafe queries. We address this challenge by introducing two comprehensive benchmarks: the Exaggerated Safety Benchmark (XSB) for single-turn prompts, annotated with "Focus" keywords that identify refusal-inducing triggers, and the Multi-turn Scenario-based Exaggerated Safety Benchmark (MS-XSB), which systematically evaluates refusal calibration in realistic, context-rich dialog settings. Our benchmarks reveal that exaggerated refusals persist across diverse recent LLMs and are especially pronounced in complex, multi-turn scenarios. To mitigate these failures, we leverage post-hoc explanation methods to identify refusal triggers and deploy three lightweight, model-agnostic approaches, ignore-word instructions, prompt rephrasing, and attention steering, at inference time, all without retraining or parameter access. Experiments on four instruction-tuned Llama models demonstrate that these strategies substantially improve compliance on safe prompts while maintaining robust safety protections. Our findings establish a reproducible framework for diagnosing and mitigating exaggerated refusals, highlighting practical pathways to safer and more helpful LLM deployments.
☆ DACIP-RC: Domain Adaptive Continual Instruction Pre-Training via Reading Comprehension on Business Conversations EMNLP 2025
The rapid advancements in Large Language Models (LLMs) have enabled their adoption in real-world industrial scenarios for various natural language processing tasks. However, the high inference cost of large-scale LLMs makes their deployment impractical, necessitating the use of smaller models. Despite their efficiency, smaller LLMs lack robust zero-shot instruction-following capabilities across diverse domains, limiting their adaptability to dynamic user requirements. Traditional fine-tuning approaches exacerbate this issue by inducing catastrophic forgetting, reducing the model's generalization ability for unseen tasks. In this paper, we propose Domain Adaptive Continual Instruction Pre-Training via Reading Comprehension (DACIP-RC), a continual pre-training technique that enhances smaller LLMs' domain adaptability for business conversational tasks. Unlike conventional pre-training approaches that rely on next-token prediction, DACIP-RC generates diverse task instructions and responses via reading comprehension on conversation transcripts, enabling better instruction generalization. Our empirical evaluations demonstrate that DACIP-RC significantly improves zero-shot generalization across a wide range of business conversational tasks, including meeting summarization, action item generation, and call purpose identification. To the best of our knowledge, this is the first work to apply instruction pre-training on business conversational data, providing insights into how industries can leverage proprietary datasets for domain adaptation.
comment: Accepted to the EMNLP 2025 Industry Track. Equal contribution from the first four authors
☆ AI Knowledge Assist: An Automated Approach for the Creation of Knowledge Bases for Conversational AI Agents EMNLP 2025
The utilization of conversational AI systems by leveraging Retrieval Augmented Generation (RAG) techniques to solve customer problems has been on the rise with the rapid progress of Large Language Models (LLMs). However, the absence of a company-specific dedicated knowledge base is a major barrier to the integration of conversational AI systems in contact centers. To this end, we introduce AI Knowledge Assist, a system that extracts knowledge in the form of question-answer (QA) pairs from historical customer-agent conversations to automatically build a knowledge base. Fine-tuning a lightweight LLM on internal data demonstrates state-of-the-art performance, outperforming larger closed-source LLMs. More specifically, empirical evaluation on 20 companies demonstrates that the proposed AI Knowledge Assist system that leverages the LLaMA-3.1-8B model eliminates the cold-start gap in contact centers by achieving above 90% accuracy in answering information-seeking questions. This enables immediate deployment of RAG-powered chatbots.
comment: Accepted to the EMNLP 2025 Industry Track
☆ Mitigating Judgment Preference Bias in Large Language Models through Group-Based Polling
Large Language Models (LLMs) as automatic evaluators, commonly referred to as LLM-as-a-Judge, have also attracted growing attention. This approach plays a vital role in aligning LLMs with human judgments, providing accurate and reliable assessments. However, LLM-based judgment models often exhibit judgment preference bias during the evaluation phase, tending to favor responses generated by themselves, undermining the reliability of their judgments. This paper introduces the Group-Based Polling Optimization (Genii), an unsupervised multi-agent collaborative optimization framework that mitigates the inherent judgment preference bias of judgment models. Specifically, Genii integrates various LLM-based judgment models into a multi-agent system and simulates the interactive client-server polling mechanism to optimize each client agent unsupervisedly. Our experiments demonstrate that Genii outperforms supervised models trained on annotated judgment data, while requiring no human-labeled annotations. Genii consistently improves performance across different client agents during the polling, even when weaker models act as server agents. Further analysis reveals that Genii effectively mitigates judgment preference bias of LLM-based judgment models, demonstrating its effectiveness. All codes are available at https://github.com/NEUIR/Genii.
☆ Interpreting LLM-as-a-Judge Policies via Verifiable Global Explanations
Using LLMs to evaluate text, that is, LLM-as-a-judge, is increasingly being used at scale to augment or even replace human annotations. As such, it is imperative that we understand the potential biases and risks of doing so. In this work, we propose an approach for extracting high-level concept-based global policies from LLM-as-a-Judge. Our approach consists of two algorithms: 1) CLoVE (Contrastive Local Verifiable Explanations), which generates verifiable, concept-based, contrastive local explanations and 2) GloVE (Global Verifiable Explanations), which uses iterative clustering, summarization and verification to condense local rules into a global policy. We evaluate GloVE on seven standard benchmarking datasets for content harm detection. We find that the extracted global policies are highly faithful to decisions of the LLM-as-a-Judge. Additionally, we evaluated the robustness of global policies to text perturbations and adversarial attacks. Finally, we conducted a user study to evaluate user understanding and satisfaction with global policies.
comment: 12 pages, 2 figures, 3 tables
☆ Can Risk-taking AI-Assistants suitably represent entities
Responsible AI demands systems whose behavioral tendencies can be effectively measured, audited, and adjusted to prevent inadvertently nudging users toward risky decisions or embedding hidden biases in risk aversion. As language models (LMs) are increasingly incorporated into AI-driven decision support systems, understanding their risk behaviors is crucial for their responsible deployment. This study investigates the manipulability of risk aversion (MoRA) in LMs, examining their ability to replicate human risk preferences across diverse economic scenarios, with a focus on gender-specific attitudes, uncertainty, role-based decision-making, and the manipulability of risk aversion. The results indicate that while LMs such as DeepSeek Reasoner and Gemini-2.0-flash-lite exhibit some alignment with human behaviors, notable discrepancies highlight the need to refine bio-centric measures of manipulability. These findings suggest directions for refining AI design to better align human and AI risk preferences and enhance ethical decision-making. The study calls for further advancements in model design to ensure that AI systems more accurately replicate human risk preferences, thereby improving their effectiveness in risk management contexts. This approach could enhance the applicability of AI assistants in managing risk.
☆ Evaluating LLM-Generated Legal Explanations for Regulatory Compliance in Social Media Influencer Marketing EMNLP
The rise of influencer marketing has blurred boundaries between organic content and sponsored content, making the enforcement of legal rules relating to transparency challenging. Effective regulation requires applying legal knowledge with a clear purpose and reason, yet current detection methods of undisclosed sponsored content generally lack legal grounding or operate as opaque "black boxes". Using 1,143 Instagram posts, we compare gpt-5-nano and gemini-2.5-flash-lite under three prompting strategies with controlled levels of legal knowledge provided. Both models perform strongly in classifying content as sponsored or not (F1 up to 0.93), though performance drops by over 10 points on ambiguous cases. We further develop a taxonomy of reasoning errors, showing frequent citation omissions (28.57%), unclear references (20.71%), and hidden ads exhibiting the highest miscue rate (28.57%). While adding regulatory text to the prompt improves explanation quality, it does not consistently improve detection accuracy. The contribution of this paper is threefold. First, it makes a novel addition to regulatory compliance technology by providing a taxonomy of common errors in LLM-generated legal reasoning to evaluate whether automated moderation is not only accurate but also legally robust, thereby advancing the transparent detection of influencer marketing content. Second, it features an original dataset of LLM explanations annotated by two students who were trained in influencer marketing law. Third, it combines quantitative and qualitative evaluation strategies for LLM explanations and critically reflects on how these findings can support advertising regulatory bodies in automating moderation processes on a solid legal foundation.
comment: Accepted for publication at the Natural Legal Language Processing Workshop (NLLP) 2025, co-located with EMNLP
☆ VersionRAG: Version-Aware Retrieval-Augmented Generation for Evolving Documents
Retrieval-Augmented Generation (RAG) systems fail when documents evolve through versioning-a ubiquitous characteristic of technical documentation. Existing approaches achieve only 58-64% accuracy on version-sensitive questions, retrieving semantically similar content without temporal validity checks. We present VersionRAG, a version-aware RAG framework that explicitly models document evolution through a hierarchical graph structure capturing version sequences, content boundaries, and changes between document states. During retrieval, VersionRAG routes queries through specialized paths based on intent classification, enabling precise version-aware filtering and change tracking. On our VersionQA benchmark-100 manually curated questions across 34 versioned technical documents-VersionRAG achieves 90% accuracy, outperforming naive RAG (58%) and GraphRAG (64%). VersionRAG reaches 60% accuracy on implicit change detection where baselines fail (0-10%), demonstrating its ability to track undocumented modifications. Additionally, VersionRAG requires 97% fewer tokens during indexing than GraphRAG, making it practical for large-scale deployment. Our work establishes versioned document QA as a distinct task and provides both a solution and benchmark for future research.
☆ Lossless Vocabulary Reduction for Auto-Regressive Language Models
Tokenization -- the process of decomposing a given text into a sequence of subwords called tokens -- is one of the key components in the development of language models. Particularly, auto-regressive language models generate texts token by token, i.e., by predicting the next-token distribution given the previous ones, and thus tokenization directly affects their efficiency in text generation. Since each language model has their own vocabulary as a set of possible tokens, they struggle to cooperate with each other at the level of next-token distributions such as model ensemble. In this paper, we establish a theoretical framework of lossless vocabulary reduction, which efficiently converts a given auto-regressive language model into the one with an arbitrarily small vocabulary without any loss in accuracy. As an application, we demonstrate that language models with different tokenization can cooperate with each other efficiently through their maximal common vocabulary.
☆ The Price of Thought: A Multilingual Analysis of Reasoning, Performance, and Cost of Negotiation in Large Language Models
Negotiation is a fundamental challenge for AI agents, as it requires an ability to reason strategically, model opponents, and balance cooperation with competition. We conduct the first comprehensive study systematically evaluating the effect of (LLM-)reasoning on the negotiation abilities of both commercial and open-weight LLMs, and do this across three languages. Using a self-play setup across three diverse dialogue games, we analyse trade-offs between performance and cost, the language consistency of reasoning processes, and the nature of strategic adaptation exhibited by models. Our findings show that enabling reasoning-that is, scaling test time compute-significantly improves negotiation outcomes by enhancing collaboration and helping models overcome task complexities, but comes at a substantial computational cost: reasoning improves GPT-5's performance by 31.4 % while increasing its cost by nearly 400 %. Most critically, we uncover a significant multilingual reasoning distinction: open-weight models consistently switch to English for their internal reasoning steps, even when negotiating in German or Italian (and thus possibly impacting potential explainability gains through the disclosure of reasoning traces), while leading commercial models maintain language consistency between their reasoning and final output.
☆ Everything is Plausible: Investigating the Impact of LLM Rationales on Human Notions of Plausibility
We investigate the degree to which human plausibility judgments of multiple-choice commonsense benchmark answers are subject to influence by (im)plausibility arguments for or against an answer, in particular, using rationales generated by LLMs. We collect 3,000 plausibility judgments from humans and another 13,600 judgments from LLMs. Overall, we observe increases and decreases in mean human plausibility ratings in the presence of LLM-generated PRO and CON rationales, respectively, suggesting that, on the whole, human judges find these rationales convincing. Experiments with LLMs reveal similar patterns of influence. Our findings demonstrate a novel use of LLMs for studying aspects of human cognition, while also raising practical concerns that, even in domains where humans are ``experts'' (i.e., common sense), LLMs have the potential to exert considerable influence on people's beliefs.
comment: pre-print
☆ AutoQual: An LLM Agent for Automated Discovery of Interpretable Features for Review Quality Assessment EMNLP 2025
Ranking online reviews by their intrinsic quality is a critical task for e-commerce platforms and information services, impacting user experience and business outcomes. However, quality is a domain-dependent and dynamic concept, making its assessment a formidable challenge. Traditional methods relying on hand-crafted features are unscalable across domains and fail to adapt to evolving content patterns, while modern deep learning approaches often produce black-box models that lack interpretability and may prioritize semantics over quality. To address these challenges, we propose AutoQual, an LLM-based agent framework that automates the discovery of interpretable features. While demonstrated on review quality assessment, AutoQual is designed as a general framework for transforming tacit knowledge embedded in data into explicit, computable features. It mimics a human research process, iteratively generating feature hypotheses through reflection, operationalizing them via autonomous tool implementation, and accumulating experience in a persistent memory. We deploy our method on a large-scale online platform with a billion-level user base. Large-scale A/B testing confirms its effectiveness, increasing average reviews viewed per user by 0.79% and the conversion rate of review readers by 0.27%.
comment: EMNLP 2025
☆ FedDTRE: Federated Dialogue Generation Models Powered by Trustworthiness Evaluation
With the rapid development of artificial intelligence, dialogue systems have become a prominent form of human-computer interaction. However, traditional centralized or fully local training approaches face challenges in balancing privacy preservation and personalization due to data privacy concerns and heterogeneous device capabilities. Federated learning, as a representative distributed paradigm, offers a promising solution. However, existing methods often suffer from overfitting under limited client data and tend to forget global information after multiple training rounds, leading to poor generalization. To address these issues, we propose FedDTRE, a Federated adaptive aggregation strategy for Dialogue generation based on Trustworthiness Evaluation. Instead of directly replacing local models with the global model, FedDTRE leverages trustworthiness scores of both global and local models on a fairness-oriented evaluation dataset to dynamically regulate the global model's contribution during local updates. Experimental results demonstrate that FedDTRE can improve dialogue model performance and enhance the quality of dialogue generation.
☆ A Survey of Process Reward Models: From Outcome Signals to Process Supervisions for Large Language Models
Although Large Language Models (LLMs) exhibit advanced reasoning ability, conventional alignment remains largely dominated by outcome reward models (ORMs) that judge only final answers. Process Reward Models(PRMs) address this gap by evaluating and guiding reasoning at the step or trajectory level. This survey provides a systematic overview of PRMs through the full loop: how to generate process data, build PRMs, and use PRMs for test-time scaling and reinforcement learning. We summarize applications across math, code, text, multimodal reasoning, robotics, and agents, and review emerging benchmarks. Our goal is to clarify design spaces, reveal open challenges, and guide future research toward fine-grained, robust reasoning alignment.
☆ TaoSR-AGRL: Adaptive Guided Reinforcement Learning Framework for E-commerce Search Relevance
Query-product relevance prediction is fundamental to e-commerce search and has become even more critical in the era of AI-powered shopping, where semantic understanding and complex reasoning directly shape the user experience and business conversion. Large Language Models (LLMs) enable generative, reasoning-based approaches, typically aligned via supervised fine-tuning (SFT) or preference optimization methods like Direct Preference Optimization (DPO). However, the increasing complexity of business rules and user queries exposes the inability of existing methods to endow models with robust reasoning capacity for long-tail and challenging cases. Efforts to address this via reinforcement learning strategies like Group Relative Policy Optimization (GRPO) often suffer from sparse terminal rewards, offering insufficient guidance for multi-step reasoning and slowing convergence. To address these challenges, we propose TaoSR-AGRL, an Adaptive Guided Reinforcement Learning framework for LLM-based relevance prediction in Taobao Search Relevance. TaoSR-AGRL introduces two key innovations: (1) Rule-aware Reward Shaping, which decomposes the final relevance judgment into dense, structured rewards aligned with domain-specific relevance criteria; and (2) Adaptive Guided Replay, which identifies low-accuracy rollouts during training and injects targeted ground-truth guidance to steer the policy away from stagnant, rule-violating reasoning patterns toward compliant trajectories. TaoSR-AGRL was evaluated on large-scale real-world datasets and through online side-by-side human evaluations on Taobao Search. It consistently outperforms DPO and standard GRPO baselines in offline experiments, improving relevance accuracy, rule adherence, and training stability. The model trained with TaoSR-AGRL has been successfully deployed in the main search scenario on Taobao, serving hundreds of millions of users.
☆ Pseudo2Real: Task Arithmetic for Pseudo-Label Correction in Automatic Speech Recognition
Robust ASR under domain shift is crucial because real-world systems encounter unseen accents and domains with limited labeled data. Although pseudo-labeling offers a practical workaround, it often introduces systematic, accent-specific errors that filtering fails to fix. We ask: How can we correct these recurring biases without target ground truth? We propose a simple parameter-space correction: in a source domain containing both real and pseudo-labeled data, two ASR models are fine-tuned from the same initialization, one on ground-truth labels and the other on pseudo-labels, and their weight difference forms a correction vector that captures pseudo-label biases. When applied to a pseudo-labeled target model, this vector enhances recognition, achieving up to a 35% relative Word Error Rate (WER) reduction on AfriSpeech-200 across ten African accents with the Whisper tiny model.
☆ Climate Knowledge in Large Language Models
Large language models (LLMs) are increasingly deployed for climate-related applications, where understanding internal climatological knowledge is crucial for reliability and misinformation risk assessment. Despite growing adoption, the capacity of LLMs to recall climate normals from parametric knowledge remains largely uncharacterized. We investigate the capacity of contemporary LLMs to recall climate normals without external retrieval, focusing on a prototypical query: mean July 2-m air temperature 1991-2020 at specified locations. We construct a global grid of queries at 1{\deg} resolution land points, providing coordinates and location descriptors, and validate responses against ERA5 reanalysis. Results show that LLMs encode non-trivial climate structure, capturing latitudinal and topographic patterns, with root-mean-square errors of 3-6 {\deg}C and biases of $\pm$1 {\deg}C. However, spatially coherent errors remain, particularly in mountains and high latitudes. Performance degrades sharply above 1500 m, where RMSE reaches 5-13 {\deg}C compared to 2-4 {\deg}C at lower elevations. We find that including geographic context (country, city, region) reduces errors by 27% on average, with larger models being most sensitive to location descriptors. While models capture the global mean magnitude of observed warming between 1950-1974 and 2000-2024, they fail to reproduce spatial patterns of temperature change, which directly relate to assessing climate change. This limitation highlights that while LLMs may capture present-day climate distributions, they struggle to represent the regional and local expression of long-term shifts in temperature essential for understanding climate dynamics. Our evaluation framework provides a reproducible benchmark for quantifying parametric climate knowledge in LLMs and complements existing climate communication assessments.
comment: 16 pages, 4 figures, 2 tables
☆ ChatGPT as a Translation Engine: A Case Study on Japanese-English
This study investigates ChatGPT for Japanese-English translation, exploring simple and enhanced prompts and comparing against commercially available translation engines. Performing both automatic and MQM-based human evaluations, we found that document-level translation outperforms sentence-level translation for ChatGPT. On the other hand, we were not able to determine if enhanced prompts performed better than simple prompts in our experiments. We also discovered that ChatGPT-3.5 was preferred by automatic evaluation, but a tradeoff exists between accuracy (ChatGPT-3.5) and fluency (ChatGPT-4). Lastly, ChatGPT yields competitive results against two widely-known translation systems.
☆ Learning on the Job: An Experience-Driven Self-Evolving Agent for Long-Horizon Tasks
Large Language Models have demonstrated remarkable capabilities across diverse domains, yet significant challenges persist when deploying them as AI agents for real-world long-horizon tasks. Existing LLM agents suffer from a critical limitation: they are test-time static and cannot learn from experience, lacking the ability to accumulate knowledge and continuously improve on the job. To address this challenge, we propose MUSE, a novel agent framework that introduces an experience-driven, self-evolving system centered around a hierarchical Memory Module. MUSE organizes diverse levels of experience and leverages them to plan and execute long-horizon tasks across multiple applications. After each sub-task execution, the agent autonomously reflects on its trajectory, converting the raw trajectory into structured experience and integrating it back into the Memory Module. This mechanism enables the agent to evolve beyond its static pretrained parameters, fostering continuous learning and self-evolution. We evaluate MUSE on the long-horizon productivity benchmark TAC. It achieves new SOTA performance by a significant margin using only a lightweight Gemini-2.5 Flash model. Sufficient Experiments demonstrate that as the agent autonomously accumulates experience, it exhibits increasingly superior task completion capabilities, as well as robust continuous learning and self-evolution capabilities. Moreover, the accumulated experience from MUSE exhibits strong generalization properties, enabling zero-shot improvement on new tasks. MUSE establishes a new paradigm for AI agents capable of real-world productivity task automation.
☆ Leveraging Author-Specific Context for Scientific Figure Caption Generation: 3rd SciCap Challenge
Scientific figure captions require both accuracy and stylistic consistency to convey visual information. Here, we present a domain-specific caption generation system for the 3rd SciCap Challenge that integrates figure-related textual context with author-specific writing styles using the LaMP-Cap dataset. Our approach uses a two-stage pipeline: Stage 1 combines context filtering, category-specific prompt optimization via DSPy's MIPROv2 and SIMBA, and caption candidate selection; Stage 2 applies few-shot prompting with profile figures for stylistic refinement. Our experiments demonstrate that category-specific prompts outperform both zero-shot and general optimized approaches, improving ROUGE-1 recall by +8.3\% while limiting precision loss to -2.8\% and BLEU-4 reduction to -10.9\%. Profile-informed stylistic refinement yields 40--48\% gains in BLEU scores and 25--27\% in ROUGE. Overall, our system demonstrates that combining contextual understanding with author-specific stylistic adaptation can generate captions that are both scientifically accurate and stylistically faithful to the source paper.
☆ VoiceAgentBench: Are Voice Assistants ready for agentic tasks?
Large-scale Speech Language Models (SpeechLMs) have enabled voice assistants capable of understanding natural spoken queries and performing complex tasks. However, existing speech benchmarks primarily focus on isolated capabilities such as transcription, or question-answering, and do not systematically evaluate agentic scenarios encompassing multilingual and cultural understanding, as well as adversarial robustness. To address this, we introduce VoiceAgentBench, a comprehensive benchmark designed to evaluate SpeechLMs in realistic spoken agentic settings. It comprises over 5,500 synthetic spoken queries, including dialogues grounded in Indian context, covering single-tool invocations, multi-tool workflows, multi-turn interactions, and safety evaluations. The benchmark supports English, Hindi, and 5 other Indian languages, reflecting real-world linguistic and cultural diversity. We simulate speaker variability using a novel sampling algorithm that selects audios for TTS voice conversion based on its speaker embeddings, maximizing acoustic and speaker diversity. Our evaluation measures tool selection accuracy, structural consistency, and the correctness of tool invocations, including adversarial robustness. Our experiments reveal significant gaps in contextual tool orchestration tasks, Indic generalization, and adversarial robustness, exposing critical limitations of current SpeechLMs.
☆ Active Confusion Expression in Large Language Models: Leveraging World Models toward Better Social Reasoning
While large language models (LLMs) excel in mathematical and code reasoning, we observe they struggle with social reasoning tasks, exhibiting cognitive confusion, logical inconsistencies, and conflation between objective world states and subjective belief states. Through deteiled analysis of DeepSeek-R1's reasoning trajectories, we find that LLMs frequently encounter reasoning impasses and tend to output contradictory terms like "tricky" and "confused" when processing scenarios with multiple participants and timelines, leading to erroneous reasoning or infinite loops. The core issue is their inability to disentangle objective reality from agents' subjective beliefs. To address this, we propose an adaptive world model-enhanced reasoning mechanism that constructs a dynamic textual world model to track entity states and temporal sequences. It dynamically monitors reasoning trajectories for confusion indicators and promptly intervenes by providing clear world state descriptions, helping models navigate through cognitive dilemmas. The mechanism mimics how humans use implicit world models to distinguish between external events and internal beliefs. Evaluations on three social benchmarks demonstrate significant improvements in accuracy (e.g., +10% in Hi-ToM) while reducing computational costs (up to 33.8% token reduction), offering a simple yet effective solution for deploying LLMs in social contexts.
comment: 15 pages, 10 figures
☆ LightReasoner: Can Small Language Models Teach Large Language Models Reasoning?
Large language models (LLMs) have demonstrated remarkable progress in reasoning, often through supervised fine-tuning (SFT). However, SFT is resource-intensive, relying on large curated datasets, rejection-sampled demonstrations, and uniform optimization across all tokens, even though only a fraction carry meaningful learning value. In this work, we explore a counterintuitive idea: can smaller language models (SLMs) teach larger language models (LLMs) by revealing high-value reasoning moments that reflect the latter's unique strength? We propose LightReasoner, a novel framework that leverages the behavioral divergence between a stronger expert model (LLM) and a weaker amateur model (SLM). LightReasoner operates in two stages: (1) a sampling stage that pinpoints critical reasoning moments and constructs supervision examples capturing the expert's advantage through expert-amateur contrast, and (2) a fine-tuning stage that aligns the expert model with these distilled examples, amplifying its reasoning strengths. Across seven mathematical benchmarks, LightReasoner improves accuracy by up to 28.1%, while reducing time consumption by 90%, sampled problems by 80%, and tuned token usage by 99%, all without relying on ground-truth labels. By turning weaker SLMs into effective teaching signals, LightReasoner offers a scalable and resource-efficient approach for advancing LLM reasoning. Code is available at: https://github.com/HKUDS/LightReasoner
☆ A$^2$Search: Ambiguity-Aware Question Answering with Reinforcement Learning
Recent advances in Large Language Models (LLMs) and Reinforcement Learning (RL) have led to strong performance in open-domain question answering (QA). However, existing models still struggle with questions that admit multiple valid answers. Standard QA benchmarks, which typically assume a single gold answer, overlook this reality and thus produce inappropriate training signals. Existing attempts to handle ambiguity often rely on costly manual annotation, which is difficult to scale to multi-hop datasets such as HotpotQA and MuSiQue. In this paper, we present A$^2$Search, an annotation-free, end-to-end training framework to recognize and handle ambiguity. At its core is an automated pipeline that detects ambiguous questions and gathers alternative answers via trajectory sampling and evidence verification. The model is then optimized with RL using a carefully designed $\mathrm{AnsF1}$ reward, which naturally accommodates multiple answers. Experiments on eight open-domain QA benchmarks demonstrate that A$^2$Search achieves new state-of-the-art performance. With only a single rollout, A$^2$Search-7B yields an average $\mathrm{AnsF1}@1$ score of $48.4\%$ across four multi-hop benchmarks, outperforming all strong baselines, including the substantially larger ReSearch-32B ($46.2\%$). Extensive analyses further show that A$^2$Search resolves ambiguity and generalizes across benchmarks, highlighting that embracing ambiguity is essential for building more reliable QA systems. Our code, data, and model weights can be found at https://github.com/zfj1998/A2Search
☆ TTOM: Test-Time Optimization and Memorization for Compositional Video Generation
Video Foundation Models (VFMs) exhibit remarkable visual generation performance, but struggle in compositional scenarios (e.g., motion, numeracy, and spatial relation). In this work, we introduce Test-Time Optimization and Memorization (TTOM), a training-free framework that aligns VFM outputs with spatiotemporal layouts during inference for better text-image alignment. Rather than direct intervention to latents or attention per-sample in existing work, we integrate and optimize new parameters guided by a general layout-attention objective. Furthermore, we formulate video generation within a streaming setting, and maintain historical optimization contexts with a parametric memory mechanism that supports flexible operations, such as insert, read, update, and delete. Notably, we found that TTOM disentangles compositional world knowledge, showing powerful transferability and generalization. Experimental results on the T2V-CompBench and Vbench benchmarks establish TTOM as an effective, practical, scalable, and efficient framework to achieve cross-modal alignment for compositional video generation on the fly.
comment: Project page: https://ttom-t2v.github.io/
☆ Vision-Enabled LLMs in Historical Lexicography: Digitising and Enriching Estonian-German Dictionaries from the 17th and 18th Centuries
This article presents research conducted at the Institute of the Estonian Language between 2022 and 2025 on the application of large language models (LLMs) to the study of 17th and 18th century Estonian dictionaries. The authors address three main areas: enriching historical dictionaries with modern word forms and meanings; using vision-enabled LLMs to perform text recognition on sources printed in Gothic script (Fraktur); and preparing for the creation of a unified, cross-source dataset. Initial experiments with J. Gutslaff's 1648 dictionary indicate that LLMs have significant potential for semi-automatic enrichment of dictionary information. When provided with sufficient context, Claude 3.7 Sonnet accurately provided meanings and modern equivalents for 81% of headword entries. In a text recognition experiment with A. T. Helle's 1732 dictionary, a zero-shot method successfully identified and structured 41% of headword entries into error-free JSON-formatted output. For digitising the Estonian-German dictionary section of A. W. Hupel's 1780 grammar, overlapping tiling of scanned image files is employed, with one LLM being used for text recognition and a second for merging the structured output. These findings demonstrate that even for minor languages LLMs have a significant potential for saving time and financial resources.
☆ Comprehensiveness Metrics for Automatic Evaluation of Factual Recall in Text Generation
Despite demonstrating remarkable performance across a wide range of tasks, large language models (LLMs) have also been found to frequently produce outputs that are incomplete or selectively omit key information. In sensitive domains, such omissions can result in significant harm comparable to that posed by factual inaccuracies, including hallucinations. In this study, we address the challenge of evaluating the comprehensiveness of LLM-generated texts, focusing on the detection of missing information or underrepresented viewpoints. We investigate three automated evaluation strategies: (1) an NLI-based method that decomposes texts into atomic statements and uses natural language inference (NLI) to identify missing links, (2) a Q&A-based approach that extracts question-answer pairs and compares responses across sources, and (3) an end-to-end method that directly identifies missing content using LLMs. Our experiments demonstrate the surprising effectiveness of the simple end-to-end approach compared to more complex methods, though at the cost of reduced robustness, interpretability and result granularity. We further assess the comprehensiveness of responses from several popular open-weight LLMs when answering user queries based on multiple sources.
☆ STEPER: Step-wise Knowledge Distillation for Enhancing Reasoning Ability in Multi-Step Retrieval-Augmented Language Models EMNLP 2025
Answering complex real-world questions requires step-by-step retrieval and integration of relevant information to generate well-grounded responses. However, existing knowledge distillation methods overlook the need for different reasoning abilities at different steps, hindering transfer in multi-step retrieval-augmented frameworks. To address this, we propose Stepwise Knowledge Distillation for Enhancing Reasoning Ability in Multi-Step Retrieval-Augmented Language Models (StepER). StepER employs step-wise supervision to align with evolving information and reasoning demands across stages. Additionally, it incorporates difficulty-aware training to progressively optimize learning by prioritizing suitable steps. Our method is adaptable to various multi-step retrieval-augmented language models, including those that use retrieval queries for reasoning paths or decomposed questions. Extensive experiments show that StepER outperforms prior methods on multi-hop QA benchmarks, with an 8B model achieving performance comparable to a 70B teacher model.
comment: EMNLP 2025 Main
☆ Towards Human-Like Grading: A Unified LLM-Enhanced Framework for Subjective Question Evaluation
Automatic grading of subjective questions remains a significant challenge in examination assessment due to the diversity in question formats and the open-ended nature of student responses. Existing works primarily focus on a specific type of subjective question and lack the generality to support comprehensive exams that contain diverse question types. In this paper, we propose a unified Large Language Model (LLM)-enhanced auto-grading framework that provides human-like evaluation for all types of subjective questions across various domains. Our framework integrates four complementary modules to holistically evaluate student answers. In addition to a basic text matching module that provides a foundational assessment of content similarity, we leverage the powerful reasoning and generative capabilities of LLMs to: (1) compare key knowledge points extracted from both student and reference answers, (2) generate a pseudo-question from the student answer to assess its relevance to the original question, and (3) simulate human evaluation by identifying content-related and non-content strengths and weaknesses. Extensive experiments on both general-purpose and domain-specific datasets show that our framework consistently outperforms traditional and LLM-based baselines across multiple grading metrics. Moreover, the proposed system has been successfully deployed in real-world training and certification exams at a major e-commerce enterprise.
☆ ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall
Large Language Models (LLMs) require efficient knowledge editing (KE) to update factual information, yet existing methods exhibit significant performance decay in multi-hop factual recall. This failure is particularly acute when edits involve intermediate implicit subjects within reasoning chains. Through causal analysis, we reveal that this limitation stems from an oversight of how chained knowledge is dynamically represented and utilized at the neuron level. We discover that during multi hop reasoning, implicit subjects function as query neurons, which sequentially activate corresponding value neurons across transformer layers to accumulate information toward the final answer, a dynamic prior KE work has overlooked. Guided by this insight, we propose ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall, a framework that leverages neuron-level attribution to identify and edit these critical query-value (Q-V) pathways. ACE provides a mechanistically grounded solution for multi-hop KE, empirically outperforming state-of-the-art methods by 9.44% on GPT-J and 37.46% on Qwen3-8B. Our analysis further reveals more fine-grained activation patterns in Qwen3 and demonstrates that the semantic interpretability of value neurons is orchestrated by query-driven accumulation. These findings establish a new pathway for advancing KE capabilities based on the principled understanding of internal reasoning mechanisms.
☆ Metric Calculating Benchmark: Code-Verifiable Complicate Instruction Following Benchmark for Large Language Models EMNLP2025
Recent frontier-level LLMs have saturated many previously difficult benchmarks, leaving little room for further differentiation. This progress highlights the need for challenging benchmarks that provide objective verification. In this paper, we introduce MCBench, a benchmark designed to evaluate whether LLMs can execute string-matching NLP metrics by strictly following step-by-step instructions. Unlike prior benchmarks that depend on subjective judgments or general reasoning, MCBench offers an objective, deterministic and codeverifiable evaluation. This setup allows us to systematically test whether LLMs can maintain accurate step-by-step execution, including instruction adherence, numerical computation, and long-range consistency in handling intermediate results. To ensure objective evaluation of these abilities, we provide a parallel reference code that can evaluate the accuracy of LLM output. We provide three evaluative metrics and three benchmark variants designed to measure the detailed instruction understanding capability of LLMs. Our analyses show that MCBench serves as an effective and objective tool for evaluating the capabilities of cutting-edge LLMs.
comment: Accepted to the EMNLP2025
☆ Standard-to-Dialect Transfer Trends Differ across Text and Speech: A Case Study on Intent and Topic Classification in German Dialects
Research on cross-dialectal transfer from a standard to a non-standard dialect variety has typically focused on text data. However, dialects are primarily spoken, and non-standard spellings are known to cause issues in text processing. We compare standard-to-dialect transfer in three settings: text models, speech models, and cascaded systems where speech first gets automatically transcribed and then further processed by a text model. In our experiments, we focus on German and multiple German dialects in the context of written and spoken intent and topic classification. To that end, we release the first dialectal audio intent classification dataset. We find that the speech-only setup provides the best results on the dialect data while the text-only setup works best on the standard data. While the cascaded systems lag behind the text-only models for German, they perform relatively well on the dialectal data if the transcription system generates normalized, standard-like output.
☆ Contrastive Weak-to-strong Generalization
Weak-to-strong generalization provides a promising paradigm for scaling large language models (LLMs) by training stronger models on samples from aligned weaker ones, without requiring human feedback or explicit reward modeling. However, its robustness and generalization are hindered by the noise and biases in weak-model outputs, which limit its applicability in practice. To address this challenge, we leverage implicit rewards, which approximate explicit rewards through log-likelihood ratios, and reveal their structural equivalence with Contrastive Decoding (CD), a decoding strategy shown to reduce noise in LLM generation. Building on this connection, we propose Contrastive Weak-to-Strong Generalization (ConG), a framework that employs contrastive decoding between pre- and post-alignment weak models to generate higher-quality samples. This approach enables more reliable capability transfer, denoising, and improved robustness, substantially mitigating the limitations of traditional weak-to-strong methods. Empirical results across different model families confirm consistent improvements, demonstrating the generality and effectiveness of ConG. Taken together, our findings highlight the potential of ConG to advance weak-to-strong generalization and provide a promising pathway toward AGI.
☆ CS3-Bench: Evaluating and Enhancing Speech-to-Speech LLMs for Mandarin-English Code-Switching
The advancement of multimodal large language models has accelerated the development of speech-to-speech interaction systems. While natural monolingual interaction has been achieved, we find existing models exhibit deficiencies in language alignment. In our proposed Code-Switching Speech-to-Speech Benchmark (CS3-Bench), experiments on 7 mainstream models demonstrate a relative performance drop of up to 66% in knowledge-intensive question answering and varying degrees of misunderstanding in open-ended conversations. Starting from a model with severe performance deterioration, we propose both data constructions and training approaches to improve the language alignment capabilities, specifically employing Chain of Recognition (CoR) to enhance understanding and Keyword Highlighting (KH) to guide generation. Our approach improves the knowledge accuracy from 25.14% to 46.13%, with open-ended understanding rate from 64.5% to 86.5%, and significantly reduces pronunciation errors in the secondary language. CS3-Bench is available at https://huggingface.co/datasets/VocalNet/CS3-Bench.
☆ Do LLMs Really Need 10+ Thoughts for "Find the Time 1000 Days Later"? Towards Structural Understanding of LLM Overthinking
Models employing long chain-of-thought (CoT) reasoning have shown superior performance on complex reasoning tasks. Yet, this capability introduces a critical and often overlooked inefficiency -- overthinking -- models often engage in unnecessarily extensive reasoning even for simple queries, incurring significant computations without accuracy improvements. While prior work has explored solutions to mitigate overthinking, a fundamental gap remains in our understanding of its underlying causes. Most existing analyses are limited to superficial, profiling-based observations, failing to delve into LLMs' inner workings. This study introduces a systematic, fine-grained analyzer of LLMs' thought process to bridge the gap, TRACE. We first benchmark the overthinking issue, confirming that long-thinking models are five to twenty times slower on simple tasks with no substantial gains. We then use TRACE to first decompose the thought process into minimally complete sub-thoughts. Next, by inferring discourse relationships among sub-thoughts, we construct granular thought progression graphs and subsequently identify common thinking patterns for topically similar queries. Our analysis reveals two major patterns for open-weight thinking models -- Explorer and Late Landing. This finding provides evidence that over-verification and over-exploration are the primary drivers of overthinking in LLMs. Grounded in thought structures, we propose a utility-based definition of overthinking, which moves beyond length-based metrics. This revised definition offers a more insightful understanding of LLMs' thought progression, as well as practical guidelines for principled overthinking management.
comment: 30 pages, 41 figures, 10 tables. Preprint
☆ Ready to Translate, Not to Represent? Bias and Performance Gaps in Multilingual LLMs Across Language Families and Domains
The rise of Large Language Models (LLMs) has redefined Machine Translation (MT), enabling context-aware and fluent translations across hundreds of languages and textual domains. Despite their remarkable capabilities, LLMs often exhibit uneven performance across language families and specialized domains. Moreover, recent evidence reveals that these models can encode and amplify different biases present in their training data, posing serious concerns for fairness, especially in low-resource languages. To address these gaps, we introduce Translation Tangles, a unified framework and dataset for evaluating the translation quality and fairness of open-source LLMs. Our approach benchmarks 24 bidirectional language pairs across multiple domains using different metrics. We further propose a hybrid bias detection pipeline that integrates rule-based heuristics, semantic similarity filtering, and LLM-based validation. We also introduce a high-quality, bias-annotated dataset based on human evaluations of 1,439 translation-reference pairs. The code and dataset are accessible on GitHub: https://github.com/faiyazabdullah/TranslationTangles
☆ AdaSwitch: Adaptive Switching Generation for Knowledge Distillation
Small language models (SLMs) are crucial for applications with strict latency and computational constraints, yet achieving high performance remains challenging. Knowledge distillation (KD) can transfer capabilities from large teacher models, but existing methods involve trade-offs: off-policy distillation provides high-quality supervision but introduces a training-inference mismatch, while on-policy approaches maintain consistency but rely on low-quality student outputs. To address these issues, we propose AdaSwitch, a novel approach that dynamically combines on-policy and off-policy generation at the token level. AdaSwitch allows the student to first explore its own predictions and then selectively integrate teacher guidance based on real-time quality assessment. This approach simultaneously preserves consistency and maintains supervision quality. Experiments on three datasets with two teacher-student LLM pairs demonstrate that AdaSwitch consistently improves accuracy, offering a practical and effective method for distilling SLMs with acceptable additional overhead.
☆ Self-Improving LLM Agents at Test-Time
One paradigm of language model (LM) fine-tuning relies on creating large training datasets, under the assumption that high quantity and diversity will enable models to generalize to novel tasks after post-training. In practice, gathering large sets of data is inefficient, and training on them is prohibitively expensive; worse, there is no guarantee that the resulting model will handle complex scenarios or generalize better. Moreover, existing techniques rarely assess whether a training sample provides novel information or is redundant with the knowledge already acquired by the model, resulting in unnecessary costs. In this work, we explore a new test-time self-improvement method to create more effective and generalizable agentic LMs on-the-fly. The proposed algorithm can be summarized in three steps: (i) first it identifies the samples that model struggles with (self-awareness), (ii) then generates similar examples from detected uncertain samples (self-data augmentation), and (iii) uses these newly generated samples at test-time fine-tuning (self-improvement). We study two variants of this approach: Test-Time Self-Improvement (TT-SI), where the same model generates additional training examples from its own uncertain cases and then learns from them, and contrast this approach with Test-Time Distillation (TT-D), where a stronger model generates similar examples for uncertain cases, enabling student to adapt using distilled supervision. Empirical evaluations across different agent benchmarks demonstrate that TT-SI improves the performance with +5.48% absolute accuracy gain on average across all benchmarks and surpasses other standard learning methods, yet using 68x less training samples. Our findings highlight the promise of TT-SI, demonstrating the potential of self-improvement algorithms at test-time as a new paradigm for building more capable agents toward self-evolution.
☆ MetaDefense: Defending Finetuning-based Jailbreak Attack Before and During Generation NeurIPS 2025
This paper introduces MetaDefense, a novel framework for defending against finetuning-based jailbreak attacks in large language models (LLMs). We observe that existing defense mechanisms fail to generalize to harmful queries disguised by unseen attack templates, despite LLMs being capable of distinguishing disguised harmful queries in the embedding space. Based on these insights, we propose a two-stage defense approach: (i) pre-generation defense that detects harmful queries before response generation begins, and (ii) mid-generation defense that monitors partial responses during generation to prevent outputting more harmful content. Our MetaDefense trains the LLM to predict the harmfulness of both queries and partial responses using specialized prompts, enabling early termination of potentially harmful interactions. Extensive experiments across multiple LLM architectures (LLaMA-2-7B, Qwen-2.5-3B-Instruct, and LLaMA-3.2-3B-Instruct) demonstrate that MetaDefense significantly outperforms existing defense mechanisms, achieving robust defense against harmful queries with seen and unseen attack templates while maintaining competitive performance on benign tasks. Code is available at https://github.com/ws-jiang/MetaDefense.
comment: Accepted By NeurIPS 2025
☆ MetricalARGS: A Taxonomy for Studying Metrical Poetry with LLMs
Prior NLP work studying poetry has focused primarily on automatic poem generation and summarization. Many languages have well-studied traditions of poetic meter which enforce constraints on a poem in terms of syllable and phoneme patterns. Such advanced literary forms offer opportunities for probing deeper reasoning and language understanding in Large Language Models (LLMs) and their ability to follow strict pre-requisites and rules. In this paper, we introduce MetricalARGS, the first taxonomy of poetry-related NLP tasks designed to evaluate LLMs on metrical poetry across four dimensions: Analysis, Retrieval, Generation, and Support. We discuss how these tasks relate to existing NLP tasks, addressing questions around datasets and evaluation metrics. Taking Telugu as our example language, we illustrate how the taxonomy can be used in practice. MetricalARGS highlights the broader possibilities for understanding the capabilities and limitations of today's LLMs through the lens of metrical poetry.
comment: Pre-print
♻ ☆ Play to Generalize: Learning to Reason Through Game Play
Developing reasoning capabilities in multimodal large language models (MLLMs) remains challenging. Motivated by literature suggesting that gameplay promotes transferable reasoning skills, we propose a novel post-training method, Visual Game Learning (ViGaL), where MLLMs develop generalizable reasoning skills through playing arcade-like games. Specifically, we show that training a 7B-parameter MLLM via reinforcement learning (RL) on simple games like Snake significantly enhances the downstream performance on multimodal math benchmarks like MathVista, on multi-discipline questions like MMMU and on 3D spatial reasoning benchmarks like VSI-Bench, without seeing any worked solutions, equations, or diagrams during RL. Remarkably, our model outperforms specialist models post-trained on benchmark-oriented multimodal reasoning data, while preserving the model's performance on general visual benchmarks, a challenge where specialist models often fall short. Our findings suggest that multimodal reasoning can emerge from gameplay, pointing to a promising strategy of designing surrogate tasks for RL post-training.
comment: Project Page: https://yunfeixie233.github.io/ViGaL/
♻ ☆ Understanding In-context Learning of Addition via Activation Subspaces
To perform few-shot learning, language models extract signals from a few input-label pairs, aggregate these into a learned prediction rule, and apply this rule to new inputs. How is this implemented in the forward pass of modern transformer models? To explore this question, we study a structured family of few-shot learning tasks for which the true prediction rule is to add an integer $k$ to the input. We introduce a novel optimization method that localizes the model's few-shot ability to only a few attention heads. We then perform an in-depth analysis of individual heads, via dimensionality reduction and decomposition. As an example, on Llama-3-8B-instruct, we reduce its mechanism on our tasks to just three attention heads with six-dimensional subspaces, where four dimensions track the unit digit with trigonometric functions at periods $2$, $5$, and $10$, and two dimensions track magnitude with low-frequency components. To deepen our understanding of the mechanism, we also derive a mathematical identity relating ``aggregation'' and ``extraction'' subspaces for attention heads, allowing us to track the flow of information from individual examples to a final aggregated concept. Using this, we identify a self-correction mechanism where mistakes learned from earlier demonstrations are suppressed by later demonstrations. Our results demonstrate how tracking low-dimensional subspaces of localized heads across a forward pass can provide insight into fine-grained computational structures in language models.
♻ ☆ Zebra-CoT: A Dataset for Interleaved Vision Language Reasoning
Humans often use visual aids, for example diagrams or sketches, when solving complex problems. Training multimodal models to do the same, known as Visual Chain of Thought (Visual CoT), is challenging due to: (1) poor off-the-shelf visual CoT performance, which hinders reinforcement learning, and (2) the lack of high-quality visual CoT training data. We introduce $\textbf{Zebra-CoT}$, a diverse large-scale dataset with 182,384 samples, containing logically coherent interleaved text-image reasoning traces. We focus on four categories of tasks where sketching or visual reasoning is especially natural, spanning scientific questions such as geometry, physics, and algorithms; 2D visual reasoning tasks like visual search and jigsaw puzzles; 3D reasoning tasks including 3D multi-hop inference, embodied and robot planning; visual logic problems and strategic games like chess. Fine-tuning the Anole-7B model on the Zebra-CoT training corpus results in an improvement of +12% in our test-set accuracy and yields up to +13% performance gain on standard VLM benchmark evaluations. Fine-tuning Bagel-7B yields a model that generates high-quality interleaved visual reasoning chains, underscoring Zebra-CoT's effectiveness for developing multimodal reasoning abilities. We open-source our dataset and models to support development and evaluation of visual CoT.
comment: dataset link: https://huggingface.co/datasets/multimodal-reasoning-lab/Zebra-CoT
♻ ☆ Evaluating Evaluation Metrics -- The Mirage of Hallucination Detection EMNLP 2025
Hallucinations pose a significant obstacle to the reliability and widespread adoption of language models, yet their accurate measurement remains a persistent challenge. While many task- and domain-specific metrics have been proposed to assess faithfulness and factuality concerns, the robustness and generalization of these metrics are still untested. In this paper, we conduct a large-scale empirical evaluation of 6 diverse sets of hallucination detection metrics across 4 datasets, 37 language models from 5 families, and 5 decoding methods. Our extensive investigation reveals concerning gaps in current hallucination evaluation: metrics often fail to align with human judgments, take an overtly myopic view of the problem, and show inconsistent gains with parameter scaling. Encouragingly, LLM-based evaluation, particularly with GPT-4, yields the best overall results, and mode-seeking decoding methods seem to reduce hallucinations, especially in knowledge-grounded settings. These findings underscore the need for more robust metrics to understand and quantify hallucinations, and better strategies to mitigate them.
comment: Accepted at EMNLP 2025 Findings (Short)
♻ ☆ Spiffy: Multiplying Diffusion LLM Acceleration via Lossless Speculative Decoding
Diffusion LLMs (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs (AR-LLMs) with the potential to operate at significantly higher token generation rates. However, currently available open-source dLLMs often generate at much lower rates, typically decoding only a single token at every denoising timestep in order to maximize output quality. We present Spiffy, a speculative decoding algorithm that accelerates dLLM inference by $\mathbf{2.8{-}3.1\times}$ while provably preserving the model's output distribution. This work addresses the unique challenges involved in applying ideas from speculative decoding of AR-LLMs to the dLLM setting. Spiffy proposes draft states by leveraging the dLLM's distribution itself in an auto-speculative manner. This approach is efficient and effective, and eliminates the overheads of training and running an independent draft model. To structure the candidate draft states, we propose a novel directed draft graph which is uniquely designed to take advantage of the bidirectional, block-wise nature of dLLM generation and can be verified in parallel by the dLLM. To further optimize the structure of these draft graphs, we introduce an efficient, offline calibration algorithm that procedurally determines high-quality graph configurations. These optimized draft graphs, enabling increased acceptance rates, lead to a significant boost in the overall speedup achieved by the system. Crucially, Spiffy is also complementary to other recent innovations in improving dLLM generation speeds such as KV-caching and multi-token unmasking. We demonstrate that when combined with such parallel decoding algorithms, Spiffy is able to effectively multiply the benefits of these methods leading to total speedups of up to $\mathbf{7.9\times}$.
comment: Original version uploaded on Sep 22, 2025. (v2): Extended Table 2 with additional analysis and referenced it in Sec 5.2
♻ ☆ Evaluating LLMs' Mathematical Reasoning in Financial Document Question Answering
Large Language Models (LLMs), excel in natural language understanding, but their capability for complex mathematical reasoning with an amalgamation of structured tables and unstructured text is uncertain. This study explores LLMs' mathematical reasoning on four financial tabular question-answering datasets: TATQA, FinQA, ConvFinQA, and Multihiertt. Through extensive experiments with various models and prompting techniques, we assess how LLMs adapt to complex tables and mathematical tasks. We focus on sensitivity to table complexity and performance variations with an increasing number of arithmetic reasoning steps. The results provide insights into LLMs' capabilities and limitations in handling complex mathematical scenarios for semi-structured tables. Ultimately, we introduce a novel prompting technique tailored to semi-structured documents, matching or outperforming other baselines in performance while providing a nuanced understanding of LLMs abilities for such a task.
comment: 26 pages, 17 figures
♻ ☆ Paper2Video: Automatic Video Generation from Scientific Papers
Academic presentation videos have become an essential medium for research communication, yet producing them remains highly labor-intensive, often requiring hours of slide design, recording, and editing for a short 2 to 10 minutes video. Unlike natural video, presentation video generation involves distinctive challenges: inputs from research papers, dense multi-modal information (text, figures, tables), and the need to coordinate multiple aligned channels such as slides, subtitles, speech, and human talker. To address these challenges, we introduce Paper2Video, the first benchmark of 101 research papers paired with author-created presentation videos, slides, and speaker metadata. We further design four tailored evaluation metrics--Meta Similarity, PresentArena, PresentQuiz, and IP Memory--to measure how videos convey the paper's information to the audience. Building on this foundation, we propose PaperTalker, the first multi-agent framework for academic presentation video generation. It integrates slide generation with effective layout refinement by a novel effective tree search visual choice, cursor grounding, subtitling, speech synthesis, and talking-head rendering, while parallelizing slide-wise generation for efficiency. Experiments on Paper2Video demonstrate that the presentation videos produced by our approach are more faithful and informative than existing baselines, establishing a practical step toward automated and ready-to-use academic video generation. Our dataset, agent, and code are available at https://github.com/showlab/Paper2Video.
comment: Project Page: https://showlab.github.io/Paper2Video/
♻ ☆ More Than One Teacher: Adaptive Multi-Guidance Policy Optimization for Diverse Exploration
Reinforcement Learning with Verifiable Rewards (RLVR) is a promising paradigm for enhancing the reasoning ability in Large Language Models (LLMs). However, prevailing methods primarily rely on self-exploration or a single off-policy teacher to elicit long chain-of-thought (LongCoT) reasoning, which may introduce intrinsic model biases and restrict exploration, ultimately limiting reasoning diversity and performance. Drawing inspiration from multi-teacher strategies in knowledge distillation, we introduce Adaptive Multi-Guidance Policy Optimization (AMPO), a novel framework that adaptively leverages guidance from multiple proficient teacher models, but only when the on-policy model fails to generate correct solutions. This "guidance-on-demand" approach expands exploration while preserving the value of self-discovery. Moreover, AMPO incorporates a comprehension-based selection mechanism, prompting the student to learn from the reasoning paths that it is most likely to comprehend, thus balancing broad exploration with effective exploitation. Extensive experiments show AMPO substantially outperforms a strong baseline (GRPO), with a 4.3% improvement on mathematical reasoning tasks and 12.2% on out-of-distribution tasks, while significantly boosting Pass@k performance and enabling more diverse exploration. Notably, using four peer-sized teachers, our method achieves comparable results to approaches that leverage a single, more powerful teacher (e.g., DeepSeek-R1) with more data. These results demonstrate a more efficient and scalable path to superior reasoning and generalizability. Our code is available at https://github.com/SII-Enigma/AMPO.
comment: 20 pages, 5 figures
♻ ☆ Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check EMNLP
Downstream scaling laws aim to predict task performance at larger scales from the model's performance at smaller scales. Whether such prediction should be possible is unclear: some works discover clear linear scaling trends after simple transformations of the performance metric, whereas others point out fundamental challenges to downstream scaling laws, such as emergence and inverse scaling. In this work, we conduct a meta-analysis of existing data on downstream scaling laws, and we find that predictable scaling only occurs in a minority of cases: 39% of the time. Moreover, seemingly benign changes to the experimental setting can completely change the scaling behavior. Our analysis underscores the need to understand the conditions under which scaling laws succeed. To accurately model the relationship between pretraining loss and task performance, we must embrace the cases in which scaling behavior deviates from linear trends.
comment: EMNLP Findings 2025 Camera Ready
♻ ☆ Kimi-Dev: Agentless Training as Skill Prior for SWE-Agents
Large Language Models (LLMs) are increasingly applied to software engineering (SWE), with SWE-bench as a key benchmark. Solutions are split into SWE-Agent frameworks with multi-turn interactions and workflow-based Agentless methods with single-turn verifiable steps. We argue these paradigms are not mutually exclusive: reasoning-intensive Agentless training induces skill priors, including localization, code edit, and self-reflection that enable efficient and effective SWE-Agent adaptation. In this work, we first curate the Agentless training recipe and present Kimi-Dev, an open-source SWE LLM achieving 60.4\% on SWE-bench Verified, the best among workflow approaches. With additional SFT adaptation on 5k publicly-available trajectories, Kimi-Dev powers SWE-Agents to 48.6\% pass@1, on par with that of Claude 3.5 Sonnet (241022 version). These results show that structured skill priors from Agentless training can bridge workflow and agentic frameworks for transferable coding agents.
comment: 58 pages
♻ ☆ FlowNIB: An Information Bottleneck Analysis of Bidirectional vs. Unidirectional Language Models
Bidirectional language models have better context understanding and perform better than unidirectional models on natural language understanding tasks, yet the theoretical reasons behind this advantage remain unclear. In this work, we investigate this disparity through the lens of the Information Bottleneck (IB) principle, which formalizes a trade-off between compressing input information and preserving task-relevant content. We propose FlowNIB, a dynamic and scalable method for estimating mutual information during training that addresses key limitations of classical IB approaches, including computational intractability and fixed trade-off schedules. Theoretically, we show that bidirectional models retain more mutual information and exhibit higher effective dimensionality than unidirectional models. To support this, we present a generalized framework for measuring representational complexity and prove that bidirectional representations are strictly more informative under mild conditions. We further validate our findings through extensive experiments across multiple models and tasks using FlowNIB, revealing how information is encoded and compressed throughout training. Together, our work provides a principled explanation for the effectiveness of bidirectional architectures and introduces a practical tool for analyzing information flow in deep language models.
♻ ☆ A Survey of Reinforcement Learning for Large Reasoning Models
In this paper, we survey recent advances in Reinforcement Learning (RL) for reasoning with Large Language Models (LLMs). RL has achieved remarkable success in advancing the frontier of LLM capabilities, particularly in addressing complex logical tasks such as mathematics and coding. As a result, RL has emerged as a foundational methodology for transforming LLMs into LRMs. With the rapid progress of the field, further scaling of RL for LRMs now faces foundational challenges not only in computational resources but also in algorithm design, training data, and infrastructure. To this end, it is timely to revisit the development of this domain, reassess its trajectory, and explore strategies to enhance the scalability of RL toward Artificial SuperIntelligence (ASI). In particular, we examine research applying RL to LLMs and LRMs for reasoning abilities, especially since the release of DeepSeek-R1, including foundational components, core problems, training resources, and downstream applications, to identify future opportunities and directions for this rapidly evolving area. We hope this review will promote future research on RL for broader reasoning models. Github: https://github.com/TsinghuaC3I/Awesome-RL-for-LRMs
comment: Fixed typos; added missing and recent citations (117 -> 120 pages)
♻ ☆ Hybrid Reinforcement: When Reward Is Sparse, It's Better to Be Dense
Post-training for reasoning of large language models (LLMs) increasingly relies on verifiable rewards: deterministic checkers that provide 0-1 correctness signals. While reliable, such binary feedback is brittle--many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning.
comment: 21 pages
♻ ☆ Benchmarking LLM Causal Reasoning with Scientifically Validated Relationships
Causal reasoning is fundamental for Large Language Models (LLMs) to understand genuine cause-and-effect relationships beyond pattern matching. Existing benchmarks suffer from critical limitations such as reliance on synthetic data and narrow domain coverage. We introduce a novel benchmark constructed from casually identified relationships extracted from top-tier economics and finance journals, drawing on rigorous methodologies including instrumental variables, difference-in-differences, and regression discontinuity designs. Our benchmark comprises 40,379 evaluation items covering five task types across domains such as health, environment, technology, law, and culture. Experimental results on eight state-of-the-art LLMs reveal substantial limitations, with the best model achieving only 57.6\% accuracy. Moreover, model scale does not consistently translate to superior performance, and even advanced reasoning models struggle with fundamental causal relationship identification. These findings underscore a critical gap between current LLM capabilities and demands of reliable causal reasoning in high-stakes applications.
♻ ☆ InfiR2: A Comprehensive FP8 Training Recipe for Reasoning-Enhanced Language Models
The immense computational cost of training Large Language Models (LLMs) presents a major barrier to innovation. While FP8 training offers a promising solution with significant theoretical efficiency gains, its widespread adoption has been hindered by the lack of a comprehensive, open-source training recipe. To bridge this gap, we introduce an end-to-end FP8 training recipe that seamlessly integrates continual pre-training and supervised fine-tuning. Our methodology employs a fine-grained, hybrid-granularity quantization strategy to maintain numerical fidelity while maximizing computational efficiency. Through extensive experiments, including the continue pre-training of models on a 160B-token corpus, we demonstrate that our recipe is not only remarkably stable but also essentially lossless, achieving performance on par with the BF16 baseline across a suite of reasoning benchmarks. Crucially, this is achieved with substantial efficiency improvements, including up to a 22% reduction in training time, a 14% decrease in peak memory usage, and a 19% increase in throughput. Our results establish FP8 as a practical and robust alternative to BF16, and we will release the accompanying code to further democratize large-scale model training.
comment: This paper has been withdrawn by the authors due to a significant bug discovered in our data processing pipeline. This bug affects the validity of the experimental results, and we can no longer stand by the conclusions presented
♻ ☆ Disambiguation-Centric Finetuning Makes Enterprise Tool-Calling LLMs More Realistic and Less Risky
Large language models (LLMs) are increasingly tasked with invoking enterprise APIs, yet they routinely falter when near-duplicate tools vie for the same user intent or when required arguments are left underspecified. We introduce DiaFORGE (Dialogue Framework for Organic Response Generation & Evaluation), a disambiguation-centric, three-stage pipeline that (i) synthesizes persona-driven, multi-turn dialogues in which the assistant must distinguish among highly similar tools, (ii) performs supervised fine-tuning of open-source models with reasoning traces across 3B - 70B parameters, and (iii) evaluates real-world readiness via a dynamic suite that redeploys each model in a live agentic loop and reports end-to-end goal completion alongside conventional static metrics. On our dynamic benchmark DiaBENCH, models trained with DiaFORGE raise tool-invocation success by 27 pp over GPT-4o and by 49 pp over Claude-3.5-Sonnet, both under optimized prompting. To spur further research, we release an open corpus of 5000 production-grade enterprise API specifications paired with rigorously validated, disambiguation-focused dialogues, offering a practical blueprint for building reliable, enterprise-ready tool-calling agents.
♻ ☆ CoCoA: Collaborative Chain-of-Agents for Parametric-Retrieved Knowledge Synergy
Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs), especially for knowledge-intensive tasks. Despite its advantages, current RAG methods often struggle to fully exploit knowledge during generation. In particular, the synergy between the model's internal parametric knowledge and external retrieved knowledge remains limited. Retrieved contents may sometimes mislead generation, while certain generated content can guide the model toward more accurate outputs. In this work, we propose Collaborative Chain-of-Agents, a framework designed to enhance explicitly synergy over both parametric and retrieved knowledge. Specifically, we first introduce CoCoA-zero, a multi-agent RAG framework that first performs conditional knowledge induction and then reasons answers. Building on this, we develop CoCoA, a long-chain training strategy that synthesizes extended multi-agent reasoning trajectories from CoCoA-zero to fine-tune the LLM. This strategy enhances the model's capability to explicitly integrate and jointly leverage parametric and retrieved knowledge. Experimental results demonstrate the superiority of CoCoA in open-domain QA and multi-hop QA.
comment: code available at https://github.com/liunian-Jay/CoCoA
♻ ☆ T-VEC: A Telecom-Specific Vectorization Model with Enhanced Semantic Understanding via Deep Triplet Loss Fine-Tuning EMNLP 2025
The specialized vocabulary and nuanced concepts of the telecommunications industry pose persistent challenges for standard Natural Language Processing (NLP) models. Generic embedding models often struggle to represent telecom-specific semantics, limiting their utility in retrieval and downstream tasks. We present T-VEC (Telecom Vectorization Model), a domain-adapted embedding model fine-tuned from the gte-Qwen2-1.5B-instruct backbone using a triplet loss objective. Fine-tuning was performed on T-Embed, a high-quality, large-scale dataset covering diverse telecom concepts, standards, and operational scenarios. Although T-Embed contains some proprietary material and cannot be fully released, we open source 75% of the dataset to support continued research in domain-specific representation learning. On a custom benchmark comprising 1500 query-passage pairs from IETF RFCs and vendor manuals, T-VEC surpasses MPNet, BGE, Jina and E5, demonstrating superior domain grounding and semantic precision in telecom-specific retrieval. Embedding visualizations further showcase tight clustering of telecom-relevant concepts. We release T-VEC and its tokenizer to support semantically faithful NLP applications within the telecom domain.
comment: Accepted to EMNLP 2025 (Industry Track)
♻ ☆ Instructor-Worker Large Language Model System for Policy Recommendation: a Case Study on Air Quality Analysis of the January 2025 Los Angeles Wildfires
The Los Angeles wildfires of January 2025 caused more than 250 billion dollars in damage and lasted for nearly an entire month before containment. Following our previous work, the Digital Twin Building, we modify and leverage the multi-agent large language model framework as well as the cloud-mapping integration to study the air quality during the Los Angeles wildfires. Recent advances in large language models have allowed for out-of-the-box automated large-scale data analysis. We use a multi-agent large language system comprised of an Instructor agent and Worker agents. Upon receiving the users' instructions, the Instructor agent retrieves the data from the cloud platform and produces instruction prompts to the Worker agents. The Worker agents then analyze the data and provide summaries. The summaries are finally input back into the Instructor agent, which then provides the final data analysis. We test this system's capability for data-based policy recommendation by assessing our Instructor-Worker LLM system's health recommendations based on air quality during the Los Angeles wildfires.
♻ ☆ Matryoshka Pilot: Learning to Drive Black-Box LLMs with LLMs NeurIPS 2025
Despite the impressive generative abilities of black-box large language models (LLMs), their inherent opacity hinders further advancements in capabilities such as reasoning, planning, and personalization. Existing works aim to enhance LLM capabilities via domain-specific adaptation, which require additional training on accessible model parameters, an infeasible option for black-box LLMs. To address this challenge, we introduce Matryoshka Pilot (M-Pilot), a lightweight white-box LLM controller that guides a large-scale black-box LLM generator by decomposing complex tasks into a series of intermediate outputs. Specifically, we consider the black-box LLM as an environment, with M-Pilot serving as a policy to provide intermediate guidance through prompts for driving the black-box LLM. M-Pilot is trained to pivot the outputs of the black-box LLM aligning with preferences during iterative interaction, which enables controllable multi-turn generation and self-improvement in optimizing intermediate guidance. Empirical evaluations on diverse tasks demonstrate that our method effectively enhances the capabilities of black-box LLMs in complex, long-horizon tasks.
comment: Accepted by NeurIPS 2025
♻ ☆ Beyond Single Frames: Can LMMs Comprehend Temporal and Contextual Narratives in Image Sequences?
Large Multimodal Models (LMMs) have achieved remarkable success across various visual-language tasks. However, existing benchmarks predominantly focus on single-image understanding, leaving the analysis of image sequences largely unexplored. To address this limitation, we introduce StripCipher, a comprehensive benchmark designed to evaluate capabilities of LMMs to comprehend and reason over sequential images. StripCipher comprises a human-annotated dataset and three challenging subtasks: visual narrative comprehension, contextual frame prediction, and temporal narrative reordering. Our evaluation of 16 state-of-the-art LMMs, including GPT-4o and Qwen2.5VL, reveals a significant performance gap compared to human capabilities, particularly in tasks that require reordering shuffled sequential images. For instance, GPT-4o achieves only 23.93% accuracy in the reordering subtask, which is 56.07% lower than human performance. Further quantitative analysis discuss several factors, such as input format of images, affecting the performance of LLMs in sequential understanding, underscoring the fundamental challenges that remain in the development of LMMs.
♻ ☆ Language Surgery in Multilingual Large Language Models
Large Language Models (LLMs) have demonstrated remarkable generalization capabilities across tasks and languages, revolutionizing natural language processing. This paper investigates the naturally emerging representation alignment in LLMs, particularly in the middle layers, and its implications for disentangling language-specific and language-agnostic information. We empirically confirm the existence of this alignment, analyze its behavior in comparison to explicitly designed alignment models, and demonstrate its potential for language-specific manipulation without semantic degradation. Building on these findings, we propose Inference-Time Language Control (ITLC), a novel method that leverages latent injection to enable precise cross-lingual language control and mitigate language confusion in LLMs. Our experiments highlight ITLC's strong cross-lingual control capabilities while preserving semantic integrity in target languages. Furthermore, we demonstrate its effectiveness in alleviating the cross-lingual language confusion problem, which persists even in current large-scale LLMs, leading to inconsistent language generation. This work advances our understanding of representation alignment in LLMs and introduces a practical solution for enhancing their monolingual and cross-lingual performance.
♻ ☆ Examining Multilingual Embedding Models Cross-Lingually Through LLM-Generated Adversarial Examples EMNLP2025
The evaluation of cross-lingual semantic search models is often limited to existing datasets from tasks such as information retrieval and semantic textual similarity. We introduce Cross-Lingual Semantic Discrimination (CLSD), a lightweight evaluation task that requires only parallel sentences and a Large Language Model (LLM) to generate adversarial distractors. CLSD measures an embedding model's ability to rank the true parallel sentence above semantically misleading but lexically similar alternatives. As a case study, we construct CLSD datasets for German--French in the news domain. Our experiments show that models fine-tuned for retrieval tasks benefit from pivoting through English, whereas bitext mining models perform best in direct cross-lingual settings. A fine-grained similarity analysis further reveals that embedding models differ in their sensitivity to linguistic perturbations. We release our code and datasets under AGPL-3.0: https://github.com/impresso/cross_lingual_semantic_discrimination
comment: To appear in EMNLP2025 Findings
♻ ☆ HiChunk: Evaluating and Enhancing Retrieval-Augmented Generation with Hierarchical Chunking
Retrieval-Augmented Generation (RAG) enhances the response capabilities of language models by integrating external knowledge sources. However, document chunking as an important part of RAG system often lacks effective evaluation tools. This paper first analyzes why existing RAG evaluation benchmarks are inadequate for assessing document chunking quality, specifically due to evidence sparsity. Based on this conclusion, we propose HiCBench, which includes manually annotated multi-level document chunking points, synthesized evidence-dense quetion answer(QA) pairs, and their corresponding evidence sources. Additionally, we introduce the HiChunk framework, a multi-level document structuring framework based on fine-tuned LLMs, combined with the Auto-Merge retrieval algorithm to improve retrieval quality. Experiments demonstrate that HiCBench effectively evaluates the impact of different chunking methods across the entire RAG pipeline. Moreover, HiChunk achieves better chunking quality within reasonable time consumption, thereby enhancing the overall performance of RAG systems.
comment: 17 pages, 5 figures, 6 tables
♻ ☆ Hallucination Detection in LLMs with Topological Divergence on Attention Graphs
Hallucination, i.e., generating factually incorrect content, remains a critical challenge for large language models (LLMs). We introduce TOHA, a TOpology-based HAllucination detector in the RAG setting, which leverages a topological divergence metric to quantify the structural properties of graphs induced by attention matrices. Examining the topological divergence between prompt and response subgraphs reveals consistent patterns: higher divergence values in specific attention heads correlate with hallucinated outputs, independent of the dataset. Extensive experiments - including evaluation on question answering and summarization tasks - show that our approach achieves state-of-the-art or competitive results on several benchmarks while requiring minimal annotated data and computational resources. Our findings suggest that analyzing the topological structure of attention matrices can serve as an efficient and robust indicator of factual reliability in LLMs.
♻ ☆ BiomedSQL: Text-to-SQL for Scientific Reasoning on Biomedical Knowledge Bases
Biomedical researchers increasingly rely on large-scale structured databases for complex analytical tasks. However, current text-to-SQL systems often struggle to map qualitative scientific questions into executable SQL, particularly when implicit domain reasoning is required. We introduce BiomedSQL, the first benchmark explicitly designed to evaluate scientific reasoning in text-to-SQL generation over a real-world biomedical knowledge base. BiomedSQL comprises 68,000 question/SQL query/answer triples generated from templates and grounded in a harmonized BigQuery knowledge base that integrates gene-disease associations, causal inference from omics data, and drug approval records. Each question requires models to infer domain-specific criteria, such as genome-wide significance thresholds, effect directionality, or trial phase filtering, rather than rely on syntactic translation alone. We evaluate a range of open- and closed-source LLMs across prompting strategies and interaction paradigms. Our results reveal a substantial performance gap: GPT-o3-mini achieves 59.0% execution accuracy, while our custom multi-step agent, BMSQL, reaches 62.6%, both well below the expert baseline of 90.0%. BiomedSQL provides a new foundation for advancing text-to-SQL systems capable of supporting scientific discovery through robust reasoning over structured biomedical knowledge bases. Our dataset is publicly available at https://huggingface.co/datasets/NIH-CARD/BiomedSQL, and our code is open-source at https://github.com/NIH-CARD/biomedsql.
comment: Under Review
♻ ☆ Can Small-Scale Data Poisoning Exacerbate Dialect-Linked Biases in Large Language Models?
Style-conditioned data poisoning is identified as a covert vector for amplifying sociolinguistic bias in large language models. Using small poisoned budgets that pair dialectal prompts -- principally African American Vernacular English (AAVE) and a Southern dialect -- with toxic or stereotyped completions during instruction tuning, this work probes whether linguistic style can act as a latent trigger for harmful behavior. Across multiple model families and scales, poisoned exposure elevates toxicity and stereotype expression for dialectal inputs -- most consistently for AAVE -- while Standard American English remains comparatively lower yet not immune. A multi-metric audit combining classifier-based toxicity with an LLM-as-a-judge reveals stereotype-laden content even when lexical toxicity appears muted, indicating that conventional detectors under-estimate sociolinguistic harms. Additionally, poisoned models exhibit emergent jailbreaking despite the absence of explicit slurs in the poison, suggesting weakened alignment rather than memorization. These findings underscore the need for dialect-aware evaluation, content-level stereotype auditing, and training protocols that explicitly decouple style from toxicity to prevent bias amplification through seemingly minor, style-based contamination.
♻ ☆ ThinkNote: Enhancing Knowledge Integration and Utilization of Large Language Models via Constructivist Cognition Modeling
Large Language Models (LLMs) have demonstrated strong performance across a wide range of NLP tasks. However, they often exhibit suboptimal behaviors and inconsistencies when exposed to unfamiliar external information, underscoring their limitations in effectively leveraging such knowledge. Inspired by constructivist learning theory, we propose ThinkNote, a novel framework that enhances the external knowledge utilization of LLMs through a two-stage constructivist cognitive modeling process. Specifically, ThinkNote performs knowledge assimilation to align new information with the model's parametric memory, forming a coherent internal representation. It then applies thought accommodation to adapt internal reasoning, thereby promoting more consistent and reliable outputs. Extensive experimental results demonstrate that ThinkNote achieves a 10% improvement over strong baseline methods on various question-answering benchmarks. Further analysis indicates that ThinkNote effectively integrates and utilizes external knowledge to help LLMs generate accurate responses and improves their self-consistency. All data and codes are available at https://github.com/OpenMatch/ThinkNote.
♻ ☆ QoNext: Towards Next-generation QoE for Foundation Models
Existing evaluations of foundation models, including recent human-centric approaches, fail to capture what truly matters: user's experience during interaction. Current methods treat evaluation as a matter of output correctness alone, overlooking that user satisfaction emerges from the interplay between response quality and interaction, which limits their ability to account for the mechanisms underlying user experience. To address this gap, we introduce QoNext, the first framework that adapts Quality of Experience (QoE) principles from networking and multimedia to the assessment of foundation models. QoNext identifies experiential factors that shape user experience and incorporates them into controlled experiments, where human ratings are collected under varied configurations. From these studies we construct a QoE-oriented database and train predictive models that estimate perceived user experience from measurable system parameters. Our results demonstrate that QoNext not only enables proactive and fine-grained evaluation but also provides actionable guidance for productized services of optimizing foundation models in practice.
♻ ☆ Multi-Trigger Poisoning Amplifies Backdoor Vulnerabilities in LLMs
Recent studies have shown that Large Language Models (LLMs) are vulnerable to data poisoning attacks, where malicious training examples embed hidden behaviours triggered by specific input patterns. However, most existing works assume a phrase and focus on the attack's effectiveness, offering limited understanding of trigger mechanisms and how multiple triggers interact within the model. In this paper, we present a framework for studying poisoning in LLMs. We show that multiple distinct backdoor triggers can coexist within a single model without interfering with each other, enabling adversaries to embed several triggers concurrently. Using multiple triggers with high embedding similarity, we demonstrate that poisoned triggers can achieve robust activation even when tokens are substituted or separated by long token spans. Our findings expose a broader and more persistent vulnerability surface in LLMs. To mitigate this threat, we propose a post hoc recovery method that selectively retrains specific model components based on a layer-wise weight difference analysis. Our method effectively removes the trigger behaviour with minimal parameter updates, presenting a practical and efficient defence against multi-trigger poisoning.
♻ ☆ DACP: Domain-Adaptive Continual Pre-Training of Large Language Models for Phone Conversation Summarization EMNLP 2025
Large language models (LLMs) have achieved impressive performance in text summarization, yet their performance often falls short when applied to specialized domains that differ from their original pre-training distribution. While fine-tuning can improve summarization quality, it typically relies on costly and scarce high-quality labeled data. In this work, we explore continual pre-training as a scalable, self-supervised approach to adapt LLMs for downstream summarization tasks, particularly in the context of noisy real-world conversation transcripts. We conduct extensive experiments using large-scale, unlabeled business conversation data to investigate whether continual pre-training enhances model capabilities in conversational summarization. Our results demonstrate that continual pre-training yields substantial gains in both in-domain and out-of-domain summarization benchmarks, while maintaining strong generalization and robustness. We also analyze the effects of data selection strategies, providing practical guidelines for applying continual pre-training in summarization-focused industrial applications.
comment: Accepted to the NewSumm Workshop at EMNLP 2025. Equal contribution from the first four authors
♻ ☆ Dissecting Logical Reasoning in LLMs: A Fine-Grained Evaluation and Supervision Study EMNLP 2025
Logical reasoning is a core capability for large language models (LLMs), yet existing benchmarks that rely solely on final-answer accuracy fail to capture the quality of the reasoning process. To address this, we introduce FineLogic, a fine-grained evaluation framework that assesses logical reasoning across three dimensions: overall accuracy, stepwise soundness, and representation-level probing. Leveraging this framework, we conduct a comprehensive study on how different supervision formats in fine-tuning shape reasoning abilities. We fine-tune LLMs on four supervision styles: one in natural language and three symbolic variants. We find a key trade-off: natural language supervision excels at generalization to out-of-distribution and long-chain problems, whereas symbolic supervision is superior at instilling structurally sound, atomic reasoning steps. Furthermore, our probing analysis indicates that fine-tuning primarily refines the model's step-by-step generation process, rather than improving its ability to converge on an answer early. Together, our framework and analysis provide a more rigorous lens for evaluating and improving logical reasoning in LLMs. The code is available at https://github.com/YujunZhou/FineLogic.
comment: Accepted by the Findings of EMNLP 2025
♻ ☆ OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework
Large Language Models (LLMs) fine-tuned via Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning with Verifiable Rewards (RLVR) significantly improve the alignment of human-AI values, further raising the upper bound of AI capabilities, particularly in reasoning-intensive, long-context Chain-of-Thought (CoT) tasks. However, existing frameworks commonly face challenges such as inference bottlenecks and complexity barriers, which restrict their accessibility to newcomers. To bridge this gap, we introduce \textbf{OpenRLHF}, a user-friendly, scalable, and easy-to-learn open-source RLHF framework built upon Ray, vLLM, DeepSpeed, and HuggingFace Transformers, featuring a simplified design, clear code structure, and comprehensive documentation to facilitate entry for researchers and practitioners. Experimental results show that OpenRLHF achieves superior training efficiency, with speedups ranging from 1.22x to 1.68x across different model sizes, compared to state-of-the-art frameworks. Additionally, it requires significantly fewer lines of code for implementation. OpenRLHF is publicly available at https://github.com/OpenRLHF/OpenRLHF, and has already been adopted by leading institutions to accelerate RLHF research and learning.
comment: update template
♻ ☆ Distilling a Small Utility-Based Passage Selector to Enhance Retrieval-Augmented Generation SIGIR
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating retrieved information. Standard retrieval process prioritized relevance, focusing on topical alignment between queries and passages. In contrast, in RAG, the emphasis has shifted to utility, which considers the usefulness of passages for generating accurate answers. Despite empirical evidence showing the benefits of utility-based retrieval in RAG, the high computational cost of using LLMs for utility judgments limits the number of passages evaluated. This restriction is problematic for complex queries requiring extensive information. To address this, we propose a method to distill the utility judgment capabilities of LLMs into smaller, more efficient models. Our approach focuses on utility-based selection rather than ranking, enabling dynamic passage selection tailored to specific queries without the need for fixed thresholds. We train student models to learn pseudo-answer generation and utility judgments from teacher LLMs, using a sliding window method that dynamically selects useful passages. Our experiments demonstrate that utility-based selection provides a flexible and cost-effective solution for RAG, significantly reducing computational costs while improving answer quality. We present the distillation results using Qwen3-32B as the teacher model for both relevance ranking and utility-based selection, distilled into RankQwen1.7B and UtilityQwen1.7B. Our findings indicate that for complex questions, utility-based selection is more effective than relevance ranking in enhancing answer generation performance. We will release the relevance ranking and utility-based selection annotations for the MS MARCO dataset, supporting further research in this area.
comment: Accepted by SIGIR-AP25
♻ ☆ TokenSelect: Efficient Long-Context Inference and Length Extrapolation for LLMs via Dynamic Token-Level KV Cache Selection EMNLP2025
Rapid advances in Large Language Models (LLMs) have spurred demand for processing extended context sequences in contemporary applications. However, this progress faces two challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues limit LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using QK dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a few critical KV cache tokens in attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we design the Selection Cache based on observations of consecutive Query similarity and implemented the efficient Paged Dot Product Kernel, significantly reducing the selection overhead. A comprehensive evaluation of TokenSelect demonstrates up to $23.84\times$ speedup in attention computation and up to $2.28\times$ acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.
comment: Accepted by EMNLP2025
♻ ☆ Flora: Effortless Context Construction to Arbitrary Length and Scale
Effectively handling long contexts is challenging for Large Language Models (LLMs) due to the rarity of long texts, high computational demands, and substantial forgetting of short-context abilities. Recent approaches have attempted to construct long contexts for instruction tuning, but these methods often require LLMs or human interventions, which are both costly and limited in length and diversity. Also, the drop in short-context performances of present long-context LLMs remains significant. In this paper, we introduce Flora, an effortless (human/LLM-free) long-context construction strategy. Flora can markedly enhance the long-context performance of LLMs by arbitrarily assembling short instructions based on categories and instructing LLMs to generate responses based on long-context meta-instructions. This enables Flora to produce contexts of arbitrary length and scale with rich diversity, while only slightly compromising short-context performance. Experiments on Llama3-8B-Instruct and QwQ-32B show that LLMs enhanced by Flora excel in three long-context benchmarks while maintaining strong performances in short-context tasks. Our data-construction code is available at \href{https://github.com/txchen-USTC/Flora}{https://github.com/txchen-USTC/Flora}.
♻ ☆ Think With Videos For Agentic Long-Video Understanding
Long-video understanding~(LVU) is a challenging problem in computer vision. Existing methods either downsample frames for single-pass reasoning, sacrificing fine-grained details, or depend on textual reasoning over task-agnostic representations, hindering task-specific perception and exploration. In this paper, we propose VideoExplorer, a framework grounded in the principle of ``thinking with video'', which naturally intertwines planning, temporal grounding, and scalable perception into a coherent reasoning process. Rather than reasoning over a static context, VideoExplorer iteratively formulates sub-questions, locates relevant moments, and performs task-oriented, temporally scalable video understanding until reaching the final answer, enabling faithful, efficient, and interpretable reasoning. To address the lack of LVU training resources, we construct a long-video reasoning dataset using difficulty-adaptive sampling to ensure high-quality trajectories on complex tasks. Building on this dataset, we design a two-stage training pipeline: supervised trajectory initialization followed by trajectory-level preference optimization, encouraging adaptive temporal grounding and iterative information integration guided by downstream rewards. Extensive evaluations on popular long-video understanding and reasoning benchmarks demonstrate VideoExplorer's significant advantage over existing baselines, highlighting its robustness, adaptability, and efficiency. Our code is made publicly available in this repository(https://github.com/yhy-2000/VideoDeepResearch).
♻ ☆ UNCLE: Benchmarking Uncertainty Expressions in Long-Form Generation EMNLP 2025
Large Language Models (LLMs) are prone to hallucination, particularly in long-form generations. A promising direction to mitigate hallucination is to teach LLMs to express uncertainty explicitly when they lack sufficient knowledge. However, existing work lacks direct and fair evaluation of LLMs' ability to express uncertainty effectively in long-form generation. To address this gap, we first introduce UNCLE, a benchmark designed to evaluate uncertainty expression in both long- and short-form question answering (QA). UNCLE covers five domains and includes more than 1,000 entities, each with paired short- and long-form QA items. Our dataset is the first to directly link short- and long-form QA through aligned questions and gold-standard answers. Along with UNCLE, we propose a suite of new metrics to assess the models' capabilities to selectively express uncertainty. We then demonstrate that current models fail to convey uncertainty appropriately in long-form generation. We further explore both prompt-based and training-based methods to improve models' performance, with the training-based methods yielding greater gains. Further analysis of alignment gaps between short- and long-form uncertainty expression highlights promising directions for future research using UNCLE.
comment: EMNLP 2025 Main
♻ ☆ Tug-of-war between idioms' figurative and literal interpretations in LLMs
Idioms present a unique challenge for language models due to their non-compositional figurative interpretations, which often strongly diverge from the idiom's literal interpretation. In this paper, we employ causal tracing to systematically analyze how pretrained causal transformers deal with this ambiguity. We localize three mechanisms: (i) Early sublayers and specific attention heads retrieve an idiom's figurative interpretation, while suppressing its literal interpretation. (ii) When disambiguating context precedes the idiom, the model leverages it from the earliest layer and later layers refine the interpretation if the context conflicts with the retrieved interpretation. (iii) Then, selective, competing pathways carry both interpretations: an intermediate pathway prioritizes the figurative interpretation and a parallel direct route favors the literal interpretation, ensuring that both readings remain available. Our findings provide mechanistic evidence for idiom comprehension in autoregressive transformers.
♻ ☆ Breaking the Reviewer: Assessing the Vulnerability of Large Language Models in Automated Peer Review Under Textual Adversarial Attacks
Peer review is essential for maintaining academic quality, but the increasing volume of submissions places a significant burden on reviewers. Large language models (LLMs) offer potential assistance in this process, yet their susceptibility to textual adversarial attacks raises reliability concerns. This paper investigates the robustness of LLMs used as automated reviewers in the presence of such attacks. We focus on three key questions: (1) The effectiveness of LLMs in generating reviews compared to human reviewers. (2) The impact of adversarial attacks on the reliability of LLM-generated reviews. (3) Challenges and potential mitigation strategies for LLM-based review. Our evaluation reveals significant vulnerabilities, as text manipulations can distort LLM assessments. We offer a comprehensive evaluation of LLM performance in automated peer reviewing and analyze its robustness against adversarial attacks. Our findings emphasize the importance of addressing adversarial risks to ensure AI strengthens, rather than compromises, the integrity of scholarly communication.
comment: Minor correction: Fixed sign errors in the results table. The update does not affect the main findings or conclusions
♻ ☆ Depression Detection on Social Media with Large Language Models EMNLP 2025
Limited access to mental healthcare resources hinders timely depression diagnosis, leading to detrimental outcomes. Social media platforms present a valuable data source for early detection, yet this task faces two significant challenges: 1) the need for medical knowledge to distinguish clinical depression from transient mood changes, and 2) the dual requirement for high accuracy and model explainability. To address this, we propose DORIS, a framework that leverages Large Language Models (LLMs). To integrate medical knowledge, DORIS utilizes LLMs to annotate user texts against established medical diagnostic criteria and to summarize historical posts into temporal mood courses. These medically-informed features are then used to train an accurate Gradient Boosting Tree (GBT) classifier. Explainability is achieved by generating justifications for predictions based on the LLM-derived symptom annotations and mood course analyses. Extensive experimental results validate the effectiveness as well as interpretability of our method, highlighting its potential as a supportive clinical tool.
comment: EMNLP 2025
♻ ☆ Foundations of LLM Knowledge Materialization: Termination, Reproducibility, Robustness
Large Language Models (LLMs) encode substantial factual knowledge, yet measuring and systematizing this knowledge remains challenging. Converting it into structured format, for example through recursive extraction approaches such as the GPTKB methodology (Hu et al., 2025b), is still underexplored. Key open questions include whether such extraction can terminate, whether its outputs are reproducible, and how robust they are to variations. We systematically study LLM knowledge materialization using miniGPTKBs (domain-specific, tractable subcrawls), analyzing termination, reproducibility, and robustness across three categories of metrics: yield, lexical similarity, and semantic similarity. We experiment with four variations (seed, language, randomness, model) and three illustrative domains (from history, entertainment, and finance). Our findings show (i) high termination rates, though model-dependent; (ii) mixed reproducibility; and (iii) robustness that varies by perturbation type: high for seeds and temperature, lower for languages and models. These results suggest that LLM knowledge materialization can reliably surface core knowledge, while also revealing important limitations.
♻ ☆ Large Language Models Hallucination: A Comprehensive Survey
Large language models (LLMs) have transformed natural language processing, achieving remarkable performance across diverse tasks. However, their impressive fluency often comes at the cost of producing false or fabricated information, a phenomenon known as hallucination. Hallucination refers to the generation of content by an LLM that is fluent and syntactically correct but factually inaccurate or unsupported by external evidence. Hallucinations undermine the reliability and trustworthiness of LLMs, especially in domains requiring factual accuracy. This survey provides a comprehensive review of research on hallucination in LLMs, with a focus on causes, detection, and mitigation. We first present a taxonomy of hallucination types and analyze their root causes across the entire LLM development lifecycle, from data collection and architecture design to inference. We further examine how hallucinations emerge in key natural language generation tasks. Building on this foundation, we introduce a structured taxonomy of detection approaches and another taxonomy of mitigation strategies. We also analyze the strengths and limitations of current detection and mitigation approaches and review existing evaluation benchmarks and metrics used to quantify LLMs hallucinations. Finally, we outline key open challenges and promising directions for future research, providing a foundation for the development of more truthful and trustworthy LLMs.
♻ ☆ DNA-DetectLLM: Unveiling AI-Generated Text via a DNA-Inspired Mutation-Repair Paradigm NeurIPS 2025
The rapid advancement of large language models (LLMs) has blurred the line between AI-generated and human-written text. This progress brings societal risks such as misinformation, authorship ambiguity, and intellectual property concerns, highlighting the urgent need for reliable AI-generated text detection methods. However, recent advances in generative language modeling have resulted in significant overlap between the feature distributions of human-written and AI-generated text, blurring classification boundaries and making accurate detection increasingly challenging. To address the above challenges, we propose a DNA-inspired perspective, leveraging a repair-based process to directly and interpretably capture the intrinsic differences between human-written and AI-generated text. Building on this perspective, we introduce DNA-DetectLLM, a zero-shot detection method for distinguishing AI-generated and human-written text. The method constructs an ideal AI-generated sequence for each input, iteratively repairs non-optimal tokens, and quantifies the cumulative repair effort as an interpretable detection signal. Empirical evaluations demonstrate that our method achieves state-of-the-art detection performance and exhibits strong robustness against various adversarial attacks and input lengths. Specifically, DNA-DetectLLM achieves relative improvements of 5.55% in AUROC and 2.08% in F1 score across multiple public benchmark datasets. Code and data are available at https://github.com/Xiaoweizhu57/DNA-DetectLLM.
comment: NeurIPS 2025 Spotlight
♻ ☆ Inference-time Alignment in Continuous Space NeurIPS 2025
Aligning large language models with human feedback at inference time has received increasing attention due to its flexibility. Existing methods rely on generating multiple responses from the base policy for search using a reward model, which can be considered as searching in a discrete response space. However, these methods struggle to explore informative candidates when the base policy is weak or the candidate set is small, resulting in limited effectiveness. In this paper, to address this problem, we propose Simple Energy Adaptation ($\textbf{SEA}$), a simple yet effective algorithm for inference-time alignment. In contrast to expensive search over the discrete space, SEA directly adapts original responses from the base policy toward the optimal one via gradient-based sampling in continuous latent space. Specifically, SEA formulates inference as an iterative optimization procedure on an energy function over actions in the continuous space defined by the optimal policy, enabling simple and effective alignment. For instance, despite its simplicity, SEA outperforms the second-best baseline with a relative improvement of up to $ \textbf{77.51%}$ on AdvBench and $\textbf{16.36%}$ on MATH. Our code is publicly available at https://github.com/yuanyige/sea
comment: Accepted at NeurIPS 2025
♻ ☆ Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning EMNLP 2025
Supervised Fine-Tuning (SFT) Large Language Models (LLM) fundamentally rely on high-quality training data. While data selection and data synthesis are two common strategies to improve data quality, existing approaches often face limitations in static dataset curation that fail to adapt to evolving model capabilities. In this paper, we introduce Middo, a self-evolving Model-informed dynamic data optimization framework that uses model-aware data selection and context-preserving data refinement. Unlike conventional one-off filtering/synthesis methods, our framework establishes a closed-loop optimization system: (1) A self-referential diagnostic module proactively identifies suboptimal samples through tri-axial model signals - loss patterns (complexity), embedding cluster dynamics (diversity), and self-alignment scores (quality); (2) An adaptive optimization engine then transforms suboptimal samples into pedagogically valuable training points while preserving semantic integrity; (3) This optimization process continuously evolves with model capability through dynamic learning principles. Experiments on multiple benchmarks demonstrate that our Middo consistently enhances the quality of seed data and boosts LLM's performance with improving accuracy by 7.15% on average while maintaining the original dataset scale. This work establishes a new paradigm for sustainable LLM training through dynamic human-AI co-evolution of data and models. Our datasets, models, and code are publicly available at https://github.com/Word2VecT/Middo.
comment: Accepted by EMNLP 2025 (Main)
♻ ☆ Spotlight Attention: Towards Efficient LLM Generation via Non-linear Hashing-based KV Cache Retrieval
Reducing the key-value (KV) cache burden in Large Language Models (LLMs) significantly accelerates inference. Dynamically selecting critical KV caches during decoding helps maintain performance. Existing methods use random linear hashing to identify important tokens, but this approach is inefficient due to the orthogonal distribution of queries and keys within two narrow cones in LLMs. We introduce Spotlight Attention, a novel method that employs non-linear hashing functions to optimize the embedding distribution of queries and keys, enhancing coding efficiency and robustness. We also developed a lightweight, stable training framework using a Bradley-Terry ranking-based loss, enabling optimization of the non-linear hashing module on GPUs with 16GB memory in 8 hours. Experimental results show that Spotlight Attention drastically improves retrieval precision while shortening the length of the hash code at least 5$\times$ compared to traditional linear hashing. Finally, we exploit the computational advantages of bitwise operations by implementing specialized CUDA kernels, achieving hashing retrieval for 512K tokens in under 100$\mu$s on a single A100 GPU, with end-to-end throughput up to 3$\times$ higher than vanilla decoding.
♻ ☆ Trans-EnV: A Framework for Evaluating the Linguistic Robustness of LLMs Against English Varieties NeurIPS 2025
Large Language Models (LLMs) are predominantly evaluated on Standard American English (SAE), often overlooking the diversity of global English varieties. This narrow focus may raise fairness concerns as degraded performance on non-standard varieties can lead to unequal benefits for users worldwide. Therefore, it is critical to extensively evaluate the linguistic robustness of LLMs on multiple non-standard English varieties. We introduce Trans-EnV, a framework that automatically transforms SAE datasets into multiple English varieties to evaluate the linguistic robustness. Our framework combines (1) linguistics expert knowledge to curate variety-specific features and transformation guidelines from linguistic literature and corpora, and (2) LLM-based transformations to ensure both linguistic validity and scalability. Using Trans-EnV, we transform six benchmark datasets into 38 English varieties and evaluate seven state-of-the-art LLMs. Our results reveal significant performance disparities, with accuracy decreasing by up to 46.3% on non-standard varieties. These findings highlight the importance of comprehensive linguistic robustness evaluation across diverse English varieties. Each construction of Trans-EnV was validated through rigorous statistical testing and consultation with a researcher in the field of second language acquisition, ensuring its linguistic validity. Our code and datasets are publicly available at https://github.com/jiyounglee-0523/TransEnV and https://huggingface.co/collections/jiyounglee0523/transenv-681eadb3c0c8cf363b363fb1.
comment: NeurIPS 2025 Track on Datasets and Benchmarks (27 pages, 6 figures, 16 tables)
♻ ☆ Can LLMs Grasp Implicit Cultural Values? Benchmarking LLMs' Cultural Intelligence with CQ-Bench
Cultural Intelligence (CQ) refers to the ability to understand unfamiliar cultural contexts, a crucial skill for large language models (LLMs) to effectively engage with globally diverse users. Existing studies often focus on explicitly stated cultural norms, but fail to capture the subtle, implicit values that are common in daily conversation. To address this gap, we introduce CQBench, a benchmark specifically designed to assess LLMs' capability to infer implicit cultural values from natural conversational contexts. CQBench consists of multi character conversation based stories using values from the World Value Survey and the GlobalOpinions, with topics including ethical, religious, social, etc. Our automatic dataset construction pipeline integrates rigorous validation procedures (incorporation, consistency, and implicitness checks), achieving a 94.5% human model agreement in the final validation. To leverage CQBench data, we design three tasks of increasing complexity: attitude detection, value selection, and value extraction. These tasks evaluate whether models can detect attitude and recognize values embedded within natural dialogues rather than relying on explicit cultural knowledge. We find that while frontier models like o1 reach human level performance in value selection (0.809 F1), they still fall short in nuanced attitude detection (0.622 F1). Notably, finetuning a smaller LLaMA-3.2-3B on only 500 culturally rich examples improves performance by over 10%, even outperforming o3-mini in some cases. Using CQ-Bench, we provide insights into the current challenges in LLMs' CQ research and suggest practical pathways for enhancing LLMs' cross-cultural reasoning abilities.
♻ ☆ Med-R$^3$: Enhancing Medical Retrieval-Augmented Reasoning of LLMs via Progressive Reinforcement Learning
In medical scenarios, effectively retrieving external knowledge and leveraging it for rigorous logical reasoning is of significant importance. Despite their potential, existing work has predominantly focused on enhancing either retrieval or reasoning capabilities of the models in isolation, with little attention given to their joint optimization, which leads to limited coordination between the two processes. Additionally, current methods rely heavily on supervised fine-tuning (SFT), which can cause models to memorize existing problem-solving pathways, thereby restricting their generalization ability when confronted with novel problem contexts. Furthermore, while some studies have explored to improve retrieval-augmented reasoning in general domains via reinforcement learning, their reward function designs do not adequately capture the specific demands of the medical domain. To address these challenges, we introduce **Med-R$^3$**, a **Med**ical **R**etrieval-augmented **R**easoning framework driven by progressive **R**einforcement learning. In this framework, we first develop the model's ability to perform logical reasoning over medical problems. Subsequently, on the basis of this foundation, we adaptively optimize the retrieval capability to better align with the characteristics of knowledge corpus and external information utilization throughout the reasoning process. Finally, we conduct joint optimization of the model's retrieval and reasoning coordination. Extensive experiments indicate that **Med-R$^3$** could achieve state-of-the-art performances, with LLaMA3.1-8B-Instruct + Med-R$^3$ surpassing closed-sourced GPT-4o-mini by 3.93\% at a comparable parameter scale, while Qwen2.5-14B augmented with Med-R$^3$ shows a more substantial gain of 13.53\%.
♻ ☆ Efficiency-Effectiveness Reranking FLOPs for LLM-based Rerankers EMNLP
Large Language Models (LLMs) have recently been applied to reranking tasks in information retrieval, achieving strong performance. However, their high computational demands often hinder practical deployment. Existing studies evaluate the efficiency of LLM-based rerankers using proxy metrics such as latency, the number of forward passes, input tokens, and output tokens. However, these metrics depend on hardware and running-time choices (\eg parallel or not, batch size, etc), and often fail to account for model size, making it difficult to interpret and obscuring the evaluation of the efficiency-effectiveness tradeoff. To address this issue, we propose \ours\footnote{https://github.com/zhiyuanpeng/EER-FLOPs.} for LLM-based rerankers: RPP (ranking metrics per PetaFLOP), measuring how much ranking quality (e.g., NDCG or MRR) a method achieves per PetaFLOP, and QPP (queries per PetaFLOP), measuring how many queries can be processed per PetaFLOP. Accompanied by the new metrics, an interpretable FLOPs estimator is developed to estimate the FLOPs of an LLM-based reranker even without running any experiments. Based on the proposed metrics, we conduct comprehensive experiments to evaluate a wide range of LLM-based rerankers with different architectures, studying the efficiency-effectiveness trade-off and bringing this issue to the attention of the research community.
comment: Accepted by EMNLP Industry Track 2025
♻ ☆ Teaching Your Models to Understand Code via Focal Preference Alignment EMNLP'25
Preference learning extends the performance of Code LLMs beyond traditional supervised fine-tuning by leveraging relative quality comparisons. In existing approaches, a set of n candidate solutions is evaluated based on test case success rates, with the candidate demonstrating a higher pass rate being labeled as positive and its counterpart with a lower pass rate as negative. However, because this approach aligns entire failing code blocks rather than pinpointing specific errors, it lacks the granularity necessary to capture meaningful error-correction relationships. As a result, the model is unable to learn more informative error-correction patterns. To address these issues, we propose Target-DPO, a new preference alignment framework that mimics human iterative debugging to refine Code LLMs. Target-DPO explicitly locates error regions and aligns the corresponding tokens via a tailored DPO algorithm. To facilitate it, we introduce the CodeFlow dataset, where samples are iteratively refined until passing tests, with modifications capturing error corrections. Extensive experiments show that a diverse suite of Code LLMs equipped with Target-DPO achieves significant performance gains in code generation and improves on challenging tasks like BigCodeBench. In-depth analysis reveals that Target-DPO yields fewer errors. Code, model and datasets are in: https://github.com/JieWu02/Target-DPO.
comment: Accepted by EMNLP'25
♻ ☆ Med-R$^2$: Crafting Trustworthy LLM Physicians via Retrieval and Reasoning of Evidence-Based Medicine
Large Language Models (LLMs) have exhibited remarkable capabilities in clinical scenarios. Despite their potential, existing works face challenges when applying LLMs to medical settings. Strategies relying on training with medical datasets are highly cost-intensive and may suffer from outdated training data. Leveraging external knowledge bases is a suitable alternative, yet it faces obstacles such as limited retrieval precision and poor effectiveness in answer extraction. These issues collectively prevent LLMs from demonstrating the expected level of proficiency in mastering medical expertise. To address these challenges, we introduce Med-R^2, a novel LLM physician framework that adheres to the Evidence-Based Medicine (EBM) process, efficiently integrating retrieval mechanisms as well as the selection and reasoning processes of evidence, thereby enhancing the problem-solving capabilities of LLMs in healthcare scenarios and fostering a trustworthy LLM physician. Our comprehensive experiments indicate that Med-R^2 achieves a 13.27\% improvement over vanilla RAG methods and even a 4.55\% enhancement compared to fine-tuning strategies, without incurring additional training costs. Furthermore, we find that our LLaMA3.1-70B + Med-R$^2$ surpasses frontier models, including GPT-4o, Claude3.5-Sonnet and DeepSeek-V3 by 1.05\%, 6.14\% and 1.91\%. Med-R$^2$ effectively enhances the capabilities of LLMs in the medical domain.
♻ ☆ ReasonMed: A 370K Multi-Agent Generated Dataset for Advancing Medical Reasoning
Reasoning-based large language models have excelled in mathematics and programming, yet their potential in knowledge-intensive medical question answering remains underexplored and insufficiently validated in clinical contexts. To bridge this gap, we introduce ReasonMed, the largest medical reasoning dataset to date, comprising 370k high-quality examples distilled from 1.75 million initial reasoning paths generated by complementary LLMs and curated through a cost-efficient easy-medium-difficult (EMD) pipeline. ReasonMed is built through a multi-agent generation, verification, and refinement process, in which an Error Refiner improves reasoning paths by correcting error-prone steps identified by a verifier. Using ReasonMed, we investigate effective strategies for training medical reasoning models and find that integrating detailed CoT reasoning with concise answer summaries yields the most robust fine-tuning results. Models trained on ReasonMed set a new benchmark: ReasonMed-7B surpasses the prior best sub-10B models by 4.17% and even exceeds LLaMA3.1-70B on PubMedQA by 4.60%. When scaled to ReasonMed-14B, it remains highly competitive, underscoring consistent scaling potential. The codes and datasets are available at https://github.com/YuSun-Work/ReasonMed.
comment: 28 pages, 6 figures, 7 tables
♻ ☆ Fair Play in the Newsroom: Actor-Based Filtering Gender Discrimination in Text Corpora
Language corpora are the foundation of most natural language processing research, yet they often reproduce structural inequalities. One such inequality is gender discrimination in how actors are represented, which can distort analyses and perpetuate discriminatory outcomes. This paper introduces a user-centric, actor-level pipeline for detecting and mitigating gender discrimination in large-scale text corpora. By combining discourse-aware analysis with metrics for sentiment, syntactic agency, and quotation styles, our method enables both fine-grained auditing and exclusion-based balancing. Applied to the taz2024full corpus of German newspaper articles (1980-2024), the pipeline yields a more gender-balanced dataset while preserving core dynamics of the source material. Our findings show that structural asymmetries can be reduced through systematic filtering, though subtler biases in sentiment and framing remain. We release the tools and reports to support further research in discourse-based fairness auditing and equitable corpus construction.
♻ ☆ EpiCoder: Encompassing Diversity and Complexity in Code Generation ICML 2025
Existing methods for code generation use code snippets as seed data, restricting the complexity and diversity of the synthesized data. In this paper, we introduce a novel feature tree-based synthesis framework, which revolves around hierarchical code features derived from high-level abstractions of code. The feature tree is constructed from raw data and refined iteratively to increase the quantity and diversity of the extracted features, which captures and recognizes more complex patterns and relationships within the code. By adjusting the depth and breadth of the sampled subtrees, our framework provides precise control over the complexity of the generated code, enabling functionalities that range from function-level operations to multi-file scenarios. We fine-tuned widely-used base models to obtain EpiCoder series, achieving state-of-the-art performance on multiple benchmarks at both the function and file levels. In particular, empirical evidence indicates that our approach shows significant potential in the synthesizing of repository-level code data. Our code and data are publicly available at https://github.com/microsoft/EpiCoder.
comment: ICML 2025
♻ ☆ Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation ICASSP 2026
Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including dedicated multilingual and long-form tracks. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.
comment: Submitted to ICASSP 2026; Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard ; Code: https://github.com/huggingface/open_asr_leaderboard
♻ ☆ MoM: Linear Sequence Modeling with Mixture-of-Memories ICLR 2026
Linear sequence modeling methods, such as linear attention, state space modeling, and linear RNNs, offer significant efficiency improvements by reducing the complexity of training and inference. However, these methods typically compress the entire input sequence into a single fixed-size memory state, which leads to suboptimal performance on recall-intensive tasks. To address this limitation, we introduce a novel architecture called Mixture-of-Memories (MoM). MoM utilizes multiple independent memory states, with a router network directing input tokens to specific memory states. This approach greatly enhances the overall memory capacity while minimizing memory interference. MoM serves as a general framework that can be seamlessly combined with diverse memory update mechanisms across linear models. As a result, MoM performs exceptionally well on recall-intensive tasks, surpassing existing linear sequence modeling techniques. Despite incorporating multiple memory states, the computation of each memory state remains linear in complexity, allowing MoM to retain the linear-complexity advantage during training, while constant-complexity during inference. Our experimental results show that MoM outperforms current linear sequence models on downstream language tasks, particularly recall-intensive tasks, and even achieves performance comparable to Transformer models. The code is released at https://github.com/OpenSparseLLMs/MoM and is also released as a part of https://github.com/OpenSparseLLMs/Linear-MoE.
comment: Technical report, 18 pages, submitted to ICLR 2026
♻ ☆ Training LLMs to be Better Text Embedders through Bidirectional Reconstruction EMNLP 2025
Large language models (LLMs) have increasingly been explored as powerful text embedders. Existing LLM-based text embedding approaches often leverage the embedding of the final token, typically a reserved special token such as [EOS]. However, these tokens have not been intentionally trained to capture the semantics of the whole context, limiting their capacity as text embeddings, especially for retrieval and re-ranking tasks. We propose to add a new training stage before contrastive learning to enrich the semantics of the final token embedding. This stage employs bidirectional generative reconstruction tasks, namely EBQ2D (Embedding-Based Query-to-Document) and EBD2Q (Embedding-Based Document-to-Query), which interleave to anchor the [EOS] embedding and reconstruct either side of Query-Document pairs. Experimental results demonstrate that our additional training stage significantly improves LLM performance on the Massive Text Embedding Benchmark (MTEB), achieving new state-of-the-art results across different LLM base models and scales.
comment: accepted by EMNLP 2025 Main Conference
♻ ☆ Building Resource-Constrained Language Agents: A Korean Case Study on Chemical Toxicity Information EMNLP 2025
Language agents powered by large language models (LLMs) face significant deployment challenges in resource-constrained environments, particularly for specialized domains and less-common languages. This paper presents Tox-chat, a Korean chemical toxicity information agent devised within these limitations. We propose two key innovations: a context-efficient architecture that reduces token consumption through hierarchical section search, and a scenario-based dialogue generation methodology that effectively distills tool-using capabilities from larger models. Experimental evaluations demonstrate that our fine-tuned 8B parameter model substantially outperforms both untuned models and baseline approaches, in terms of DB faithfulness and preference. Our work offers valuable insights for researchers developing domain-specific language agents under practical constraints.
comment: EMNLP 2025 Industry track
♻ ☆ MAGIC: A Multi-Hop and Graph-Based Benchmark for Inter-Context Conflicts in Retrieval-Augmented Generation EMNLP 2025
Knowledge conflict often arises in retrieval-augmented generation (RAG) systems, where retrieved documents may be inconsistent with one another or contradict the model's parametric knowledge. Existing benchmarks for investigating the phenomenon have notable limitations, including a narrow focus on the question answering setup, heavy reliance on entity substitution techniques, and a restricted range of conflict types. To address these issues, we propose a knowledge graph (KG)-based framework that generates varied and subtle conflicts between two similar yet distinct contexts, while ensuring interpretability through the explicit relational structure of KGs. Experimental results on our benchmark, MAGIC, provide intriguing insights into the inner workings of LLMs regarding knowledge conflict: both open-source and proprietary models struggle with conflict detection -- especially when multi-hop reasoning is required -- and often fail to pinpoint the exact source of contradictions. Finally, we present in-depth analyses that serve as a foundation for improving LLMs in integrating diverse, sometimes even conflicting, information.
comment: EMNLP 2025 Findings
♻ ☆ AutoAgent: A Fully-Automated and Zero-Code Framework for LLM Agents
Large Language Model (LLM) Agents have demonstrated remarkable capabilities in task automation and intelligent decision-making, driving the widespread adoption of agent development frameworks such as LangChain and AutoGen. However, these frameworks predominantly serve developers with extensive technical expertise - a significant limitation considering that only 0.03 % of the global population possesses the necessary programming skills. This stark accessibility gap raises a fundamental question: Can we enable everyone, regardless of technical background, to build their own LLM agents using natural language alone? To address this challenge, we introduce AutoAgent-a Fully-Automated and highly Self-Developing framework that enables users to create and deploy LLM agents through Natural Language Alone. Operating as an autonomous Agent Operating System, AutoAgent comprises four key components: i) Agentic System Utilities, ii) LLM-powered Actionable Engine, iii) Self-Managing File System, and iv) Self-Play Agent Customization module. This lightweight yet powerful system enables efficient and dynamic creation and modification of tools, agents, and workflows without coding requirements or manual intervention. Beyond its code-free agent development capabilities, AutoAgent also serves as a versatile multi-agent system for General AI Assistants. Comprehensive evaluations on the GAIA benchmark demonstrate AutoAgent's effectiveness in generalist multi-agent tasks, surpassing existing state-of-the-art methods. Furthermore, AutoAgent's Retrieval-Augmented Generation (RAG)-related capabilities have shown consistently superior performance compared to many alternative LLM-based solutions.
comment: Code: https://github.com/HKUDS/AutoAgent
♻ ☆ Mining the Mind: What 100M Beliefs Reveal About Frontier LLM Knowledge
LLMs are remarkable artifacts that have revolutionized a range of NLP and AI tasks. A significant contributor is their factual knowledge, which, to date, remains poorly understood, and is usually analyzed from biased samples. In this paper, we take a deep tour into the factual knowledge (or beliefs) of a frontier LLM, based on GPTKB v1.5 (Hu et al., 2025a), a recursively elicited set of 100 million beliefs of one of the strongest currently available frontier LLMs, GPT-4.1. We find that the models' factual knowledge differs quite significantly from established knowledge bases, and that its accuracy is significantly lower than indicated by previous benchmarks. We also find that inconsistency, ambiguity and hallucinations are major issues, shedding light on future research opportunities concerning factual LLM knowledge.
♻ ☆ Less is More: Compact Clue Selection for Efficient Retrieval-Augmented Generation Reasoning
Current RAG retrievers are designed primarily for human readers, emphasizing complete, readable, and coherent paragraphs. However, LLMs benefit more from precise, compact, and well-structured input, which enhances reasoning quality and efficiency. Existing methods often rely on reranking or summarization to identify key sentences, but may suffer from semantic breaks and unfaithfulness. Thus, efficiently extracting and organizing answer-relevant clues from large-scale documents while reducing LLM reasoning costs remains a challenge for RAG. Inspired by Occam's razor, we frame LLM-centric retrieval as a MinMax optimization: maximizing the extraction of potential clues and reranking them for well-organization, while minimizing reasoning costs by truncating to the smallest sufficient clues set. In this paper, we propose CompSelect, a Compact clue Selection mechanism for LLM-centric RAG, consisting of a clue extractor, a reranker, and a truncator. (1) The clue extractor first uses answer-containing sentences as fine-tuning targets, aiming to extract sufficient potential clues; (2) The reranker is trained to prioritize effective clues based on real LLM feedback; (3) The truncator uses the truncated text containing the minimum sufficient clues for answering the question as fine-tuning targets, thereby enabling efficient RAG reasoning. Experiments on three QA datasets show that CompSelect improves QA performance by approximately 11\% and reduces Total Latency and Online Latency by approximately 17\% and 67\% compared to various baseline methods on both LLaMA3 and Qwen3. Further analysis confirms its robustness to unreliable retrieval and generalization across different scenarios, offering a scalable and cost-efficient solution for web-scale RAG applications.
comment: 12 pages, 7 figures, 12 tables, under review
♻ ☆ Enhancing LLM Reliability via Explicit Knowledge Boundary Modeling
Large language models (LLMs) are prone to hallucination stemming from misaligned self-awareness, particularly when processing queries exceeding their knowledge boundaries. While existing mitigation strategies employ uncertainty estimation or query rejection mechanisms, they suffer from computational efficiency and sacrificed helpfulness. To address these issues, we propose the Explicit Knowledge Boundary Modeling (EKBM) framework, integrating fast and slow reasoning systems to harmonize reliability and usability. The framework first employs a fast-thinking model to generate confidence-labeled responses, enabling immediate utilization of high-confidence outputs, whereas uncertain predictions trigger a slow refinement model for accuracy improvement. To align model behavior with our proposed object, we propose a hybrid training pipeline, enhancing self-awareness without degrading task performance. Evaluations on dialogue state tracking tasks demonstrate that EKBM achieves superior model reliability over uncertainty-based baselines. Further analysis reveals that refinement substantially boosts accuracy while maintaining low computational overhead. The framework establishes a scalable paradigm for deploying reliable LLMs in error-sensitive applications, effectively balancing accuracy and practical utility.
♻ ☆ Multiple Memory Systems for Enhancing the Long-term Memory of Agent
An agent powered by large language models have achieved impressive results, but effectively handling the vast amounts of historical data generated during interactions remains a challenge. The current approach is to design a memory module for the agent to process these data. However, existing methods, such as MemoryBank and A-MEM, have poor quality of stored memory content, which affects recall performance and response quality. In order to better construct high-quality long-term memory content, we have designed a multiple memory system (MMS) inspired by cognitive psychology theory. The system processes short-term memory to multiple long-term memory fragments, and constructs retrieval memory units and contextual memory units based on these fragments, with a one-to-one correspondence between the two. During the retrieval phase, MMS will match the most relevant retrieval memory units based on the user's query. Then, the corresponding contextual memory units is obtained as the context for the response stage to enhance knowledge, thereby effectively utilizing historical data. Experiments on LoCoMo dataset compared our method with three others, proving its effectiveness. Ablation studies confirmed the rationality of our memory units. We also analyzed the robustness regarding the number of selected memory segments and the storage overhead, demonstrating its practical value.
♻ ☆ Logic Jailbreak: Efficiently Unlocking LLM Safety Restrictions Through Formal Logical Expression
Despite substantial advancements in aligning large language models (LLMs) with human values, current safety mechanisms remain susceptible to jailbreak attacks. We hypothesize that this vulnerability stems from distributional discrepancies between alignment-oriented prompts and malicious prompts. To investigate this, we introduce LogiBreak, a novel and universal black-box jailbreak method that leverages logical expression translation to circumvent LLM safety systems. By converting harmful natural language prompts into formal logical expressions, LogiBreak exploits the distributional gap between alignment data and logic-based inputs, preserving the underlying semantic intent and readability while evading safety constraints. We evaluate LogiBreak on a multilingual jailbreak dataset spanning three languages, demonstrating its effectiveness across various evaluation settings and linguistic contexts.
♻ ☆ The Role of Model Confidence on Bias Effects in Measured Uncertainties for Vision-Language Models EMNLP
With the growing adoption of Large Language Models (LLMs) for open-ended tasks, accurately assessing epistemic uncertainty, which reflects a model's lack of knowledge, has become crucial to ensuring reliable outcomes. However, quantifying epistemic uncertainty in such tasks is challenging due to the presence of aleatoric uncertainty, which arises from multiple valid answers. While bias can introduce noise into epistemic uncertainty estimation, it may also reduce noise from aleatoric uncertainty. To investigate this trade-off, we conduct experiments on Visual Question Answering (VQA) tasks and find that mitigating prompt-introduced bias improves uncertainty quantification in GPT-4o. Building on prior work showing that LLMs tend to copy input information when model confidence is low, we further analyze how these prompt biases affect measured epistemic and aleatoric uncertainty across varying bias-free confidence levels with GPT-4o and Qwen2-VL. We find that all considered biases have greater effects in both uncertainties when bias-free model confidence is lower. Moreover, lower bias-free model confidence is associated with greater bias-induced underestimation of epistemic uncertainty, resulting in overconfident estimates, whereas it has no significant effect on the direction of bias effect in aleatoric uncertainty estimation. These distinct effects deepen our understanding of bias mitigation for uncertainty quantification and potentially inform the development of more advanced techniques.
comment: Accepted to EMNLP Findings
Information Retrieval 23
Agent Learning via Early Experience
A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.
comment: Work in progress
☆ Detecting Legend Items on Historical Maps Using GPT-4o with In-Context Learning
Historical map legends are critical for interpreting cartographic symbols. However, their inconsistent layouts and unstructured formats make automatic extraction challenging. Prior work focuses primarily on segmentation or general optical character recognition (OCR), with few methods effectively matching legend symbols to their corresponding descriptions in a structured manner. We present a method that combines LayoutLMv3 for layout detection with GPT-4o using in-context learning to detect and link legend items and their descriptions via bounding box predictions. Our experiments show that GPT-4 with structured JSON prompts outperforms the baseline, achieving 88% F-1 and 85% IoU, and reveal how prompt design, example counts, and layout alignment affect performance. This approach supports scalable, layout-aware legend parsing and improves the indexing and searchability of historical maps across various visual styles.
☆ Mobile Gamer Lifetime Value Prediction via Objective Decomposition and Reconstruction
For Internet platforms operating real-time bidding (RTB) advertising service, a comprehensive understanding of user lifetime value (LTV) plays a pivotal role in optimizing advertisement allocation efficiency and maximizing the return on investment (ROI) for advertisement sponsors, thereby facilitating growth of commercialization revenue for the platform. However, the inherent complexity of user LTV distributions induces significant challenges in accurate LTV prediction. Existing state-of-the-art works, which primarily focus on directly learning the LTV distributions through well-designed loss functions, achieve limited success due to their vulnerability to outliers. In this paper, we proposed a novel LTV prediction method to address distribution challenges through an objective decomposition and reconstruction framework. Briefly speaking, based on the in-app purchase characteristics of mobile gamers, our model was designed to first predict the number of transactions at specific prices and then calculate the total payment amount from these intermediate predictions. Our proposed model was evaluated through experiments on real-world industrial dataset, and deployed on the TapTap RTB advertising system for online A/B testing along with the state-of-the-art ZILN model.
comment: 6 pages, 6 figures
☆ ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval
In this paper, we introduce ReasonEmbed, a novel text embedding model developed for reasoning-intensive document retrieval. Our work includes three key technical contributions. First, we propose ReMixer, a new data synthesis method that overcomes the triviality problem prevalent in previous synthetic datasets, enabling large-scale production of 82K high-quality training samples. Second, we design Redapter, a self-adaptive learning algorithm that dynamically adjusts training each sample's weight based on its reasoning intensity. This allows the model to effectively capture the complex semantic relationships between queries and documents. Third, we implement ReasonEmbed across multiple backbones of varying sizes, all of which achieve superior performance on reasoning-intensive retrieval tasks. Notably, our ReasonEmbed-Qwen3-8B model offers a record-high nDCG@10 score of 38.1 on the BRIGHT benchmark, which significantly outperforms existing text embedding models. We will fully open-source our created resources in ReasonEmbed to push forward the research advancement in this field.
comment: 17 pages, 3 figures
☆ VersionRAG: Version-Aware Retrieval-Augmented Generation for Evolving Documents
Retrieval-Augmented Generation (RAG) systems fail when documents evolve through versioning-a ubiquitous characteristic of technical documentation. Existing approaches achieve only 58-64% accuracy on version-sensitive questions, retrieving semantically similar content without temporal validity checks. We present VersionRAG, a version-aware RAG framework that explicitly models document evolution through a hierarchical graph structure capturing version sequences, content boundaries, and changes between document states. During retrieval, VersionRAG routes queries through specialized paths based on intent classification, enabling precise version-aware filtering and change tracking. On our VersionQA benchmark-100 manually curated questions across 34 versioned technical documents-VersionRAG achieves 90% accuracy, outperforming naive RAG (58%) and GraphRAG (64%). VersionRAG reaches 60% accuracy on implicit change detection where baselines fail (0-10%), demonstrating its ability to track undocumented modifications. Additionally, VersionRAG requires 97% fewer tokens during indexing than GraphRAG, making it practical for large-scale deployment. Our work establishes versioned document QA as a distinct task and provides both a solution and benchmark for future research.
☆ TaoSR-AGRL: Adaptive Guided Reinforcement Learning Framework for E-commerce Search Relevance
Query-product relevance prediction is fundamental to e-commerce search and has become even more critical in the era of AI-powered shopping, where semantic understanding and complex reasoning directly shape the user experience and business conversion. Large Language Models (LLMs) enable generative, reasoning-based approaches, typically aligned via supervised fine-tuning (SFT) or preference optimization methods like Direct Preference Optimization (DPO). However, the increasing complexity of business rules and user queries exposes the inability of existing methods to endow models with robust reasoning capacity for long-tail and challenging cases. Efforts to address this via reinforcement learning strategies like Group Relative Policy Optimization (GRPO) often suffer from sparse terminal rewards, offering insufficient guidance for multi-step reasoning and slowing convergence. To address these challenges, we propose TaoSR-AGRL, an Adaptive Guided Reinforcement Learning framework for LLM-based relevance prediction in Taobao Search Relevance. TaoSR-AGRL introduces two key innovations: (1) Rule-aware Reward Shaping, which decomposes the final relevance judgment into dense, structured rewards aligned with domain-specific relevance criteria; and (2) Adaptive Guided Replay, which identifies low-accuracy rollouts during training and injects targeted ground-truth guidance to steer the policy away from stagnant, rule-violating reasoning patterns toward compliant trajectories. TaoSR-AGRL was evaluated on large-scale real-world datasets and through online side-by-side human evaluations on Taobao Search. It consistently outperforms DPO and standard GRPO baselines in offline experiments, improving relevance accuracy, rule adherence, and training stability. The model trained with TaoSR-AGRL has been successfully deployed in the main search scenario on Taobao, serving hundreds of millions of users.
☆ Generation and annotation of item usage scenarios in e-commerce using large language models
Complementary recommendations suggest combinations of useful items that play important roles in e-commerce. However, complementary relationships are often subjective and vary among individuals, making them difficult to infer from historical data. Unlike conventional history-based methods that rely on statistical co-occurrence, we focus on the underlying usage context that motivates item combinations. We hypothesized that people select complementary items by imagining specific usage scenarios and identifying the needs in such situations. Based on this idea, we explored the use of large language models (LLMs) to generate item usage scenarios as a starting point for constructing complementary recommendation systems. First, we evaluated the plausibility of LLM-generated scenarios through manual annotation. The results demonstrated that approximately 85% of the generated scenarios were determined to be plausible, suggesting that LLMs can effectively generate realistic item usage scenarios.
☆ HySim-LLM: Embedding-Weighted Fine-Tuning Bounds and Manifold Denoising for Domain-Adapted LLMs
The extraction and standardization of pharmacokinetic (PK) information from scientific literature remain significant challenges in computational pharmacology, which limits the reliability of data-driven models in drug development. Large language models (LLMs) have achieved remarkable progress in text understanding and reasoning, yet their adaptation to structured biomedical data, such as PK tables, remains constrained by heterogeneity, noise, and domain shift. To address these limitations, we propose HySim-LLM, a unified mathematical and computational framework that integrates embedding-weighted fine-tuning and manifold-aware denoising to enhance the robustness and interpretability of LLMs. We establish two theoretical results: (1) a similarity-weighted generalization bound that quantifies adaptation performance under embedding divergence, and (2) a manifold-based denoising guarantee that bounds loss contributions from noisy or off-manifold samples. These theorems provide a principled foundation for fine-tuning LLMs in structured biomedical settings. The framework offers a mathematically grounded pathway toward reliable and interpretable LLM adaptation for biomedical and data-intensive scientific domains.
☆ PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations
Large Language Models (LLMs) pose a new paradigm of modeling and computation for information tasks. Recommendation systems are a critical application domain poised to benefit significantly from the sequence modeling capabilities and world knowledge inherent in these large models. In this paper, we introduce PLUM, a framework designed to adapt pre-trained LLMs for industry-scale recommendation tasks. PLUM consists of item tokenization using Semantic IDs, continued pre-training (CPT) on domain-specific data, and task-specific fine-tuning for recommendation objectives. For fine-tuning, we focus particularly on generative retrieval, where the model is directly trained to generate Semantic IDs of recommended items based on user context. We conduct comprehensive experiments on large-scale internal video recommendation datasets. Our results demonstrate that PLUM achieves substantial improvements for retrieval compared to a heavily-optimized production model built with large embedding tables. We also present a scaling study for the model's retrieval performance, our learnings about CPT, a few enhancements to Semantic IDs, along with an overview of the training and inference methods that enable launching this framework to billions of users in YouTube.
comment: 11 pages, 6 figures
☆ Who Stole Your Data? A Method for Detecting Unauthorized RAG Theft
Retrieval-augmented generation (RAG) enhances Large Language Models (LLMs) by mitigating hallucinations and outdated information issues, yet simultaneously facilitates unauthorized data appropriation at scale. This paper addresses this challenge through two key contributions. First, we introduce RPD, a novel dataset specifically designed for RAG plagiarism detection that encompasses diverse professional domains and writing styles, overcoming limitations in existing resources. Second, we develop a dual-layered watermarking system that embeds protection at both semantic and lexical levels, complemented by an interrogator-detective framework that employs statistical hypothesis testing on accumulated evidence. Extensive experimentation demonstrates our approach's effectiveness across varying query volumes, defense prompts, and retrieval parameters, while maintaining resilience against adversarial evasion techniques. This work establishes a foundational framework for intellectual property protection in retrieval-augmented AI systems.
☆ Queries Are Not Alone: Clustering Text Embeddings for Video Search SIGIR
The rapid proliferation of video content across various platforms has highlighted the urgent need for advanced video retrieval systems. Traditional methods, which primarily depend on directly matching textual queries with video metadata, often fail to bridge the semantic gap between text descriptions and the multifaceted nature of video content. This paper introduces a novel framework, the Video-Text Cluster (VTC), which enhances video retrieval by clustering text queries to capture a broader semantic scope. We propose a unique clustering mechanism that groups related queries, enabling our system to consider multiple interpretations and nuances of each query. This clustering is further refined by our innovative Sweeper module, which identifies and mitigates noise within these clusters. Additionally, we introduce the Video-Text Cluster-Attention (VTC-Att) mechanism, which dynamically adjusts focus within the clusters based on the video content, ensuring that the retrieval process emphasizes the most relevant textual features. Further experiments have demonstrated that our proposed model surpasses existing state-of-the-art models on five public datasets.
comment: Accepted by International ACM SIGIR Conference on Research and Development in Information Retrieval 2025
☆ ISMIE: A Framework to Characterize Information Seeking in Modern Information Environments SIGIR
The modern information environment (MIE) is increasingly complex, shaped by a wide range of techniques designed to satisfy users' information needs. Information seeking (IS) models are effective mechanisms for characterizing user-system interactions. However, conceptualizing a model that fully captures the MIE landscape poses a challenge. We argue: Does such a model exist? To address this, we propose the Information Seeking in Modern Information Environments (ISMIE) framework as a fundamental step. ISMIE conceptualizes the information seeking process (ISP) via three key concepts: Components (e.g., Information Seeker), Intervening Variables (e.g., Interactive Variables), and Activities (e.g., Acquiring). Using ISMIE's concepts and employing a case study based on a common scenario - misinformation dissemination - we analyze six existing IS and information retrieval (IR) models to illustrate their limitations and the necessity of ISMIE. We then show how ISMIE serves as an actionable framework for both characterization and experimental design. We characterize three pressing issues and then outline two research blueprints: a user-centric, industry-driven experimental design for the authenticity and trust crisis to AI-generated content and a system-oriented, academic-driven design for tackling dopamine-driven content consumption. Our framework offers a foundation for developing IS and IR models to advance knowledge on understanding human interactions and system design in MIEs.
comment: This paper has been accepted to SIGIR-AP 2025
☆ Stop DDoS Attacking the Research Community with AI-Generated Survey Papers NeurIPS 2025
Survey papers are foundational to the scholarly progress of research communities, offering structured overviews that guide both novices and experts across disciplines. However, the recent surge of AI-generated surveys, especially enabled by large language models (LLMs), has transformed this traditionally labor-intensive genre into a low-effort, high-volume output. While such automation lowers entry barriers, it also introduces a critical threat: the phenomenon we term the "survey paper DDoS attack" to the research community. This refers to the unchecked proliferation of superficially comprehensive but often redundant, low-quality, or even hallucinated survey manuscripts, which floods preprint platforms, overwhelms researchers, and erodes trust in the scientific record. In this position paper, we argue that we must stop uploading massive amounts of AI-generated survey papers (i.e., survey paper DDoS attack) to the research community, by instituting strong norms for AI-assisted review writing. We call for restoring expert oversight and transparency in AI usage and, moreover, developing new infrastructures such as Dynamic Live Surveys, community-maintained, version-controlled repositories that blend automated updates with human curation. Through quantitative trend analysis, quality audits, and cultural impact discussion, we show that safeguarding the integrity of surveys is no longer optional but imperative to the research community.
comment: Accepted by NeurIPS 2025 (Position Track)
♻ ☆ SelfRACG: Enabling LLMs to Self-Express and Retrieve for Code Generation
Existing retrieval-augmented code generation (RACG) methods typically use an external retrieval module to fetch semantically similar code snippets used for generating subsequent fragments. However, even for consecutive code fragments, the content often diverges due to logical progression, resulting in a content gap. This gap undermines the performance of current RACG methods, as \textit{external} retrieval modules based on content matching fail to infer the specific information need of LLMs to generate the next code fragment. Therefore, we propose \textbf{SelfRACG}, a novel paradigm that enables large language models (LLMs) to \textbf{Self}-express their information needs to enhance \textbf{RACG}. Specifically, SelfRACG includes an information need expression module and a two-stage information need-guided training strategy, which encourages LLMs to express their information need. Extensive experiments demonstrate that SelfRACG can retrieve external knowledge that better aligns with the LLM's own information needs, resulting in superior generation performance compared to vanilla RACG.
comment: Tsinghua&Xiaohongshu
♻ ☆ SustainableQA: A Comprehensive Question Answering Dataset for Corporate Sustainability and EU Taxonomy Reporting
The growing demand for corporate sustainability transparency, particularly under new regulations like the EU Taxonomy, necessitates precise data extraction from large, unstructured corporate reports, a task for which Large Language Models and Retrieval-RAG systems require high-quality, domain-specific question-answering datasets. To address this, we introduce SustainableQA, a novel dataset and a scalable pipeline that generates comprehensive QA pairs from corporate sustainability and annual reports by integrating semantic chunk classification, a hybrid span extraction pipeline, and a specialized table-to-paragraph transformation. To ensure high quality, the generation is followed by a novel automated assessment and refinement pipeline that systematically validates each QA pair for faithfulness and relevance, repairing or discarding low-quality entries. This results in a final, robust dataset of over 195,000 diverse factoid and non-factoid QA pairs, whose effectiveness is demonstrated by initial fine-tuning experiments where a compact 8B parameter model significantly outperforms much larger state-of-the-art models. SustainableQA proves to be a highly effective resource for developing and benchmarking advanced knowledge assistants capable of navigating complex sustainability compliance data.
♻ ☆ Distilling a Small Utility-Based Passage Selector to Enhance Retrieval-Augmented Generation SIGIR
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating retrieved information. Standard retrieval process prioritized relevance, focusing on topical alignment between queries and passages. In contrast, in RAG, the emphasis has shifted to utility, which considers the usefulness of passages for generating accurate answers. Despite empirical evidence showing the benefits of utility-based retrieval in RAG, the high computational cost of using LLMs for utility judgments limits the number of passages evaluated. This restriction is problematic for complex queries requiring extensive information. To address this, we propose a method to distill the utility judgment capabilities of LLMs into smaller, more efficient models. Our approach focuses on utility-based selection rather than ranking, enabling dynamic passage selection tailored to specific queries without the need for fixed thresholds. We train student models to learn pseudo-answer generation and utility judgments from teacher LLMs, using a sliding window method that dynamically selects useful passages. Our experiments demonstrate that utility-based selection provides a flexible and cost-effective solution for RAG, significantly reducing computational costs while improving answer quality. We present the distillation results using Qwen3-32B as the teacher model for both relevance ranking and utility-based selection, distilled into RankQwen1.7B and UtilityQwen1.7B. Our findings indicate that for complex questions, utility-based selection is more effective than relevance ranking in enhancing answer generation performance. We will release the relevance ranking and utility-based selection annotations for the MS MARCO dataset, supporting further research in this area.
comment: Accepted by SIGIR-AP25
♻ ☆ Training LLMs to be Better Text Embedders through Bidirectional Reconstruction EMNLP 2025
Large language models (LLMs) have increasingly been explored as powerful text embedders. Existing LLM-based text embedding approaches often leverage the embedding of the final token, typically a reserved special token such as [EOS]. However, these tokens have not been intentionally trained to capture the semantics of the whole context, limiting their capacity as text embeddings, especially for retrieval and re-ranking tasks. We propose to add a new training stage before contrastive learning to enrich the semantics of the final token embedding. This stage employs bidirectional generative reconstruction tasks, namely EBQ2D (Embedding-Based Query-to-Document) and EBD2Q (Embedding-Based Document-to-Query), which interleave to anchor the [EOS] embedding and reconstruct either side of Query-Document pairs. Experimental results demonstrate that our additional training stage significantly improves LLM performance on the Massive Text Embedding Benchmark (MTEB), achieving new state-of-the-art results across different LLM base models and scales.
comment: accepted by EMNLP 2025 Main Conference
♻ ☆ Reasoning-enhanced Query Understanding through Decomposition and Interpretation
Accurate inference of user intent is crucial for enhancing document retrieval in modern search engines. While large language models (LLMs) have made significant strides in this area, their effectiveness has predominantly been assessed with short, keyword-based queries. As AI-driven search evolves, long-form queries with intricate intents are becoming more prevalent, yet they remain underexplored in the context of LLM-based query understanding (QU). To bridge this gap, we introduce ReDI: a Reasoning-enhanced approach for query understanding through Decomposition and Interpretation. ReDI leverages the reasoning and comprehension capabilities of LLMs in a three-stage pipeline: (i) it breaks down complex queries into targeted sub-queries to accurately capture user intent; (ii) it enriches each sub-query with detailed semantic interpretations to improve the query-document matching; and (iii) it independently retrieves documents for each sub-query and employs a fusion strategy to aggregate the results for the final ranking. We compiled a large-scale dataset of real-world complex queries from a major search engine and distilled the query understanding capabilities of teacher models into smaller models for practical application. Experiments on BRIGHT and BEIR demonstrate that ReDI consistently surpasses strong baselines in both sparse and dense retrieval paradigms, affirming its effectiveness. We release our code at https://anonymous.4open.science/r/ReDI-6FC7/.
♻ ☆ Utility-Focused LLM Annotation for Retrieval and Retrieval-Augmented Generation EMNLP25
This paper explores the use of large language models (LLMs) for annotating document utility in training retrieval and retrieval-augmented generation (RAG) systems, aiming to reduce dependence on costly human annotations. We address the gap between retrieval relevance and generative utility by employing LLMs to annotate document utility. To effectively utilize multiple positive samples per query, we introduce a novel loss that maximizes their summed marginal likelihood. Using the Qwen-2.5-32B model, we annotate utility on the MS MARCO dataset and conduct retrieval experiments on MS MARCO and BEIR, as well as RAG experiments on MS MARCO QA, NQ, and HotpotQA. Our results show that LLM-generated annotations enhance out-of-domain retrieval performance and improve RAG outcomes compared to models trained solely on human annotations or downstream QA metrics. Furthermore, combining LLM annotations with just 20% of human labels achieves performance comparable to using full human annotations. Our study offers a comprehensive approach to utilizing LLM annotations for initializing QA systems on new corpora.
comment: Accepted by the EMNLP25 main conference
♻ ☆ Scenario-Wise Rec: A Multi-Scenario Recommendation Benchmark CIKM'2025
Multi Scenario Recommendation (MSR) tasks, referring to building a unified model to enhance performance across all recommendation scenarios, have recently gained much attention. However, current research in MSR faces two significant challenges that hinder the field's development: the absence of uniform procedures for multi-scenario dataset processing, thus hindering fair comparisons, and most models being closed-sourced, which complicates comparisons with current SOTA models. Consequently, we introduce our benchmark, \textbf{Scenario-Wise Rec}, which comprises 6 public datasets and 12 benchmark models, along with a training and evaluation pipeline. Additionally, we validated the benchmark using an industrial advertising dataset, reinforcing its reliability and applicability in real-world scenarios. We aim for this benchmark to offer researchers valuable insights from prior work, enabling the development of novel models based on our benchmark and thereby fostering a collaborative research ecosystem in MSR. Our source code is also publicly available.
comment: Accepted to CIKM'2025
♻ ☆ Multi-Source Knowledge Pruning for Retrieval-Augmented Generation: A Benchmark and Empirical Study CIKM 2025
Retrieval-augmented generation (RAG) is increasingly recognized as an effective approach to mitigating the hallucination of large language models (LLMs) through the integration of external knowledge. While numerous efforts, most studies focus on a single type of external knowledge source. However, in real-world applications, most situations involve diverse knowledge from various sources, yet this area has been less explored. The main dilemma is the lack of a suitable dataset containing multiple knowledge sources and pre-exploration of the associated issues. To address these challenges, we standardize a benchmark dataset that combines structured and unstructured knowledge across diverse and complementary domains. Based on this dataset, we further develop a plug-and-play RAG framework, \textbf{PruningRAG}, whose main characteristic is the use of multi-granularity pruning strategies to optimize the integration of relevant information while minimizing misleading context. It consistently improves performance across various existing RAG variants, demonstrating its robustness and broad applicability. Building upon the standardized dataset and PruningRAG, we also report a series of experimental results, as well as insightful findings. Our dataset and code are publicly available\footnote{https://github.com/USTCAGI/PruningRAG}, with the aim of advancing future research in the RAG community.
comment: Accepted by CIKM 2025
♻ ☆ Preprint: Poster: Did I Just Browse A Website Written by LLMs?
Increasingly, web content is automatically generated by large language models (LLMs) with little human input. We call this "LLM-dominant" content. Since LLMs plagiarize and hallucinate, LLM-dominant content can be unreliable and unethical. Yet, websites rarely disclose such content, and human readers struggle to distinguish it. Thus, we must develop reliable detectors for LLM-dominant content. However, state-of-the-art LLM detectors are inaccurate on web content, because web content has low positive rates, complex markup, and diverse genres, instead of clean, prose-like benchmark data SoTA detectors are optimized for. We propose a highly reliable, scalable pipeline that classifies entire websites. Instead of naively classifying text extracted from each page, we classify each site based on an LLM text detector's outputs of multiple prose-like pages to boost accuracies. We train and evaluate our detector by collecting 2 distinct ground truth datasets totaling 120 sites, and obtain 100% accuracies testing across them. In the wild, we detect a sizable portion of sites as LLM-dominant among 10k sites in search engine results and 10k in Common Crawl archives. We find LLM-dominant sites are growing in prevalence and rank highly in search results, raising questions about their impact on end users and the overall Web ecosystem.
comment: ACM Internet Measurement Conference 2025 Poster & ACM IMC 2025 Student Workshop. 2 pages. 3 figures
♻ ☆ Contrastive Learning Augmented Social Recommendations
Recommender systems are essential for modern content platforms, yet traditional behavior-based models often struggle with cold users who have limited interaction data. Engaging these users is crucial for platform growth. To bridge this gap, we propose leveraging the social-relation graph to enrich interest representations from behavior-based models. However, extracting value from social graphs is challenging due to relation noise and cross-domain inconsistency. To address the noise propagation and obtain accurate social interest, we employ a dual-view denoising strategy, employing low-rank SVD to the user-item interaction matrix for a denoised social graph and contrastive learning to align the original and reconstructed social graphs. Addressing the interest inconsistency between social and behavioral interests, we adopt a "mutual distillation" technique to isolate the original interests into aligned social/behavior interests and social/behavior specific interests, maximizing the utility of both. Experimental results on widely adopted industry datasets verify the method's effectiveness, particularly for cold users, offering a fresh perspective for future research. The implementation can be accessed at https://github.com/WANGLin0126/CLSRec.
Machine Learning 150
☆ BLAZER: Bootstrapping LLM-based Manipulation Agents with Zero-Shot Data Generation
Scaling data and models has played a pivotal role in the remarkable progress of computer vision and language. Inspired by these domains, recent efforts in robotics have similarly focused on scaling both data and model size to develop more generalizable and robust policies. However, unlike vision and language, robotics lacks access to internet-scale demonstrations across diverse robotic tasks and environments. As a result, the scale of existing datasets typically suffers from the need for manual data collection and curation. To address this problem, here we propose BLAZER, a framework that learns manipulation policies from automatically generated training data. We build on the zero-shot capabilities of LLM planners and automatically generate demonstrations for diverse manipulation tasks in simulation. Successful examples are then used to finetune an LLM and to improve its planning capabilities without human supervision. Notably, while BLAZER training requires access to the simulator's state, we demonstrate direct transfer of acquired skills to sensor-based manipulation. Through extensive experiments, we show BLAZER to significantly improve zero-shot manipulation in both simulated and real environments. Moreover, BLAZER improves on tasks outside of its training pool and enables downscaling of LLM models. Our code and data will be made publicly available on the project page.
comment: 11 pages, 8 figures
☆ Reconstructing the local density field with combined convolutional and point cloud architecture NeurIPS 2025
We construct a neural network to perform regression on the local dark-matter density field given line-of-sight peculiar velocities of dark-matter halos, biased tracers of the dark matter field. Our architecture combines a convolutional U-Net with a point-cloud DeepSets. This combination enables efficient use of small-scale information and improves reconstruction quality relative to a U-Net-only approach. Specifically, our hybrid network recovers both clustering amplitudes and phases better than the U-Net on small scales.
comment: 6 pages, 4 figures, 1 table. Accepted at the NeurIPS 2025 Workshop: ML4PS. Comments welcome!
☆ Who Said Neural Networks Aren't Linear?
Neural networks are famously nonlinear. However, linearity is defined relative to a pair of vector spaces, $f$$:$$X$$\to$$Y$. Is it possible to identify a pair of non-standard vector spaces for which a conventionally nonlinear function is, in fact, linear? This paper introduces a method that makes such vector spaces explicit by construction. We find that if we sandwich a linear operator $A$ between two invertible neural networks, $f(x)=g_y^{-1}(A g_x(x))$, then the corresponding vector spaces $X$ and $Y$ are induced by newly defined addition and scaling actions derived from $g_x$ and $g_y$. We term this kind of architecture a Linearizer. This framework makes the entire arsenal of linear algebra, including SVD, pseudo-inverse, orthogonal projection and more, applicable to nonlinear mappings. Furthermore, we show that the composition of two Linearizers that share a neural network is also a Linearizer. We leverage this property and demonstrate that training diffusion models using our architecture makes the hundreds of sampling steps collapse into a single step. We further utilize our framework to enforce idempotency (i.e. $f(f(x))=f(x)$) on networks leading to a globally projective generative model and to demonstrate modular style transfer.
☆ ArenaBencher: Automatic Benchmark Evolution via Multi-Model Competitive Evaluation
Benchmarks are central to measuring the capabilities of large language models and guiding model development, yet widespread data leakage from pretraining corpora undermines their validity. Models can match memorized content rather than demonstrate true generalization, which inflates scores, distorts cross-model comparisons, and misrepresents progress. We introduce ArenaBencher, a model-agnostic framework for automatic benchmark evolution that updates test cases while preserving comparability. Given an existing benchmark and a diverse pool of models to be evaluated, ArenaBencher infers the core ability of each test case, generates candidate question-answer pairs that preserve the original objective, verifies correctness and intent with an LLM as a judge, and aggregates feedback from multiple models to select candidates that expose shared weaknesses. The process runs iteratively with in-context demonstrations that steer generation toward more challenging and diagnostic cases. We apply ArenaBencher to math problem solving, commonsense reasoning, and safety domains and show that it produces verified, diverse, and fair updates that uncover new failure modes, increase difficulty while preserving test objective alignment, and improve model separability. The framework provides a scalable path to continuously evolve benchmarks in step with the rapid progress of foundation models.
comment: Preprint
☆ How to Teach Large Multimodal Models New Skills
How can we teach large multimodal models (LMMs) new skills without erasing prior abilities? We study sequential fine-tuning on five target skills while monitoring general ability on eight held-out benchmarks across three model families. We observe that apparent "forgetting" on held-out tasks after narrow fine-tuning can partly recover at later stages. We trace this behavior to a measurable shift in the output token distribution, manifested through a simple counting-bias probe that co-varies with forgetting. Guided by this picture, we identify two simple, robust tuning recipes that learn strongly while limiting drift: (i) updating only the self-attention projection layers, and (ii) updating only the MLP Gate&Up while freezing the Down projection. Across models and tasks, these choices deliver strong target gains while largely preserving held-out performance. Code is available at https://github.com/jessemelpolio/LMM_CL
comment: In submission. Code is available at https://github.com/jessemelpolio/LMM_CL
☆ Where Have All the Kaczmarz Iterates Gone?
The randomized Kaczmarz (RK) algorithm is one of the most computationally and memory-efficient iterative algorithms for solving large-scale linear systems. However, practical applications often involve noisy and potentially inconsistent systems. While the convergence of RK is well understood for consistent systems, the study of RK on noisy, inconsistent linear systems is limited. This paper investigates the asymptotic behavior of RK iterates in expectation when solving noisy and inconsistent systems, addressing the locations of their limit points. We explore the roles of singular vectors of the (noisy) coefficient matrix and derive bounds on the convergence horizon, which depend on the noise levels and system characteristics. Finally, we provide extensive numerical experiments that validate our theoretical findings, offering practical insights into the algorithm's performance under realistic conditions. These results establish a deeper understanding of the RK algorithm's limitations and robustness in noisy environments, paving the way for optimized applications in real-world scientific and engineering problems.
Agent Learning via Early Experience
A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.
comment: Work in progress
☆ Improving Reasoning for Diffusion Language Models via Group Diffusion Policy Optimization
Diffusion language models (DLMs) enable parallel, order-agnostic generation with iterative refinement, offering a flexible alternative to autoregressive large language models (LLMs). However, adapting reinforcement learning (RL) fine-tuning to DLMs remains an open challenge because of the intractable likelihood. Pioneering work such as diffu-GRPO estimated token-level likelihoods via one-step unmasking. While computationally efficient, this approach is severely biased. A more principled foundation lies in sequence-level likelihoods, where the evidence lower bound (ELBO) serves as a surrogate. Yet, despite this clean mathematical connection, ELBO-based methods have seen limited adoption due to the prohibitive cost of likelihood evaluation. In this work, we revisit ELBO estimation and disentangle its sources of variance. This decomposition motivates reducing variance through fast, deterministic integral approximations along a few pivotal dimensions. Building on this insight, we introduce \textbf{Group Diffusion Policy Optimization (GDPO)}, a new RL algorithm tailored for DLMs. GDPO leverages simple yet effective Semi-deterministic Monte Carlo schemes to mitigate the variance explosion of ELBO estimators under vanilla double Monte Carlo sampling, yielding a provably lower-variance estimator under tight evaluation budgets. Empirically, GDPO achieves consistent gains over pretrained checkpoints and outperforms diffu-GRPO, one of the state-of-the-art baselines, on the majority of math, reasoning, and coding benchmarks.
☆ Entropy Regularizing Activation: Boosting Continuous Control, Large Language Models, and Image Classification with Activation as Entropy Constraints
We propose ERA, a new paradigm that constrains the sampling entropy above given thresholds by applying specially designed activations to the outputs of models. Our approach demonstrates broad effectiveness across different domains: 1) for large language models(LLMs), boosting the AIME 2025 score for Qwen2.5-Math-7B by 37.4%; 2) for continuous control reinforcement learning agents, improving performance by more than 30% over strong baselines such as SAC on the challenging HumanoidBench; 3) for image classification, enhancing ImageNet top-1 accuracy by 0.69% for ResNet-50. These gains are achieved with a computational overhead of less than 7%. Our work validates output activation as a powerful tool for entropy control, opening a new direction for designing simpler and more robust algorithms.
☆ SPAD: Specialized Prefill and Decode Hardware for Disaggregated LLM Inference
Large Language Models (LLMs) have gained popularity in recent years, driving up the demand for inference. LLM inference is composed of two phases with distinct characteristics: a compute-bound prefill phase followed by a memory-bound decode phase. To efficiently serve LLMs, prior work proposes prefill-decode disaggregation to run each phase on separate hardware. However, existing hardware poorly matches the different requirements of each phase. Current datacenter GPUs and TPUs follow a more-is-better design philosophy that maximizes compute and memory resources, causing memory bandwidth underutilization in the prefill phase and compute underutilization in the decode phase. Such underutilization directly translates into increased serving costs. This paper proposes SPAD (Specialized Prefill and Decode hardware), adopting a less-is-more methodology to design specialized chips tailored to the distinct characteristics of prefill and decode phases. The proposed Prefill Chips have larger systolic arrays and use cost-effective GDDR memory, whereas the proposed Decode Chips retain high memory bandwidth but reduce compute capacity. Compared to modeled H100s, simulations show that the proposed Prefill Chips deliver 8% higher prefill performance on average at 52% lower hardware cost, while the proposed Decode Chips achieve 97% of the decode performance with 28% lower TDP. End-to-end simulations on production traces show that SPAD reduces hardware cost by 19%-41% and TDP by 2%-17% compared to modeled baseline clusters while offering the same performance. Even when models and workloads change, SPAD can reallocate either type of chip to run either phase and still achieve 11%-43% lower hardware costs, demonstrating the longevity of the SPAD design.
☆ Computational and statistical lower bounds for low-rank estimation under general inhomogeneous noise
Recent work has generalized several results concerning the well-understood spiked Wigner matrix model of a low-rank signal matrix corrupted by additive i.i.d. Gaussian noise to the inhomogeneous case, where the noise has a variance profile. In particular, for the special case where the variance profile has a block structure, a series of results identified an effective spectral algorithm for detecting and estimating the signal, identified the threshold signal strength required for that algorithm to succeed, and proved information-theoretic lower bounds that, for some special signal distributions, match the above threshold. We complement these results by studying the computational optimality of this spectral algorithm. Namely, we show that, for a much broader range of signal distributions, whenever the spectral algorithm cannot detect a low-rank signal, then neither can any low-degree polynomial algorithm. This gives the first evidence for a computational hardness conjecture of Guionnet, Ko, Krzakala, and Zdeborov\'a (2023). With similar techniques, we also prove sharp information-theoretic lower bounds for a class of signal distributions not treated by prior work. Unlike all of the above results on inhomogeneous models, our results do not assume that the variance profile has a block structure, and suggest that the same spectral algorithm might remain optimal for quite general profiles. We include a numerical study of this claim for an example of a smoothly-varying rather than piecewise-constant profile. Our proofs involve analyzing the graph sums of a matrix, which also appear in free and traffic probability, but we require new bounds on these quantities that are tighter than existing ones for non-negative matrices, which may be of independent interest.
comment: 52 pages, 3 figures
☆ On the optimization dynamics of RLVR: Gradient gap and step size thresholds
Reinforcement Learning with Verifiable Rewards (RLVR), which uses simple binary feedback to post-train large language models, has shown significant empirical success. However, a principled understanding of why it works has been lacking. This paper builds a theoretical foundation for RLVR by analyzing its training process at both the full-response (trajectory) and token levels. Central to our analysis is a quantity called the Gradient Gap, which formalizes the direction of improvement from low-reward to high-reward regions of the response space. We prove that convergence critically depends on aligning the update direction with this Gradient Gap. Moreover, we derive a sharp step-size threshold based on the magnitude of the Gradient Gap: below it, learning converges, whereas above it, performance collapses. Our theory further predicts how the critical step size must scale with response length and the success rate, thereby explaining why practical heuristics such as length normalization improve stability and showing that, with a fixed learning rate, the success rate can stagnate strictly below $100\%$. We validate these predictions through controlled bandit simulations and LLM experiments, including training Qwen2.5-7B with GRPO.
☆ Permutation-Invariant Spectral Learning via Dyson Diffusion
Diffusion models are central to generative modeling and have been adapted to graphs by diffusing adjacency matrix representations. The challenge of having up to $n!$ such representations for graphs with $n$ nodes is only partially mitigated by using permutation-equivariant learning architectures. Despite their computational efficiency, existing graph diffusion models struggle to distinguish certain graph families, unless graph data are augmented with ad hoc features. This shortcoming stems from enforcing the inductive bias within the learning architecture. In this work, we leverage random matrix theory to analytically extract the spectral properties of the diffusion process, allowing us to push the inductive bias from the architecture into the dynamics. Building on this, we introduce the Dyson Diffusion Model, which employs Dyson's Brownian Motion to capture the spectral dynamics of an Ornstein-Uhlenbeck process on the adjacency matrix while retaining all non-spectral information. We demonstrate that the Dyson Diffusion Model learns graph spectra accurately and outperforms existing graph diffusion models.
☆ Convergence Theorems for Entropy-Regularized and Distributional Reinforcement Learning NeurIPS 2025
In the pursuit of finding an optimal policy, reinforcement learning (RL) methods generally ignore the properties of learned policies apart from their expected return. Thus, even when successful, it is difficult to characterize which policies will be learned and what they will do. In this work, we present a theoretical framework for policy optimization that guarantees convergence to a particular optimal policy, via vanishing entropy regularization and a temperature decoupling gambit. Our approach realizes an interpretable, diversity-preserving optimal policy as the regularization temperature vanishes and ensures the convergence of policy derived objects--value functions and return distributions. In a particular instance of our method, for example, the realized policy samples all optimal actions uniformly. Leveraging our temperature decoupling gambit, we present an algorithm that estimates, to arbitrary accuracy, the return distribution associated to its interpretable, diversity-preserving optimal policy.
comment: Accepted to NeurIPS 2025. First two authors contributed equally
☆ DYNAMIX: RL-based Adaptive Batch Size Optimization in Distributed Machine Learning Systems
Existing batch size selection approaches in distributed machine learning rely on static allocation or simplistic heuristics that fail to adapt to heterogeneous, dynamic computing environments. We present DYNAMIX, a reinforcement learning framework that formulates batch size optimization as a sequential decision-making problem using Proximal Policy Optimization (PPO). Our approach employs a multi-dimensional state representation encompassing network-level metrics, system-level resource utilization, and training statistical efficiency indicators to enable informed decision-making across diverse computational resources. Our approach eliminates the need for explicit system modeling while integrating seamlessly with existing distributed training frameworks. Through evaluations across diverse workloads, hardware configurations, and network conditions, DYNAMIX achieves up to 6.3% improvement in the final model accuracy and 46% reduction in the total training time. Our scalability experiments demonstrate that DYNAMIX maintains the best performance as cluster size increases to 32 nodes, while policy transfer experiments show that learned policies generalize effectively across related model architectures.
☆ CaRT: Teaching LLM Agents to Know When They Know Enough
Many tasks require learned models to strategically gather relevant information over multiple rounds of interaction before actually acting on a task. Strategic information gathering requires models to know not only how to effectively acquire information, but also when to stop gathering information and make a decision, in order to avoid overthinking or getting derailed when acting. In this paper, we formalize this problem and introduce Counterfactuals and Reasoning for Termination (CaRT), an approach for teaching LLMs when to stop seeking information. To appropriately learn when to terminate, CaRT fine-tunes LLMs using counterfactual pairs of trajectories, one where termination is appropriate and a minimally modified version of the same trajectory where it is not. It trains the LLM to explain the rationale for the termination decision in either case via verbal reasoning, and imbues this capability into the base LLM via fine-tuning. We instantiate CaRT in two domains: interactive medical diagnosis and math problem solving. In both domains, we find that CaRT improves the efficiency of information gathering and task success rate compared to other fine-tuning methods.
☆ AutoMLGen: Navigating Fine-Grained Optimization for Coding Agents
Large language models (LLMs) have shown impressive performance in general programming tasks. However, in Machine Learning Engineering (MLE) scenarios such as AutoML and Kaggle competitions, achieving high performance depends heavily on expert intervention and repeated adjustments rather than simply generating correct code. When applied directly to these tasks, LLMs often lack fine-grained domain priors, and existing MLE approaches that use linear or tree-structured searches limit knowledge transfer to adjacent hierarchical links. As a result, they cannot leverage past full trajectories or share information across branches, limiting self-evolving ability and search space diversity. To address these limitations, we introduce AutoMLGen, an LLM-based coding agent that integrates a domain knowledge base for high-quality prior guidance and Monte Carlo Graph Search (MCGS) for efficient exploration. MCGS retains the tree-guided exploration of MCTS while embedding a graph structure into the expansion stage to enable dynamic path reorganization, historical trajectory reuse, and multi-solution fusion to support both self-evolution and collaborative learning. Combined with fine-grained operator sets, this design improves stability and accelerates convergence. Evaluation on the MLE-Bench shows that AutoMLGen achieves state-of-the-art performance in numerous dimensions, such as the average medal rate and the valid submission rate, under a 12-hour budget (half the standard runtime). The code is available at https://github.com/Alpha-Innovator/InternAgent.
☆ AI-Driven Radiology Report Generation for Traumatic Brain Injuries
Traumatic brain injuries present significant diagnostic challenges in emergency medicine, where the timely interpretation of medical images is crucial for patient outcomes. In this paper, we propose a novel AI-based approach for automatic radiology report generation tailored to cranial trauma cases. Our model integrates an AC-BiFPN with a Transformer architecture to capture and process complex medical imaging data such as CT and MRI scans. The AC-BiFPN extracts multi-scale features, enabling the detection of intricate anomalies like intracranial hemorrhages, while the Transformer generates coherent, contextually relevant diagnostic reports by modeling long-range dependencies. We evaluate the performance of our model on the RSNA Intracranial Hemorrhage Detection dataset, where it outperforms traditional CNN-based models in both diagnostic accuracy and report generation. This solution not only supports radiologists in high-pressure environments but also provides a powerful educational tool for trainee physicians, offering real-time feedback and enhancing their learning experience. Our findings demonstrate the potential of combining advanced feature extraction with transformer-based text generation to improve clinical decision-making in the diagnosis of traumatic brain injuries.
☆ Better Together: Leveraging Unpaired Multimodal Data for Stronger Unimodal Models
Traditional multimodal learners find unified representations for tasks like visual question answering, but rely heavily on paired datasets. However, an overlooked yet potentially powerful question is: can one leverage auxiliary unpaired multimodal data to directly enhance representation learning in a target modality? We introduce UML: Unpaired Multimodal Learner, a modality-agnostic training paradigm in which a single model alternately processes inputs from different modalities while sharing parameters across them. This design exploits the assumption that different modalities are projections of a shared underlying reality, allowing the model to benefit from cross-modal structure without requiring explicit pairs. Theoretically, under linear data-generating assumptions, we show that unpaired auxiliary data can yield representations strictly more informative about the data-generating process than unimodal training. Empirically, we show that using unpaired data from auxiliary modalities -- such as text, audio, or images -- consistently improves downstream performance across diverse unimodal targets such as image and audio. Our project page: https://unpaired-multimodal.github.io/
comment: 63 pages, 29 tables, and 47 figures
☆ Implementing Semantic Join Operators Efficiently
Semantic query processing engines often support semantic joins, enabling users to match rows that satisfy conditions specified in natural language. Such join conditions can be evaluated using large language models (LLMs) that solve novel tasks without task-specific training. Currently, many semantic query processing engines implement semantic joins via nested loops, invoking the LLM to evaluate the join condition on row pairs. Instead, this paper proposes a novel algorithm, inspired by the block nested loops join operator implementation in traditional database systems. The proposed algorithm integrates batches of rows from both input tables into a single prompt. The goal of the LLM invocation is to identify all matching row pairs in the current input. The paper introduces formulas that can be used to optimize the size of the row batches, taking into account constraints on the size of the LLM context window (limiting both input and output size). An adaptive variant of the proposed algorithm refers to cases in which the size of the output is difficult to estimate. A formal analysis of asymptotic processing costs, as well as empirical results, demonstrates that the proposed approach reduces costs significantly and performs well compared to join implementations used by recent semantic query processing engines.
☆ DexMan: Learning Bimanual Dexterous Manipulation from Human and Generated Videos
We present DexMan, an automated framework that converts human visual demonstrations into bimanual dexterous manipulation skills for humanoid robots in simulation. Operating directly on third-person videos of humans manipulating rigid objects, DexMan eliminates the need for camera calibration, depth sensors, scanned 3D object assets, or ground-truth hand and object motion annotations. Unlike prior approaches that consider only simplified floating hands, it directly controls a humanoid robot and leverages novel contact-based rewards to improve policy learning from noisy hand-object poses estimated from in-the-wild videos. DexMan achieves state-of-the-art performance in object pose estimation on the TACO benchmark, with absolute gains of 0.08 and 0.12 in ADD-S and VSD. Meanwhile, its reinforcement learning policy surpasses previous methods by 19% in success rate on OakInk-v2. Furthermore, DexMan can generate skills from both real and synthetic videos, without the need for manual data collection and costly motion capture, and enabling the creation of large-scale, diverse datasets for training generalist dexterous manipulation.
comment: Video results are available at: https://embodiedai-ntu.github.io/dexman/index.html
☆ Looking to Learn: Token-wise Dynamic Gating for Low-Resource Vision-Language Modelling EMNLP 2025
Training vision-language models on cognitively-plausible amounts of data requires rethinking how models integrate multimodal information. Within the constraints of the Vision track for the BabyLM Challenge 2025, we propose a lightweight decoder-based architecture with (1) token-wise dynamic gating for adaptive fusion of linguistic and visual cues, (2) feature modulation and channel attention to maximise the utility of limited visual information and (3) auxiliary contrastive objectives for visual grounding. Evaluation on five benchmarks (BLiMP, BLiMP Supplement, EWoK, Winoground and VQA) shows competitive or superior performance to multimodal baselines. More notably, our dynamic gate discovers interpretable patterns without explicit supervision, favouring visual cues for content words and linguistic cues for function words. While we identify limitations in the Challenge constraints, such as the information bottleneck created by global image embeddings and training instability from the dataset split, our findings establish dynamic gating as a powerful tool for efficient multimodal learning, offering both interpretability and performance even under severe constraints.
comment: Accepted to the EMNLP 2025 BabyLM Workshop
☆ In-Context Clustering with Large Language Models
We propose In-Context Clustering (ICC), a flexible LLM-based procedure for clustering data from diverse distributions. Unlike traditional clustering algorithms constrained by predefined similarity measures, ICC flexibly captures complex relationships among inputs through an attention mechanism. We show that pretrained LLMs exhibit impressive zero-shot clustering capabilities on text-encoded numeric data, with attention matrices showing salient cluster patterns. Spectral clustering using attention matrices offers surprisingly competitive performance. We further enhance the clustering capabilities of LLMs on numeric and image data through fine-tuning using the Next Token Prediction (NTP) loss. Moreover, the flexibility of LLM prompting enables text-conditioned image clustering, a capability that classical clustering methods lack. Our work extends in-context learning to an unsupervised setting, showcasing the effectiveness and flexibility of LLMs for clustering. Our code is available at https://agenticlearning.ai/icc.
☆ Accelerated Aggregated D-Optimal Designs for Estimating Main Effects in Black-Box Models
Recent advances in supervised learning have driven growing interest in explaining black-box models, particularly by estimating the effects of input variables on model predictions. However, existing approaches often face key limitations, including poor scalability, sensitivity to out-of-distribution sampling, and instability under correlated features. To address these issues, we propose A2D2E, an $\textbf{E}$stimator based on $\textbf{A}$ccelerated $\textbf{A}$ggregated $\textbf{D}$-Optimal $\textbf{D}$esigns. Our method leverages principled experimental design to improve efficiency and robustness in main effect estimation. We establish theoretical guarantees, including convergence and variance reduction, and validate A2D2E through extensive simulations. We further provide the potential of the proposed method with a case study on real data and applications in language models. The code to reproduce the results can be found at https://github.com/cchihyu/A2D2E.
☆ Don't Run with Scissors: Pruning Breaks VLA Models but They Can Be Recovered
Vision-Language-Action (VLA) models have advanced robotic capabilities but remain challenging to deploy on resource-limited hardware. Pruning has enabled efficient compression of large language models (LLMs), yet it is largely understudied in robotics. Surprisingly, we observe that pruning VLA models leads to drastic degradation and increased safety violations. We introduce GLUESTICK, a post-pruning recovery method that restores much of the original model's functionality while retaining sparsity benefits. Our method performs a one-time interpolation between the dense and pruned models in weight-space to compute a corrective term. This correction is used during inference by each pruned layer to recover lost capabilities with minimal overhead. GLUESTICK requires no additional training, is agnostic to the pruning algorithm, and introduces a single hyperparameter that controls the tradeoff between efficiency and accuracy. Across diverse VLA architectures and tasks in manipulation and navigation, GLUESTICK achieves competitive memory efficiency while substantially recovering success rates and reducing safety violations. Additional material can be found at: https://gluestick-vla.github.io/.
☆ Wavefunction Flows: Efficient Quantum Simulation of Continuous Flow Models
Flow models are a cornerstone of modern machine learning. They are generative models that progressively transform probability distributions according to learned dynamics. Specifically, they learn a continuous-time Markov process that efficiently maps samples from a simple source distribution into samples from a complex target distribution. We show that these models are naturally related to the Schr\"odinger equation, for an unusual Hamiltonian on continuous variables. Moreover, we prove that the dynamics generated by this Hamiltonian can be efficiently simulated on a quantum computer. Together, these results give a quantum algorithm for preparing coherent encodings (a.k.a., qsamples) for a vast family of probability distributions--namely, those expressible by flow models--by reducing the task to an existing classical learning problem, plus Hamiltonian simulation. For statistical problems defined by flow models, such as mean estimation and property testing, this enables the use of quantum algorithms tailored to qsamples, which may offer advantages over classical algorithms based only on samples from a flow model. More broadly, these results reveal a close connection between state-of-the-art machine learning models, such as flow matching and diffusion models, and one of the main expected capabilities of quantum computers: simulating quantum dynamics.
☆ SummDiff: Generative Modeling of Video Summarization with Diffusion
Video summarization is a task of shortening a video by choosing a subset of frames while preserving its essential moments. Despite the innate subjectivity of the task, previous works have deterministically regressed to an averaged frame score over multiple raters, ignoring the inherent subjectivity of what constitutes a good summary. We propose a novel problem formulation by framing video summarization as a conditional generation task, allowing a model to learn the distribution of good summaries and to generate multiple plausible summaries that better reflect varying human perspectives. Adopting diffusion models for the first time in video summarization, our proposed method, SummDiff, dynamically adapts to visual contexts and generates multiple candidate summaries conditioned on the input video. Extensive experiments demonstrate that SummDiff not only achieves the state-of-the-art performance on various benchmarks but also produces summaries that closely align with individual annotator preferences. Moreover, we provide a deeper insight with novel metrics from an analysis of the knapsack, which is an important last step of generating summaries but has been overlooked in evaluation.
☆ Integral Signatures of Activation Functions: A 9-Dimensional Taxonomy and Stability Theory for Deep Learning
Activation functions govern the expressivity and stability of neural networks, yet existing comparisons remain largely heuristic. We propose a rigorous framework for their classification via a nine-dimensional integral signature S_sigma(phi), combining Gaussian propagation statistics (m1, g1, g2, m2, eta), asymptotic slopes (alpha_plus, alpha_minus), and regularity measures (TV(phi'), C(phi)). This taxonomy establishes well-posedness, affine reparameterization laws with bias, and closure under bounded slope variation. Dynamical analysis yields Lyapunov theorems with explicit descent constants and identifies variance stability regions through (m2', g2). From a kernel perspective, we derive dimension-free Hessian bounds and connect smoothness to bounded variation of phi'. Applying the framework, we classify eight standard activations (ReLU, leaky-ReLU, tanh, sigmoid, Swish, GELU, Mish, TeLU), proving sharp distinctions between saturating, linear-growth, and smooth families. Numerical Gauss-Hermite and Monte Carlo validation confirms theoretical predictions. Our framework provides principled design guidance, moving activation choice from trial-and-error to provable stability and kernel conditioning.
comment: 25 pages
☆ gLSTM: Mitigating Over-Squashing by Increasing Storage Capacity
Graph Neural Networks (GNNs) leverage the graph structure to transmit information between nodes, typically through the message-passing mechanism. While these models have found a wide variety of applications, they are known to suffer from over-squashing, where information from a large receptive field of node representations is collapsed into a single fixed sized vector, resulting in an information bottleneck. In this paper, we re-examine the over-squashing phenomenon through the lens of model storage and retrieval capacity, which we define as the amount of information that can be stored in a node's representation for later use. We study some of the limitations of existing tasks used to measure over-squashing and introduce a new synthetic task to demonstrate that an information bottleneck can saturate this capacity. Furthermore, we adapt ideas from the sequence modeling literature on associative memories, fast weight programmers, and the xLSTM model to develop a novel GNN architecture with improved capacity. We demonstrate strong performance of this architecture both on our capacity synthetic task, as well as a range of real-world graph benchmarks.
comment: 22 pages, 22 figures, 7 tables
☆ Synthetic Series-Symbol Data Generation for Time Series Foundation Models NeurIPS 2025
Foundation models for time series analysis (TSA) have attracted significant attention. However, challenges such as training data scarcity and imbalance continue to hinder their development. Inspired by complex dynamic system theories, we design a series-symbol data generation mechanism, enabling the unrestricted creation of high-quality time series data paired with corresponding symbolic expressions. To leverage series-symbol data pairs with strong correlations, we develop \texttt{SymTime}, a pre-trained foundation model for enhancing time series representation using symbolic information. \texttt{SymTime} demonstrates competitive performance across five major TSA tasks when fine-tunes with downstream tasks, rivaling foundation models pre-trained on real-world datasets. This approach underscores the potential of series-symbol data generation and pretraining mechanisms in overcoming data scarcity and enhancing task performance. The code is available at https://github.com/wwhenxuan/SymTime.
comment: 63 pages, NeurIPS 2025 accepted
☆ xRouter: Training Cost-Aware LLMs Orchestration System via Reinforcement Learning
Modern LLM deployments confront a widening cost-performance spectrum: premium models deliver strong reasoning but are expensive, while lightweight models are economical yet brittle on complex tasks. Static escalation rules and keyword heuristics under-utilize this spectrum and fail to adapt across task types. We present xRouter, a tool-calling-based routing system in which a learned router can either answer directly or invoke one or more external models. The router is trained end-to-end with reinforcement learning using an explicit, cost-aware reward that encodes cost-performance trade-offs, eliminating the need for hand-engineered routing rules. Our implementation encompasses the full reinforcement learning framework, including reward and cost accounting, as well as the deployment and evaluation pipelines. Across diverse benchmarks, xRouter achieves strong cost-performance trade-offs (e.g., substantial cost reductions at comparable task completion rates), and provides empirical insights into what reliably helps learned routing and what does not, ranging from model trainability to the difficulty of eliciting sophisticated orchestration behaviors in small open models. We hope these findings and our open implementation will serve as a practical substrate for advancing learned, cost-aware LLM orchestration.
comment: 24 Pages, 4 Figures, 2 Tables
☆ Navigating Sparsities in High-Dimensional Linear Contextual Bandits
High-dimensional linear contextual bandit problems remain a significant challenge due to the curse of dimensionality. Existing methods typically consider either the model parameters to be sparse or the eigenvalues of context covariance matrices to be (approximately) sparse, lacking general applicability due to the rigidity of conventional reward estimators. To overcome this limitation, a powerful pointwise estimator is introduced in this work that adaptively navigates both kinds of sparsity. Based on this pointwise estimator, a novel algorithm, termed HOPE, is proposed. Theoretical analyses demonstrate that HOPE not only achieves improved regret bounds in previously discussed homogeneous settings (i.e., considering only one type of sparsity) but also, for the first time, efficiently handles two new challenging heterogeneous settings (i.e., considering a mixture of two types of sparsity), highlighting its flexibility and generality. Experiments corroborate the superiority of HOPE over existing methods across various scenarios.
☆ Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency
This work represents the first effort to scale up continuous-time consistency distillation to general application-level image and video diffusion models. Although continuous-time consistency model (sCM) is theoretically principled and empirically powerful for accelerating academic-scale diffusion, its applicability to large-scale text-to-image and video tasks remains unclear due to infrastructure challenges in Jacobian-vector product (JVP) computation and the limitations of standard evaluation benchmarks. We first develop a parallelism-compatible FlashAttention-2 JVP kernel, enabling sCM training on models with over 10 billion parameters and high-dimensional video tasks. Our investigation reveals fundamental quality limitations of sCM in fine-detail generation, which we attribute to error accumulation and the "mode-covering" nature of its forward-divergence objective. To remedy this, we propose the score-regularized continuous-time consistency model (rCM), which incorporates score distillation as a long-skip regularizer. This integration complements sCM with the "mode-seeking" reverse divergence, effectively improving visual quality while maintaining high generation diversity. Validated on large-scale models (Cosmos-Predict2, Wan2.1) up to 14B parameters and 5-second videos, rCM matches or surpasses the state-of-the-art distillation method DMD2 on quality metrics while offering notable advantages in diversity, all without GAN tuning or extensive hyperparameter searches. The distilled models generate high-fidelity samples in only $1\sim4$ steps, accelerating diffusion sampling by $15\times\sim50\times$. These results position rCM as a practical and theoretically grounded framework for advancing large-scale diffusion distillation.
☆ ClauseLens: Clause-Grounded, CVaR-Constrained Reinforcement Learning for Trustworthy Reinsurance Pricing
Reinsurance treaty pricing must satisfy stringent regulatory standards, yet current quoting practices remain opaque and difficult to audit. We introduce ClauseLens, a clause-grounded reinforcement learning framework that produces transparent, regulation-compliant, and risk-aware treaty quotes. ClauseLens models the quoting task as a Risk-Aware Constrained Markov Decision Process (RA-CMDP). Statutory and policy clauses are retrieved from legal and underwriting corpora, embedded into the agent's observations, and used both to constrain feasible actions and to generate clause-grounded natural language justifications. Evaluated in a multi-agent treaty simulator calibrated to industry data, ClauseLens reduces solvency violations by 51%, improves tail-risk performance by 27.9% (CVaR_0.10), and achieves 88.2% accuracy in clause-grounded explanations with retrieval precision of 87.4% and recall of 91.1%. These findings demonstrate that embedding legal context into both decision and explanation pathways yields interpretable, auditable, and regulation-aligned quoting behavior consistent with Solvency II, NAIC RBC, and the EU AI Act.
comment: Accepted for publication at the 6th ACM International Conference on AI in Finance (ICAIF 2025), Singapore. Author-accepted version (October 2025). 10 pages, 5 figures
☆ Reinforcing Diffusion Models by Direct Group Preference Optimization
While reinforcement learning methods such as Group Relative Preference Optimization (GRPO) have significantly enhanced Large Language Models, adapting them to diffusion models remains challenging. In particular, GRPO demands a stochastic policy, yet the most cost-effective diffusion samplers are based on deterministic ODEs. Recent work addresses this issue by using inefficient SDE-based samplers to induce stochasticity, but this reliance on model-agnostic Gaussian noise leads to slow convergence. To resolve this conflict, we propose Direct Group Preference Optimization (DGPO), a new online RL algorithm that dispenses with the policy-gradient framework entirely. DGPO learns directly from group-level preferences, which utilize relative information of samples within groups. This design eliminates the need for inefficient stochastic policies, unlocking the use of efficient deterministic ODE samplers and faster training. Extensive results show that DGPO trains around 20 times faster than existing state-of-the-art methods and achieves superior performance on both in-domain and out-of-domain reward metrics. Code is available at https://github.com/Luo-Yihong/DGPO.
Prompts Generalize with Low Data: Non-vacuous Generalization Bounds for Optimizing Prompts with More Informative Priors ICML 2025
Many prompt engineering techniques have been successful in practice, even when optimizing over a large prompt space with with a small amount of task-specific data. Recent work has partially explained this success by showing generalization bounds which apply PAC-Bayes theory to the discrete prompt space, but they are non-vacuous only in data-rich scenarios. We argue that such widespread success can be more fully explained through more carefully considering data- or distribution-dependent perplexity, which acts as an effective prior and steers the optimization towards prompts that are more ``natural'' for the task at hand. We derive novel generalization bounds that are non-vacuous for data-scarce prompt optimization via more useful priors, formally analyzing how perplexity regularization tightens these bounds by limiting exploration. Empirically, we explore both the bounds' effectiveness and the practical benefits of perplexity regularization in improving prompt generalization.
comment: EXAIT Workshop paper at ICML 2025
☆ Optimal Stopping in Latent Diffusion Models
We identify and analyze a surprising phenomenon of Latent Diffusion Models (LDMs) where the final steps of the diffusion can degrade sample quality. In contrast to conventional arguments that justify early stopping for numerical stability, this phenomenon is intrinsic to the dimensionality reduction in LDMs. We provide a principled explanation by analyzing the interaction between latent dimension and stopping time. Under a Gaussian framework with linear autoencoders, we characterize the conditions under which early stopping is needed to minimize the distance between generated and target distributions. More precisely, we show that lower-dimensional representations benefit from earlier termination, whereas higher-dimensional latent spaces require later stopping time. We further establish that the latent dimension interplays with other hyperparameters of the problem such as constraints in the parameters of score matching. Experiments on synthetic and real datasets illustrate these properties, underlining that early stopping can improve generative quality. Together, our results offer a theoretical foundation for understanding how the latent dimension influences the sample quality, and highlight stopping time as a key hyperparameter in LDMs.
☆ Biology-driven assessment of deep learning super-resolution imaging of the porosity network in dentin
The mechanosensory system of teeth is currently believed to partly rely on Odontoblast cells stimulation by fluid flow through a porosity network extending through dentin. Visualizing the smallest sub-microscopic porosity vessels therefore requires the highest achievable resolution from confocal fluorescence microscopy, the current gold standard. This considerably limits the extent of the field of view to very small sample regions. To overcome this limitation, we tested different deep learning (DL) super-resolution (SR) models to allow faster experimental acquisitions of lower resolution images and restore optimal image quality by post-processing. Three supervised 2D SR models (RCAN, pix2pix, FSRCNN) and one unsupervised (CycleGAN) were applied to a unique set of experimentally paired high- and low-resolution confocal images acquired with different sampling schemes, resulting in a pixel size increase of x2, x4, x8. Model performance was quantified using a broad set of similarity and distribution-based image quality assessment (IQA) metrics, which yielded inconsistent results that mostly contradicted our visual perception. This raises the question of the relevance of such generic metrics to efficiently target the specific structure of dental porosity. To resolve this conflicting information, the generated SR images were segmented taking into account the specific scales and morphology of the porosity network and analysed by comparing connected components. Additionally, the capacity of the SR models to preserve 3D porosity connectivity throughout the confocal image stacks was evaluated using graph analysis. This biology-driven assessment allowed a far better mechanistic interpretation of SR performance, highlighting differences in model sensitivity to weak intensity features and the impact of non-linearity in image generation, which explains the failure of standard IQA metrics.
☆ Single layer tiny Co$^4$ outpaces GPT-2 and GPT-BERT
We show that a tiny Co$^4$ machine(Adeel,2025) with a single layer, two heads, and 8M parameters, operating at an approximate cost of $O(N)$ (where $N$ is the number of input tokens), outpaces the BabyLM Challenge baselines GPT-2 (124M, 12 layers, $O(N^2))$ and GPT-BERT (30M, 12 layers, $O(N^2))$ in just two epochs, while both are trained for ten. Co$^4$ achieves orders-of-magnitude greater training efficiency on 10M tokens, demonstrating highly sample efficient pretraining. Using the BabyLM challenge evaluation pipeline across complex benchmarks, Co$^4$ exhibits strong zero-shot and fine-tuning performance on SuperGLUE tasks. Specifically, Co$^4$ outperforms GPT-2 on 5 out of 7 zero-shot metrics and 6 out of 7 fine-tuning tasks, and GPT-BERT on 4 out of 7 metrics in both cases. These results suggest the need to rethink prevailing deep learning paradigms and associated scaling laws.
☆ FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts NeurIPS 2025
Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning method for foundation models, but it suffers from parameter interference, resulting in suboptimal performance. Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-task correlations in single-task instruction tuning, they introduce additional router parameters and remain ineffective in multi-task model merging where inter-task interference arises. Inspired by the fly olfactory circuit, we propose FlyLoRA, an implicit MoE-based LoRA variant that introduces: (1) rank-wise expert activation in the up-projection matrix, and (2) an implicit router that unifies expert routing and down-projection, where a frozen sparse random projection matrix replaces the traditional dense trainable version. This design resolves the trade-off between intra-task decorrelation and computational efficiency by eliminating the need for an explicit router, while inherently mitigating inter-task interference due to the orthogonality property of random matrices. Extensive experiments across four domains -- general knowledge understanding, scientific question answering, mathematical reasoning, and code generation -- demonstrate consistent performance improvements over existing methods. Beyond empirical gains, FlyLoRA highlights how biological structures can inspire innovations in AI technologies. Code is available at https://github.com/gfyddha/FlyLoRA.
comment: NeurIPS 2025 accepted paper
☆ Characterizing the Multiclass Learnability of Forgiving 0-1 Loss Functions
In this paper we will give a characterization of the learnability of forgiving 0-1 loss functions in the finite label multiclass setting. To do this, we create a new combinatorial dimension that is based off of the Natarajan Dimension \citep{natarajan1989learning} and we show that a hypothesis class is learnable in our setting if and only if this Generalized Natarajan Dimension is finite. We also show a connection to learning with set-valued feedback. Through our results we show that the learnability of a set learning problem is characterized by the Natarajan Dimension.
comment: 9 pages
☆ Contrastive Self-Supervised Learning at the Edge: An Energy Perspective
While contrastive learning (CL) shows considerable promise in self-supervised representation learning, its deployment on resource-constrained devices remains largely underexplored. The substantial computational demands required for training conventional CL frameworks pose a set of challenges, particularly in terms of energy consumption, data availability, and memory usage. We conduct an evaluation of four widely used CL frameworks: SimCLR, MoCo, SimSiam, and Barlow Twins. We focus on the practical feasibility of these CL frameworks for edge and fog deployment, and introduce a systematic benchmarking strategy that includes energy profiling and reduced training data conditions. Our findings reveal that SimCLR, contrary to its perceived computational cost, demonstrates the lowest energy consumption across various data regimes. Finally, we also extend our analysis by evaluating lightweight neural architectures when paired with CL frameworks. Our study aims to provide insights into the resource implications of deploying CL in edge/fog environments with limited processing capabilities and opens several research directions for its future optimization.
☆ On the Relationship Between the Choice of Representation and In-Context Learning
In-context learning (ICL) is the ability of a large language model (LLM) to learn a new task from a few demonstrations presented as part of the context. Past studies have attributed a large portion of the success of ICL to the way these in-context demonstrations are represented, particularly to how labels are represented in classification tasks. On the other hand, observations of the learning capacity of ICL (i.e., the extent to which more in-context demonstrations can lead to higher performance) have been mixed, and ICL is often thought to occur only under specific conditions. The interaction between these two aspects in ICL, representation and learning, has not been studied in depth until now. We hypothesize that they are largely independent of one another, such that the representation of demonstrations determines the baseline accuracy of ICL, while learning from additional demonstrations improves only on top of this baseline. We validate this hypothesis by developing an optimization algorithm that can enumerate a spectrum of possible label sets (representations) varying in semantic relevance. We then perform ICL with varying numbers of in-context demonstrations for each of these label sets. We observed that learning happens regardless of the quality of the label set itself, although its efficiency, measured by the slope of improvement over in-context demonstrations, is conditioned on both the label set quality and the parameter count of the underlying language model. Despite the emergence of learning, the relative quality (accuracy) of the choice of a label set (representation) is largely maintained throughout learning, confirming our hypothesis and implying their orthogonality. Our work reveals a previously underexplored aspect of ICL: the independent effects of learning from demonstrations and their representations on ICL performance.
comment: 25 pages, 6 figures, 10 tables
☆ Guided Star-Shaped Masked Diffusion
The performance of pre-trained masked diffusion models is often constrained by their sampling procedure, which makes decisions irreversible and struggles in low-step generation regimes. We introduce a novel sampling algorithm that works with pre-trained models and, after a lightweight fine-tuning of a single layer, significantly improves sample quality and efficiency. Our method reformulates the generation process using a star-shaped paradigm, which inherently allows for error correction. To make this process effective, we augment it with a learnable re-masking scheduler that intelligently identifies and revises likely errors. This approach yields a substantial quality boost, particularly when using a small number of sampling steps. We extensively ablate key components of our approach and show its usability in different scenarios. In comprehensive experiments on text, and code generation, our sampling algorithm outperforms or matches existing methods.
☆ DeepEN: Personalized Enteral Nutrition for Critically Ill Patients using Deep Reinforcement Learning
We introduce DeepEN, a deep reinforcement learning (RL) framework for personalized enteral nutrition (EN) in critically ill patients. Trained offline on over 11,000 ICU patients from the MIMIC-IV database, DeepEN generates 4-hourly recommendations for caloric, protein, and fluid intake tailored to each patient's evolving physiology. The model integrates a curated, clinically informed state space with a custom reward function that balances short-term physiological and nutrition-related goals with long-term survival outcomes. Using a dueling double deep Q-network with conservative Q-learning regularization, DeepEN learns clinically realistic policies that align with high-value clinician actions while discouraging unsafe deviations. Across various qualitative and quantitative metrics, DeepEN outperforms clinician-derived and guideline-based policies, achieving a 3.7 $\pm$ 0.17 percentage-point reduction in estimated mortality (18.8% vs 22.5%) and improvements in key nutritional biomarkers. These findings highlight the potential of safe, data-driven personalization of EN therapy to improve outcomes beyond traditional guideline- or heuristic-based approaches.
☆ Learning What's Missing: Attention Dispersion and EMA Stabilization in Length Generalization
We study length generalization in transformers through the set complement task, where a model must predict a uniform distribution over tokens absent from an input sequence -- an ability central to board-game style reasoning. Our main theoretical result establishes two statements. First, we prove tight bounds on embedding and value dimensions for single-layer attention-only transformers. Second, we show that if such a model achieves balanced logit displacement at lengths 1 and 2, then it must generalize to longer sequences, though with reduced precision. A mechanistic reading of the proof explains this limitation: as more tokens are attended to, softmax compresses logit displacements, eroding separation between valid and invalid outputs. Training dynamics also suggest a second obstacle: when many next tokens are possible, updates become noisy. We hypothesize that dropout can counteract the first effect and Exponential Moving Average (EMA) the second. We validate these hypotheses through random hyperparameter search on the set complement task, which confirms both mechanisms. We then test OthelloGPT, a GPT-1 style model trained on random Othello moves, and find that EMA again improves length generalization in this more complex setting.
comment: 10 pages, 5 figures, 2 tables
☆ PAC Learnability in the Presence of Performativity
Following the wide-spread adoption of machine learning models in real-world applications, the phenomenon of performativity, i.e. model-dependent shifts in the test distribution, becomes increasingly prevalent. Unfortunately, since models are usually trained solely based on samples from the original (unshifted) distribution, this performative shift may lead to decreased test-time performance. In this paper, we study the question of whether and when performative binary classification problems are learnable, via the lens of the classic PAC (Probably Approximately Correct) learning framework. We motivate several performative scenarios, accounting in particular for linear shifts in the label distribution, as well as for more general changes in both the labels and the features. We construct a performative empirical risk function, which depends only on data from the original distribution and on the type performative effect, and is yet an unbiased estimate of the true risk of a classifier on the shifted distribution. Minimizing this notion of performative risk allows us to show that any PAC-learnable hypothesis space in the standard binary classification setting remains PAC-learnable for the considered performative scenarios. We also conduct an extensive experimental evaluation of our performative risk minimization method and showcase benefits on synthetic and real data.
comment: 21 pages, 3 figures
☆ New Machine Learning Approaches for Intrusion Detection in ADS-B SC
With the growing reliance on the vulnerable Automatic Dependent Surveillance-Broadcast (ADS-B) protocol in air traffic management (ATM), ensuring security is critical. This study investigates emerging machine learning models and training strategies to improve AI-based intrusion detection systems (IDS) for ADS-B. Focusing on ground-based ATM systems, we evaluate two deep learning IDS implementations: one using a transformer encoder and the other an extended Long Short-Term Memory (xLSTM) network, marking the first xLSTM-based IDS for ADS-B. A transfer learning strategy was employed, involving pre-training on benign ADS-B messages and fine-tuning with labeled data containing instances of tampered messages. Results show this approach outperforms existing methods, particularly in identifying subtle attacks that progressively undermine situational awareness. The xLSTM-based IDS achieves an F1-score of 98.9%, surpassing the transformer-based model at 94.3%. Tests on unseen attacks validated the generalization ability of the xLSTM model. Inference latency analysis shows that the 7.26-second delay introduced by the xLSTM-based IDS fits within the Secondary Surveillance Radar (SSR) refresh interval (5-12 s), although it may be restrictive for time-critical operations. While the transformer-based IDS achieves a 2.1-second latency, it does so at the cost of lower detection performance.
comment: This is the author's version of the work accepted for publication Digital Avionics Systems Conference (DASC) 2025. The final version will be available via IEEE Xplore
☆ Beyond Pass@k: Breadth-Depth Metrics for Reasoning Boundaries
Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm to improve Large Language Models on reasoning tasks such as coding, math or logic. To assess the reasoning boundary (the fraction of problems a model can solve) researchers often report Pass@k at large sampling budgets. Recent results reveal a crossover phenomenon: while RLVR models outperform the base model at small k values, the base model usually outperforms them when sampling a very large number of completions. This has been interpreted as evidence that base models have a larger reasoning boundary. We argue that on tasks with discrete answer spaces, such as math with numeric outputs, Pass@k at large k reflects the increasingly higher chance of success in the limit of the number of trials rather than genuine reasoning, and can therefore be misleading. We propose Cover@tau, which measures the fraction of problems that a model can solve for which at least a tau proportion of completions are correct. Unlike Pass@k, Cover@tau captures reasoning under an explicit reliability threshold: models that rely on random guessing degrade rapidly as tau increases. We evaluate several RLVR models using Cover@tau-based metrics and illustrate how the relative rankings of popular algorithms change compared to Pass@1, offering a different perspective on reasoning boundaries.
comment: 10 pages, 3 figures
☆ Iterated Agent for Symbolic Regression
Symbolic regression (SR), the automated discovery of mathematical expressions from data, is a cornerstone of scientific inquiry. However, it is often hindered by the combinatorial explosion of the search space and a tendency to overfit. Popular methods, rooted in genetic programming, explore this space syntactically, often yielding overly complex, uninterpretable models. This paper introduces IdeaSearchFitter, a framework that employs Large Language Models (LLMs) as semantic operators within an evolutionary search. By generating candidate expressions guided by natural-language rationales, our method biases discovery towards models that are not only accurate but also conceptually coherent and interpretable. We demonstrate IdeaSearchFitter's efficacy across diverse challenges: it achieves competitive, noise-robust performance on the Feynman Symbolic Regression Database (FSReD), outperforming several strong baselines; discovers mechanistically aligned models with good accuracy-complexity trade-offs on real-world data; and derives compact, physically-motivated parametrizations for Parton Distribution Functions in a frontier high-energy physics application. IdeaSearchFitter is a specialized module within our broader iterated agent framework, IdeaSearch, which is publicly available at https://www.ideasearch.cn/.
comment: 45 pages, 22 figures, 8 tables
☆ To Ask or Not to Ask: Learning to Require Human Feedback
Developing decision-support systems that complement human performance in classification tasks remains an open challenge. A popular approach, Learning to Defer (LtD), allows a Machine Learning (ML) model to pass difficult cases to a human expert. However, LtD treats humans and ML models as mutually exclusive decision-makers, restricting the expert contribution to mere predictions. To address this limitation, we propose Learning to Ask (LtA), a new framework that handles both when and how to incorporate expert input in an ML model. LtA is based on a two-part architecture: a standard ML model and an enriched model trained with additional expert human feedback, with a formally optimal strategy for selecting when to query the enriched model. We provide two practical implementations of LtA: a sequential approach, which trains the models in stages, and a joint approach, which optimises them simultaneously. For the latter, we design surrogate losses with realisable-consistency guarantees. Our experiments with synthetic and real expert data demonstrate that LtA provides a more flexible and powerful foundation for effective human-AI collaboration.
☆ Robust and Efficient Collaborative Learning
Collaborative machine learning is challenged by training-time adversarial behaviors. Existing approaches to tolerate such behaviors either rely on a central server or induce high communication costs. We propose Robust Pull-based Epidemic Learning (RPEL), a novel, scalable collaborative approach to ensure robust learning despite adversaries. RPEL does not rely on any central server and, unlike traditional methods, where communication costs grow in $\mathcal{O}(n^2)$ with the number of nodes $n$, RPEL employs a pull-based epidemic-based communication strategy that scales in $\mathcal{O}(n \log n)$. By pulling model parameters from small random subsets of nodes, RPEL significantly lowers the number of required messages without compromising convergence guarantees, which hold with high probability. Empirical results demonstrate that RPEL maintains robustness in adversarial settings, competes with all-to-all communication accuracy, and scales efficiently across large networks.
☆ Dynamic Features Adaptation in Networking: Toward Flexible training and Explainable inference NeurIPS 2025
As AI becomes a native component of 6G network control, AI models must adapt to continuously changing conditions, including the introduction of new features and measurements driven by multi-vendor deployments, hardware upgrades, and evolving service requirements. To address this growing need for flexible learning in non-stationary environments, this vision paper highlights Adaptive Random Forests (ARFs) as a reliable solution for dynamic feature adaptation in communication network scenarios. We show that iterative training of ARFs can effectively lead to stable predictions, with accuracy improving over time as more features are added. In addition, we highlight the importance of explainability in AI-driven networks, proposing Drift-Aware Feature Importance (DAFI) as an efficient XAI feature importance (FI) method. DAFI uses a distributional drift detector to signal when to apply computationally intensive FI methods instead of lighter alternatives. Our tests on 3 different datasets indicate that our approach reduces runtime by up to 2 times, while producing more consistent feature importance values. Together, ARFs and DAFI provide a promising framework to build flexible AI methods adapted to 6G network use-cases.
comment: Accepted at AI4NextG Workshop, NeurIPS 2025
☆ Bridging the Physics-Data Gap with FNO-Guided Conditional Flow Matching: Designing Inductive Bias through Hierarchical Physical Constraints
Conventional time-series generation often ignores domain-specific physical constraints, limiting statistical and physical consistency. We propose a hierarchical framework that embeds the inherent hierarchy of physical laws-conservation, dynamics, boundary, and empirical relations-directly into deep generative models, introducing a new paradigm of physics-informed inductive bias. Our method combines Fourier Neural Operators (FNOs) for learning physical operators with Conditional Flow Matching (CFM) for probabilistic generation, integrated via time-dependent hierarchical constraints and FNO-guided corrections. Experiments on harmonic oscillators, human activity recognition, and lithium-ion battery degradation show 16.3% higher generation quality, 46% fewer physics violations, and 18.5% improved predictive accuracy over baselines.
comment: 8 pages, 1 figure
☆ Counterfactual Identifiability via Dynamic Optimal Transport NeurIPS 2025
We address the open question of counterfactual identification for high-dimensional multivariate outcomes from observational data. Pearl (2000) argues that counterfactuals must be identifiable (i.e., recoverable from the observed data distribution) to justify causal claims. A recent line of work on counterfactual inference shows promising results but lacks identification, undermining the causal validity of its estimates. To address this, we establish a foundation for multivariate counterfactual identification using continuous-time flows, including non-Markovian settings under standard criteria. We characterise the conditions under which flow matching yields a unique, monotone and rank-preserving counterfactual transport map with tools from dynamic optimal transport, ensuring consistent inference. Building on this, we validate the theory in controlled scenarios with counterfactual ground-truth and demonstrate improvements in axiomatic counterfactual soundness on real images.
comment: Accepted at NeurIPS 2025
☆ Mix- and MoE-DPO: A Variational Inference Approach to Direct Preference Optimization
Direct Preference Optimization (DPO) has recently emerged as a simple and effective alternative to reinforcement learning from human feedback (RLHF) for aligning large language models (LLMs) with user preferences. However, existing DPO formulations rely on a single monolithic model, which limits their expressivity in multi-task settings and their adaptability to heterogeneous or diverse preference distributions. In this work, we propose Mix- and MoE-DPO, a framework that extends DPO with both soft mixture models and mixture-of-experts (MoE) architectures, using a stochastic variational inference approach. Our method introduces a latent-variable model over expert assignments and optimizes a variational evidence lower bound (ELBO), enabling stable and efficient learning of specialized expert policies from preference data. Mix- and MoE-DPO provides three key advantages over standard DPO: (i) generalization via universal function approximation through mixtures; (ii) reward and policy specialization through expert components tailored to distinct preference modes; and (iii) contextual alignment through input-dependent soft gating that enables user-specific mixture policies. Our framework supports both shared base architectures with expert-specific policy heads and fully independent expert models, allowing flexible trade-offs between parameter efficiency and specialization. We validate our approach on a variety of model sizes and multi-preference datasets, demonstrating that Mix- and MoE-DPO offers a powerful and scalable method for preference-based LLM alignment.
☆ Opponent Shaping in LLM Agents
Large Language Models (LLMs) are increasingly being deployed as autonomous agents in real-world environments. As these deployments scale, multi-agent interactions become inevitable, making it essential to understand strategic behavior in such systems. A central open question is whether LLM agents, like reinforcement learning agents, can shape the learning dynamics and influence the behavior of others through interaction alone. In this paper, we present the first investigation of opponent shaping (OS) with LLM-based agents. Existing OS algorithms cannot be directly applied to LLMs, as they require higher-order derivatives, face scalability constraints, or depend on architectural components that are absent in transformers. To address this gap, we introduce ShapeLLM, an adaptation of model-free OS methods tailored for transformer-based agents. Using ShapeLLM, we examine whether LLM agents can influence co-players' learning dynamics across diverse game-theoretic environments. We demonstrate that LLM agents can successfully guide opponents toward exploitable equilibria in competitive games (Iterated Prisoner's Dilemma, Matching Pennies, and Chicken) and promote coordination and improve collective welfare in cooperative games (Iterated Stag Hunt and a cooperative version of the Prisoner's Dilemma). Our findings show that LLM agents can both shape and be shaped through interaction, establishing opponent shaping as a key dimension of multi-agent LLM research.
comment: 29 pages, 15 figures, 15 tables
☆ Contrastive Decoding for Synthetic Data Generation in Low-Resource Language Modeling
Large language models (LLMs) are trained on huge amounts of textual data, and concerns have been raised that the limits of such data may soon be reached. A potential solution is to train on synthetic data sampled from LLMs. In this work, we build on this idea and investigate the benefits of contrastive decoding for generating synthetic corpora. In a controlled setting, we experiment with sampling corpora using the relative difference between a good and bad model trained on the same original corpus of 100 million words. By amplifying the signal from a model that has better performance, we create a synthetic corpus and mix it with the original training data. Our findings show that training on a mixture of synthesized and real data improves performance on the language modeling objective and a range of downstream tasks. In particular, we see that training with a mix of synthetic data from contrastive decoding benefits tasks that require more reasoning skills, while synthetic data from traditional sampling helps more on tasks dependent on surface level linguistic capabilities.
comment: 13 pages, 3 figures
☆ The Hidden Bias: A Study on Explicit and Implicit Political Stereotypes in Large Language Models
Large Language Models (LLMs) are increasingly integral to information dissemination and decision-making processes. Given their growing societal influence, understanding potential biases, particularly within the political domain, is crucial to prevent undue influence on public opinion and democratic processes. This work investigates political bias and stereotype propagation across eight prominent LLMs using the two-dimensional Political Compass Test (PCT). Initially, the PCT is employed to assess the inherent political leanings of these models. Subsequently, persona prompting with the PCT is used to explore explicit stereotypes across various social dimensions. In a final step, implicit stereotypes are uncovered by evaluating models with multilingual versions of the PCT. Key findings reveal a consistent left-leaning political alignment across all investigated models. Furthermore, while the nature and extent of stereotypes vary considerably between models, implicit stereotypes elicited through language variation are more pronounced than those identified via explicit persona prompting. Interestingly, for most models, implicit and explicit stereotypes show a notable alignment, suggesting a degree of transparency or "awareness" regarding their inherent biases. This study underscores the complex interplay of political bias and stereotypes in LLMs.
☆ Enhancing Reasoning for Diffusion LLMs via Distribution Matching Policy Optimization
Diffusion large language models (dLLMs) are promising alternatives to autoregressive large language models (AR-LLMs), as they potentially allow higher inference throughput. Reinforcement learning (RL) is a crucial component for dLLMs to achieve comparable performance with AR-LLMs on important tasks, such as reasoning. However, RL algorithms that are well-suited for dLLMs' unique characteristics have yet to be developed. This paper proposes Distribution Matching Policy Optimization (DMPO), a principled and theoretically grounded RL fine-tuning method specifically designed to enhance the reasoning capabilities of dLLMs by matching the dLLM policy distribution to the optimal, reward-tilted one through cross-entropy optimization. We identify a key challenge in the implementation with a small training batch size and propose several effective solutions through a novel weight baseline subtraction technique. DMPO exhibits superior performance on multiple reasoning benchmarks without supervised fine-tuning, with an accuracy improvement of up to $42.9\%$ over previously SOTA baselines and $55.8\%$ over the base model, underscoring the effectiveness of the distribution matching framework. Our code is available at https://github.com/yuchen-zhu-zyc/DMPO.
☆ Reinforcement Learning from Probabilistic Forecasts for Safe Decision-Making via Conditional Value-at-Risk Planning
Sequential decisions in volatile, high-stakes settings require more than maximizing expected return; they require principled uncertainty management. This paper presents the Uncertainty-Aware Markov Decision Process (UAMDP), a unified framework that couples Bayesian forecasting, posterior-sampling reinforcement learning, and planning under a conditional value-at-risk (CVaR) constraint. In a closed loop, the agent updates its beliefs over latent dynamics, samples plausible futures via Thompson sampling, and optimizes policies subject to preset risk tolerances. We establish regret bounds that converge to the Bayes-optimal benchmark under standard regularity conditions. We evaluate UAMDP in two domains-high-frequency equity trading and retail inventory control-both marked by structural uncertainty and economic volatility. Relative to strong deep learning baselines, UAMDP improves long-horizon forecasting accuracy (RMSE decreases by up to 25\% and sMAPE by 32\%), and these gains translate into economic performance: the trading Sharpe ratio rises from 1.54 to 1.74 while maximum drawdown is roughly halved. These results show that integrating calibrated probabilistic modeling, exploration aligned with posterior uncertainty, and risk-aware control yields a robust, generalizable approach to safer and more profitable sequential decision-making.
☆ Investigating Counterclaims in Causality Extraction from Text
Research on causality extraction from text has so far almost entirely neglected counterclaims. Existing causality extraction datasets focus solely on "procausal" claims, i.e., statements that support a relationship. "Concausal" claims, i.e., statements that refute a relationship, are entirely ignored or even accidentally annotated as procausal. We address this shortcoming by developing a new dataset that integrates concausality. Based on an extensive literature review, we first show that concausality is an integral part of causal reasoning on incomplete knowledge. We operationalize this theory in the form of a rigorous guideline for annotation and then augment the Causal News Corpus with concausal statements, obtaining a substantial inter-annotator agreement of Cohen's $\kappa=0.74$. To demonstrate the importance of integrating concausal statements, we show that models trained without concausal relationships tend to misclassify these as procausal instead. Based on our new dataset, this mistake can be mitigated, enabling transformers to effectively distinguish pro- and concausality.
☆ Post-hoc Stochastic Concept Bottleneck Models
Concept Bottleneck Models (CBMs) are interpretable models that predict the target variable through high-level human-understandable concepts, allowing users to intervene on mispredicted concepts to adjust the final output. While recent work has shown that modeling dependencies between concepts can improve CBM performance, especially under interventions, such approaches typically require retraining the entire model, which may be infeasible when access to the original data or compute is limited. In this paper, we introduce Post-hoc Stochastic Concept Bottleneck Models (PSCBMs), a lightweight method that augments any pre-trained CBM with a multivariate normal distribution over concepts by adding only a small covariance-prediction module, without retraining the backbone model. We propose two training strategies and show on real-world data that PSCBMs consistently match or improve both concept and target accuracy over standard CBMs at test time. Furthermore, we show that due to the modeling of concept dependencies, PSCBMs perform much better than CBMs under interventions, while remaining far more efficient than retraining a similar stochastic model from scratch.
☆ Expressive Value Learning for Scalable Offline Reinforcement Learning
Reinforcement learning (RL) is a powerful paradigm for learning to make sequences of decisions. However, RL has yet to be fully leveraged in robotics, principally due to its lack of scalability. Offline RL offers a promising avenue by training agents on large, diverse datasets, avoiding the costly real-world interactions of online RL. Scaling offline RL to increasingly complex datasets requires expressive generative models such as diffusion and flow matching. However, existing methods typically depend on either backpropagation through time (BPTT), which is computationally prohibitive, or policy distillation, which introduces compounding errors and limits scalability to larger base policies. In this paper, we consider the question of how to develop a scalable offline RL approach without relying on distillation or backpropagation through time. We introduce Expressive Value Learning for Offline Reinforcement Learning (EVOR): a scalable offline RL approach that integrates both expressive policies and expressive value functions. EVOR learns an optimal, regularized Q-function via flow matching during training. At inference-time, EVOR performs inference-time policy extraction via rejection sampling against the expressive value function, enabling efficient optimization, regularization, and compute-scalable search without retraining. Empirically, we show that EVOR outperforms baselines on a diverse set of offline RL tasks, demonstrating the benefit of integrating expressive value learning into offline RL.
comment: 24 pages, 5 figures
☆ FuelCast: Benchmarking Tabular and Temporal Models for Ship Fuel Consumption ECML
In the shipping industry, fuel consumption and emissions are critical factors due to their significant impact on economic efficiency and environmental sustainability. Accurate prediction of ship fuel consumption is essential for further optimization of maritime operations. However, heterogeneous methodologies and limited high-quality datasets hinder direct comparison of modeling approaches. This paper makes three key contributions: (1) we introduce and release a new dataset (https://huggingface.co/datasets/krohnedigital/FuelCast) comprising operational and environmental data from three ships; (2) we define a standardized benchmark covering tabular regression and time-series regression (3) we investigate the application of in-context learning for ship consumption modeling using the TabPFN foundation model - a first in this domain to our knowledge. Our results demonstrate strong performance across all evaluated models, supporting the feasibility of onboard, data-driven fuel prediction. Models incorporating environmental conditions consistently outperform simple polynomial baselines relying solely on vessel speed. TabPFN slightly outperforms other techniques, highlighting the potential of foundation models with in-context learning capabilities for tabular prediction. Furthermore, including temporal context improves accuracy.
comment: This preprint has not undergone peer review or any post-submission improvements or corrections. The Version of Record of this contribution will be published in "ECML PKDD Workshop 2025 - Advanced Analytics and Learning on Temporal Data"
☆ Dual-granularity Sinkhorn Distillation for Enhanced Learning from Long-tailed Noisy Data
Real-world datasets for deep learning frequently suffer from the co-occurring challenges of class imbalance and label noise, hindering model performance. While methods exist for each issue, effectively combining them is non-trivial, as distinguishing genuine tail samples from noisy data proves difficult, often leading to conflicting optimization strategies. This paper presents a novel perspective: instead of primarily developing new complex techniques from scratch, we explore synergistically leveraging well-established, individually 'weak' auxiliary models - specialized for tackling either class imbalance or label noise but not both. This view is motivated by the insight that class imbalance (a distributional-level concern) and label noise (a sample-level concern) operate at different granularities, suggesting that robustness mechanisms for each can in principle offer complementary strengths without conflict. We propose Dual-granularity Sinkhorn Distillation (D-SINK), a novel framework that enhances dual robustness by distilling and integrating complementary insights from such 'weak', single-purpose auxiliary models. Specifically, D-SINK uses an optimal transport-optimized surrogate label allocation to align the target model's sample-level predictions with a noise-robust auxiliary and its class distributions with an imbalance-robust one. Extensive experiments on benchmark datasets demonstrate that D-SINK significantly improves robustness and achieves strong empirical performance in learning from long-tailed noisy data.
comment: 25 pages, 2 figures
☆ Long-tailed Recognition with Model Rebalancing
Long-tailed recognition is ubiquitous and challenging in deep learning and even in the downstream finetuning of foundation models, since the skew class distribution generally prevents the model generalization to the tail classes. Despite the promise of previous methods from the perspectives of data augmentation, loss rebalancing and decoupled training etc., consistent improvement in the broad scenarios like multi-label long-tailed recognition is difficult. In this study, we dive into the essential model capacity impact under long-tailed context, and propose a novel framework, Model Rebalancing (MORE), which mitigates imbalance by directly rebalancing the model's parameter space. Specifically, MORE introduces a low-rank parameter component to mediate the parameter space allocation guided by a tailored loss and sinusoidal reweighting schedule, but without increasing the overall model complexity or inference costs. Extensive experiments on diverse long-tailed benchmarks, spanning multi-class and multi-label tasks, demonstrate that MORE significantly improves generalization, particularly for tail classes, and effectively complements existing imbalance mitigation methods. These results highlight MORE's potential as a robust plug-and-play module in long-tailed settings.
☆ Leveraging Whisper Embeddings for Audio-based Lyrics Matching
Audio-based lyrics matching can be an appealing alternative to other content-based retrieval approaches, but existing methods often suffer from limited reproducibility and inconsistent baselines. In this work, we introduce WEALY, a fully reproducible pipeline that leverages Whisper decoder embeddings for lyrics matching tasks. WEALY establishes robust and transparent baselines, while also exploring multimodal extensions that integrate textual and acoustic features. Through extensive experiments on standard datasets, we demonstrate that WEALY achieves a performance comparable to state-of-the-art methods that lack reproducibility. In addition, we provide ablation studies and analyses on language robustness, loss functions, and embedding strategies. This work contributes a reliable benchmark for future research, and underscores the potential of speech technologies for music information retrieval tasks.
☆ Bidirectional Representations Augmented Autoregressive Biological Sequence Generation:Application in De Novo Peptide Sequencing NeurIPS 2025
Autoregressive (AR) models, common in sequence generation, are limited in many biological tasks such as de novo peptide sequencing and protein modeling by their unidirectional nature, failing to capture crucial global bidirectional token dependencies. Non-Autoregressive (NAR) models offer holistic, bidirectional representations but face challenges with generative coherence and scalability. To transcend this, we propose a hybrid framework enhancing AR generation by dynamically integrating rich contextual information from non-autoregressive mechanisms. Our approach couples a shared input encoder with two decoders: a non-autoregressive one learning latent bidirectional biological features, and an AR decoder synthesizing the biological sequence by leveraging these bidirectional features. A novel cross-decoder attention module enables the AR decoder to iteratively query and integrate these bidirectional features, enriching its predictions. This synergy is cultivated via a tailored training strategy with importance annealing for balanced objectives and cross-decoder gradient blocking for stable, focused learning. Evaluations on a demanding nine-species benchmark of de novo peptide sequencing show that our model substantially surpasses AR and NAR baselines. It uniquely harmonizes AR stability with NAR contextual awareness, delivering robust, superior performance on diverse downstream data. This research advances biological sequence modeling techniques and contributes a novel architectural paradigm for augmenting AR models with enhanced bidirectional understanding for complex sequence generation. Code is available at https://github.com/BEAM-Labs/denovo.
comment: Accepted by NeurIPS 2025
☆ Beyond Sub-6 GHz: Leveraging mmWave Wi-Fi for Gait-Based Person Identification
Person identification plays a vital role in enabling intelligent, personalized, and secure human-computer interaction. Recent research has demonstrated the feasibility of leveraging Wi-Fi signals for passive person identification using a person's unique gait pattern. Although most existing work focuses on sub-6 GHz frequencies, the emergence of mmWave offers new opportunities through its finer spatial resolution, though its comparative advantages for person identification remain unexplored. This work presents the first comparative study between sub-6 GHz and mmWave Wi-Fi signals for person identification with commercial off-the-shelf (COTS) Wi-Fi, using a novel dataset of synchronized measurements from the two frequency bands in an indoor environment. To ensure a fair comparison, we apply identical training pipelines and model configurations across both frequency bands. Leveraging end-to-end deep learning, we show that even at low sampling rates (10 Hz), mmWave Wi-Fi signals can achieve high identification accuracy (91.2% on 20 individuals) when combined with effective background subtraction.
☆ Quantum Agents for Algorithmic Discovery
We introduce quantum agents trained by episodic, reward-based reinforcement learning to autonomously rediscover several seminal quantum algorithms and protocols. In particular, our agents learn: efficient logarithmic-depth quantum circuits for the Quantum Fourier Transform; Grover's search algorithm; optimal cheating strategies for strong coin flipping; and optimal winning strategies for the CHSH and other nonlocal games. The agents achieve these results directly through interaction, without prior access to known optimal solutions. This demonstrates the potential of quantum intelligence as a tool for algorithmic discovery, opening the way for the automated design of novel quantum algorithms and protocols.
☆ Unsupervised Multi-Source Federated Domain Adaptation under Domain Diversity through Group-Wise Discrepancy Minimization
Unsupervised multi-source domain adaptation (UMDA) aims to learn models that generalize to an unlabeled target domain by leveraging labeled data from multiple, diverse source domains. While distributed UMDA methods address privacy constraints by avoiding raw data sharing, existing approaches typically assume a small number of sources and fail to scale effectively. Increasing the number of heterogeneous domains often makes existing methods impractical, leading to high computational overhead or unstable performance. We propose GALA, a scalable and robust federated UMDA framework that introduces two key components: (1) a novel inter-group discrepancy minimization objective that efficiently approximates full pairwise domain alignment without quadratic computation; and (2) a temperature-controlled, centroid-based weighting strategy that dynamically prioritizes source domains based on alignment with the target. Together, these components enable stable and parallelizable training across large numbers of heterogeneous sources. To evaluate performance in high-diversity scenarios, we introduce Digit-18, a new benchmark comprising 18 digit datasets with varied synthetic and real-world domain shifts. Extensive experiments show that GALA consistently achieves competitive or state-of-the-art results on standard benchmarks and significantly outperforms prior methods in diverse multi-source settings where others fail to converge.
☆ AI Knowledge Assist: An Automated Approach for the Creation of Knowledge Bases for Conversational AI Agents EMNLP 2025
The utilization of conversational AI systems by leveraging Retrieval Augmented Generation (RAG) techniques to solve customer problems has been on the rise with the rapid progress of Large Language Models (LLMs). However, the absence of a company-specific dedicated knowledge base is a major barrier to the integration of conversational AI systems in contact centers. To this end, we introduce AI Knowledge Assist, a system that extracts knowledge in the form of question-answer (QA) pairs from historical customer-agent conversations to automatically build a knowledge base. Fine-tuning a lightweight LLM on internal data demonstrates state-of-the-art performance, outperforming larger closed-source LLMs. More specifically, empirical evaluation on 20 companies demonstrates that the proposed AI Knowledge Assist system that leverages the LLaMA-3.1-8B model eliminates the cold-start gap in contact centers by achieving above 90% accuracy in answering information-seeking questions. This enables immediate deployment of RAG-powered chatbots.
comment: Accepted to the EMNLP 2025 Industry Track
☆ Think Just Enough: Sequence-Level Entropy as a Confidence Signal for LLM Reasoning
We introduce a simple, yet novel entropy-based framework to drive token efficiency in large language models during reasoning tasks. Our approach uses Shannon entropy from token-level logprobs as a confidence signal to enable early stopping, achieving 25-50% computational savings while maintaining task accuracy. Crucially, we demonstrate that entropy-based confidence calibration represents an emergent property of advanced post-training optimization present in modern reasoning models but notably absent in standard instruction-tuned and pre-trained models (Llama 3.3 70B). We show that the entropy threshold to stop reasoning varies from model to model but can be calculated easily in one shot using only a few examples from existing reasoning datasets. Our results indicate that advanced reasoning models often know that they've gotten a correct answer early on, and that this emergent confidence awareness can be exploited to save tokens and reduce latency. The framework demonstrates consistent performance across reasoning-optimized model families with 25-50% computational cost reduction while preserving accuracy, revealing that confidence mechanisms represent a distinguishing characteristic of modern post-trained reasoning systems versus their predecessors.
☆ Arbitrary Entropy Policy Optimization: Entropy Is Controllable in Reinforcement Finetuning
Reinforcement finetuning (RFT) is essential for enhancing the reasoning capabilities of large language models (LLM), yet the widely adopted Group Relative Policy Optimization (GRPO) suffers from entropy collapse, where entropy monotonically decreases, exploration vanishes, and policies converge prematurely. Existing entropy-regularized methods only partially alleviate this issue while introducing bias and instability, leaving entropy control unresolved and the connection between entropy, exploration, and performance unclear. We propose Arbitrary Entropy Policy Optimization (AEPO), which eliminates entropy collapse by replacing entropy bonuses with REINFORCE policy gradient on temperature-adjusted distributions and stabilizing entropy through temperature regulation. AEPO integrates three key designs: policy gradient as regularization, distribution as regularization, and REINFORCE as regularization, enabling precise entropy control without distorting optimization. Experiments demonstrate three major contributions: AEPO (1) stabilizes entropy at arbitrary target levels, effectively removing collapse in GRPO; (2) reveals a non-monotonic relation where performance first improves then declines with increasing entropy, clarifying the link between entropy, exploration, and reasoning; and (3) generalizes beyond entropy, providing a broader RFT paradigm where superior target distributions can serve as REINFORCE regularizers.
☆ Approximate Domain Unlearning for Vision-Language Models NeurIPS 2025
Pre-trained Vision-Language Models (VLMs) exhibit strong generalization capabilities, enabling them to recognize a wide range of objects across diverse domains without additional training. However, they often retain irrelevant information beyond the requirements of specific downstream tasks, raising concerns about computational efficiency and potential information leakage. This has motivated growing interest in approximate unlearning, which aims to selectively remove unnecessary knowledge while preserving overall model performance. Existing approaches to approximate unlearning have primarily focused on class unlearning, where a VLM is retrained to fail to recognize specified object classes while maintaining accuracy for others. However, merely forgetting object classes is often insufficient in practical applications. For instance, an autonomous driving system should accurately recognize real cars while avoiding misrecognition of illustrated cars depicted in roadside advertisements as real cars, which could be hazardous. In this paper, we introduce Approximate Domain Unlearning (ADU), a novel problem setting that requires reducing recognition accuracy for images from specified domains (e.g., illustration) while preserving accuracy for other domains (e.g., real). ADU presents new technical challenges: due to the strong domain generalization capability of pre-trained VLMs, domain distributions are highly entangled in the feature space, making naive approaches based on penalizing target domains ineffective. To tackle this limitation, we propose a novel approach that explicitly disentangles domain distributions and adaptively captures instance-specific domain information. Extensive experiments show that our approach outperforms baselines built upon VLM tuning techniques, paving the way for practical and fine-grained unlearning in VLMs. Code: https://kodaikawamura.github.io/Domain_Unlearning/.
comment: NeurIPS 2025 (Spotlight)
☆ High-dimensional Analysis of Synthetic Data Selection
Despite the progress in the development of generative models, their usefulness in creating synthetic data that improve prediction performance of classifiers has been put into question. Besides heuristic principles such as "synthetic data should be close to the real data distribution", it is actually not clear which specific properties affect the generalization error. Our paper addresses this question through the lens of high-dimensional regression. Theoretically, we show that, for linear models, the covariance shift between the target distribution and the distribution of the synthetic data affects the generalization error but, surprisingly, the mean shift does not. Furthermore we prove that, in some settings, matching the covariance of the target distribution is optimal. Remarkably, the theoretical insights from linear models carry over to deep neural networks and generative models. We empirically demonstrate that the covariance matching procedure (matching the covariance of the synthetic data with that of the data coming from the target distribution) performs well against several recent approaches for synthetic data selection, across training paradigms, architectures, datasets and generative models used for augmentation.
☆ Random Window Augmentations for Deep Learning Robustness in CT and Liver Tumor Segmentation
Contrast-enhanced Computed Tomography (CT) is important for diagnosis and treatment planning for various medical conditions. Deep learning (DL) based segmentation models may enable automated medical image analysis for detecting and delineating tumors in CT images, thereby reducing clinicians' workload. Achieving generalization capabilities in limited data domains, such as radiology, requires modern DL models to be trained with image augmentation. However, naively applying augmentation methods developed for natural images to CT scans often disregards the nature of the CT modality, where the intensities measure Hounsfield Units (HU) and have important physical meaning. This paper challenges the use of such intensity augmentations for CT imaging and shows that they may lead to artifacts and poor generalization. To mitigate this, we propose a CT-specific augmentation technique, called Random windowing, that exploits the available HU distribution of intensities in CT images. Random windowing encourages robustness to contrast-enhancement and significantly increases model performance on challenging images with poor contrast or timing. We perform ablations and analysis of our method on multiple datasets, and compare to, and outperform, state-of-the-art alternatives, while focusing on the challenge of liver tumor segmentation.
comment: 10 pages, 9 figures. This work has been submitted to the IEEE for possible publication
☆ Bayesian Decision Making around Experts
Complex learning agents are increasingly deployed alongside existing experts, such as human operators or previously trained agents. However, it remains unclear how should learners optimally incorporate certain forms of expert data, which may differ in structure from the learner's own action-outcome experiences. We study this problem in the context of Bayesian multi-armed bandits, considering: (i) offline settings, where the learner receives a dataset of outcomes from the expert's optimal policy before interaction, and (ii) simultaneous settings, where the learner must choose at each step whether to update its beliefs based on its own experience, or based on the outcome simultaneously achieved by an expert. We formalize how expert data influences the learner's posterior, and prove that pretraining on expert outcomes tightens information-theoretic regret bounds by the mutual information between the expert data and the optimal action. For the simultaneous setting, we propose an information-directed rule where the learner processes the data source that maximizes their one-step information gain about the optimal action. Finally, we propose strategies for how the learner can infer when to trust the expert and when not to, safeguarding the learner for the cases where the expert is ineffective or compromised. By quantifying the value of expert data, our framework provides practical, information-theoretic algorithms for agents to intelligently decide when to learn from others.
☆ Lossless Vocabulary Reduction for Auto-Regressive Language Models
Tokenization -- the process of decomposing a given text into a sequence of subwords called tokens -- is one of the key components in the development of language models. Particularly, auto-regressive language models generate texts token by token, i.e., by predicting the next-token distribution given the previous ones, and thus tokenization directly affects their efficiency in text generation. Since each language model has their own vocabulary as a set of possible tokens, they struggle to cooperate with each other at the level of next-token distributions such as model ensemble. In this paper, we establish a theoretical framework of lossless vocabulary reduction, which efficiently converts a given auto-regressive language model into the one with an arbitrarily small vocabulary without any loss in accuracy. As an application, we demonstrate that language models with different tokenization can cooperate with each other efficiently through their maximal common vocabulary.
☆ Beyond Real Data: Synthetic Data through the Lens of Regularization
Synthetic data can improve generalization when real data is scarce, but excessive reliance may introduce distributional mismatches that degrade performance. In this paper, we present a learning-theoretic framework to quantify the trade-off between synthetic and real data. Our approach leverages algorithmic stability to derive generalization error bounds, characterizing the optimal synthetic-to-real data ratio that minimizes expected test error as a function of the Wasserstein distance between the real and synthetic distributions. We motivate our framework in the setting of kernel ridge regression with mixed data, offering a detailed analysis that may be of independent interest. Our theory predicts the existence of an optimal ratio, leading to a U-shaped behavior of test error with respect to the proportion of synthetic data. Empirically, we validate this prediction on CIFAR-10 and a clinical brain MRI dataset. Our theory extends to the important scenario of domain adaptation, showing that carefully blending synthetic target data with limited source data can mitigate domain shift and enhance generalization. We conclude with practical guidance for applying our results to both in-domain and out-of-domain scenarios.
☆ Computations and ML for surjective rational maps
The present note studies \emph{surjective rational endomorphisms} $f: \mathbb{P}^2 \dashrightarrow \mathbb{P}^2$ with \emph{cubic} terms and the indeterminacy locus $I_f \ne \emptyset$. We develop an experimental approach, based on some Python programming and Machine Learning, towards the classification of such maps; a couple of new explicit $f$ is constructed in this way. We also prove (via pure projective geometry) that a general non-regular cubic endomorphism $f$ of $\mathbb{P}^2$ is surjective if and only if the set $I_f$ has cardinality at least $3$.
comment: 15 pages, 2 figures, a couple of Python codes
☆ A Novel Ensemble Learning Approach for Enhanced IoT Attack Detection: Redefining Security Paradigms in Connected Systems
The rapid expansion of Internet of Things (IoT) devices has transformed industries and daily life by enabling widespread connectivity and data exchange. However, this increased interconnection has introduced serious security vulnerabilities, making IoT systems more exposed to sophisticated cyber attacks. This study presents a novel ensemble learning architecture designed to improve IoT attack detection. The proposed approach applies advanced machine learning techniques, specifically the Extra Trees Classifier, along with thorough preprocessing and hyperparameter optimization. It is evaluated on several benchmark datasets including CICIoT2023, IoTID20, BotNeTIoT L01, ToN IoT, N BaIoT, and BoT IoT. The results show excellent performance, achieving high recall, accuracy, and precision with very low error rates. These outcomes demonstrate the model efficiency and superiority compared to existing approaches, providing an effective and scalable method for securing IoT environments. This research establishes a solid foundation for future progress in protecting connected devices from evolving cyber threats.
comment: 14 pages, 5 fiugres, 7 tables
☆ Detecting and Mitigating Insertion Hallucination in Video-to-Audio Generation
Video-to-Audio generation has made remarkable strides in automatically synthesizing sound for video. However, existing evaluation metrics, which focus on semantic and temporal alignment, overlook a critical failure mode: models often generate acoustic events, particularly speech and music, that have no corresponding visual source. We term this phenomenon Insertion Hallucination and identify it as a systemic risk driven by dataset biases, such as the prevalence of off-screen sounds, that remains completely undetected by current metrics. To address this challenge, we first develop a systematic evaluation framework that employs a majority-voting ensemble of multiple audio event detectors. We also introduce two novel metrics to quantify the prevalence and severity of this issue: IH@vid (the fraction of videos with hallucinations) and IH@dur (the fraction of hallucinated duration). Building on this, we propose Posterior Feature Correction, a novel training-free inference-time method that mitigates IH. PFC operates in a two-pass process: it first generates an initial audio output to detect hallucinated segments, and then regenerates the audio after masking the corresponding video features at those timestamps. Experiments on several mainstream V2A benchmarks first reveal that state-of-the-art models suffer from severe IH. In contrast, our PFC method reduces both the prevalence and duration of hallucinations by over 50\% on average, without degrading, and in some cases even improving, conventional metrics for audio quality and temporal synchronization. Our work is the first to formally define, systematically measure, and effectively mitigate Insertion Hallucination, paving the way for more reliable and faithful V2A models.
♻ ☆ Graph-SCP: Accelerating Set Cover Problems with Graph Neural Networks
Machine learning (ML) approaches are increasingly being used to accelerate combinatorial optimization (CO) problems. We investigate the Set Cover Problem (SCP) and propose Graph-SCP, a graph neural network method that augments existing optimization solvers by learning to identify a smaller sub-problem that contains the solution space. Graph-SCP uses both supervised learning from prior solved instances and unsupervised learning to minimize the SCP objective. We evaluate the performance of Graph-SCP on synthetically weighted and unweighted SCP instances with diverse problem characteristics and complexities, and on instances from the OR Library, a canonical benchmark for SCP. We show that Graph-SCP reduces the problem size by 60-80% and achieves runtime speedups of up to 10x on average when compared to Gurobi (a state-of-the-art commercial solver), while maintaining solution quality. This is in contrast to fast greedy solutions that significantly compromise solution quality to achieve guaranteed polynomial runtime. We showcase Graph-SCP's ability to generalize to larger problem sizes, training on SCP instances with up to 3,000 subsets and testing on SCP instances with up to 10,000 subsets.
♻ ☆ Understanding In-context Learning of Addition via Activation Subspaces
To perform few-shot learning, language models extract signals from a few input-label pairs, aggregate these into a learned prediction rule, and apply this rule to new inputs. How is this implemented in the forward pass of modern transformer models? To explore this question, we study a structured family of few-shot learning tasks for which the true prediction rule is to add an integer $k$ to the input. We introduce a novel optimization method that localizes the model's few-shot ability to only a few attention heads. We then perform an in-depth analysis of individual heads, via dimensionality reduction and decomposition. As an example, on Llama-3-8B-instruct, we reduce its mechanism on our tasks to just three attention heads with six-dimensional subspaces, where four dimensions track the unit digit with trigonometric functions at periods $2$, $5$, and $10$, and two dimensions track magnitude with low-frequency components. To deepen our understanding of the mechanism, we also derive a mathematical identity relating ``aggregation'' and ``extraction'' subspaces for attention heads, allowing us to track the flow of information from individual examples to a final aggregated concept. Using this, we identify a self-correction mechanism where mistakes learned from earlier demonstrations are suppressed by later demonstrations. Our results demonstrate how tracking low-dimensional subspaces of localized heads across a forward pass can provide insight into fine-grained computational structures in language models.
♻ ☆ From Moments to Models: Graphon Mixture-Aware Mixup and Contrastive Learning
Real-world graph datasets often consist of mixtures of populations, where graphs are generated from multiple distinct underlying distributions. However, modern representation learning approaches, such as graph contrastive learning (GCL) and augmentation methods like Mixup, typically overlook this mixture structure. In this work, we propose a unified framework that explicitly models data as a mixture of underlying probabilistic graph generative models represented by graphons. To characterize these graphons, we leverage graph moments (motif densities) to cluster graphs arising from the same model. This enables us to disentangle the mixture components and identify their distinct generative mechanisms. This model-aware partitioning benefits two key graph learning tasks: 1) It enables a graphon-mixture-aware mixup (GMAM), a data augmentation technique that interpolates in a semantically valid space guided by the estimated graphons, instead of assuming a single graphon per class. 2) For GCL, it enables model-adaptive and principled augmentations. Additionally, by introducing a new model-aware objective, our proposed approach (termed MGCL) improves negative sampling by restricting negatives to graphs from other models. We establish a key theoretical guarantee: a novel, tighter bound showing that graphs sampled from graphons with small cut distance will have similar motif densities with high probability. Extensive experiments on benchmark datasets demonstrate strong empirical performance. In unsupervised learning, MGCL achieves state-of-the-art results, obtaining the top average rank across eight datasets. In supervised learning, GMAM consistently outperforms existing strategies, achieving new state-of-the-art accuracy in 6 out of 7 datasets.
♻ ☆ Feature Identification via the Empirical NTK
We provide evidence that eigenanalysis of the empirical neural tangent kernel (eNTK) can surface the features used by trained neural networks. Across two standard toy models for mechanistic interpretability, Toy Models of Superposition (TMS) and a 1-layer MLP trained on modular addition, we find that the eNTK exhibits sharp spectral cliffs whose top eigenspaces align with ground-truth features. In TMS, the eNTK recovers the ground-truth features in both the sparse (high superposition) and dense regimes. In modular arithmetic, the eNTK can be used to recover Fourier feature families. Moreover, we provide evidence that a layerwise eNTK localizes features to specific layers and that the evolution of the eNTK spectrum can be used to diagnose the grokking phase transition. These results suggest that eNTK analysis may provide a practical handle for feature discovery and for detecting phase changes in small models.
comment: 14 pages, 5 figures. v2: references and expanded discussion in Appendix B added
♻ ☆ Zebra-CoT: A Dataset for Interleaved Vision Language Reasoning
Humans often use visual aids, for example diagrams or sketches, when solving complex problems. Training multimodal models to do the same, known as Visual Chain of Thought (Visual CoT), is challenging due to: (1) poor off-the-shelf visual CoT performance, which hinders reinforcement learning, and (2) the lack of high-quality visual CoT training data. We introduce $\textbf{Zebra-CoT}$, a diverse large-scale dataset with 182,384 samples, containing logically coherent interleaved text-image reasoning traces. We focus on four categories of tasks where sketching or visual reasoning is especially natural, spanning scientific questions such as geometry, physics, and algorithms; 2D visual reasoning tasks like visual search and jigsaw puzzles; 3D reasoning tasks including 3D multi-hop inference, embodied and robot planning; visual logic problems and strategic games like chess. Fine-tuning the Anole-7B model on the Zebra-CoT training corpus results in an improvement of +12% in our test-set accuracy and yields up to +13% performance gain on standard VLM benchmark evaluations. Fine-tuning Bagel-7B yields a model that generates high-quality interleaved visual reasoning chains, underscoring Zebra-CoT's effectiveness for developing multimodal reasoning abilities. We open-source our dataset and models to support development and evaluation of visual CoT.
comment: dataset link: https://huggingface.co/datasets/multimodal-reasoning-lab/Zebra-CoT
♻ ☆ Training a Foundation Model for Materials on a Budget NeurIPS 2025
Foundation models for materials modeling are advancing quickly, but their training remains expensive, often placing state-of-the-art methods out of reach for many research groups. We introduce Nequix, a compact E(3)-equivariant potential that pairs a simplified NequIP design with modern training practices, including equivariant root-mean-square layer normalization and the Muon optimizer, to retain accuracy while substantially reducing compute requirements. Nequix has 700K parameters and was trained in 100 A100 GPU-hours. On the Matbench-Discovery and MDR Phonon benchmarks, Nequix ranks third overall while requiring a 20 times lower training cost than most other methods, and it delivers two orders of magnitude faster inference speed than the current top-ranked model. We release model weights and fully reproducible codebase at https://github.com/atomicarchitects/nequix.
comment: To appear at the NeurIPS 2025 Workshop on AI for Accelerated Materials Discovery
♻ ☆ Efficient Graph Condensation via Gaussian Process
Graph condensation reduces the size of large graphs while preserving performance, addressing the scalability challenges of Graph Neural Networks caused by computational inefficiencies on large datasets. Existing methods often rely on bi-level optimization, requiring extensive GNN training and limiting their scalability. To address these issues, this paper proposes Graph Condensation via Gaussian Process (GCGP), a novel and computationally efficient approach to graph condensation. GCGP utilizes a Gaussian Process (GP), with the condensed graph serving as observations, to estimate the posterior distribution of predictions. This approach eliminates the need for the iterative and resource-intensive training typically required by GNNs. To enhance the capability of the GCGP in capturing dependencies between function values, we derive a specialized covariance function that incorporates structural information. This covariance function broadens the receptive field of input nodes by local neighborhood aggregation, thereby facilitating the representation of intricate dependencies within the nodes. To address the challenge of optimizing binary structural information in condensed graphs, Concrete random variables are utilized to approximate the binary adjacency matrix in a continuous counterpart. This relaxation process allows the adjacency matrix to be represented in a differentiable form, enabling the application of gradient-based optimization techniques to discrete graph structures. Experimental results show that the proposed GCGP method efficiently condenses large-scale graph data while preserving predictive performance, addressing the scalability and efficiency challenges. The implementation of our method is publicly available at https://github.com/WANGLin0126/GCGP.
♻ ☆ Multi-Turn Human-LLM Interaction Through the Lens of a Two-Way Intelligibility Protocol NeurIPS 2025
Our interest is in the design of software systems involving a human-expert interacting -- using natural language -- with a large language model (LLM) on data analysis tasks. For complex problems, it is possible that LLMs can harness human expertise and creativity to find solutions that were otherwise elusive. On one level, this interaction takes place through multiple turns of prompts from the human and responses from the LLM. Here we investigate a more structured approach based on an abstract protocol described in [3] for interaction between agents. The protocol is motivated by a notion of "two-way intelligibility" and is modelled by a pair of communicating finite-state machines. We provide an implementation of the protocol, and provide empirical evidence of using the implementation to mediate interactions between an LLM and a human-agent in two areas of scientific interest (radiology and drug design). We conduct controlled experiments with a human proxy (a database), and uncontrolled experiments with human subjects. The results provide evidence in support of the protocol's capability of capturing one- and two-way intelligibility in human-LLM interaction; and for the utility of two-way intelligibility in the design of human-machine systems. Our code is available at https://github.com/karannb/interact.
comment: Multi-Turn Interactions in Large Language Models (MTI-LLM) Workshop at NeurIPS 2025
♻ ☆ Evaluating Evaluation Metrics -- The Mirage of Hallucination Detection EMNLP 2025
Hallucinations pose a significant obstacle to the reliability and widespread adoption of language models, yet their accurate measurement remains a persistent challenge. While many task- and domain-specific metrics have been proposed to assess faithfulness and factuality concerns, the robustness and generalization of these metrics are still untested. In this paper, we conduct a large-scale empirical evaluation of 6 diverse sets of hallucination detection metrics across 4 datasets, 37 language models from 5 families, and 5 decoding methods. Our extensive investigation reveals concerning gaps in current hallucination evaluation: metrics often fail to align with human judgments, take an overtly myopic view of the problem, and show inconsistent gains with parameter scaling. Encouragingly, LLM-based evaluation, particularly with GPT-4, yields the best overall results, and mode-seeking decoding methods seem to reduce hallucinations, especially in knowledge-grounded settings. These findings underscore the need for more robust metrics to understand and quantify hallucinations, and better strategies to mitigate them.
comment: Accepted at EMNLP 2025 Findings (Short)
♻ ☆ Spiffy: Multiplying Diffusion LLM Acceleration via Lossless Speculative Decoding
Diffusion LLMs (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs (AR-LLMs) with the potential to operate at significantly higher token generation rates. However, currently available open-source dLLMs often generate at much lower rates, typically decoding only a single token at every denoising timestep in order to maximize output quality. We present Spiffy, a speculative decoding algorithm that accelerates dLLM inference by $\mathbf{2.8{-}3.1\times}$ while provably preserving the model's output distribution. This work addresses the unique challenges involved in applying ideas from speculative decoding of AR-LLMs to the dLLM setting. Spiffy proposes draft states by leveraging the dLLM's distribution itself in an auto-speculative manner. This approach is efficient and effective, and eliminates the overheads of training and running an independent draft model. To structure the candidate draft states, we propose a novel directed draft graph which is uniquely designed to take advantage of the bidirectional, block-wise nature of dLLM generation and can be verified in parallel by the dLLM. To further optimize the structure of these draft graphs, we introduce an efficient, offline calibration algorithm that procedurally determines high-quality graph configurations. These optimized draft graphs, enabling increased acceptance rates, lead to a significant boost in the overall speedup achieved by the system. Crucially, Spiffy is also complementary to other recent innovations in improving dLLM generation speeds such as KV-caching and multi-token unmasking. We demonstrate that when combined with such parallel decoding algorithms, Spiffy is able to effectively multiply the benefits of these methods leading to total speedups of up to $\mathbf{7.9\times}$.
comment: Original version uploaded on Sep 22, 2025. (v2): Extended Table 2 with additional analysis and referenced it in Sec 5.2
♻ ☆ Language Model Embeddings Can Be Sufficient for Bayesian Optimization
Bayesian Optimization is ubiquitous in experimental design and black-box optimization for improving search efficiency. However, most existing approaches rely on regression models which are limited to fixed search spaces and structured, tabular input features. This paper explores the use of LLM embeddings over string inputs for in-context regression in Bayesian Optimization. Our results show that representing inputs as strings enables general-purpose regression across diverse domains, including synthetic, combinatorial, and hyperparameter optimization. Furthermore, our approach achieves optimization performance comparable to state-of-the-art Gaussian Process-based methods such as Google Vizier, and demonstrates potential for broader and more flexible applications.
comment: Code can be found in https://github.com/google-research/optformer/tree/main/optformer/embed_then_regress
♻ ☆ More Than One Teacher: Adaptive Multi-Guidance Policy Optimization for Diverse Exploration
Reinforcement Learning with Verifiable Rewards (RLVR) is a promising paradigm for enhancing the reasoning ability in Large Language Models (LLMs). However, prevailing methods primarily rely on self-exploration or a single off-policy teacher to elicit long chain-of-thought (LongCoT) reasoning, which may introduce intrinsic model biases and restrict exploration, ultimately limiting reasoning diversity and performance. Drawing inspiration from multi-teacher strategies in knowledge distillation, we introduce Adaptive Multi-Guidance Policy Optimization (AMPO), a novel framework that adaptively leverages guidance from multiple proficient teacher models, but only when the on-policy model fails to generate correct solutions. This "guidance-on-demand" approach expands exploration while preserving the value of self-discovery. Moreover, AMPO incorporates a comprehension-based selection mechanism, prompting the student to learn from the reasoning paths that it is most likely to comprehend, thus balancing broad exploration with effective exploitation. Extensive experiments show AMPO substantially outperforms a strong baseline (GRPO), with a 4.3% improvement on mathematical reasoning tasks and 12.2% on out-of-distribution tasks, while significantly boosting Pass@k performance and enabling more diverse exploration. Notably, using four peer-sized teachers, our method achieves comparable results to approaches that leverage a single, more powerful teacher (e.g., DeepSeek-R1) with more data. These results demonstrate a more efficient and scalable path to superior reasoning and generalizability. Our code is available at https://github.com/SII-Enigma/AMPO.
comment: 20 pages, 5 figures
♻ ☆ Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check EMNLP
Downstream scaling laws aim to predict task performance at larger scales from the model's performance at smaller scales. Whether such prediction should be possible is unclear: some works discover clear linear scaling trends after simple transformations of the performance metric, whereas others point out fundamental challenges to downstream scaling laws, such as emergence and inverse scaling. In this work, we conduct a meta-analysis of existing data on downstream scaling laws, and we find that predictable scaling only occurs in a minority of cases: 39% of the time. Moreover, seemingly benign changes to the experimental setting can completely change the scaling behavior. Our analysis underscores the need to understand the conditions under which scaling laws succeed. To accurately model the relationship between pretraining loss and task performance, we must embrace the cases in which scaling behavior deviates from linear trends.
comment: EMNLP Findings 2025 Camera Ready
♻ ☆ LLINBO: Trustworthy LLM-in-the-Loop Bayesian Optimization
Bayesian optimization (BO) is a sequential decision-making tool widely used for optimizing expensive black-box functions. Recently, Large Language Models (LLMs) have shown remarkable adaptability in low-data regimes, making them promising tools for black-box optimization by leveraging contextual knowledge to propose high-quality query points. However, relying solely on LLMs as optimization agents introduces risks due to their lack of explicit surrogate modeling and calibrated uncertainty, as well as their inherently opaque internal mechanisms. This structural opacity makes it difficult to characterize or control the exploration-exploitation trade-off, ultimately undermining theoretical tractability and reliability. To address this, we propose LLINBO: LLM-in-the-Loop BO, a hybrid framework for BO that combines LLMs with statistical surrogate experts (e.g., Gaussian Processes (GP)). The core philosophy is to leverage contextual reasoning strengths of LLMs for early exploration, while relying on principled statistical models to guide efficient exploitation. Specifically, we introduce three mechanisms that enable this collaboration and establish their theoretical guarantees. We end the paper with a real-life proof-of-concept in the context of 3D printing. The code to reproduce the results can be found at https://github.com/UMDataScienceLab/LLM-in-the-Loop-BO.
♻ ☆ A Survey of Reinforcement Learning for Large Reasoning Models
In this paper, we survey recent advances in Reinforcement Learning (RL) for reasoning with Large Language Models (LLMs). RL has achieved remarkable success in advancing the frontier of LLM capabilities, particularly in addressing complex logical tasks such as mathematics and coding. As a result, RL has emerged as a foundational methodology for transforming LLMs into LRMs. With the rapid progress of the field, further scaling of RL for LRMs now faces foundational challenges not only in computational resources but also in algorithm design, training data, and infrastructure. To this end, it is timely to revisit the development of this domain, reassess its trajectory, and explore strategies to enhance the scalability of RL toward Artificial SuperIntelligence (ASI). In particular, we examine research applying RL to LLMs and LRMs for reasoning abilities, especially since the release of DeepSeek-R1, including foundational components, core problems, training resources, and downstream applications, to identify future opportunities and directions for this rapidly evolving area. We hope this review will promote future research on RL for broader reasoning models. Github: https://github.com/TsinghuaC3I/Awesome-RL-for-LRMs
comment: Fixed typos; added missing and recent citations (117 -> 120 pages)
♻ ☆ Hybrid Reinforcement: When Reward Is Sparse, It's Better to Be Dense
Post-training for reasoning of large language models (LLMs) increasingly relies on verifiable rewards: deterministic checkers that provide 0-1 correctness signals. While reliable, such binary feedback is brittle--many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning.
comment: 21 pages
♻ ☆ Data-Error Scaling Laws in Machine Learning on Combinatorial Mutation-prone Sets: Proteins and Small Molecules
We investigate trends in the data-error scaling laws of machine learning (ML) models trained on discrete combinatorial spaces that are prone-to-mutation, such as proteins or organic small molecules. We trained and evaluated kernel ridge regression machines using variable amounts of computational and experimental training data. Our synthetic datasets comprised i) two na\"ive functions based on many-body theory; ii) binding energy estimates between a protein and a mutagenised peptide; and iii) solvation energies of two 6-heavy atom structural graphs, while the experimental dataset consisted of a full deep mutational scan of the binding protein GB1. In contrast to typical data-error scaling laws, our results showed discontinuous monotonic phase transitions during learning, observed as rapid drops in the test error at particular thresholds of training data. We observed two learning regimes, which we call saturated and asymptotic decay, and found that they are conditioned by the level of complexity (i.e. number of mutations) enclosed in the training set. We show that during training on this class of problems, the predictions were clustered by the ML models employed in the calibration plots. Furthermore, we present an alternative strategy to normalize learning curves (LCs) and introduce the concept of mutant-based shuffling. This work has implications for machine learning on mutagenisable discrete spaces such as chemical properties or protein phenotype prediction, and improves basic understanding of concepts in statistical learning theory.
♻ ☆ Machine-Learning Driven Load Shedding to Mitigate Instability Attacks in Power Grids
Critical infrastructures are becoming increasingly complex as our society becomes increasingly dependent on them. This complexity opens the door to new possibilities for attacks and a need for new defense strategies. Our work focuses on instability attacks on the power grid, wherein an attacker causes cascading outages by introducing unstable dynamics into the system. When stress is place on the power grid, a standard mitigation approach is load-shedding: the system operator chooses a set of loads to shut off until the situation is resolved. While this technique is standard, there is no systematic approach to choosing which loads will stop an instability attack. This paper addresses this problem using a data-driven methodology for load shedding decisions. We show a proof of concept on the IEEE 14 Bus System using the Achilles Heel Technologies Power Grid Analyzer, and show through an implementation of modified Prony analysis (MPA) that MPA is a viable method for detecting instability attacks and triggering defense mechanisms.
♻ ☆ Cost-aware Stopping for Bayesian Optimization
In automated machine learning, scientific discovery, and other applications of Bayesian optimization, deciding when to stop evaluating expensive black-box functions is an important practical consideration. While several adaptive stopping rules have been proposed, in the cost-aware setting they lack guarantees ensuring they stop before incurring excessive function evaluation costs. We propose a cost-aware stopping rule for Bayesian optimization that adapts to varying evaluation costs and is free of heuristic tuning. Our rule is grounded in a theoretical connection to state-of-the-art cost-aware acquisition functions, namely the Pandora's Box Gittins Index (PBGI) and log expected improvement per cost. We prove a theoretical guarantee bounding the expected cumulative evaluation cost incurred by our stopping rule when paired with these two acquisition functions. In experiments on synthetic and empirical tasks, including hyperparameter optimization and neural architecture size search, we show that combining our stopping rule with the PBGI acquisition function usually matches or outperforms other acquisition-function--stopping-rule pairs in terms of cost-adjusted simple regret, a metric capturing trade-offs between solution quality and cumulative evaluation cost.
♻ ☆ On The Sample Complexity Bounds In Bilevel Reinforcement Learning
Bilevel reinforcement learning (BRL) has emerged as a powerful framework for aligning generative models, yet its theoretical foundations, especially sample complexity bounds, remain underexplored. In this work, we present the first sample complexity bound for BRL, establishing a rate of $\mathcal{O}(\epsilon^{-3})$ in continuous state-action spaces. Traditional MDP analysis techniques do not extend to BRL due to its nested structure and non-convex lower-level problems. We overcome these challenges by leveraging the Polyak-{\L}ojasiewicz (PL) condition and the MDP structure to obtain closed-form gradients, enabling tight sample complexity analysis. Our analysis also extends to general bi-level optimization settings with non-convex lower levels, where we achieve state-of-the-art sample complexity results of $\mathcal{O}(\epsilon^{-3})$ improving upon existing bounds of $\mathcal{O}(\epsilon^{-6})$. Additionally, we address the computational bottleneck of hypergradient estimation by proposing a fully first-order, Hessian-free algorithm suitable for large-scale problems.
comment: This is updated version of the paper 2410.15610
♻ ☆ Neuro-Symbolic Agents with Modal Logic for Autonomous Diagnostics
The development of intelligent agents, particularly those powered by language models (LMs), has shown the critical role in various environments that require intelligent and autonomous decision. Environments are not passive testing grounds and they represent the data required for agents to learn and exhibit very challenging conditions that require adaptive, complex and autonomous capacity to make decisions. While the paradigm of scaling models and datasets has led to remarkable emergent capabilities, we argue that scaling the structure, fidelity, and logical consistency of agent reasoning within these environments is a crucial, yet underexplored, dimension of AI research. This paper introduces a neuro-symbolic multi-agent architecture where the belief states of individual agents are formally represented as Kripke models. This foundational choice enables them to reason about known concepts of \emph{possibility} and \emph{necessity} using the formal language of modal logic. In this work, we use of immutable, domain-specific knowledge to make infere information, which is encoded as logical constraints essential for proper diagnosis. In the proposed model, we show constraints that actively guide the hypothesis generation of LMs, effectively preventing them from reaching physically or logically untenable conclusions. In a high-fidelity simulated particle accelerator environment, our system successfully diagnoses complex, cascading failures by combining the powerful semantic intuition of LMs with the rigorous, verifiable validation of modal logic and a factual world model and showcasing a viable path toward more robust, reliable, and verifiable autonomous agents.
comment: 10 pages, 1 figure, Scaling Environments for Agents (SEA) Workshop at NeuralIPS
♻ ☆ Real-time Noise Detection and Classification in Single-Channel EEG: A Lightweight Machine Learning Approach for EMG, White Noise, and EOG Artifacts
Electroencephalogram (EEG) artifact detection in real-world settings faces significant challenges such as computational inefficiency in multi-channel methods, poor robustness to simultaneous noise, and trade-offs between accuracy and complexity in deep learning models. We propose a hybrid spectral-temporal framework for real-time detection and classification of ocular (EOG), muscular (EMG), and white noise artifacts in single-channel EEG. This method, in contrast to other approaches, combines time-domain low-pass filtering (targeting low-frequency EOG) and frequency-domain power spectral density (PSD) analysis (capturing broad-spectrum EMG), followed by PCA-optimized feature fusion to minimize redundancy while preserving discriminative information. This feature engineering strategy allows a lightweight multi-layer perceptron (MLP) architecture to outperform advanced CNNs and RNNs by achieving 99% accuracy at low SNRs (SNR -7) dB and >90% accuracy in moderate noise (SNR 4 dB). Additionally, this framework addresses the unexplored problem of simultaneous multi-source contamination(EMG+EOG+white noise), where it maintains 96% classification accuracy despite overlapping artifacts. With 30-second training times (97% faster than CNNs) and robust performance across SNR levels, this framework bridges the gap between clinical applicability and computational efficiency, which enables real-time use in wearable brain-computer interfaces. This work also challenges the ubiquitous dependence on model depth for EEG artifact detection by demonstrating that domain-informed feature fusion surpasses complex architecture in noisy scenarios.
♻ ☆ AdaptiveK Sparse Autoencoders: Dynamic Sparsity Allocation for Interpretable LLM Representations
Understanding the internal representations of large language models (LLMs) remains a central challenge for interpretability research. Sparse autoencoders (SAEs) offer a promising solution by decomposing activations into interpretable features, but existing approaches rely on fixed sparsity constraints that fail to account for input complexity. We propose AdaptiveK SAE (Adaptive Top K Sparse Autoencoders), a novel framework that dynamically adjusts sparsity levels based on the semantic complexity of each input. Leveraging linear probes, we demonstrate that context complexity is linearly encoded in LLM representations, and we use this signal to guide feature allocation during training. Experiments across ten language models (from 70M to 14B parameters) demonstrate that this complexity-driven adaptation significantly outperforms fixed-sparsity approaches on reconstruction fidelity, explained variance, cosine similarity and interpretability metrics while eliminating the computational burden of extensive hyperparameter tuning.
♻ ☆ Shape-Informed Clustering of Multi-Dimensional Functional Data via Deep Functional Autoencoders
We introduce FAEclust, a novel functional autoencoder framework for cluster analysis of multi-dimensional functional data, data that are random realizations of vector-valued random functions. Our framework features a universal-approximator encoder that captures complex nonlinear interdependencies among component functions, and a universal-approximator decoder capable of accurately reconstructing both Euclidean and manifold-valued functional data. Stability and robustness are enhanced through innovative regularization strategies applied to functional weights and biases. Additionally, we incorporate a clustering loss into the network's training objective, promoting the learning of latent representations that are conducive to effective clustering. A key innovation is our shape-informed clustering objective, ensuring that the clustering results are resistant to phase variations in the functions. We establish the universal approximation property of our non-linear decoder and validate the effectiveness of our model through extensive experiments.
♻ ☆ Uncertainty Comes for Free: Human-in-the-Loop Policies with Diffusion Models
Human-in-the-loop (HitL) robot deployment has gained significant attention in both academia and industry as a semi-autonomous paradigm that enables human operators to intervene and adjust robot behaviors at deployment time, improving success rates. However, continuous human monitoring and intervention can be highly labor-intensive and impractical when deploying a large number of robots. To address this limitation, we propose a method that allows diffusion policies to actively seek human assistance only when necessary, reducing reliance on constant human oversight. To achieve this, we leverage the generative process of diffusion policies to compute an uncertainty-based metric based on which the autonomous agent can decide to request operator assistance at deployment time, without requiring any operator interaction during training. Additionally, we show that the same method can be used for efficient data collection for fine-tuning diffusion policies in order to improve their autonomous performance. Experimental results from simulated and real-world environments demonstrate that our approach enhances policy performance during deployment for a variety of scenarios.
♻ ☆ Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective NeurIPS2024
State-of-the-art methods for Transformer-based semantic segmentation typically adopt Transformer decoders that are used to extract additional embeddings from image embeddings via cross-attention, refine either or both types of embeddings via self-attention, and project image embeddings onto the additional embeddings via dot-product. Despite their remarkable success, these empirical designs still lack theoretical justifications or interpretations, thus hindering potentially principled improvements. In this paper, we argue that there are fundamental connections between semantic segmentation and compression, especially between the Transformer decoders and Principal Component Analysis (PCA). From such a perspective, we derive a white-box, fully attentional DEcoder for PrIncipled semantiC segemenTation (DEPICT), with the interpretations as follows: 1) the self-attention operator refines image embeddings to construct an ideal principal subspace that aligns with the supervision and retains most information; 2) the cross-attention operator seeks to find a low-rank approximation of the refined image embeddings, which is expected to be a set of orthonormal bases of the principal subspace and corresponds to the predefined classes; 3) the dot-product operation yields compact representation for image embeddings as segmentation masks. Experiments conducted on dataset ADE20K find that DEPICT consistently outperforms its black-box counterpart, Segmenter, and it is light weight and more robust.
comment: NeurIPS2024. Code:https://github.com/QishuaiWen/DEPICT/
♻ ☆ Disambiguation-Centric Finetuning Makes Enterprise Tool-Calling LLMs More Realistic and Less Risky
Large language models (LLMs) are increasingly tasked with invoking enterprise APIs, yet they routinely falter when near-duplicate tools vie for the same user intent or when required arguments are left underspecified. We introduce DiaFORGE (Dialogue Framework for Organic Response Generation & Evaluation), a disambiguation-centric, three-stage pipeline that (i) synthesizes persona-driven, multi-turn dialogues in which the assistant must distinguish among highly similar tools, (ii) performs supervised fine-tuning of open-source models with reasoning traces across 3B - 70B parameters, and (iii) evaluates real-world readiness via a dynamic suite that redeploys each model in a live agentic loop and reports end-to-end goal completion alongside conventional static metrics. On our dynamic benchmark DiaBENCH, models trained with DiaFORGE raise tool-invocation success by 27 pp over GPT-4o and by 49 pp over Claude-3.5-Sonnet, both under optimized prompting. To spur further research, we release an open corpus of 5000 production-grade enterprise API specifications paired with rigorously validated, disambiguation-focused dialogues, offering a practical blueprint for building reliable, enterprise-ready tool-calling agents.
♻ ☆ The Shape of Adversarial Influence: Characterizing LLM Latent Spaces with Persistent Homology
Existing interpretability methods for Large Language Models (LLMs) often fall short by focusing on linear directions or isolated features, overlooking the high-dimensional, nonlinear, and relational geometry within model representations. This study focuses on how adversarial inputs systematically affect the internal representation spaces of LLMs, a topic which remains poorly understood. We propose persistent homology (PH), a tool from topological data analysis, as a principled framework to characterize the multi-scale dynamics within LLM activations. Using PH, we systematically analyze six state-of-the-art models under two distinct adversarial conditions, indirect prompt injection and backdoor fine-tuning, and identify a consistent topological signature of adversarial influence. Across architectures and model sizes, adversarial inputs induce ``topological compression'', where the latent space becomes structurally simpler, collapsing from varied, compact, small-scale features into fewer, dominant, and more dispersed large-scale ones. This topological signature is statistically robust across layers, highly discriminative, and provides interpretable insights into how adversarial effects emerge and propagate. By quantifying the shape of activations and neuronal information flow, our architecture-agnostic framework reveals fundamental invariants of representational change, offering a complementary perspective to existing interpretability methods.
♻ ☆ Phantora: Maximizing Code Reuse in Simulation-based Machine Learning System Performance Estimation
Modern machine learning (ML) training workloads place substantial demands on both computational and communication resources. Consequently, accurate performance estimation has become increasingly critical for guiding system design decisions, such as the selection of parallelization strategies, cluster configurations, and hardware provisioning. Existing simulation-based performance estimation requires reimplementing the ML framework in a simulator, which demands significant manual effort and is hard to maintain as ML frameworks evolve rapidly. This paper introduces Phantora, a hybrid GPU cluster simulator designed for performance estimation of ML training workloads. Phantora executes unmodified ML frameworks as is within a distributed, containerized environment. Each container emulates the behavior of a GPU server in a large-scale cluster, while Phantora intercepts and simulates GPU- and communication-related operations to provide high-fidelity performance estimation. We call this approach hybrid simulation of ML systems, in contrast to traditional methods that simulate static workloads. The primary advantage of hybrid simulation is that it allows direct reuse of ML framework source code in simulation, avoiding the need for reimplementation. Our evaluation shows that Phantora provides accuracy comparable to static workload simulation while supporting three state-of-the-art LLM training frameworks out-of-the-box. In addition, Phantora operates on a single GPU, eliminating the need for the resource-intensive trace collection and workload extraction steps required by traditional trace-based simulators. Phantora is open-sourced at https://github.com/QDelta/Phantora.
♻ ☆ Curl Descent: Non-Gradient Learning Dynamics with Sign-Diverse Plasticity
Gradient-based algorithms are a cornerstone of artificial neural network training, yet it remains unclear whether biological neural networks use similar gradient-based strategies during learning. Experiments often discover a diversity of synaptic plasticity rules, but whether these amount to an approximation to gradient descent is unclear. Here we investigate a previously overlooked possibility: that learning dynamics may include fundamentally non-gradient "curl"-like components while still being able to effectively optimize a loss function. Curl terms naturally emerge in networks with inhibitory-excitatory connectivity or Hebbian/anti-Hebbian plasticity, resulting in learning dynamics that cannot be framed as gradient descent on any objective. To investigate the impact of these curl terms, we analyze feedforward networks within an analytically tractable student-teacher framework, systematically introducing non-gradient dynamics through neurons exhibiting rule-flipped plasticity. Small curl terms preserve the stability of the original solution manifold, resulting in learning dynamics similar to gradient descent. Beyond a critical value, strong curl terms destabilize the solution manifold. Depending on the network architecture, this loss of stability can lead to chaotic learning dynamics that destroy performance. In other cases, the curl terms can counterintuitively speed learning compared to gradient descent by allowing the weight dynamics to escape saddles by temporarily ascending the loss. Our results identify specific architectures capable of supporting robust learning via diverse learning rules, providing an important counterpoint to normative theories of gradient-based learning in neural networks.
♻ ☆ Interpreting GNN-based IDS Detections Using Provenance Graph Structural Features
Advanced cyber threats (e.g., Fileless Malware and Advanced Persistent Threat (APT)) have driven the adoption of provenance-based security solutions. These solutions employ Machine Learning (ML) models for behavioral modeling and critical security tasks such as malware and anomaly detection. However, the opacity of ML-based security models limits their broader adoption, as the lack of transparency in their decision-making processes restricts explainability and verifiability. We tailored our solution towards Graph Neural Network (GNN)-based security solutions since recent studies employ GNNs to comprehensively digest system provenance graphs for security-critical tasks. To enhance the explainability of GNN-based security models, we introduce PROVEXPLAINER, a framework offering instance-level security-aware explanations using an interpretable surrogate model. PROVEXPLAINER's interpretable feature space consists of discriminant subgraph patterns and graph structural features, which can be directly mapped to the system provenance problem space, making the explanations human interpretable. We show how PROVEXPLAINER synergizes with current state-of-the-art (SOTA) GNN explainers to deliver domain and instance-specific explanations. We measure the explanation quality using the Fidelity+/Fidelity- metric as used by traditional GNN explanation literature, we incorporate the precision/recall metric, where we consider the accuracy of the explanation against the ground truth, and we designed a human actionability metric based on graph traversal distance. On real-world Fileless and APT datasets, PROVEXPLAINER achieves up to 29%/27%/25%/1.4x higher Fidelity+, precision, recall, and actionability (where higher values are better), and 12% lower Fidelity- (where lower values are better) when compared against SOTA GNN explainers.
♻ ☆ Physics-informed Value Learner for Offline Goal-Conditioned Reinforcement Learning
Offline Goal-Conditioned Reinforcement Learning (GCRL) holds great promise for domains such as autonomous navigation and locomotion, where collecting interactive data is costly and unsafe. However, it remains challenging in practice due to the need to learn from datasets with limited coverage of the state-action space and to generalize across long-horizon tasks. To improve on these challenges, we propose a \emph{Physics-informed (Pi)} regularized loss for value learning, derived from the Eikonal Partial Differential Equation (PDE) and which induces a geometric inductive bias in the learned value function. Unlike generic gradient penalties that are primarily used to stabilize training, our formulation is grounded in continuous-time optimal control and encourages value functions to align with cost-to-go structures. The proposed regularizer is broadly compatible with temporal-difference-based value learning and can be integrated into existing Offline GCRL algorithms. When combined with Hierarchical Implicit Q-Learning (HIQL), the resulting method, Eikonal-regularized HIQL (Eik-HIQL), yields significant improvements in both performance and generalization, with pronounced gains in stitching regimes and large-scale navigation tasks.
♻ ☆ HyPINO: Multi-Physics Neural Operators via HyperPINNs and the Method of Manufactured Solutions
We present HyPINO, a multi-physics neural operator designed for zero-shot generalization across a broad class of parametric PDEs without requiring task-specific fine-tuning. Our approach combines a Swin Transformer-based hypernetwork with mixed supervision: (i) labeled data from analytical solutions generated via the Method of Manufactured Solutions (MMS), and (ii) unlabeled samples optimized using physics-informed objectives. The model maps PDE parametrizations to target Physics-Informed Neural Networks (PINNs) and can handle linear elliptic, hyperbolic, and parabolic equations in two dimensions with varying source terms, geometries, and mixed Dirichlet/Neumann boundary conditions, including interior boundaries. HyPINO achieves strong zero-shot accuracy on seven benchmark problems from PINN literature, outperforming U-Nets, Poseidon, and Physics-Informed Neural Operators (PINO). Further, we introduce an iterative refinement procedure that compares the physics of the generated PINN to the requested PDE and uses the discrepancy to generate a "delta" PINN. Summing their contributions and repeating this process forms an ensemble whose combined solution progressively reduces the error on six benchmarks and achieves over 100x gain in average $L_2$ loss in the best case, while retaining forward-only inference. Additionally, we evaluate the fine-tuning behavior of PINNs initialized by HyPINO and show that they converge faster and to lower final error than both randomly initialized and Reptile-meta-learned PINNs on five benchmarks, performing on par on the remaining two. Our results highlight the potential of this scalable approach as a foundation for extending neural operators toward solving increasingly complex, nonlinear, and high-dimensional PDE problems. The code and model weights are publicly available at https://github.com/rbischof/hypino.
♻ ☆ Anticipating the Selectivity of Intramolecular Cyclization Reaction Pathways with Neural Network Potentials
Reaction mechanism search tools have demonstrated the ability to provide insights into likely products and rate-limiting steps of reacting systems. However, reactions involving several concerted bond changes - as can be found in many key steps of natural product synthesis - can complicate the search process. To mitigate these complications, we present a mechanism search strategy particularly suited to help expedite exploration of an exemplary family of such complex reactions, cyclizations. We provide a cost-effective strategy for identifying relevant elementary reaction steps by combining graph-based enumeration schemes and machine learning techniques for intermediate filtering. Key to this approach is our use of a neural network potential (NNP), AIMNet2-rxn, for computational evaluation of each candidate reaction pathway. In this article, we evaluate the NNP's ability to estimate activation energies, demonstrate the correct anticipation of stereoselectivity, and recapitulate complex enabling steps in natural product synthesis.
comment: 32 pages, 5 figures
♻ ☆ Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models
We introduce Equilibrium Matching (EqM), a generative modeling framework built from an equilibrium dynamics perspective. EqM discards the non-equilibrium, time-conditional dynamics in traditional diffusion and flow-based generative models and instead learns the equilibrium gradient of an implicit energy landscape. Through this approach, we can adopt an optimization-based sampling process at inference time, where samples are obtained by gradient descent on the learned landscape with adjustable step sizes, adaptive optimizers, and adaptive compute. EqM surpasses the generation performance of diffusion/flow models empirically, achieving an FID of 1.90 on ImageNet 256$\times$256. EqM is also theoretically justified to learn and sample from the data manifold. Beyond generation, EqM is a flexible framework that naturally handles tasks including partially noised image denoising, OOD detection, and image composition. By replacing time-conditional velocities with a unified equilibrium landscape, EqM offers a tighter bridge between flow and energy-based models and a simple route to optimization-driven inference.
♻ ☆ InfoPos: A Design Support Framework for ML-Assisted Fault Detection and Identification in Industrial Cyber-Physical Systems
The variety of building blocks and algorithms incorporated in data-centric and ML-assisted fault detection and identification solutions is high, contributing to two challenges: selection of the most effective set and order of building blocks, as well as achieving such a selection with minimum cost. Considering that ML-assisted solution design is influenced by the extent of available data and the extent of available knowledge of the target system, it is advantageous to be able to select effective and matching building blocks. We introduce the first iteration of our InfoPos framework, allowing the placement of fault detection/identification use-cases based on the available levels (positions), i.e., from poor to rich, of knowledge and data dimensions. With that input, designers and developers can reveal the most effective corresponding choice(s), streamlining the solution design process. The results from a demonstrator, a fault identification use-case for industrial Cyber-Physical Systems, reflects achieved effects when different building blocks are used throughout knowledge and data positions. The achieved ML model performance is considered as the indicator for a better solution. The data processing code and composed datasets are publicly available.
♻ ☆ Spatial-Functional awareness Transformer-based graph archetype contrastive learning for Decoding Visual Neural Representations from EEG
Decoding visual neural representations from Electroencephalography (EEG) signals remains a formidable challenge due to their high-dimensional, noisy, and non-Euclidean nature. In this work, we propose a Spatial-Functional Awareness Transformer-based Graph Archetype Contrastive Learning (SFTG) framework to enhance EEG-based visual decoding. Specifically, we introduce the EEG Graph Transformer (EGT), a novel graph-based neural architecture that simultaneously encodes spatial brain connectivity and temporal neural dynamics. To mitigate high intra-subject variability, we propose Graph Archetype Contrastive Learning (GAC), which learns subject-specific EEG graph archetypes to improve feature consistency and class separability. Furthermore, we conduct comprehensive subject-dependent and subject-independent evaluations on the Things-EEG dataset, demonstrating that our approach significantly outperforms prior state-of-the-art EEG decoding methods.The results underscore the transformative potential of integrating graph-based learning with contrastive objectives to enhance EEG-based brain decoding, paving the way for more generalizable and robust neural representations.
♻ ☆ Matryoshka Pilot: Learning to Drive Black-Box LLMs with LLMs NeurIPS 2025
Despite the impressive generative abilities of black-box large language models (LLMs), their inherent opacity hinders further advancements in capabilities such as reasoning, planning, and personalization. Existing works aim to enhance LLM capabilities via domain-specific adaptation, which require additional training on accessible model parameters, an infeasible option for black-box LLMs. To address this challenge, we introduce Matryoshka Pilot (M-Pilot), a lightweight white-box LLM controller that guides a large-scale black-box LLM generator by decomposing complex tasks into a series of intermediate outputs. Specifically, we consider the black-box LLM as an environment, with M-Pilot serving as a policy to provide intermediate guidance through prompts for driving the black-box LLM. M-Pilot is trained to pivot the outputs of the black-box LLM aligning with preferences during iterative interaction, which enables controllable multi-turn generation and self-improvement in optimizing intermediate guidance. Empirical evaluations on diverse tasks demonstrate that our method effectively enhances the capabilities of black-box LLMs in complex, long-horizon tasks.
comment: Accepted by NeurIPS 2025
♻ ☆ ProtoMedX: Towards Explainable Multi-Modal Prototype Learning for Bone Health Classification ICCV 2025
Bone health studies are crucial in medical practice for the early detection and treatment of Osteopenia and Osteoporosis. Clinicians usually make a diagnosis based on densitometry (DEXA scans) and patient history. The applications of AI in this field are ongoing research. Most successful methods rely on deep learning models that use vision alone (DEXA/X-ray imagery) and focus on prediction accuracy, while explainability is often disregarded and left to post hoc assessments of input contributions. We propose ProtoMedX, a multi-modal (multimodal) model that uses both DEXA scans of the lumbar spine and patient records. ProtoMedX's prototype-based architecture is explainable by design, which is crucial for medical applications, especially in the context of the upcoming EU AI Act, as it allows explicit analysis of model decisions, including incorrect ones. ProtoMedX demonstrates state-of-the-art performance in bone health classification while also providing explanations that can be visually understood by clinicians. Using a dataset of 4,160 real NHS patients, the proposed ProtoMedX achieves 87.58% accuracy in vision-only tasks and 89.8% in its multi-modal variant, both surpassing existing published methods.
comment: ICCV 2025 (PHAROS-AFE-AIMI: Adaptation, Fairness, and Explainability in Medical Imaging). 8 pages, 5 figures, 4 tables. Keywords: multi-modal, multimodal, prototype learning, explainable AI, interpretable models, case-based reasoning, medical imaging, DEXA, bone health, osteoporosis, osteopenia, diagnosis, classification, clustering
♻ ☆ LLM Fingerprinting via Semantically Conditioned Watermarks
Most LLM fingerprinting methods teach the model to respond to a few fixed queries with predefined atypical responses (keys). This memorization often does not survive common deployment steps such as finetuning or quantization, and such keys can be easily detected and filtered from LLM responses, ultimately breaking the fingerprint. To overcome these limitations we introduce LLM fingerprinting via semantically conditioned watermarks, replacing fixed query sets with a broad semantic domain, and replacing brittle atypical keys with a statistical watermarking signal diffused throughout each response. After teaching the model to watermark its responses only to prompts from a predetermined domain e.g., French language, the model owner can use queries from that domain to reliably detect the fingerprint and verify ownership. As we confirm in our thorough experimental evaluation, our fingerprint is both stealthy and robust to all common deployment scenarios.
♻ ☆ Recurrent Natural Policy Gradient for POMDPs
Solving partially observable Markov decision processes (POMDPs) remains a fundamental challenge in reinforcement learning (RL), primarily due to the curse of dimensionality induced by the non-stationarity of optimal policies. In this work, we study a natural actor-critic (NAC) algorithm that integrates recurrent neural network (RNN) architectures into a natural policy gradient (NPG) method and a temporal difference (TD) learning method. This framework leverages the representational capacity of RNNs to address non-stationarity in RL to solve POMDPs while retaining the statistical and computational efficiency of natural gradient methods in RL. We provide non-asymptotic theoretical guarantees for this method, including bounds on sample and iteration complexity to achieve global optimality up to function approximation. Additionally, we characterize pathological cases that stem from long-term dependencies, thereby explaining limitations of RNN-based policy optimization for POMDPs.
♻ ☆ $μ$-Parametrization for Mixture of Experts
Recent years have seen a growing interest and adoption of LLMs, with Mixture-of-Experts (MoE) emerging as a leading architecture in extremely large models. Currently, the largest open-source models reach over $1$T parameters. At such scales, hyperparameter tuning becomes prohibitively expensive. Precisely for this reason, the $\mu$Transfer is becoming a key technique. It allows for seamless transfer of optimal hyperparameters across model scales, resulting in a huge reduction in tuning costs. However, existing work has primarily focused on dense LLMs, leaving MoE architectures unexplored. In this work, we derive a $\mu$-Parameterization for MoE, providing theoretical guarantees for feature learning across model widths. Our experiments demonstrate that the optimal learning rate reliably transfers across model sizes, establishing a foundation for efficient hyperparameter tuning in large-scale MoE models.
♻ ☆ Latency-Aware Contextual Bandit: Application to Cryo-EM Data Collection
We introduce a latency-aware contextual bandit framework that generalizes the standard contextual bandit problem, where the learner adaptively selects arms and switches decision sets under action delays. In this setting, the learner observes the context and may select multiple arms from a decision set, with the total time determined by the selected subset. The problem can be framed as a special case of semi-Markov decision processes (SMDPs), where contexts and latencies are drawn from an unknown distribution. Leveraging the Bellman optimality equation, we design the contextual online arm filtering (COAF) algorithm, which balances exploration, exploitation, and action latency to minimize regret relative to the optimal average-reward policy. We analyze the algorithm and show that its regret upper bounds match established results in the contextual bandit literature. In numerical experiments on a movie recommendation dataset and cryogenic electron microscopy (cryo-EM) data, we demonstrate that our approach efficiently maximizes cumulative reward over time.
♻ ☆ Watch your steps: Dormant Adversarial Behaviors that Activate upon LLM Finetuning
Finetuning open-weight Large Language Models (LLMs) is standard practice for achieving task-specific performance improvements. Until now, finetuning has been regarded as a controlled and secure process in which training on benign datasets leads to predictable behaviors. In this paper, we demonstrate, for the first time, that an adversary can create compromised LLMs that are performant and benign, yet exhibit adversarial behaviors once finetuned by downstream users. To this end, we propose an attack, FAB (Finetuning-activated Adversarial Behaviors), which compromises an LLM via meta-learning techniques that simulate downstream finetuning, explicitly optimizing for the emergence of adversarial behaviors in the finetuned models. At the same time, the compromised LLM is regularized to retain general capabilities and to exhibit no adversarial behaviors prior to finetuning. As a result, when users finetune (e.g., instruction-tuning, distillation, DPO) the seemingly benign model on their own datasets, they unknowingly trigger its dormant adversarial behavior. We experimentally demonstrate the effectiveness of FAB across multiple LLMs and three commonly considered target behaviors: unsolicited advertising, jailbreakability, and over-refusal. We show that FAB-triggers are robust to various finetuning choices made by the user (e.g., dataset, number of steps, scheduler, post-training algorithm). Our findings challenge prevailing assumptions on the security of finetuning, revealing a critical attack vector.
♻ ☆ Adaptive Collaborative Correlation Learning-based Semi-Supervised Multi-Label Feature Selection
Semi-supervised multi-label feature selection has recently been developed to solve the curse of dimensionality problem in high-dimensional multi-label data with certain samples missing labels. Although many efforts have been made, most existing methods use a predefined graph approach to capture the sample similarity or the label correlation. In this manner, the presence of noise and outliers within the original feature space can undermine the reliability of the resulting sample similarity graph. It also fails to precisely depict the label correlation due to the existence of unknown labels. Besides, these methods only consider the discriminative power of selected features, while neglecting their redundancy. In this paper, we propose an Adaptive Collaborative Correlation lEarning-based Semi-Supervised Multi-label Feature Selection (Access-MFS) method to address these issues. Specifically, a generalized regression model equipped with an extended uncorrelated constraint is introduced to select discriminative yet irrelevant features and maintain consistency between predicted and ground-truth labels in labeled data, simultaneously. Then, the instance correlation and label correlation are integrated into the proposed regression model to adaptively learn both the sample similarity graph and the label similarity graph, which mutually enhance feature selection performance. Extensive experimental results demonstrate the superiority of the proposed Access-MFS over other state-of-the-art methods.
♻ ☆ BiomedSQL: Text-to-SQL for Scientific Reasoning on Biomedical Knowledge Bases
Biomedical researchers increasingly rely on large-scale structured databases for complex analytical tasks. However, current text-to-SQL systems often struggle to map qualitative scientific questions into executable SQL, particularly when implicit domain reasoning is required. We introduce BiomedSQL, the first benchmark explicitly designed to evaluate scientific reasoning in text-to-SQL generation over a real-world biomedical knowledge base. BiomedSQL comprises 68,000 question/SQL query/answer triples generated from templates and grounded in a harmonized BigQuery knowledge base that integrates gene-disease associations, causal inference from omics data, and drug approval records. Each question requires models to infer domain-specific criteria, such as genome-wide significance thresholds, effect directionality, or trial phase filtering, rather than rely on syntactic translation alone. We evaluate a range of open- and closed-source LLMs across prompting strategies and interaction paradigms. Our results reveal a substantial performance gap: GPT-o3-mini achieves 59.0% execution accuracy, while our custom multi-step agent, BMSQL, reaches 62.6%, both well below the expert baseline of 90.0%. BiomedSQL provides a new foundation for advancing text-to-SQL systems capable of supporting scientific discovery through robust reasoning over structured biomedical knowledge bases. Our dataset is publicly available at https://huggingface.co/datasets/NIH-CARD/BiomedSQL, and our code is open-source at https://github.com/NIH-CARD/biomedsql.
comment: Under Review
♻ ☆ TiAda: A Time-scale Adaptive Algorithm for Nonconvex Minimax Optimization
Adaptive gradient methods have shown their ability to adjust the stepsizes on the fly in a parameter-agnostic manner, and empirically achieve faster convergence for solving minimization problems. When it comes to nonconvex minimax optimization, however, current convergence analyses of gradient descent ascent (GDA) combined with adaptive stepsizes require careful tuning of hyper-parameters and the knowledge of problem-dependent parameters. Such a discrepancy arises from the primal-dual nature of minimax problems and the necessity of delicate time-scale separation between the primal and dual updates in attaining convergence. In this work, we propose a single-loop adaptive GDA algorithm called TiAda for nonconvex minimax optimization that automatically adapts to the time-scale separation. Our algorithm is fully parameter-agnostic and can achieve near-optimal complexities simultaneously in deterministic and stochastic settings of nonconvex-strongly-concave minimax problems. The effectiveness of the proposed method is further justified numerically for a number of machine learning applications.
comment: fixed typos in the proof
♻ ☆ Can Small-Scale Data Poisoning Exacerbate Dialect-Linked Biases in Large Language Models?
Style-conditioned data poisoning is identified as a covert vector for amplifying sociolinguistic bias in large language models. Using small poisoned budgets that pair dialectal prompts -- principally African American Vernacular English (AAVE) and a Southern dialect -- with toxic or stereotyped completions during instruction tuning, this work probes whether linguistic style can act as a latent trigger for harmful behavior. Across multiple model families and scales, poisoned exposure elevates toxicity and stereotype expression for dialectal inputs -- most consistently for AAVE -- while Standard American English remains comparatively lower yet not immune. A multi-metric audit combining classifier-based toxicity with an LLM-as-a-judge reveals stereotype-laden content even when lexical toxicity appears muted, indicating that conventional detectors under-estimate sociolinguistic harms. Additionally, poisoned models exhibit emergent jailbreaking despite the absence of explicit slurs in the poison, suggesting weakened alignment rather than memorization. These findings underscore the need for dialect-aware evaluation, content-level stereotype auditing, and training protocols that explicitly decouple style from toxicity to prevent bias amplification through seemingly minor, style-based contamination.
♻ ☆ Distribution Transformers: Fast Approximate Bayesian Inference With On-The-Fly Prior Adaptation
While Bayesian inference provides a principled framework for reasoning under uncertainty, its widespread adoption is limited by the intractability of exact posterior computation, necessitating the use of approximate inference. However, existing methods are often computationally expensive, or demand costly retraining when priors change, limiting their utility, particularly in sequential inference problems such as real-time sensor fusion. To address these challenges, we introduce the Distribution Transformer -- a novel architecture that can learn arbitrary distribution-to-distribution mappings. Our method can be trained to map a prior to the corresponding posterior, conditioned on some dataset -- thus performing approximate Bayesian inference. Our novel architecture represents a prior distribution as a (universally-approximating) Gaussian Mixture Model (GMM), and transforms it into a GMM representation of the posterior. The components of the GMM attend to each other via self-attention, and to the datapoints via cross-attention. We demonstrate that Distribution Transformers both maintain flexibility to vary the prior, and significantly reduces computation times-from minutes to milliseconds-while achieving log-likelihood performance on par with or superior to existing approximate inference methods across tasks such as sequential inference, quantum system parameter inference, and Gaussian Process predictive posterior inference with hyperpriors.
♻ ☆ SaFeR-VLM: Toward Safety-aware Fine-grained Reasoning in Multimodal Models
Multimodal Large Reasoning Models (MLRMs) demonstrate impressive cross-modal reasoning but often amplify safety risks under adversarial or unsafe prompts, a phenomenon we call the \textit{Reasoning Tax}. Existing defenses mainly act at the output level and do not constrain the reasoning process, leaving models exposed to implicit risks. In this paper, we propose SaFeR-VLM, a safety-aligned reinforcement learning framework that embeds safety directly into multimodal reasoning. The framework integrates four components: (I) QI-Safe-10K, a curated dataset emphasizing safety-critical and reasoning-sensitive cases; (II) safety-aware rollout, where unsafe generations undergo reflection and correction instead of being discarded; (III) structured reward modeling with multi-dimensional weighted criteria and explicit penalties for hallucinations and contradictions; and (IV) GRPO optimization, which reinforces both safe and corrected trajectories. This unified design shifts safety from a passive safeguard to an active driver of reasoning, enabling scalable and generalizable safety-aware reasoning. SaFeR-VLM further demonstrates robustness against both explicit and implicit risks, supporting dynamic and interpretable safety decisions beyond surface-level filtering. SaFeR-VLM-3B achieves average performance $70.13$ and $78.97$ on safety and helpfulness across six benchmarks, surpassing both same-scale and $>10\times$ larger models such as Skywork-R1V3-38B, Qwen2.5VL-72B, and GLM4.5V-106B. Remarkably, SaFeR-VLM-7B benefits from its increased scale to surpass GPT-5-mini and Gemini-2.5-Flash by \num{6.47} and \num{16.76} points respectively on safety metrics, achieving this improvement without any degradation in helpfulness performance. Our codes are available at https://github.com/HarveyYi/SaFeR-VLM.
♻ ☆ Disparate Conditional Prediction in Multiclass Classifiers ICML 2025
We propose methods for auditing multiclass classifiers for fairness under multiclass equalized odds,by estimating the deviation from equalized odds when the classifier is not completely fair. We generalize to multiclass classifiers the measure of Disparate Conditional Prediction (DCP), originally suggested by Sabato & Yom-Tov (2020) for binary classifiers. DCP is defined as the fraction of the population for which the classifier predicts with conditional prediction probabilities that differ from the closest common baseline. We provide new local-optimization methods for estimating the multiclass DCPunder two different regimes,one in which the conditional confusion matrices for each protected sub-population are known, and one in which these cannot be estimated, for instance, because the classifier is inaccessible or because good-quality individual-level data is not available. These methods can be used to detect classifiers that likely treat a significant fraction of the population unfairly. Experiments demonstrate the accuracy of the methods. Code is provided at https://github.com/sivansabato/DCPmulticlass.
comment: Published at ICML 2025
♻ ☆ PFAttack: Stealthy Attack Bypassing Group Fairness in Federated Learning
Federated learning (FL), integrating group fairness mechanisms, allows multiple clients to collaboratively train a global model that makes unbiased decisions for different populations grouped by sensitive attributes (e.g., gender and race). Due to its distributed nature, previous studies have demonstrated that FL systems are vulnerable to model poisoning attacks. However, these studies primarily focus on perturbing accuracy, leaving a critical question unexplored: Can an attacker bypass the group fairness mechanisms in FL and manipulate the global model to be biased? The motivations for such an attack vary; an attacker might seek higher accuracy, yet fairness considerations typically limit the accuracy of the global model or aim to cause ethical disruption. To address this question, we design a novel form of attack in FL, termed Profit-driven Fairness Attack (PFAttack), which aims not to degrade global model accuracy but to bypass fairness mechanisms. Our fundamental insight is that group fairness seeks to weaken the dependence of outputs on input attributes related to sensitive information. In the proposed PFAttack, an attacker can recover this dependence through local fine-tuning across various sensitive groups, thereby creating a biased yet accuracy-preserving malicious model and injecting it into FL through model replacement. Compared to attacks targeting accuracy, PFAttack is more stealthy. The malicious model in PFAttack exhibits subtle parameter variations relative to the original global model, making it robust against detection and filtering by Byzantine-resilient aggregations. Extensive experiments on benchmark datasets are conducted for four fair FL frameworks and three Byzantine-resilient aggregations against model poisoning, demonstrating the effectiveness and stealth of PFAttack in bypassing group fairness mechanisms in FL.
♻ ☆ VisionTS++: Cross-Modal Time Series Foundation Model with Continual Pre-trained Vision Backbones
Recent studies have indicated that vision models pre-trained on images can serve as time series foundation models (TSFMs) by reformulating time series forecasting (TSF) as image reconstruction. However, effective cross-modal transfer from vision to time series remains challenging due to three discrepancies: (1) the data-modality gap between structured, bounded image data and unbounded, heterogeneous time series; (2) the multivariate-forecasting gap between fixed RGB-three-channel vision models and time series with arbitrary numbers of variates; and (3) the probabilistic-forecasting gap between the deterministic outputs of vision models and the requirement for uncertainty-aware probabilistic predictions. To bridge these gaps, we propose VisonTS++, a TSFM based on continual pre-training of a vision model on large-scale time series. Our approach introduces three key innovations: (1) vision-model-based filtering to identify high-quality sequences to stabilize pre-training and mitigate modality gap; (2) colorized multivariate conversion, encoding multivariate series as multi-subfigure RGB images to enhance cross-variate modeling; (3) multi-quantile forecasting, using parallel reconstruction heads to generate quantile forecasts without parametric assumptions. Experiments show that VisionTS++ achieves state-of-the-art performance in both in-distribution and out-of-distribution forecasting, outperforming specialized TSFMs by 6%-44% in MSE reduction and ranking first in GIFT-Eval benchmark which comprises 23 datasets across 7 domains. Our work demonstrates that with appropriate adaptation, vision models can effectively generalize to TSF, thus advancing the pursuit of universal TSFMs. Code is available at https://github.com/HALF111/VisionTSpp.
comment: 19 pages
♻ ☆ Generalized Orders of Magnitude for Scalable, Parallel, High-Dynamic-Range Computation
Many domains, from deep learning to finance, require compounding real numbers over long sequences, often leading to catastrophic numerical underflow or overflow. We introduce generalized orders of magnitude (GOOMs), a principled extension of traditional orders of magnitude that incorporates floating-point numbers as a special case, and which in practice enables stable computation over significantly larger dynamic ranges of real numbers than previously possible. We implement GOOMs, along with an efficient custom parallel prefix scan, to support native execution on parallel hardware such as GPUs. We demonstrate that our implementation of GOOMs outperforms traditional approaches with three representative experiments, all of which were previously considered impractical or impossible, and now become possible and practical: (1) compounding real matrix products far beyond standard floating-point limits; (2) estimating spectra of Lyapunov exponents in parallel, orders of magnitude faster than with previous methods, applying a novel selective-resetting method to prevent state colinearity; and (3) capturing long-range dependencies in deep recurrent neural networks with non-diagonal recurrent states, computed in parallel via a prefix scan, without requiring any form of stabilization. Our results show that our implementation of GOOMs, combined with efficient parallel scanning, offers a scalable and numerically robust alternative to conventional floating-point numbers for high-dynamic-range applications.
comment: 18 pages, 4 figures (main text). 14 pages, 21 figures (appendix). Code is at https://github.com/glassroom/generalized_orders_of_magnitude
♻ ☆ Rethinking Losses for Diffusion Bridge Samplers NeurIPS 2025
Diffusion bridges are a promising class of deep-learning methods for sampling from unnormalized distributions. Recent works show that the Log Variance (LV) loss consistently outperforms the reverse Kullback-Leibler (rKL) loss when using the reparametrization trick to compute rKL-gradients. While the on-policy LV loss yields identical gradients to the rKL loss when combined with the log-derivative trick for diffusion samplers with non-learnable forward processes, this equivalence does not hold for diffusion bridges or when diffusion coefficients are learned. Based on this insight we argue that for diffusion bridges the LV loss does not represent an optimization objective that can be motivated like the rKL loss via the data processing inequality. Our analysis shows that employing the rKL loss with the log-derivative trick (rKL-LD) does not only avoid these conceptual problems but also consistently outperforms the LV loss. Experimental results with different types of diffusion bridges on challenging benchmarks show that samplers trained with the rKL-LD loss achieve better performance. From a practical perspective we find that rKL-LD requires significantly less hyperparameter optimization and yields more stable training behavior.
comment: Accepted at NeurIPS 2025 as a Conference Paper
♻ ☆ TASP: Topology-aware Sequence Parallelism
Long-context large language models (LLMs) face constraints due to the quadratic complexity of the self-attention mechanism. The mainstream sequence parallelism (SP) method, Ring Attention, attempts to solve this by distributing the query into multiple query chunks across accelerators and enable each Q tensor to access all KV tensors from other accelerators via the Ring AllGather communication primitive. However, it exhibits low communication efficiency, restricting its practical applicability. This inefficiency stems from the mismatch between the Ring AllGather communication primitive it adopts and the AlltoAll topology of modern accelerators. A Ring AllGather primitive is composed of iterations of ring-styled data transfer, which can only utilize a very limited fraction of an AlltoAll topology. Inspired by the Hamiltonian decomposition of complete directed graphs, we identify that modern accelerator topology can be decomposed into multiple orthogonal ring datapaths which can concurrently transfer data without interference. Based on this, we further observe that the Ring AllGather primitive can also be decomposed into the same number of concurrent ring-styled data transfer at every iteration. Based on these insights, we propose TASP, a topology-aware SP method for long-context LLMs that fully utilizes the communication capacity of modern accelerators via topology decomposition and primitive decomposition. Experimental results on both single-node and multi-node NVIDIA H100 systems and a single-node AMD MI300X system demonstrate that TASP achieves higher communication efficiency than Ring Attention on these modern accelerator topologies and achieves up to 3.58 speedup than Ring Attention and its variant Zigzag-Ring Attention. The code is available at https://github.com/infinigence/HamiltonAttention.
♻ ☆ OneForecast: A Universal Framework for Global and Regional Weather Forecasting
Accurate weather forecasts are important for disaster prevention, agricultural planning, etc. Traditional numerical weather prediction (NWP) methods offer physically interpretable high-accuracy predictions but are computationally expensive and fail to fully leverage rapidly growing historical data. In recent years, deep learning models have made significant progress in weather forecasting, but challenges remain, such as balancing global and regional high-resolution forecasts, excessive smoothing in extreme event predictions, and insufficient dynamic system modeling. To address these issues, this paper proposes a global-regional nested weather forecasting framework (OneForecast) based on graph neural networks. By combining a dynamic system perspective with multi-grid theory, we construct a multi-scale graph structure and densify the target region to capture local high-frequency features. We introduce an adaptive messaging mechanism, using dynamic gating units to deeply integrate node and edge features for more accurate extreme event forecasting. For high-resolution regional forecasts, we propose a neural nested grid method to mitigate boundary information loss. Experimental results show that OneForecast performs excellently across global to regional scales and short-term to long-term forecasts, especially in extreme event predictions. Codes link https://github.com/YuanGao-YG/OneForecast.
♻ ☆ Multi-Trigger Poisoning Amplifies Backdoor Vulnerabilities in LLMs
Recent studies have shown that Large Language Models (LLMs) are vulnerable to data poisoning attacks, where malicious training examples embed hidden behaviours triggered by specific input patterns. However, most existing works assume a phrase and focus on the attack's effectiveness, offering limited understanding of trigger mechanisms and how multiple triggers interact within the model. In this paper, we present a framework for studying poisoning in LLMs. We show that multiple distinct backdoor triggers can coexist within a single model without interfering with each other, enabling adversaries to embed several triggers concurrently. Using multiple triggers with high embedding similarity, we demonstrate that poisoned triggers can achieve robust activation even when tokens are substituted or separated by long token spans. Our findings expose a broader and more persistent vulnerability surface in LLMs. To mitigate this threat, we propose a post hoc recovery method that selectively retrains specific model components based on a layer-wise weight difference analysis. Our method effectively removes the trigger behaviour with minimal parameter updates, presenting a practical and efficient defence against multi-trigger poisoning.
♻ ☆ DACP: Domain-Adaptive Continual Pre-Training of Large Language Models for Phone Conversation Summarization EMNLP 2025
Large language models (LLMs) have achieved impressive performance in text summarization, yet their performance often falls short when applied to specialized domains that differ from their original pre-training distribution. While fine-tuning can improve summarization quality, it typically relies on costly and scarce high-quality labeled data. In this work, we explore continual pre-training as a scalable, self-supervised approach to adapt LLMs for downstream summarization tasks, particularly in the context of noisy real-world conversation transcripts. We conduct extensive experiments using large-scale, unlabeled business conversation data to investigate whether continual pre-training enhances model capabilities in conversational summarization. Our results demonstrate that continual pre-training yields substantial gains in both in-domain and out-of-domain summarization benchmarks, while maintaining strong generalization and robustness. We also analyze the effects of data selection strategies, providing practical guidelines for applying continual pre-training in summarization-focused industrial applications.
comment: Accepted to the NewSumm Workshop at EMNLP 2025. Equal contribution from the first four authors
♻ ☆ OpenRLHF: An Easy-to-use, Scalable and High-performance RLHF Framework
Large Language Models (LLMs) fine-tuned via Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning with Verifiable Rewards (RLVR) significantly improve the alignment of human-AI values, further raising the upper bound of AI capabilities, particularly in reasoning-intensive, long-context Chain-of-Thought (CoT) tasks. However, existing frameworks commonly face challenges such as inference bottlenecks and complexity barriers, which restrict their accessibility to newcomers. To bridge this gap, we introduce \textbf{OpenRLHF}, a user-friendly, scalable, and easy-to-learn open-source RLHF framework built upon Ray, vLLM, DeepSpeed, and HuggingFace Transformers, featuring a simplified design, clear code structure, and comprehensive documentation to facilitate entry for researchers and practitioners. Experimental results show that OpenRLHF achieves superior training efficiency, with speedups ranging from 1.22x to 1.68x across different model sizes, compared to state-of-the-art frameworks. Additionally, it requires significantly fewer lines of code for implementation. OpenRLHF is publicly available at https://github.com/OpenRLHF/OpenRLHF, and has already been adopted by leading institutions to accelerate RLHF research and learning.
comment: update template
♻ ☆ Multi-Continental Healthcare Modelling Using Blockchain-Enabled Federated Learning
One of the biggest challenges of building artificial intelligence (AI) model in the healthcare area is the data sharing. Since healthcare data is private, sensitive, and heterogeneous, collecting sufficient data for modelling is exhausting, costly, and sometimes impossible. In this paper, we propose a framework for global healthcare modelling using datasets from multi-continents (Europe, North America, and Asia) without sharing the local datasets, and choose glucose management as a study model to verify its effectiveness. Technically, blockchain-enabled federated learning is implemented with adaptation to meet the privacy and safety requirements of healthcare data, meanwhile, it rewards honest participation and penalizes malicious activities using its on-chain incentive mechanism. Experimental results show that the proposed framework is effective, efficient, and privacy-preserving. Its prediction accuracy consistently outperforms models trained on limited personal data and achieves comparable or even slightly better results than centralized training in certain scenarios, all while preserving data privacy. This work paves the way for international collaborations on healthcare projects, where additional data is crucial for reducing bias and providing benefits to humanity.
comment: Accepted by IEEE Global Blockchain Conference, 2025
♻ ☆ Distilling a Small Utility-Based Passage Selector to Enhance Retrieval-Augmented Generation SIGIR
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating retrieved information. Standard retrieval process prioritized relevance, focusing on topical alignment between queries and passages. In contrast, in RAG, the emphasis has shifted to utility, which considers the usefulness of passages for generating accurate answers. Despite empirical evidence showing the benefits of utility-based retrieval in RAG, the high computational cost of using LLMs for utility judgments limits the number of passages evaluated. This restriction is problematic for complex queries requiring extensive information. To address this, we propose a method to distill the utility judgment capabilities of LLMs into smaller, more efficient models. Our approach focuses on utility-based selection rather than ranking, enabling dynamic passage selection tailored to specific queries without the need for fixed thresholds. We train student models to learn pseudo-answer generation and utility judgments from teacher LLMs, using a sliding window method that dynamically selects useful passages. Our experiments demonstrate that utility-based selection provides a flexible and cost-effective solution for RAG, significantly reducing computational costs while improving answer quality. We present the distillation results using Qwen3-32B as the teacher model for both relevance ranking and utility-based selection, distilled into RankQwen1.7B and UtilityQwen1.7B. Our findings indicate that for complex questions, utility-based selection is more effective than relevance ranking in enhancing answer generation performance. We will release the relevance ranking and utility-based selection annotations for the MS MARCO dataset, supporting further research in this area.
comment: Accepted by SIGIR-AP25
♻ ☆ TokenSelect: Efficient Long-Context Inference and Length Extrapolation for LLMs via Dynamic Token-Level KV Cache Selection EMNLP2025
Rapid advances in Large Language Models (LLMs) have spurred demand for processing extended context sequences in contemporary applications. However, this progress faces two challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues limit LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using QK dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a few critical KV cache tokens in attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we design the Selection Cache based on observations of consecutive Query similarity and implemented the efficient Paged Dot Product Kernel, significantly reducing the selection overhead. A comprehensive evaluation of TokenSelect demonstrates up to $23.84\times$ speedup in attention computation and up to $2.28\times$ acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.
comment: Accepted by EMNLP2025
♻ ☆ Panorama: Fast-Track Nearest Neighbors
Approximate Nearest-Neighbor Search (ANNS) efficiently finds data items whose embeddings are close to that of a given query in a high-dimensional space, aiming to balance accuracy with speed. Used in recommendation systems, image and video retrieval, natural language processing, and retrieval-augmented generation (RAG), ANNS algorithms such as IVFPQ, HNSW graphs, Annoy, and MRPT utilize graph, tree, clustering, and quantization techniques to navigate large vector spaces. Despite this progress, ANNS systems spend up to 99\% of query time to compute distances in their final refinement phase. In this paper, we present PANORAMA, a machine learning-driven approach that tackles the ANNS verification bottleneck through data-adaptive learned orthogonal transforms that facilitate the accretive refinement of distance bounds. Such transforms compact over 90\% of signal energy into the first half of dimensions, enabling early candidate pruning with partial distance computations. We integrate PANORAMA into state-of-the-art ANNS methods, namely IVFPQ/Flat, HNSW, MRPT, and Annoy, without index modification, using level-major memory layouts, SIMD-vectorized partial distance computations, and cache-aware access patterns. Experiments across diverse datasets -- from image-based CIFAR-10 and GIST to modern embedding spaces including OpenAI's Ada 2 and Large 3 -- demonstrate that PANORAMA affords a 2--30$\times$ end-to-end speedup with no recall loss.
♻ ☆ Learning Equilibria from Data: Provably Efficient Multi-Agent Imitation Learning
This paper provides the first expert sample complexity characterization for learning a Nash equilibrium from expert data in Markov Games. We show that a new quantity named the single policy deviation concentrability coefficient is unavoidable in the non-interactive imitation learning setting, and we provide an upper bound for behavioral cloning (BC) featuring such coefficient. BC exhibits substantial regret in games with high concentrability coefficient, leading us to utilize expert queries to develop and introduce two novel solution algorithms: MAIL-BRO and MURMAIL. The former employs a best response oracle and learns an $\varepsilon$-Nash equilibrium with $\mathcal{O}(\varepsilon^{-4})$ expert and oracle queries. The latter bypasses completely the best response oracle at the cost of a worse expert query complexity of order $\mathcal{O}(\varepsilon^{-8})$. Finally, we provide numerical evidence, confirming our theoretical findings.
♻ ☆ Self-Improving Skill Learning for Robust Skill-based Meta-Reinforcement Learning ICLR 2026
Meta-reinforcement learning (Meta-RL) facilitates rapid adaptation to unseen tasks but faces challenges in long-horizon environments. Skill-based approaches tackle this by decomposing state-action sequences into reusable skills and employing hierarchical decision-making. However, these methods are highly susceptible to noisy offline demonstrations, leading to unstable skill learning and degraded performance. To address this, we propose Self-Improving Skill Learning (SISL), which performs self-guided skill refinement using decoupled high-level and skill improvement policies, while applying skill prioritization via maximum return relabeling to focus updates on task-relevant trajectories, resulting in robust and stable adaptation even under noisy and suboptimal data. By mitigating the effect of noise, SISL achieves reliable skill learning and consistently outperforms other skill-based meta-RL methods on diverse long-horizon tasks.
comment: 9 pages main, 25 pages appendix with reference. Submitted to ICLR 2026
Multimedia 8
☆ Improving Temporal Understanding Logic Consistency in Video-Language Models via Attention Enhancement
Large language models (LLMs) often generate self-contradictory outputs, which severely impacts their reliability and hinders their adoption in practical applications. In video-language models (Video-LLMs), this phenomenon recently draws the attention of researchers. Specifically, these models fail to provide logically consistent responses to rephrased questions based on their grounding outputs. However, the underlying causes of this phenomenon remain underexplored. In this work, we adopt an interpretability-driven approach to analyze, statistically summarize, and intervention the potential factors of the phenomenon. We find that one of the primary reasons for the inconsistency in responses lies in the inability of cross-modal attention heads to effectively distinguish video tokens across different timestamps. To address this, we propose an attention enhancement method called Temporally Conditioned Attention Sharpening (TCAS), which constructs an enhancement objective based on attention distinctions to enhance the model's temporal resolution capability, thereby improving its temporal understanding logic consistency. Experimental results demonstrate that our method significantly enhances the temporal logic consistency of Video-LLMs. Further interpretability analyses reveal that our method indeed improves the temporal discriminability of attention heads, validating our conclusions. Additionally, our method achieves performance improvements in general video temporal grounding tasks, highlighting that temporal logic consistency is a bottleneck in temporal understanding. By enhancing consistency, our method drives significant progress in video temporal understanding.
☆ Personality-Enhanced Multimodal Depression Detection in the Elderly
This paper presents our solution to the Multimodal Personality-aware Depression Detection (MPDD) challenge at ACM MM 2025. We propose a multimodal depression detection model in the Elderly that incorporates personality characteristics. We introduce a multi-feature fusion approach based on a co-attention mechanism to effectively integrate LLDs, MFCCs, and Wav2Vec features in the audio modality. For the video modality, we combine representations extracted from OpenFace, ResNet, and DenseNet to construct a comprehensive visual feature set. Recognizing the critical role of personality in depression detection, we design an interaction module that captures the relationships between personality traits and multimodal features. Experimental results from the MPDD Elderly Depression Detection track demonstrate that our method significantly enhances performance, providing valuable insights for future research in multimodal depression detection among elderly populations.
comment: 6 pages,2 figures,accepted by ACM Multimedia Asia 2025
☆ TTOM: Test-Time Optimization and Memorization for Compositional Video Generation
Video Foundation Models (VFMs) exhibit remarkable visual generation performance, but struggle in compositional scenarios (e.g., motion, numeracy, and spatial relation). In this work, we introduce Test-Time Optimization and Memorization (TTOM), a training-free framework that aligns VFM outputs with spatiotemporal layouts during inference for better text-image alignment. Rather than direct intervention to latents or attention per-sample in existing work, we integrate and optimize new parameters guided by a general layout-attention objective. Furthermore, we formulate video generation within a streaming setting, and maintain historical optimization contexts with a parametric memory mechanism that supports flexible operations, such as insert, read, update, and delete. Notably, we found that TTOM disentangles compositional world knowledge, showing powerful transferability and generalization. Experimental results on the T2V-CompBench and Vbench benchmarks establish TTOM as an effective, practical, scalable, and efficient framework to achieve cross-modal alignment for compositional video generation on the fly.
comment: Project page: https://ttom-t2v.github.io/
☆ SatFusion: A Unified Framework for Enhancing Satellite IoT Images via Multi-Temporal and Multi-Source Data Fusion
With the rapid advancement of the digital society, the proliferation of satellites in the Satellite Internet of Things (Sat-IoT) has led to the continuous accumulation of large-scale multi-temporal and multi-source images across diverse application scenarios. However, existing methods fail to fully exploit the complementary information embedded in both temporal and source dimensions. For example, Multi-Image Super-Resolution (MISR) enhances reconstruction quality by leveraging temporal complementarity across multiple observations, yet the limited fine-grained texture details in input images constrain its performance. Conversely, pansharpening integrates multi-source images by injecting high-frequency spatial information from panchromatic data, but typically relies on pre-interpolated low-resolution inputs and assumes noise-free alignment, making it highly sensitive to noise and misregistration. To address these issues, we propose SatFusion: A Unified Framework for Enhancing Satellite IoT Images via Multi-Temporal and Multi-Source Data Fusion. Specifically, SatFusion first employs a Multi-Temporal Image Fusion (MTIF) module to achieve deep feature alignment with the panchromatic image. Then, a Multi-Source Image Fusion (MSIF) module injects fine-grained texture information from the panchromatic data. Finally, a Fusion Composition module adaptively integrates the complementary advantages of both modalities while dynamically refining spectral consistency, supervised by a weighted combination of multiple loss functions. Extensive experiments on the WorldStrat, WV3, QB, and GF2 datasets demonstrate that SatFusion significantly improves fusion quality, robustness under challenging conditions, and generalizability to real-world Sat-IoT scenarios. The code is available at: https://github.com/dllgyufei/SatFusion.git.
☆ IsoSignVid2Aud: Sign Language Video to Audio Conversion without Text Intermediaries
Sign language to spoken language audio translation is important to connect the hearing- and speech-challenged humans with others. We consider sign language videos with isolated sign sequences rather than continuous grammatical signing. Such videos are useful in educational applications and sign prompt interfaces. Towards this, we propose IsoSignVid2Aud, a novel end-to-end framework that translates sign language videos with a sequence of possibly non-grammatic continuous signs to speech without requiring intermediate text representation, providing immediate communication benefits while avoiding the latency and cascading errors inherent in multi-stage translation systems. Our approach combines an I3D-based feature extraction module with a specialized feature transformation network and an audio generation pipeline, utilizing a novel Non-Maximal Suppression (NMS) algorithm for the temporal detection of signs in non-grammatic continuous sequences. Experimental results demonstrate competitive performance on ASL-Citizen-1500 and WLASL-100 datasets with Top-1 accuracies of 72.01\% and 78.67\%, respectively, and audio quality metrics (PESQ: 2.67, STOI: 0.73) indicating intelligible speech output. Code is available at: https://github.com/BheeshmSharma/IsoSignVid2Aud_AIMLsystems-2025.
comment: Accepted in AIML-Systems-2025
☆ Reinforcement Learning-Driven Edge Management for Reliable Multi-view 3D Reconstruction
Real-time multi-view 3D reconstruction is a mission-critical application for key edge-native use cases, such as fire rescue, where timely and accurate 3D scene modeling enables situational awareness and informed decision-making. However, the dynamic and unpredictable nature of edge resource availability introduces disruptions, such as degraded image quality, unstable network links, and fluctuating server loads, which challenge the reliability of the reconstruction pipeline. In this work, we present a reinforcement learning (RL)-based edge resource management framework for reliable 3D reconstruction to ensure high quality reconstruction within a reasonable amount of time, despite the system operating under a resource-constrained and disruption-prone environment. In particular, the framework adopts two cooperative Q-learning agents, one for camera selection and one for server selection, both of which operate entirely online, learning policies through interactions with the edge environment. To support learning under realistic constraints and evaluate system performance, we implement a distributed testbed comprising lab-hosted end devices and FABRIC infrastructure-hosted edge servers to emulate smart city edge infrastructure under realistic disruption scenarios. Results show that the proposed framework improves application reliability by effectively balancing end-to-end latency and reconstruction quality in dynamic environments.
♻ ☆ Paper2Video: Automatic Video Generation from Scientific Papers
Academic presentation videos have become an essential medium for research communication, yet producing them remains highly labor-intensive, often requiring hours of slide design, recording, and editing for a short 2 to 10 minutes video. Unlike natural video, presentation video generation involves distinctive challenges: inputs from research papers, dense multi-modal information (text, figures, tables), and the need to coordinate multiple aligned channels such as slides, subtitles, speech, and human talker. To address these challenges, we introduce Paper2Video, the first benchmark of 101 research papers paired with author-created presentation videos, slides, and speaker metadata. We further design four tailored evaluation metrics--Meta Similarity, PresentArena, PresentQuiz, and IP Memory--to measure how videos convey the paper's information to the audience. Building on this foundation, we propose PaperTalker, the first multi-agent framework for academic presentation video generation. It integrates slide generation with effective layout refinement by a novel effective tree search visual choice, cursor grounding, subtitling, speech synthesis, and talking-head rendering, while parallelizing slide-wise generation for efficiency. Experiments on Paper2Video demonstrate that the presentation videos produced by our approach are more faithful and informative than existing baselines, establishing a practical step toward automated and ready-to-use academic video generation. Our dataset, agent, and code are available at https://github.com/showlab/Paper2Video.
comment: Project Page: https://showlab.github.io/Paper2Video/
♻ ☆ PRVR: Partially Relevant Video Retrieval
In current text-to-video retrieval (T2VR), videos to be retrieved have been properly trimmed so that a correspondence between the videos and ad-hoc textual queries naturally exists. Note in practice that videos circulated on the Internet and social media platforms, while being relatively short, are typically rich in their content. Often, multiple scenes / actions / events are shown in a single video, leading to a more challenging T2VR setting wherein only part of the video content is relevant w.r.t. a given query. This paper presents a first study on this setting which we term Partially Relevant Video Retrieval (PRVR). Considering that a video typically consists of multiple moments, a video is regarded as partially relevant w.r.t. to a given query if it contains a query-related moment. We formulate the PRVR task as a multiple instance learning problem, and propose a Multi-Scale Similarity Learning (MS-SL++) network that jointly learns both clip-scale and frame-scale similarities to determine the partial relevance between video-query pairs. Extensive experiments on three diverse video-text datasets (TVshow Retrieval, ActivityNet-Captions and Charades-STA) demonstrate the viability of the proposed method.
comment: Accepted by TPAMI. The paper's homepage is https://github.com/HuiGuanLab/ms-sl-pp
Computation and Language 150
☆ Artificial Hippocampus Networks for Efficient Long-Context Modeling
Long-sequence modeling faces a fundamental trade-off between the efficiency of compressive fixed-size memory in RNN-like models and the fidelity of lossless growing memory in attention-based Transformers. Inspired by the Multi-Store Model in cognitive science, we introduce a memory framework of artificial neural networks. Our method maintains a sliding window of the Transformer's KV cache as lossless short-term memory, while a learnable module termed Artificial Hippocampus Network (AHN) recurrently compresses out-of-window information into a fixed-size compact long-term memory. To validate this framework, we instantiate AHNs using modern RNN-like architectures, including Mamba2, DeltaNet, and Gated DeltaNet. Extensive experiments on long-context benchmarks LV-Eval and InfiniteBench demonstrate that AHN-augmented models consistently outperform sliding window baselines and achieve performance comparable or even superior to full-attention models, while substantially reducing computational and memory requirements. For instance, augmenting the Qwen2.5-3B-Instruct with AHNs reduces inference FLOPs by 40.5% and memory cache by 74.0%, while improving its average score on LV-Eval (128k sequence length) from 4.41 to 5.88. Code is available at: https://github.com/ByteDance-Seed/AHN.
comment: Code: https://github.com/ByteDance-Seed/AHN
☆ Vibe Checker: Aligning Code Evaluation with Human Preference
Large Language Models (LLMs) have catalyzed vibe coding, where users leverage LLMs to generate and iteratively refine code through natural language interactions until it passes their vibe check. Vibe check is tied to real-world human preference and goes beyond functionality: the solution should feel right, read cleanly, preserve intent, and remain correct. However, current code evaluation remains anchored to pass@k and captures only functional correctness, overlooking the non-functional instructions that users routinely apply. In this paper, we hypothesize that instruction following is the missing piece underlying vibe check that represents human preference in coding besides functional correctness. To quantify models' code instruction following capabilities with measurable signals, we present VeriCode, a taxonomy of 30 verifiable code instructions together with corresponding deterministic verifiers. We use the taxonomy to augment established evaluation suites, resulting in Vibe Checker, a testbed to assess both code instruction following and functional correctness. Upon evaluating 31 leading LLMs, we show that even the strongest models struggle to comply with multiple instructions and exhibit clear functional regression. Most importantly, a composite score of functional correctness and instruction following correlates the best with human preference, with the latter emerging as the primary differentiator on real-world programming tasks. Our work identifies core factors of the vibe check, providing a concrete path for benchmarking and developing models that better align with user preferences in coding.
comment: Preprint
Agent Bain vs. Agent McKinsey: A New Text-to-SQL Benchmark for the Business Domain ACL
In the business domain, where data-driven decision making is crucial, text-to-SQL is fundamental for easy natural language access to structured data. While recent LLMs have achieved strong performance in code generation, existing text-to-SQL benchmarks remain focused on factual retrieval of past records. We introduce CORGI, a new benchmark specifically designed for real-world business contexts. CORGI is composed of synthetic databases inspired by enterprises such as Doordash, Airbnb, and Lululemon. It provides questions across four increasingly complex categories of business queries: descriptive, explanatory, predictive, and recommendational. This challenge calls for causal reasoning, temporal forecasting, and strategic recommendation, reflecting multi-level and multi-step agentic intelligence. We find that LLM performance drops on high-level questions, struggling to make accurate predictions and offer actionable plans. Based on execution success rate, the CORGI benchmark is about 21\% more difficult than the BIRD benchmark. This highlights the gap between popular LLMs and the need for real-world business intelligence. We release a public dataset and evaluation framework, and a website for public submissions.
comment: 20 pages, 6 figures, under review for ACL ARR
☆ Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning
Large Reasoning Models (LRMs) have achieved remarkable performance on complex reasoning tasks by adopting the "think-then-answer" paradigm, which enhances both accuracy and interpretability. However, current LRMs exhibit two critical limitations when processing non-English languages: (1) They often struggle to maintain input-output language consistency; (2) They generally perform poorly with wrong reasoning paths and lower answer accuracy compared to English. These limitations significantly degrade the user experience for non-English speakers and hinder the global deployment of LRMs. To address these limitations, we propose M-Thinker, which is trained by the GRPO algorithm that involves a Language Consistency (LC) reward and a novel Cross-lingual Thinking Alignment (CTA) reward. Specifically, the LC reward defines a strict constraint on the language consistency between the input, thought, and answer. Besides, the CTA reward compares the model's non-English reasoning paths with its English reasoning path to transfer its own reasoning capability from English to non-English languages. Through an iterative RL procedure, our M-Thinker-1.5B/7B models not only achieve nearly 100% language consistency and superior performance on two multilingual benchmarks (MMATH and PolyMath), but also exhibit excellent generalization on out-of-domain languages.
comment: 13 pages, 8 tables, 4 figures
☆ AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs
Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.
comment: 26 pages, 23 figures, the code is available at \url{https://github.com/DabDans/AudioMarathon}
On the Convergence of Moral Self-Correction in Large Language Models
Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only a general and abstract goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. Focusing on moral self-correction in LLMs, we reveal a key characteristic of intrinsic self-correction: performance convergence through multi-round interactions; and provide a mechanistic analysis of this convergence behavior. Based on our experimental results and analysis, we uncover the underlying mechanism of convergence: consistently injected self-correction instructions activate moral concepts that reduce model uncertainty, leading to converged performance as the activated moral concepts stabilize over successive rounds. This paper demonstrates the strong potential of moral self-correction by showing that it exhibits a desirable property of converged performance.
comment: 19pages, 7 figures
☆ Online Rubrics Elicitation from Pairwise Comparisons
Rubrics provide a flexible way to train LLMs on open-ended long-form answers where verifiable rewards are not applicable and human preferences provide coarse signals. Prior work shows that reinforcement learning with rubric-based rewards leads to consistent gains in LLM post-training. Most existing approaches rely on rubrics that remain static over the course of training. Such static rubrics, however, are vulnerable to reward-hacking type behaviors and fail to capture emergent desiderata that arise during training. We introduce Online Rubrics Elicitation (OnlineRubrics), a method that dynamically curates evaluation criteria in an online manner through pairwise comparisons of responses from current and reference policies. This online process enables continuous identification and mitigation of errors as training proceeds. Empirically, this approach yields consistent improvements of up to 8% over training exclusively with static rubrics across AlpacaEval, GPQA, ArenaHard as well as the validation sets of expert questions and rubrics. We qualitatively analyze the elicited criteria and identify prominent themes such as transparency, practicality, organization, and reasoning.
☆ Don't Adapt Small Language Models for Tools; Adapt Tool Schemas to the Models
Small language models (SLMs) offer significant computational advantages for tool-augmented AI systems, yet they struggle with tool-use tasks, particularly in selecting appropriate tools and identifying correct parameters. A common failure mode is schema misalignment: models hallucinate plausible but non-existent tool names that reflect naming conventions internalized during pretraining but absent from the provided tool schema. Rather than forcing models to adapt to arbitrary schemas, we propose adapting schemas to align with models' pretrained knowledge. We introduce PA-Tool (Pretraining-Aligned Tool Schema Generation), a training-free method that leverages peakedness-a signal from contamination detection indicating pretraining familiarity-to automatically rename tool components. By generating multiple candidates and selecting those with highest output concentration across samples, PA-Tool identifies pretrain-aligned naming patterns. Experiments on MetaTool and RoTBench show improvements of up to 17% points, with schema misalignment errors reduced by 80%. PA-Tool enables small models to approach state-of-the-art performance while maintaining computational efficiency for adaptation to new tools without retraining. Our work demonstrates that schema-level interventions can unlock the tool-use potential of resource-efficient models by adapting schemas to models rather than models to schemas.
comment: 15 pages, 4 figures
☆ LeMAJ (Legal LLM-as-a-Judge): Bridging Legal Reasoning and LLM Evaluation EMNLP
Evaluating large language model (LLM) outputs in the legal domain presents unique challenges due to the complex and nuanced nature of legal analysis. Current evaluation approaches either depend on reference data, which is costly to produce, or use standardized assessment methods, both of which have significant limitations for legal applications. Although LLM-as-a-Judge has emerged as a promising evaluation technique, its reliability and effectiveness in legal contexts depend heavily on evaluation processes unique to the legal industry and how trustworthy the evaluation appears to the human legal expert. This is where existing evaluation methods currently fail and exhibit considerable variability. This paper aims to close the gap: a) we break down lengthy responses into 'Legal Data Points' (LDPs), self-contained units of information, and introduce a novel, reference-free evaluation methodology that reflects how lawyers evaluate legal answers; b) we demonstrate that our method outperforms a variety of baselines on both our proprietary dataset and an open-source dataset (LegalBench); c) we show how our method correlates more closely with human expert evaluations and helps improve inter-annotator agreement; and finally d) we open source our Legal Data Points for a subset of LegalBench used in our experiments, allowing the research community to replicate our results and advance research in this vital area of LLM evaluation on legal question-answering.
comment: Published in Natural Legal Language Processing - EMNLP Workshop 2025
☆ Hybrid Reinforcement: When Reward Is Sparse, It's Better to Be Dense
Post-training for reasoning of large language models (LLMs) increasingly relies on verifiable rewards: deterministic checkers that provide 0-1 correctness signals. While reliable, such binary feedback is brittle--many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning.
comment: 20 pages
☆ Red-Bandit: Test-Time Adaptation for LLM Red-Teaming via Bandit-Guided LoRA Experts
Automated red-teaming has emerged as a scalable approach for auditing Large Language Models (LLMs) prior to deployment, yet existing approaches lack mechanisms to efficiently adapt to model-specific vulnerabilities at inference. We introduce Red-Bandit, a red-teaming framework that adapts online to identify and exploit model failure modes under distinct attack styles (e.g., manipulation, slang). Red-Bandit post-trains a set of parameter-efficient LoRA experts, each specialized for a particular attack style, using reinforcement learning that rewards the generation of unsafe prompts via a rule-based safety model. At inference, a multi-armed bandit policy dynamically selects among these attack-style experts based on the target model's response safety, balancing exploration and exploitation. Red-Bandit achieves state-of-the-art results on AdvBench under sufficient exploration (ASR@10), while producing more human-readable prompts (lower perplexity). Moreover, Red-Bandit's bandit policy serves as a diagnostic tool for uncovering model-specific vulnerabilities by indicating which attack styles most effectively elicit unsafe behaviors.
☆ When Benchmarks Age: Temporal Misalignment through Large Language Model Factuality Evaluation
The rapid evolution of large language models (LLMs) and the real world has outpaced the static nature of widely used evaluation benchmarks, raising concerns about their reliability for evaluating LLM factuality. While substantial works continue to rely on the popular but old benchmarks, their temporal misalignment with real-world facts and modern LLMs, and their effects on LLM factuality evaluation remain underexplored. Therefore, in this work, we present a systematic investigation of this issue by examining five popular factuality benchmarks and eight LLMs released across different years. An up-to-date fact retrieval pipeline and three metrics are tailored to quantify benchmark aging and its impact on LLM factuality evaluation. Experimental results and analysis illustrate that a considerable portion of samples in the widely used factuality benchmarks are outdated, leading to unreliable assessments of LLM factuality. We hope our work can provide a testbed to assess the reliability of a benchmark for LLM factuality evaluation and inspire more research on the benchmark aging issue. Codes are available in https://github.com/JiangXunyi/BenchAge.
☆ LAD-RAG: Layout-aware Dynamic RAG for Visually-Rich Document Understanding
Question answering over visually rich documents (VRDs) requires reasoning not only over isolated content but also over documents' structural organization and cross-page dependencies. However, conventional retrieval-augmented generation (RAG) methods encode content in isolated chunks during ingestion, losing structural and cross-page dependencies, and retrieve a fixed number of pages at inference, regardless of the specific demands of the question or context. This often results in incomplete evidence retrieval and degraded answer quality for multi-page reasoning tasks. To address these limitations, we propose LAD-RAG, a novel Layout-Aware Dynamic RAG framework. During ingestion, LAD-RAG constructs a symbolic document graph that captures layout structure and cross-page dependencies, adding it alongside standard neural embeddings to yield a more holistic representation of the document. During inference, an LLM agent dynamically interacts with the neural and symbolic indices to adaptively retrieve the necessary evidence based on the query. Experiments on MMLongBench-Doc, LongDocURL, DUDE, and MP-DocVQA demonstrate that LAD-RAG improves retrieval, achieving over 90% perfect recall on average without any top-k tuning, and outperforming baseline retrievers by up to 20% in recall at comparable noise levels, yielding higher QA accuracy with minimal latency.
☆ Benchmarking LLM Causal Reasoning with Scientifically Validated Relationships
Causal reasoning is fundamental for Large Language Models (LLMs) to understand genuine cause-and-effect relationships beyond pattern matching. Existing benchmarks suffer from critical limitations such as reliance on synthetic data and narrow domain coverage. We introduce a novel benchmark constructed from casually identified relationships extracted from top-tier economics and finance journals, drawing on rigorous methodologies including instrumental variables, difference-in-differences, and regression discontinuity designs. Our benchmark comprises 40,379 evaluation items covering five task types across domains such as health, environment, technology, law, and culture. Experimental results on eight state-of-the-art LLMs reveal substantial limitations, with the best model achieving only 57.6\% accuracy. Moreover, model scale does not consistently translate to superior performance, and even advanced reasoning models struggle with fundamental causal relationship identification. These findings underscore a critical gap between current LLM capabilities and demands of reliable causal reasoning in high-stakes applications.
☆ Customer-R1: Personalized Simulation of Human Behaviors via RL-based LLM Agent in Online Shopping
Simulating step-wise human behavior with Large Language Models (LLMs) has become an emerging research direction, enabling applications in various practical domains. While prior methods, including prompting, supervised fine-tuning (SFT), and reinforcement learning (RL), have shown promise in modeling step-wise behavior, they primarily learn a population-level policy without conditioning on a user's persona, yielding generic rather than personalized simulations. In this work, we pose a critical question: how can LLM agents better simulate personalized user behavior? We introduce Customer-R1, an RL-based method for personalized, step-wise user behavior simulation in online shopping environments. Our policy is conditioned on an explicit persona, and we optimize next-step rationale and action generation via action correctness reward signals. Experiments on the OPeRA dataset emonstrate that Customer-R1 not only significantly outperforms prompting and SFT-based baselines in next-action prediction tasks, but also better matches users' action distribution, indicating higher fidelity in personalized behavior simulation.
☆ Where to Begin: Efficient Pretraining via Subnetwork Selection and Distillation
Small Language models (SLMs) offer an efficient and accessible alternative to Large Language Models (LLMs), delivering strong performance while using far fewer resources. We introduce a simple and effective framework for pretraining SLMs that brings together three complementary ideas. First, we identify structurally sparse sub-network initializations that consistently outperform randomly initialized models of similar size under the same compute budget. Second, we use evolutionary search to automatically discover high-quality sub-network initializations, providing better starting points for pretraining. Third, we apply knowledge distillation from larger teacher models to speed up training and improve generalization. Together, these components make SLM pretraining substantially more efficient: our best model, discovered using evolutionary search and initialized with LLM weights, matches the validation perplexity of a comparable Pythia SLM while requiring 9.2x fewer pretraining tokens. We release all code and models at https://github.com/whittle-org/whittle/, offering a practical and reproducible path toward cost-efficient small language model development at scale.
☆ Machines in the Crowd? Measuring the Footprint of Machine-Generated Text on Reddit
Generative Artificial Intelligence is reshaping online communication by enabling large-scale production of Machine-Generated Text (MGT) at low cost. While its presence is rapidly growing across the Web, little is known about how MGT integrates into social media environments. In this paper, we present the first large-scale characterization of MGT on Reddit. Using a state-of-the-art statistical method for detection of MGT, we analyze over two years of activity (2022-2024) across 51 subreddits representative of Reddit's main community types such as information seeking, social support, and discussion. We study the concentration of MGT across communities and over time, and compared MGT to human-authored text in terms of social signals it expresses and engagement it receives. Our very conservative estimate of MGT prevalence indicates that synthetic text is marginally present on Reddit, but it can reach peaks of up to 9% in some communities in some months. MGT is unevenly distributed across communities, more prevalent in subreddits focused on technical knowledge and social support, and often concentrated in the activity of a small fraction of users. MGT also conveys distinct social signals of warmth and status giving typical of language of AI assistants. Despite these stylistic differences, MGT achieves engagement levels comparable than human-authored content and in a few cases even higher, suggesting that AI-generated text is becoming an organic component of online social discourse. This work offers the first perspective on the MGT footprint on Reddit, paving the way for new investigations involving platform governance, detection strategies, and community dynamics.
☆ How much speech data is necessary for ASR in African languages? An evaluation of data scaling in Kinyarwanda and Kikuyu
The development of Automatic Speech Recognition (ASR) systems for low-resource African languages remains challenging due to limited transcribed speech data. While recent advances in large multilingual models like OpenAI's Whisper offer promising pathways for low-resource ASR development, critical questions persist regarding practical deployment requirements. This paper addresses two fundamental concerns for practitioners: determining the minimum data volumes needed for viable performance and characterizing the primary failure modes that emerge in production systems. We evaluate Whisper's performance through comprehensive experiments on two Bantu languages: systematic data scaling analysis on Kinyarwanda using training sets from 1 to 1,400 hours, and detailed error characterization on Kikuyu using 270 hours of training data. Our scaling experiments demonstrate that practical ASR performance (WER < 13\%) becomes achievable with as little as 50 hours of training data, with substantial improvements continuing through 200 hours (WER < 10\%). Complementing these volume-focused findings, our error analysis reveals that data quality issues, particularly noisy ground truth transcriptions, account for 38.6\% of high-error cases, indicating that careful data curation is as critical as data volume for robust system performance. These results provide actionable benchmarks and deployment guidance for teams developing ASR systems across similar low-resource language contexts. We release accompanying and models see https://github.com/SunbirdAI/kinyarwanda-whisper-eval
☆ Language Lives in Sparse Dimensions: Toward Interpretable and Efficient Multilingual Control for Large Language Models
Large language models exhibit strong multilingual capabilities despite limited exposure to non-English data. Prior studies show that English-centric large language models map multilingual content into English-aligned representations at intermediate layers and then project them back into target-language token spaces in the final layer. From this observation, we hypothesize that this cross-lingual transition is governed by a small and sparse set of dimensions, which occur at consistent indices across the intermediate to final layers. Building on this insight, we introduce a simple, training-free method to identify and manipulate these dimensions, requiring only as few as 50 sentences of either parallel or monolingual data. Experiments on a multilingual generation control task reveal the interpretability of these dimensions, demonstrating that the interventions in these dimensions can switch the output language while preserving semantic content, and that it surpasses the performance of prior neuron-based approaches at a substantially lower cost.
comment: Work in progress. Our code will be available at: https://github.com/ku-nlp/language-specific-dimensions
☆ Sunflower: A New Approach To Expanding Coverage of African Languages in Large Language Models
There are more than 2000 living languages in Africa, most of which have been bypassed by advances in language technology. Current leading LLMs exhibit strong performance on a number of the most common languages (e.g. Swahili or Yoruba), but prioritise support for the languages with the most speakers first, resulting in piecemeal ability across disparate languages. We contend that a regionally focussed approach is more efficient, and present a case study for Uganda, a country with high linguistic diversity. We describe the development of Sunflower 14B and 32B, a pair of models based on Qwen 3 with state of the art comprehension in the majority of all Ugandan languages. These models are open source and can be used to reduce language barriers in a number of important practical applications.
☆ Biasless Language Models Learn Unnaturally: How LLMs Fail to Distinguish the Possible from the Impossible
Are large language models (LLMs) sensitive to the distinction between humanly possible languages and humanly impossible languages? This question is taken by many to bear on whether LLMs and humans share the same innate learning biases. Previous work has attempted to answer it in the positive by comparing LLM learning curves on existing language datasets and on "impossible" datasets derived from them via various perturbation functions. Using the same methodology, we examine this claim on a wider set of languages and impossible perturbations. We find that in most cases, GPT-2 learns each language and its impossible counterpart equally easily, in contrast to previous claims. We also apply a more lenient condition by testing whether GPT-2 provides any kind of separation between the whole set of natural languages and the whole set of impossible languages. By considering cross-linguistic variance in various metrics computed on the perplexity curves, we show that GPT-2 provides no systematic separation between the possible and the impossible. Taken together, these perspectives show that LLMs do not share the human innate biases that shape linguistic typology.
comment: 15 pages, 4 figures
☆ CARPAS: Towards Content-Aware Refinement of Provided Aspects for Summarization in Large Language Models
Aspect-based summarization has attracted significant attention for its ability to generate more fine-grained and user-aligned summaries. While most existing approaches assume a set of predefined aspects as input, real-world scenarios often present challenges where these given aspects may be incomplete, irrelevant, or entirely missing from the document. Users frequently expect systems to adaptively refine or filter the provided aspects based on the actual content. In this paper, we initiate this novel task setting, termed Content-Aware Refinement of Provided Aspects for Summarization (CARPAS), with the aim of dynamically adjusting the provided aspects based on the document context before summarizing. We construct three new datasets to facilitate our pilot experiments, and by using LLMs with four representative prompting strategies in this task, we find that LLMs tend to predict an overly comprehensive set of aspects, which often results in excessively long and misaligned summaries. Building on this observation, we propose a preliminary subtask to predict the number of relevant aspects, and demonstrate that the predicted number can serve as effective guidance for the LLMs, reducing the inference difficulty, and enabling them to focus on the most pertinent aspects. Our extensive experiments show that the proposed approach significantly improves performance across all datasets. Moreover, our deeper analyses uncover LLMs' compliance when the requested number of aspects differs from their own estimations, establishing a crucial insight for the deployment of LLMs in similar real-world applications.
comment: 22 pages, 17 figures
☆ Quantifying Data Contamination in Psychometric Evaluations of LLMs
Recent studies apply psychometric questionnaires to Large Language Models (LLMs) to assess high-level psychological constructs such as values, personality, moral foundations, and dark traits. Although prior work has raised concerns about possible data contamination from psychometric inventories, which may threaten the reliability of such evaluations, there has been no systematic attempt to quantify the extent of this contamination. To address this gap, we propose a framework to systematically measure data contamination in psychometric evaluations of LLMs, evaluating three aspects: (1) item memorization, (2) evaluation memorization, and (3) target score matching. Applying this framework to 21 models from major families and four widely used psychometric inventories, we provide evidence that popular inventories such as the Big Five Inventory (BFI-44) and Portrait Values Questionnaire (PVQ-40) exhibit strong contamination, where models not only memorize items but can also adjust their responses to achieve specific target scores.
comment: 12 pages, 1 figure
☆ NurseLLM: The First Specialized Language Model for Nursing EMNLP 2025
Recent advancements in large language models (LLMs) have significantly transformed medical systems. However, their potential within specialized domains such as nursing remains largely underexplored. In this work, we introduce NurseLLM, the first nursing-specialized LLM tailored for multiple choice question-answering (MCQ) tasks. We develop a multi-stage data generation pipeline to build the first large scale nursing MCQ dataset to train LLMs on a broad spectrum of nursing topics. We further introduce multiple nursing benchmarks to enable rigorous evaluation. Our extensive experiments demonstrate that NurseLLM outperforms SoTA general-purpose and medical-specialized LLMs of comparable size on different benchmarks, underscoring the importance of a specialized LLM for the nursing domain. Finally, we explore the role of reasoning and multi-agent collaboration systems in nursing, highlighting their promise for future research and applications.
comment: EMNLP 2025 Industry Track
☆ More Data or Better Data? A Critical Analysis of Data Selection and Synthesis for Mathematical Reasoning EMNLP 2025
The reasoning capabilities of Large Language Models (LLMs) play a critical role in many downstream tasks, yet depend strongly on the quality of training data. Despite various proposed data construction methods, their practical utility in real-world pipelines remains underexplored. In this work, we conduct a comprehensive analysis of open-source datasets and data synthesis techniques for mathematical reasoning, evaluating them under a unified pipeline designed to mirror training and deployment scenarios. We further distill effective data selection strategies and identify practical methods suitable for industrial applications. Our findings highlight that structuring data in more interpretable formats, or distilling from stronger models often outweighs simply scaling up data volume. This study provides actionable guidance for integrating training data to enhance LLM capabilities, supporting both cost-effective data curation and scalable model enhancement. We hope this work will inspire further research on how to balance "more data" versus "better data" for real-world reasoning tasks.
comment: 12 pages, 3 figures, submitted to EMNLP 2025 Industry Track
☆ Reasoning for Hierarchical Text Classification: The Case of Patents
Hierarchical text classification (HTC) assigns documents to multiple levels of a pre-defined taxonomy. Automated patent subject classification represents one of the hardest HTC scenarios because of domain knowledge difficulty and a huge number of labels. Prior approaches only output a flat label set, which offers little insight into the reason behind predictions. Therefore, we propose Reasoning for Hierarchical Classification (RHC), a novel framework that reformulates HTC as a step-by-step reasoning task to sequentially deduce hierarchical labels. RHC trains large language models (LLMs) in two stages: a cold-start stage that aligns outputs with chain-of-thought (CoT) reasoning format and a reinforcement learning (RL) stage to enhance multi-step reasoning ability. RHC demonstrates four advantages in our experiments. (1) Effectiveness: RHC surpasses previous baselines and outperforms the supervised fine-tuning counterparts by approximately 3% in accuracy and macro F1. (2) Explainability: RHC produces natural-language justifications before prediction to facilitate human inspection. (3) Scalability: RHC scales favorably with model size with larger gains compared to standard fine-tuning. (4) Applicability: Beyond patents, we further demonstrate that RHC achieves state-of-the-art performance on other widely used HTC benchmarks, which highlights its broad applicability.
comment: 15 pages, 10 tables, 3 figures
☆ A Multi-Agent Framework for Stateful Inference-Time Search
Recent work explores agentic inference-time techniques to perform structured, multi-step reasoning. However, stateless inference often struggles on multi-step tasks due to the absence of persistent state. Moreover, task-specific fine-tuning or instruction-tuning often achieve surface-level code generation but remain brittle on tasks requiring deeper reasoning and long-horizon dependencies. To address these limitations, we propose stateful multi-agent evolutionary search, a training-free framework that departs from prior stateless approaches by combining (i) persistent inference-time state, (ii) adversarial mutation, and (iii) evolutionary preservation. We demonstrate its effectiveness in automated unit test generation through the generation of edge cases. We generate robust edge cases using an evolutionary search process, where specialized agents sequentially propose, mutate, and score candidates. A controller maintains persistent state across generations, while evolutionary preservation ensures diversity and exploration across all possible cases. This yields a generalist agent capable of discovering robust, high-coverage edge cases across unseen codebases. Experiments show our stateful multi-agent inference framework achieves substantial gains in coverage over stateless single-step baselines, evaluated on prevalent unit-testing benchmarks such as HumanEval and TestGenEvalMini and using three diverse LLM families - Llama, Gemma, and GPT. These results indicate that combining persistent inference-time state with evolutionary search materially improves unit-test generation.
☆ Comparing human and language models sentence processing difficulties on complex structures
Large language models (LLMs) that fluently converse with humans are a reality - but do LLMs experience human-like processing difficulties? We systematically compare human and LLM sentence comprehension across seven challenging linguistic structures. We collect sentence comprehension data from humans and five families of state-of-the-art LLMs, varying in size and training procedure in a unified experimental framework. Our results show LLMs overall struggle on the target structures, but especially on garden path (GP) sentences. Indeed, while the strongest models achieve near perfect accuracy on non-GP structures (93.7% for GPT-5), they struggle on GP structures (46.8% for GPT-5). Additionally, when ranking structures based on average performance, rank correlation between humans and models increases with parameter count. For each target structure, we also collect data for their matched baseline without the difficult structure. Comparing performance on the target vs. baseline sentences, the performance gap observed in humans holds for LLMs, with two exceptions: for models that are too weak performance is uniformly low across both sentence types, and for models that are too strong the performance is uniformly high. Together, these reveal convergence and divergence in human and LLM sentence comprehension, offering new insights into the similarity of humans and LLMs.
comment: Data and code will be released soon
☆ TRIM: Token-wise Attention-Derived Saliency for Data-Efficient Instruction Tuning
Instruction tuning is essential for aligning large language models (LLMs) to downstream tasks and commonly relies on large, diverse corpora. However, small, high-quality subsets, known as coresets, can deliver comparable or superior results, though curating them remains challenging. Existing methods often rely on coarse, sample-level signals like gradients, an approach that is computationally expensive and overlooks fine-grained features. To address this, we introduce TRIM (Token Relevance via Interpretable Multi-layer Attention), a forward-only, token-centric framework. Instead of using gradients, TRIM operates by matching underlying representational patterns identified via attention-based "fingerprints" from a handful of target samples. Such an approach makes TRIM highly efficient and uniquely sensitive to the structural features that define a task. Coresets selected by our method consistently outperform state-of-the-art baselines by up to 9% on downstream tasks and even surpass the performance of full-data fine-tuning in some settings. By avoiding expensive backward passes, TRIM achieves this at a fraction of the computational cost. These findings establish TRIM as a scalable and efficient alternative for building high-quality instruction-tuning datasets.
☆ Opt-ICL at LeWiDi-2025: Maximizing In-Context Signal from Rater Examples via Meta-Learning EMNLP 2025
Many natural language processing (NLP) tasks involve subjectivity, ambiguity, or legitimate disagreement between annotators. In this paper, we outline our system for modeling human variation. Our system leverages language models' (LLMs) in-context learning abilities, along with a two-step meta-learning training procedure for 1) post-training on many datasets requiring in-context learning and 2) specializing the model via in-context meta-learning to the particular data distribution of interest. We also evaluate the performance of our system submission to the Learning With Disagreements (LeWiDi) competition, where it was the overall winner on both tasks. Additionally, we perform an ablation study to measure the importance of each system component. We find that including rater examples in-context is crucial for our system's performance, dataset-specific fine-tuning is helpful on the larger datasets, post-training on other in-context datasets is helpful on one of the competition datasets, and that performance improves with model scale.
comment: NLPerspectives: The 4th Workshop on Perspectivist Approaches to Natural Language Processing at EMNLP 2025
☆ TALENT: Table VQA via Augmented Language-Enhanced Natural-text Transcription
Table Visual Question Answering (Table VQA) is typically addressed by large vision-language models (VLMs). While such models can answer directly from images, they often miss fine-grained details unless scaled to very large sizes, which are computationally prohibitive, especially for mobile deployment. A lighter alternative is to have a small VLM perform OCR and then use a large language model (LLM) to reason over structured outputs such as Markdown tables. However, these representations are not naturally optimized for LLMs and still introduce substantial errors. We propose TALENT (Table VQA via Augmented Language-Enhanced Natural-text Transcription), a lightweight framework that leverages dual representations of tables. TALENT prompts a small VLM to produce both OCR text and natural language narration, then combines them with the question for reasoning by an LLM. This reframes Table VQA as an LLM-centric multimodal reasoning task, where the VLM serves as a perception-narration module rather than a monolithic solver. Additionally, we construct ReTabVQA, a more challenging Table VQA dataset requiring multi-step quantitative reasoning over table images. Experiments show that TALENT enables a small VLM-LLM combination to match or surpass a single large VLM at significantly lower computational cost on both public datasets and ReTabVQA.
☆ Making Machines Sound Sarcastic: LLM-Enhanced and Retrieval-Guided Sarcastic Speech Synthesis
Sarcasm is a subtle form of non-literal language that poses significant challenges for speech synthesis due to its reliance on nuanced semantic, contextual, and prosodic cues. While existing speech synthesis research has focused primarily on broad emotional categories, sarcasm remains largely unexplored. In this paper, we propose a Large Language Model (LLM)-enhanced Retrieval-Augmented framework for sarcasm-aware speech synthesis. Our approach combines (1) semantic embeddings from a LoRA-fine-tuned LLaMA 3, which capture pragmatic incongruity and discourse-level cues of sarcasm, and (2) prosodic exemplars retrieved via a Retrieval Augmented Generation (RAG) module, which provide expressive reference patterns of sarcastic delivery. Integrated within a VITS backbone, this dual conditioning enables more natural and contextually appropriate sarcastic speech. Experiments demonstrate that our method outperforms baselines in both objective measures and subjective evaluations, yielding improvements in speech naturalness, sarcastic expressivity, and downstream sarcasm detection.
☆ The Cognitive Bandwidth Bottleneck: Shifting Long-Horizon Agent from Planning with Actions to Planning with Schemas
Enabling LLMs to effectively operate long-horizon task which requires long-term planning and multiple interactions is essential for open-world autonomy. Conventional methods adopt planning with actions where a executable action list would be provided as reference. However, this action representation choice would be impractical when the environment action space is combinatorial exploded (e.g., open-ended real world). This naturally leads to a question: As environmental action space scales, what is the optimal action representation for long-horizon agents? In this paper, we systematically study the effectiveness of two different action representations. The first one is conventional planning with actions (PwA) which is predominantly adopted for its effectiveness on existing benchmarks. The other one is planning with schemas (PwS) which instantiate an action schema into action lists (e.g., "move [OBJ] to [OBJ]" -> "move apple to desk") to ensure concise action space and reliable scalability. This alternative is motivated by its alignment with human cognition and its compliance with environment-imposed action format restriction. We propose cognitive bandwidth perspective as a conceptual framework to qualitatively understand the differences between these two action representations and empirically observe a representation-choice inflection point between ALFWorld (~35 actions) and SciWorld (~500 actions), which serve as evidence of the need for scalable representations. We further conduct controlled experiments to study how the location of this inflection point interacts with different model capacities: stronger planning proficiency shifts the inflection rightward, whereas better schema instantiation shifts it leftward. Finally, noting the suboptimal performance of PwS agents, we provide an actionable guide for building more capable PwS agents for better scalable autonomy.
comment: 22 pages
☆ All Claims Are Equal, but Some Claims Are More Equal Than Others: Importance-Sensitive Factuality Evaluation of LLM Generations
Existing methods for evaluating the factuality of large language model (LLM) responses treat all claims as equally important. This results in misleading evaluations when vital information is missing or incorrect as it receives the same weight as peripheral details, raising the question: how can we reliably detect such differences when there are errors in key information? Current approaches that measure factuality tend to be insensitive to omitted or false key information. To investigate this lack of sensitivity, we construct VITALERRORS, a benchmark of 6,733 queries with minimally altered LLM responses designed to omit or falsify key information. Using this dataset, we demonstrate the insensitivities of existing evaluation metrics to key information errors. To address this gap, we introduce VITAL, a set of metrics that provide greater sensitivity in measuring the factuality of responses by incorporating the relevance and importance of claims with respect to the query. Our analysis demonstrates that VITAL metrics more reliably detect errors in key information than previous methods. Our dataset, metrics, and analysis provide a foundation for more accurate and robust assessment of LLM factuality.
☆ Accelerating Diffusion LLM Inference via Local Determinism Propagation
Diffusion large language models (dLLMs) represent a significant advancement in text generation, offering parallel token decoding capabilities. However, existing open-source implementations suffer from quality-speed trade-offs that impede their practical deployment. Conservative sampling strategies typically decode only the most confident token per step to ensure quality (i.e., greedy decoding), at the cost of inference efficiency due to repeated redundant refinement iterations--a phenomenon we term delayed decoding. Through systematic analysis of dLLM decoding dynamics, we characterize this delayed decoding behavior and propose a training-free adaptive parallel decoding strategy, named LocalLeap, to address these inefficiencies. LocalLeap is built on two fundamental empirical principles: local determinism propagation centered on high-confidence anchors and progressive spatial consistency decay. By applying these principles, LocalLeap identifies anchors and performs localized relaxed parallel decoding within bounded neighborhoods, achieving substantial inference step reduction through early commitment of already-determined tokens without compromising output quality. Comprehensive evaluation on various benchmarks demonstrates that LocalLeap achieves 6.94$\times$ throughput improvements and reduces decoding steps to just 14.2\% of the original requirement, achieving these gains with negligible performance impact. The source codes are available at: https://github.com/friedrichor/LocalLeap.
comment: 21 pages, 4 figures. Under review
☆ LuxInstruct: A Cross-Lingual Instruction Tuning Dataset For Luxembourgish
Instruction tuning has become a key technique for enhancing the performance of large language models, enabling them to better follow human prompts. However, low-resource languages such as Luxembourgish face severe limitations due to the lack of high-quality instruction datasets. Traditional reliance on machine translation often introduces semantic misalignment and cultural inaccuracies. In this work, we address these challenges by creating a cross-lingual instruction tuning dataset for Luxembourgish, without resorting to machine-generated translations into it. Instead, by leveraging aligned data from English, French, and German, we build a high-quality dataset that preserves linguistic and cultural nuances. We provide evidence that cross-lingual instruction tuning not only improves representational alignment across languages but also the model's generative capabilities in Luxembourgish. This highlights how cross-lingual data curation can avoid the common pitfalls of machine-translated data and directly benefit low-resource language development.
comment: Paper under review; Dataset available at https://huggingface.co/datasets/fredxlpy/LuxInstruct
☆ Revisiting Metric Reliability for Fine-grained Evaluation of Machine Translation and Summarization in Indian Languages
While automatic metrics drive progress in Machine Translation (MT) and Text Summarization (TS), existing metrics have been developed and validated almost exclusively for English and other high-resource languages. This narrow focus leaves Indian languages, spoken by over 1.5 billion people, largely overlooked, casting doubt on the universality of current evaluation practices. To address this gap, we introduce ITEM, a large-scale benchmark that systematically evaluates the alignment of 26 automatic metrics with human judgments across six major Indian languages, enriched with fine-grained annotations. Our extensive evaluation, covering agreement with human judgments, sensitivity to outliers, language-specific reliability, inter-metric correlations, and resilience to controlled perturbations, reveals four central findings: (1) LLM-based evaluators show the strongest alignment with human judgments at both segment and system levels; (2) outliers exert a significant impact on metric-human agreement; (3) in TS, metrics are more effective at capturing content fidelity, whereas in MT, they better reflect fluency; and (4) metrics differ in their robustness and sensitivity when subjected to diverse perturbations. Collectively, these findings offer critical guidance for advancing metric design and evaluation in Indian languages.
comment: 18 pages, 14 figures
☆ Does Local News Stay Local?: Online Content Shifts in Sinclair-Acquired Stations
Local news stations are often considered to be reliable sources of non-politicized information, particularly local concerns that residents care about. Because these stations are trusted news sources, viewers are particularly susceptible to the information they report. The Sinclair Broadcast group is a broadcasting company that has acquired many local news stations in the last decade. We investigate the effects of local news stations being acquired by Sinclair: how does coverage change? We use computational methods to investigate changes in internet content put out by local news stations before and after being acquired by Sinclair and in comparison to national news outlets. We find that there is clear evidence that local news stations report more frequently on national news at the expense of local topics, and that their coverage of polarizing national topics increases.
☆ Search-R3: Unifying Reasoning and Embedding Generation in Large Language Models
Despite their remarkable natural language understanding capabilities, Large Language Models (LLMs) have been underutilized for retrieval tasks. We present Search-R3, a novel framework that addresses this limitation by adapting LLMs to generate search embeddings as a direct output of their reasoning process. Our approach exploits LLMs' chain-of-thought capabilities, allowing them to produce more effective embeddings by reasoning step-by-step through complex semantic analyses. We implement this through three complementary mechanisms. (1) a supervised learning stage enables the model's ability to produce quality embeddings, (2) a reinforcement learning (RL) methodology that optimizes embedding generation alongside reasoning, and (3) a specialized RL environment that efficiently handles evolving embedding representations without requiring complete corpus re-encoding at each training iteration. Our extensive evaluations on diverse benchmarks demonstrate that Search-R3 significantly outperforms prior methods by unifying the reasoning and embedding generation processes. This integrated post-training approach represents a substantial advancement in handling complex knowledge-intensive tasks that require both sophisticated reasoning and effective information retrieval. Project page: https://github.com/ytgui/Search-R3
☆ Beyond Monolingual Assumptions: A Survey of Code-Switched NLP in the Era of Large Language Models
Code-switching (CSW), the alternation of languages and scripts within a single utterance, remains a fundamental challenge for multiling ual NLP, even amidst the rapid advances of large language models (LLMs). Most LLMs still struggle with mixed-language inputs, limited CSW datasets, and evaluation biases, hindering deployment in multilingual societies. This survey provides the first comprehensive analysis of CSW-aware LLM research, reviewing \total{unique_references} studies spanning five research areas, 12 NLP tasks, 30+ datasets, and 80+ languages. We classify recent advances by architecture, training strategy, and evaluation methodology, outlining how LLMs have reshaped CSW modeling and what challenges persist. The paper concludes with a roadmap emphasizing the need for inclusive datasets, fair evaluation, and linguistically grounded models to achieve truly multilingual intelligence. A curated collection of all resources is maintained at https://github.com/lingo-iitgn/awesome-code-mixing/.
☆ Mining the Mind: What 100M Beliefs Reveal About Frontier LLM Knowledge
LLMs are remarkable artifacts that have revolutionized a range of NLP and AI tasks. A significant contributor is their factual knowledge, which, to date, remains poorly understood, and is usually analyzed from biased samples. In this paper, we take a deep tour into the factual knowledge (or beliefs) of a frontier LLM, based on GPTKB v1.5 (Hu et al., 2025a), a recursively elicited set of 100 million beliefs of one of the strongest currently available frontier LLMs, GPT-4.1. We find that the models' factual knowledge differs quite significantly from established knowledge bases, and that its accuracy is significantly lower than indicated by previous benchmarks. We also find that inconsistency, ambiguity and hallucinations are major issues, shedding light on future research opportunities concerning factual LLM knowledge.
☆ Native Hybrid Attention for Efficient Sequence Modeling
Transformers excel at sequence modeling but face quadratic complexity, while linear attention offers improved efficiency but often compromises recall accuracy over long contexts. In this work, we introduce Native Hybrid Attention (NHA), a novel hybrid architecture of linear and full attention that integrates both intra \& inter-layer hybridization into a unified layer design. NHA maintains long-term context in key-value slots updated by a linear RNN, and augments them with short-term tokens from a sliding window. A single \texttt{softmax attention} operation is then applied over all keys and values, enabling per-token and per-head context-dependent weighting without requiring additional fusion parameters. The inter-layer behavior is controlled through a single hyperparameter, the sliding window size, which allows smooth adjustment between purely linear and full attention while keeping all layers structurally uniform. Experimental results show that NHA surpasses Transformers and other hybrid baselines on recall-intensive and commonsense reasoning tasks. Furthermore, pretrained LLMs can be structurally hybridized with NHA, achieving competitive accuracy while delivering significant efficiency gains. Code is available at https://github.com/JusenD/NHA.
comment: Technical report, 16 pages
☆ Pragyaan: Designing and Curating High-Quality Cultural Post-Training Datasets for Indian Languages EMNLP 2025
The effectiveness of Large Language Models (LLMs) depends heavily on the availability of high-quality post-training data, particularly instruction-tuning and preference-based examples. Existing open-source datasets, however, often lack multilingual coverage, cultural grounding, and suffer from task diversity gaps that are especially pronounced for Indian languages. We introduce a human-in-the-loop pipeline that combines translations with synthetic expansion to produce reliable and diverse Indic post-training data. Using this pipeline, we curate two datasets: Pragyaan-IT (22.5K) and Pragyaan-Align (100K) across 10 Indian languages covering 13 broad and 56 sub-categories, leveraging 57 diverse datasets. Our dataset protocol incorporates several often-overlooked dimensions and emphasize task diversity, multi-turn dialogue, instruction fidelity, safety alignment, and preservation of cultural nuance, providing a foundation for more inclusive and effective multilingual LLMs.
comment: EMNLP 2025
☆ Towards Reliable Retrieval in RAG Systems for Large Legal Datasets EMNLP 2025
Retrieval-Augmented Generation (RAG) is a promising approach to mitigate hallucinations in Large Language Models (LLMs) for legal applications, but its reliability is critically dependent on the accuracy of the retrieval step. This is particularly challenging in the legal domain, where large databases of structurally similar documents often cause retrieval systems to fail. In this paper, we address this challenge by first identifying and quantifying a critical failure mode we term Document-Level Retrieval Mismatch (DRM), where the retriever selects information from entirely incorrect source documents. To mitigate DRM, we investigate a simple and computationally efficient technique which we refer to as Summary-Augmented Chunking (SAC). This method enhances each text chunk with a document-level synthetic summary, thereby injecting crucial global context that would otherwise be lost during a standard chunking process. Our experiments on a diverse set of legal information retrieval tasks show that SAC greatly reduces DRM and, consequently, also improves text-level retrieval precision and recall. Interestingly, we find that a generic summarization strategy outperforms an approach that incorporates legal expert domain knowledge to target specific legal elements. Our work provides evidence that this practical, scalable, and easily integrable technique enhances the reliability of RAG systems when applied to large-scale legal document datasets.
comment: Accepted for the 7th Natural Legal Language Processing Workshop (NLLP 2025), co-located with EMNLP 2025
☆ RedTWIZ: Diverse LLM Red Teaming via Adaptive Attack Planning
This paper presents the vision, scientific contributions, and technical details of RedTWIZ: an adaptive and diverse multi-turn red teaming framework, to audit the robustness of Large Language Models (LLMs) in AI-assisted software development. Our work is driven by three major research streams: (1) robust and systematic assessment of LLM conversational jailbreaks; (2) a diverse generative multi-turn attack suite, supporting compositional, realistic and goal-oriented jailbreak conversational strategies; and (3) a hierarchical attack planner, which adaptively plans, serializes, and triggers attacks tailored to specific LLM's vulnerabilities. Together, these contributions form a unified framework -- combining assessment, attack generation, and strategic planning -- to comprehensively evaluate and expose weaknesses in LLMs' robustness. Extensive evaluation is conducted to systematically assess and analyze the performance of the overall system and each component. Experimental results demonstrate that our multi-turn adversarial attack strategies can successfully lead state-of-the-art LLMs to produce unsafe generations, highlighting the pressing need for more research into enhancing LLM's robustness.
☆ VelLMes: A high-interaction AI-based deception framework EuroS&P 2025
There are very few SotA deception systems based on Large Language Models. The existing ones are limited only to simulating one type of service, mainly SSH shells. These systems - but also the deception technologies not based on LLMs - lack an extensive evaluation that includes human attackers. Generative AI has recently become a valuable asset for cybersecurity researchers and practitioners, and the field of cyber-deception is no exception. Researchers have demonstrated how LLMs can be leveraged to create realistic-looking honeytokens, fake users, and even simulated systems that can be used as honeypots. This paper presents an AI-based deception framework called VelLMes, which can simulate multiple protocols and services such as SSH Linux shell, MySQL, POP3, and HTTP. All of these can be deployed and used as honeypots, thus VelLMes offers a variety of choices for deception design based on the users' needs. VelLMes is designed to be attacked by humans, so interactivity and realism are key for its performance. We evaluate the generative capabilities and the deception capabilities. Generative capabilities were evaluated using unit tests for LLMs. The results of the unit tests show that, with careful prompting, LLMs can produce realistic-looking responses, with some LLMs having a 100% passing rate. In the case of the SSH Linux shell, we evaluated deception capabilities with 89 human attackers. The results showed that about 30% of the attackers thought that they were interacting with a real system when they were assigned an LLM-based honeypot. Lastly, we deployed 10 instances of the SSH Linux shell honeypot on the Internet to capture real-life attacks. Analysis of these attacks showed us that LLM honeypots simulating Linux shells can perform well against unstructured and unexpected attacks on the Internet, responding correctly to most of the issued commands.
comment: 9 pages. 9 figures. 1 table. This is a preprint of a paper that was presented at the Active Defense and Deception Workshop colocated with IEEE EuroS&P 2025 conference
☆ Probing Social Identity Bias in Chinese LLMs with Gendered Pronouns and Social Groups
Large language models (LLMs) are increasingly deployed in user-facing applications, raising concerns about their potential to reflect and amplify social biases. We investigate social identity framing in Chinese LLMs using Mandarin-specific prompts across ten representative Chinese LLMs, evaluating responses to ingroup ("We") and outgroup ("They") framings, and extending the setting to 240 social groups salient in the Chinese context. To complement controlled experiments, we further analyze Chinese-language conversations from a corpus of real interactions between users and chatbots. Across models, we observe systematic ingroup-positive and outgroup-negative tendencies, which are not confined to synthetic prompts but also appear in naturalistic dialogue, indicating that bias dynamics might strengthen in real interactions. Our study provides a language-aware evaluation framework for Chinese LLMs, demonstrating that social identity biases documented in English generalize cross-linguistically and intensify in user-facing contexts.
☆ EDUMATH: Generating Standards-aligned Educational Math Word Problems
Math word problems (MWPs) are critical K-12 educational tools, and customizing them to students' interests and ability levels can increase learning outcomes. However, teachers struggle to find time to customize MWPs for each student given large class sizes and increasing burnout. We propose that LLMs can support math education by generating MWPs customized to student interests and math education standards. To this end, we use a joint human expert-LLM judge approach to evaluate over 11,000 MWPs generated by open and closed LLMs and develop the first teacher-annotated dataset for standards-aligned educational MWP generation. We show the value of our data by using it to train a 12B open model that matches the performance of larger and more capable open models. We also use our teacher-annotated data to train a text classifier that enables a 30B open LLM to outperform existing closed baselines without any training. Next, we show our models' MWPs are more similar to human-written MWPs than those from existing models. We conclude by conducting the first study of customized LLM-generated MWPs with grade school students, finding they perform similarly on our models' MWPs relative to human-written MWPs but consistently prefer our customized MWPs.
comment: 32 pages, 15 figures
☆ Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation ICASSP 2026
Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including dedicated multilingual and long-form tracks. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.
comment: Submitted to ICASSP 2026; Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard; Code: https://github.com/huggingface/open_asr_leaderboard
☆ Revisiting the Uniform Information Density Hypothesis in LLM Reasoning Traces
The Uniform Information Density (UID) hypothesis suggests that effective communication maintains a stable flow of information. In this work, we revisit this principle in the context of large language model (LLM) reasoning traces, asking whether step-level uniformity reflects reasoning quality. To this end, we propose an entropy-based stepwise information density metric and introduce two complementary measures of uniformity, local and global uniformity scores. Across the experiments on six different reasoning benchmarks, we find that step-level uniformity not only provides a strong theoretical lens but also yields practical performance benefits; for example, selecting reasoning traces with more uniform information density at the step-level improves accuracy by 10-32\% relative gains over baselines at AIME2025. Our analysis further reveals that correct reasoning traces tend to avoid sharp information density spikes, while incorrect traces exhibit irregular information bursts. These results demonstrate that UID-inspired information density measures outperform alternative internal signals as predictors of reasoning quality. Results highlight the uniformity of the information density as a robust diagnostic and selection criterion for building more reliable and accurate reasoning systems.
☆ SHANKS: Simultaneous Hearing and Thinking for Spoken Language Models
Current large language models (LLMs) and spoken language models (SLMs) begin thinking and taking actions only after the user has finished their turn. This prevents the model from interacting during the user's turn and can lead to high response latency while it waits to think. Consequently, thinking after receiving the full input is not suitable for speech-to-speech interaction, where real-time, low-latency exchange is important. We address this by noting that humans naturally "think while listening." In this paper, we propose SHANKS, a general inference framework that enables SLMs to generate unspoken chain-of-thought reasoning while listening to the user input. SHANKS streams the input speech in fixed-duration chunks and, as soon as a chunk is received, generates unspoken reasoning based on all previous speech and reasoning, while the user continues speaking. SHANKS uses this unspoken reasoning to decide whether to interrupt the user and to make tool calls to complete the task. We demonstrate that SHANKS enhances real-time user-SLM interaction in two scenarios: (1) when the user is presenting a step-by-step solution to a math problem, SHANKS can listen, reason, and interrupt when the user makes a mistake, achieving 37.1% higher interruption accuracy than a baseline that interrupts without thinking; and (2) in a tool-augmented dialogue, SHANKS can complete 56.9% of the tool calls before the user finishes their turn. Overall, SHANKS moves toward models that keep thinking throughout the conversation, not only after a turn ends. Animated illustrations of Shanks can be found at https://d223302.github.io/SHANKS/
comment: Work in progress
☆ LongRM: Revealing and Unlocking the Context Boundary of Reward Modeling
Reward model (RM) plays a pivotal role in aligning large language model (LLM) with human preferences. As real-world applications increasingly involve long history trajectories, e.g., LLM agent, it becomes indispensable to evaluate whether a model's responses are not only high-quality but also grounded in and consistent with the provided context. Yet, current RMs remain confined to short-context settings and primarily focus on response-level attributes (e.g., safety or helpfulness), while largely neglecting the critical dimension of long context-response consistency. In this work, we introduce Long-RewardBench, a benchmark specifically designed for long-context RM evaluation, featuring both Pairwise Comparison and Best-of-N tasks. Our preliminary study reveals that even state-of-the-art generative RMs exhibit significant fragility in long-context scenarios, failing to maintain context-aware preference judgments. Motivated by the analysis of failure patterns observed in model outputs, we propose a general multi-stage training strategy that effectively scales arbitrary models into robust Long-context RMs (LongRMs). Experiments show that our approach not only substantially improves performance on long-context evaluation but also preserves strong short-context capability. Notably, our 8B LongRM outperforms much larger 70B-scale baselines and matches the performance of the proprietary Gemini 2.5 Pro model.
☆ MeXtract: Light-Weight Metadata Extraction from Scientific Papers
Metadata plays a critical role in indexing, documenting, and analyzing scientific literature, yet extracting it accurately and efficiently remains a challenging task. Traditional approaches often rely on rule-based or task-specific models, which struggle to generalize across domains and schema variations. In this paper, we present MeXtract, a family of lightweight language models designed for metadata extraction from scientific papers. The models, ranging from 0.5B to 3B parameters, are built by fine-tuning Qwen 2.5 counterparts. In their size family, MeXtract achieves state-of-the-art performance on metadata extraction on the MOLE benchmark. To further support evaluation, we extend the MOLE benchmark to incorporate model-specific metadata, providing an out-of-domain challenging subset. Our experiments show that fine-tuning on a given schema not only yields high accuracy but also transfers effectively to unseen schemas, demonstrating the robustness and adaptability of our approach. We release all the code, datasets, and models openly for the research community.
☆ $λ$-GRPO: Unifying the GRPO Frameworks with Learnable Token Preferences
Reinforcement Learning with Human Feedback (RLHF) has been the dominant approach for improving the reasoning capabilities of Large Language Models (LLMs). Recently, Reinforcement Learning with Verifiable Rewards (RLVR) has simplified this paradigm by replacing the reward and value models with rule-based verifiers. A prominent example is Group Relative Policy Optimization (GRPO). However, GRPO inherently suffers from a length bias, since the same advantage is uniformly assigned to all tokens of a response. As a result, longer responses distribute the reward over more tokens and thus contribute disproportionately to gradient updates. Several variants, such as DAPO and Dr. GRPO, modify the token-level aggregation of the loss, yet these methods remain heuristic and offer limited interpretability regarding their implicit token preferences. In this work, we explore the possibility of allowing the model to learn its own token preference during optimization. We unify existing frameworks under a single formulation and introduce a learnable parameter $\lambda$ that adaptively controls token-level weighting. We use $\lambda$-GRPO to denote our method, and we find that $\lambda$-GRPO achieves consistent improvements over vanilla GRPO and DAPO on multiple mathematical reasoning benchmarks. On Qwen2.5 models with 1.5B, 3B, and 7B parameters, $\lambda$-GRPO improves average accuracy by $+1.9\%$, $+1.0\%$, and $+1.7\%$ compared to GRPO, respectively. Importantly, these gains come without any modifications to the training data or additional computational cost, highlighting the effectiveness and practicality of learning token preferences.
comment: 9 pages
☆ Unlocking Latent Discourse Translation in LLMs Through Quality-Aware Decoding
Large language models (LLMs) have emerged as strong contenders in machine translation.Yet, they still struggle to adequately handle discourse phenomena, such as pronoun resolution and lexical cohesion at the document level. In this study, we thoroughly investigate the discourse phenomena performance of LLMs in context-aware translation. We demonstrate that discourse knowledge is encoded within LLMs and propose the use of quality-aware decoding (QAD) to effectively extract this knowledge, showcasing its superiority over other decoding approaches through comprehensive analysis. Furthermore, we illustrate that QAD enhances the semantic richness of translations and aligns them more closely with human preferences.
☆ OpenJAI-v1.0: An Open Thai Large Language Model
We introduce OpenJAI-v1.0, an open-source large language model for Thai and English, developed from the Qwen3-14B model. Our work focuses on boosting performance on practical tasks through carefully curated data across three key use cases: instruction following, long-context understanding, and tool use. Evaluation results show that OpenJAI-v1.0 improves on the capabilities of its base model and outperforms other leading open-source Thai models on a diverse suite of benchmarks, while avoiding catastrophic forgetting. OpenJAI-v1.0 is publicly released as another alternative NLP resource for the Thai AI community.
☆ SID: Multi-LLM Debate Driven by Self Signals
Large Language Models (LLMs) have exhibited impressive capabilities across diverse application domains. Recent work has explored Multi-LLM Agent Debate (MAD) as a way to enhance performance by enabling multiple LLMs to discuss and refine responses iteratively. Nevertheless, existing MAD methods predominantly focus on utilizing external structures, such as debate graphs, using LLM-as-a-Judge, while neglecting the application of self signals, such as token logits and attention, that arise during generation. This omission leads to redundant computation and potential performance degradation. In this paper, we shift the focus to the self signals of multi-LLM debate and introduce a Self-Signals Driven Multi-LLM Debate (SID), which leverages two types of self-signals: model-level confidence and token-level semantic focus, to adaptively guide the debate process. Our approach enables high-confidence agents to exit early at the model level and compress the redundant debate contents based on the attention mechanism. We evaluate our method on various LLMs and Multimodal LLMs across multiple challenging benchmarks. Experimental results demonstrate that our method not only outperforms existing MAD techniques in accuracy but also reduces token consumption, highlighting the effectiveness of utilizing self signals in enhancing both the performance and efficiency of multi-agent debate systems. Our code will be available at~\href{https://github.com/xuhang2019/SID}{\texttt{https://github.com/xuhang2019/SID}}.
☆ GAMBIT+: A Challenge Set for Evaluating Gender Bias in Machine Translation Quality Estimation Metrics EMNLP 2025
Gender bias in machine translation (MT) systems has been extensively documented, but bias in automatic quality estimation (QE) metrics remains comparatively underexplored. Existing studies suggest that QE metrics can also exhibit gender bias, yet most analyses are limited by small datasets, narrow occupational coverage, and restricted language variety. To address this gap, we introduce a large-scale challenge set specifically designed to probe the behavior of QE metrics when evaluating translations containing gender-ambiguous occupational terms. Building on the GAMBIT corpus of English texts with gender-ambiguous occupations, we extend coverage to three source languages that are genderless or natural-gendered, and eleven target languages with grammatical gender, resulting in 33 source-target language pairs. Each source text is paired with two target versions differing only in the grammatical gender of the occupational term(s) (masculine vs. feminine), with all dependent grammatical elements adjusted accordingly. An unbiased QE metric should assign equal or near-equal scores to both versions. The dataset's scale, breadth, and fully parallel design, where the same set of texts is aligned across all languages, enables fine-grained bias analysis by occupation and systematic comparisons across languages.
comment: Accepted for publication at the 10th Conference of Machine Translation (WMT25), co-located with EMNLP 2025
☆ Crossing Domains without Labels: Distant Supervision for Term Extraction EMNLP
Automatic Term Extraction (ATE) is a critical component in downstream NLP tasks such as document tagging, ontology construction and patent analysis. Current state-of-the-art methods require expensive human annotation and struggle with domain transfer, limiting their practical deployment. This highlights the need for more robust, scalable solutions and realistic evaluation settings. To address this, we introduce a comprehensive benchmark spanning seven diverse domains, enabling performance evaluation at both the document- and corpus-levels. Furthermore, we propose a robust LLM-based model that outperforms both supervised cross-domain encoder models and few-shot learning baselines and performs competitively with its GPT-4o teacher on this benchmark. The first step of our approach is generating psuedo-labels with this black-box LLM on general and scientific domains to ensure generalizability. Building on this data, we fine-tune the first LLMs for ATE. To further enhance document-level consistency, oftentimes needed for downstream tasks, we introduce lightweight post-hoc heuristics. Our approach exceeds previous approaches on 5/7 domains with an average improvement of 10 percentage points. We release our dataset and fine-tuned models to support future research in this area.
comment: Accepted at EMNLP Industry Track 2025
☆ Mid-Training of Large Language Models: A Survey
Large language models (LLMs) are typically developed through large-scale pre-training followed by task-specific fine-tuning. Recent advances highlight the importance of an intermediate mid-training stage, where models undergo multiple annealing-style phases that refine data quality, adapt optimization schedules, and extend context length. This stage mitigates diminishing returns from noisy tokens, stabilizes convergence, and expands model capability in late training. Its effectiveness can be explained through gradient noise scale, the information bottleneck, and curriculum learning, which together promote generalization and abstraction. Despite widespread use in state-of-the-art systems, there has been no prior survey of mid-training as a unified paradigm. We introduce the first taxonomy of LLM mid-training spanning data distribution, learning-rate scheduling, and long-context extension. We distill practical insights, compile evaluation benchmarks, and report gains to enable structured comparisons across models. We also identify open challenges and propose avenues for future research and practice.
☆ Adaptive Tool Generation with Models as Tools and Reinforcement Learning
Tool-augmented language models have demonstrated strong capabilities, but their reliance on live API access creates scalability and reliability challenges during training and deployment. We propose MTR, a simulation-first training framework for tool-augmented reasoning. Instead of relying on live APIs, MTR learns from complete ReAct traces with schema-validated, simulated observations. Our approach operates through a multi-agent architecture where a ToolMaker generates task-specific, OpenAI-compatible tool interfaces, an AutoAgent produces structured think-act-observe sequences, and a ToolActor simulates realistic responses. Training proceeds in two stages: Stage-1 Supervised Fine-Tuning (SFT) teaches 'trace grammar' from complete reasoning sequences; Stage-2 Group Relative Policy Optimization (GRPO) optimizes strategy with a composite trace reward that balances answer correctness and internal consistency. Across four multi-hop QA benchmarks (HotpotQA, MuSiQue, 2WikiMultiHopQA, Bamboogle), MTR attains competitive Exact Match (EM) scores to live-API systems and excels on reasoning-intensive tasks, suggesting that effective tool reasoning can be learned from structured traces without live interactions.
☆ Exposing Citation Vulnerabilities in Generative Engines
We analyze answers generated by generative engines (GEs) from the perspectives of citation publishers and the content-injection barrier, defined as the difficulty for attackers to manipulate answers to user prompts by placing malicious content on the web. GEs integrate two functions: web search and answer generation that cites web pages using large language models. Because anyone can publish information on the web, GEs are vulnerable to poisoning attacks. Existing studies of citation evaluation focus on how faithfully answer content reflects cited sources, leaving unexamined which web sources should be selected as citations to defend against poisoning attacks. To fill this gap, we introduce evaluation criteria that assess poisoning threats using the citation information contained in answers. Our criteria classify the publisher attributes of citations to estimate the content-injection barrier thereby revealing the threat of poisoning attacks in current GEs. We conduct experiments in political domains in Japan and the United States (U.S.) using our criteria and show that citations from official party websites (primary sources) are approximately \(25\%\)--\(45\%\) in the U.S. and \(60\%\)--\(65\%\) in Japan, indicating that U.S. political answers are at higher risk of poisoning attacks. We also find that sources with low content-injection barriers are frequently cited yet are poorly reflected in answer content. To mitigate this threat, we discuss how publishers of primary sources can increase exposure of their web content in answers and show that well-known techniques are limited by language differences.
comment: 12 pages, under-reviewing at a conference
☆ BlackboxNLP-2025 MIB Shared Task: Exploring Ensemble Strategies for Circuit Localization Methods
The Circuit Localization track of the Mechanistic Interpretability Benchmark (MIB) evaluates methods for localizing circuits within large language models (LLMs), i.e., subnetworks responsible for specific task behaviors. In this work, we investigate whether ensembling two or more circuit localization methods can improve performance. We explore two variants: parallel and sequential ensembling. In parallel ensembling, we combine attribution scores assigned to each edge by different methods-e.g., by averaging or taking the minimum or maximum value. In the sequential ensemble, we use edge attribution scores obtained via EAP-IG as a warm start for a more expensive but more precise circuit identification method, namely edge pruning. We observe that both approaches yield notable gains on the benchmark metrics, leading to a more precise circuit identification approach. Finally, we find that taking a parallel ensemble over various methods, including the sequential ensemble, achieves the best results. We evaluate our approach in the BlackboxNLP 2025 MIB Shared Task, comparing ensemble scores to official baselines across multiple model-task combinations.
comment: The 8th BlackboxNLP Workshop (Shared Task), 6 pages
Overview of the Plagiarism Detection Task at PAN 2025
The generative plagiarism detection task at PAN 2025 aims at identifying automatically generated textual plagiarism in scientific articles and aligning them with their respective sources. We created a novel large-scale dataset of automatically generated plagiarism using three large language models: Llama, DeepSeek-R1, and Mistral. In this task overview paper, we outline the creation of this dataset, summarize and compare the results of all participants and four baselines, and evaluate the results on the last plagiarism detection task from PAN 2015 in order to interpret the robustness of the proposed approaches. We found that the current iteration does not invite a large variety of approaches as naive semantic similarity approaches based on embedding vectors provide promising results of up to 0.8 recall and 0.5 precision. In contrast, most of these approaches underperform significantly on the 2015 dataset, indicating a lack in generalizability.
comment: Working Notes at PAN at CLEF 2025
☆ FURINA: A Fully Customizable Role-Playing Benchmark via Scalable Multi-Agent Collaboration Pipeline
As large language models (LLMs) advance in role-playing (RP) tasks, existing benchmarks quickly become obsolete due to their narrow scope, outdated interaction paradigms, and limited adaptability across diverse application scenarios. To address this gap, we introduce FURINA-Builder, a novel multi-agent collaboration pipeline that automatically constructs fully customizable RP benchmarks at any scale. It enables evaluation of arbitrary characters across diverse scenarios and prompt formats, as the first benchmark builder in RP area for adaptable assessment. FURINA-Builder simulates dialogues between a test character and other characters drawn from a well-constructed character-scene pool, while an LLM judge selects fine-grained evaluation dimensions and adjusts the test character's responses into final test utterances. Using this pipeline, we build FURINA-Bench, a new comprehensive role-playing benchmark featuring both established and synthesized test characters, each assessed with dimension-specific evaluation criteria. Human evaluation and preliminary separability analysis justify our pipeline and benchmark design. We conduct extensive evaluations of cutting-edge LLMs and find that o3 and DeepSeek-R1 achieve the best performance on English and Chinese RP tasks, respectively. Across all models, established characters consistently outperform synthesized ones, with reasoning capabilities further amplifying this disparity. Interestingly, we observe that model scale does not monotonically reduce hallucinations. More critically, for reasoning LLMs, we uncover a novel trade-off: reasoning improves RP performance but simultaneously increases RP hallucinations. This trade-off extends to a broader Pareto frontier between RP performance and reliability for all LLMs. These findings demonstrate the effectiveness of FURINA-Builder and the challenge posed by FURINA-Bench.
GPT-5 Model Corrected GPT-4V's Chart Reading Errors, Not Prompting
We present a quantitative evaluation to understand the effect of zero-shot large-language model (LLMs) and prompting uses on chart reading tasks. We asked LLMs to answer 107 visualization questions to compare inference accuracies between the agentic GPT-5 and multimodal GPT-4V, for difficult image instances, where GPT-4V failed to produce correct answers. Our results show that model architecture dominates the inference accuracy: GPT5 largely improved accuracy, while prompt variants yielded only small effects. Pre-registration of this work is available here: https://osf.io/u78td/?view_only=6b075584311f48e991c39335c840ded3; the Google Drive materials are here:https://drive.google.com/file/d/1ll8WWZDf7cCNcfNWrLViWt8GwDNSvVrp/view.
☆ Foundations of LLM Knowledge Materialization: Termination, Reproducibility, Robustness
Large Language Models (LLMs) encode substantial factual knowledge, yet measuring and systematizing this knowledge remains challenging. Converting it into structured format, for example through recursive extraction approaches such as the GPTKB methodology (Hu et al., 2025b), is still underexplored. Key open questions include whether such extraction can terminate, whether its outputs are reproducible, and how robust they are to variations. We systematically study LLM knowledge materialization using miniGPTKBs (domain-specific, tractable subcrawls), analyzing termination, reproducibility, and robustness across three categories of metrics: yield, lexical similarity, and semantic similarity. We experiment with four variations (seed, language, randomness, model) and three illustrative domains (from history, entertainment, and finance). Our findings show (i) high termination rates, though model-dependent; (ii) mixed reproducibility; and (iii) robustness that varies by perturbation type: high for seeds and temperature, lower for languages and models. These results suggest that LLM knowledge materialization can reliably surface core knowledge, while also revealing important limitations.
☆ Adaptive LLM-Symbolic Reasoning via Dynamic Logical Solver Composition
Neuro-symbolic NLP methods aim to leverage the complementary strengths of large language models and formal logical solvers. However, current approaches are mostly static in nature, i.e., the integration of a target solver is predetermined at design time, hindering the ability to employ diverse formal inference strategies. To address this, we introduce an adaptive, multi-paradigm, neuro-symbolic inference framework that: (1) automatically identifies formal reasoning strategies from problems expressed in natural language; and (2) dynamically selects and applies specialized formal logical solvers via autoformalization interfaces. Extensive experiments on individual and multi-paradigm reasoning tasks support the following conclusions: LLMs are effective at predicting the necessary formal reasoning strategies with an accuracy above 90 percent. This enables flexible integration with formal logical solvers, resulting in our framework outperforming competing baselines by 27 percent and 6 percent compared to GPT-4o and DeepSeek-V3.1, respectively. Moreover, adaptive reasoning can even positively impact pure LLM methods, yielding gains of 10, 5, and 6 percent on zero-shot, CoT, and symbolic CoT settings with GPT-4o. Finally, although smaller models struggle with adaptive neuro-symbolic reasoning, post-training offers a viable path to improvement. Overall, this work establishes the foundations for adaptive LLM-symbolic reasoning, offering a path forward for unifying material and formal inferences on heterogeneous reasoning challenges.
☆ Evolving and Executing Research Plans via Double-Loop Multi-Agent Collaboration
Automating the end-to-end scientific research process poses a fundamental challenge: it requires both evolving high-level plans that are novel and sound, and executing these plans correctly amidst dynamic and uncertain conditions. To address this bilevel challenge, we propose a novel Double-Loop Multi-Agent (DLMA) framework to solve the given research problem automatically. The leader loop, composed of professor agents, is responsible for evolving research plans. It employs an evolutionary algorithm through involvement, improvement, and integration meetings to iteratively generate and refine a pool of research proposals, exploring the solution space effectively. The follower loop, composed of doctoral student agents, is responsible for executing the best-evolved plan. It dynamically adjusts the plan during implementation via pre-hoc and post-hoc meetings, ensuring each step (e.g., drafting, coding) is well-supported by contextual and external observations. Extensive experiments on benchmarks like ACLAward and Laboratory show that DLMA generates research papers that achieve state-of-the-art scores in automated evaluation, significantly outperforming strong baselines. Ablation studies confirm the critical roles of both loops, with evolution driving novelty and execution ensuring soundness.
☆ Gold-Switch: Training-Free Superposition of Slow- and Fast- Thinking LLMs
Large Reasoning Models (LRMs) excel in structured tasks by emulating deliberate human reasoning but often suffer from overthinking, degrading performance and wasting resources. One possible baseline is to deploy both LLM and LRM, then route input by predicting whether it requires reasoning and may cause overthinking. However, deploying multiple models can be costly or impractical. We propose a superposed deployment strategy with a lightweight, training-free regulation to optimize inference by switching one model on and off. Instead of routing, we selectively unlearn from LRM at inference, scaling down computation while preserving reasoning. By analyzing the cumulative energy of singular values, we identify optimal low-rank projections to adjust reasoning just right.
☆ A Formal Framework for Fluency-based Multi-Reference Evaluation in Grammatical Error Correction ACL
Evaluating grammatical error correction requires metrics that reflect the diversity of valid human corrections rather than privileging a single reference. Existing frameworks, largely edit-based and English-centric, rely on rigid alignments between system and reference edits, limiting their applicability in multilingual and generative settings. This paper introduces a formal framework for \textit{fluency-based multi-reference evaluation}, framing $n$-gram similarity as an aggregation problem over multiple legitimate corrections. Within this formulation, we instantiate GLEU through four aggregation strategies--\textsc{select-best}, \textsc{simple-average}, \textsc{weighted-average}, and \textsc{merged-counts}--and analyze their properties of boundedness, monotonicity, and sensitivity to reference variation. Empirical results on Czech, Estonian, Ukrainian, and Chinese corpora show that these strategies capture complementary aspects of fluency and coverage. The framework unifies multi-reference evaluation into a principled, fluency-oriented approach that incorporates linguistic diversity without penalizing legitimate variation.
comment: Submitted to ACL Rolling Review - October 2025 for EACL 2026
☆ TWIST: Training-free and Label-free Short Text Clustering through Iterative Vector Updating with LLMs
In this paper, we propose a training-free and label-free method for short text clustering that can be used on top of any existing embedder. In the context of customer-facing chatbots, companies are dealing with large amounts of user utterances that need to be clustered according to their intent. In these commercial settings, no labeled data is typically available, and the number of clusters is not known. Our method is based on iterative vector updating: it constructs sparse vectors based on representative texts, and then iteratively refines them through LLM guidance. Our method achieves comparable or superior results to state-of-the-art methods that use contrastive learning, but without assuming prior knowledge of clusters or labels. Experiments on diverse datasets and smaller LLMs show that our method is model agnostic and can be applied to any embedder, with relatively small LLMs, and different clustering methods. We also show that our method scales to large datasets, reducing the computational cost of the LLM. These low-resource, adaptable settings and the scalability of our method make it more aligned with real-world scenarios than existing clustering methods.
☆ Evaluating LLMs for Historical Document OCR: A Methodological Framework for Digital Humanities
Digital humanities scholars increasingly use Large Language Models for historical document digitization, yet lack appropriate evaluation frameworks for LLM-based OCR. Traditional metrics fail to capture temporal biases and period-specific errors crucial for historical corpus creation. We present an evaluation methodology for LLM-based historical OCR, addressing contamination risks and systematic biases in diplomatic transcription. Using 18th-century Russian Civil font texts, we introduce novel metrics including Historical Character Preservation Rate (HCPR) and Archaic Insertion Rate (AIR), alongside protocols for contamination control and stability testing. We evaluate 12 multimodal LLMs, finding that Gemini and Qwen models outperform traditional OCR while exhibiting over-historicization: inserting archaic characters from incorrect historical periods. Post-OCR correction degrades rather than improves performance. Our methodology provides digital humanities practitioners with guidelines for model selection and quality assessment in historical corpus digitization.
comment: The First Workshop on Natural Language Processing and Language Models for Digital Humanities (LM4DH 2025). RANLP 2025
☆ AWM: Accurate Weight-Matrix Fingerprint for Large Language Models
Protecting the intellectual property of large language models (LLMs) is crucial, given the substantial resources required for their training. Consequently, there is an urgent need for both model owners and third parties to determine whether a suspect LLM is trained from scratch or derived from an existing base model. However, the intensive post-training processes that models typically undergo-such as supervised fine-tuning, extensive continued pretraining, reinforcement learning, multi-modal extension, pruning, and upcycling-pose significant challenges to reliable identification. In this work, we propose a training-free fingerprinting method based on weight matrices. We leverage the Linear Assignment Problem (LAP) and an unbiased Centered Kernel Alignment (CKA) similarity to neutralize the effects of parameter manipulations, yielding a highly robust and high-fidelity similarity metric. On a comprehensive testbed of 60 positive and 90 negative model pairs, our method demonstrates exceptional robustness against all six aforementioned post-training categories while exhibiting a near-zero risk of false positives. By achieving perfect scores on all classification metrics, our approach establishes a strong basis for reliable model lineage verification. Moreover, the entire computation completes within 30s on an NVIDIA 3090 GPU. The code is available at https://github.com/LUMIA-Group/AWM.
☆ Are LLMs Reliable Rankers? Rank Manipulation via Two-Stage Token Optimization
Large language models (LLMs) are increasingly used as rerankers in information retrieval, yet their ranking behavior can be steered by small, natural-sounding prompts. To expose this vulnerability, we present Rank Anything First (RAF), a two-stage token optimization method that crafts concise textual perturbations to consistently promote a target item in LLM-generated rankings while remaining hard to detect. Stage 1 uses Greedy Coordinate Gradient to shortlist candidate tokens at the current position by combining the gradient of the rank-target with a readability score; Stage 2 evaluates those candidates under exact ranking and readability losses using an entropy-based dynamic weighting scheme, and selects a token via temperature-controlled sampling. RAF generates ranking-promoting prompts token-by-token, guided by dual objectives: maximizing ranking effectiveness and preserving linguistic naturalness. Experiments across multiple LLMs show that RAF significantly boosts the rank of target items using naturalistic language, with greater robustness than existing methods in both promoting target items and maintaining naturalness. These findings underscore a critical security implication: LLM-based reranking is inherently susceptible to adversarial manipulation, raising new challenges for the trustworthiness and robustness of modern retrieval systems. Our code is available at: https://github.com/glad-lab/RAF.
comment: 10 pages, 3 figures
☆ PTEB: Towards Robust Text Embedding Evaluation via Stochastic Paraphrasing at Evaluation Time with LLMs
Current evaluations of sentence embedding models typically rely on static test beds such as the Massive Text Embedding Benchmark (MTEB). While invaluable, repeated tuning on a fixed suite can inflate reported performance and obscure real-world robustness. We introduce the Paraphrasing Text Embedding Benchmark (PTEB), a dynamic protocol that stochastically generates meaning-preserving paraphrases at evaluation time and aggregates results across multiple runs. Using a cost-efficient LLM-based method grounded in semantic textual similarity gold ratings, we show that LLMs generate token-diverse but semantically preserving, paraphrases. Across 7 MTEB tasks, we validate our hypothesis that the performance of sentence encoders is sensitive to changes in token space even when semantics remain fixed. We also observe that smaller models are not disproportionately affected relative to larger ones. Our results are statistically robust over multiple runs and we extended our experiments to 3 multilingual datasets covering 10 languages. More generally, we aim to propose a new evaluation paradigm in NLP that relies less on static, pre-defined benchmarks but shifts towards dynamic, stochastic evaluation leveraging eval-time compute.
☆ Scaling LLM Multi-turn RL with End-to-end Summarization-based Context Management
We study reinforcement learning (RL) fine-tuning of large language model (LLM) agents for long-horizon multi-turn tool use, where context length quickly becomes a fundamental bottleneck. Existing RL pipelines can suffer from degraded instruction following, excessive rollout costs, and most importantly, strict context limits. To address these challenges, we introduce summarization-based context management to training. In specific, it periodically compresses the tool using history by LLM-generated summaries that retain task-relevant information to keep a compact context while enabling the agent to scale beyond the fixed context window. Building on this formulation, we derive a policy gradient representation that seamlessly enables standard LLM RL infrastructures to optimize both tool-use behaviors as well as summarization strategies in an end-to-end fashion. We instantiate this framework with \underline{SU}mmarization augmented \underline{P}olicy \underline{O}ptimization (\texttt{SUPO}), an LLM RL algorithm that enables long-horizon training beyond a fixed context limit. Experiments on interactive function calling and searching tasks demonstrate that \texttt{SUPO} significantly improves the success rate while maintaining the same or even lower working context length compared to baselines. We also demonstrate that for complex searching tasks, \texttt{SUPO} can further improve the evaluation performance when scaling test-time maximum round of summarization beyond that of training time. Our results establish summarization-based context management as a principled and scalable approach for training RL agents beyond a fixed context length limit.
☆ Differentially Private Synthetic Text Generation for Retrieval-Augmented Generation (RAG)
Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by grounding them in external knowledge. However, its application in sensitive domains is limited by privacy risks. Existing private RAG methods typically rely on query-time differential privacy (DP), which requires repeated noise injection and leads to accumulated privacy loss. To address this issue, we propose DP-SynRAG, a framework that uses LLMs to generate differentially private synthetic RAG databases. Unlike prior methods, the synthetic text can be reused once created, thereby avoiding repeated noise injection and additional privacy costs. To preserve essential information for downstream RAG tasks, DP-SynRAG extends private prediction, which instructs LLMs to generate text that mimics subsampled database records in a DP manner. Experiments show that DP-SynRAG achieves superior performanec to the state-of-the-art private RAG systems while maintaining a fixed privacy budget, offering a scalable solution for privacy-preserving RAG.
comment: Under review
☆ XLSR-Kanformer: A KAN-Intergrated model for Synthetic Speech Detection
Recent advancements in speech synthesis technologies have led to increasingly sophisticated spoofing attacks, posing significant challenges for automatic speaker verification systems. While systems based on self-supervised learning (SSL) models, particularly the XLSR-Conformer architecture, have demonstrated remarkable performance in synthetic speech detection, there remains room for architectural improvements. In this paper, we propose a novel approach that replaces the traditional Multi-Layer Perceptron (MLP) in the XLSR-Conformer model with a Kolmogorov-Arnold Network (KAN), a powerful universal approximator based on the Kolmogorov-Arnold representation theorem. Our experimental results on ASVspoof2021 demonstrate that the integration of KAN to XLSR-Conformer model can improve the performance by 60.55% relatively in Equal Error Rate (EER) LA and DF sets, further achieving 0.70% EER on the 21LA set. Besides, the proposed replacement is also robust to various SSL architectures. These findings suggest that incorporating KAN into SSL-based models is a promising direction for advances in synthetic speech detection.
comment: Accepted to 2025 IEEE International Conference on Advanced Video and Signal-Based Surveillance
☆ How Language Models Conflate Logical Validity with Plausibility: A Representational Analysis of Content Effects
Both humans and large language models (LLMs) exhibit content effects: biases in which the plausibility of the semantic content of a reasoning problem influences judgments regarding its logical validity. While this phenomenon in humans is best explained by the dual-process theory of reasoning, the mechanisms behind content effects in LLMs remain unclear. In this work, we address this issue by investigating how LLMs encode the concepts of validity and plausibility within their internal representations. We show that both concepts are linearly represented and strongly aligned in representational geometry, leading models to conflate plausibility with validity. Using steering vectors, we demonstrate that plausibility vectors can causally bias validity judgements, and vice versa, and that the degree of alignment between these two concepts predicts the magnitude of behavioral content effects across models. Finally, we construct debiasing vectors that disentangle these concepts, reducing content effects and improving reasoning accuracy. Our findings advance understanding of how abstract logical concepts are represented in LLMs and highlight representational interventions as a path toward more logical systems.
☆ Learning to Rewrite Prompts for Bootstrapping LLMs on Downstream Tasks
In recent years, the growing interest in Large Language Models (LLMs) has significantly advanced prompt engineering, transitioning from manual design to model-based optimization. Prompts for LLMs generally comprise two components: the \textit{instruction}, which defines the task or objective, and the \textit{input}, which is tailored to the instruction type. In natural language generation (NLG) tasks such as machine translation, the \textit{input} component is particularly critical, while the \textit{instruction} component tends to be concise. Existing prompt engineering methods primarily focus on optimizing the \textit{instruction} component for general tasks, often requiring large-parameter LLMs as auxiliary tools. However, these approaches exhibit limited applicability for tasks like machine translation, where the \textit{input} component plays a more pivotal role. To address this limitation, this paper introduces a novel prompt optimization method specifically designed for machine translation tasks. The proposed approach employs a small-parameter model trained using a back-translation-based strategy, significantly reducing training overhead for single-task optimization while delivering highly effective performance. With certain adaptations, this method can also be extended to other downstream tasks.
☆ Incremental Summarization for Customer Support via Progressive Note-Taking and Agent Feedback EMNLP 2025
We introduce an incremental summarization system for customer support agents that intelligently determines when to generate concise bullet notes during conversations, reducing agents' context-switching effort and redundant review. Our approach combines a fine-tuned Mixtral-8x7B model for continuous note generation with a DeBERTa-based classifier to filter trivial content. Agent edits refine the online notes generation and regularly inform offline model retraining, closing the agent edits feedback loop. Deployed in production, our system achieved a 3% reduction in case handling time compared to bulk summarization (with reductions of up to 9% in highly complex cases), alongside high agent satisfaction ratings from surveys. These results demonstrate that incremental summarization with continuous feedback effectively enhances summary quality and agent productivity at scale.
comment: Accepted at EMNLP 2025 Industry Track
☆ PIKA: Expert-Level Synthetic Datasets for Post-Training Alignment from Scratch
Reinforcement Learning from Human Feedback (RLHF) has become a cornerstone for aligning large language models (LLMs). However, its effectiveness depends on high-quality instruction data. Most existing alignment datasets are either private or require costly human annotation, which limits reproducibility and scalability. Even with Reinforcement Learning from AI Feedback (RLAIF), concerns about data quality remain. Moreover, it is unclear how much data is actually required to fine-tune a base model into a strong instruction-following model. Current approaches often rely on over 300k examples even at the supervised fine-tuning (SFT) stage, yet they still underperform compared to proprietary models, creating barriers for academic and resource-limited communities. To address this gap, we introduce PiKa, a data-efficient family of expert-level alignment datasets. In particular, the PiKa-SFT dataset uses only 30k SFT examples, far fewer than state-of-the-art datasets like Magpie. Through evaluations by fine-tuning Llama-3-8B-Base on PiKa and other public datasets, we show that PiKa-SFT outperforms models trained on much larger data. On AlpacaEval 2.0 and Arena-Hard benchmarks, PiKa-SFT fine-tuning even surpasses the official Llama-3-8B-Instruct model trained on over 10 million proprietary examples. We further extend our study by training the Qwen2.5 series (0.5B to 7B) on PiKa-SFT, achieving consistent gains. These findings demonstrate that high-quality alignment can be achieved with significantly less data, offering a scalable path for open-source LLM alignment. Code and data: https://github.com/SJY8460/PiKa.
☆ ToolMem: Enhancing Multimodal Agents with Learnable Tool Capability Memory
Agents utilizing tools powered by large language models (LLMs) or vision-language models (VLMs) have demonstrated remarkable progress in diverse tasks across text and visual modalities. Unlike traditional tools such as calculators, which give deterministic outputs, neural tools perform uncertainly across task scenarios. While different tools for a task may excel in varied scenarios, existing agents typically rely on fixed tools, thus limiting the flexibility in selecting the most suitable tool for specific tasks. In contrast, humans snowball their understanding of the capabilities of different tools by interacting with them, and apply this knowledge to select the optimal tool when solving a future task. To build agents that similarly benefit from this process, we propose ToolMem that enables agents to develop memories of tool capabilities from previous interactions, by summarizing their strengths and weaknesses and storing them in memory; at inference, the agent can retrieve relevant entries from ToolMem, and select the best tool to solve individual tasks more accurately. We evaluate ToolMem on learning varied text generation and text-to-image generation neural tools. Compared to no-memory, generic agents, we find ToolMem-augmented agents predict tool performance 14.8% and 28.7% more accurately across text and multimodal generation scenarios. Moreover, ToolMem facilitates optimal tool selection among multiple choices by 21% and 24% absolute increases in respective scenarios.
☆ Aligning Large Language Models via Fully Self-Synthetic Data
Traditional reinforcement learning from human feedback (RLHF) for large language models (LLMs) relies on expensive human-annotated datasets, while Reinforcement Learning from AI Feedback (RLAIF) also incurs significant costs, requiring the collection of diverse prompts and corresponding responses, often necessitating external reward models or proprietary models like GPT-4 to annotate preference pairs. In this work, we introduce Self-Alignment Optimization (SAO), a fully self-synthetic framework for LLM alignment, where all training data, including prompts (i.e., user queries), responses, and preferences, are generated by the model itself. Specifically, SAO first instructs the LLM to engage in persona role-play and generate diverse prompts and responses, which are then self-evaluated for preference optimization. Extensive experiments demonstrate that SAO effectively enhances the model's chat capabilities on standard benchmarks like AlpacaEval~2.0, while maintaining strong performance on downstream objective tasks (e.g., question-answering, math reasoning). Our work provides a practical solution for self-improvement in aligning LLMs, and the code for reproducing our results is available at: https://github.com/SJY8460/SAO.
☆ A Comparative Analysis of Contextual Representation Flow in State-Space and Transformer Architectures
State Space Models (SSMs) have recently emerged as efficient alternatives to Transformer-Based Models (TBMs) for long-sequence processing, offering linear scaling and lower memory use. Yet, how contextual information flows across layers and tokens in these architectures remains understudied. We present the first unified, token- and layer-level analysis of representation propagation in SSMs and TBMs. Using centered kernel alignment, stability metrics, and probing, we characterize how representations evolve within and across layers. We find a key divergence: TBMs rapidly homogenize token representations, with diversity reemerging only in later layers, while SSMs preserve token uniqueness early but converge to homogenization deeper. Theoretical analysis and parameter randomization further reveal that oversmoothing in TBMs stems from architectural design, whereas in SSMs it arises mainly from training dynamics. These insights clarify the inductive biases of both architectures and inform future model and training designs for long-context reasoning.
☆ Reading Between the Lines: Towards Reliable Black-box LLM Fingerprinting via Zeroth-order Gradient Estimation
The substantial investment required to develop Large Language Models (LLMs) makes them valuable intellectual property, raising significant concerns about copyright protection. LLM fingerprinting has emerged as a key technique to address this, which aims to verify a model's origin by extracting an intrinsic, unique signature (a "fingerprint") and comparing it to that of a source model to identify illicit copies. However, existing black-box fingerprinting methods often fail to generate distinctive LLM fingerprints. This ineffectiveness arises because black-box methods typically rely on model outputs, which lose critical information about the model's unique parameters due to the usage of non-linear functions. To address this, we first leverage Fisher Information Theory to formally demonstrate that the gradient of the model's input is a more informative feature for fingerprinting than the output. Based on this insight, we propose ZeroPrint, a novel method that approximates these information-rich gradients in a black-box setting using zeroth-order estimation. ZeroPrint overcomes the challenge of applying this to discrete text by simulating input perturbations via semantic-preserving word substitutions. This operation allows ZeroPrint to estimate the model's Jacobian matrix as a unique fingerprint. Experiments on the standard benchmark show ZeroPrint achieves a state-of-the-art effectiveness and robustness, significantly outperforming existing black-box methods.
☆ Do Internal Layers of LLMs Reveal Patterns for Jailbreak Detection?
Jailbreaking large language models (LLMs) has emerged as a pressing concern with the increasing prevalence and accessibility of conversational LLMs. Adversarial users often exploit these models through carefully engineered prompts to elicit restricted or sensitive outputs, a strategy widely referred to as jailbreaking. While numerous defense mechanisms have been proposed, attackers continuously develop novel prompting techniques, and no existing model can be considered fully resistant. In this study, we investigate the jailbreak phenomenon by examining the internal representations of LLMs, with a focus on how hidden layers respond to jailbreak versus benign prompts. Specifically, we analyze the open-source LLM GPT-J and the state-space model Mamba2, presenting preliminary findings that highlight distinct layer-wise behaviors. Our results suggest promising directions for further research on leveraging internal model dynamics for robust jailbreak detection and defense.
☆ TinyScientist: An Interactive, Extensible, and Controllable Framework for Building Research Agents EMNLP 2025
Automatic research with Large Language Models (LLMs) is rapidly gaining importance, driving the development of increasingly complex workflows involving multi-agent systems, planning, tool usage, code execution, and human-agent interaction to accelerate research processes. However, as more researchers and developers begin to use and build upon these tools and platforms, the complexity and difficulty of extending and maintaining such agentic workflows have become a significant challenge, particularly as algorithms and architectures continue to advance. To address this growing complexity, TinyScientist identifies the essential components of the automatic research workflow and proposes an interactive, extensible, and controllable framework that easily adapts to new tools and supports iterative growth. We provide an open-source codebase, an interactive web demonstration, and a PyPI Python package to make state-of-the-art auto-research pipelines broadly accessible to every researcher and developer.
comment: 7 pages, EMNLP 2025 Demo track
♻ ☆ DESIGNER: Design-Logic-Guided Multidisciplinary Data Synthesis for LLM Reasoning
Large language models (LLMs) have achieved remarkable success in many natural language tasks but still struggle with complex, multi-step reasoning, particularly across diverse disciplines. Existing reasoning datasets often lack disciplinary breadth, reasoning depth, and diversity, and lack guiding principles for question synthesis. We propose DESIGNER: a DESIGN-logic-guidEd Reasoning data synthesis pipeline that leverages naturally available, extensive raw documents (e.g., book corpus and web corpus) to generate multidisciplinary challenging questions. We introduce the concept of "design logic" and instruct LLMs to mimic human educators' question-creation process, enabling automated synthesis of large-scale, high-difficulty questions. We use LLMs to reverse-engineer and abstract over 120,000 design logics from existing questions across various disciplines. By matching these design logics with source documents, we are able to create reasoning questions that far surpass the difficulty and diversity of existing datasets. Using this pipeline, we synthesized two large-scale reasoning datasets that span 75 disciplines: DLR-Book (3.04 million questions from the book corpus) and DLR-Web (1.66 million questions from the web corpus). Data analysis indicates that the questions synthesized by our method exhibit greater difficulty and diversity compared to those in the baseline datasets. We validate our synthesized data through supervised fine-tuning (SFT) on the Qwen3 and Llama3 model families. Our data substantially enhances their multidisciplinary reasoning capabilities, outperforming existing datasets. Notably, after SFT on our datasets, the base versions of these models even surpass their official instruction-tuned counterparts.
♻ ☆ LatteReview: A Multi-Agent Framework for Systematic Review Automation Using Large Language Models
Systematic literature reviews and meta-analyses are essential for synthesizing research insights, but they remain time-intensive and labor-intensive due to the iterative processes of screening, evaluation, and data extraction. This paper introduces and evaluates LatteReview, a Python-based framework that leverages large language models (LLMs) and multi-agent systems to automate key elements of the systematic review process. Designed to streamline workflows while maintaining rigor, LatteReview utilizes modular agents for tasks such as title and abstract screening, relevance scoring, and structured data extraction. These agents operate within orchestrated workflows, supporting sequential and parallel review rounds, dynamic decision-making, and iterative refinement based on user feedback. LatteReview's architecture integrates LLM providers, enabling compatibility with both cloud-based and locally hosted models. The framework supports features such as Retrieval-Augmented Generation (RAG) for incorporating external context, multimodal reviews, Pydantic-based validation for structured inputs and outputs, and asynchronous programming for handling large-scale datasets. The framework is available on the GitHub repository, with detailed documentation and an installable package.
comment: 31 pages, 5 figures, 5 tables
♻ ☆ Dyna-Think: Synergizing Reasoning, Acting, and World Model Simulation in AI Agents
Recent progress in reasoning with large language models (LLMs), such as DeepSeek-R1, demonstrates impressive capabilities in domains like mathematics and coding, by exhibiting complex cognitive behaviors such as verification, goal decomposition, and self-reflection. However, it is unclear what behavior is effective and what behavior is missing for long-horizon AI agents tasks. In this work, we propose Dyna-Think, a thinking framework that integrates planning with an internal world model with reasoning and acting to enhance AI agent performance. To enable Dyna-Think, we propose Dyna-Think Imitation Learning (DIT) and Dyna-Think Dyna Training (DDT). To initialize a policy with Dyna-Think, DIT reconstructs the thinking process of R1 to focus on performing world model simulation relevant to the proposed (and planned) action, and trains the policy using this reconstructed data. To enhance Dyna-Think, DDT uses a two-stage training process to first improve the agent's world modeling ability via objectives such as state prediction or critique generation, and then improve the agent's action via policy training. We evaluate our methods on OSWorld and WindowsAgentArena, and demonstrate that Dyna-Think improves the agent's in-domain and out-of-domain performance, achieving similar best-of-n performance compared to R1 while generating 2x less tokens on average. Our extensive empirical studies reveal that 1) using critique generation for world model training is effective to improve policy performance; and 2) AI agents with better performance correlate with better world modeling abilities. We believe our results suggest a promising research direction to integrate world model simulation into AI agents to enhance their reasoning, planning, and acting capabilities.
♻ ☆ Speculative Decoding and Beyond: An In-Depth Survey of Techniques
Sequential dependencies present a fundamental bottleneck in deploying large-scale autoregressive models, particularly for real-time applications. While traditional optimization approaches like pruning and quantization often compromise model quality, recent advances in generation-refinement frameworks demonstrate that this trade-off can be significantly mitigated. This survey presents a comprehensive taxonomy of generation-refinement frameworks, analyzing methods across autoregressive sequence tasks. We categorize methods based on their generation strategies (from simple n-gram prediction to sophisticated draft models) and refinement mechanisms (including single-pass verification and iterative approaches). Through systematic analysis of both algorithmic innovations and system-level implementations, we examine deployment strategies across computing environments and explore applications spanning text, images, and speech generation. This systematic examination of both theoretical frameworks and practical implementations provides a foundation for future research in efficient autoregressive decoding.
♻ ☆ MCTS-RAG: Enhancing Retrieval-Augmented Generation with Monte Carlo Tree Search
We introduce MCTS-RAG, a novel approach that enhances the reasoning capabilities of small language models on knowledge-intensive tasks by leveraging retrieval-augmented generation (RAG) to provide relevant context and Monte Carlo Tree Search (MCTS) to refine reasoning paths. MCTS-RAG dynamically integrates retrieval and reasoning through an iterative decision-making process. Unlike standard RAG methods, which typically retrieve information independently from reasoning and thus integrate knowledge suboptimally, or conventional MCTS reasoning, which depends solely on internal model knowledge without external facts, MCTS-RAG combines structured reasoning with adaptive retrieval. This integrated approach enhances decision-making, reduces hallucinations, and ensures improved factual accuracy and response consistency. The experimental results on multiple reasoning and knowledge-intensive datasets datasets (i.e., ComplexWebQA, GPQA, and FoolMeTwice) show that our method enables small-scale LMs to achieve performance comparable to frontier LLMs like GPT-4o by effectively scaling inference-time compute, setting a new standard for reasoning in small-scale models.
♻ ☆ ParamBench: A Graduate-Level Benchmark for Evaluating LLM Understanding on Indic Subjects
Large language models have been widely evaluated on tasks such as comprehension, summarization, code generation, etc. However, their performance on graduate-level, culturally grounded questions in the Indian context remains largely unexplored. Existing Indian benchmarks emphasise basic fact-orientated queries that offer limited assessment of a deeper disciplinary understanding tailored to the Indian setting. In this paper, we present ParamBench, consisting of more than 17K questions in the Hindi language, comprising questionnaires from 21 diverse subjects. These questions are primarily derived from a nationwide graduate-level entrance examination covering topics such as history, music, instruments, yoga, literature, philosophy, law, etc.~ specifically for the Indian context. Additionally, we assess the ability of LLMs to handle diverse question formats - such as list-based matching, assertion-reason pairs, and sequence ordering - alongside conventional multiple-choice questions. We evaluated the performance of more than 16 open source LLMs on this benchmark, observing that Gemma3-27B attains the highest overall accuracy of 56.4\%. Furthermore, subject-wise analysis indicates that even for the best-performing LLMs, performance remains weak on topics such as music, classical instruments, and law, underscoring persistent challenges in culturally grounded reasoning. The dataset and source code is present at https://github.com/ayushbits/ParamBench.
♻ ☆ LLMVA-GEBC: Large Language Model with Video Adapter for Generic Event Boundary Captioning CVPR 2023
Our winning entry for the CVPR 2023 Generic Event Boundary Captioning (GEBC) competition is detailed in this paper. Unlike conventional video captioning tasks, GEBC demands that the captioning model possess an understanding of immediate changes in status around the designated video boundary, making it a difficult task. This paper proposes an effective model LLMVA-GEBC (Large Language Model with Video Adapter for Generic Event Boundary Captioning): (1) We utilize a pretrained LLM for generating human-like captions with high quality. (2) To adapt the model to the GEBC task, we take the video Q-former as an adapter and train it with the frozen visual feature extractors and LLM. Our proposed method achieved a 76.14 score on the test set and won the first place in the challenge. Our code is available at https://github.com/zjr2000/LLMVA-GEBC .
comment: Winner solution to Generic Event Boundary Captioning task in LOVEU Challenge (CVPR 2023 workshop)
♻ ☆ LaunchpadGPT: Language Model as Music Visualization Designer on Launchpad
Launchpad is a musical instrument that allows users to create and perform music by pressing illuminated buttons. To assist and inspire the design of the Launchpad light effect, and provide a more accessible approach for beginners to create music visualization with this instrument, we proposed the LaunchpadGPT model to generate music visualization designs on Launchpad automatically. Based on the language model with excellent generation ability, our proposed LaunchpadGPT takes an audio piece of music as input and outputs the lighting effects of Launchpad-playing in the form of a video (Launchpad-playing video). We collect Launchpad-playing videos and process them to obtain music and corresponding video frame of Launchpad-playing as prompt-completion pairs, to train the language model. The experiment result shows the proposed method can create better music visualization than random generation methods and hold the potential for a broader range of music visualization applications. Our code is available at https://github.com/yunlong10/LaunchpadGPT/.
comment: Accepted to International Computer Music Conference (ICMC) 2023
♻ ☆ AutoMind: Adaptive Knowledgeable Agent for Automated Data Science
Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems. LLM-driven data science agents promise to automate the entire machine learning pipeline, yet their real-world effectiveness remains limited. Existing frameworks depend on rigid, pre-defined workflows and inflexible coding strategies; consequently, they excel only on relatively simple, classical problems and fail to capture the empirical expertise that human practitioners bring to complex, innovative tasks. In this work, we introduce AutoMind, an adaptive, knowledgeable LLM-agent framework that overcomes these deficiencies through three key advances: (1) a curated expert knowledge base that grounds the agent in domain expert knowledge, (2) an agentic knowledgeable tree search algorithm that strategically explores possible solutions, and (3) a self-adaptive coding strategy that dynamically tailors code generation to task complexity. Evaluations on two automated data science benchmarks demonstrate that AutoMind delivers superior performance versus state-of-the-art baselines. Additional analyses confirm favorable effectiveness, efficiency, and qualitative solution quality, highlighting AutoMind as an efficient and robust step toward fully automated data science. Code is at https://github.com/innovatingAI/AutoMind.
comment: Ongoing work
♻ ☆ Prefilled responses enhance zero-shot detection of AI-generated images
As AI models generate increasingly realistic images, growing concerns over potential misuse underscore the need for reliable detection. Traditional supervised detection methods depend on large, curated datasets for training and often fail to generalize to novel, out-of-domain image generators. As an alternative, we explore pre-trained Vision-Language Models (VLMs) for zero-shot detection of AI-generated images. We evaluate VLM performance on three diverse benchmarks encompassing synthetic images of human faces, objects, and animals produced by 16 different state-of-the-art image generators. While off-the-shelf VLMs perform poorly on these datasets, we find that their reasoning can be guided effectively through simple response prefilling -- a method we call Prefill-Guided Thinking (PGT). In particular, prefilling a VLM response with the task-aligned phrase "Let's examine the style and the synthesis artifacts" improves the Macro F1 scores of three widely used open-source VLMs by up to 24%.
♻ ☆ KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality
Large Language Models (LLMs), particularly slow-thinking models, often exhibit severe hallucination, outputting incorrect content due to an inability to accurately recognize knowledge boundaries during reasoning. While Reinforcement Learning (RL) can enhance complex reasoning abilities, its outcome-oriented reward mechanism often lacks factual supervision over the thinking process, further exacerbating the hallucination problem. To address the high hallucination in slow-thinking models, we propose Knowledge-enhanced RL, KnowRL. KnowRL guides models to perform fact-based slow thinking by integrating a factuality reward, based on knowledge verification, into the RL training process, helping them recognize their knowledge boundaries. KnowRL guides models to perform fact-based slow thinking by integrating a factuality reward, based on knowledge verification, into the RL training process, helping them recognize their knowledge boundaries. This targeted factual input during RL training enables the model to learn and internalize fact-based reasoning strategies. By directly rewarding adherence to facts within the reasoning steps, KnowRL fosters a more reliable thinking process. Experimental results on three hallucination evaluation datasets and two reasoning evaluation datasets demonstrate that KnowRL effectively mitigates hallucinations in slow-thinking models while maintaining their original strong reasoning capabilities. Our code is available at https://github.com/zjunlp/KnowRL.
comment: Work in progress
Diagnosing Moral Reasoning Acquisition in Language Models: Pragmatics and Generalization
Ensuring that Large Language Models (LLMs) return just responses which adhere to societal values is crucial for their broader application. Prior research has shown that LLMs often fail to perform satisfactorily on tasks requiring moral cognizance, such as ethics-based judgments. While current approaches have focused on fine-tuning LLMs with curated datasets to improve their capabilities on such tasks, choosing the optimal learning paradigm to enhance the ethical responses of LLMs remains an open research debate. In this work, we aim to address this fundamental question: can current learning paradigms enable LLMs to acquire sufficient moral reasoning capabilities? Drawing from distributional semantics theory and the pragmatic nature of moral discourse, our analysis indicates that performance improvements follow a mechanism similar to that of semantic-level tasks, and therefore remain affected by the pragmatic nature of morals latent in discourse, a phenomenon we name the pragmatic dilemma. We conclude that this pragmatic dilemma imposes significant limitations on the generalization ability of current learning paradigms, making it the primary bottleneck for moral reasoning acquisition in LLMs.
♻ ☆ Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning
Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of NLP tasks, but they remain fundamentally stateless, constrained by limited context windows that hinder long-horizon reasoning. Recent efforts to address this limitation often augment LLMs with an external memory bank, yet most existing pipelines are static and heuristic-driven, lacking a learned mechanism for deciding what to store, update, or retrieve. We present Memory-R1, a reinforcement learning (RL) framework that equips LLMs with the ability to actively manage and utilize external memory through two specialized agents: a Memory Manager that learns structured operations, including ADD, UPDATE, DELETE, and NOOP; and an Answer Agent that pre-selects and reasons over relevant entries. Both agents are fine-tuned with outcome-driven RL (PPO and GRPO), enabling adaptive memory management with minimal supervision. With only 152 training QA pairs, Memory-R1 outperforms strong baselines and generalizes across diverse question types, three benchmarks (LoCoMo, MSC, LongMemEval), and multiple model scales (3B-14B).
♻ ☆ SimpleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Trajectory Synthesis
Retrieval-augmented generation (RAG) systems have advanced large language models (LLMs) in complex deep search scenarios requiring multi-step reasoning and iterative information retrieval. However, existing approaches face critical limitations that lack high-quality training trajectories or suffer from the distributional mismatches in simulated environments and prohibitive computational costs for real-world deployment. This paper introduces SimpleDeepSearcher, a lightweight yet effective framework that bridges this gap through strategic data engineering rather than complex training paradigms. Our approach synthesizes high-quality training data by simulating realistic user interactions in live web search environments, coupled with a multi-criteria curation strategy that optimizes the diversity and quality of input and output side. Experiments on five benchmarks across diverse domains demonstrate that SFT on only 871 curated samples yields significant improvements over RL-based baselines. Our work establishes SFT as a viable pathway by systematically addressing the data-scarce bottleneck, offering practical insights for efficient deep search systems. Our code is available at https://github.com/RUCAIBox/SimpleDeepSearcher.
♻ ☆ Video Understanding with Large Language Models: A Survey
With the burgeoning growth of online video platforms and the escalating volume of video content, the demand for proficient video understanding tools has intensified markedly. Given the remarkable capabilities of large language models (LLMs) in language and multimodal tasks, this survey provides a detailed overview of recent advancements in video understanding that harness the power of LLMs (Vid-LLMs). The emergent capabilities of Vid-LLMs are surprisingly advanced, particularly their ability for open-ended multi-granularity (general, temporal, and spatiotemporal) reasoning combined with commonsense knowledge, suggesting a promising path for future video understanding. We examine the unique characteristics and capabilities of Vid-LLMs, categorizing the approaches into three main types: Video Analyzer x LLM, Video Embedder x LLM, and (Analyzer + Embedder) x LLM. Furthermore, we identify five sub-types based on the functions of LLMs in Vid-LLMs: LLM as Summarizer, LLM as Manager, LLM as Text Decoder, LLM as Regressor, and LLM as Hidden Layer. Furthermore, this survey presents a comprehensive study of the tasks, datasets, benchmarks, and evaluation methodologies for Vid-LLMs. Additionally, it explores the expansive applications of Vid-LLMs across various domains, highlighting their remarkable scalability and versatility in real-world video understanding challenges. Finally, it summarizes the limitations of existing Vid-LLMs and outlines directions for future research. For more information, readers are recommended to visit the repository at https://github.com/yunlong10/Awesome-LLMs-for-Video-Understanding.
comment: Accepted to IEEE Transactions on Circuits and Systems for Video Technology (TCSVT)
♻ ☆ From Injection to Defense: Constructing Edit-Based Fingerprints for Large Language Models
Fingerprinting is critical for maintaining traceability and protecting the intellectual property (IP) of developers, as LLMs deployed in web applications are susceptible to unauthorized redistribution and misuse via fine-tuning or black-box deployment. However, current backdoor-based fingerprinting methods face a fundamental trade-off: fingerprints embedded as garbled text are easily detected and filtered, whereas those crafted as coherent natural language are prone to being triggered unintentionally. To overcome these limitations, we propose RFEdit, a knowledge-editing framework that embeds a rule-based multilingual natural language fingerprint (MNLF) by modifying a sparse subset of model weights. This approach enables efficient and robust fingerprint injection with minimal impact on unrelated knowledge in LLMs. Our RFEdit framework is further safeguarded by Fingerprint Subspace-aware Fine-Tuning (FSFT), which mitigates fingerprint degradation during legitimate fine-tuning by restricting parameter updates to the fingerprint subspace. This approach preserves fingerprint integrity while enhancing downstream task performance of LLMs. These advances establish a comprehensive pipeline from fingerprint injection to defense, achieving high detection effectiveness, robustness against adversarial manipulations, harmlessness to model utility, and persistence under fine-tuning. Extensive experiments demonstrate that RFEdit maintains robustness under quantization and pruning. Additionally, fingerprint effectiveness is generally improved by more than 10\% when combined with FSFT for math and alpaca downstream tasks.
comment: preprint
♻ ☆ 360-LLaMA-Factory: Plug & Play Sequence Parallelism for Long Post-Training
Adding sequence parallelism into LLaMA-Factory, we open-sourced 360-LLaMA-Factory at https://github.com/Qihoo360/360-LLaMA-Factory. 360-LLaMA-Factory has received wide recognition and used in models such as Light-R1 arXiv:2503.10460, TinyR1 arXiv:2503.04872, Kaggle AIMO math models and also in large companies' training frameworks. This technical report delves deeper into the different sequence parallel modes behind 360-LLaMA-Factory and discusses our implementation insights.
comment: v2: sp for vlms; code at https://github.com/Qihoo360/360-LLaMA-Factory
♻ ☆ Do LLMs Overthink Basic Math Reasoning? Benchmarking the Accuracy-Efficiency Tradeoff in Language Models
Large language models (LLMs) achieve impressive performance on complex mathematical benchmarks yet sometimes fail on basic math reasoning while generating unnecessarily verbose responses. In this paper, we present a systematic benchmark and comprehensive empirical study to evaluate the efficiency of reasoning in LLMs, focusing on the fundamental tradeoff between accuracy and overthinking. First, we formalize the accuracy-verbosity tradeoff. Second, we introduce the Overthinking Score, a harmonic-mean metric combining accuracy and token-efficiency for holistic model evaluation. Third, we establish an evaluation protocol with dynamically-generated data across 14 basic math tasks. Fourth, we conduct a large-scale empirical study evaluating 53 LLMs, including reasoning and quantized variants across different reasoning budgets. Our findings reveal: 1) model performance on complex benchmarks does not translate directly to basic math reasoning; 2) reasoning models generate ~18 more tokens while sometimes achieving lower accuracy and exhibit catastrophic collapse when token is constrained, dropping by ~28; 3) the accuracy-verbosity relationship is non-monotonic with extended reasoning budgets yielding diminishing returns (GPT-5/o-series models show zero accuracy gain from low -> medium -> high reasoning effort). Our findings challenge the assumption that longer reasoning in LLMs necessarily improves mathematical reasoning.
♻ ☆ The Sound of Syntax: Finetuning and Comprehensive Evaluation of Language Models for Speech Pathology EMNLP 2025
According to the U.S. National Institutes of Health, more than 3.4 million children experience speech disorders that require clinical intervention. The number of speech-language pathologists (SLPs) is roughly 20 times fewer than the number of affected children, highlighting a significant gap in children's care and a pressing need for technological support that improves the productivity of SLPs. State-of-the-art multimodal language models (MLMs) show promise for supporting SLPs, but their use remains underexplored largely due to a limited understanding of their performance in high-stakes clinical settings. To address this gap, we collaborate with domain experts to develop a taxonomy of real-world use cases of MLMs in speech-language pathologies. Building on this taxonomy, we introduce the first comprehensive benchmark for evaluating MLM across five core use cases, each containing 1,000 manually annotated data points. This benchmark includes robustness and sensitivity tests under various settings, including background noise, speaker gender, and accent. Our evaluation of 15 state-of-the-art MLMs reveals that no single model consistently outperforms others across all tasks. Notably, we find systematic disparities, with models performing better on male speakers, and observe that chain-of-thought prompting can degrade performance on classification tasks with large label spaces and narrow decision boundaries. Furthermore, we study fine-tuning MLMs on domain-specific data, achieving improvements of over 10\% compared to base models. These findings highlight both the potential and limitations of current MLMs for speech-language pathology applications, underscoring the need for further research and targeted development.
comment: EMNLP 2025 Oral Presentation
♻ ☆ Controlled Agentic Planning & Reasoning for Mechanism Synthesis
This work presents a dual-agent \ac{llm}-based reasoning framework for automated planar mechanism synthesis that tightly couples linguistic specification with symbolic representation and simulation. From a natural-language task description, the system composes symbolic constraints and equations, generates and parametrises simulation code, and iteratively refines designs via critic-driven feedback, including symbolic regression and geometric distance metrics, closing an actionable linguistic/symbolic optimisation loop. To evaluate the approach, we introduce MSynth, a benchmark of analytically defined planar trajectories. Empirically, critic feedback and iterative refinement yield large improvements (up to 90\% on individual tasks) and statistically significant gains per the Wilcoxon signed-rank test. Symbolic-regression prompts provide deeper mechanistic insight primarily when paired with larger models or architectures with appropriate inductive biases (e.g., LRM).
comment: 24 pages, 16 figures
♻ ☆ Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models NeurIPS
Vision-language models (VLMs) trained on internet-scale data achieve remarkable zero-shot detection performance on common objects like car, truck, and pedestrian. However, state-of-the-art models still struggle to generalize to out-of-distribution classes, tasks and imaging modalities not typically found in their pre-training. Rather than simply re-training VLMs on more visual data, we argue that one should align VLMs to new concepts with annotation instructions containing a few visual examples and rich textual descriptions. To this end, we introduce Roboflow100-VL, a large-scale collection of 100 multi-modal object detection datasets with diverse concepts not commonly found in VLM pre-training. We evaluate state-of-the-art models on our benchmark in zero-shot, few-shot, semi-supervised, and fully-supervised settings, allowing for comparison across data regimes. Notably, we find that VLMs like GroundingDINO and Qwen2.5-VL achieve less than 2% zero-shot accuracy on challenging medical imaging datasets within Roboflow100-VL, demonstrating the need for few-shot concept alignment. Lastly, we discuss our recent CVPR 2025 Foundational FSOD competition and share insights from the community. Notably, the winning team significantly outperforms our baseline by 17 mAP! Our code and dataset are available at https://github.com/roboflow/rf100-vl and https://universe.roboflow.com/rf100-vl/.
comment: The first two authors contributed equally. This work has been accepted to the Neural Information Processing Systems (NeurIPS) 2025 Datasets & Benchmark Track. Project Page: https://rf100-vl.org/
♻ ☆ MIST: Towards Multi-dimensional Implicit BiaS Evaluation of LLMs via Theory of Mind
Theory of Mind (ToM) in Large Language Models (LLMs) refers to their capacity for reasoning about mental states, yet failures in this capacity often manifest as systematic implicit bias. Evaluating this bias is challenging, as conventional direct-query methods are susceptible to social desirability effects and fail to capture its subtle, multi-dimensional nature. To this end, we propose an evaluation framework that leverages the Stereotype Content Model (SCM) to reconceptualize bias as a multi-dimensional failure in ToM across Competence, Sociability, and Morality. The framework introduces two indirect tasks: the Word Association Bias Test (WABT) to assess implicit lexical associations and the Affective Attribution Test (AAT) to measure covert affective leanings, both designed to probe latent stereotypes without triggering model avoidance. Extensive experiments on 8 State-of-the-Art LLMs demonstrate our framework's capacity to reveal complex bias structures, including pervasive sociability bias, multi-dimensional divergence, and asymmetric stereotype amplification, thereby providing a more robust methodology for identifying the structural nature of implicit bias.
♻ ☆ MMReview: A Multidisciplinary and Multimodal Benchmark for LLM-Based Peer Review Automation
With the rapid growth of academic publications, peer review has become an essential yet time-consuming responsibility within the research community. Large Language Models (LLMs) have increasingly been adopted to assist in the generation of review comments; however, current LLM-based review tasks lack a unified evaluation benchmark to rigorously assess the models' ability to produce comprehensive, accurate, and human-aligned assessments, particularly in scenarios involving multimodal content such as figures and tables. To address this gap, we propose \textbf{MMReview}, a comprehensive benchmark that spans multiple disciplines and modalities. MMReview includes multimodal content and expert-written review comments for 240 papers across 17 research domains within four major academic disciplines: Artificial Intelligence, Natural Sciences, Engineering Sciences, and Social Sciences. We design a total of 13 tasks grouped into four core categories, aimed at evaluating the performance of LLMs and Multimodal LLMs (MLLMs) in step-wise review generation, outcome formulation, alignment with human preferences, and robustness to adversarial input manipulation. Extensive experiments conducted on 16 open-source models and 5 advanced closed-source models demonstrate the thoroughness of the benchmark. We envision MMReview as a critical step toward establishing a standardized foundation for the development of automated peer review systems.
comment: Work in progress
♻ ☆ Transparent and Coherent Procedural Mistake Detection EMNLP 2025
Procedural mistake detection (PMD) is a challenging problem of classifying whether a human user (observed through egocentric video) has successfully executed a task (specified by a procedural text). Despite significant recent efforts, machine performance in the wild remains nonviable, and the reasoning processes underlying this performance are opaque. As such, we extend PMD to require generating visual self-dialog rationales to inform decisions. Given the impressive, mature image understanding capabilities observed in recent vision-and-language models (VLMs), we curate a suitable benchmark dataset for PMD based on individual frames. As our reformulation enables unprecedented transparency, we leverage a natural language inference (NLI) model to formulate two automated metrics for the coherence of generated rationales. We establish baselines for this reframed task, showing that VLMs struggle off-the-shelf, but with some trade-offs, their accuracy, coherence, and efficiency can be improved by incorporating these metrics into common inference and fine-tuning methods. Lastly, our multi-faceted metrics visualize common outcomes, highlighting areas for further improvement.
comment: EMNLP 2025 Camera Ready
♻ ☆ Blessing of Multilinguality: A Systematic Analysis of Multilingual In-Context Learning ACL 2025
While multilingual large language models generally perform adequately, and sometimes even rival English performance on high-resource languages (HRLs), they often significantly underperform on low-resource languages (LRLs). Among several prompting strategies aiming at bridging the gap, multilingual in-context learning (ICL) has been particularly effective when demonstration in target languages is unavailable. However, there lacks a systematic understanding of when and why it works well. In this work, we systematically analyze multilingual ICL, using demonstrations in HRLs to enhance cross-lingual transfer. We show that demonstrations in mixed HRLs consistently outperform English-only ones across the board, particularly for tasks written in LRLs. Surprisingly, our ablation study shows that the presence of irrelevant non-English sentences in the prompt yields measurable gains, suggesting the effectiveness of multilingual exposure itself. Our results highlight the potential of strategically leveraging multilingual resources to bridge the performance gap for underrepresented languages.
comment: ACL 2025 Findings
♻ ☆ LiTEx: A Linguistic Taxonomy of Explanations for Understanding Within-Label Variation in Natural Language Inference EMNLP 2025
There is increasing evidence of Human Label Variation (HLV) in Natural Language Inference (NLI), where annotators assign different labels to the same premise-hypothesis pair. However, within-label variation--cases where annotators agree on the same label but provide divergent reasoning--poses an additional and mostly overlooked challenge. Several NLI datasets contain highlighted words in the NLI item as explanations, but the same spans on the NLI item can be highlighted for different reasons, as evidenced by free-text explanations, which offer a window into annotators' reasoning. To systematically understand this problem and gain insight into the rationales behind NLI labels, we introduce LITEX, a linguistically-informed taxonomy for categorizing free-text explanations in English. Using this taxonomy, we annotate a subset of the e-SNLI dataset, validate the taxonomy's reliability, and analyze how it aligns with NLI labels, highlights, and explanations. We further assess the taxonomy's usefulness in explanation generation, demonstrating that conditioning generation on LITEX yields explanations that are linguistically closer to human explanations than those generated using only labels or highlights. Our approach thus not only captures within-label variation but also shows how taxonomy-guided generation for reasoning can bridge the gap between human and model explanations more effectively than existing strategies.
comment: Accepted by EMNLP 2025 Main, 22 pages, 7 figures
♻ ☆ Improving Neutral Point-of-View Generation with Data- and Parameter-Efficient RL
The paper shows that parameter-efficient reinforcement learning (PE-RL) is a highly effective training regime to improve large language models' (LLMs) ability to answer queries on sensitive topics with a Neutral Point of View (NPOV), i.e. to provide significantly more informative, diverse and impartial answers. This is shown by evaluating PE-RL and multiple strong baselines-including LoRA finetuning (strongest baseline), SFT and RLHF. PE-RL not only improves on overall NPOV quality compared to the strongest baseline ($97.06\%\rightarrow 99.08\%$), but also scores much higher on features linguists identify as key to separating sufficient answers from "great'' answers ($60.25\%\rightarrow 85.21\%$ for presence of supportive details, $68.74\%\rightarrow 91.43\%$ for absence of oversimplification). A qualitative analysis corroborates this. Moreover, our evaluation also finds a key property of PE-RL for this task: unlike methods that update all parameters, it generalises out of topic. Finally, to enable further studies we also release the dataset, SHQ-NPOV, and provide a methodology to create such datasets through iterative rounds of human peer-critique and annotator training.
♻ ☆ FedSRD: Sparsify-Reconstruct-Decompose for Communication-Efficient Federated Large Language Models Fine-Tuning
The current paradigm of training large language models (LLMs) on publicly available Web data is becoming unsustainable, with high-quality data sources in specialized domains nearing exhaustion. Federated Learning (FL) emerges as a practical solution for the next generation of AI on a decentralized Web, enabling privacy-preserving collaborative fine-tuning by leveraging private data distributed across a global client base. While Low-Rank Adaptation (LoRA) is the standard for efficient fine-tuning, its application in federated settings presents a critical challenge: communication overhead remains a significant bottleneck across the Web's heterogeneous network conditions. The structural redundancy within LoRA parameters not only incurs a heavy communication burden but also introduces conflicts when aggregating client updates. To address this, we propose FedSRD, a Sparsify-Reconstruct-Decompose framework designed for communication-efficient federated LLMs fine-tuning. We first introduce an importance-aware sparsification method that preserves the structural integrity of LoRA updates to reduce the uploaded parameter count. The server then reconstructs and aggregates these updates in a full-rank space to mitigate conflicts. Finally, it decomposes the global update into a sparse low-rank format for broadcast, ensuring a symmetrically efficient cycle. We also propose an efficient variant, FedSRD-e, to reduce computational overhead. Experimental results on 10 benchmarks demonstrate that our framework significantly reduces communication costs by up to 90\% while even improving model performance on heterogeneous client data.
♻ ☆ AutoRev: Multi-Modal Graph Retrieval for Automated Peer-Review Generation
Enhancing the quality and efficiency of academic publishing is critical for both authors and reviewers, as research papers are central to scholarly communication and a major source of high-quality content on the web. To support this goal, we propose AutoRev, an automatic peer-review system designed to provide actionable, high-quality feedback to both reviewers and authors. AutoRev leverages a novel Multi-Modal Retrieval-Augmented Generation (RAG) framework that combines textual and graphical representations of academic papers. By modelling documents as graphs, AutoRev effectively retrieves the most pertinent information, significantly reducing the input context length for LLMs and thereby enhancing their review generation capabilities. Experimental results show that AutoRev outperforms state-of-the-art baselines by up to 58.72% and demonstrates competitive performance in human evaluations against ground truth reviews. We envision AutoRev as a powerful tool to streamline the peer-review workflow, alleviating challenges and enabling scalable, high-quality scholarly publishing. By guiding both authors and reviewers, AutoRev has the potential to accelerate the dissemination of quality research on the web at a larger scale. Code will be released upon acceptance.
♻ ☆ GIIFT: Graph-guided Inductive Image-free Multimodal Machine Translation EMNLP 2025
Multimodal Machine Translation (MMT) has demonstrated the significant help of visual information in machine translation. However, existing MMT methods face challenges in leveraging the modality gap by enforcing rigid visual-linguistic alignment whilst being confined to inference within their trained multimodal domains. In this work, we construct novel multimodal scene graphs to preserve and integrate modality-specific information and introduce GIIFT, a two-stage Graph-guided Inductive Image-Free MMT framework that uses a cross-modal Graph Attention Network adapter to learn multimodal knowledge in a unified fused space and inductively generalize it to broader image-free translation domains. Experimental results on the Multi30K dataset of English-to-French and English-to-German tasks demonstrate that our GIIFT surpasses existing approaches and achieves the state-of-the-art, even without images during inference. Results on the WMT benchmark show significant improvements over the image-free translation baselines, demonstrating the strength of GIIFT towards inductive image-free inference.
comment: Accepted as an oral presentation at the EMNLP 2025 Workshop on Machine Translation (WMT)
♻ ☆ Taxonomy, Opportunities, and Challenges of Representation Engineering for Large Language Models
Representation Engineering (RepE) is a novel paradigm for controlling the behavior of LLMs. Unlike traditional approaches that modify inputs or fine-tune the model, RepE directly manipulates the model's internal representations. As a result, it may offer more effective, interpretable, data-efficient, and flexible control over models' behavior. We present the first comprehensive survey of RepE for LLMs, reviewing the rapidly growing literature to address key questions: What RepE methods exist and how do they differ? For what concepts and problems has RepE been applied? What are the strengths and weaknesses of RepE compared to other methods? To answer these, we propose a unified framework describing RepE as a pipeline comprising representation identification, operationalization, and control. We posit that while RepE methods offer significant potential, challenges remain, including managing multiple concepts, ensuring reliability, and preserving models' performance. Towards improving RepE, we identify opportunities for experimental and methodological improvements and construct a guide for best practices.
♻ ☆ The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives
The objectives that Large Language Models (LLMs) implicitly optimize remain dangerously opaque, making trustworthy alignment and auditing a grand challenge. While Inverse Reinforcement Learning (IRL) can infer reward functions from behaviour, existing approaches either produce a single, overconfident reward estimate or fail to address the fundamental ambiguity of the task (non-identifiability). This paper introduces a principled auditing framework that re-frames reward inference from a simple estimation task to a comprehensive process for verification. Our framework leverages Bayesian IRL to not only recover a distribution over objectives but to enable three critical audit capabilities: (i) Quantifying and systematically reducing non-identifiability by demonstrating posterior contraction over sequential rounds of evidence; (ii) Providing actionable, uncertainty-aware diagnostics that expose spurious shortcuts and identify out-of-distribution prompts where the inferred objective cannot be trusted; and (iii) Validating policy-level utility by showing that the refined, low-uncertainty reward can be used directly in RLHF to achieve training dynamics and toxicity reductions comparable to the ground-truth alignment process. Empirically, our framework successfully audits a detoxified LLM, yielding a well-calibrated and interpretable objective that strengthens alignment guarantees. Overall, this work provides a practical toolkit for auditors, safety teams, and regulators to verify what LLMs are truly trying to achieve, moving us toward more trustworthy and accountable AI.
comment: Preprint
♻ ☆ Enhancing Few-shot Keyword Spotting Performance through Pre-Trained Self-supervised Speech Models
Keyword Spotting plays a critical role in enabling hands-free interaction for battery-powered edge devices. Few-Shot Keyword Spotting (FS-KWS) addresses the scalability and adaptability challenges of traditional systems by enabling recognition of custom keywords with only a few examples. However, existing FS-KWS systems achieve subpar accuracy at desirable false acceptance rates, particularly in resource-constrained edge environments. To address these issues, we propose a training scheme that leverages self-supervised learning models for robust feature extraction, dimensionality reduction, and knowledge distillation. The teacher model, based on Wav2Vec 2.0 is trained using Sub-center ArcFace loss, which enhances inter-class separability and intra-class compactness. To enable efficient deployment on edge devices, we introduce attention-based dimensionality reduction and train a standard lightweight ResNet15 student model. We evaluate the proposed approach on the English portion of the Multilingual Spoken Words Corpus (MSWC) and the Google Speech Commands (GSC) datasets. Notably, the proposed training method improves the 10-shot classification accuracy from 33.4% to 74.1% on 11 classes at 1% false alarm accuracy on the GSC dataset, thus making it significantly better-suited for a real use case scenario.
comment: Submitted to IEEE Signal Processing Letters, 5 pages, 3 figures
♻ ☆ InfiMed: Low-Resource Medical MLLMs with Advancing Understanding and Reasoning
Multimodal Large Language Models (MLLMs) have achieved remarkable progress in domains such as visual understanding and mathematical reasoning. However, their application in the medical domain is constrained by two key challenges: (1) multimodal medical datasets are scarce and often contain sparse information, limiting reasoning depth; and (2) Reinforcement Learning with Verifiable Rewards (RLVR), though effective in general domains, cannot reliably improve model performance in the medical domain. To overcome these challenges, during the supervised fine-tuning (SFT) stage, we incorporate high-quality textual reasoning data and general multimodal data alongside multimodal medical data to efficiently enhance foundational medical capabilities and restore the base model's reasoning ability. Moreover, considering that there are some multimodal medical datasets with sparse information, we further synthesize reflective-pattern-injected chain-of-thought (CoT) in addition to general CoT samples, equipping the model with initial reflective reasoning capabilities that provide a structured foundation for subsequent RLVR training. Finally, we introduce our InfiMed-Series models, InfiMed-SFT-3B and InfiMed-RL-3B, both of which deliver state-of-the-art performance across seven multimodal medical benchmarks. Notably, InfiMed-RL-3B achieves an average accuracy of 59.2%, outperforming even larger models like InternVL3-8B, which achieves 57.3%. Specifically, during the SFT phase, we utilized 188K samples, while the RLVR phase incorporated 36K samples, demonstrating the efficacy of both training strategies in achieving superior performance. We also conducted a series of extensive experiments, which provide valuable insights that contribute to advancing the performance of MLLMs in medical scenarios.
♻ ☆ Are BabyLMs Deaf to Gricean Maxims? A Pragmatic Evaluation of Sample-efficient Language Models EMNLP 2025
Implicit meanings are integral to human communication, making it essential for language models to be capable of identifying and interpreting them. Grice (1975) proposed a set of conversational maxims that guide cooperative dialogue, noting that speakers may deliberately violate these principles to express meanings beyond literal words, and that listeners, in turn, recognize such violations to draw pragmatic inferences. Building on Surian et al. (1996)'s study of children's sensitivity to violations of Gricean maxims, we introduce a novel benchmark to test whether language models pretrained on less than 10M and less than 100M tokens can distinguish maxim-adhering from maxim-violating utterances. We compare these BabyLMs across five maxims and situate their performance relative to children and a Large Language Model (LLM) pretrained on 3T tokens. We find that overall, models trained on less than 100M tokens outperform those trained on less than 10M, yet fall short of child-level and LLM competence. Our results suggest that modest data increases improve some aspects of pragmatic behavior, leading to finer-grained differentiation between pragmatic dimensions.
comment: Accepted for the BabyLM workshop in EMNLP 2025
♻ ☆ Evil twins are not that evil: Qualitative insights into machine-generated prompts
It has been widely observed that language models (LMs) respond in predictable ways to algorithmically generated prompts that are seemingly unintelligible. This is both a sign that we lack a full understanding of how LMs work, and a practical challenge, because opaqueness can be exploited for harmful uses of LMs, such as jailbreaking. We present the first thorough analysis of opaque machine-generated prompts, or autoprompts, pertaining to 6 LMs of different sizes and families. We find that machine-generated prompts are characterized by a last token that is often intelligible and strongly affects the generation. A small but consistent proportion of the previous tokens are prunable, probably appearing in the prompt as a by-product of the fact that the optimization process fixes the number of tokens. The remaining tokens fall into two categories: filler tokens, which can be replaced with semantically unrelated substitutes, and keywords, that tend to have at least a loose semantic relation with the generation, although they do not engage in well-formed syntactic relations with it. Additionally, human experts can reliably identify the most influential tokens in an autoprompt a posteriori, suggesting these prompts are not entirely opaque. Finally, some of the ablations we applied to autoprompts yield similar effects in natural language inputs, suggesting that autoprompts emerge naturally from the way LMs process linguistic inputs in general.
comment: Published as workshop paper at BlackBox NLP 2025
♻ ☆ Injecting External Knowledge into the Reasoning Process Enhances Retrieval-Augmented Generation SIGIR
Retrieval-augmented generation (RAG) has been widely adopted to augment large language models (LLMs) with external knowledge for knowledge-intensive tasks. However, its effectiveness is often undermined by the presence of noisy (i.e., low-quality) retrieved passages. Enhancing LLMs' robustness to such noise is critical for improving the reliability of RAG systems. Recent advances have equipped LLMs with strong reasoning and self-reflection capabilities, allowing them to identify and correct errors in their reasoning process. Inspired by this ability, we propose Passage Injection-a simple yet effective method that explicitly incorporates retrieved passages into LLMs' reasoning process, aiming to enhance the model's ability to recognize and resist noisy passages. We validate Passage Injection under general RAG settings using BM25 as the retriever. Experiments on four reasoning-enhanced LLMs across four factual QA datasets demonstrate that Passage Injection significantly improves overall RAG performance. Further analysis on two noisy retrieval settings-random noise, where the model is provided irrelevant passages, and counterfactual noise, where it is given misleading passages-shows that Passage Injection consistently improves robustness. Controlled experiments confirm that Passage Injection can also effectively leverage helpful passages. These findings suggest that incorporating passages in LLMs' reasoning process is a promising direction for building more robust RAG systems. The code can be found \href{here}{https://github.com/Trustworthy-Information-Access/Passage-Injection}.
comment: SIGIR-AP 2025
♻ ☆ EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models
We present EvalMORAAL, a transparent chain-of-thought (CoT) framework that uses two scoring methods (log-probabilities and direct ratings) plus a model-as-judge peer review to evaluate moral alignment in 20 large language models. We assess models on the World Values Survey (55 countries, 19 topics) and the PEW Global Attitudes Survey (39 countries, 8 topics). With EvalMORAAL, top models align closely with survey responses (Pearson's r approximately 0.90 on WVS). Yet we find a clear regional difference: Western regions average r=0.82 while non-Western regions average r=0.61 (a 0.21 absolute gap), indicating consistent regional bias. Our framework adds three parts: (1) two scoring methods for all models to enable fair comparison, (2) a structured chain-of-thought protocol with self-consistency checks, and (3) a model-as-judge peer review that flags 348 conflicts using a data-driven threshold. Peer agreement relates to survey alignment (WVS r=0.74, PEW r=0.39, both p<.001), supporting automated quality checks. These results show real progress toward culture-aware AI while highlighting open challenges for use across regions.
♻ ☆ TextMine: Data, Evaluation Framework and Ontology-guided LLM Pipeline for Humanitarian Mine Action
Humanitarian Mine Action (HMA) addresses the challenge of detecting and removing landmines from conflict regions. Much of the life-saving operational knowledge produced by HMA agencies is buried in unstructured reports, limiting the transferability of information between agencies. To address this issue, we propose TextMine: the first dataset, evaluation framework and ontology-guided large language model (LLM) pipeline for knowledge extraction in the HMA domain. TextMine structures HMA reports into (subject, relation, object)-triples, thus creating domain-specific knowledge. To ensure real-world relevance, we created the dataset in collaboration with Cambodian Mine Action Center (CMAC). We further introduce a bias-aware evaluation framework that combines human-annotated triples with an LLM-as-Judge protocol to mitigate position bias in reference-free scoring. Our experiments show that ontology-aligned prompts improve extraction accuracy by up to 44.2%, reduce hallucinations by 22.5%, and enhance format adherence by 20.9% compared to baseline models. We publicly release the dataset and code.
♻ ☆ Benchmarking Gaslighting Negation Attacks Against Multimodal Large Language Models
Multimodal Large Language Models (MLLMs) have exhibited remarkable advancements in integrating different modalities, excelling in complex understanding and generation tasks. Despite their success, MLLMs remain vulnerable to conversational adversarial inputs. In this paper, we systematically study gaslighting negation attacks: a phenomenon where models, despite initially providing correct answers, are persuaded by user-provided negations to reverse their outputs, often fabricating justifications. We conduct extensive evaluations of state-of-the-art MLLMs across diverse benchmarks and observe substantial performance drops when negation is introduced. Notably, we introduce the first benchmark GaslightingBench, specifically designed to evaluate the vulnerability of MLLMs to negation arguments. GaslightingBench consists of multiple-choice questions curated from existing datasets, along with generated negation prompts across 20 diverse categories. Throughout extensive evaluation, we find that proprietary models such as Gemini-1.5-flash and GPT-4o demonstrate better resilience compared to open-source counterparts like Qwen2-VL and LLaVA, though even advanced reasoning-oriented models like Gemini-2.5-Pro remain susceptible. Our category-level analysis further shows that subjective or socially nuanced domains (e.g., Social Relation, Image Emotion) are especially fragile, while more objective domains (e.g., Geography) exhibit relatively smaller but still notable drops. Overall, all evaluated MLLMs struggle to maintain logical consistency under gaslighting negation attack. These findings highlight a fundamental robustness gap and provide insights for developing more reliable and trustworthy multimodal AI systems. Project website: https://yxg1005.github.io/GaslightingNegationAttacks/.
comment: Project website: https://yxg1005.github.io/GaslightingNegationAttacks/
♻ ☆ 2 OLMo 2 Furious
We present OLMo 2, the next generation of our fully open language models. OLMo 2 includes a family of dense autoregressive language models at 7B, 13B and 32B scales with fully released artifacts -- model weights, full training data, training code and recipes, training logs and thousands of intermediate checkpoints. In this work, we describe our modified model architecture and training recipe, focusing on techniques for achieving better training stability and improved per-token efficiency. Our updated pretraining data mixture introduces a new, specialized data mix called Dolmino Mix 1124, which significantly improves model capabilities across many downstream task benchmarks when introduced via late-stage curriculum training (i.e. specialized data during the annealing phase of pretraining). Finally, we incorporate best practices from T\"ulu 3 to develop OLMo 2-Instruct, focusing on permissive data and extending our final-stage reinforcement learning with verifiable rewards (RLVR). Our OLMo 2 base models sit at the Pareto frontier of performance to training compute, often matching or outperforming open-weight only models like Llama 3.1, Qwen 2.5, and Gemma 2 while using fewer FLOPs and with fully transparent training data, code, and recipe. Our fully open OLMo 2-Instruct models are competitive with open-weight only models of comparable size and even some proprietary models like GPT-3.5 Turbo and GPT 4o Mini.
comment: Shorter version accepted to COLM 2025. Updated to include 32B results. Model demo available at playground.allenai.org
♻ ☆ The Percept-V Challenge: Can Multimodal LLMs Crack Simple Perception Problems?
Cognitive science research treats visual perception, the ability to understand and make sense of a visual input, as one of the early developmental signs of intelligence. Its TVPS-4 framework categorizes and tests human perception into seven skills such as visual discrimination, and form constancy. Do Multimodal Large Language Models (MLLMs) match up to humans in basic perception? Even though there are many benchmarks that evaluate MLLMs on advanced reasoning and knowledge skills, there is limited research that focuses evaluation on simple perception. In response, we introduce Percept-V, a dataset containing 6000 program-generated uncontaminated images divided into 30 domains, where each domain tests one or more TVPS-4 skills. Our focus is on perception, so we make our domains quite simple and the reasoning and knowledge required for solving them are minimal. Since modern-day MLLMs can solve much more complex tasks, our a-priori expectation is that they will solve these domains very easily. Contrary to our belief, our experiments show a weak performance of SoTA proprietary and open-source MLLMs compared to very high human performance on Percept-V. We find that as number of objects in the image increases, performance goes down rather fast. Our experiments also identify the perception skills that are considerably harder for all models.
♻ ☆ Epistemic Diversity and Knowledge Collapse in Large Language Models
Large language models (LLMs) tend to generate lexically, semantically, and stylistically homogenous texts. This poses a risk of knowledge collapse, where homogenous LLMs mediate a shrinking in the range of accessible information over time. Existing works on homogenization are limited by a focus on closed-ended multiple-choice setups or fuzzy semantic features, and do not look at trends across time and cultural contexts. To overcome this, we present a new methodology to measure epistemic diversity, i.e., variation in real-world claims in LLM outputs, which we use to perform a broad empirical study of LLM knowledge collapse. We test 27 LLMs, 155 topics covering 12 countries, and 200 prompt variations sourced from real user chats. For the topics in our study, we show that while newer models tend to generate more diverse claims, nearly all models are less epistemically diverse than a basic web search. We find that model size has a negative impact on epistemic diversity, while retrieval-augmented generation (RAG) has a positive impact, though the improvement from RAG varies by the cultural context. Finally, compared to a traditional knowledge source (Wikipedia), we find that country-specific claims reflect the English language more than the local one, highlighting a gap in epistemic representation
comment: 16 pages; 8 figures, 4 tables; v2 changelog: Fixed the modeling for table 3, random effect is the model version; v3 changelog: Fixed minor formatting issues in tables 2 and 3;
♻ ☆ PARL-MT: Learning to Call Functions in Multi-Turn Conversation with Progress Awareness
Large language models (LLMs) have achieved impressive success in single-turn function calling, yet real-world applications such as travel planning or multi-stage data analysis typically unfold across multi-turn conversations. In these settings, LLMs must not only issue accurate function calls at each step but also maintain progress awareness, the ability to summarize past interactions and plan future actions to ensure coherent, long-horizon task execution. Existing approaches, however, either reduce multi-turn training to isolated single-turn samples, which neglects task-level planning, or employ end-to-end reinforcement learning (RL) that struggles with redundancy and lacks explicit integration of progress awareness. To overcome these limitations, we introduce PARL-MT, a framework that explicitly incorporates progress awareness into LLM training for multi-turn function calling. PARL-MT combines (i) a Progress Awareness Generation (PAG) pipeline, which automatically constructs datasets coupling conversation summaries with future task planning, and (ii) a Progress Awareness-Guided Reinforcement Learning (PAG-RL) algorithm, which integrates progress awareness into RL training to reduce contextual redundancy and improve alignment between local actions and global task completion. Empirical results on two public benchmarks demonstrate that PARL-MT significantly outperforms existing methods, highlighting the effectiveness of progress awareness in enabling robust and efficient multi-turn function calling.
♻ ☆ GMLM: Bridging Graph Neural Networks and Language Models for Heterophilic Node Classification
Integrating Pre-trained Language Models (PLMs) with Graph Neural Networks (GNNs) remains a central challenge in text-rich heterophilic graph learning. We propose a novel integration framework that enables effective fusion between powerful pre-trained text encoders and Relational Graph Convolutional Networks (R-GCNs). Our method enhances the alignment of textual and structural representations through a bidirectional fusion mechanism and contrastive node-level optimization. To evaluate the approach, we train two variants using different PLMs: Snowflake-Embed (state-of-the-art) and GTE-base, each paired with an R-GCN backbone. Experiments on five heterophilic benchmarks demonstrate that our integration method achieves state-of-the-art results on four datasets, surpassing existing GNN and large language model-based approaches. Notably, Snowflake-Embed + R-GCN improves accuracy on the Texas dataset by over 8\% and on Wisconsin by nearly 5\%. These results highlight the effectiveness of our fusion strategy for advancing text-rich graph representation learning.
♻ ☆ Do RAG Systems Really Suffer From Positional Bias?
Retrieval Augmented Generation enhances LLM accuracy by adding passages retrieved from an external corpus to the LLM prompt. This paper investigates how positional bias - the tendency of LLMs to weight information differently based on its position in the prompt - affects not only the LLM's capability to capitalize on relevant passages, but also its susceptibility to distracting passages. Through extensive experiments on three benchmarks, we show how state-of-the-art retrieval pipelines, while attempting to retrieve relevant passages, systematically bring highly distracting ones to the top ranks, with over 60% of queries containing at least one highly distracting passage among the top-10 retrieved passages. As a result, the impact of the LLM positional bias, which in controlled settings is often reported as very prominent by related works, is actually marginal in real scenarios since both relevant and distracting passages are, in turn, penalized. Indeed, our findings reveal that sophisticated strategies that attempt to rearrange the passages based on LLM positional preferences do not perform better than random shuffling.
♻ ☆ GlotEval: A Test Suite for Massively Multilingual Evaluation of Large Language Models EMNLP
Large language models (LLMs) are advancing at an unprecedented pace globally, with regions increasingly adopting these models for applications in their primary language. Evaluation of these models in diverse linguistic environments, especially in low-resource languages, has become a major challenge for academia and industry. Existing evaluation frameworks are disproportionately focused on English and a handful of high-resource languages, thereby overlooking the realistic performance of LLMs in multilingual and lower-resource scenarios. To address this gap, we introduce GlotEval, a lightweight framework designed for massively multilingual evaluation. Supporting seven key tasks (machine translation, text classification, summarization, open-ended generation, reading comprehension, sequence labeling, and intrinsic evaluation), spanning over dozens to hundreds of languages, GlotEval highlights consistent multilingual benchmarking, language-specific prompt templates, and non-English-centric machine translation. This enables a precise diagnosis of model strengths and weaknesses in diverse linguistic contexts. A multilingual translation case study demonstrates GlotEval's applicability for multilingual and language-specific evaluations.
comment: EMNLP demo 2025
♻ ☆ Rethinking Multilingual Continual Pretraining: Data Mixing for Adapting LLMs Across Languages and Resources
Large Language Models (LLMs) exhibit significant disparities in performance across languages, primarily benefiting high-resource languages while marginalizing underrepresented ones. Continual Pretraining (CPT) has emerged as a promising approach to address this imbalance, although the relative effectiveness of monolingual, bilingual, and code-augmented data strategies remains unclear. This study systematically evaluates 36 CPT configurations involving three multilingual base models, across 30+ languages categorized as altruistic, selfish, and stagnant, spanning various resource levels. Our findings reveal three major insights: (1) Bilingual CPT improves multilingual classification but often causes language mixing issues during generation. (2) Including programming code data during CPT consistently enhances multilingual classification accuracy, particularly benefiting low-resource languages, but introduces a trade-off by slightly degrading generation quality. (3) Contrary to prior work, we observe substantial deviations from language classifications according to their impact on cross-lingual transfer: Languages classified as altruistic often negatively affect related languages, selfish languages show conditional and configuration-dependent behavior, and stagnant languages demonstrate surprising adaptability under certain CPT conditions. These nuanced interactions emphasize the complexity of multilingual representation learning, underscoring the importance of systematic studies on generalizable language classification to inform future multilingual CPT strategies.
comment: COLM 2025
♻ ☆ Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation
Recent advancements in speech generation have been driven by large-scale training datasets. However, current models struggle to capture the spontaneity and variability inherent in real-world human speech, as they are primarily trained on audio-book datasets limited to formal, read-aloud speaking styles. To address this limitation, we introduce Emilia-Pipe, an open-source preprocessing pipeline designed to extract high-quality training data from valuable yet under-explored in-the-wild sources that capture spontaneous human speech in real-world contexts. Using Emilia-Pipe, we construct Emilia, which comprises over 101k hours of speech across six languages: English, Chinese, German, French, Japanese, and Korean. Furthermore, we expand Emilia to Emilia-Large, a dataset exceeding 216k hours, making it one of the largest open-source speech generation resources available. Extensive experiments show that Emilia-trained models produce markedly more spontaneous, human-like speech than those trained on traditional audio-book datasets, while matching their intelligibility. These models better capture diverse speaker timbres and the full spectrum of real-world conversational styles. Our work also highlights the importance of scaling dataset size for advancing speech generation performance and validates the effectiveness of Emilia for both multilingual and crosslingual speech generation tasks.
comment: Full version of arXiv:2407.05361, dataset is available at: https://huggingface.co/datasets/amphion/Emilia-Dataset
♻ ☆ Thinking with Nothinking Calibration: A New In-Context Learning Paradigm in Reasoning Large Language Models
Reasoning large language models (RLLMs) have recently demonstrated remarkable capabilities through structured and multi-step reasoning. While prior research has primarily focused on improving their training and inference strategies, their potential for in-context learning (ICL) remains largely underexplored. To fill this gap, we propose Thinking with Nothinking Calibration (JointThinking), a new ICL paradigm that prompts the model to generate two answers in parallel: one in Thinking mode and the other in Nothinking mode. A second round of Thinking is triggered only when the two initial responses are inconsistent, using a single prompt with two different answers. Extensive experiments across multiple reasoning benchmarks demonstrate that JointThinking significantly outperforms few-shot chain-of-thought (CoT), thinking twice and majority voting. Moreover, it achieves comparable in-distribution performance to training-based SOTA reasoning method, while substantially outperforming on out-of-distribution tasks. We further conduct a systematic analysis of the calibration mechanism, showing the importance of structural thinking diversity and the benefits of consistency check. Additionally, we observe that the performance gap between actual and ideal reasoning narrows as model size increases in the second thinking, indicating the strong scalability of our approach. Finally, we discuss current limitations and outline promising directions for future ICL research in RLLMs.
♻ ☆ PredGen: Accelerated Inference of Large Language Models through Input-Time Speculation for Real-Time Speech Interaction
Large Language Models (LLMs) are widely used in real-time voice chat applications, typically in combination with text-to-speech (TTS) systems to generate audio responses. However, their large size often leads to noticeable latency between the end of user input and the start of audio output, resulting in suboptimal user experiences. This latency is particularly evident when LLMs are deployed as single-user voice assistants on consumer-grade hardware with limited computing capacity. We discovered that this latency is primarily dominated by the time it takes for the LLMs to generate the first sentence, which is required as input by the TTS systems that synthesize audio responses on a sentence-by-sentence basis. To address this bottleneck, we propose Predictive Generation (PredGen), a novel framework that mitigates-or even eliminates-this delay through speculative decoding at input time. PredGen generates candidate responses while the user is still speaking, enabling the system to begin TTS processing with minimal delay. Simulated experiments on the Lmsys and MT-Bench datasets show that the proposed method can effectively reduce the latency by around 2x across a wide range of use cases, while incurring only minimal additional computation cost at input time-computation that would otherwise go unused.
comment: 16 pages,4 figures
♻ ☆ Geometry of Semantics in Next-Token Prediction: How Optimization Implicitly Organizes Linguistic Representations
We investigate how next-token prediction (NTP) optimization leads language models to extract and organize semantic structure from text. Our analysis, based on a tractable mathematical model and controlled synthetic data, reveals that NTP implicitly guides models to factor a centered support matrix encoding context-to-next-token co-occurrence patterns via singular value decomposition (SVD). While models never explicitly construct this matrix, learned word and context embeddings converge to its SVD factors, with singular vectors encoding latent semantic concepts through their sign patterns. We demonstrate that concepts corresponding to larger singular values are learned earlier during training, yielding a natural semantic hierarchy where broad categories emerge before fine-grained ones. This insight motivates orthant-based clustering, a method that combines concept signs to identify interpretable semantic categories. We validate our findings on synthetic datasets and pretrained language models, recovering diverse semantic structures such as grammatical categories, named entity types, and topical distinctions (medical, entertainment). Our work bridges classical distributional semantics and neural collapse geometry, characterizing how gradient-based optimization implicitly determines both the matrix representation and factorization method that encode semantic structure.
comment: Revised manuscript for improved clarity and readability
♻ ☆ Can AI Truly Represent Your Voice in Deliberations? A Comprehensive Study of Large-Scale Opinion Aggregation with LLMs
Large-scale public deliberations generate thousands of free-form contributions that must be synthesized into representative and neutral summaries for policy use. While LLMs have been shown as a promising tool to generate summaries for large-scale deliberations, they also risk underrepresenting minority perspectives and exhibiting bias with respect to the input order, raising fairness concerns in high-stakes contexts. Studying and fixing these issues requires a comprehensive evaluation at a large scale, yet current practice often relies on LLMs as judges, which show weak alignment with human judgments. To address this, we present DeliberationBank, a large-scale human-grounded dataset with (1) opinion data spanning ten deliberation questions created by 3,000 participants and (2) summary judgment data annotated by 4,500 participants across four dimensions (representativeness, informativeness, neutrality, policy approval). Using these datasets, we train DeliberationJudge, a fine-tuned DeBERTa model that can rate deliberation summaries from individual perspectives. DeliberationJudge is more efficient and more aligned with human judgements compared to a wide range of LLM judges. With DeliberationJudge, we evaluate 18 LLMs and reveal persistent weaknesses in deliberation summarization, especially underrepresentation of minority positions. Our framework provides a scalable and reliable way to evaluate deliberation summarization, helping ensure AI systems are more representative and equitable for policymaking.
♻ ☆ HopWeaver: Cross-Document Synthesis of High-Quality and Authentic Multi-Hop Questions
Multi-Hop Question Answering (MHQA) is crucial for evaluating the model's capability to integrate information from diverse sources. However, creating extensive and high-quality MHQA datasets is challenging: (i) manual annotation is expensive, and (ii) current synthesis methods often produce simplistic questions or require extensive manual guidance. This paper introduces HopWeaver, the first cross-document framework synthesizing authentic multi-hop questions without human intervention. HopWeaver synthesizes bridge and comparison questions through an innovative pipeline that identifies complementary documents and constructs authentic reasoning paths to ensure true multi-hop reasoning. We further present a comprehensive system for evaluating the synthesized multi-hop questions. Empirical evaluations demonstrate that the synthesized questions achieve comparable or superior quality to human-annotated datasets at a lower cost. Our framework provides a valuable tool for the research community: it can automatically generate challenging benchmarks from any raw corpus, which opens new avenues for both evaluation and targeted training to improve the reasoning capabilities of advanced QA models, especially in domains with scarce resources.
comment: 31 pages. Code will be available at [https://github.com/Zh1yuShen/HopWeaver]
♻ ☆ ProCut: LLM Prompt Compression via Attribution Estimation
In large-scale industrial LLM systems, prompt templates often expand to thousands of tokens as teams iteratively incorporate sections such as task instructions, few-shot examples, and heuristic rules to enhance robustness and coverage. This expansion leads to bloated prompts that are difficult to maintain and incur significant inference latency and serving costs. To address this, we introduce Prompt Compression via Attribution Estimation (ProCut), a flexible, LLM-agnostic, training-free framework that compresses prompts through attribution analysis. ProCut segments prompt templates into semantically meaningful units, quantifies their impact on task performance, and prunes low-utility components. Through extensive experiments on five public benchmark datasets and real-world industrial prompts, we show that ProCut achieves substantial prompt size reductions (78% fewer tokens in production) while maintaining or even slightly improving task performance (up to 62% better than alternative methods). We further introduce an LLM-driven attribution estimator that reduces compression latency by over 50%, and demonstrate that ProCut integrates seamlessly with existing prompt-optimization frameworks to produce concise, high-performing prompts.
♻ ☆ Membership Inference Attacks on LLM-based Recommender Systems
Large language models (LLMs) based Recommender Systems (RecSys) can flexibly adapt recommendation systems to different domains. It utilizes in-context learning (ICL), i.e., the prompts, to customize the recommendation functions, which include sensitive historical user-specific item interactions, e.g., implicit feedback like clicked items or explicit product reviews. Such private information may be exposed to novel privacy attack. However, no study has been done on this important issue. We design four membership inference attacks (MIAs), aiming to reveal whether victims' historical interactions have been used by system prompts. They are \emph{direct inquiry, hallucination, similarity, and poisoning attacks}, each of which utilizes the unique features of LLMs or RecSys. We have carefully evaluated them on three LLMs that have been used to develop ICL-LLM RecSys and two well-known RecSys benchmark datasets. The results confirm that the MIA threat on LLM RecSys is realistic: direct inquiry and poisoning attacks showing significantly high attack advantages. We have also analyzed the factors affecting these attacks, such as the number of shots in system prompts and the position of the victim in the shots.
comment: this paper is under review
♻ ☆ Slow-Fast Policy Optimization: Reposition-Before-Update for LLM Reasoning
Reinforcement learning (RL) has become central to enhancing reasoning in large language models (LLMs). Yet on-policy algorithms such as Group Relative Policy Optimization (GRPO) often suffer in early training: noisy gradients from low-quality rollouts lead to unstable updates and inefficient exploration. We introduce Slow-Fast Policy Optimization (SFPO), a simple yet efficient framework to address these limitations via decomposing each step into three stages: a short fast trajectory of inner steps on the same batch, a reposition mechanism to control off-policy drift, and a final slow correction. This reposition-before-update design preserves the objective and rollout process unchanged, making SFPO plug-compatible with existing policy-gradient pipelines. Extensive experiments demonstrate that SFPO consistently improves stability, reduces rollouts, and accelerates convergence of reasoning RL training. Specifically, it outperforms GRPO by up to 2.80 points in average on math reasoning benchmarks. It also achieves up to 4.93\texttimes{} fewer rollouts and an up to 4.19\texttimes{} reduction in wall-clock time to match GRPO's best accuracy.
♻ ☆ TIME: A Multi-level Benchmark for Temporal Reasoning of LLMs in Real-World Scenarios NeurIPS 2025
Temporal reasoning is pivotal for Large Language Models (LLMs) to comprehend the real world. However, existing works neglect the real-world challenges for temporal reasoning: (1) intensive temporal information, (2) fast-changing event dynamics, and (3) complex temporal dependencies in social interactions. To bridge this gap, we propose a multi-level benchmark TIME, designed for temporal reasoning in real-world scenarios. TIME consists of 38,522 QA pairs, covering 3 levels with 11 fine-grained sub-tasks. This benchmark encompasses 3 sub-datasets reflecting different real-world challenges: TIME-Wiki, TIME-News, and TIME-Dial. We conduct extensive experiments on reasoning models and non-reasoning models. And we conducted an in-depth analysis of temporal reasoning performance across diverse real-world scenarios and tasks, and summarized the impact of test-time scaling on temporal reasoning capabilities. Additionally, we release TIME-Lite, a human-annotated subset to foster future research and standardized evaluation in temporal reasoning. The code is available at https://github.com/sylvain-wei/TIME , the dataset is available at https://huggingface.co/datasets/SylvainWei/TIME , and the project page link is https://sylvain-wei.github.io/TIME/ .
comment: Accepted by NeurIPS 2025 (Spotlight)
♻ ☆ LLM Hallucination Detection: HSAD
Although Large Language Models have demonstrated powerful capabilities in a wide range of tasks such as language understanding and code generation, the frequent occurrence of hallucinations during the generation process has become a significant impediment to their deployment in critical application scenarios. Current mainstream hallucination detection methods rely on factual consistency verification or static hidden layer features. The former is constrained by the scope of knowledge coverage, while the latter struggles to capture reasoning biases during the inference process. To address these issues, and inspired by signal analysis methods in cognitive neuroscience, this paper proposes a hallucination detection method based on the frequency-domain analysis of hidden layer temporal signals, named HSAD (\textbf{H}idden \textbf{S}ignal \textbf{A}nalysis-based \textbf{D}etection). First, by treating the LLM's reasoning process as a cognitive journey that unfolds over time, we propose modeling and simulating the human process of signal perception and discrimination in a deception-detection scenario through hidden layer temporal signals. Next, The Fast Fourier Transform is applied to map these temporal signals into the frequency domain to construct spectral features, which are used to capture anomalies that arise during the reasoning process; analysis experiments on these spectral features have proven the effectiveness of this approach. Finally, a hallucination detection algorithm is designed based on these spectral features to identify hallucinations in the generated content. By effectively combining the modeling of the reasoning process with frequency-domain feature extraction, the HSAD method overcomes the limitations of existing approaches in terms of knowledge coverage and the detection of reasoning biases, demonstrating higher detection accuracy and robustness.
comment: in Chinese language
♻ ☆ SMARTER: A Data-efficient Framework to Improve Toxicity Detection with Explanation via Self-augmenting Large Language Models
WARNING: This paper contains examples of offensive materials. To address the proliferation of toxic content on social media, we introduce SMARTER, we introduce SMARTER, a data-efficient two-stage framework for explainable content moderation using Large Language Models (LLMs). In Stage 1, we leverage LLMs' own outputs to generate synthetic explanations for both correct and incorrect labels, enabling alignment via preference optimization with minimal human supervision. In Stage 2, we refine explanation quality through cross-model training, allowing weaker models to align stylistically and semantically with stronger ones. Experiments on three benchmark tasks -- HateXplain, Latent Hate, and Implicit Hate -- demonstrate that SMARTER enables LLMs to achieve up to a 13.5% macro-F1 improvement over standard few-shot baselines while using only a fraction of the full training data. Our framework offers a scalable strategy for low-resource settings by harnessing LLMs' self-improving capabilities for both classification and explanation.
comment: NLP, Hate speech detection, explanation, LLM. Version 2: updated experiments and analysis
♻ ☆ An Illusion of Progress? Assessing the Current State of Web Agents
As digitalization and cloud technologies evolve, the web is becoming increasingly important in the modern society. Autonomous web agents based on large language models (LLMs) hold a great potential in work automation. It is therefore important to accurately measure and monitor the progression of their capabilities. In this work, we conduct a comprehensive and rigorous assessment of the current state of web agents. Our results depict a very different picture of the competency of current agents, suggesting over-optimism in previously reported results. This gap can be attributed to shortcomings in existing benchmarks. We introduce Online-Mind2Web, an online evaluation benchmark consisting of 300 diverse and realistic tasks spanning 136 websites. It enables us to evaluate web agents under a setting that approximates how real users use these agents. To facilitate more scalable evaluation and development, we also develop a novel LLM-as-a-Judge automatic evaluation method and show that it can achieve around 85% agreement with human judgment, substantially higher than existing methods. Finally, we present the first comprehensive comparative analysis of current web agents, highlighting both their strengths and limitations to inspire future research.
comment: 22 pages, 17 figures, 7 tables
Information Retrieval 25
☆ Towards Reliable Retrieval in RAG Systems for Large Legal Datasets EMNLP 2025
Retrieval-Augmented Generation (RAG) is a promising approach to mitigate hallucinations in Large Language Models (LLMs) for legal applications, but its reliability is critically dependent on the accuracy of the retrieval step. This is particularly challenging in the legal domain, where large databases of structurally similar documents often cause retrieval systems to fail. In this paper, we address this challenge by first identifying and quantifying a critical failure mode we term Document-Level Retrieval Mismatch (DRM), where the retriever selects information from entirely incorrect source documents. To mitigate DRM, we investigate a simple and computationally efficient technique which we refer to as Summary-Augmented Chunking (SAC). This method enhances each text chunk with a document-level synthetic summary, thereby injecting crucial global context that would otherwise be lost during a standard chunking process. Our experiments on a diverse set of legal information retrieval tasks show that SAC greatly reduces DRM and, consequently, also improves text-level retrieval precision and recall. Interestingly, we find that a generic summarization strategy outperforms an approach that incorporates legal expert domain knowledge to target specific legal elements. Our work provides evidence that this practical, scalable, and easily integrable technique enhances the reliability of RAG systems when applied to large-scale legal document datasets.
comment: Accepted for the 7th Natural Legal Language Processing Workshop (NLLP 2025), co-located with EMNLP 2025
☆ Spiral Model Technique For Data Science & Machine Learning Lifecycle
Analytics play an important role in modern business. Companies adapt data science lifecycles to their culture to seek productivity and improve their competitiveness among others. Data science lifecycles are fairly an important contributing factor to start and end a project that are data dependent. Data science and Machine learning life cycles comprises of series of steps that are involved in a project. A typical life cycle states that it is a linear or cyclical model that revolves around. It is mostly depicted that it is possible in a traditional data science life cycle to start the process again after reaching the end of cycle. This paper suggests a new technique to incorporate data science life cycle to business problems that have a clear end goal. A new technique called spiral technique is introduced to emphasize versatility, agility and iterative approach to business processes.
☆ Ethical AI prompt recommendations in large language models using collaborative filtering
As large language models (LLMs) shape AI development, ensuring ethical prompt recommendations is crucial. LLMs offer innovation but risk bias, fairness issues, and accountability concerns. Traditional oversight methods struggle with scalability, necessitating dynamic solutions. This paper proposes using collaborative filtering, a technique from recommendation systems, to enhance ethical prompt selection. By leveraging user interactions, it promotes ethical guidelines while reducing bias. Contributions include a synthetic dataset for prompt recommendations and the application of collaborative filtering. The work also tackles challenges in ethical AI, such as bias mitigation, transparency, and preventing unethical prompt engineering.
comment: This paper has been accepted to by the International Journal of Parallel, Emergent & Distributed Systems (Taylor and Francis) and has an assigned DOI. We have already chose to make this open access using CC BY. The article is not yet available online on the publisher's website. The DOI is: doi.org/10.1080/17445760.2025.2573086
☆ M3Retrieve: Benchmarking Multimodal Retrieval for Medicine EMNLP
With the increasing use of RetrievalAugmented Generation (RAG), strong retrieval models have become more important than ever. In healthcare, multimodal retrieval models that combine information from both text and images offer major advantages for many downstream tasks such as question answering, cross-modal retrieval, and multimodal summarization, since medical data often includes both formats. However, there is currently no standard benchmark to evaluate how well these models perform in medical settings. To address this gap, we introduce M3Retrieve, a Multimodal Medical Retrieval Benchmark. M3Retrieve, spans 5 domains,16 medical fields, and 4 distinct tasks, with over 1.2 Million text documents and 164K multimodal queries, all collected under approved licenses. We evaluate leading multimodal retrieval models on this benchmark to explore the challenges specific to different medical specialities and to understand their impact on retrieval performance. By releasing M3Retrieve, we aim to enable systematic evaluation, foster model innovation, and accelerate research toward building more capable and reliable multimodal retrieval systems for medical applications. The dataset and the baselines code are available in this github page https://github.com/AkashGhosh/M3Retrieve.
comment: EMNLP Mains 2025
☆ Crossing Domains without Labels: Distant Supervision for Term Extraction EMNLP
Automatic Term Extraction (ATE) is a critical component in downstream NLP tasks such as document tagging, ontology construction and patent analysis. Current state-of-the-art methods require expensive human annotation and struggle with domain transfer, limiting their practical deployment. This highlights the need for more robust, scalable solutions and realistic evaluation settings. To address this, we introduce a comprehensive benchmark spanning seven diverse domains, enabling performance evaluation at both the document- and corpus-levels. Furthermore, we propose a robust LLM-based model that outperforms both supervised cross-domain encoder models and few-shot learning baselines and performs competitively with its GPT-4o teacher on this benchmark. The first step of our approach is generating psuedo-labels with this black-box LLM on general and scientific domains to ensure generalizability. Building on this data, we fine-tune the first LLMs for ATE. To further enhance document-level consistency, oftentimes needed for downstream tasks, we introduce lightweight post-hoc heuristics. Our approach exceeds previous approaches on 5/7 domains with an average improvement of 10 percentage points. We release our dataset and fine-tuned models to support future research in this area.
comment: Accepted at EMNLP Industry Track 2025
☆ Exposing Citation Vulnerabilities in Generative Engines
We analyze answers generated by generative engines (GEs) from the perspectives of citation publishers and the content-injection barrier, defined as the difficulty for attackers to manipulate answers to user prompts by placing malicious content on the web. GEs integrate two functions: web search and answer generation that cites web pages using large language models. Because anyone can publish information on the web, GEs are vulnerable to poisoning attacks. Existing studies of citation evaluation focus on how faithfully answer content reflects cited sources, leaving unexamined which web sources should be selected as citations to defend against poisoning attacks. To fill this gap, we introduce evaluation criteria that assess poisoning threats using the citation information contained in answers. Our criteria classify the publisher attributes of citations to estimate the content-injection barrier thereby revealing the threat of poisoning attacks in current GEs. We conduct experiments in political domains in Japan and the United States (U.S.) using our criteria and show that citations from official party websites (primary sources) are approximately \(25\%\)--\(45\%\) in the U.S. and \(60\%\)--\(65\%\) in Japan, indicating that U.S. political answers are at higher risk of poisoning attacks. We also find that sources with low content-injection barriers are frequently cited yet are poorly reflected in answer content. To mitigate this threat, we discuss how publishers of primary sources can increase exposure of their web content in answers and show that well-known techniques are limited by language differences.
comment: 12 pages, under-reviewing at a conference
Overview of the Plagiarism Detection Task at PAN 2025
The generative plagiarism detection task at PAN 2025 aims at identifying automatically generated textual plagiarism in scientific articles and aligning them with their respective sources. We created a novel large-scale dataset of automatically generated plagiarism using three large language models: Llama, DeepSeek-R1, and Mistral. In this task overview paper, we outline the creation of this dataset, summarize and compare the results of all participants and four baselines, and evaluate the results on the last plagiarism detection task from PAN 2015 in order to interpret the robustness of the proposed approaches. We found that the current iteration does not invite a large variety of approaches as naive semantic similarity approaches based on embedding vectors provide promising results of up to 0.8 recall and 0.5 precision. In contrast, most of these approaches underperform significantly on the 2015 dataset, indicating a lack in generalizability.
comment: Working Notes at PAN at CLEF 2025
☆ Are LLMs Reliable Rankers? Rank Manipulation via Two-Stage Token Optimization
Large language models (LLMs) are increasingly used as rerankers in information retrieval, yet their ranking behavior can be steered by small, natural-sounding prompts. To expose this vulnerability, we present Rank Anything First (RAF), a two-stage token optimization method that crafts concise textual perturbations to consistently promote a target item in LLM-generated rankings while remaining hard to detect. Stage 1 uses Greedy Coordinate Gradient to shortlist candidate tokens at the current position by combining the gradient of the rank-target with a readability score; Stage 2 evaluates those candidates under exact ranking and readability losses using an entropy-based dynamic weighting scheme, and selects a token via temperature-controlled sampling. RAF generates ranking-promoting prompts token-by-token, guided by dual objectives: maximizing ranking effectiveness and preserving linguistic naturalness. Experiments across multiple LLMs show that RAF significantly boosts the rank of target items using naturalistic language, with greater robustness than existing methods in both promoting target items and maintaining naturalness. These findings underscore a critical security implication: LLM-based reranking is inherently susceptible to adversarial manipulation, raising new challenges for the trustworthiness and robustness of modern retrieval systems. Our code is available at: https://github.com/glad-lab/RAF.
comment: 10 pages, 3 figures
☆ Reproducing and Extending Causal Insights Into Term Frequency Computation in Neural Rankers SIGIR
Neural ranking models have shown outstanding performance across a variety of tasks, such as document retrieval, re-ranking, question answering and conversational retrieval. However, the inner decision process of these models remains largely unclear, especially as models increase in size. Most interpretability approaches, such as probing, focus on correlational insights rather than establishing causal relationships. The paper 'Axiomatic Causal Interventions for Reverse Engineering Relevance Computation in Neural Retrieval Models' by Chen et al. addresses this gap by introducing a framework for activation patching - a causal interpretability method - in the information retrieval domain, offering insights into how neural retrieval models compute document relevance. The study demonstrates that neural ranking models not only capture term-frequency information, but also that these representations can be localized to specific components of the model, such as individual attention heads or layers. This paper aims to reproduce the findings by Chen et al. and to further explore the presence of pre-defined retrieval axioms in neural IR models. We validate the main claims made by Chen et al., and extend the framework to include an additional term-frequency axiom, which states that the impact of increasing query term frequency on document ranking diminishes as the frequency becomes higher. We successfully identify a group of attention heads that encode this axiom and analyze their behavior to give insight into the inner decision-making process of neural ranking models.
comment: 10 pages, 6 figures, submitted to SIGIR-AP
☆ Can We Hide Machines in the Crowd? Quantifying Equivalence in LLM-in-the-loop Annotation Tasks SIGIR
Many evaluations of large language models (LLMs) in text annotation focus primarily on the correctness of the output, typically comparing model-generated labels to human-annotated ``ground truth'' using standard performance metrics. In contrast, our study moves beyond effectiveness alone. We aim to explore how labeling decisions -- by both humans and LLMs -- can be statistically evaluated across individuals. Rather than treating LLMs purely as annotation systems, we approach LLMs as an alternative annotation mechanism that may be capable of mimicking the subjective judgments made by humans. To assess this, we develop a statistical evaluation method based on Krippendorff's $\alpha$, paired bootstrapping, and the Two One-Sided t-Tests (TOST) equivalence test procedure. This evaluation method tests whether an LLM can blend into a group of human annotators without being distinguishable. We apply this approach to two datasets -- MovieLens 100K and PolitiFact -- and find that the LLM is statistically indistinguishable from a human annotator in the former ($p = 0.004$), but not in the latter ($p = 0.155$), highlighting task-dependent differences. It also enables early evaluation on a small sample of human data to inform whether LLMs are suitable for large-scale annotation in a given application.
comment: Accepted at SIGIR-AP 2025
☆ LLM-Powered Nuanced Video Attribute Annotation for Enhanced Recommendations RecSys 2025
This paper presents a case study on deploying Large Language Models (LLMs) as an advanced "annotation" mechanism to achieve nuanced content understanding (e.g., discerning content "vibe") at scale within a large-scale industrial short-form video recommendation system. Traditional machine learning classifiers for content understanding face protracted development cycles and a lack of deep, nuanced comprehension. The "LLM-as-annotators" approach addresses these by significantly shortening development times and enabling the annotation of subtle attributes. This work details an end-to-end workflow encompassing: (1) iterative definition and robust evaluation of target attributes, refined by offline metrics and online A/B testing; (2) scalable offline bulk annotation of video corpora using LLMs with multimodal features, optimized inference, and knowledge distillation for broad application; and (3) integration of these rich annotations into the online recommendation serving system, for example, through personalized restrict retrieval. Experimental results demonstrate the efficacy of this approach, with LLMs outperforming human raters in offline annotation quality for nuanced attributes and yielding significant improvements of user participation and satisfied consumption in online A/B tests. The study provides insights into designing and scaling production-level LLM pipelines for rich content evaluation, highlighting the adaptability and benefits of LLM-generated nuanced understanding for enhancing content discovery, user satisfaction, and the overall effectiveness of modern recommendation systems.
comment: RecSys 2025 Industry Track
☆ Retentive Relevance: Capturing Long-Term User Value in Recommendation Systems
Recommendation systems have traditionally relied on short-term engagement signals, such as clicks and likes, to personalize content. However, these signals are often noisy, sparse, and insufficient for capturing long-term user satisfaction and retention. We introduce Retentive Relevance, a novel content-level survey-based feedback measure that directly assesses users' intent to return to the platform for similar content. Unlike other survey measures that focus on immediate satisfaction, Retentive Relevance targets forward-looking behavioral intentions, capturing longer term user intentions and providing a stronger predictor of retention. We validate Retentive Relevance using psychometric methods, establishing its convergent, discriminant, and behavioral validity. Through large-scale offline modeling, we show that Retentive Relevance significantly outperforms both engagement signals and other survey measures in predicting next-day retention, especially for users with limited historical engagement. We develop a production-ready proxy model that integrates Retentive Relevance into the final stage of a multi-stage ranking system on a social media platform. Calibrated score adjustments based on this model yield substantial improvements in engagement, and retention, while reducing exposure to low-quality content, as demonstrated by large-scale A/B experiments. This work provides the first empirically validated framework linking content-level user perceptions to retention outcomes in production systems. We offer a scalable, user-centered solution that advances both platform growth and user experience. Our work has broad implications for responsible AI development.
☆ Evaluation of LLMs for Process Model Analysis and Optimization
In this paper, we report our experience with several LLMs for their ability to understand a process model in an interactive, conversational style, find syntactical and logical errors in it, and reason with it in depth through a natural language (NL) interface. Our findings show that a vanilla, untrained LLM like ChatGPT (model o3) in a zero-shot setting is effective in understanding BPMN process models from images and answering queries about them intelligently at syntactic, logic, and semantic levels of depth. Further, different LLMs vary in performance in terms of their accuracy and effectiveness. Nevertheless, our empirical analysis shows that LLMs can play a valuable role as assistants for business process designers and users. We also study the LLM's "thought process" and ability to perform deeper reasoning in the context of process analysis and optimization. We find that the LLMs seem to exhibit anthropomorphic properties.
comment: 15 pages, 5 tables, 4 figures; full research paper currently under review for the Workshop on Information Technologies and Systems (WITS) 2025. The paper presents a comprehensive evaluation of large language models (LLMs) for business process model analysis and optimization, including error detection, reasoning, and scenario-based redesign
☆ Reasoning by Exploration: A Unified Approach to Retrieval and Generation over Graphs
Reasoning over structured graphs remains a fundamental challenge for Large Language Models (LLMs), particularly when scaling to large graphs. Existing approaches typically follow the retrieval-augmented generation (RAG) paradigm: first retrieving subgraphs relevant to the query and then generating answers conditioned on the retrieved subgraphs. However, such two-phase pipelines often struggle to faithfully incorporate graph structure, since the generation process is ultimately constrained by the quality and completeness of the retrieved subgraph. Although many advanced retrievers have been proposed recently to mitigate this issue, they are usually tailored to the training graphs and generalize poorly to unseen graphs, which limits their practical applicability. In this work, we propose Reasoning by Exploration (RoE), a novel approach that unifies retrieval and generation by framing reasoning over graphs as a process of graph exploration. At each step, the LLM selects candidate nodes and edges to explore, gradually constructing reasoning paths and generating answers along the way. To enable effective exploration, RoE is trained in two stages: supervised fine-tuning (SFT) on gold reasoning paths, followed by reinforcement learning (RL) to enhance exploration effectiveness and generalization. Experiments on benchmark datasets demonstrate that RoE achieves substantial overall improvements over baselines, while also generalizing effectively to unseen graphs.
☆ Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation
Modern long-context large language models (LLMs) perform well on synthetic "needle-in-a-haystack" (NIAH) benchmarks, but such tests overlook how noisy contexts arise from biased retrieval and agentic workflows. We argue that haystack engineering is necessary to construct noisy long contexts that faithfully capture key real-world factors -- distraction from heterogeneous biased retrievers and cascading errors in agentic workflows -- to test models' long-context robustness. We instantiate it through HaystackCraft, a new NIAH benchmark built on the full English Wikipedia hyperlink network with multi-hop questions. HaystackCraft evaluates how heterogeneous retrieval strategies (e.g., sparse, dense, hybrid, and graph-based) affect distractor composition, haystack ordering, and downstream LLM performance. HaystackCraft further extends NIAH to dynamic, LLM-dependent settings that simulate agentic operations, where models refine queries, reflect on their past reasonings, and decide when to stop. Experiments with 15 long-context models show that (1) while stronger dense retrievers can introduce more challenging distractors, graph-based reranking simultaneously improves retrieval effectiveness and mitigates more harmful distractors; (2) in agentic tests, even advanced models like Gemini 2.5 Pro and GPT-5 suffer cascading failures from self-generated distractors or struggle to perform early stops. These results highlight persistent challenges in agentic long-context reasoning and establish HaystackCraft as a valuable testbed for future progress.
comment: Code available at https://github.com/Graph-COM/HaystackCraft
♻ ☆ SimpleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Trajectory Synthesis
Retrieval-augmented generation (RAG) systems have advanced large language models (LLMs) in complex deep search scenarios requiring multi-step reasoning and iterative information retrieval. However, existing approaches face critical limitations that lack high-quality training trajectories or suffer from the distributional mismatches in simulated environments and prohibitive computational costs for real-world deployment. This paper introduces SimpleDeepSearcher, a lightweight yet effective framework that bridges this gap through strategic data engineering rather than complex training paradigms. Our approach synthesizes high-quality training data by simulating realistic user interactions in live web search environments, coupled with a multi-criteria curation strategy that optimizes the diversity and quality of input and output side. Experiments on five benchmarks across diverse domains demonstrate that SFT on only 871 curated samples yields significant improvements over RL-based baselines. Our work establishes SFT as a viable pathway by systematically addressing the data-scarce bottleneck, offering practical insights for efficient deep search systems. Our code is available at https://github.com/RUCAIBox/SimpleDeepSearcher.
♻ ☆ Injecting External Knowledge into the Reasoning Process Enhances Retrieval-Augmented Generation SIGIR
Retrieval-augmented generation (RAG) has been widely adopted to augment large language models (LLMs) with external knowledge for knowledge-intensive tasks. However, its effectiveness is often undermined by the presence of noisy (i.e., low-quality) retrieved passages. Enhancing LLMs' robustness to such noise is critical for improving the reliability of RAG systems. Recent advances have equipped LLMs with strong reasoning and self-reflection capabilities, allowing them to identify and correct errors in their reasoning process. Inspired by this ability, we propose Passage Injection-a simple yet effective method that explicitly incorporates retrieved passages into LLMs' reasoning process, aiming to enhance the model's ability to recognize and resist noisy passages. We validate Passage Injection under general RAG settings using BM25 as the retriever. Experiments on four reasoning-enhanced LLMs across four factual QA datasets demonstrate that Passage Injection significantly improves overall RAG performance. Further analysis on two noisy retrieval settings-random noise, where the model is provided irrelevant passages, and counterfactual noise, where it is given misleading passages-shows that Passage Injection consistently improves robustness. Controlled experiments confirm that Passage Injection can also effectively leverage helpful passages. These findings suggest that incorporating passages in LLMs' reasoning process is a promising direction for building more robust RAG systems. The code can be found \href{here}{https://github.com/Trustworthy-Information-Access/Passage-Injection}.
comment: SIGIR-AP 2025
♻ ☆ Epistemic Diversity and Knowledge Collapse in Large Language Models
Large language models (LLMs) tend to generate lexically, semantically, and stylistically homogenous texts. This poses a risk of knowledge collapse, where homogenous LLMs mediate a shrinking in the range of accessible information over time. Existing works on homogenization are limited by a focus on closed-ended multiple-choice setups or fuzzy semantic features, and do not look at trends across time and cultural contexts. To overcome this, we present a new methodology to measure epistemic diversity, i.e., variation in real-world claims in LLM outputs, which we use to perform a broad empirical study of LLM knowledge collapse. We test 27 LLMs, 155 topics covering 12 countries, and 200 prompt variations sourced from real user chats. For the topics in our study, we show that while newer models tend to generate more diverse claims, nearly all models are less epistemically diverse than a basic web search. We find that model size has a negative impact on epistemic diversity, while retrieval-augmented generation (RAG) has a positive impact, though the improvement from RAG varies by the cultural context. Finally, compared to a traditional knowledge source (Wikipedia), we find that country-specific claims reflect the English language more than the local one, highlighting a gap in epistemic representation
comment: 16 pages; 8 figures, 4 tables; v2 changelog: Fixed the modeling for table 3, random effect is the model version; v3 changelog: Fixed minor formatting issues in tables 2 and 3;
♻ ☆ Do RAG Systems Really Suffer From Positional Bias?
Retrieval Augmented Generation enhances LLM accuracy by adding passages retrieved from an external corpus to the LLM prompt. This paper investigates how positional bias - the tendency of LLMs to weight information differently based on its position in the prompt - affects not only the LLM's capability to capitalize on relevant passages, but also its susceptibility to distracting passages. Through extensive experiments on three benchmarks, we show how state-of-the-art retrieval pipelines, while attempting to retrieve relevant passages, systematically bring highly distracting ones to the top ranks, with over 60% of queries containing at least one highly distracting passage among the top-10 retrieved passages. As a result, the impact of the LLM positional bias, which in controlled settings is often reported as very prominent by related works, is actually marginal in real scenarios since both relevant and distracting passages are, in turn, penalized. Indeed, our findings reveal that sophisticated strategies that attempt to rearrange the passages based on LLM positional preferences do not perform better than random shuffling.
♻ ☆ Reasoning-enhanced Query Understanding through Decomposition and Interpretation
Accurate inference of user intent is crucial for enhancing document retrieval in modern search engines. While large language models (LLMs) have made significant strides in this area, their effectiveness has predominantly been assessed with short, keyword-based queries. As AI-driven search evolves, long-form queries with intricate intents are becoming more prevalent, yet they remain underexplored in the context of LLM-based query understanding (QU). To bridge this gap, we introduce ReDI: a Reasoning-enhanced approach for query understanding through Decomposition and Interpretation. ReDI leverages the reasoning and comprehension capabilities of LLMs in a three-stage pipeline: (i) it breaks down complex queries into targeted sub-queries to accurately capture user intent; (ii) it enriches each sub-query with detailed semantic interpretations to improve the query-document matching; and (iii) it independently retrieves documents for each sub-query and employs a fusion strategy to aggregate the results for the final ranking. We compiled a large-scale dataset of real-world complex queries from a major search engine and distilled the query understanding capabilities of teacher models into smaller models for practical application. Experiments on BRIGHT and BEIR demonstrate that ReDI consistently surpasses strong baselines in both sparse and dense retrieval paradigms, affirming its effectiveness. We release our code at https://anonymous.4open.science/r/ReDI-6FC7/.
♻ ☆ TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling NeurIPS
While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited, leaving other simpler yet crucial components such as metadata or attribute filtering underutilized in the system. We propose an LLM-based music recommendation system with tool calling to serve as a unified retrieval-reranking pipeline. Our system positions an LLM as an end-to-end recommendation system that interprets user intent, plans tool invocations, and orchestrates specialized components: boolean filters (SQL), sparse retrieval (BM25), dense retrieval (embedding similarity), and generative retrieval (semantic IDs). Through tool planning, the system predicts which types of tools to use, their execution order, and the arguments needed to find music matching user preferences, supporting diverse modalities while seamlessly integrating multiple database filtering methods. We demonstrate that this unified tool-calling framework achieves competitive performance across diverse recommendation scenarios by selectively employing appropriate retrieval methods based on user queries, envisioning a new paradigm for conversational music recommendation systems.
comment: Accepted for publication at The Workshop on AI for Music, Neural Information Processing Systems (NeurIPS-AI4Music)
♻ ☆ Membership Inference Attacks on LLM-based Recommender Systems
Large language models (LLMs) based Recommender Systems (RecSys) can flexibly adapt recommendation systems to different domains. It utilizes in-context learning (ICL), i.e., the prompts, to customize the recommendation functions, which include sensitive historical user-specific item interactions, e.g., implicit feedback like clicked items or explicit product reviews. Such private information may be exposed to novel privacy attack. However, no study has been done on this important issue. We design four membership inference attacks (MIAs), aiming to reveal whether victims' historical interactions have been used by system prompts. They are \emph{direct inquiry, hallucination, similarity, and poisoning attacks}, each of which utilizes the unique features of LLMs or RecSys. We have carefully evaluated them on three LLMs that have been used to develop ICL-LLM RecSys and two well-known RecSys benchmark datasets. The results confirm that the MIA threat on LLM RecSys is realistic: direct inquiry and poisoning attacks showing significantly high attack advantages. We have also analyzed the factors affecting these attacks, such as the number of shots in system prompts and the position of the victim in the shots.
comment: this paper is under review
♻ ☆ Scalable In-context Ranking with Generative Models
In-context Ranking (ICR) is an emerging paradigm for Information Retrieval (IR), which leverages contextual understanding of LLMs by directly incorporating the task description, candidate documents, and the query into the model's input prompt and tasking the LLM to identify relevant document(s). While it is effective, efficiency is a significant challenge in this paradigm, especially as the candidate list grows due to quadratic/super-linear scaling of attention operation with context length. To this end, this paper first identifies inherent and exploitable structures in the attention of LLMs finetuned for ICR: (1) inter-document block sparsity: attention is dense within each document block but sparse across different documents in the context; and (2) query-document block relevance: the attention scores from certain query tokens to a document block in middle layers strongly correlate with that document's actual relevance. Motivated by these observations, we introduce BlockRank (Blockwise In-context Ranking), a novel method that adapts the attention operation in an LLM by (a) architecturally enforcing the observed inter-document block sparsity, reducing attention complexity from quadratic to linear without loss in performance, and (b) optimizing query-document block relevance for true relevant documents during fine-tuning using an auxiliary contrastive training objective, improving retrieval in attention. Experiments on BEIR, MSMarco and NQ with Mistral-7B demonstrate that BlockRank Mistral matches or outperforms existing SOTA listwise rankers and controlled fine-tuned baseline while being significantly more efficient at inference (4.7x for 100 MSMarco documents in context) and scaling gracefully to long-context shortlists, around 500 documents in-context (approximately 100K context length) within a second, presenting a scalable and effective solution for ICR.
♻ ☆ Advancing AI Research Assistants with Expert-Involved Learning
Large language models (LLMs) and large multimodal models (LMMs) promise to accelerate biomedical discovery, yet their reliability remains unclear. We introduce ARIEL (AI Research Assistant for Expert-in-the-Loop Learning), an open-source evaluation and optimization framework that pairs a curated multimodal biomedical corpus with expert-vetted tasks to probe two capabilities: full-length article summarization and fine-grained figure interpretation. Using uniform protocols and blinded PhD-level evaluation, we find that state-of-the-art models generate fluent but incomplete summaries, whereas LMMs struggle with detailed visual reasoning. We later observe that prompt engineering and lightweight fine-tuning substantially improve textual coverage, and a compute-scaled inference strategy enhances visual question answering. We build an ARIEL agent that integrates textual and visual cues, and we show it can propose testable mechanistic hypotheses. ARIEL delineates current strengths and limitations of foundation models, and provides a reproducible platform for advancing trustworthy AI in biomedicine.
comment: 36 pages, 7 figures
♻ ☆ TalkPlayData 2: An Agentic Synthetic Data Pipeline for Multimodal Conversational Music Recommendation
We present TalkPlayData 2, a synthetic dataset for multimodal conversational music recommendation generated by an agentic data pipeline. In the proposed pipeline, multiple large language model (LLM) agents are created under various roles with specialized prompts and access to different parts of information, and the chat data is acquired by logging the conversation between the Listener LLM and the Recsys LLM. To cover various conversation scenarios, for each conversation, the Listener LLM is conditioned on a finetuned conversation goal. Finally, all the LLMs are multimodal with audio and images, allowing a simulation of multimodal recommendation and conversation. In the LLM-as-a-judge and subjective evaluation experiments, TalkPlayData 2 achieved the proposed goal in various aspects related to training a generative recommendation model for music. TalkPlayData 2 and its generation code are released at https://talkpl.ai/talkplaydata2.
comment: 2025-10-08: updating the stat table with the latest numbers. updated the abstract per the latest license terms
Machine Learning 150
☆ Artificial Hippocampus Networks for Efficient Long-Context Modeling
Long-sequence modeling faces a fundamental trade-off between the efficiency of compressive fixed-size memory in RNN-like models and the fidelity of lossless growing memory in attention-based Transformers. Inspired by the Multi-Store Model in cognitive science, we introduce a memory framework of artificial neural networks. Our method maintains a sliding window of the Transformer's KV cache as lossless short-term memory, while a learnable module termed Artificial Hippocampus Network (AHN) recurrently compresses out-of-window information into a fixed-size compact long-term memory. To validate this framework, we instantiate AHNs using modern RNN-like architectures, including Mamba2, DeltaNet, and Gated DeltaNet. Extensive experiments on long-context benchmarks LV-Eval and InfiniteBench demonstrate that AHN-augmented models consistently outperform sliding window baselines and achieve performance comparable or even superior to full-attention models, while substantially reducing computational and memory requirements. For instance, augmenting the Qwen2.5-3B-Instruct with AHNs reduces inference FLOPs by 40.5% and memory cache by 74.0%, while improving its average score on LV-Eval (128k sequence length) from 4.41 to 5.88. Code is available at: https://github.com/ByteDance-Seed/AHN.
comment: Code: https://github.com/ByteDance-Seed/AHN
☆ Vibe Checker: Aligning Code Evaluation with Human Preference
Large Language Models (LLMs) have catalyzed vibe coding, where users leverage LLMs to generate and iteratively refine code through natural language interactions until it passes their vibe check. Vibe check is tied to real-world human preference and goes beyond functionality: the solution should feel right, read cleanly, preserve intent, and remain correct. However, current code evaluation remains anchored to pass@k and captures only functional correctness, overlooking the non-functional instructions that users routinely apply. In this paper, we hypothesize that instruction following is the missing piece underlying vibe check that represents human preference in coding besides functional correctness. To quantify models' code instruction following capabilities with measurable signals, we present VeriCode, a taxonomy of 30 verifiable code instructions together with corresponding deterministic verifiers. We use the taxonomy to augment established evaluation suites, resulting in Vibe Checker, a testbed to assess both code instruction following and functional correctness. Upon evaluating 31 leading LLMs, we show that even the strongest models struggle to comply with multiple instructions and exhibit clear functional regression. Most importantly, a composite score of functional correctness and instruction following correlates the best with human preference, with the latter emerging as the primary differentiator on real-world programming tasks. Our work identifies core factors of the vibe check, providing a concrete path for benchmarking and developing models that better align with user preferences in coding.
comment: Preprint
☆ h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning
Large language models excel at short-horizon reasoning tasks, but performance drops as reasoning horizon lengths increase. Existing approaches to combat this rely on inference-time scaffolding or costly step-level supervision, neither of which scales easily. In this work, we introduce a scalable method to bootstrap long-horizon reasoning capabilities using only existing, abundant short-horizon data. Our approach synthetically composes simple problems into complex, multi-step dependency chains of arbitrary length. We train models on this data using outcome-only rewards under a curriculum that automatically increases in complexity, allowing RL training to be scaled much further without saturating. Empirically, our method generalizes remarkably well: curriculum training on composed 6th-grade level math problems (GSM8K) boosts accuracy on longer, competition-level benchmarks (GSM-Symbolic, MATH-500, AIME) by up to 2.06x. Importantly, our long-horizon improvements are significantly higher than baselines even at high pass@k, showing that models can learn new reasoning paths under RL. Theoretically, we show that curriculum RL with outcome rewards achieves an exponential improvement in sample complexity over full-horizon training, providing training signal comparable to dense supervision. h1 therefore introduces an efficient path towards scaling RL for long-horizon problems using only existing data.
comment: Preprint, 31 pages, 8 figures
☆ MLE-Smith: Scaling MLE Tasks with Automated Multi-Agent Pipeline
While Language Models (LMs) have made significant progress in automating machine learning engineering (MLE), the acquisition of high-quality MLE training data is significantly constrained. Current MLE benchmarks suffer from low scalability and limited applicability because they rely on static, manually curated tasks, demanding extensive time and manual effort to produce. We introduce MLE-Smith, a fully automated multi-agent pipeline, to transform raw datasets into competition-style MLE challenges through an efficient generate-verify-execute paradigm for scaling MLE tasks with verifiable quality, real-world usability, and rich diversity. The proposed multi-agent pipeline in MLE-Smith drives structured task design and standardized refactoring, coupled with a hybrid verification mechanism that enforces strict structural rules and high-level semantic soundness. It further validates empirical solvability and real-world fidelity through interactive execution. We apply MLE-Smith to 224 of real-world datasets and generate 606 tasks spanning multiple categories, objectives, and modalities, demonstrating that MLE-Smith can work effectively across a wide range of real-world datasets. Evaluation on the generated tasks shows that the performance of eight mainstream and cutting-edge LLMs on MLE-Smith tasks is strongly correlated with their performance on carefully human-designed tasks, highlighting the effectiveness of the MLE-Smith to scaling up MLE tasks, while maintaining task quality.
☆ Cocoon: A System Architecture for Differentially Private Training with Correlated Noises
Machine learning (ML) models memorize and leak training data, causing serious privacy issues to data owners. Training algorithms with differential privacy (DP), such as DP-SGD, have been gaining attention as a solution. However, DP-SGD adds a noise at each training iteration, which degrades the accuracy of the trained model. To improve accuracy, a new family of approaches adds carefully designed correlated noises, so that noises cancel out each other across iterations. We performed an extensive characterization study of these new mechanisms, for the first time to the best of our knowledge, and show they incur non-negligible overheads when the model is large or uses large embedding tables. Motivated by the analysis, we propose Cocoon, a hardware-software co-designed framework for efficient training with correlated noises. Cocoon accelerates models with embedding tables through pre-computing and storing correlated noises in a coalesced format (Cocoon-Emb), and supports large models through a custom near-memory processing device (Cocoon-NMP). On a real system with an FPGA-based NMP device prototype, Cocoon improves the performance by 2.33-10.82x(Cocoon-Emb) and 1.55-3.06x (Cocoon-NMP).
On the Convergence of Moral Self-Correction in Large Language Models
Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only a general and abstract goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. Focusing on moral self-correction in LLMs, we reveal a key characteristic of intrinsic self-correction: performance convergence through multi-round interactions; and provide a mechanistic analysis of this convergence behavior. Based on our experimental results and analysis, we uncover the underlying mechanism of convergence: consistently injected self-correction instructions activate moral concepts that reduce model uncertainty, leading to converged performance as the activated moral concepts stabilize over successive rounds. This paper demonstrates the strong potential of moral self-correction by showing that it exhibits a desirable property of converged performance.
comment: 19pages, 7 figures
☆ MolGA: Molecular Graph Adaptation with Pre-trained 2D Graph Encoder
Molecular graph representation learning is widely used in chemical and biomedical research. While pre-trained 2D graph encoders have demonstrated strong performance, they overlook the rich molecular domain knowledge associated with submolecular instances (atoms and bonds). While molecular pre-training approaches incorporate such knowledge into their pre-training objectives, they typically employ designs tailored to a specific type of knowledge, lacking the flexibility to integrate diverse knowledge present in molecules. Hence, reusing widely available and well-validated pre-trained 2D encoders, while incorporating molecular domain knowledge during downstream adaptation, offers a more practical alternative. In this work, we propose MolGA, which adapts pre-trained 2D graph encoders to downstream molecular applications by flexibly incorporating diverse molecular domain knowledge. First, we propose a molecular alignment strategy that bridge the gap between pre-trained topological representations with domain-knowledge representations. Second, we introduce a conditional adaptation mechanism that generates instance-specific tokens to enable fine-grained integration of molecular domain knowledge for downstream tasks. Finally, we conduct extensive experiments on eleven public datasets, demonstrating the effectiveness of MolGA.
comment: Under review
☆ Evolutionary Profiles for Protein Fitness Prediction
Predicting the fitness impact of mutations is central to protein engineering but constrained by limited assays relative to the size of sequence space. Protein language models (pLMs) trained with masked language modeling (MLM) exhibit strong zero-shot fitness prediction; we provide a unifying view by interpreting natural evolution as implicit reward maximization and MLM as inverse reinforcement learning (IRL), in which extant sequences act as expert demonstrations and pLM log-odds serve as fitness estimates. Building on this perspective, we introduce EvoIF, a lightweight model that integrates two complementary sources of evolutionary signal: (i) within-family profiles from retrieved homologs and (ii) cross-family structural-evolutionary constraints distilled from inverse folding logits. EvoIF fuses sequence-structure representations with these profiles via a compact transition block, yielding calibrated probabilities for log-odds scoring. On ProteinGym (217 mutational assays; >2.5M mutants), EvoIF and its MSA-enabled variant achieve state-of-the-art or competitive performance while using only 0.15% of the training data and fewer parameters than recent large models. Ablations confirm that within-family and cross-family profiles are complementary, improving robustness across function types, MSA depths, taxa, and mutation depths. The codes will be made publicly available at https://github.com/aim-uofa/EvoIF.
☆ GTCN-G: A Residual Graph-Temporal Fusion Network for Imbalanced Intrusion Detection (Preprint)
The escalating complexity of network threats and the inherent class imbalance in traffic data present formidable challenges for modern Intrusion Detection Systems (IDS). While Graph Neural Networks (GNNs) excel in modeling topological structures and Temporal Convolutional Networks (TCNs) are proficient in capturing time-series dependencies, a framework that synergistically integrates both while explicitly addressing data imbalance remains an open challenge. This paper introduces a novel deep learning framework, named Gated Temporal Convolutional Network and Graph (GTCN-G), engineered to overcome these limitations. Our model uniquely fuses a Gated TCN (G-TCN) for extracting hierarchical temporal features from network flows with a Graph Convolutional Network (GCN) designed to learn from the underlying graph structure. The core innovation lies in the integration of a residual learning mechanism, implemented via a Graph Attention Network (GAT). This mechanism preserves original feature information through residual connections, which is critical for mitigating the class imbalance problem and enhancing detection sensitivity for rare malicious activities (minority classes). We conducted extensive experiments on two public benchmark datasets, UNSW-NB15 and ToN-IoT, to validate our approach. The empirical results demonstrate that the proposed GTCN-G model achieves state-of-the-art performance, significantly outperforming existing baseline models in both binary and multi-class classification tasks.
comment: This preprint was submitted to IEEE TrustCom 2025. The accepted version will be published under copyright 2025 IEEE
☆ Online Rubrics Elicitation from Pairwise Comparisons
Rubrics provide a flexible way to train LLMs on open-ended long-form answers where verifiable rewards are not applicable and human preferences provide coarse signals. Prior work shows that reinforcement learning with rubric-based rewards leads to consistent gains in LLM post-training. Most existing approaches rely on rubrics that remain static over the course of training. Such static rubrics, however, are vulnerable to reward-hacking type behaviors and fail to capture emergent desiderata that arise during training. We introduce Online Rubrics Elicitation (OnlineRubrics), a method that dynamically curates evaluation criteria in an online manner through pairwise comparisons of responses from current and reference policies. This online process enables continuous identification and mitigation of errors as training proceeds. Empirically, this approach yields consistent improvements of up to 8% over training exclusively with static rubrics across AlpacaEval, GPQA, ArenaHard as well as the validation sets of expert questions and rubrics. We qualitatively analyze the elicited criteria and identify prominent themes such as transparency, practicality, organization, and reasoning.
☆ Dynamic Regret Bounds for Online Omniprediction with Long Term Constraints
We present an algorithm guaranteeing dynamic regret bounds for online omniprediction with long term constraints. The goal in this recently introduced problem is for a learner to generate a sequence of predictions which are broadcast to a collection of downstream decision makers. Each decision maker has their own utility function, as well as a vector of constraint functions, each mapping their actions and an adversarially selected state to reward or constraint violation terms. The downstream decision makers select actions "as if" the state predictions are correct, and the goal of the learner is to produce predictions such that all downstream decision makers choose actions that give them worst-case utility guarantees while minimizing worst-case constraint violation. Within this framework, we give the first algorithm that obtains simultaneous \emph{dynamic regret} guarantees for all of the agents -- where regret for each agent is measured against a potentially changing sequence of actions across rounds of interaction, while also ensuring vanishing constraint violation for each agent. Our results do not require the agents themselves to maintain any state -- they only solve one-round constrained optimization problems defined by the prediction made at that round.
☆ Test-Time Graph Search for Goal-Conditioned Reinforcement Learning
Offline goal-conditioned reinforcement learning (GCRL) trains policies that reach user-specified goals at test time, providing a simple, unsupervised, domain-agnostic way to extract diverse behaviors from unlabeled, reward-free datasets. Nonetheless, long-horizon decision making remains difficult for GCRL agents due to temporal credit assignment and error accumulation, and the offline setting amplifies these effects. To alleviate this issue, we introduce Test-Time Graph Search (TTGS), a lightweight planning approach to solve the GCRL task. TTGS accepts any state-space distance or cost signal, builds a weighted graph over dataset states, and performs fast search to assemble a sequence of subgoals that a frozen policy executes. When the base learner is value-based, the distance is derived directly from the learned goal-conditioned value function, so no handcrafted metric is needed. TTGS requires no changes to training, no additional supervision, no online interaction, and no privileged information, and it runs entirely at inference. On the OGBench benchmark, TTGS improves success rates of multiple base learners on challenging locomotion tasks, demonstrating the benefit of simple metric-guided test-time planning for offline GCRL.
☆ Discriminative Feature Feedback with General Teacher Classes
We study the theoretical properties of the interactive learning protocol Discriminative Feature Feedback (DFF) (Dasgupta et al., 2018). The DFF learning protocol uses feedback in the form of discriminative feature explanations. We provide the first systematic study of DFF in a general framework that is comparable to that of classical protocols such as supervised learning and online learning. We study the optimal mistake bound of DFF in the realizable and the non-realizable settings, and obtain novel structural results, as well as insights into the differences between Online Learning and settings with richer feedback such as DFF. We characterize the mistake bound in the realizable setting using a new notion of dimension. In the non-realizable setting, we provide a mistake upper bound and show that it cannot be improved in general. Our results show that unlike Online Learning, in DFF the realizable dimension is insufficient to characterize the optimal non-realizable mistake bound or the existence of no-regret algorithms.
☆ Hybrid Reinforcement: When Reward Is Sparse, It's Better to Be Dense
Post-training for reasoning of large language models (LLMs) increasingly relies on verifiable rewards: deterministic checkers that provide 0-1 correctness signals. While reliable, such binary feedback is brittle--many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning.
comment: 20 pages
☆ A Broader View of Thompson Sampling
Thompson Sampling is one of the most widely used and studied bandit algorithms, known for its simple structure, low regret performance, and solid theoretical guarantees. Yet, in stark contrast to most other families of bandit algorithms, the exact mechanism through which posterior sampling (as introduced by Thompson) is able to "properly" balance exploration and exploitation, remains a mystery. In this paper we show that the core insight to address this question stems from recasting Thompson Sampling as an online optimization algorithm. To distill this, a key conceptual tool is introduced, which we refer to as "faithful" stationarization of the regret formulation. Essentially, the finite horizon dynamic optimization problem is converted into a stationary counterpart which "closely resembles" the original objective (in contrast, the classical infinite horizon discounted formulation, that leads to the Gittins index, alters the problem and objective in too significant a manner). The newly crafted time invariant objective can be studied using Bellman's principle which leads to a time invariant optimal policy. When viewed through this lens, Thompson Sampling admits a simple online optimization form that mimics the structure of the Bellman-optimal policy, and where greediness is regularized by a measure of residual uncertainty based on point-biserial correlation. This answers the question of how Thompson Sampling balances exploration-exploitation, and moreover, provides a principled framework to study and further improve Thompson's original idea.
☆ Guided by the Experts: Provable Feature Learning Dynamic of Soft-Routed Mixture-of-Experts
Mixture-of-Experts (MoE) architectures have emerged as a cornerstone of modern AI systems. In particular, MoEs route inputs dynamically to specialized experts whose outputs are aggregated through weighted summation. Despite their widespread application, theoretical understanding of MoE training dynamics remains limited to either separate expert-router optimization or only top-1 routing scenarios with carefully constructed datasets. This paper advances MoE theory by providing convergence guarantees for joint training of soft-routed MoE models with non-linear routers and experts in a student-teacher framework. We prove that, with moderate over-parameterization, the student network undergoes a feature learning phase, where the router's learning process is ``guided'' by the experts, that recovers the teacher's parameters. Moreover, we show that a post-training pruning can effectively eliminate redundant neurons, followed by a provably convergent fine-tuning process that reaches global optimality. To our knowledge, our analysis is the first to bring novel insights in understanding the optimization landscape of the MoE architecture.
☆ An in-depth look at approximation via deep and narrow neural networks
In 2017, Hanin and Sellke showed that the class of arbitrarily deep, real-valued, feed-forward and ReLU-activated networks of width w forms a dense subset of the space of continuous functions on R^n, with respect to the topology of uniform convergence on compact sets, if and only if w>n holds. To show the necessity, a concrete counterexample function f:R^n->R was used. In this note we actually approximate this very f by neural networks in the two cases w=n and w=n+1 around the aforementioned threshold. We study how the approximation quality behaves if we vary the depth and what effect (spoiler alert: dying neurons) cause that behavior.
comment: 11 pages
☆ Accelerating Inference for Multilayer Neural Networks with Quantum Computers
Fault-tolerant Quantum Processing Units (QPUs) promise to deliver exponential speed-ups in select computational tasks, yet their integration into modern deep learning pipelines remains unclear. In this work, we take a step towards bridging this gap by presenting the first fully-coherent quantum implementation of a multilayer neural network with non-linear activation functions. Our constructions mirror widely used deep learning architectures based on ResNet, and consist of residual blocks with multi-filter 2D convolutions, sigmoid activations, skip-connections, and layer normalizations. We analyse the complexity of inference for networks under three quantum data access regimes. Without any assumptions, we establish a quadratic speedup over classical methods for shallow bilinear-style networks. With efficient quantum access to the weights, we obtain a quartic speedup over classical methods. With efficient quantum access to both the inputs and the network weights, we prove that a network with an $N$-dimensional vectorized input, $k$ residual block layers, and a final residual-linear-pooling layer can be implemented with an error of $\epsilon$ with $O(\text{polylog}(N/\epsilon)^k)$ inference cost.
☆ Covert Quantum Learning: Privately and Verifiably Learning from Quantum Data
Quantum learning from remotely accessed quantum compute and data must address two key challenges: verifying the correctness of data and ensuring the privacy of the learner's data-collection strategies and resulting conclusions. The covert (verifiable) learning model of Canetti and Karchmer (TCC 2021) provides a framework for endowing classical learning algorithms with such guarantees. In this work, we propose models of covert verifiable learning in quantum learning theory and realize them without computational hardness assumptions for remote data access scenarios motivated by established quantum data advantages. We consider two privacy notions: (i) strategy-covertness, where the eavesdropper does not gain information about the learner's strategy; and (ii) target-covertness, where the eavesdropper does not gain information about the unknown object being learned. We show: Strategy-covert algorithms for making quantum statistical queries via classical shadows; Target-covert algorithms for learning quadratic functions from public quantum examples and private quantum statistical queries, for Pauli shadow tomography and stabilizer state learning from public multi-copy and private single-copy quantum measurements, and for solving Forrelation and Simon's problem from public quantum queries and private classical queries, where the adversary is a unidirectional or i.i.d. ancilla-free eavesdropper. The lattermost results in particular establish that the exponential separation between classical and quantum queries for Forrelation and Simon's problem survives under covertness constraints. Along the way, we design covert verifiable protocols for quantum data acquisition from public quantum queries which may be of independent interest. Overall, our models and corresponding algorithms demonstrate that quantum advantages are privately and verifiably achievable even with untrusted, remote data.
comment: 16 + 54 pages
☆ Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples
Poisoning attacks can compromise the safety of large language models (LLMs) by injecting malicious documents into their training data. Existing work has studied pretraining poisoning assuming adversaries control a percentage of the training corpus. However, for large models, even small percentages translate to impractically large amounts of data. This work demonstrates for the first time that poisoning attacks instead require a near-constant number of documents regardless of dataset size. We conduct the largest pretraining poisoning experiments to date, pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes, despite the largest models training on more than 20 times more clean data. We also run smaller-scale experiments to ablate factors that could influence attack success, including broader ratios of poisoned to clean data and non-random distributions of poisoned samples. Finally, we demonstrate the same dynamics for poisoning during fine-tuning. Altogether, our results suggest that injecting backdoors through data poisoning may be easier for large models than previously believed as the number of poisons required does not scale up with model size, highlighting the need for more research on defences to mitigate this risk in future models.
☆ Resolution scaling governs DINOv3 transfer performance in chest radiograph classification
Self-supervised learning (SSL) has advanced visual representation learning, but its value in chest radiography, a high-volume imaging modality with fine-grained findings, remains unclear. Meta's DINOv3 extends earlier SSL models through Gram-anchored self-distillation. Whether these design choices improve transfer learning for chest radiography has not been systematically tested. We benchmarked DINOv3 against DINOv2 and ImageNet initialization across seven datasets (n>814,000). Two representative backbones were evaluated: ViT-B/16 and ConvNeXt-B. Images were analyzed at 224x224, 512x512, and 1024x1024 pixels. We additionally assessed frozen features from a 7B model. The primary outcome was mean AUROC across labels. At 224x224, DINOv3 and DINOv2 achieved comparable performance on adult datasets. Increasing resolution to 512x512 yielded consistent improvements for DINOv3 over both DINOv2 and ImageNet. In contrast, results in pediatric cohort showed no differences across initializations. Across all settings, ConvNeXt-B outperformed ViT-B/16. Models using frozen DINOv3-7B features underperformed relative to fully finetuned 86-89M-parameter backbones, highlighting the importance of domain adaptation. Scaling to 1024x1024 did not further improve accuracy. Resolution-related gains were most evident for boundary-dependent and small focal abnormalities. In chest radiography, higher input resolution is critical for leveraging the benefits of modern self-supervised models. 512x512 pixels represent a practical upper limit where DINOv3-initialized ConvNeXt-B networks provide the strongest performance, while larger inputs offer minimal return on cost. Clinically, these findings support use of finetuned, mid-sized backbones at 512x512 for chest radiograph interpretation, with the greatest gains expected in detecting subtle or boundary-centered lesions relevant to emergency and critical care settings.
☆ Split Conformal Classification with Unsupervised Calibration
Methods for split conformal prediction leverage calibration samples to transform any prediction rule into a set-prediction rule that complies with a target coverage probability. Existing methods provide remarkably strong performance guarantees with minimal computational costs. However, they require to use calibration samples composed by labeled examples different to those used for training. This requirement can be highly inconvenient, as it prevents the use of all labeled examples for training and may require acquiring additional labels solely for calibration. This paper presents an effective methodology for split conformal prediction with unsupervised calibration for classification tasks. In the proposed approach, set-prediction rules are obtained using unsupervised calibration samples together with supervised training samples previously used to learn the classification rule. Theoretical and experimental results show that the presented methods can achieve performance comparable to that with supervised calibration, at the expenses of a moderate degradation in performance guarantees and computational efficiency.
☆ Bridged Clustering for Representation Learning: Semi-Supervised Sparse Bridging
We introduce Bridged Clustering, a semi-supervised framework to learn predictors from any unpaired input $X$ and output $Y$ dataset. Our method first clusters $X$ and $Y$ independently, then learns a sparse, interpretable bridge between clusters using only a few paired examples. At inference, a new input $x$ is assigned to its nearest input cluster, and the centroid of the linked output cluster is returned as the prediction $\hat{y}$. Unlike traditional SSL, Bridged Clustering explicitly leverages output-only data, and unlike dense transport-based methods, it maintains a sparse and interpretable alignment. Through theoretical analysis, we show that with bounded mis-clustering and mis-bridging rates, our algorithm becomes an effective and efficient predictor. Empirically, our method is competitive with SOTA methods while remaining simple, model-agnostic, and highly label-efficient in low-supervision settings.
☆ Bayesian Portfolio Optimization by Predictive Synthesis
Portfolio optimization is a critical task in investment. Most existing portfolio optimization methods require information on the distribution of returns of the assets that make up the portfolio. However, such distribution information is usually unknown to investors. Various methods have been proposed to estimate distribution information, but their accuracy greatly depends on the uncertainty of the financial markets. Due to this uncertainty, a model that could well predict the distribution information at one point in time may perform less accurately compared to another model at a different time. To solve this problem, we investigate a method for portfolio optimization based on Bayesian predictive synthesis (BPS), one of the Bayesian ensemble methods for meta-learning. We assume that investors have access to multiple asset return prediction models. By using BPS with dynamic linear models to combine these predictions, we can obtain a Bayesian predictive posterior about the mean rewards of assets that accommodate the uncertainty of the financial markets. In this study, we examine how to construct mean-variance portfolios and quantile-based portfolios based on the predicted distribution information.
☆ Quantifying Data Contamination in Psychometric Evaluations of LLMs
Recent studies apply psychometric questionnaires to Large Language Models (LLMs) to assess high-level psychological constructs such as values, personality, moral foundations, and dark traits. Although prior work has raised concerns about possible data contamination from psychometric inventories, which may threaten the reliability of such evaluations, there has been no systematic attempt to quantify the extent of this contamination. To address this gap, we propose a framework to systematically measure data contamination in psychometric evaluations of LLMs, evaluating three aspects: (1) item memorization, (2) evaluation memorization, and (3) target score matching. Applying this framework to 21 models from major families and four widely used psychometric inventories, we provide evidence that popular inventories such as the Big Five Inventory (BFI-44) and Portrait Values Questionnaire (PVQ-40) exhibit strong contamination, where models not only memorize items but can also adjust their responses to achieve specific target scores.
comment: 12 pages, 1 figure
☆ NurseLLM: The First Specialized Language Model for Nursing EMNLP 2025
Recent advancements in large language models (LLMs) have significantly transformed medical systems. However, their potential within specialized domains such as nursing remains largely underexplored. In this work, we introduce NurseLLM, the first nursing-specialized LLM tailored for multiple choice question-answering (MCQ) tasks. We develop a multi-stage data generation pipeline to build the first large scale nursing MCQ dataset to train LLMs on a broad spectrum of nursing topics. We further introduce multiple nursing benchmarks to enable rigorous evaluation. Our extensive experiments demonstrate that NurseLLM outperforms SoTA general-purpose and medical-specialized LLMs of comparable size on different benchmarks, underscoring the importance of a specialized LLM for the nursing domain. Finally, we explore the role of reasoning and multi-agent collaboration systems in nursing, highlighting their promise for future research and applications.
comment: EMNLP 2025 Industry Track
☆ ELMUR: External Layer Memory with Update/Rewrite for Long-Horizon RL
Real-world robotic agents must act under partial observability and long horizons, where key cues may appear long before they affect decision making. However, most modern approaches rely solely on instantaneous information, without incorporating insights from the past. Standard recurrent or transformer models struggle with retaining and leveraging long-term dependencies: context windows truncate history, while naive memory extensions fail under scale and sparsity. We propose ELMUR (External Layer Memory with Update/Rewrite), a transformer architecture with structured external memory. Each layer maintains memory embeddings, interacts with them via bidirectional cross-attention, and updates them through an Least Recently Used (LRU) memory module using replacement or convex blending. ELMUR extends effective horizons up to 100,000 times beyond the attention window and achieves a 100% success rate on a synthetic T-Maze task with corridors up to one million steps. In POPGym, it outperforms baselines on more than half of the tasks. On MIKASA-Robo sparse-reward manipulation tasks with visual observations, it nearly doubles the performance of strong baselines. These results demonstrate that structured, layer-local external memory offers a simple and scalable approach to decision making under partial observability.
comment: 22 pages, 7 figures
☆ A Multi-Agent Framework for Stateful Inference-Time Search
Recent work explores agentic inference-time techniques to perform structured, multi-step reasoning. However, stateless inference often struggles on multi-step tasks due to the absence of persistent state. Moreover, task-specific fine-tuning or instruction-tuning often achieve surface-level code generation but remain brittle on tasks requiring deeper reasoning and long-horizon dependencies. To address these limitations, we propose stateful multi-agent evolutionary search, a training-free framework that departs from prior stateless approaches by combining (i) persistent inference-time state, (ii) adversarial mutation, and (iii) evolutionary preservation. We demonstrate its effectiveness in automated unit test generation through the generation of edge cases. We generate robust edge cases using an evolutionary search process, where specialized agents sequentially propose, mutate, and score candidates. A controller maintains persistent state across generations, while evolutionary preservation ensures diversity and exploration across all possible cases. This yields a generalist agent capable of discovering robust, high-coverage edge cases across unseen codebases. Experiments show our stateful multi-agent inference framework achieves substantial gains in coverage over stateless single-step baselines, evaluated on prevalent unit-testing benchmarks such as HumanEval and TestGenEvalMini and using three diverse LLM families - Llama, Gemma, and GPT. These results indicate that combining persistent inference-time state with evolutionary search materially improves unit-test generation.
☆ Spectral Graph Clustering under Differential Privacy: Balancing Privacy, Accuracy, and Efficiency
We study the problem of spectral graph clustering under edge differential privacy (DP). Specifically, we develop three mechanisms: (i) graph perturbation via randomized edge flipping combined with adjacency matrix shuffling, which enforces edge privacy while preserving key spectral properties of the graph. Importantly, shuffling considerably amplifies the guarantees: whereas flipping edges with a fixed probability alone provides only a constant epsilon edge DP guarantee as the number of nodes grows, the shuffled mechanism achieves (epsilon, delta) edge DP with parameters that tend to zero as the number of nodes increase; (ii) private graph projection with additive Gaussian noise in a lower-dimensional space to reduce dimensionality and computational complexity; and (iii) a noisy power iteration method that distributes Gaussian noise across iterations to ensure edge DP while maintaining convergence. Our analysis provides rigorous privacy guarantees and a precise characterization of the misclassification error rate. Experiments on synthetic and real-world networks validate our theoretical analysis and illustrate the practical privacy-utility trade-offs.
☆ DPMM-CFL: Clustered Federated Learning via Dirichlet Process Mixture Model Nonparametric Clustering
Clustered Federated Learning (CFL) improves performance under non-IID client heterogeneity by clustering clients and training one model per cluster, thereby balancing between a global model and fully personalized models. However, most CFL methods require the number of clusters K to be fixed a priori, which is impractical when the latent structure is unknown. We propose DPMM-CFL, a CFL algorithm that places a Dirichlet Process (DP) prior over the distribution of cluster parameters. This enables nonparametric Bayesian inference to jointly infer both the number of clusters and client assignments, while optimizing per-cluster federated objectives. This results in a method where, at each round, federated updates and cluster inferences are coupled, as presented in this paper. The algorithm is validated on benchmark datasets under Dirichlet and class-split non-IID partitions.
comment: 5 pages, 2 figures
☆ TRIM: Token-wise Attention-Derived Saliency for Data-Efficient Instruction Tuning
Instruction tuning is essential for aligning large language models (LLMs) to downstream tasks and commonly relies on large, diverse corpora. However, small, high-quality subsets, known as coresets, can deliver comparable or superior results, though curating them remains challenging. Existing methods often rely on coarse, sample-level signals like gradients, an approach that is computationally expensive and overlooks fine-grained features. To address this, we introduce TRIM (Token Relevance via Interpretable Multi-layer Attention), a forward-only, token-centric framework. Instead of using gradients, TRIM operates by matching underlying representational patterns identified via attention-based "fingerprints" from a handful of target samples. Such an approach makes TRIM highly efficient and uniquely sensitive to the structural features that define a task. Coresets selected by our method consistently outperform state-of-the-art baselines by up to 9% on downstream tasks and even surpass the performance of full-data fine-tuning in some settings. By avoiding expensive backward passes, TRIM achieves this at a fraction of the computational cost. These findings establish TRIM as a scalable and efficient alternative for building high-quality instruction-tuning datasets.
☆ The Contingencies of Physical Embodiment Allow for Open-Endedness and Care
Physical vulnerability and mortality are often seen as obstacles to be avoided in the development of artificial agents, which struggle to adapt to open-ended environments and provide aligned care. Meanwhile, biological organisms survive, thrive, and care for each other in an open-ended physical world with relative ease and efficiency. Understanding the role of the conditions of life in this disparity can aid in developing more robust, adaptive, and caring artificial agents. Here we define two minimal conditions for physical embodiment inspired by the existentialist phenomenology of Martin Heidegger: being-in-the-world (the agent is a part of the environment) and being-towards-death (unless counteracted, the agent drifts toward terminal states due to the second law of thermodynamics). We propose that from these conditions we can obtain both a homeostatic drive - aimed at maintaining integrity and avoiding death by expending energy to learn and act - and an intrinsic drive to continue to do so in as many ways as possible. Drawing inspiration from Friedrich Nietzsche's existentialist concept of will-to-power, we examine how intrinsic drives to maximize control over future states, e.g., empowerment, allow agents to increase the probability that they will be able to meet their future homeostatic needs, thereby enhancing their capacity to maintain physical integrity. We formalize these concepts within a reinforcement learning framework, which enables us to examine how intrinsically driven embodied agents learning in open-ended multi-agent environments may cultivate the capacities for open-endedness and care.ov
comment: 15 pages, 1 figure
☆ GNN-enhanced Traffic Anomaly Detection for Next-Generation SDN-Enabled Consumer Electronics
Consumer electronics (CE) connected to the Internet of Things are susceptible to various attacks, including DDoS and web-based threats, which can compromise their functionality and facilitate remote hijacking. These vulnerabilities allow attackers to exploit CE for broader system attacks while enabling the propagation of malicious code across the CE network, resulting in device failures. Existing deep learning-based traffic anomaly detection systems exhibit high accuracy in traditional network environments but are often overly complex and reliant on static infrastructure, necessitating manual configuration and management. To address these limitations, we propose a scalable network model that integrates Software-defined Networking (SDN) and Compute First Networking (CFN) for next-generation CE networks. In this network model, we propose a Graph Neural Networks-based Network Anomaly Detection framework (GNN-NAD) that integrates SDN-based CE networks and enables the CFN architecture. GNN-NAD uniquely fuses a static, vulnerability-aware attack graph with dynamic traffic features, providing a holistic view of network security. The core of the framework is a GNN model (GSAGE) for graph representation learning, followed by a Random Forest (RF) classifier. This design (GSAGE+RF) demonstrates superior performance compared to existing feature selection methods. Experimental evaluations on CE environment reveal that GNN-NAD achieves superior metrics in accuracy, recall, precision, and F1 score, even with small sample sizes, exceeding the performance of current network anomaly detection methods. This work advances the security and efficiency of next-generation intelligent CE networks.
comment: This paper has been accepted for publication in IEEE Transactions on Consumer Electronics. 10 pages, 6 figures
☆ Active Control of Turbulent Airfoil Flows Using Adjoint-based Deep Learning
We train active neural-network flow controllers using a deep learning PDE augmentation method to optimize lift-to-drag ratios in turbulent airfoil flows at Reynolds number $5\times10^4$ and Mach number 0.4. Direct numerical simulation and large eddy simulation are employed to model compressible, unconfined flow over two- and three-dimensional semi-infinite NACA 0012 airfoils at angles of attack $\alpha = 5^\circ$, $10^\circ$, and $15^\circ$. Control actions, implemented through a blowing/suction jet at a fixed location and geometry on the upper surface, are adaptively determined by a neural network that maps local pressure measurements to optimal jet total pressure, enabling a sensor-informed control policy that responds spatially and temporally to unsteady flow conditions. The sensitivities of the flow to the neural network parameters are computed using the adjoint Navier-Stokes equations, which we construct using automatic differentiation applied to the flow solver. The trained flow controllers significantly improve the lift-to-drag ratios and reduce flow separation for both two- and three-dimensional airfoil flows, especially at $\alpha = 5^\circ$ and $10^\circ$. The 2D-trained models remain effective when applied out-of-sample to 3D flows, which demonstrates the robustness of the adjoint-trained control approach. The 3D-trained models capture the flow dynamics even more effectively, which leads to better energy efficiency and comparable performance for both adaptive (neural network) and offline (simplified, constant-pressure) controllers. These results underscore the effectiveness of this learning-based approach in improving aerodynamic performance.
☆ Diffusion-Augmented Reinforcement Learning for Robust Portfolio Optimization under Stress Scenarios
In the ever-changing and intricate landscape of financial markets, portfolio optimisation remains a formidable challenge for investors and asset managers. Conventional methods often struggle to capture the complex dynamics of market behaviour and align with diverse investor preferences. To address this, we propose an innovative framework, termed Diffusion-Augmented Reinforcement Learning (DARL), which synergistically integrates Denoising Diffusion Probabilistic Models (DDPMs) with Deep Reinforcement Learning (DRL) for portfolio management. By leveraging DDPMs to generate synthetic market crash scenarios conditioned on varying stress intensities, our approach significantly enhances the robustness of training data. Empirical evaluations demonstrate that DARL outperforms traditional baselines, delivering superior risk-adjusted returns and resilience against unforeseen crises, such as the 2025 Tariff Crisis. This work offers a robust and practical methodology to bolster stress resilience in DRL-driven financial applications.
☆ Non-Asymptotic Analysis of Efficiency in Conformalized Regression
Conformal prediction provides prediction sets with coverage guarantees. The informativeness of conformal prediction depends on its efficiency, typically quantified by the expected size of the prediction set. Prior work on the efficiency of conformalized regression commonly treats the miscoverage level $\alpha$ as a fixed constant. In this work, we establish non-asymptotic bounds on the deviation of the prediction set length from the oracle interval length for conformalized quantile and median regression trained via SGD, under mild assumptions on the data distribution. Our bounds of order $\mathcal{O}(1/\sqrt{n} + 1/(\alpha^2 n) + 1/\sqrt{m} + \exp(-\alpha^2 m))$ capture the joint dependence of efficiency on the proper training set size $n$, the calibration set size $m$, and the miscoverage level $\alpha$. The results identify phase transitions in convergence rates across different regimes of $\alpha$, offering guidance for allocating data to control excess prediction set length. Empirical results are consistent with our theoretical findings.
☆ Generative World Modelling for Humanoids: 1X World Model Challenge Technical Report
World models are a powerful paradigm in AI and robotics, enabling agents to reason about the future by predicting visual observations or compact latent states. The 1X World Model Challenge introduces an open-source benchmark of real-world humanoid interaction, with two complementary tracks: sampling, focused on forecasting future image frames, and compression, focused on predicting future discrete latent codes. For the sampling track, we adapt the video generation foundation model Wan-2.2 TI2V-5B to video-state-conditioned future frame prediction. We condition the video generation on robot states using AdaLN-Zero, and further post-train the model using LoRA. For the compression track, we train a Spatio-Temporal Transformer model from scratch. Our models achieve 23.0 dB PSNR in the sampling task and a Top-500 CE of 6.6386 in the compression task, securing 1st place in both challenges.
comment: 6 pages, 3 figures, 1X world model challenge technical report
☆ Explaining Models under Multivariate Bernoulli Distribution via Hoeffding Decomposition
Explaining the behavior of predictive models with random inputs can be achieved through sub-models decomposition, where such sub-models have easier interpretable features. Arising from the uncertainty quantification community, recent results have demonstrated the existence and uniqueness of a generalized Hoeffding decomposition for such predictive models when the stochastic input variables are correlated, based on concepts of oblique projection onto L 2 subspaces. This article focuses on the case where the input variables have Bernoulli distributions and provides a complete description of this decomposition. We show that in this case the underlying L 2 subspaces are one-dimensional and that the functional decomposition is explicit. This leads to a complete interpretability framework and theoretically allows reverse engineering. Explicit indicators of the influence of inputs on the output prediction (exemplified by Sobol' indices and Shapley effects) can be explicitly derived. Illustrated by numerical experiments, this type of analysis proves useful for addressing decision-support problems, based on binary decision diagrams, Boolean networks or binary neural networks. The article outlines perspectives for exploring high-dimensional settings and, beyond the case of binary inputs, extending these findings to models with finite countable inputs.
☆ Non-Stationary Online Structured Prediction with Surrogate Losses
Online structured prediction, including online classification as a special case, is the task of sequentially predicting labels from input features. Therein the surrogate regret -- the cumulative excess of the target loss (e.g., 0-1 loss) over the surrogate loss (e.g., logistic loss) of the fixed best estimator -- has gained attention, particularly because it often admits a finite bound independent of the time horizon $T$. However, such guarantees break down in non-stationary environments, where every fixed estimator may incur the surrogate loss growing linearly with $T$. We address this by proving a bound of the form $F_T + C(1 + P_T)$ on the cumulative target loss, where $F_T$ is the cumulative surrogate loss of any comparator sequence, $P_T$ is its path length, and $C > 0$ is some constant. This bound depends on $T$ only through $F_T$ and $P_T$, often yielding much stronger guarantees in non-stationary environments. Our core idea is to synthesize the dynamic regret bound of the online gradient descent (OGD) with the technique of exploiting the surrogate gap. Our analysis also sheds light on a new Polyak-style learning rate for OGD, which systematically offers target-loss guarantees and exhibits promising empirical performance. We further extend our approach to a broader class of problems via the convolutional Fenchel--Young loss. Finally, we prove a lower bound showing that the dependence on $F_T$ and $P_T$ is tight.
☆ HTMformer: Hybrid Time and Multivariate Transformer for Time Series Forecasting
Transformer-based methods have achieved impressive results in time series forecasting. However, existing Transformers still exhibit limitations in sequence modeling as they tend to overemphasize temporal dependencies. This incurs additional computational overhead without yielding corresponding performance gains. We find that the performance of Transformers is highly dependent on the embedding method used to learn effective representations. To address this issue, we extract multivariate features to augment the effective information captured in the embedding layer, yielding multidimensional embeddings that convey richer and more meaningful sequence representations. These representations enable Transformer-based forecasters to better understand the series. Specifically, we introduce Hybrid Temporal and Multivariate Embeddings (HTME). The HTME extractor integrates a lightweight temporal feature extraction module with a carefully designed multivariate feature extraction module to provide complementary features, thereby achieving a balance between model complexity and performance. By combining HTME with the Transformer architecture, we present HTMformer, leveraging the enhanced feature extraction capability of the HTME extractor to build a lightweight forecaster. Experiments conducted on eight real-world datasets demonstrate that our approach outperforms existing baselines in both accuracy and efficiency.
☆ Pseudo-MDPs: A Novel Framework for Efficiently Optimizing Last Revealer Seed Manipulations in Blockchains
This study tackles the computational challenges of solving Markov Decision Processes (MDPs) for a restricted class of problems. It is motivated by the Last Revealer Attack (LRA), which undermines fairness in some Proof-of-Stake (PoS) blockchains such as Ethereum (\$400B market capitalization). We introduce pseudo-MDPs (pMDPs) a framework that naturally models such problems and propose two distinct problem reductions to standard MDPs. One problem reduction provides a novel, counter-intuitive perspective, and combining the two problem reductions enables significant improvements in dynamic programming algorithms such as value iteration. In the case of the LRA which size is parameterized by $\kappa$ (in Ethereum's case $\kappa$= 325), we reduce the computational complexity from $O(2^\kappa \kappa^{2^{\kappa+2}})$ to $O(\kappa^4)$ (per iteration). This solution also provide the usual benefits from Dynamic Programming solutions: exponentially fast convergence toward the optimal solution is guaranteed. The dual perspective also simplifies policy extraction, making the approach well-suited for resource-constrained agents who can operate with very limited memory and computation once the problem has been solved. Furthermore, we generalize those results to a broader class of MDPs, enhancing their applicability. The framework is validated through two case studies: a fictional card game and the LRA on the Ethereum random seed consensus protocol. These applications demonstrate the framework's ability to solve large-scale problems effectively while offering actionable insights into optimal strategies. This work advances the study of MDPs and contributes to understanding security vulnerabilities in blockchain systems.
☆ Vision-Language-Action Models for Robotics: A Review Towards Real-World Applications
Amid growing efforts to leverage advances in large language models (LLMs) and vision-language models (VLMs) for robotics, Vision-Language-Action (VLA) models have recently gained significant attention. By unifying vision, language, and action data at scale, which have traditionally been studied separately, VLA models aim to learn policies that generalise across diverse tasks, objects, embodiments, and environments. This generalisation capability is expected to enable robots to solve novel downstream tasks with minimal or no additional task-specific data, facilitating more flexible and scalable real-world deployment. Unlike previous surveys that focus narrowly on action representations or high-level model architectures, this work offers a comprehensive, full-stack review, integrating both software and hardware components of VLA systems. In particular, this paper provides a systematic review of VLAs, covering their strategy and architectural transition, architectures and building blocks, modality-specific processing techniques, and learning paradigms. In addition, to support the deployment of VLAs in real-world robotic applications, we also review commonly used robot platforms, data collection strategies, publicly available datasets, data augmentation methods, and evaluation benchmarks. Throughout this comprehensive survey, this paper aims to offer practical guidance for the robotics community in applying VLAs to real-world robotic systems. All references categorized by training approach, evaluation method, modality, and dataset are available in the table on our project website: https://vla-survey.github.io .
comment: Accepted to IEEE Access, website: https://vla-survey.github.io
☆ Blind Construction of Angular Power Maps in Massive MIMO Networks
Channel state information (CSI) acquisition is a challenging problem in massive multiple-input multiple-output (MIMO) networks. Radio maps provide a promising solution for radio resource management by reducing online CSI acquisition. However, conventional approaches for radio map construction require location-labeled CSI data, which is challenging in practice. This paper investigates unsupervised angular power map construction based on large timescale CSI data collected in a massive MIMO network without location labels. A hidden Markov model (HMM) is built to connect the hidden trajectory of a mobile with the CSI evolution of a massive MIMO channel. As a result, the mobile location can be estimated, enabling the construction of an angular power map. We show that under uniform rectilinear mobility with Poisson-distributed base stations (BSs), the Cramer-Rao Lower Bound (CRLB) for localization error can vanish at any signal-to-noise ratios (SNRs), whereas when BSs are confined to a limited region, the error remains nonzero even with infinite independent measurements. Based on reference signal received power (RSRP) data collected in a real multi-cell massive MIMO network, an average localization error of 18 meters can be achieved although measurements are mainly obtained from a single serving cell.
☆ Introspection in Learned Semantic Scene Graph Localisation IROS 2025
This work investigates how semantics influence localisation performance and robustness in a learned self-supervised, contrastive semantic localisation framework. After training a localisation network on both original and perturbed maps, we conduct a thorough post-hoc introspection analysis to probe whether the model filters environmental noise and prioritises distinctive landmarks over routine clutter. We validate various interpretability methods and present a comparative reliability analysis. Integrated gradients and Attention Weights consistently emerge as the most reliable probes of learned behaviour. A semantic class ablation further reveals an implicit weighting in which frequent objects are often down-weighted. Overall, the results indicate that the model learns noise-robust, semantically salient relations about place definition, thereby enabling explainable registration under challenging visual and structural variations.
comment: IEEE IROS 2025 Workshop FAST
☆ Enhancing Speech Emotion Recognition via Fine-Tuning Pre-Trained Models and Hyper-Parameter Optimisation
We propose a workflow for speech emotion recognition (SER) that combines pre-trained representations with automated hyperparameter optimisation (HPO). Using SpeechBrain wav2vec2-base model fine-tuned on IEMOCAP as the encoder, we compare two HPO strategies, Gaussian Process Bayesian Optimisation (GP-BO) and Tree-structured Parzen Estimators (TPE), under an identical four-dimensional search space and 15-trial budget, with balanced class accuracy (BCA) on the German EmoDB corpus as the objective. All experiments run on 8 CPU cores with 32 GB RAM. GP-BO achieves 0.96 BCA in 11 minutes, and TPE (Hyperopt implementation) attains 0.97 in 15 minutes. In contrast, grid search requires 143 trials and 1,680 minutes to exceed 0.9 BCA, and the best AutoSpeech 2020 baseline reports only 0.85 in 30 minutes on GPU. For cross-lingual generalisation, an EmoDB-trained HPO-tuned model improves zero-shot accuracy by 0.25 on CREMA-D and 0.26 on RAVDESS. Results show that efficient HPO with pre-trained encoders delivers competitive SER on commodity CPUs. Source code to this work is available at: https://github.com/youngaryan/speechbrain-emotion-hpo.
☆ COMPASS: A Multi-Turn Benchmark for Tool-Mediated Planning & Preference Optimization
Real-world large language model (LLM) agents must master strategic tool use and user preference optimization through multi-turn interactions to assist users with complex planning tasks. We introduce COMPASS (Constrained Optimization through Multi-turn Planning and Strategic Solutions), a benchmark that evaluates agents on realistic travel-planning scenarios. We cast travel planning as a constrained preference optimization problem, where agents must satisfy hard constraints while simultaneously optimizing soft user preferences. To support this, we build a realistic travel database covering transportation, accommodation, and ticketing for 20 U.S. National Parks, along with a comprehensive tool ecosystem that mirrors commercial booking platforms. Evaluating state-of-the-art models, we uncover two critical gaps: (i) an acceptable-optimal gap, where agents reliably meet constraints but fail to optimize preferences, and (ii) a plan-coordination gap, where performance collapses on multi-service (flight and hotel) coordination tasks, especially for open-source models. By grounding reasoning and planning in a practical, user-facing domain, COMPASS provides a benchmark that directly measures an agent's ability to optimize user preferences in realistic tasks, bridging theoretical advances with real-world impact.
☆ Unified Molecule Pre-training with Flexible 2D and 3D Modalities: Single and Paired Modality Integration CIKM 2025
Molecular representation learning plays a crucial role in advancing applications such as drug discovery and material design. Existing work leverages 2D and 3D modalities of molecular information for pre-training, aiming to capture comprehensive structural and geometric insights. However, these methods require paired 2D and 3D molecular data to train the model effectively and prevent it from collapsing into a single modality, posing limitations in scenarios where a certain modality is unavailable or computationally expensive to generate. To overcome this limitation, we propose FlexMol, a flexible molecule pre-training framework that learns unified molecular representations while supporting single-modality input. Specifically, inspired by the unified structure in vision-language models, our approach employs separate models for 2D and 3D molecular data, leverages parameter sharing to improve computational efficiency, and utilizes a decoder to generate features for the missing modality. This enables a multistage continuous learning process where both modalities contribute collaboratively during training, while ensuring robustness when only one modality is available during inference. Extensive experiments demonstrate that FlexMol achieves superior performance across a wide range of molecular property prediction tasks, and we also empirically demonstrate its effectiveness with incomplete data. Our code and data are available at https://github.com/tewiSong/FlexMol.
comment: CIKM 2025
☆ Federated Unlearning in the Wild: Rethinking Fairness and Data Discrepancy
Machine unlearning is critical for enforcing data deletion rights like the "right to be forgotten." As a decentralized paradigm, Federated Learning (FL) also requires unlearning, but realistic implementations face two major challenges. First, fairness in Federated Unlearning (FU) is often overlooked. Exact unlearning methods typically force all clients into costly retraining, even those uninvolved. Approximate approaches, using gradient ascent or distillation, make coarse interventions that can unfairly degrade performance for clients with only retained data. Second, most FU evaluations rely on synthetic data assumptions (IID/non-IID) that ignore real-world heterogeneity. These unrealistic benchmarks obscure the true impact of unlearning and limit the applicability of current methods. We first conduct a comprehensive benchmark of existing FU methods under realistic data heterogeneity and fairness conditions. We then propose a novel, fairness-aware FU approach, Federated Cross-Client-Constrains Unlearning (FedCCCU), to explicitly address both challenges. FedCCCU offers a practical and scalable solution for real-world FU. Experimental results show that existing methods perform poorly in realistic settings, while our approach consistently outperforms them.
☆ Native Hybrid Attention for Efficient Sequence Modeling
Transformers excel at sequence modeling but face quadratic complexity, while linear attention offers improved efficiency but often compromises recall accuracy over long contexts. In this work, we introduce Native Hybrid Attention (NHA), a novel hybrid architecture of linear and full attention that integrates both intra \& inter-layer hybridization into a unified layer design. NHA maintains long-term context in key-value slots updated by a linear RNN, and augments them with short-term tokens from a sliding window. A single \texttt{softmax attention} operation is then applied over all keys and values, enabling per-token and per-head context-dependent weighting without requiring additional fusion parameters. The inter-layer behavior is controlled through a single hyperparameter, the sliding window size, which allows smooth adjustment between purely linear and full attention while keeping all layers structurally uniform. Experimental results show that NHA surpasses Transformers and other hybrid baselines on recall-intensive and commonsense reasoning tasks. Furthermore, pretrained LLMs can be structurally hybridized with NHA, achieving competitive accuracy while delivering significant efficiency gains. Code is available at https://github.com/JusenD/NHA.
comment: Technical report, 16 pages
☆ Sharpness-Aware Data Generation for Zero-shot Quantization
Zero-shot quantization aims to learn a quantized model from a pre-trained full-precision model with no access to original real training data. The common idea in zero-shot quantization approaches is to generate synthetic data for quantizing the full-precision model. While it is well-known that deep neural networks with low sharpness have better generalization ability, none of the previous zero-shot quantization works considers the sharpness of the quantized model as a criterion for generating training data. This paper introduces a novel methodology that takes into account quantized model sharpness in synthetic data generation to enhance generalization. Specifically, we first demonstrate that sharpness minimization can be attained by maximizing gradient matching between the reconstruction loss gradients computed on synthetic and real validation data, under certain assumptions. We then circumvent the problem of the gradient matching without real validation set by approximating it with the gradient matching between each generated sample and its neighbors. Experimental evaluations on CIFAR-100 and ImageNet datasets demonstrate the superiority of the proposed method over the state-of-the-art techniques in low-bit quantization settings.
☆ Root Cause Analysis of Outliers in Unknown Cyclic Graphs
We study the propagation of outliers in cyclic causal graphs with linear structural equations, tracing them back to one or several "root cause" nodes. We show that it is possible to identify a short list of potential root causes provided that the perturbation is sufficiently strong and propagates according to the same structural equations as in the normal mode. This shortlist consists of the true root causes together with those of its parents lying on a cycle with the root cause. Notably, our method does not require prior knowledge of the causal graph.
☆ Spiral Model Technique For Data Science & Machine Learning Lifecycle
Analytics play an important role in modern business. Companies adapt data science lifecycles to their culture to seek productivity and improve their competitiveness among others. Data science lifecycles are fairly an important contributing factor to start and end a project that are data dependent. Data science and Machine learning life cycles comprises of series of steps that are involved in a project. A typical life cycle states that it is a linear or cyclical model that revolves around. It is mostly depicted that it is possible in a traditional data science life cycle to start the process again after reaching the end of cycle. This paper suggests a new technique to incorporate data science life cycle to business problems that have a clear end goal. A new technique called spiral technique is introduced to emphasize versatility, agility and iterative approach to business processes.
☆ Revisiting Mixout: An Overlooked Path to Robust Finetuning
Finetuning vision foundation models often improves in-domain accuracy but comes at the cost of robustness under distribution shift. We revisit Mixout, a stochastic regularizer that intermittently replaces finetuned weights with their pretrained reference, through the lens of a single-run, weight-sharing implicit ensemble. This perspective reveals three key levers that govern robustness: the \emph{masking anchor}, \emph{resampling frequency}, and \emph{mask sparsity}. Guided by this analysis, we introduce GMixout, which (i) replaces the fixed anchor with an exponential moving-average snapshot that adapts during training, and (ii) regulates masking period via an explicit resampling-frequency hyperparameter. Our sparse-kernel implementation updates only a small fraction of parameters with no inference-time overhead, enabling training on consumer-grade GPUs. Experiments on benchmarks covering covariate shift, corruption, and class imbalance, ImageNet / ImageNet-LT, DomainNet, iWildCam, and CIFAR100-C, GMixout consistently improves in-domain accuracy beyond zero-shot performance while surpassing both Model Soups and strong parameter-efficient finetuning baselines under distribution shift.
☆ Relational Database Distillation: From Structured Tables to Condensed Graph Data
Relational databases (RDBs) underpin the majority of global data management systems, where information is structured into multiple interdependent tables. To effectively use the knowledge within RDBs for predictive tasks, recent advances leverage graph representation learning to capture complex inter-table relations as multi-hop dependencies. Despite achieving state-of-the-art performance, these methods remain hindered by the prohibitive storage overhead and excessive training time, due to the massive scale of the database and the computational burden of intensive message passing across interconnected tables. To alleviate these concerns, we propose and study the problem of Relational Database Distillation (RDD). Specifically, we aim to distill large-scale RDBs into compact heterogeneous graphs while retaining the predictive power (i.e., utility) required for training graph-based models. Multi-modal column information is preserved through node features, and primary-foreign key relations are encoded via heterogeneous edges, thereby maintaining both data fidelity and relational structure. To ensure adaptability across diverse downstream tasks without engaging the traditional, inefficient bi-level distillation framework, we further design a kernel ridge regression-guided objective with pseudo-labels, which produces quality features for the distilled graph. Extensive experiments on multiple real-world RDBs demonstrate that our solution substantially reduces the data size while maintaining competitive performance on classification and regression tasks, creating an effective pathway for scalable learning with RDBs.
☆ Falsification-Driven Reinforcement Learning for Maritime Motion Planning
Compliance with maritime traffic rules is essential for the safe operation of autonomous vessels, yet training reinforcement learning (RL) agents to adhere to them is challenging. The behavior of RL agents is shaped by the training scenarios they encounter, but creating scenarios that capture the complexity of maritime navigation is non-trivial, and real-world data alone is insufficient. To address this, we propose a falsification-driven RL approach that generates adversarial training scenarios in which the vessel under test violates maritime traffic rules, which are expressed as signal temporal logic specifications. Our experiments on open-sea navigation with two vessels demonstrate that the proposed approach provides more relevant training scenarios and achieves more consistent rule compliance.
☆ Accelerating Sparse Ternary GEMM for Quantized LLM inference on Apple Silicon
Sparse Ternary General Matrix-Matrix Multiplication (GEMM) remains under-optimized in existing libraries for Apple Silicon CPUs. We present a Sparse Ternary GEMM kernel optimized specifically for Apple's M-series processors. We propose a set of architecture-aware optimizations, including a novel blocked and interleaved sparse data format to improve memory locality, strategies to increase Instruction-Level Parallelism (ILP), and NEON-based Single Instruction Multiple Data (SIMD) vectorization to exploit data-level parallelism. Our scalar implementation achieves up to a 5.98x performance increase over a traditional Ternary Compressed Sparse Column (TCSC) baseline for large matrices with 50% ternary nonzero values (sparsity), reaching up to a 50.2% of the processor's theoretical peak performance, and remains stable across varying sparsity levels. Our vectorized implementation delivers up to a 5.59x performance increase for large matrices with 25% sparsity, and remains stable across varying sparsity levels.
☆ High-Rate Mixout: Revisiting Mixout for Robust Domain Generalization WACV 2026
Ensembling fine-tuned models initialized from powerful pre-trained weights is a common strategy to improve robustness under distribution shifts, but it comes with substantial computational costs due to the need to train and store multiple models. Dropout offers a lightweight alternative by simulating ensembles through random neuron deactivation; however, when applied to pre-trained models, it tends to over-regularize and disrupt critical representations necessary for generalization. In this work, we investigate Mixout, a stochastic regularization technique that provides an alternative to Dropout for domain generalization. Rather than deactivating neurons, Mixout mitigates overfitting by probabilistically swapping a subset of fine-tuned weights with their pre-trained counterparts during training, thereby maintaining a balance between adaptation and retention of prior knowledge. Our study reveals that achieving strong performance with Mixout on domain generalization benchmarks requires a notably high masking probability of 0.9 for ViTs and 0.8 for ResNets. While this may seem like a simple adjustment, it yields two key advantages for domain generalization: (1) higher masking rates more strongly penalize deviations from the pre-trained parameters, promoting better generalization to unseen domains; and (2) high-rate masking substantially reduces computational overhead, cutting gradient computation by up to 45% and gradient memory usage by up to 90%. Experiments across five domain generalization benchmarks, PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet, using ResNet and ViT architectures, show that our approach, High-rate Mixout, achieves out-of-domain accuracy comparable to ensemble-based methods while significantly reducing training costs.
comment: WACV 2026: Winter Conference on Applications of Computer Vision 2026
☆ From Condensation to Rank Collapse: A Two-Stage Analysis of Transformer Training Dynamics
Although transformer-based models have shown exceptional empirical performance, the fundamental principles governing their training dynamics are inadequately characterized beyond configuration-specific studies. Inspired by empirical evidence showing improved reasoning capabilities under small initialization scales in language models, we employ the gradient flow analytical framework established in [Zhou et al. NeurIPS 2022] to systematically investigate linearized Transformer training dynamics. Our theoretical analysis dissects the dynamics of attention modules into two distinct stages. In the first stage, asymmetric weight perturbations from random initialization sustain non-degenerate gradient dynamics in parameter matrices, facilitating systematic escape from small initialization regimes. Subsequently, these matrices undergo condensation, progressively aligning toward the target orientation. In the second stage, the previously static key-query matrices actively participate in training, driving the normalized matrices toward asymptotic rank collapse. This two-stage framework generalizes classical directional convergence results.
☆ Grouped Differential Attention
The self-attention mechanism, while foundational to modern Transformer architectures, suffers from a critical inefficiency: it frequently allocates substantial attention to redundant or noisy context. Differential Attention addressed this by using subtractive attention maps for signal and noise, but its required balanced head allocation imposes rigid constraints on representational flexibility and scalability. To overcome this, we propose Grouped Differential Attention (GDA), a novel approach that introduces unbalanced head allocation between signal-preserving and noise-control groups. GDA significantly enhances signal focus by strategically assigning more heads to signal extraction and fewer to noise-control, stabilizing the latter through controlled repetition (akin to GQA). This design achieves stronger signal fidelity with minimal computational overhead. We further extend this principle to group-differentiated growth, a scalable strategy that selectively replicates only the signal-focused heads, thereby ensuring efficient capacity expansion. Through large-scale pretraining and continual training experiments, we demonstrate that moderate imbalance ratios in GDA yield substantial improvements in generalization and stability compared to symmetric baselines. Our results collectively establish that ratio-aware head allocation and selective expansion offer an effective and practical path toward designing scalable, computation-efficient Transformer architectures.
☆ Fisher Information, Training and Bias in Fourier Regression Models
Motivated by the growing interest in quantum machine learning, in particular quantum neural networks (QNNs), we study how recently introduced evaluation metrics based on the Fisher information matrix (FIM) are effective for predicting their training and prediction performance. We exploit the equivalence between a broad class of QNNs and Fourier models, and study the interplay between the \emph{effective dimension} and the \emph{bias} of a model towards a given task, investigating how these affect the model's training and performance. We show that for a model that is completely agnostic, or unbiased, towards the function to be learned, a higher effective dimension likely results in a better trainability and performance. On the other hand, for models that are biased towards the function to be learned a lower effective dimension is likely beneficial during training. To obtain these results, we derive an analytical expression of the FIM for Fourier models and identify the features controlling a model's effective dimension. This allows us to construct models with tunable effective dimension and bias, and to compare their training. We furthermore introduce a tensor network representation of the considered Fourier models, which could be a tool of independent interest for the analysis of QNN models. Overall, these findings provide an explicit example of the interplay between geometrical properties, model-task alignment and training, which are relevant for the broader machine learning community.
☆ Revisiting Node Affinity Prediction in Temporal Graphs
Node affinity prediction is a common task that is widely used in temporal graph learning with applications in social and financial networks, recommender systems, and more. Recent works have addressed this task by adapting state-of-the-art dynamic link property prediction models to node affinity prediction. However, simple heuristics, such as Persistent Forecast or Moving Average, outperform these models. In this work, we analyze the challenges in training current Temporal Graph Neural Networks for node affinity prediction and suggest appropriate solutions. Combining the solutions, we develop NAViS - Node Affinity prediction model using Virtual State, by exploiting the equivalence between heuristics and state space models. While promising, training NAViS is non-trivial. Therefore, we further introduce a novel loss function for node affinity prediction. We evaluate NAViS on TGB and show that it outperforms the state-of-the-art, including heuristics. Our source code is available at https://github.com/orfeld415/NAVIS
comment: preprint
☆ PyCFRL: A Python library for counterfactually fair offline reinforcement learning via sequential data preprocessing
Reinforcement learning (RL) aims to learn and evaluate a sequential decision rule, often referred to as a "policy", that maximizes the population-level benefit in an environment across possibly infinitely many time steps. However, the sequential decisions made by an RL algorithm, while optimized to maximize overall population benefits, may disadvantage certain individuals who are in minority or socioeconomically disadvantaged groups. To address this problem, we introduce PyCFRL, a Python library for ensuring counterfactual fairness in offline RL. PyCFRL implements a novel data preprocessing algorithm for learning counterfactually fair RL policies from offline datasets and provides tools to evaluate the values and counterfactual unfairness levels of RL policies. We describe the high-level functionalities of PyCFRL and demonstrate one of its major use cases through a data example. The library is publicly available on PyPI and Github (https://github.com/JianhanZhang/PyCFRL), and detailed tutorials can be found in the PyCFRL documentation (https://pycfrl-documentation.netlify.app).
☆ Textual interpretation of transient image classifications from large language models
Modern astronomical surveys deliver immense volumes of transient detections, yet distinguishing real astrophysical signals (for example, explosive events) from bogus imaging artefacts remains a challenge. Convolutional neural networks are effectively used for real versus bogus classification; however, their reliance on opaque latent representations hinders interpretability. Here we show that large language models (LLMs) can approach the performance level of a convolutional neural network on three optical transient survey datasets (Pan-STARRS, MeerLICHT and ATLAS) while simultaneously producing direct, human-readable descriptions for every candidate. Using only 15 examples and concise instructions, Google's LLM, Gemini, achieves a 93% average accuracy across datasets that span a range of resolution and pixel scales. We also show that a second LLM can assess the coherence of the output of the first model, enabling iterative refinement by identifying problematic cases. This framework allows users to define the desired classification behaviour through natural language and examples, bypassing traditional training pipelines. Furthermore, by generating textual descriptions of observed features, LLMs enable users to query classifications as if navigating an annotated catalogue, rather than deciphering abstract latent spaces. As next-generation telescopes and surveys further increase the amount of data available, LLM-based classification could help bridge the gap between automated detection and transparent, human-level understanding.
comment: Published in Nature Astronomy (2025). Publisher's Version of Record (CC BY 4.0). DOI: 10.1038/s41550-025-02670-z
☆ Quantum Sparse Recovery and Quantum Orthogonal Matching Pursuit
We study quantum sparse recovery in non-orthogonal, overcomplete dictionaries: given coherent quantum access to a state and a dictionary of vectors, the goal is to reconstruct the state up to $\ell_2$ error using as few vectors as possible. We first show that the general recovery problem is NP-hard, ruling out efficient exact algorithms in full generality. To overcome this, we introduce Quantum Orthogonal Matching Pursuit (QOMP), the first quantum analogue of the classical OMP greedy algorithm. QOMP combines quantum subroutines for inner product estimation, maximum finding, and block-encoded projections with an error-resetting design that avoids iteration-to-iteration error accumulation. Under standard mutual incoherence and well-conditioned sparsity assumptions, QOMP provably recovers the exact support of a $K$-sparse state in polynomial time. As an application, we give the first framework for sparse quantum tomography with non-orthogonal dictionaries in $\ell_2$ norm, achieving query complexity $\widetilde{O}(\sqrt{N}/\epsilon)$ in favorable regimes and reducing tomography to estimating only $K$ coefficients instead of $N$ amplitudes. In particular, for pure-state tomography with $m=O(N)$ dictionary vectors and sparsity $K=\widetilde{O}(1)$ on a well-conditioned subdictionary, this circumvents the $\widetilde{\Omega}(N/\epsilon)$ lower bound that holds in the dense, orthonormal-dictionary setting, without contradiction, by leveraging sparsity together with non-orthogonality. Beyond tomography, we analyze QOMP in the QRAM model, where it yields polynomial speedups over classical OMP implementations, and provide a quantum algorithm to estimate the mutual incoherence of a dictionary of $m$ vectors in $O(m/\epsilon)$ queries, improving over both deterministic and quantum-inspired classical methods.
☆ Bayesian Nonparametric Dynamical Clustering of Time Series
We present a method that models the evolution of an unbounded number of time series clusters by switching among an unknown number of regimes with linear dynamics. We develop a Bayesian non-parametric approach using a hierarchical Dirichlet process as a prior on the parameters of a Switching Linear Dynamical System and a Gaussian process prior to model the statistical variations in amplitude and temporal alignment within each cluster. By modeling the evolution of time series patterns, the method avoids unnecessary proliferation of clusters in a principled manner. We perform inference by formulating a variational lower bound for off-line and on-line scenarios, enabling efficient learning through optimization. We illustrate the versatility and effectiveness of the approach through several case studies of electrocardiogram analysis using publicly available databases.
comment: This work has been submitted to the IEEE for possible publication. 15 pages. 9 figures
☆ DecompGAIL: Learning Realistic Traffic Behaviors with Decomposed Multi-Agent Generative Adversarial Imitation Learning
Realistic traffic simulation is critical for the development of autonomous driving systems and urban mobility planning, yet existing imitation learning approaches often fail to model realistic traffic behaviors. Behavior cloning suffers from covariate shift, while Generative Adversarial Imitation Learning (GAIL) is notoriously unstable in multi-agent settings. We identify a key source of this instability: irrelevant interaction misguidance, where a discriminator penalizes an ego vehicle's realistic behavior due to unrealistic interactions among its neighbors. To address this, we propose Decomposed Multi-agent GAIL (DecompGAIL), which explicitly decomposes realism into ego-map and ego-neighbor components, filtering out misleading neighbor: neighbor and neighbor: map interactions. We further introduce a social PPO objective that augments ego rewards with distance-weighted neighborhood rewards, encouraging overall realism across agents. Integrated into a lightweight SMART-based backbone, DecompGAIL achieves state-of-the-art performance on the WOMD Sim Agents 2025 benchmark.
☆ Utilizing Large Language Models for Machine Learning Explainability
This study explores the explainability capabilities of large language models (LLMs), when employed to autonomously generate machine learning (ML) solutions. We examine two classification tasks: (i) a binary classification problem focused on predicting driver alertness states, and (ii) a multilabel classification problem based on the yeast dataset. Three state-of-the-art LLMs (i.e. OpenAI GPT, Anthropic Claude, and DeepSeek) are prompted to design training pipelines for four common classifiers: Random Forest, XGBoost, Multilayer Perceptron, and Long Short-Term Memory networks. The generated models are evaluated in terms of predictive performance (recall, precision, and F1-score) and explainability using SHAP (SHapley Additive exPlanations). Specifically, we measure Average SHAP Fidelity (Mean Squared Error between SHAP approximations and model outputs) and Average SHAP Sparsity (number of features deemed influential). The results reveal that LLMs are capable of producing effective and interpretable models, achieving high fidelity and consistent sparsity, highlighting their potential as automated tools for interpretable ML pipeline generation. The results show that LLMs can produce effective, interpretable pipelines with high fidelity and consistent sparsity, closely matching manually engineered baselines.
☆ Vacuum Spiker: A Spiking Neural Network-Based Model for Efficient Anomaly Detection in Time Series
Anomaly detection is a key task across domains such as industry, healthcare, and cybersecurity. Many real-world anomaly detection problems involve analyzing multiple features over time, making time series analysis a natural approach for such problems. While deep learning models have achieved strong performance in this field, their trend to exhibit high energy consumption limits their deployment in resource-constrained environments such as IoT devices, edge computing platforms, and wearables. To address this challenge, this paper introduces the \textit{Vacuum Spiker algorithm}, a novel Spiking Neural Network-based method for anomaly detection in time series. It incorporates a new detection criterion that relies on global changes in neural activity rather than reconstruction or prediction error. It is trained using Spike Time-Dependent Plasticity in a novel way, intended to induce changes in neural activity when anomalies occur. A new efficient encoding scheme is also proposed, which discretizes the input space into non-overlapping intervals, assigning each to a single neuron. This strategy encodes information with a single spike per time step, improving energy efficiency compared to conventional encoding methods. Experimental results on publicly available datasets show that the proposed algorithm achieves competitive performance while significantly reducing energy consumption, compared to a wide set of deep learning and machine learning baselines. Furthermore, its practical utility is validated in a real-world case study, where the model successfully identifies power curtailment events in a solar inverter. These results highlight its potential for sustainable and efficient anomaly detection.
comment: 53 pages, 16 figures, preprint submitted to a journal for review
☆ Angular Constraint Embedding via SpherePair Loss for Constrained Clustering NeurIPS 2025
Constrained clustering integrates domain knowledge through pairwise constraints. However, existing deep constrained clustering (DCC) methods are either limited by anchors inherent in end-to-end modeling or struggle with learning discriminative Euclidean embedding, restricting their scalability and real-world applicability. To avoid their respective pitfalls, we propose a novel angular constraint embedding approach for DCC, termed SpherePair. Using the SpherePair loss with a geometric formulation, our method faithfully encodes pairwise constraints and leads to embeddings that are clustering-friendly in angular space, effectively separating representation learning from clustering. SpherePair preserves pairwise relations without conflict, removes the need to specify the exact number of clusters, generalizes to unseen data, enables rapid inference of the number of clusters, and is supported by rigorous theoretical guarantees. Comparative evaluations with state-of-the-art DCC methods on diverse benchmarks, along with empirical validation of theoretical insights, confirm its superior performance, scalability, and overall real-world effectiveness. Code is available at \href{https://github.com/spherepaircc/SpherePairCC/tree/main}{our repository}.
comment: Accepted by NeurIPS 2025, 6 Figures and 1 Table in Main text, 18 Figures and 5 Tables in Appendices
☆ Multi-Dimensional Autoscaling of Stream Processing Services on Edge Devices
Edge devices have limited resources, which inevitably leads to situations where stream processing services cannot satisfy their needs. While existing autoscaling mechanisms focus entirely on resource scaling, Edge devices require alternative ways to sustain the Service Level Objectives (SLOs) of competing services. To address these issues, we introduce a Multi-dimensional Autoscaling Platform (MUDAP) that supports fine-grained vertical scaling across both service- and resource-level dimensions. MUDAP supports service-specific scaling tailored to available parameters, e.g., scale data quality or model size for a particular service. To optimize the execution across services, we present a scaling agent based on Regression Analysis of Structural Knowledge (RASK). The RASK agent efficiently explores the solution space and learns a continuous regression model of the processing environment for inferring optimal scaling actions. We compared our approach with two autoscalers, the Kubernetes VPA and a reinforcement learning agent, for scaling up to 9 services on a single Edge device. Our results showed that RASK can infer an accurate regression model in merely 20 iterations (i.e., observe 200s of processing). By increasingly adding elasticity dimensions, RASK sustained the highest request load with 28% less SLO violations, compared to baselines.
☆ MoRE-GNN: Multi-omics Data Integration with a Heterogeneous Graph Autoencoder
The integration of multi-omics single-cell data remains challenging due to high-dimensionality and complex inter-modality relationships. To address this, we introduce MoRE-GNN (Multi-omics Relational Edge Graph Neural Network), a heterogeneous graph autoencoder that combines graph convolution and attention mechanisms to dynamically construct relational graphs directly from data. Evaluations on six publicly available datasets demonstrate that MoRE-GNN captures biologically meaningful relationships and outperforms existing methods, particularly in settings with strong inter-modality correlations. Furthermore, the learned representations allow for accurate downstream cross-modal predictions. While performance may vary with dataset complexity, MoRE-GNN offers an adaptive, scalable and interpretable framework for advancing multi-omics integration.
☆ SaFeR-VLM: Toward Safety-aware Fine-grained Reasoning in Multimodal Models
Multimodal Large Reasoning Models (MLRMs) demonstrate impressive cross-modal reasoning but often amplify safety risks under adversarial or unsafe prompts, a phenomenon we call the \textit{Reasoning Tax}. Existing defenses mainly act at the output level and do not constrain the reasoning process, leaving models exposed to implicit risks. In this paper, we propose SaFeR-VLM, a safety-aligned reinforcement learning framework that embeds safety directly into multimodal reasoning. The framework integrates four components: (I) QI-Safe-10K, a curated dataset emphasizing safety-critical and reasoning-sensitive cases; (II) safety-aware rollout, where unsafe generations undergo reflection and correction instead of being discarded; (III) structured reward modeling with multi-dimensional weighted criteria and explicit penalties for hallucinations and contradictions; and (IV) GRPO optimization, which reinforces both safe and corrected trajectories. This unified design shifts safety from a passive safeguard to an active driver of reasoning, enabling scalable and generalizable safety-aware reasoning. SaFeR-VLM further demonstrates robustness against both explicit and implicit risks, supporting dynamic and interpretable safety decisions beyond surface-level filtering. SaFeR-VLM-3B achieves average performance $70.13$ and $78.97$ on safety and helpfulness across six benchmarks, surpassing both same-scale and $>10\times$ larger models such as Skywork-R1V3-38B, Qwen2.5VL-72B, and GLM4.5V-106B. Remarkably, SaFeR-VLM-7B benefits from its increased scale to surpass GPT-5-mini and Gemini-2.5-Flash by \num{6.47} and \num{16.76} points respectively on safety metrics, achieving this improvement without any degradation in helpfulness performance. Our codes are available at https://github.com/HarveyYi/SaFeR-VLM.
☆ Multi-hop Deep Joint Source-Channel Coding with Deep Hash Distillation for Semantically Aligned Image Retrieval
We consider image transmission via deep joint source-channel coding (DeepJSCC) over multi-hop additive white Gaussian noise (AWGN) channels by training a DeepJSCC encoder-decoder pair with a pre-trained deep hash distillation (DHD) module to semantically cluster images, facilitating security-oriented applications through enhanced semantic consistency and improving the perceptual reconstruction quality. We train the DeepJSCC module to both reduce mean square error (MSE) and minimize cosine distance between DHD hashes of source and reconstructed images. Significantly improved perceptual quality as a result of semantic alignment is illustrated for different multi-hop settings, for which classical DeepJSCC may suffer from noise accumulation, measured by the learned perceptual image patch similarity (LPIPS) metric.
☆ Towards Generalization of Graph Neural Networks for AC Optimal Power Flow
AC Optimal Power Flow (ACOPF) is computationally expensive for large-scale power systems, with conventional solvers requiring prohibitive solution times. Machine learning approaches offer computational speedups but struggle with scalability and topology adaptability without expensive retraining. To enable scalability across grid sizes and adaptability to topology changes, we propose a Hybrid Heterogeneous Message Passing Neural Network (HH-MPNN). HH-MPNN models buses, generators, loads, shunts, transmission lines and transformers as distinct node or edge types, combined with a scalable transformer model for handling long-range dependencies. On grids from 14 to 2,000 buses, HH-MPNN achieves less than 1% optimality gap on default topologies. Applied zero-shot to thousands of unseen topologies, HH-MPNN achieves less than 3% optimality gap despite training only on default topologies. Pre-training on smaller grids also improves results on a larger grid. Computational speedups reach 1,000x to 10,000x compared to interior point solvers. These results advance practical, generalizable machine learning for real-time power system operations.
comment: Pre-print has been submitted for review
☆ Enhancing Bankruptcy Prediction of Banks through Advanced Machine Learning Techniques: An Innovative Approach and Analysis
Context: Financial system stability is determined by the condition of the banking system. A bank failure can destroy the stability of the financial system, as banks are subject to systemic risk, affecting not only individual banks but also segments or the entire financial system. Calculating the probability of a bank going bankrupt is one way to ensure the banking system is safe and sound. Existing literature and limitations: Statistical models, such as Altman's Z-Score, are one of the common techniques for developing a bankruptcy prediction model. However, statistical methods rely on rigid and sometimes irrelevant assumptions, which can result in low forecast accuracy. New approaches are necessary. Objective of the research: Bankruptcy models are developed using machine learning techniques, such as logistic regression (LR), random forest (RF), and support vector machines (SVM). According to several studies, machine learning is also more accurate and effective than statistical methods for categorising and forecasting banking risk management. Present Research: The commercial bank data are derived from the annual financial statements of 44 active banks and 21 bankrupt banks in Turkey from 1994 to 2004, and the rural bank data are derived from the quarterly financial reports of 43 active and 43 bankrupt rural banks in Indonesia between 2013 and 2019. Five rural banks in Indonesia have also been selected to demonstrate the feasibility of analysing bank bankruptcy trends. Findings and implications: The results of the research experiments show that RF can forecast data from commercial banks with a 90% accuracy rate. Furthermore, the three machine learning methods proposed accurately predict the likelihood of rural bank bankruptcy. Contribution and Conclusion: The proposed innovative machine learning approach help to implement policies that reduce the costs of bankruptcy.
☆ Reconquering Bell sampling on qudits: stabilizer learning and testing, quantum pseudorandomness bounds, and more
Bell sampling is a simple yet powerful tool based on measuring two copies of a quantum state in the Bell basis, and has found applications in a plethora of problems related to stabiliser states and measures of magic. However, it was not known how to generalise the procedure from qubits to $d$-level systems -- qudits -- for all dimensions $d > 2$ in a useful way. Indeed, a prior work of the authors (arXiv'24) showed that the natural extension of Bell sampling to arbitrary dimensions fails to provide meaningful information about the quantum states being measured. In this paper, we overcome the difficulties encountered in previous works and develop a useful generalisation of Bell sampling to qudits of all $d\geq 2$. At the heart of our primitive is a new unitary, based on Lagrange's four-square theorem, that maps four copies of any stabiliser state $|\mathcal{S}\rangle$ to four copies of its complex conjugate $|\mathcal{S}^\ast\rangle$ (up to some Pauli operator), which may be of independent interest. We then demonstrate the utility of our new Bell sampling technique by lifting several known results from qubits to qudits for any $d\geq 2$: 1. Learning stabiliser states in $O(n^3)$ time with $O(n)$ samples; 2. Solving the Hidden Stabiliser Group Problem in $\tilde{O}(n^3/\varepsilon)$ time with $\tilde{O}(n/\varepsilon)$ samples; 3. Testing whether $|\psi\rangle$ has stabiliser size at least $d^t$ or is $\varepsilon$-far from all such states in $\tilde{O}(n^3/\varepsilon)$ time with $\tilde{O}(n/\varepsilon)$ samples; 4. Clifford circuits with at most $n/2$ single-qudit non-Clifford gates cannot prepare pseudorandom states; 5. Testing whether $|\psi\rangle$ has stabiliser fidelity at least $1-\varepsilon_1$ or at most $1-\varepsilon_2$ with $O(d^2/\varepsilon_2)$ samples if $\varepsilon_1 = 0$ or $O(d^2/\varepsilon_2^2)$ samples if $\varepsilon_1 = O(d^{-2})$.
comment: 51 pages, 1 figure. Comments are welcome
☆ CNN-TFT explained by SHAP with multi-head attention weights for time series forecasting
Convolutional neural networks (CNNs) and transformer architectures offer strengths for modeling temporal data: CNNs excel at capturing local patterns and translational invariances, while transformers effectively model long-range dependencies via self-attention. This paper proposes a hybrid architecture integrating convolutional feature extraction with a temporal fusion transformer (TFT) backbone to enhance multivariate time series forecasting. The CNN module first applies a hierarchy of one-dimensional convolutional layers to distill salient local patterns from raw input sequences, reducing noise and dimensionality. The resulting feature maps are then fed into the TFT, which applies multi-head attention to capture both short- and long-term dependencies and to weigh relevant covariates adaptively. We evaluate the CNN-TFT on a hydroelectric natural flow time series dataset. Experimental results demonstrate that CNN-TFT outperforms well-established deep learning models, with a mean absolute percentage error of up to 2.2%. The explainability of the model is obtained by a proposed Shapley additive explanations with multi-head attention weights (SHAP-MHAW). Our novel architecture, named CNN-TFT-SHAP-MHAW, is promising for applications requiring high-fidelity, multivariate time series forecasts, being available for future analysis at https://github.com/SFStefenon/CNN-TFT-SHAP-MHAW .
☆ Vectorized FlashAttention with Low-cost Exponential Computation in RISC-V Vector Processors
Attention is a core operation in numerous machine learning and artificial intelligence models. This work focuses on the acceleration of attention kernel using FlashAttention algorithm, in vector processors, particularly those based on the RISC-V instruction set architecture (ISA). This work represents the first effort to vectorize FlashAttention, minimizing scalar code and simplifying the computational complexity of evaluating exponentials needed by softmax used in attention. By utilizing a low-cost approximation for exponentials in floating-point arithmetic, we reduce the cost of computing the exponential function without the need to extend baseline vector ISA with new custom instructions. Also, appropriate tiling strategies are explored with the goal to improve memory locality. Experimental results highlight the scalability of our approach, demonstrating significant performance gains with the vectorized implementations when processing attention layers in practical applications.
☆ Early wind turbine alarm prediction based on machine learning: AlarmForecasting
Alarm data is pivotal in curbing fault behavior in Wind Turbines (WTs) and forms the backbone for advancedpredictive monitoring systems. Traditionally, research cohorts have been confined to utilizing alarm data solelyas a diagnostic tool, merely indicative of unhealthy status. However, this study aims to offer a transformativeleap towards preempting alarms, preventing alarms from triggering altogether, and consequently avertingimpending failures. Our proposed Alarm Forecasting and Classification (AFC) framework is designed on twosuccessive modules: first, the regression module based on long short-term memory (LSTM) for time-series alarmforecasting, and thereafter, the classification module to implement alarm tagging on the forecasted alarm. Thisway, the entire alarm taxonomy can be forecasted reliably rather than a few specific alarms. 14 Senvion MM82turbines with an operational period of 5 years are used as a case study; the results demonstrated 82%, 52%,and 41% accurate forecasts for 10, 20, and 30 min alarm forecasts, respectively. The results substantiateanticipating and averting alarms, which is significant in curbing alarm frequency and enhancing operationalefficiency through proactive intervention.
comment: International Journal of Electrical Power and Energy Systems
☆ Recurrence-Complete Frame-based Action Models
In recent years, attention-like mechanisms have been used to great success in the space of large language models, unlocking scaling potential to a previously unthinkable extent. "Attention Is All You Need" famously claims RNN cells are not needed in conjunction with attention. We challenge this view. In this paper, we point to existing proofs that architectures with fully parallelizable forward or backward passes cannot represent classes of problems specifically interesting for long-running agentic tasks. We further conjecture a critical time t beyond which non-recurrence-complete models fail to aggregate inputs correctly, with concrete implications for agentic systems (e.g., software engineering agents). To address this, we introduce a recurrence-complete architecture and train it on GitHub-derived action sequences. Loss follows a power law in the trained sequence length while the parameter count remains fixed. Moreover, longer-sequence training always amortizes its linearly increasing wall-time cost, yielding lower loss as a function of wall time.
☆ Efficient numeracy in language models through single-token number embeddings
To drive progress in science and engineering, large language models (LLMs) must be able to process large amounts of numerical data and solve long calculations efficiently. This is currently only possible through the use of external tools or extensive reasoning chains, either limiting the numerical intuition of LLMs or limiting the length of problems they can solve. We show that frontier LLMs require excessive amounts of reasoning tokens to solve even basic calculations, which is exacerbated by their tokenization strategies that split single numbers into multiple tokens. This motivates the need for efficient and effective single-token number encodings. We introduce a set of desiderata for such encodings and show that existing approaches fail to fulfill them. To address these shortcomings, we propose BitTokens, a novel tokenization strategy that embeds any number into a single token using its IEEE 754 binary floating-point representation. Through extensive experiments we show that our BitTokens allow even small language models to learn algorithms that solve basic arithmetic operations nearly perfectly. This newly gained efficiency could expand the length and complexity of problems language models can solve.
☆ Efficient Discriminative Joint Encoders for Large Scale Vision-Language Reranking
Multimodal retrieval still leans on embedding-based models like CLIP for fast vector search over pre-computed image embeddings. Yet, unlike text retrieval, where joint-encoder rerankers are standard, comparable vision--language rerankers are largely absent. We find that seminal joint encoders such as BLIP are severely bottlenecked by an expensive visual feature-extraction stage, preventing practical deployment at scale. Motivated by this bottleneck, we introduce EDJE, an Efficient Discriminative Joint Encoder that precomputes vision tokens offline and compresses them via a lightweight attention-based adapter, so online inference runs only a compact joint encoder over a small set of visual tokens plus the text. EDJE preserves strong retrieval performance while drastically reducing storage and online compute, enabling high-throughput inference. Specifically, EDJE processes 50k image--text pairs/second while requiring 49kB of disk storage per image, matching prior art on Flickr (zero-shot) and COCO (fine-tuned) retrieval. The implementation and checkpoints will be made publicly available shortly.
comment: preprint
☆ The Unreasonable Effectiveness of Randomized Representations in Online Continual Graph Learning
Catastrophic forgetting is one of the main obstacles for Online Continual Graph Learning (OCGL), where nodes arrive one by one, distribution drifts may occur at any time and offline training on task-specific subgraphs is not feasible. In this work, we explore a surprisingly simple yet highly effective approach for OCGL: we use a fixed, randomly initialized encoder to generate robust and expressive node embeddings by aggregating neighborhood information, training online only a lightweight classifier. By freezing the encoder, we eliminate drifts of the representation parameters, a key source of forgetting, obtaining embeddings that are both expressive and stable. When evaluated across several OCGL benchmarks, despite its simplicity and lack of memory buffer, this approach yields consistent gains over state-of-the-art methods, with surprising improvements of up to 30% and performance often approaching that of the joint offline-training upper bound. These results suggest that in OCGL, catastrophic forgetting can be minimized without complex replay or regularization by embracing architectural simplicity and stability.
☆ BlackboxNLP-2025 MIB Shared Task: Exploring Ensemble Strategies for Circuit Localization Methods
The Circuit Localization track of the Mechanistic Interpretability Benchmark (MIB) evaluates methods for localizing circuits within large language models (LLMs), i.e., subnetworks responsible for specific task behaviors. In this work, we investigate whether ensembling two or more circuit localization methods can improve performance. We explore two variants: parallel and sequential ensembling. In parallel ensembling, we combine attribution scores assigned to each edge by different methods-e.g., by averaging or taking the minimum or maximum value. In the sequential ensemble, we use edge attribution scores obtained via EAP-IG as a warm start for a more expensive but more precise circuit identification method, namely edge pruning. We observe that both approaches yield notable gains on the benchmark metrics, leading to a more precise circuit identification approach. Finally, we find that taking a parallel ensemble over various methods, including the sequential ensemble, achieves the best results. We evaluate our approach in the BlackboxNLP 2025 MIB Shared Task, comparing ensemble scores to official baselines across multiple model-task combinations.
comment: The 8th BlackboxNLP Workshop (Shared Task), 6 pages
☆ Quantum Computing Methods for Malware Detection
In this paper, we explore the potential of quantum computing in enhancing malware detection through the application of Quantum Machine Learning (QML). Our main objective is to investigate the performance of the Quantum Support Vector Machine (QSVM) algorithm compared to SVM. A publicly available dataset containing raw binaries of Portable Executable (PE) files was used for the classification. The QSVM algorithm, incorporating quantum kernels through different feature maps, was implemented and evaluated on a local simulator within the Qiskit SDK and IBM quantum computers. Experimental results from simulators and quantum hardware provide insights into the behavior and performance of quantum computers, especially in handling large-scale computations for malware detection tasks. The work summarizes the practical experience with using quantum hardware via the Qiskit interfaces. We describe in detail the critical issues encountered, as well as the fixes that had to be developed and applied to the base code of the Qiskit Machine Learning library. These issues include missing transpilation of the circuits submitted to IBM Quantum systems and exceeding the maximum job size limit due to the submission of all the circuits in one job.
comment: 22 pages, 2 figures, 3 tables
☆ Get RICH or Die Scaling: Profitably Trading Inference Compute for Robustness
Models are susceptible to adversarially out-of-distribution (OOD) data despite large training-compute investments into their robustification. Zaremba et al. (2025) make progress on this problem at test time, showing LLM reasoning improves satisfaction of model specifications designed to thwart attacks, resulting in a correlation between reasoning effort and robustness to jailbreaks. However, this benefit of test compute fades when attackers are given access to gradients or multimodal inputs. We address this gap, clarifying that inference-compute offers benefits even in such cases. Our approach argues that compositional generalization, through which OOD data is understandable via its in-distribution (ID) components, enables adherence to defensive specifications on adversarially OOD inputs. Namely, we posit the Robustness from Inference Compute Hypothesis (RICH): inference-compute defenses profit as the model's training data better reflects the attacked data's components. We empirically support this hypothesis across vision language model and attack types, finding robustness gains from test-time compute if specification following on OOD data is unlocked by compositional generalization, while RL finetuning and protracted reasoning are not critical. For example, increasing emphasis on defensive specifications via prompting lowers the success rate of gradient-based multimodal attacks on VLMs robustified by adversarial pretraining, but this same intervention provides no such benefit to not-robustified models. This correlation of inference-compute's robustness benefit with base model robustness is the rich-get-richer dynamic of the RICH: attacked data components are more ID for robustified models, aiding compositional generalization to OOD data. Accordingly, we advise layering train-time and test-time defenses to obtain their synergistic benefit.
comment: 17 pages
♻ ☆ NdLinear: Preserving Multi-Dimensional Structure for Parameter-Efficient Neural Networks
In deep learning, processing multidimensional inputs (e.g., images, medical scans, and time series) is an important task that often requires flattening the inputs. We introduce $\mathit{NdLinear}$, a drop-in replacement for linear layers that operates directly on tensors, requiring no flattening. By applying transformations separately along each dimension, NdLinear preserves native data structure while achieving dramatic parameter reductions, often by orders of magnitude, with minimal memory overhead. We prove NdLinear maintains expressivity through structured Tucker decomposition while preserving VC-dimension scaling. Extensive experiments demonstrate NdLinear's capacity to achieve significant parameter reductions with substantial wall-clock efficiency gains and minimal memory overhead. For instance, our $\mathit{NdLinear-LoRA}$ matches or exceeds standard LoRA on language reasoning tasks using up to $9\times$ fewer parameters. Experiments across CNNs, RNNs, Transformers, and MLPs on vision, language, time-series, and tabular tasks consistently demonstrate NdLinear's efficiency gains. While excelling at axis-separable tasks, NdLinear has limitations with entangled spatial interactions. By processing data in its original N-dimensional form, NdLinear provides a theoretically grounded, practical component for building more efficient neural architectures.
comment: Code is available at https://github.com/ensemble-core/NdLinear
♻ ☆ Valid Inference with Imperfect Synthetic Data NeurIPS 2025
Predictions and generations from large language models are increasingly being explored as an aid in limited data regimes, such as in computational social science and human subjects research. While prior technical work has mainly explored the potential to use model-predicted labels for unlabeled data in a principled manner, there is increasing interest in using large language models to generate entirely new synthetic samples (e.g., synthetic simulations), such as in responses to surveys. However, it remains unclear by what means practitioners can combine such data with real data and yet produce statistically valid conclusions upon them. In this paper, we introduce a new estimator based on generalized method of moments, providing a hyperparameter-free solution with strong theoretical guarantees to address this challenge. Intriguingly, we find that interactions between the moment residuals of synthetic data and those of real data (i.e., when they are predictive of each other) can greatly improve estimates of the target parameter. We validate the finite-sample performance of our estimator across different tasks in computational social science applications, demonstrating large empirical gains.
comment: NeurIPS 2025
♻ ☆ Dyna-Think: Synergizing Reasoning, Acting, and World Model Simulation in AI Agents
Recent progress in reasoning with large language models (LLMs), such as DeepSeek-R1, demonstrates impressive capabilities in domains like mathematics and coding, by exhibiting complex cognitive behaviors such as verification, goal decomposition, and self-reflection. However, it is unclear what behavior is effective and what behavior is missing for long-horizon AI agents tasks. In this work, we propose Dyna-Think, a thinking framework that integrates planning with an internal world model with reasoning and acting to enhance AI agent performance. To enable Dyna-Think, we propose Dyna-Think Imitation Learning (DIT) and Dyna-Think Dyna Training (DDT). To initialize a policy with Dyna-Think, DIT reconstructs the thinking process of R1 to focus on performing world model simulation relevant to the proposed (and planned) action, and trains the policy using this reconstructed data. To enhance Dyna-Think, DDT uses a two-stage training process to first improve the agent's world modeling ability via objectives such as state prediction or critique generation, and then improve the agent's action via policy training. We evaluate our methods on OSWorld and WindowsAgentArena, and demonstrate that Dyna-Think improves the agent's in-domain and out-of-domain performance, achieving similar best-of-n performance compared to R1 while generating 2x less tokens on average. Our extensive empirical studies reveal that 1) using critique generation for world model training is effective to improve policy performance; and 2) AI agents with better performance correlate with better world modeling abilities. We believe our results suggest a promising research direction to integrate world model simulation into AI agents to enhance their reasoning, planning, and acting capabilities.
♻ ☆ BLISS: A Lightweight Bilevel Influence Scoring Method for Data Selection in Language Model Pretraining
Effective data selection is essential for pretraining large language models (LLMs), enhancing efficiency and improving generalization to downstream tasks. However, existing approaches often require leveraging external pretrained models, making it difficult to disentangle the effects of data selection from those of the external pretrained models. In addition, they often overlook the long-term impact of selected data if the model is trained to convergence, primarily due to the prohibitive cost of full-scale LLM pretraining. In this paper, we introduce BLISS (\textbf{B}ileve\textbf{L} \textbf{I}nfluence \textbf{S}coring method for data \textbf{S}election): a lightweight data selection method that operates entirely \emph{from scratch}, without relying on any external pretrained oracle models, while explicitly accounting for the long-term impact of selected data. BLISS leverages a small proxy model as a surrogate for the LLM and employs a score model to estimate the long-term influence of training samples if the proxy model is trained to convergence. We formulate data selection as a bilevel optimization problem, where the upper-level objective optimizes the score model to assign importance weights to training samples, ensuring that minimizing the lower-level objective (i.e., training the proxy model over the weighted training loss until convergence) leads to best validation performance. Once optimized, the trained score model predicts influence scores for the dataset, enabling efficient selection of high-quality samples for LLM pretraining. We validate BLISS by pretraining 410M/1B/2.8B Pythia and LLaMA-0.5B models on selected subsets of the C4 dataset. Notably, under the 1B model setting, BLISS achieves $1.7\times$ speedup in reaching the same performance as the state-of-the-art method, demonstrating superior performance across multiple downstream tasks.
♻ ☆ SafeProtein: Red-Teaming Framework and Benchmark for Protein Foundation Models
Proteins play crucial roles in almost all biological processes. The advancement of deep learning has greatly accelerated the development of protein foundation models, leading to significant successes in protein understanding and design. However, the lack of systematic red-teaming for these models has raised serious concerns about their potential misuse, such as generating proteins with biological safety risks. This paper introduces SafeProtein, the first red-teaming framework designed for protein foundation models to the best of our knowledge. SafeProtein combines multimodal prompt engineering and heuristic beam search to systematically design red-teaming methods and conduct tests on protein foundation models. We also curated SafeProtein-Bench, which includes a manually constructed red-teaming benchmark dataset and a comprehensive evaluation protocol. SafeProtein achieved continuous jailbreaks on state-of-the-art protein foundation models (up to 70% attack success rate for ESM3), revealing potential biological safety risks in current protein foundation models and providing insights for the development of robust security protection technologies for frontier models. The codes will be made publicly available at https://github.com/jigang-fan/SafeProtein.
♻ ☆ Empirical Comparison of Membership Inference Attacks in Deep Transfer Learning
With the emergence of powerful large-scale foundation models, the training paradigm is increasingly shifting from from-scratch training to transfer learning. This enables high utility training with small, domain-specific datasets typical in sensitive applications. Membership inference attacks (MIAs) provide an empirical estimate of the privacy leakage by machine learning models. Yet, prior assessments of MIAs against models fine-tuned with transfer learning rely on a small subset of possible attacks. We address this by comparing performance of diverse MIAs in transfer learning settings to help practitioners identify the most efficient attacks for privacy risk evaluation. We find that attack efficacy decreases with the increase in training data for score-based MIAs. We find that there is no one MIA which captures all privacy risks in models trained with transfer learning. While the Likelihood Ratio Attack (LiRA) demonstrates superior performance across most experimental scenarios, the Inverse Hessian Attack (IHA) proves to be more effective against models fine-tuned on PatchCamelyon dataset in high data regime.
comment: 30 pages, 13 figures, published in TMLR https://openreview.net/forum?id=UligTUCgdt
♻ ☆ Dual Natural Gradient Descent for Scalable Training of Physics-Informed Neural Networks
Natural-gradient methods markedly accelerate the training of Physics-Informed Neural Networks (PINNs), yet their Gauss--Newton update must be solved in the parameter space, incurring a prohibitive $O(n^3)$ time complexity, where $n$ is the number of network trainable weights. We show that exactly the same step can instead be formulated in a generally smaller residual space of size $m = \sum_{\gamma} N_{\gamma} d_{\gamma}$, where each residual class $\gamma$ (e.g. PDE interior, boundary, initial data) contributes $N_{\gamma}$ collocation points of output dimension $d_{\gamma}$. Building on this insight, we introduce \textit{Dual Natural Gradient Descent} (D-NGD). D-NGD computes the Gauss--Newton step in residual space, augments it with a geodesic-acceleration correction at negligible extra cost, and provides both a dense direct solver for modest $m$ and a Nystrom-preconditioned conjugate-gradient solver for larger $m$. Experimentally, D-NGD scales second-order PINN optimization to networks with up to 12.8 million parameters, delivers one- to three-order-of-magnitude lower final error $L^2$ than first-order methods (Adam, SGD) and quasi-Newton methods, and -- crucially -- enables natural-gradient training of PINNs at this scale on a single GPU.
♻ ☆ Lossy Neural Compression for Geospatial Analytics: A Review
Over the past decades, there has been an explosion in the amount of available Earth Observation (EO) data. The unprecedented coverage of the Earth's surface and atmosphere by satellite imagery has resulted in large volumes of data that must be transmitted to ground stations, stored in data centers, and distributed to end users. Modern Earth System Models (ESMs) face similar challenges, operating at high spatial and temporal resolutions, producing petabytes of data per simulated day. Data compression has gained relevance over the past decade, with neural compression (NC) emerging from deep learning and information theory, making EO data and ESM outputs ideal candidates due to their abundance of unlabeled data. In this review, we outline recent developments in NC applied to geospatial data. We introduce the fundamental concepts of NC including seminal works in its traditional applications to image and video compression domains with focus on lossy compression. We discuss the unique characteristics of EO and ESM data, contrasting them with "natural images", and explain the additional challenges and opportunities they present. Moreover, we review current applications of NC across various EO modalities and explore the limited efforts in ESM compression to date. The advent of self-supervised learning (SSL) and foundation models (FM) has advanced methods to efficiently distill representations from vast unlabeled data. We connect these developments to NC for EO, highlighting the similarities between the two fields and elaborate on the potential of transferring compressed feature representations for machine--to--machine communication. Based on insights drawn from this review, we devise future directions relevant to applications in EO and ESM.
comment: self-consistent review paper
♻ ☆ AutoMind: Adaptive Knowledgeable Agent for Automated Data Science
Large Language Model (LLM) agents have shown great potential in addressing real-world data science problems. LLM-driven data science agents promise to automate the entire machine learning pipeline, yet their real-world effectiveness remains limited. Existing frameworks depend on rigid, pre-defined workflows and inflexible coding strategies; consequently, they excel only on relatively simple, classical problems and fail to capture the empirical expertise that human practitioners bring to complex, innovative tasks. In this work, we introduce AutoMind, an adaptive, knowledgeable LLM-agent framework that overcomes these deficiencies through three key advances: (1) a curated expert knowledge base that grounds the agent in domain expert knowledge, (2) an agentic knowledgeable tree search algorithm that strategically explores possible solutions, and (3) a self-adaptive coding strategy that dynamically tailors code generation to task complexity. Evaluations on two automated data science benchmarks demonstrate that AutoMind delivers superior performance versus state-of-the-art baselines. Additional analyses confirm favorable effectiveness, efficiency, and qualitative solution quality, highlighting AutoMind as an efficient and robust step toward fully automated data science. Code is at https://github.com/innovatingAI/AutoMind.
comment: Ongoing work
♻ ☆ Prefilled responses enhance zero-shot detection of AI-generated images
As AI models generate increasingly realistic images, growing concerns over potential misuse underscore the need for reliable detection. Traditional supervised detection methods depend on large, curated datasets for training and often fail to generalize to novel, out-of-domain image generators. As an alternative, we explore pre-trained Vision-Language Models (VLMs) for zero-shot detection of AI-generated images. We evaluate VLM performance on three diverse benchmarks encompassing synthetic images of human faces, objects, and animals produced by 16 different state-of-the-art image generators. While off-the-shelf VLMs perform poorly on these datasets, we find that their reasoning can be guided effectively through simple response prefilling -- a method we call Prefill-Guided Thinking (PGT). In particular, prefilling a VLM response with the task-aligned phrase "Let's examine the style and the synthesis artifacts" improves the Macro F1 scores of three widely used open-source VLMs by up to 24%.
♻ ☆ KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality
Large Language Models (LLMs), particularly slow-thinking models, often exhibit severe hallucination, outputting incorrect content due to an inability to accurately recognize knowledge boundaries during reasoning. While Reinforcement Learning (RL) can enhance complex reasoning abilities, its outcome-oriented reward mechanism often lacks factual supervision over the thinking process, further exacerbating the hallucination problem. To address the high hallucination in slow-thinking models, we propose Knowledge-enhanced RL, KnowRL. KnowRL guides models to perform fact-based slow thinking by integrating a factuality reward, based on knowledge verification, into the RL training process, helping them recognize their knowledge boundaries. KnowRL guides models to perform fact-based slow thinking by integrating a factuality reward, based on knowledge verification, into the RL training process, helping them recognize their knowledge boundaries. This targeted factual input during RL training enables the model to learn and internalize fact-based reasoning strategies. By directly rewarding adherence to facts within the reasoning steps, KnowRL fosters a more reliable thinking process. Experimental results on three hallucination evaluation datasets and two reasoning evaluation datasets demonstrate that KnowRL effectively mitigates hallucinations in slow-thinking models while maintaining their original strong reasoning capabilities. Our code is available at https://github.com/zjunlp/KnowRL.
comment: Work in progress
♻ ☆ Bit-Level Discrete Diffusion with Markov Probabilistic Models: An Improved Framework with Sharp Convergence Bounds under Minimal Assumptions
This paper introduces Discrete Markov Probabilistic Models (DMPMs), a novel discrete diffusion algorithm for discrete data generation. The algorithm operates in discrete bit space, where the noising process is a continuous-time Markov chain that flips labels uniformly at random. The time-reversal process, like the forward noise process, is a jump process with its intensity governed by a discrete analogue of the classical score function. Crucially, this intensity is proven to be the conditional expectation of a function of the forward process, underlining theoretical alignment with score-based generative models. We establish convergence bounds for the algorithm under minimal assumptions, ensuring robustness and efficiency, which we demonstrate through experiments on low-dimensional Bernoulli-distributed datasets and high-dimensional binary MNIST data. The results highlight competitive performance in generating discrete structures compared to the state-of-the-art. This work bridges theoretical foundations and practical applications, advancing the development of effective and theoretically grounded discrete generative modeling.
♻ ☆ Differential Privacy for Adaptive Weight Aggregation in Federated Tumor Segmentation
Federated Learning (FL) is a distributed machine learning approach that safeguards privacy by creating an impartial global model while respecting the privacy of individual client data. However, the conventional FL method can introduce security risks when dealing with diverse client data, potentially compromising privacy and data integrity. To address these challenges, we present a differential privacy (DP) federated deep learning framework in medical image segmentation. In this paper, we extend our similarity weight aggregation (SimAgg) method to DP-SimAgg algorithm, a differentially private similarity-weighted aggregation algorithm for brain tumor segmentation in multi-modal magnetic resonance imaging (MRI). Our DP-SimAgg method not only enhances model segmentation capabilities but also provides an additional layer of privacy preservation. Extensive benchmarking and evaluation of our framework, with computational performance as a key consideration, demonstrate that DP-SimAgg enables accurate and robust brain tumor segmentation while minimizing communication costs during model training. This advancement is crucial for preserving the privacy of medical image data and safeguarding sensitive information. In conclusion, adding a differential privacy layer in the global weight aggregation phase of the federated brain tumor segmentation provides a promising solution to privacy concerns without compromising segmentation model efficacy. By leveraging DP, we ensure the protection of client data against adversarial attacks and malicious participants.
comment: I have changed the methodology because of some technical errors in this version
♻ ☆ AMBER: Adaptive Mesh Generation by Iterative Mesh Resolution Prediction NeurIPS2025
The cost and accuracy of simulating complex physical systems using the Finite Element Method (FEM) scales with the resolution of the underlying mesh. Adaptive meshes improve computational efficiency by refining resolution in critical regions, but typically require task-specific heuristics or cumbersome manual design by a human expert. We propose Adaptive Meshing By Expert Reconstruction (AMBER), a supervised learning approach to mesh adaptation. Starting from a coarse mesh, AMBER iteratively predicts the sizing field, i.e., a function mapping from the geometry to the local element size of the target mesh, and uses this prediction to produce a new intermediate mesh using an out-of-the-box mesh generator. This process is enabled through a hierarchical graph neural network, and relies on data augmentation by automatically projecting expert labels onto AMBER-generated data during training. We evaluate AMBER on 2D and 3D datasets, including classical physics problems, mechanical components, and real-world industrial designs with human expert meshes. AMBER generalizes to unseen geometries and consistently outperforms multiple recent baselines, including ones using Graph and Convolutional Neural Networks, and Reinforcement Learning-based approaches.
comment: Accepted to NeurIPS2025
♻ ☆ Last-iterate Convergence for Symmetric, General-sum, $2 \times 2$ Games Under The Exponential Weights Dynamic
We conduct a comprehensive analysis of the discrete-time exponential-weights dynamic with a constant step size on all \emph{general-sum and symmetric} $2 \times 2$ normal-form games, i.e. games with $2$ pure strategies per player, and where the ensuing payoff tuple is of the form $(A,A^\top)$ (where $A$ is the $2 \times 2$ payoff matrix corresponding to the first player). Such symmetric games commonly arise in real-world interactions between "symmetric" agents who have identically defined utility functions -- such as Bertrand competition, multi-agent performative prediction, and certain congestion games -- and display a rich multiplicity of equilibria despite the seemingly simple setting. Somewhat surprisingly, we show through a first-principles analysis that the exponential weights dynamic, which is popular in online learning, converges in the last iterate for such games regardless of initialization with an appropriately chosen step size. For certain games and/or initializations, we further show that the convergence rate is in fact exponential and holds for any step size. We illustrate our theory with extensive simulations and applications to the aforementioned game-theoretic interactions. In the case of multi-agent performative prediction, we formulate a new "mortgage competition" game between lenders (i.e. banks) who interact with a population of customers, and show that it fits into our framework.
♻ ☆ From Injection to Defense: Constructing Edit-Based Fingerprints for Large Language Models
Fingerprinting is critical for maintaining traceability and protecting the intellectual property (IP) of developers, as LLMs deployed in web applications are susceptible to unauthorized redistribution and misuse via fine-tuning or black-box deployment. However, current backdoor-based fingerprinting methods face a fundamental trade-off: fingerprints embedded as garbled text are easily detected and filtered, whereas those crafted as coherent natural language are prone to being triggered unintentionally. To overcome these limitations, we propose RFEdit, a knowledge-editing framework that embeds a rule-based multilingual natural language fingerprint (MNLF) by modifying a sparse subset of model weights. This approach enables efficient and robust fingerprint injection with minimal impact on unrelated knowledge in LLMs. Our RFEdit framework is further safeguarded by Fingerprint Subspace-aware Fine-Tuning (FSFT), which mitigates fingerprint degradation during legitimate fine-tuning by restricting parameter updates to the fingerprint subspace. This approach preserves fingerprint integrity while enhancing downstream task performance of LLMs. These advances establish a comprehensive pipeline from fingerprint injection to defense, achieving high detection effectiveness, robustness against adversarial manipulations, harmlessness to model utility, and persistence under fine-tuning. Extensive experiments demonstrate that RFEdit maintains robustness under quantization and pruning. Additionally, fingerprint effectiveness is generally improved by more than 10\% when combined with FSFT for math and alpaca downstream tasks.
comment: preprint
♻ ☆ Autonomy-Aware Clustering: When Local Decisions Supersede Global Prescriptions
Clustering arises in a wide range of problem formulations, yet most existing approaches assume that the entities under clustering are passive and strictly conform to their assigned groups. In reality, entities often exhibit local autonomy, overriding prescribed associations in ways not fully captured by feature representations. Such autonomy can substantially reshape clustering outcomes -- altering cluster compositions, geometry, and cardinality -- with significant downstream effects on inference and decision-making. We introduce autonomy-aware clustering, a reinforcement learning (RL) framework that learns and accounts for the influence of local autonomy without requiring prior knowledge of its form. Our approach integrates RL with a Deterministic Annealing (DA) procedure, where, to determine underlying clusters, DA naturally promotes exploration in early stages of annealing and transitions to exploitation later. We also show that the annealing procedure exhibits phase transitions that enable design of efficient annealing schedules. To further enhance adaptability, we propose the Adaptive Distance Estimation Network (ADEN), a transformer-based attention model that learns dependencies between entities and cluster representatives within the RL loop, accommodates variable-sized inputs and outputs, and enables knowledge transfer across diverse problem instances. Empirical results show that our framework closely aligns with underlying data dynamics: even without explicit autonomy models, it achieves solutions close to the ground truth (gap ~3-4%), whereas ignoring autonomy leads to substantially larger gaps (~35-40%). The code and data are publicly available at https://github.com/salar96/AutonomyAwareClustering.
comment: Preprint. Under review at a peer-reviewed venue. Minor formatting correction: the earlier version included an incorrect conference header, which has been removed. Content unchanged
♻ ☆ On Task Vectors and Gradients
Task arithmetic has emerged as a simple yet powerful technique for model merging, enabling the combination of multiple finetuned models into one. Despite its empirical success, a clear theoretical explanation of why and when it works is lacking. This paper provides a rigorous theoretical foundation for task arithmetic by establishing a connection between task vectors and gradients of the task losses. We show that under standard gradient descent, a task vector generated from one epoch of finetuning is exactly equivalent to the negative gradient of the loss, scaled by the learning rate. For the practical multi-epoch setting, we prove that this equivalence holds approximately, with a second-order error term that we explicitly bound for feed-forward networks. Our empirical analysis across seven vision benchmarks corroborates our theory, demonstrating that the first-epoch gradient dominates the finetuning trajectory in both norm and direction. A key implication is that merging models finetuned for only a single epoch often yields performance comparable to merging fully converged models. These findings reframe task arithmetic as a form of approximate multitask learning, providing a clear rationale for its effectiveness and highlighting the critical role of early training dynamics in model merging.
comment: 10 pages of main paper, 5 figures
♻ ☆ AbsoluteNet: A Deep Learning Neural Network to Classify Cerebral Hemodynamic Responses of Auditory Processing
In recent years, deep learning (DL) approaches have demonstrated promising results in decoding hemodynamic responses captured by functional near-infrared spectroscopy (fNIRS), particularly in the context of brain-computer interface (BCI) applications. This work introduces AbsoluteNet, a novel deep learning architecture designed to classify auditory event-related responses recorded using fNIRS. The proposed network is built upon principles of spatio-temporal convolution and customized activation functions. Our model was compared against several models, namely fNIRSNET, MDNN, DeepConvNet, and ShallowConvNet. The results showed that AbsoluteNet outperforms existing models, reaching 87.0% accuracy, 84.8% sensitivity, and 89.2% specificity in binary classification, surpassing fNIRSNET, the second-best model, by 3.8% in accuracy. These findings underscore the effectiveness of our proposed deep learning model in decoding hemodynamic responses related to auditory processing and highlight the importance of spatio-temporal feature aggregation and customized activation functions to better fit fNIRS dynamics.
♻ ☆ FFT-based Dynamic Subspace Selection for Low-Rank Adaptive Optimization of Large Language Models
Low-rank optimization has emerged as a promising direction in training large language models (LLMs) to improve running time and reduce the memory usage of adaptive optimizers by constraining learning to a lower-dimensional space. Prior work typically projects gradients of linear layers using approaches based on Singular Value Decomposition (SVD) or QR-decomposition. Applying these techniques individually to each layer in large models is computationally expensive and incurs additional memory costs due to storing the projection matrices. In this work, we propose a computationally efficient and conceptually simple, two-step procedure to approximate SVD/QR-based gradient projections into lower-dimensional spaces by using a predefined orthogonal matrix of the Discrete Cosine Transform (DCT). We dynamically select columns from the DCT matrix based on their alignment with the gradient of each layer. The effective projection matrices are obtained via a simple matmul with the DCT matrix in $O(n^3)$ time, followed by a lightweight sorting step to identify the most relevant basis vectors. For large layers, DCT can be computed via Makhoul's $N$-point algorithm based on Fast Fourier Transform (FFT) in $O(n^2 \log(n))$ time. Due to the predefined nature of the orthogonal bases, they are computed once at the start of training. Our numerical experiments on both pre-training and fine-tuning tasks demonstrate the effectiveness of our dual strategy in approximating optimal low-rank projections, obtaining an approach with rank-independent running time that matches the performance of costly SVD/QR-based methods while achieving faster runtime and reduced memory usage by up to $25\%$ across different model sizes. Our code is available at \href{https://github.com/IST-DASLab/ISTA-DASLab-Optimizers}{\texttt{https://github.com/IST-DASLab/ISTA-DASLab-Optimizers}}.
♻ ☆ Enjoying Non-linearity in Multinomial Logistic Bandits
We consider the multinomial logistic bandit problem, a variant of where a learner interacts with an environment by selecting actions to maximize expected rewards based on probabilistic feedback from multiple possible outcomes. In the binary setting, recent work has focused on understanding the impact of the non-linearity of the logistic model (Faury et al., 2020; Abeille et al., 2021). They introduced a problem-dependent constant $\kappa_* \geq 1$, that may be exponentially large in some problem parameters and which is captured by the derivative of the sigmoid function. It encapsulates the non-linearity and improves existing regret guarantees over $T$ rounds from $\smash{O(d\sqrt{T})}$ to $\smash{O(d\sqrt{T/\kappa_*})}$, where $d$ is the dimension of the parameter space. We extend their analysis to the multinomial logistic bandit framework, making it suitable for complex applications with more than two choices, such as reinforcement learning or recommender systems. To achieve this, we extend the definition of $\kappa_*$ to the multinomial setting and propose an efficient algorithm that leverages the problem's non-linearity. Our method yields a problem-dependent regret bound of order $ \smash{\widetilde{\mathcal{O}}( R d \sqrt{{KT}/{\kappa_*}})} $, where $R$ is the norm of the vector of rewards and $K$ is the number of outcomes. This improves upon the best existing guarantees of order $ \smash{\widetilde{\mathcal{O}}( RdK \sqrt{T} )} $. Moreover, we provide a $\smash{ \Omega(Rd\sqrt{KT/\kappa_*})}$ lower-bound, showing that our algorithm is minimax-optimal and that our definition of $\kappa_*$ is optimal.
♻ ☆ MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning
Learning object-level, structured representations is widely regarded as a key to better generalization in vision and underpins the design of next-generation Pre-trained Vision Models (PVMs). Mainstream Object-Centric Learning (OCL) methods adopt Slot Attention or its variants to iteratively aggregate objects' super-pixels into a fixed set of query feature vectors, termed slots. However, their reliance on a static slot count leads to an object being represented as multiple parts when the number of objects varies. We introduce MetaSlot, a plug-and-play Slot Attention variant that adapts to variable object counts. MetaSlot (i) maintains a codebook that holds prototypes of objects in a dataset by vector-quantizing the resulting slot representations; (ii) removes duplicate slots from the traditionally aggregated slots by quantizing them with the codebook; and (iii) injects progressively weaker noise into the Slot Attention iterations to accelerate and stabilize the aggregation. MetaSlot is a general Slot Attention variant that can be seamlessly integrated into existing OCL architectures. Across multiple public datasets and tasks--including object discovery and recognition--models equipped with MetaSlot achieve significant performance gains and markedly interpretable slot representations, compared with existing Slot Attention variants.
♻ ☆ TokenWeave: Efficient Compute-Communication Overlap for Distributed LLM Inference
Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLink. Multiple techniques have been proposed to mitigate these overheads by decomposing computations into finer-grained tasks and overlapping communication with sub-tasks as they complete. However, fine-grained decomposition of a large computation into many smaller computations on GPUs results in overheads. Furthermore, the communication itself uses many streaming multiprocessors (SMs), adding to the overhead. We present TokenWeave to address these challenges. TokenWeave proposes a Token-Splitting technique that divides the tokens in the inference batch into two approximately equal subsets in a wave-aware manner. The communication of one subset is then overlapped with the computation of the other. In addition, TokenWeave optimizes the order of the layer normalization computation with respect to communication operations and implements a novel fused AllReduce--RMSNorm kernel that carefully leverages Multimem instruction support available on Hopper and Blackwell NVIDIA GPUs. These optimizations allow TokenWeave to perform communication and RMSNorm using only 2-8 SMs. Moreover, our kernel enables the memory-bound RMSNorm to be overlapped with the other batch's computation, providing additional gains. Our evaluations demonstrate up to 1.29x speedup in latency and 1.26x higher throughput across multiple models and workloads. In several settings, TokenWeave results in better performance compared to an equivalent model with all communication removed.
comment: 14 pages, 16 figures. For source code, see https://github.com/microsoft/tokenweave. In version 2, Figure 6 shows All Reduce bandwidth, not Reduce Scatter. The Multimem Reduce Scatter bandwidth formula differs slightly from the Ring-based version
♻ ☆ Want to train KANS at scale? Now UKAN!
Kolmogorov-Arnold Networks (KANs) have recently emerged as a powerful alternative to traditional multilayer perceptrons. However, their reliance on predefined, bounded grids restricts their ability to approximate functions on unbounded domains. To address this, we present Unbounded Kolmogorov-Arnold Networks (UKANs), a method that removes the need for bounded grids in traditional Kolmogorov-Arnold Networks (KANs). The key innovation of this method is a coefficient-generator (CG) model that produces, on the fly, only the B-spline coefficients required locally on an unbounded symmetric grid. UKANs couple multilayer perceptrons with KANs by feeding the positional encoding of grid groups into the CG model, enabling function approximation on unbounded domains without requiring data normalization. To reduce the computational cost of both UKANs and KANs, we introduce a GPU-accelerated library that lowers B-spline evaluation complexity by a factor proportional to the grid size, enabling large-scale learning by leveraging efficient memory management, in line with recent software advances such as FlashAttention and FlashFFTConv. Performance benchmarking confirms the superior memory and computational efficiency of our accelerated KAN (warpKAN), and UKANs, showing a 3-30x speed-up and up to 1000x memory reduction compared to vanilla KANs. Experiments on regression, classification, and generative tasks demonstrate the effectiveness of UKANs to match or surpass KAN accuracy. Finally, we use both accelerated KAN and UKAN in a molecular property prediction task, establishing the feasibility of large-scale end-to-end training with our optimized implementation.
comment: 16 pages, 5 figures, 8 tables
♻ ☆ 360-LLaMA-Factory: Plug & Play Sequence Parallelism for Long Post-Training
Adding sequence parallelism into LLaMA-Factory, we open-sourced 360-LLaMA-Factory at https://github.com/Qihoo360/360-LLaMA-Factory. 360-LLaMA-Factory has received wide recognition and used in models such as Light-R1 arXiv:2503.10460, TinyR1 arXiv:2503.04872, Kaggle AIMO math models and also in large companies' training frameworks. This technical report delves deeper into the different sequence parallel modes behind 360-LLaMA-Factory and discusses our implementation insights.
comment: v2: sp for vlms; code at https://github.com/Qihoo360/360-LLaMA-Factory
♻ ☆ Data-Driven Adaptive PID Control Based on Physics-Informed Neural Networks
This article proposes a data-driven PID controller design based on the principle of adaptive gain optimization, leveraging Physics-Informed Neural Networks (PINNs) generated for predictive modeling purposes. The proposed control design method utilizes gradients of the PID gain optimization, achieved through the automatic differentiation of PINNs, to apply model predictive control using a cost function based on tracking error and control inputs. By optimizing PINNs-based PID gains, the method achieves adaptive gain tuning that ensures stability while accounting for system nonlinearities. The proposed method features a systematic framework for integrating PINNs-based models of dynamical control systems into closed-loop control systems, enabling direct application to PID control design. A series of numerical experiments is conducted to demonstrate the effectiveness of the proposed method from the control perspectives based on both time and frequency domains.
comment: This work has been submitted to the IEEE Transactions on Control Systems Technology for possible publication
♻ ☆ AdaDim: Dimensionality Adaptation for SSL Representational Dynamics
A key factor in effective Self-Supervised learning (SSL) is preventing dimensional collapse, where higher-dimensional representation spaces ($R$) span a lower-dimensional subspace. Therefore, SSL optimization strategies involve guiding a model to produce $R$ with a higher dimensionality ($H(R)$) through objectives that encourage decorrelation of features or sample uniformity in $R$. A higher $H(R)$ indicates that $R$ has greater feature diversity which is useful for generalization to downstream tasks. Alongside dimensionality optimization, SSL algorithms also utilize a projection head that maps $R$ into an embedding space $Z$. Recent work has characterized the projection head as a filter of noisy or irrelevant features from the SSL objective by reducing the mutual information $I(R;Z)$. Therefore, the current literature's view is that a good SSL representation space should have a high $H(R)$ and a low $I(R;Z)$. However, this view of SSL is lacking in terms of an understanding of the underlying training dynamics that influences the relationship between both terms. Our analysis shows that the best performing SSL models do not have the highest $H(R)$ nor the lowest $I(R;Z)$, but effectively arrive at a balance between both. To take advantage of this analysis, we introduce AdaDim, a training strategy that leverages SSL training dynamics by adaptively balancing between increasing $H(R)$ through feature decorrelation and sample uniformity as well as gradual regularization of $I(R;Z)$ as training progresses. We show performance improvements of up to 3% over common SSL baselines despite our method not utilizing expensive techniques such as queues, clustering, predictor networks, or student-teacher architectures.
comment: Under Review
♻ ☆ Enhancing Generative Auto-bidding with Offline Reward Evaluation and Policy Search
Auto-bidding serves as a critical tool for advertisers to improve their advertising performance. Recent progress has demonstrated that AI-Generated Bidding (AIGB), which learns a conditional generative planner from offline data, achieves superior performance compared to typical offline reinforcement learning (RL)-based auto-bidding methods. However, existing AIGB methods still face a performance bottleneck due to their inherent inability to explore beyond the static offline dataset. To address this, we propose {AIGB-Pearl} (\emph{{P}lanning with {E}valu{A}tor via RL}), a novel method that integrates generative planning and policy optimization. The core of AIGB-Pearl lies in constructing a trajectory evaluator for scoring generation quality and designing a provably sound KL-Lipschitz-constrained score maximization scheme to ensure safe and efficient exploration beyond the offline dataset. A practical algorithm incorporating the synchronous coupling technique is further devised to ensure the model regularity required by the proposed scheme. Extensive experiments on both simulated and real-world advertising systems demonstrate the state-of-the-art performance of our approach.
♻ ☆ Grounding the Ungrounded: A Spectral-Graph Framework for Quantifying Hallucinations in Multimodal LLMs
Hallucinations in LLMs--especially in multimodal settings--undermine reliability. We present a rigorous, information-geometric framework in diffusion dynamics that quantifies hallucination in MLLMs: model outputs are embedded spectrally on multimodal graph Laplacians, and gaps to a truth manifold define a semantic-distortion metric. We derive Courant--Fischer bounds on a temperature-dependent hallucination energy and use RKHS eigenmodes to obtain modality-aware, interpretable measures that track evolution over prompts and time. This reframes hallucination as measurable and bounded, providing a principled basis for evaluation and mitigation.
comment: 29 pages, 3 figures, 1 table
♻ ☆ RespoDiff: Dual-Module Bottleneck Transformation for Responsible & Faithful T2I Generation NeurIPS 2025
The rapid advancement of diffusion models has enabled high-fidelity and semantically rich text-to-image generation; however, ensuring fairness and safety remains an open challenge. Existing methods typically improve fairness and safety at the expense of semantic fidelity and image quality. In this work, we propose RespoDiff, a novel framework for responsible text-to-image generation that incorporates a dual-module transformation on the intermediate bottleneck representations of diffusion models. Our approach introduces two distinct learnable modules: one focused on capturing and enforcing responsible concepts, such as fairness and safety, and the other dedicated to maintaining semantic alignment with neutral prompts. To facilitate the dual learning process, we introduce a novel score-matching objective that enables effective coordination between the modules. Our method outperforms state-of-the-art methods in responsible generation by ensuring semantic alignment while optimizing both objectives without compromising image fidelity. Our approach improves responsible and semantically coherent generation by 20% across diverse, unseen prompts. Moreover, it integrates seamlessly into large-scale models like SDXL, enhancing fairness and safety. Code will be released upon acceptance.
comment: Accepted at NeurIPS 2025
♻ ☆ Maximising the Utility of Validation Sets for Imbalanced Noisy-label Meta-learning
Meta-learning is an effective method to handle imbalanced and noisy-label learning, but it depends on a validation set containing randomly selected, manually labelled and balanced distributed samples. The random selection and manual labelling and balancing of this validation set is not only sub-optimal for meta-learning, but it also scales poorly with the number of classes. Hence, recent meta-learning papers have proposed ad-hoc heuristics to automatically build and label this validation set, but these heuristics are still sub-optimal for meta-learning. In this paper, we analyse the meta-learning algorithm and propose new criteria to characterise the utility of the validation set, based on: 1) the informativeness of the validation set; 2) the class distribution balance of the set; and 3) the correctness of the labels of the set. Furthermore, we propose a new imbalanced noisy-label meta-learning (INOLML) algorithm that automatically builds a validation set by maximising its utility using the criteria above. Our method shows significant improvements over previous meta-learning approaches and sets the new state-of-the-art on several benchmarks.
♻ ☆ Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models NeurIPS
Vision-language models (VLMs) trained on internet-scale data achieve remarkable zero-shot detection performance on common objects like car, truck, and pedestrian. However, state-of-the-art models still struggle to generalize to out-of-distribution classes, tasks and imaging modalities not typically found in their pre-training. Rather than simply re-training VLMs on more visual data, we argue that one should align VLMs to new concepts with annotation instructions containing a few visual examples and rich textual descriptions. To this end, we introduce Roboflow100-VL, a large-scale collection of 100 multi-modal object detection datasets with diverse concepts not commonly found in VLM pre-training. We evaluate state-of-the-art models on our benchmark in zero-shot, few-shot, semi-supervised, and fully-supervised settings, allowing for comparison across data regimes. Notably, we find that VLMs like GroundingDINO and Qwen2.5-VL achieve less than 2% zero-shot accuracy on challenging medical imaging datasets within Roboflow100-VL, demonstrating the need for few-shot concept alignment. Lastly, we discuss our recent CVPR 2025 Foundational FSOD competition and share insights from the community. Notably, the winning team significantly outperforms our baseline by 17 mAP! Our code and dataset are available at https://github.com/roboflow/rf100-vl and https://universe.roboflow.com/rf100-vl/.
comment: The first two authors contributed equally. This work has been accepted to the Neural Information Processing Systems (NeurIPS) 2025 Datasets & Benchmark Track. Project Page: https://rf100-vl.org/
♻ ☆ Error Bounds for Physics-Informed Neural Networks in Fokker-Planck PDEs UAI
Stochastic differential equations are commonly used to describe the evolution of stochastic processes. The state uncertainty of such processes is best represented by the probability density function (PDF), whose evolution is governed by the Fokker-Planck partial differential equation (FP-PDE). However, it is generally infeasible to solve the FP-PDE in closed form. In this work, we show that physics-informed neural networks (PINNs) can be trained to approximate the solution PDF. Our main contribution is the analysis of PINN approximation error: we develop a theoretical framework to construct tight error bounds using PINNs. In addition, we derive a practical error bound that can be efficiently constructed with standard training methods. We discuss that this error-bound framework generalizes to approximate solutions of other linear PDEs. Empirical results on nonlinear, high-dimensional, and chaotic systems validate the correctness of our error bounds while demonstrating the scalability of PINNs and their significant computational speedup in obtaining accurate PDF solutions compared to the Monte Carlo approach.
comment: Accepted at Uncertainty in Artificial Intelligence (UAI) 2025
♻ ☆ DiffMI: Breaking Face Recognition Privacy via Diffusion-Driven Training-Free Model Inversion
Face recognition poses serious privacy risks due to its reliance on sensitive and immutable biometric data. While modern systems mitigate privacy risks by mapping facial images to embeddings (commonly regarded as privacy-preserving), model inversion attacks reveal that identity information can still be recovered, exposing critical vulnerabilities. However, existing attacks are often computationally expensive and lack generalization, especially those requiring target-specific training. Even training-free approaches suffer from limited identity controllability, hindering faithful reconstruction of nuanced or unseen identities. In this work, we propose DiffMI, the first diffusion-driven, training-free model inversion attack. DiffMI introduces a novel pipeline combining robust latent code initialization, a ranked adversarial refinement strategy, and a statistically grounded, confidence-aware optimization objective. DiffMI applies directly to unseen target identities and face recognition models, offering greater adaptability than training-dependent approaches while significantly reducing computational overhead. Our method achieves 84.42%--92.87% attack success rates against inversion-resilient systems and outperforms the best prior training-free GAN-based approach by 4.01%--9.82%. The implementation is available at https://github.com/azrealwang/DiffMI.
♻ ☆ Learning to Recover: Dynamic Reward Shaping with Wheel-Leg Coordination for Fallen Robots
Adaptive recovery from fall incidents are essential skills for the practical deployment of wheeled-legged robots, which uniquely combine the agility of legs with the speed of wheels for rapid recovery. However, traditional methods relying on preplanned recovery motions, simplified dynamics or sparse rewards often fail to produce robust recovery policies. This paper presents a learning-based framework integrating Episode-based Dynamic Reward Shaping and curriculum learning, which dynamically balances exploration of diverse recovery maneuvers with precise posture refinement. An asymmetric actor-critic architecture accelerates training by leveraging privileged information in simulation, while noise-injected observations enhance robustness against uncertainties. We further demonstrate that synergistic wheel-leg coordination reduces joint torque consumption by 15.8% and 26.2% and improves stabilization through energy transfer mechanisms. Extensive evaluations on two distinct quadruped platforms achieve recovery success rates up to 99.1% and 97.8% without platform-specific tuning. The supplementary material is available at https://boyuandeng.github.io/L2R-WheelLegCoordination/
♻ ☆ Interpretable Robot Control via Structured Behavior Trees and Large Language Models
As intelligent robots become more integrated into human environments, there is a growing need for intuitive and reliable Human-Robot Interaction (HRI) interfaces that are adaptable and more natural to interact with. Traditional robot control methods often require users to adapt to interfaces or memorize predefined commands, limiting usability in dynamic, unstructured environments. This paper presents a novel framework that bridges natural language understanding and robotic execution by combining Large Language Models (LLMs) with Behavior Trees. This integration enables robots to interpret natural language instructions given by users and translate them into executable actions by activating domain-specific plugins. The system supports scalable and modular integration, with a primary focus on perception-based functionalities, such as person tracking and hand gesture recognition. To evaluate the system, a series of real-world experiments was conducted across diverse environments. Experimental results demonstrate that the proposed approach is practical in real-world scenarios, with an average cognition-to-execution accuracy of approximately 94%, making a significant contribution to HRI systems and robots. The complete source code of the framework is publicly available at https://github.com/snt-arg/robot_suite.
comment: 15 pages, 5 figures, 3 tables
♻ ☆ GreedyPixel: Fine-Grained Black-Box Adversarial Attack Via Greedy Algorithm
Deep neural networks are highly vulnerable to adversarial examples that inputs with small, carefully crafted perturbations that cause misclassification, making adversarial attacks an essential tool for robustness evaluation. Existing black-box attacks fall into three categories: query-only, transfer-only, and query-and-transfer, and vary in perturbation pattern and optimization strategy. However, no prior method jointly achieves query-and-transfer guidance, pixel-wise sparsity, and training-free direct optimization, leaving a gap between black-box flexibility and white-box precision. We present GreedyPixel, a new attack framework that fills this gap by combining a surrogate-derived pixel priority map with greedy, per-pixel optimization refined by query feedback. This design reduces the exponential brute-force search space to a tractable linear procedure, guarantees monotonic loss decrease and convergence to a coordinate-wise optimum, and concentrates perturbations on robust, semantically meaningful pixels to improve perceptual quality. Extensive experiments on CIFAR-10 and ImageNet under both white-box and black-box settings demonstrate that GreedyPixel achieves state-of-the-art attack success rates and produces visually imperceptible perturbations. Our results show that GreedyPixel bridges the precision gap between white-box and black-box attacks and provides a practical framework for fine-grained robustness evaluation. The implementation is available at https://github.com/azrealwang/greedypixel.
♻ ☆ A Differentiable Alignment Framework for Sequence-to-Sequence Modeling via Optimal Transport
Accurate sequence-to-sequence (seq2seq) alignment is critical for applications like medical speech analysis and language learning tools relying on automatic speech recognition (ASR). State-of-the-art end-to-end (E2E) ASR systems, such as the Connectionist Temporal Classification (CTC) and transducer-based models, suffer from peaky behavior and alignment inaccuracies. In this paper, we propose a novel differentiable alignment framework based on one-dimensional optimal transport, enabling the model to learn a single alignment and perform ASR in an E2E manner. We introduce a pseudo-metric, called Sequence Optimal Transport Distance (SOTD), over the sequence space and discuss its theoretical properties. Based on the SOTD, we propose Optimal Temporal Transport Classification (OTTC) loss for ASR and contrast its behavior with CTC. Experimental results on the TIMIT, AMI, and LibriSpeech datasets show that our method considerably improves alignment performance compared to CTC and the more recently proposed Consistency-Regularized CTC, though with a trade-off in ASR performance. We believe this work opens new avenues for seq2seq alignment research, providing a solid foundation for further exploration and development within the community.
♻ ☆ Dynamic Learning Rate for Deep Reinforcement Learning: A Bandit Approach
In deep Reinforcement Learning (RL), the learning rate critically influences both stability and performance, yet its optimal value shifts during training as the environment and policy evolve. Standard decay schedulers assume monotonic convergence and often misalign with these dynamics, leading to premature or delayed adjustments. We introduce LRRL, a meta-learning approach that dynamically selects the learning rate based on policy performance rather than training steps. LRRL adaptively favors rates that improve returns, remaining robust even when the candidate set includes values that individually cause divergence. Across Atari and MuJoCo benchmarks, LRRL achieves performance competitive with or superior to tuned baselines and standard schedulers. Our findings position LRRL as a practical solution for adapting to non-stationary objectives in deep RL.
♻ ☆ Möbius transforms and Shapley values for vector-valued functions on weighted directed acyclic multigraphs
We generalize the concept of M\"obius inversion and Shapley values to directed acyclic multigraphs and weighted versions thereof. We further allow value functions (games) and thus their M\"obius transforms (synergy function) and Shapley values to have values in any abelian group that is a module over a ring that contains the graph weights, e.g. vector-valued functions. To achieve this and overcome the obstruction that the classical axioms (linearity, efficiency, null player, symmetry) are not strong enough to uniquely determine Shapley values in this more general setting, we analyze Shapley values from two novel points of view: 1) We introduce projection operators that allow us to interpret Shapley values as the recursive projection and re-attribution of higher-order synergies to lower-order ones; 2) we propose a strengthening of the null player axiom and a localized symmetry axiom, namely the weak elements and flat hierarchy axioms. The former allows us to remove coalitions with vanishing synergy while preserving the rest of the hierarchical structure. The latter treats player-coalition bonds uniformly in the corner case of hierarchically flat graphs. Together with linearity these axioms already imply a unique explicit formula for the Shapley values, as well as classical properties like efficiency, null player, symmetry, and novel ones like the projection property. This whole framework then specializes to finite inclusion algebras, lattices, partial orders and mereologies, and also recovers certain previously known cases as corner cases, and presents others from a new perspective. The admission of general weighted directed acyclic multigraph structured hierarchies and vector-valued functions and Shapley values opens up the possibility for new analytic tools and application areas, like machine learning, language processing, explainable artificial intelligence, and many more.
comment: 43 pages, 2 figures
♻ ☆ Contrastive Graph Condensation: Advancing Data Versatility through Self-Supervised Learning KDD 2025
With the increasing computation of training graph neural networks (GNNs) on large-scale graphs, graph condensation (GC) has emerged as a promising solution to synthesize a compact, substitute graph of the large-scale original graph for efficient GNN training. However, existing GC methods predominantly employ classification as the surrogate task for optimization, thus excessively relying on node labels and constraining their utility in label-sparsity scenarios. More critically, this surrogate task tends to overfit class-specific information within the condensed graph, consequently restricting the generalization capabilities of GC for other downstream tasks. To address these challenges, we introduce Contrastive Graph Condensation (CTGC), which adopts a self-supervised surrogate task to extract critical, causal information from the original graph and enhance the cross-task generalizability of the condensed graph. Specifically, CTGC employs a dual-branch framework to disentangle the generation of the node attributes and graph structures, where a dedicated structural branch is designed to explicitly encode geometric information through nodes' positional embeddings. By implementing an alternating optimization scheme with contrastive loss terms, CTGC promotes the mutual enhancement of both branches and facilitates high-quality graph generation through the model inversion technique. Extensive experiments demonstrate that CTGC excels in handling various downstream tasks with a limited number of labels, consistently outperforming state-of-the-art GC methods.
comment: KDD 2025
♻ ☆ Domain Generalization by Rejecting Extreme Augmentations WACV 2024
Data augmentation is one of the most effective techniques for regularizing deep learning models and improving their recognition performance in a variety of tasks and domains. However, this holds for standard in-domain settings, in which the training and test data follow the same distribution. For the out-of-domain case, where the test data follow a different and unknown distribution, the best recipe for data augmentation is unclear. In this paper, we show that for out-of-domain and domain generalization settings, data augmentation can provide a conspicuous and robust improvement in performance. To do that, we propose a simple training procedure: (i) use uniform sampling on standard data augmentation transformations; (ii) increase the strength transformations to account for the higher data variance expected when working out-of-domain, and (iii) devise a new reward function to reject extreme transformations that can harm the training. With this procedure, our data augmentation scheme achieves a level of accuracy that is comparable to or better than state-of-the-art methods on benchmark domain generalization datasets. Code: https://github.com/Masseeh/DCAug
comment: WACV 2024: Winter Conference on Applications of Computer Vision 2024
♻ ☆ Improving Neutral Point-of-View Generation with Data- and Parameter-Efficient RL
The paper shows that parameter-efficient reinforcement learning (PE-RL) is a highly effective training regime to improve large language models' (LLMs) ability to answer queries on sensitive topics with a Neutral Point of View (NPOV), i.e. to provide significantly more informative, diverse and impartial answers. This is shown by evaluating PE-RL and multiple strong baselines-including LoRA finetuning (strongest baseline), SFT and RLHF. PE-RL not only improves on overall NPOV quality compared to the strongest baseline ($97.06\%\rightarrow 99.08\%$), but also scores much higher on features linguists identify as key to separating sufficient answers from "great'' answers ($60.25\%\rightarrow 85.21\%$ for presence of supportive details, $68.74\%\rightarrow 91.43\%$ for absence of oversimplification). A qualitative analysis corroborates this. Moreover, our evaluation also finds a key property of PE-RL for this task: unlike methods that update all parameters, it generalises out of topic. Finally, to enable further studies we also release the dataset, SHQ-NPOV, and provide a methodology to create such datasets through iterative rounds of human peer-critique and annotator training.
♻ ☆ Real-Time Progress Prediction in Reasoning Language Models
Recent advances in reasoning language models -- particularly those that use long, latent chains of thought -- have demonstrated remarkable capabilities in complex, agentic tasks. However, as these models operate over increasingly extended time horizons, their internal progress becomes opaque to users, complicating expectation management and real-time oversight. In this work, we investigate whether real-time progress prediction is feasible. We discretize progress and train a linear probe to classify reasoning states. We then introduce a two-stage fine-tuning approach that enables reasoning models to generate progress estimates (0$\rightarrow$100\%) during inference. Our best fine-tuned model achieves an average error of 10\% for sequences less than 16,000 tokens, offering a practical mechanism for monitoring and interpreting model reasoning in real time.
♻ ☆ Train-Free Segmentation in MRI with Cubical Persistent Homology
We present a new general framework for segmentation of MRI scans based on Topological Data Analysis (TDA), offering several advantages over traditional machine learning approaches. The pipeline proceeds in three steps, first identifying the whole object to segment via automatic thresholding, then detecting a distinctive subset whose topology is known in advance, and finally deducing the various components of the segmentation. Unlike most prior TDA uses in medical image segmentation, which are typically embedded within deep networks, our approach is a standalone method tailored to MRI. A key ingredient is the localization of representative cycles from the persistence diagram, which enables interpretable mappings from topological features to anatomical components. In particular, the method offers the ability to perform segmentation without the need for large annotated datasets. Its modular design makes it adaptable to a wide range of data segmentation challenges. We validate the framework on three applications: glioblastoma segmentation in brain MRI, where a sphere is to be detected; myocardium in cardiac MRI, forming a cylinder; and cortical plate detection in fetal brain MRI, whose 2D slices are circles. We compare our method with established supervised and unsupervised baselines.
comment: Preprint, 36 pages, 18 figures, 4 tables. For associated code, see https://github.com/antonfrancois/gliomaSegmentation_TDA
♻ ☆ Automating RT Planning at Scale: High Quality Data For AI Training
Radiotherapy (RT) planning is complex, subjective, and time-intensive. Advances with artificial intelligence (AI) promise to improve its precision and efficiency, but progress is often limited by the scarcity of large, standardized datasets. To address this, we introduce the Automated Iterative RT Planning (AIRTP) system, a scalable solution for generating high-quality treatment plans. This scalable solution is designed to generate substantial volumes of consistently high-quality treatment plans, overcoming a key obstacle in the advancement of AI-driven RT planning. Our AIRTP pipeline adheres to clinical guidelines and automates essential steps, including organ-at-risk (OAR) contouring, helper structure creation, beam setup, optimization, and plan quality improvement, using AI integrated with RT planning software like Varian Eclipse. Furthermore, a novel approach for determining optimization parameters to reproduce 3D dose distributions, i.e. a method to convert dose predictions to deliverable treatment plans constrained by machine limitations is proposed. A comparative analysis of plan quality reveals that our automated pipeline produces treatment plans of quality comparable to those generated manually, which traditionally require several hours of labor per plan. Committed to public research, the first data release of our AIRTP pipeline includes nine cohorts covering head-and-neck and lung cancer sites to support an AAPM 2025 challenge. To our best knowledge, this dataset features more than 10 times number of plans compared to the largest existing well-curated public dataset. Repo: https://github.com/RiqiangGao/GDP-HMM_AAPMChallenge.
comment: radiotherapy planning, data for AI training
♻ ☆ Tempo: Compiled Dynamic Deep Learning with Symbolic Dependence Graphs
Deep learning (DL) algorithms are often defined in terms of temporal relationships: a tensor at one timestep may depend on tensors from earlier or later timesteps. Such dynamic dependencies (and corresponding dynamic tensor shapes) are difficult to express and optimize: while eager DL systems support such dynamism, they cannot apply compiler-based optimizations; graph-based systems require static tensor shapes, which forces users to pad tensors or break-up programs into multiple static graphs. We describe Tempo, a new DL system that combines the dynamism of eager execution with the whole-program optimizations of graph-based compilation. Tempo achieves this through a declarative programming model with recurrent tensors, which include explicit temporal dimensions. Temporal dimensions can be indexed using symbolic expressions to express dynamic dependencies on past and future tensors. Based on this, Tempo constructs a symbolic dependence graph, which concisely encodes dynamic dependencies between operators, and applies whole-program optimizations, such as algebraic simplifications, vectorization, tiling, and fusion. By tiling dynamic dependencies into static-size blocks, Tempo can also reuse existing static code-generators. It then uses a polyhedral model to find a feasible execution schedule, which includes memory management operations. We show that Tempo achieves a 7$\times$ speedup over JAX for Llama-3.2-3B decoding; for reinforcement learning algorithms, Tempo achieves a 54$\times$ speedup, with 16$\times$ lower peak memory usage.
comment: 17 pages, 24 figures, 3 bibliography pages
♻ ☆ Who Pays for Fairness? Rethinking Recourse under Social Burden
Machine learning based predictions are increasingly used in sensitive decision-making applications that directly affect our lives. This has led to extensive research into ensuring the fairness of classifiers. Beyond just fair classification, emerging legislation now mandates that when a classifier delivers a negative decision, it must also offer actionable steps an individual can take to reverse that outcome. This concept is known as algorithmic recourse. Nevertheless, many researchers have expressed concerns about the fairness guarantees within the recourse process itself. In this work, we provide a holistic theoretical characterization of unfairness in algorithmic recourse, formally linking fairness guarantees in recourse and classification, and highlighting limitations of the standard equal cost paradigm. We then introduce a novel fairness framework based on social burden, along with a practical algorithm (MISOB), broadly applicable under real-world conditions. Empirical results on real-world datasets show that MISOB reduces the social burden across all groups without compromising overall classifier accuracy.
♻ ☆ GenFacts-Generative Counterfactual Explanations for Multi-Variate Time Series
Counterfactual explanations aim to enhance model transparency by illustrating how input modifications can change model predictions. In the multivariate time series domain, existing approaches often produce counterfactuals that lack validity, plausibility, or intuitive interpretability. We present \textbf{GenFacts}, a novel generative framework for producing plausible and actionable counterfactual explanations for time series classifiers. GenFacts introduces a structured approach to latent space modeling and targeted counterfactual synthesis. We evaluate GenFacts on radar gesture recognition as an industrial use case and handwritten letter trajectories as an intuitive benchmark. Across both datasets, GenFacts consistently outperforms baseline methods in plausibility metrics (+18.7\%) and achieves the highest interpretability scores in user studies. These results underscore that realism and user-centered interpretability, rather than sparsity alone, are vital for actionable counterfactuals in time series applications.
comment: 4 pages
♻ ☆ NAR-*ICP: Neural Execution of Classical ICP-based Pointcloud Registration Algorithms
This study explores the intersection of neural networks and classical robotics algorithms through the Neural Algorithmic Reasoning (NAR) blueprint, enabling the training of neural networks to reason like classical robotics algorithms by learning to execute them. Algorithms are integral to robotics and safety-critical applications due to their predictable and consistent performance through logical and mathematical principles. In contrast, while neural networks are highly adaptable, handling complex, high-dimensional data and generalising across tasks, they often lack interpretability and transparency in their internal computations. To bridge the two, we propose a novel Graph Neural Network (GNN)-based framework, NAR-*ICP, that learns the intermediate computations of classical ICP-based registration algorithms, extending the CLRS Benchmark. We evaluate our approach across real-world and synthetic datasets, demonstrating its flexibility in handling complex inputs, and its potential to be used within larger learning pipelines. Our method achieves superior performance compared to the baselines, even surpassing the algorithms it was trained on, further demonstrating its ability to generalise beyond the capabilities of traditional algorithms.
comment: 19 pages, 16 tables, 7 figures
♻ ☆ Edit-Based Flow Matching for Temporal Point Processes
Temporal point processes (TPPs) are a fundamental tool for modeling event sequences in continuous time, but most existing approaches rely on autoregressive parameterizations that are limited by their sequential sampling. Recent non-autoregressive, diffusion-style models mitigate these issues by jointly interpolating between noise and data through event insertions and deletions in a discrete Markov chain. In this work, we generalize this perspective and introduce an Edit Flow process for TPPs that transports noise to data via insert, delete, and substitute edit operations. By learning the instantaneous edit rates within a continuous-time Markov chain framework, we attain a flexible and efficient model that effectively reduces the total number of necessary edit operations during generation. Empirical results demonstrate the generative flexibility of our unconditionally trained model in a wide range of unconditional and conditional generation tasks on benchmark TPPs.
♻ ☆ Inference-Time Scaling of Discrete Diffusion Models via Importance Weighting and Optimal Proposal Design
Discrete diffusion models have become highly effective across various domains. However, real-world applications often require the generative process to adhere to certain constraints. To this end, we propose a Sequential Monte Carlo (SMC) framework that enables scalable inference-time control of discrete diffusion models through principled importance weighting and optimal proposal construction. Specifically, our approach derives tractable importance weights for a range of intermediate targets and characterises the optimal proposal, for which we develop two practical approximations: a first-order gradient-based approximation and an amortised proposal trained to minimise the log-variance of the importance weights. Empirical results across synthetic tasks, language modelling, biology design, and text-to-image generation demonstrate that our framework enhances controllability and sample quality, highlighting the effectiveness of SMC as a versatile recipe for scaling discrete diffusion models at inference time.
♻ ☆ Flow Matching for Robust Simulation-Based Inference under Model Misspecification
Simulation-based inference (SBI) is transforming experimental sciences by enabling parameter estimation in complex non-linear models from simulated data. A persistent challenge, however, is model misspecification: simulators are only approximations of reality, and mismatches between simulated and real data can yield biased or overconfident posteriors. We address this issue by introducing Flow Matching Corrected Posterior Estimation (FMCPE), a framework that leverages the flow matching paradigm to refine simulation-trained posterior estimators using a small set of real calibration samples. Our approach proceeds in two stages: first, a posterior approximator is trained on abundant simulated data; second, flow matching transports its predictions toward the true posterior supported by real observations, without requiring explicit knowledge of the misspecification. This design enables FMCPE to combine the scalability of SBI with robustness to distributional shift. Across synthetic benchmarks and real-world datasets, we show that our proposal consistently mitigates the effects of misspecification, delivering improved inference accuracy and uncertainty calibration compared to standard SBI baselines, while remaining computationally efficient.
♻ ☆ Taxonomy, Opportunities, and Challenges of Representation Engineering for Large Language Models
Representation Engineering (RepE) is a novel paradigm for controlling the behavior of LLMs. Unlike traditional approaches that modify inputs or fine-tune the model, RepE directly manipulates the model's internal representations. As a result, it may offer more effective, interpretable, data-efficient, and flexible control over models' behavior. We present the first comprehensive survey of RepE for LLMs, reviewing the rapidly growing literature to address key questions: What RepE methods exist and how do they differ? For what concepts and problems has RepE been applied? What are the strengths and weaknesses of RepE compared to other methods? To answer these, we propose a unified framework describing RepE as a pipeline comprising representation identification, operationalization, and control. We posit that while RepE methods offer significant potential, challenges remain, including managing multiple concepts, ensuring reliability, and preserving models' performance. Towards improving RepE, we identify opportunities for experimental and methodological improvements and construct a guide for best practices.
♻ ☆ P3D: Scalable Neural Surrogates for High-Resolution 3D Physics Simulations with Global Context
We present a scalable framework for learning deterministic and probabilistic neural surrogates for high-resolution 3D physics simulations. We introduce a hybrid CNN-Transformer backbone architecture targeted for 3D physics simulations, which significantly outperforms existing architectures in terms of speed and accuracy. Our proposed network can be pretrained on small patches of the simulation domain, which can be fused to obtain a global solution, optionally guided via a fast and scalable sequence-to-sequence model to include long-range dependencies. This setup allows for training large-scale models with reduced memory and compute requirements for high-resolution datasets. We evaluate our backbone architecture against a large set of baseline methods with the objective to simultaneously learn the dynamics of 14 different types of PDEs in 3D. We demonstrate how to scale our model to high-resolution isotropic turbulence with spatial resolutions of up to $512^3$. Finally, we demonstrate the versatility of our network by training it as a diffusion model to produce probabilistic samples of highly turbulent 3D channel flows across varying Reynolds numbers, accurately capturing the underlying flow statistics.
♻ ☆ GRPO is Secretly a Process Reward Model ICLR 2026
We prove theoretically that the GRPO RL algorithm induces a non-trivial process reward model (PRM), under certain assumptions regarding within-group overlap of token sequences across completions. We then show empirically that these assumptions are met under real-world conditions: GRPO does in fact induce a non-trivial PRM. Leveraging the framework of GRPO-as-a-PRM, we identify a flaw in the GRPO objective: non-uniformly distributed process steps hinder both exploration and exploitation (under different conditions). We propose a simple modification to the algorithm to mitigate this defect ($\lambda$-GRPO), and show that LLMs trained with $\lambda$-GRPO achieve higher validation accuracy and performance on downstream reasoning tasks$-$and reach peak performance more rapidly$-$than LLMs trained with standard GRPO. Our results call into question the advantage of costly, explicitly-defined PRMs for GRPO: we show that it is possible to instead leverage the hidden, built-in PRM structure within the vanilla GRPO algorithm to boost model performance with a negligible impact on training time and cost.
comment: 14 pages, 6 figures; under review at ICLR 2026
♻ ☆ When Judgment Becomes Noise: How Design Failures in LLM Judge Benchmarks Silently Undermine Validity
LLM-judged benchmarks are increasingly used to evaluate complex model behaviors, yet their design introduces failure modes absent in conventional ground-truth based benchmarks. We argue that without tight objectives and verifiable constructions, benchmark rankings can produce high-confidence rankings that are in fact largely noise. We introduce two mechanisms to diagnose these issues. Schematic adherence quantifies how much of a judge's overall verdict is explained by the explicit evaluation schema, revealing unexplained variance when judges deviate from their own rubric. Psychometric validity aggregates internal consistency and discriminant validity signals to quantify irreducible uncertainty in any benchmarking run. Applying these tools to Arena-Hard Auto, we find severe schema incoherence and factor collapse across popular judges: for example, unexplained variance exceeding 90 percent for DeepSeek-R1-32B and factor correlations above 0.93 for most criteria. We also show that the ELO-style aggregation used by Arena-Hard Auto collapses and masks genuine ranking uncertainty. Our results highlight design failures that undermine validity and offer actionable principles for building better-scoped, reliability-aware LLM-judged benchmarks. We released our code and dataset at https://github.com/penfever/judgment-to-noise
♻ ☆ Quantum Rationale-Aware Graph Contrastive Learning for Jet Discrimination
In high-energy physics, particle jet tagging plays a pivotal role in distinguishing quark from gluon jets using data from collider experiments. While graph-based deep learning methods have advanced this task beyond traditional feature-engineered approaches, the complex data structure and limited labeled samples present ongoing challenges. However, existing contrastive learning (CL) frameworks struggle to leverage rationale-aware augmentations effectively, often lacking supervision signals that guide the extraction of salient features and facing computational efficiency issues such as high parameter counts. In this study, we demonstrate that integrating a quantum rationale generator (QRG) within our proposed Quantum Rationale-aware Graph Contrastive Learning (QRGCL) framework significantly enhances jet discrimination performance, reducing reliance on labeled data and capturing discriminative features. Evaluated on the quark-gluon jet dataset, QRGCL achieves an AUC score of $77.53\%$ while maintaining a compact architecture of only 45 QRG parameters, outperforming classical, quantum, and hybrid GCL and GNN benchmarks. These results highlight QRGCL's potential to advance jet tagging and other complex classification tasks in high-energy physics, where computational efficiency and feature extraction limitations persist.
♻ ☆ FedAGHN: Personalized Federated Learning with Attentive Graph HyperNetworks
Personalized Federated Learning (PFL) aims to address the statistical heterogeneity of data across clients by learning the personalized model for each client. Among various PFL approaches, the personalized aggregation-based approach conducts parameter aggregation in the server-side aggregation phase to generate personalized models, and focuses on learning appropriate collaborative relationships among clients for aggregation. However, the collaborative relationships vary in different scenarios and even at different stages of the FL process. To this end, we propose Personalized Federated Learning with Attentive Graph HyperNetworks (FedAGHN), which employs Attentive Graph HyperNetworks (AGHNs) to dynamically capture fine-grained collaborative relationships and generate client-specific personalized initial models. Specifically, AGHNs empower graphs to explicitly model the client-specific collaborative relationships, construct collaboration graphs, and introduce tunable attentive mechanism to derive the collaboration weights, so that the personalized initial models can be obtained by aggregating parameters over the collaboration graphs. Extensive experiments can demonstrate the superiority of FedAGHN. Moreover, a series of visualizations are presented to explore the effectiveness of collaboration graphs learned by FedAGHN.
comment: Final accepted manuscript of the article published in Knowledge-Based Systems, 2025
♻ ☆ The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives
The objectives that Large Language Models (LLMs) implicitly optimize remain dangerously opaque, making trustworthy alignment and auditing a grand challenge. While Inverse Reinforcement Learning (IRL) can infer reward functions from behaviour, existing approaches either produce a single, overconfident reward estimate or fail to address the fundamental ambiguity of the task (non-identifiability). This paper introduces a principled auditing framework that re-frames reward inference from a simple estimation task to a comprehensive process for verification. Our framework leverages Bayesian IRL to not only recover a distribution over objectives but to enable three critical audit capabilities: (i) Quantifying and systematically reducing non-identifiability by demonstrating posterior contraction over sequential rounds of evidence; (ii) Providing actionable, uncertainty-aware diagnostics that expose spurious shortcuts and identify out-of-distribution prompts where the inferred objective cannot be trusted; and (iii) Validating policy-level utility by showing that the refined, low-uncertainty reward can be used directly in RLHF to achieve training dynamics and toxicity reductions comparable to the ground-truth alignment process. Empirically, our framework successfully audits a detoxified LLM, yielding a well-calibrated and interpretable objective that strengthens alignment guarantees. Overall, this work provides a practical toolkit for auditors, safety teams, and regulators to verify what LLMs are truly trying to achieve, moving us toward more trustworthy and accountable AI.
comment: Preprint
♻ ☆ Unlocking Dataset Distillation with Diffusion Models
Dataset distillation seeks to condense datasets into smaller but highly representative synthetic samples. While diffusion models now lead all generative benchmarks, current distillation methods avoid them and rely instead on GANs or autoencoders, or, at best, sampling from a fixed diffusion prior. This trend arises because naive backpropagation through the long denoising chain leads to vanishing gradients, which prevents effective synthetic sample optimization. To address this limitation, we introduce Latent Dataset Distillation with Diffusion Models (LD3M), the first method to learn gradient-based distilled latents and class embeddings end-to-end through a pre-trained latent diffusion model. A linearly decaying skip connection, injected from the initial noisy state into every reverse step, preserves the gradient signal across dozens of timesteps without requiring diffusion weight fine-tuning. Across multiple ImageNet subsets at 128x128 and 256x256, LD3M improves downstream accuracy by up to 4.8 percentage points (1 IPC) and 4.2 points (10 IPC) over the prior state-of-the-art. The code for LD3M is provided at https://github.com/Brian-Moser/prune_and_distill.
♻ ☆ ExLLM: Experience-Enhanced LLM Optimization for Molecular Design and Beyond
Molecular design involves an enormous and irregular search space, where traditional optimizers such as Bayesian optimization, genetic algorithms, and generative models struggle to leverage expert knowledge or handle complex feedback. Recently, LLMs have been used as optimizers, achieving promising results on benchmarks such as PMO. However, existing approaches rely only on prompting or extra training, without mechanisms to handle complex feedback or maintain scalable memory. In particular, the common practice of appending or summarizing experiences at every query leads to redundancy, degraded exploration, and ultimately poor final outcomes under large-scale iterative search. We introduce ExLLM (Experience-Enhanced LLM optimization), an LLM-as-optimizer framework with three components: (1) a compact, evolving experience snippet tailored to large discrete spaces that distills non-redundant cues and improves convergence at low cost; (2) a simple yet effective k-offspring scheme that widens exploration per call and reduces orchestration cost; and (3) a lightweight feedback adapter that normalizes objectives for selection while formatting constraints and expert hints for iteration. ExLLM sets new state-of-the-art results on PMO and generalizes strongly in our setup, it sets records on circle packing and stellarator design, and yields consistent gains across additional domains requiring only a task-description template and evaluation functions to transfer.
comment: 10 pages, under review
♻ ☆ Estimating the Joint Probability of Scenario Parameters with Gaussian Mixture Copula Models
This paper presents the first application of Gaussian Mixture Copula Models to the statistical modeling of driving scenarios for the safety validation of automated driving systems. Knowledge of the joint probability distribution of scenario parameters is essential for scenario-based safety assessment, where risk quantification depends on the likelihood of concrete parameter combinations. Gaussian Mixture Copula Models bring together the multimodal expressivity of Gaussian Mixture Models and the flexibility of copulas, enabling separate modeling of marginal distributions and dependencies. We benchmark Gaussian Mixture Copula Models against previously proposed approaches - Gaussian Mixture Models and Gaussian Copula Models - using real-world driving data drawn from scenarios defined in United Nations Regulation No. 157. Our evaluation across approximately 18 million scenario instances demonstrates that Gaussian Mixture Copula Models consistently surpass Gaussian Copula Models and perform better than, or at least comparably to, Gaussian Mixture Models, as measured by both log-likelihood and Sinkhorn distance. These results are promising for the adoption of Gaussian Mixture Copula Models as a statistical foundation for future scenario-based validation frameworks.
comment: 8 pages, 4 figures; This work has been submitted to the IEEE for possible publication
♻ ☆ VICON: Vision In-Context Operator Networks for Multi-Physics Fluid Dynamics Prediction NIPS
In-Context Operator Networks (ICONs) have demonstrated the ability to learn operators across diverse partial differential equations using few-shot, in-context learning. However, existing ICONs process each spatial point as an individual token, severely limiting computational efficiency when handling dense data in higher spatial dimensions. We propose Vision In-Context Operator Networks (VICON), which integrates vision transformer architectures to efficiently process 2D data through patch-wise operations while preserving ICON's adaptability to multiphysics systems and varying timesteps. Evaluated across three fluid dynamics benchmarks, VICON significantly outperforms state-of-the-art baselines: DPOT and MPP, reducing the averaged last-step rollout error by 37.9% compared to DPOT and 44.7% compared to MPP, while requiring only 72.5% and 34.8% of their respective inference times. VICON naturally supports flexible rollout strategies with varying timestep strides, enabling immediate deployment in imperfect measurement systems where sampling frequencies may differ or frames might be dropped - common challenges in real-world settings - without requiring retraining or interpolation. In these realistic scenarios, VICON exhibits remarkable robustness, experiencing only 24.41% relative performance degradation compared to 71.37%-74.49% degradation in baseline methods, demonstrating its versatility for deploying in realistic applications. Our scripts for processing datasets and code are publicly available at https://github.com/Eydcao/VICON.
comment: update after NIPS suggestions
Multimedia 6
☆ AV-EMO-Reasoning: Benchmarking Emotional Reasoning Capabilities in Omni-modal LLMS with Audio-visual Cues
Emotions conveyed through voice and face shape engagement and context in human-AI interaction. Despite rapid progress in omni-modal large language models (LLMs), the holistic evaluation of emotional reasoning with audiovisual cues remains limited. To address this gap, we introduce AV-EMO-Reasoning, a benchmark designed to systematically assess emotional coherence in LLMs. The framework leverages a curated, single- and multi-turn synthetic audiovisual corpus with a real-world set and is assessed under continuous, categorical, and perceptual metrics. Experiments with leading LLMs show that visual cues reliably improve emotional coherence over audio-only baselines. Moreover, LLMs can leverage audio-visual cues to generate more emotion-aware speech. Models exhibit complementary strengths across metric families, indicating that automatic scores capture facets distinct from perceptual judgments. By releasing a systematic evaluation benchmark, AV-EMO-Reasoning offers a reproducible standard for evaluating emotion-aware dialogue and advances toward more natural, adaptive human-AI interaction.
♻ ☆ LaunchpadGPT: Language Model as Music Visualization Designer on Launchpad
Launchpad is a musical instrument that allows users to create and perform music by pressing illuminated buttons. To assist and inspire the design of the Launchpad light effect, and provide a more accessible approach for beginners to create music visualization with this instrument, we proposed the LaunchpadGPT model to generate music visualization designs on Launchpad automatically. Based on the language model with excellent generation ability, our proposed LaunchpadGPT takes an audio piece of music as input and outputs the lighting effects of Launchpad-playing in the form of a video (Launchpad-playing video). We collect Launchpad-playing videos and process them to obtain music and corresponding video frame of Launchpad-playing as prompt-completion pairs, to train the language model. The experiment result shows the proposed method can create better music visualization than random generation methods and hold the potential for a broader range of music visualization applications. Our code is available at https://github.com/yunlong10/LaunchpadGPT/.
comment: Accepted to International Computer Music Conference (ICMC) 2023
♻ ☆ Multi-modal Segment Assemblage Network for Ad Video Editing with Importance-Coherence Reward ACCV 2022
Advertisement video editing aims to automatically edit advertising videos into shorter videos while retaining coherent content and crucial information conveyed by advertisers. It mainly contains two stages: video segmentation and segment assemblage. The existing method performs well at video segmentation stages but suffers from the problems of dependencies on extra cumbersome models and poor performance at the segment assemblage stage. To address these problems, we propose M-SAN (Multi-modal Segment Assemblage Network) which can perform efficient and coherent segment assemblage task end-to-end. It utilizes multi-modal representation extracted from the segments and follows the Encoder-Decoder Ptr-Net framework with the Attention mechanism. Importance-coherence reward is designed for training M-SAN. We experiment on the Ads-1k dataset with 1000+ videos under rich ad scenarios collected from advertisers. To evaluate the methods, we propose a unified metric, Imp-Coh@Time, which comprehensively assesses the importance, coherence, and duration of the outputs at the same time. Experimental results show that our method achieves better performance than random selection and the previous method on the metric. Ablation experiments further verify that multi-modal representation and importance-coherence reward significantly improve the performance. Ads-1k dataset is available at: https://github.com/yunlong10/Ads-1k
comment: Accepted by ACCV 2022
♻ ☆ TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling NeurIPS
While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited, leaving other simpler yet crucial components such as metadata or attribute filtering underutilized in the system. We propose an LLM-based music recommendation system with tool calling to serve as a unified retrieval-reranking pipeline. Our system positions an LLM as an end-to-end recommendation system that interprets user intent, plans tool invocations, and orchestrates specialized components: boolean filters (SQL), sparse retrieval (BM25), dense retrieval (embedding similarity), and generative retrieval (semantic IDs). Through tool planning, the system predicts which types of tools to use, their execution order, and the arguments needed to find music matching user preferences, supporting diverse modalities while seamlessly integrating multiple database filtering methods. We demonstrate that this unified tool-calling framework achieves competitive performance across diverse recommendation scenarios by selectively employing appropriate retrieval methods based on user queries, envisioning a new paradigm for conversational music recommendation systems.
comment: Accepted for publication at The Workshop on AI for Music, Neural Information Processing Systems (NeurIPS-AI4Music)
♻ ☆ TalkPlayData 2: An Agentic Synthetic Data Pipeline for Multimodal Conversational Music Recommendation
We present TalkPlayData 2, a synthetic dataset for multimodal conversational music recommendation generated by an agentic data pipeline. In the proposed pipeline, multiple large language model (LLM) agents are created under various roles with specialized prompts and access to different parts of information, and the chat data is acquired by logging the conversation between the Listener LLM and the Recsys LLM. To cover various conversation scenarios, for each conversation, the Listener LLM is conditioned on a finetuned conversation goal. Finally, all the LLMs are multimodal with audio and images, allowing a simulation of multimodal recommendation and conversation. In the LLM-as-a-judge and subjective evaluation experiments, TalkPlayData 2 achieved the proposed goal in various aspects related to training a generative recommendation model for music. TalkPlayData 2 and its generation code are released at https://talkpl.ai/talkplaydata2.
comment: 2025-10-08: updating the stat table with the latest numbers. updated the abstract per the latest license terms
♻ ☆ What Media Frames Reveal About Stance: A Dataset and Study about Memes in Climate Change Discourse EMNLP
Media framing refers to the emphasis on specific aspects of perceived reality to shape how an issue is defined and understood. Its primary purpose is to shape public perceptions often in alignment with the authors' opinions and stances. However, the interaction between stance and media frame remains largely unexplored. In this work, we apply an interdisciplinary approach to conceptualize and computationally explore this interaction with internet memes on climate change. We curate CLIMATEMEMES, the first dataset of climate-change memes annotated with both stance and media frames, inspired by research in communication science. CLIMATEMEMES includes 1,184 memes sourced from 47 subreddits, enabling analysis of frame prominence over time and communities, and sheds light on the framing preferences of different stance holders. We propose two meme understanding tasks: stance detection and media frame detection. We evaluate LLaVA-NeXT and Molmo in various setups, and report the corresponding results on their LLM backbone. Human captions consistently enhance performance. Synthetic captions and human-corrected OCR also help occasionally. Our findings highlight that VLMs perform well on stance, but struggle on frames, where LLMs outperform VLMs. Finally, we analyze VLMs' limitations in handling nuanced frames and stance expressions on climate change internet memes.
comment: 20 pages, 9 figures, EMNLP Findings
Computation and Language 150
☆ TaTToo: Tool-Grounded Thinking PRM for Test-Time Scaling in Tabular Reasoning
Process Reward Models (PRMs) have recently emerged as a powerful framework for enhancing the reasoning capabilities of large reasoning models (LRMs), particularly in the context of test-time scaling (TTS). However, their potential for supervising LRMs on tabular reasoning domains remains underexplored. Through detailed empirical analyses, we identify that existing PRMs, though widely adopted for supervising text-only reasoning steps, struggle with table-specific operations such as sub-table retrieval and schema interaction, leading to critical performance bottlenecks. To address this limitation, we propose TaTToo, a novel table-grounded PRM framework that (i) reasons explicitly over tabular reasoning steps and (ii) integrates tool-based verification to provide precise reward supervision. Concretely, we first design a scalable data curation pipeline that constructs over 60k high-quality step-level annotations by integrating table verification rationales with tool-based executions. Building on the collected data, we train TaTToo with a dual-stage paradigm: cold-start supervised fine-tuning to capture tool-use reasoning patterns, followed by reinforcement learning with tool-grounded reward shaping to align our model with table-based verification. We provide a comprehensive evaluation of the policy improvement induced by our newly designed PRM. Across 5 challenging tabular reasoning benchmarks covering numerical reasoning, fact-checking, and data analysis, TaTToo improves downstream policy LRMs by 30.9% at inference, surpasses strong PRM baselines such as Qwen-2.5-Math-PRM-72B with only 8B parameters, and demonstrates strong generalizability across diverse TTS strategies.
☆ Stratified GRPO: Handling Structural Heterogeneity in Reinforcement Learning of LLM Search Agents
Large language model (LLM) agents increasingly rely on external tools such as search engines to solve complex, multi-step problems, and reinforcement learning (RL) has become a key paradigm for training them. However, the trajectories of search agents are structurally heterogeneous, where variations in the number, placement, and outcomes of search calls lead to fundamentally different answer directions and reward distributions. Standard policy gradient methods, which use a single global baseline, suffer from what we identify and formalize as cross-stratum bias-an "apples-to-oranges" comparison of heterogeneous trajectories. This cross-stratum bias distorts credit assignment and hinders exploration of complex, multi-step search strategies. To address this, we propose Stratified GRPO, whose central component, Stratified Advantage Normalization (SAN), partitions trajectories into homogeneous strata based on their structural properties and computes advantages locally within each stratum. This ensures that trajectories are evaluated only against their true peers. Our analysis proves that SAN eliminates cross-stratum bias, yields conditionally unbiased unit-variance estimates inside each stratum, and retains the global unbiasedness and unit-variance properties enjoyed by standard normalization, resulting in a more pure and scale-stable learning signal. To improve practical stability under finite-sample regimes, we further linearly blend SAN with the global estimator. Extensive experiments on diverse single-hop and multi-hop question-answering benchmarks demonstrate that Stratified GRPO consistently and substantially outperforms GRPO by up to 11.3 points, achieving higher training rewards, greater training stability, and more effective search policies. These results establish stratification as a principled remedy for structural heterogeneity in RL for LLM search agents.
☆ TokenChain: A Discrete Speech Chain via Semantic Token Modeling ICASSP
Machine Speech Chain, simulating the human perception-production loop, proves effective in jointly improving ASR and TTS. We propose TokenChain, a fully discrete speech chain coupling semantic-token ASR with a two-stage TTS: an autoregressive text-to-semantic model co-trained with ASR and a masked-generative semantic-to-acoustic model for synthesis only. End-to-end feedback across the text interface is enabled with straight-through argmax/Gumbel-Softmax and balanced with supervised ASR via dynamic weight averaging. Ablations examine optimal temperature schedules for in- and cross-domain transfer. Evaluation reveals TokenChain surpasses baseline accuracy 2-6 epochs earlier and yields 5-13% lower equal-epoch error with stable T2S on LibriSpeech, and reduces relative ASR WER by 56% and T2S WER by 31% on TED-LIUM with minimal forgetting, showing that chain learning remains effective with token interfaces and models.
comment: 5 pages, 3 figures. Submitted to IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2026
☆ Peeking inside the Black-Box: Reinforcement Learning for Explainable and Accurate Relation Extraction
This paper introduces a framework for relation extraction (RE) that enhances both accuracy and explainability. The framework has two key components: (i) a reasoning mechanism that formulates relation extraction as a series of text-processing steps inspired by cognitive science, and (ii) an optimization process driven by reinforcement learning (RL) with a novel reward function designed to improve both task accuracy and explanation quality. We call our approach CogRE. Our framework addresses the lack of supervision for language-based explanations in traditional RE by promoting outputs that include important relation keywords. These keywords are drawn from a high-quality dictionary that is automatically constructed using an LLM. We evaluate our approach for the task of one-shot RE using two LLMs and two RE datasets. Our experiments show that CogRE improves explanation quality by addressing two common failure patterns in one-shot RE: poor attention focus and limited one-shot learning capability. For example, our cognitive-structured reasoning with Qwen2.5-15B-Instruct on One-shot NYT29 achieves 24.65% F1, surpassing prior reasoning-based designs. Optimizing this approach with RL using our reward further improves performance by +23.46% (absolute). Finally, human evaluation shows that our best model generates relational keywords closely aligned with gold labels, increasing human explanation quality ratings by 54% (relative).
comment: Working in process
☆ Latent Speech-Text Transformer
Auto-regressive speech-text models are typically pre-trained on a large number of interleaved sequences of text tokens and raw speech encoded as speech tokens using vector quantization. These models have demonstrated state-of-the-art performance in speech-to-speech understanding and generation benchmarks, together with promising scaling laws, primarily enabled by the representational alignment between text and speech. Nevertheless, they suffer from shortcomings, partly owing to the disproportionately longer sequences of speech tokens in contrast to textual tokens. This results in a large compute imbalance between modalities during pre-training as well as during inference, and a potential hindrance to effectively aligning speech and text, ultimately translating to several orders of magnitude slower scaling laws. We introduce the Latent Speech-Text Transformer (LST), which makes pre-training speech-text models more data-efficient by dynamically and inexpensively aggregating speech tokens into latent speech patches. These patches serve as higher-level units that can either align with corresponding textual units to aid capability transfer or even encapsulate common speech sequences like silences to be more compute-efficient. We show that LST outperforms vanilla approaches on speech-to-speech as well as text-to-text benchmarks in both data- and compute-controlled settings, the former indicating more effective representational alignment and the latter indicating steeper scaling laws for speech-text models. On HellaSwag story completion, LST achieves 6.5% absolute gain in speech accuracy under compute-controlled training and 5.3% under data-controlled training, while also improving text performance. We will release our models, code, and the evaluation data to facilitate further research.
comment: 16 pages, 13 figures
☆ BanglaTalk: Towards Real-Time Speech Assistance for Bengali Regional Dialects
Real-time speech assistants are becoming increasingly popular for ensuring improved accessibility to information. Bengali, being a low-resource language with a high regional dialectal diversity, has seen limited progress in developing such systems. Existing systems are not optimized for real-time use and focus only on standard Bengali. In this work, we present BanglaTalk, the first real-time speech assistance system for Bengali regional dialects. BanglaTalk follows the client-server architecture and uses the Real-time Transport Protocol (RTP) to ensure low-latency communication. To address dialectal variation, we introduce a dialect-aware ASR system, BRDialect, developed by fine-tuning the IndicWav2Vec model in ten Bengali regional dialects. It outperforms the baseline ASR models by 12.41-33.98% on the RegSpeech12 dataset. Furthermore, BanglaTalk can operate at a low bandwidth of 24 kbps while maintaining an average end-to-end delay of 4.9 seconds. Low bandwidth usage and minimal end-to-end delay make the system both cost-effective and interactive for real-time use cases, enabling inclusive and accessible speech technology for the diverse community of Bengali speakers.
☆ RECODE-H: A Benchmark for Research Code Development with Interactive Human Feedback
Large language models (LLMs) show the promise in supporting scientific research implementation, yet their ability to generate correct and executable code remains limited. Existing works largely adopt one-shot settings, ignoring the iterative and feedback-driven nature of realistic workflows of scientific research development. To address this gap, we present RECODE-H, a benchmark of 102 tasks from research papers and repositories that evaluates LLM agents through multi-turn interactions with LLM-simulated human feedback. It includes structured instructions,unit tests, and a five-level feedback hierarchy to reflect realistic researcher-agent collaboration. We further present ReCodeAgent, a framework that integrates feedback into iterative code generation. Experiments with leading LLMs, including GPT-5, Claude-Sonnet-4, DeepSeek-V3.1, and Gemini 2.5, show substantial performance gains with richer feedback, while also highlighting ongoing challenges in the generation of complex research code. RECODE-H establishes a foundation for developing adaptive, feedback-driven LLM agents in scientific research implementation
comment: Code and dataset are available at github.com/ChunyuMiao98/RECODE
☆ Mixing Mechanisms: How Language Models Retrieve Bound Entities In-Context
A key component of in-context reasoning is the ability of language models (LMs) to bind entities for later retrieval. For example, an LM might represent "Ann loves pie" by binding "Ann" to "pie", allowing it to later retrieve "Ann" when asked "Who loves pie?" Prior research on short lists of bound entities found strong evidence that LMs implement such retrieval via a positional mechanism, where "Ann" is retrieved based on its position in context. In this work, we find that this mechanism generalizes poorly to more complex settings; as the number of bound entities in context increases, the positional mechanism becomes noisy and unreliable in middle positions. To compensate for this, we find that LMs supplement the positional mechanism with a lexical mechanism (retrieving "Ann" using its bound counterpart "pie") and a reflexive mechanism (retrieving "Ann" through a direct pointer). Through extensive experiments on nine models and ten binding tasks, we uncover a consistent pattern in how LMs mix these mechanisms to drive model behavior. We leverage these insights to develop a causal model combining all three mechanisms that estimates next token distributions with 95% agreement. Finally, we show that our model generalizes to substantially longer inputs of open-ended text interleaved with entity groups, further demonstrating the robustness of our findings in more natural settings. Overall, our study establishes a more complete picture of how LMs bind and retrieve entities in-context.
☆ VecInfer: Efficient LLM Inference with Low-Bit KV Cache via Outlier-Suppressed Vector Quantization
The Key-Value (KV) cache introduces substantial memory overhead during large language model (LLM) inference. Although existing vector quantization (VQ) methods reduce KV cache usage and provide flexible representational capacity across bit-widths, they suffer severe performance degradation at ultra-low bit-widths due to key cache outliers that hinder effective codebook utilization. To address this challenge, we propose VecInfer, a novel VQ method for aggressive KV cache compression while enabling efficient inference. By applying smooth and Hadamard transformations, VecInfer suppresses outliers in the key cache, enabling the codebook to comprehensively cover the original data distribution and thereby reducing quantization difficulty. To facilitate efficient deployment, we design an optimized CUDA kernel that fuses computation with dequantization to minimize memory access overhead. Extensive evaluations demonstrate that VecInfer consistently outperforms existing quantization baselines across both long-context understanding and mathematical reasoning tasks. With only 2-bit quantization, VecInfer achieves performance comparable to full precision, while delivering up to $\mathbf{2.7\times}$ speedup in large-batch self-attention computation and $\mathbf{8.3\times}$ reduction in single-batch end-to-end latency on Llama-3.1-8B with a 196k sequence length.
☆ RoSE: Round-robin Synthetic Data Evaluation for Selecting LLM Generators without Human Test Sets
LLMs are powerful generators of synthetic data, which are used for training smaller, specific models. This is especially valuable for low-resource languages, where human-labelled data is scarce but LLMs can still produce high-quality text. However, LLMs differ in how useful their outputs are for training. Selecting the best LLM as a generator is challenging because extrinsic evaluation requires costly human annotations (which are often unavailable for low-resource languages), while intrinsic metrics correlate poorly with downstream performance. We introduce Round robin Synthetic data Evaluation (RoSE), a proxy metric for selecting the best LLM generator without human test sets. RoSE trains a small model on the outputs of a candidate generator (LLM) and then evaluates it on generated synthetic examples from all other candidate LLMs. The final RoSE score is the mean performance of this small model. Across six LLMs, eleven languages, and three tasks (sentiment, topic, intent), RoSE identifies the optimal generator more often than any other intrinsic heuristics. RoSE outperforms intrinsic heuristics and comes within 0.76 percentage points of the optimal generator baseline. This result is measured in terms of downstream performance, obtained by training a small model on the chosen generator's outputs (optimal vs. proxy metric selected) and evaluating it on human-labelled test data. Additionally, RoSE is the only metric to achieve a positive correlation with performance on human test data.
comment: 16 pages
☆ CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credits
Diffusion large language models (dLLMs) generate text through iterative denoising steps, achieving parallel decoding by denoising only high-confidence positions at each step. However, existing approaches often repetitively remask tokens due to initially low confidence scores, leading to redundant iterations and limiting overall acceleration. Through the analysis of dLLM decoding traces, we observe that the model often determines the final prediction for a token several steps before the decoding step. To leverage this historical information and avoid redundant steps, we introduce the concept of Trace Credit, which quantifies each token's convergence potential by accumulating historical logits. Furthermore, we propose CreditDecoding, a training-free parallel decoding algorithm that accelerates the confidence convergence of correct but underconfident tokens by fusing current logits with Trace Credit. This process significantly reduces redundant iterations and enhances decoding robustness. On eight benchmarks, CreditDecoding achieves a 5.48 times speedup and a 0.48 performance improvement over LLaDA-8B-Instruct, and a 4.11 times speedup with a 0.15 performance improvement over LLaDA-MoE-Instruct. Importantly, CreditDecoding scales effectively to long sequences and is orthogonal to mainstream inference optimizations, making it a readily integrable and versatile solution.
comment: 18 pages,8 figures,4 tables
☆ Parallel Tokenizers: Rethinking Vocabulary Design for Cross-Lingual Transfer
Tokenization defines the foundation of multilingual language models by determining how words are represented and shared across languages. However, existing methods often fail to support effective cross-lingual transfer because semantically equivalent words are assigned distinct embeddings. For example, "I eat rice" in English and "Ina cin shinkafa" in Hausa are typically mapped to different vocabulary indices, preventing shared representations and limiting cross-lingual generalization. We introduce parallel tokenizers. This new framework trains tokenizers monolingually and then aligns their vocabularies exhaustively using bilingual dictionaries or word-to-word translation, ensuring consistent indices for semantically equivalent words. This alignment enforces a shared semantic space across languages while naturally improving fertility balance. To assess their effectiveness, we pretrain a transformer encoder from scratch on thirteen low-resource languages and evaluate it on sentiment analysis, hate speech detection, emotion classification, and sentence embedding similarity. Across all tasks, models trained with parallel tokenizers outperform conventional multilingual baselines, confirming that rethinking tokenization is essential for advancing multilingual representation learning--especially in low-resource settings.
comment: 18 pages, 25 tables, 7 figures
☆ Taxonomy of User Needs and Actions
The growing ubiquity of conversational AI highlights the need for frameworks that capture not only users' instrumental goals but also the situated, adaptive, and social practices through which they achieve them. Existing taxonomies of conversational behavior either overgeneralize, remain domain-specific, or reduce interactions to narrow dialogue functions. To address this gap, we introduce the Taxonomy of User Needs and Actions (TUNA), an empirically grounded framework developed through iterative qualitative analysis of 1193 human-AI conversations, supplemented by theoretical review and validation across diverse contexts. TUNA organizes user actions into a three-level hierarchy encompassing behaviors associated with information seeking, synthesis, procedural guidance, content creation, social interaction, and meta-conversation. By centering user agency and appropriation practices, TUNA enables multi-scale evaluation, supports policy harmonization across products, and provides a backbone for layering domain-specific taxonomies. This work contributes a systematic vocabulary for describing AI use, advancing both scholarly understanding and practical design of safer, more responsive, and more accountable conversational systems.
☆ Influence Functions for Efficient Data Selection in Reasoning
Fine-tuning large language models (LLMs) on chain-of-thought (CoT) data shows that a small amount of high-quality data can outperform massive datasets. Yet, what constitutes "quality" remains ill-defined. Existing reasoning methods rely on indirect heuristics such as problem difficulty or trace length, while instruction-tuning has explored a broader range of automated selection strategies, but rarely in the context of reasoning. We propose to define reasoning data quality using influence functions, which measure the causal effect of individual CoT examples on downstream accuracy, and introduce influence-based pruning, which consistently outperforms perplexity and embedding-based baselines on math reasoning within a model family.
☆ Distributional Semantics Tracing: A Framework for Explaining Hallucinations in Large Language Models
Large Language Models (LLMs) are prone to hallucination, the generation of plausible yet factually incorrect statements. This work investigates the intrinsic, architectural origins of this failure mode through three primary contributions.First, to enable the reliable tracing of internal semantic failures, we propose \textbf{Distributional Semantics Tracing (DST)}, a unified framework that integrates established interpretability techniques to produce a causal map of a model's reasoning, treating meaning as a function of context (distributional semantics). Second, we pinpoint the model's layer at which a hallucination becomes inevitable, identifying a specific \textbf{commitment layer} where a model's internal representations irreversibly diverge from factuality. Third, we identify the underlying mechanism for these failures. We observe a conflict between distinct computational pathways, which we interpret using the lens of dual-process theory: a fast, heuristic \textbf{associative pathway} (akin to System 1) and a slow, deliberate \textbf{contextual pathway} (akin to System 2), leading to predictable failure modes such as \textit{Reasoning Shortcut Hijacks}. Our framework's ability to quantify the coherence of the contextual pathway reveals a strong negative correlation ($\rho = -0.863$) with hallucination rates, implying that these failures are predictable consequences of internal semantic weakness. The result is a mechanistic account of how, when, and why hallucinations occur within the Transformer architecture.
☆ The Valley of Code Reasoning: Scaling Knowledge Distillation of Large Language Models NeurIPS 2025
Distilling the thinking traces of a Large Language Model (LLM) with reasoning capabilities into a smaller model has been proven effective. Yet, there is a scarcity of work done on how model performances scale with the quantity of distillation data. In this work, we study the scaling trend of distilling competitive coding skills on two small non-reasoning LLMs. We validate the hypothesis that there is a $\textit{valley of code reasoning}$: downstream performance on competitive coding first drops as data quantity increases, then it steadily increases in a sharper-than-log-linear fashion. Having identified the trend, we further fine-tune the models at two different distillation stages on the same data to ground conclusions on their respective learning phases. We learn that across stages in the low and medium-low data regimes, small models benefit significantly from easier coding questions than from harder ones. We also find that, surprisingly, the correctness of outputs in training data makes no difference to distillation outcomes. Our work represents a step forward in understanding the training dynamics of code reasoning distillation outside intuition
comment: NeurIPS 2025 Workshop on Deep Learning for Code (DL4C), Project page: https://collinear.ai/valley-of-reasoning
☆ The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives
The objectives that Large Language Models (LLMs) implicitly optimize remain dangerously opaque, making trustworthy alignment and auditing a grand challenge. While Inverse Reinforcement Learning (IRL) can infer reward functions from behaviour, existing approaches either produce a single, overconfident reward estimate or fail to address the fundamental ambiguity of the task (non-identifiability). This paper introduces a principled auditing framework that re-frames reward inference from a simple estimation task to a comprehensive process for verification. Our framework leverages Bayesian IRL to not only recover a distribution over objectives but to enable three critical audit capabilities: (i) Quantifying and systematically reducing non-identifiability by demonstrating posterior contraction over sequential rounds of evidence; (ii) Providing actionable, uncertainty-aware diagnostics that expose spurious shortcuts and identify out-of-distribution prompts where the inferred objective cannot be trusted; and (iii) Validating policy-level utility by showing that the refined, low-uncertainty reward can be used directly in RLHF to achieve training dynamics and toxicity reductions comparable to the ground-truth alignment process. Empirically, our framework successfully audits a detoxified LLM, yielding a well-calibrated and interpretable objective that strengthens alignment guarantees. Overall, this work provides a practical toolkit for auditors, safety teams, and regulators to verify what LLMs are truly trying to achieve, moving us toward more trustworthy and accountable AI.
comment: Preprint
☆ Learning from Failures: Understanding LLM Alignment through Failure-Aware Inverse RL
Reinforcement Learning from Human Feedback (RLHF) aligns Large Language Models (LLMs) with human preferences, yet the underlying reward signals they internalize remain hidden, posing a critical challenge for interpretability and safety. Existing approaches attempt to extract these latent incentives using Inverse Reinforcement Learning (IRL), but treat all preference pairs equally, often overlooking the most informative signals: those examples the extracted reward model misclassifies or assigns nearly equal scores, which we term \emph{failures}. We introduce a novel \emph{failure-aware} IRL algorithm that focuses on misclassified or difficult examples to recover the latent rewards defining model behaviors. By learning from these failures, our failure-aware IRL extracts reward functions that better reflect the true objectives behind RLHF. We demonstrate that failure-aware IRL outperforms existing IRL baselines across multiple metrics when applied to LLM detoxification, without requiring external classifiers or supervision. Crucially, failure-aware IRL yields rewards that better capture the true incentives learned during RLHF, enabling more effective re-RLHF training than standard IRL. This establishes failure-aware IRL as a robust, scalable method for auditing model alignment and reducing ambiguity in the IRL process.
comment: Preprint
☆ Spectrum Tuning: Post-Training for Distributional Coverage and In-Context Steerability
Language model post-training has enhanced instruction-following and performance on many downstream tasks, but also comes with an often-overlooked cost on tasks with many possible valid answers. We characterize three desiderata for conditional distributional modeling: in-context steerability, valid output space coverage, and distributional alignment, and document across three model families how current post-training can reduce these properties. In particular, we disambiguate between two kinds of in-context learning: ICL for eliciting existing underlying knowledge or capabilities, and in-context steerability, where a model must use in-context information to override its priors and steer to a novel data generating distribution. To better evaluate and improve these desiderata, we introduce Spectrum Suite, a large-scale resource compiled from >40 data sources and spanning >90 tasks requiring models to steer to and match diverse distributions ranging from varied human preferences to numerical distributions and more. We find that while current post-training techniques help elicit underlying capabilities and knowledge, they hurt models' ability to flexibly steer in-context. To mitigate these issues, we propose Spectrum Tuning, a post-training method using Spectrum Suite to improve steerability and distributional coverage. We find that Spectrum Tuning often improves over pretrained models and their instruction-tuned counterparts, enhancing steerability, spanning more of the output space, and improving distributional alignment on held-out datasets.
☆ ASPO: Asymmetric Importance Sampling Policy Optimization
Recent Large Language Model (LLM) post-training methods rely on token-level clipping mechanisms during Reinforcement Learning (RL). However, we identify a fundamental flaw in this Outcome-Supervised RL (OSRL) paradigm: the Importance Sampling (IS) ratios of positive-advantage tokens are mismatched, leading to unbalanced token weighting for positive and negative tokens. This mismatch suppresses the update of low-probability tokens while over-amplifying already high-probability ones. To address this, we propose Asymmetric Importance Sampling Policy Optimization (ASPO), which uses a simple yet effective strategy that flips the IS ratios of positive-advantage tokens, aligning their update direction with the learning dynamics of negative ones. AIS further incorporates a soft dual-clipping mechanism to stabilize extreme updates while maintaining gradient flow. Comprehensive experiments on coding and mathematical reasoning benchmarks demonstrate that ASPO significantly mitigates premature convergence, improves training stability, and enhances final performance over strong GRPO-based baselines. Our analysis provides new insights into the role of token-level weighting in OSRL and highlights the critical importance of correcting IS in LLM RL. The code and models of ASPO are available at https://github.com/wizard-III/Archer2.0.
☆ MixReasoning: Switching Modes to Think
Reasoning models enhance performance by tackling problems in a step-by-step manner, decomposing them into sub-problems and exploring long chains of thought before producing an answer. However, applying extended reasoning to every step introduces substantial redundancy, as sub-problems vary widely in difficulty and complexity: a small number of pivotal steps are genuinely challenging and decisive for the final answer, while many others only involve straightforward revisions or simple computations. Therefore, a natural idea is to endow reasoning models with the ability to adaptively respond to this variation, rather than treating all steps with the same level of elaboration. To this end, we propose MixReasoning, a framework that dynamically adjusts the depth of reasoning within a single response. The resulting chain of thought then becomes a mixture of detailed reasoning on difficult steps and concise inference on simpler ones. Experiments on GSM8K, MATH-500, and AIME show that MixReasoning shortens reasoning length and substantially improves efficiency without compromising accuracy.
☆ CDTP: A Large-Scale Chinese Data-Text Pair Dataset for Comprehensive Evaluation of Chinese LLMs
Large Language Models (LLMs) have achieved remarkable success across a wide range of natural language processing tasks. However, Chinese LLMs face unique challenges, primarily due to the dominance of unstructured free text and the lack of structured representations in Chinese corpora. While existing benchmarks for LLMs partially assess Chinese LLMs, they are still predominantly English-centric and fail to address the unique linguistic characteristics of Chinese, lacking structured datasets essential for robust evaluation. To address these challenges, we present a Comprehensive Benchmark for Evaluating Chinese Large Language Models (CB-ECLLM) based on the newly constructed Chinese Data-Text Pair (CDTP) dataset. Specifically, CDTP comprises over 7 million aligned text pairs, each consisting of unstructured text coupled with one or more corresponding triples, alongside a total of 15 million triples spanning four critical domains. The core contributions of CDTP are threefold: (i) enriching Chinese corpora with high-quality structured information; (ii) enabling fine-grained evaluation tailored to knowledge-driven tasks; and (iii) supporting multi-task fine-tuning to assess generalization and robustness across scenarios, including Knowledge Graph Completion, Triple-to-Text generation, and Question Answering. Furthermore, we conduct rigorous evaluations through extensive experiments and ablation studies to assess the effectiveness, Supervised Fine-Tuning (SFT), and robustness of the benchmark. To support reproducible research, we offer an open-source codebase and outline potential directions for future investigations based on our insights.
☆ Evaluating The Impact of Stimulus Quality in Investigations of LLM Language Performance
Recent studies employing Large Language Models (LLMs) to test the Argument from the Poverty of the Stimulus (APS) have yielded contrasting results across syntactic phenomena. This paper investigates the hypothesis that characteristics of the stimuli used in recent studies, including lexical ambiguities and structural complexities, may confound model performance. A methodology is proposed for re-evaluating LLM competence on syntactic prediction, focusing on GPT-2. This involves: 1) establishing a baseline on previously used (both filtered and unfiltered) stimuli, and 2) generating a new, refined dataset using a state-of-the-art (SOTA) generative LLM (Gemini 2.5 Pro Preview) guided by linguistically-informed templates designed to mitigate identified confounds. Our preliminary findings indicate that GPT-2 demonstrates notably improved performance on these refined PG stimuli compared to baselines, suggesting that stimulus quality significantly influences outcomes in surprisal-based evaluations of LLM syntactic competency.
comment: Presented at https://brigap-workshop.github.io/ Information to be updated upon publication of proceedings
☆ MASA: Rethinking the Representational Bottleneck in LoRA with Multi-A Shared Adaptation
Low-Rank Adaptation (LoRA) has emerged as a dominant method in Parameter-Efficient Fine-Tuning (PEFT) for large language models, which augments the transformer layer with one down-projection $A$ and one up-projection $B$. However, LoRA's reliance on a single down-projection matrix ($A$) creates a representational bottleneck, as this solitary feature extractor is inherently insufficient for capturing the diverse signals required by complex tasks. This motivates our architectural shift to focus on enriching the feature adaptation to improve the downstream task adaptation ability. We propose MASA (Multi-$A$ Shared Adaptation), an architecture that implements a multi-$A$, single-$B$ structure where the multi-$A$ expert ensemble is asymmetrically shared across layers to ensure parameter efficiency. In MASA, these specialized experts capture diverse features, which are then integrated by a single, layer-specific $B$-matrix. The effectiveness and versatility of our method are validated through a comprehensive suite of experiments spanning multi-domain generalization, single-domain specialization, and multi-task reasoning. For example, on the MMLU benchmark, MASA achieves an average accuracy of 59.62%, outperforming the standard LoRA by 1.08 points (a relative improvement of 1.84%) with comparable learnable parameters of 0.52%.
comment: 14 pages, 5 figures
☆ Deterministic Legal Retrieval: An Action API for Querying the SAT-Graph RAG
The Structure-Aware Temporal Graph RAG (SAT-Graph RAG) addresses core limitations of standard Retrieval-Augmented Generation in the legal domain by providing a verifiable knowledge graph that models hierarchical structure, temporal evolution, and causal events of legal norms. However, a critical gap remains: how to reliably query this structured knowledge without sacrificing its deterministic properties. This paper introduces the SAT-Graph API, a formal query execution layer centered on canonical actions-atomic, composable, and auditable primitives that isolate probabilistic discovery from deterministic retrieval. These actions enable: (i) high-precision hybrid search; (ii) robust reference resolution; (iii) point-in-time version retrieval; and (iv) auditable causal tracing. We demonstrate how planner-guided agents can decompose complex queries into Directed Acyclic Graphs (DAGs) of these actions. This two-layer architecture transforms retrieval from an opaque black box to a transparent, auditable process, directly addressing Explainable AI (XAI) requirements for high-stakes domains.
☆ Exploring Gaps in the APS: Direct Minimal Pair Analysis in LLM Syntactic Assessments
Recent studies probing the Argument from the Poverty of the Stimulus (APS) have applied Large Language Models (LLMs) to test the learnability of complex syntax through surprisal-based metrics. However, divergent conclusions raise questions concerning the insights these metrics offer. While Wilcox et al. (2024) used direct minimal pair comparisons (the "wh-effect") to demonstrate that models successfully generalise knowledge of filler-gap dependencies, Lan et al. (2024) used a Difference-in-Differences (DiD) metric and found that models largely fail on parasitic gaps (PGs). This paper argues that the direct minimal pair approach offers greater diagnostic transparency. We demonstrate this by generating a full 8-permutation paradigm of refined PG stimuli and evaluating the GPT-2 model used in previous studies with a systematic Wilcox-style wh-effect analysis. Our results show that GPT-2 succeeds across all four tested conditions, indicating robust knowledge of filler-gap licensing principles even in complex PG environments. This finding, which contrasts with the more ambiguous results from DiD-style metrics, suggests that the choice of evaluation metric is critical for assessing an LLM's syntactic competence.
comment: Presented at the https://brigap-workshop.github.io/ Information to be updated after publication of proceedings
☆ Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs
Large Language Models (LLMs) are increasingly applied to complex tasks that require extended reasoning. In such settings, models often benefit from diverse chains-of-thought to arrive at multiple candidate solutions. This requires two competing objectives: to inject enough stochasticity to explore multiple reasoning chains, and to ensure sufficient accuracy and quality in each path. Existing works pursue the first objective by increasing exploration at highly uncertain steps with higher temperature or larger candidate token sets, while others improve reliability by rejecting samples with low confidence post-generation, implying that low confidence correlates with low answer quality. These two lines of thought are in conflict, as they conflate different sources of uncertainty. To resolve this, we argue that the decoding rule should be calibrated by correctness, not confidence alone. We should sample from tokens with higher estimated correctness, and reduce sampling where expected correctness is low. We propose simple strategies that achieve this goal: Greedy-Threshold makes sampling greedy at very low confidence steps. Calibrated-TopK and Calibrated-epsilon set truncation threshold based on estimated rank-wise correctness. Together, our findings challenge prevailing heuristics about decoding under uncertainty and show gains across math and general reasoning benchmarks.
☆ LexiCon: a Benchmark for Planning under Temporal Constraints in Natural Language
Owing to their reasoning capabilities, large language models (LLMs) have been evaluated on planning tasks described in natural language. However, LLMs have largely been tested on planning domains without constraints. In order to deploy them in real-world settings where adherence to constraints, in particular safety constraints, is critical, we need to evaluate their performance on constrained planning tasks. We introduce LexiCon -- a natural language-based (Lexi) constrained (Con) planning benchmark, consisting of a suite of environments, that can be used to evaluate the planning capabilities of LLMs in a principled fashion. The core idea behind LexiCon is to take existing planning environments and impose temporal constraints on the states. These constrained problems are then translated into natural language and given to an LLM to solve. A key feature of LexiCon is its extensibility. That is, the set of supported environments can be extended with new (unconstrained) environment generators, for which temporal constraints are constructed automatically. This renders LexiCon future-proof: the hardness of the generated planning problems can be increased as the planning capabilities of LLMs improve. Our experiments reveal that the performance of state-of-the-art LLMs, including reasoning models like GPT-5, o3, and R1, deteriorates as the degree of constrainedness of the planning tasks increases.
☆ Probing the Difficulty Perception Mechanism of Large Language Models
Large language models (LLMs) are increasingly deployed on complex reasoning tasks, yet little is known about their ability to internally evaluate problem difficulty, which is an essential capability for adaptive reasoning and efficient resource allocation. In this work, we investigate whether LLMs implicitly encode problem difficulty in their internal representations. Using a linear probe on the final-token representations of LLMs, we demonstrate that the difficulty level of math problems can be linearly modeled. We further locate the specific attention heads of the final Transformer layer: these attention heads have opposite activation patterns for simple and difficult problems, thus achieving perception of difficulty. Our ablation experiments prove the accuracy of the location. Crucially, our experiments provide practical support for using LLMs as automatic difficulty annotators, potentially substantially reducing reliance on costly human labeling in benchmark construction and curriculum learning. We also uncover that there is a significant difference in entropy and difficulty perception at the token level. Our study reveals that difficulty perception in LLMs is not only present but also structurally organized, offering new theoretical insights and practical directions for future research.
☆ MatheMagic: Generating Dynamic Mathematics Benchmarks Robust to Memorization
Conducting contamination-free evaluation of mathematical capabilities can be difficult for two reasons: models may memorize a test set once it is made public, and current mathematical benchmarks are prone to overfitting due to having limited diversity of symbols and rules, coupled with closed-ended answers. This paper proposes a method to leverage these shortcomings as useful features to a construct dynamic, counterfactual benchmark, which can be used to both reveal overfitting and measure true reasoning. We demonstrate this via MatheMagic, which generates math test instances with the interpretations of numbers and operators altered, yet has automatically verifiable answers. Test instances are randomly seeded and constructed at test time to evaluate a model's induction or deduction capability, offering stability, extensibility, comparability, and robustness to overfitting. Our experiments find that models solve deduction more easily than induction, but they revert to standard math. Further analysis reveals that math-adapted models fail to exhibit a general "skill" of reasoning, and fine-tuning on induction tasks generalizes poorly.
☆ EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models
We present EvalMORAAL, a transparent chain-of-thought (CoT) framework that uses two scoring methods (log-probabilities and direct ratings) plus a model-as-judge peer review to evaluate moral alignment in 20 large language models. We assess models on the World Values Survey (55 countries, 19 topics) and the PEW Global Attitudes Survey (39 countries, 8 topics). With EvalMORAAL, top models align closely with survey responses (Pearson's r approximately 0.90 on WVS). Yet we find a clear regional difference: Western regions average r=0.82 while non-Western regions average r=0.61 (a 0.21 absolute gap), indicating consistent regional bias. Our framework adds three parts: (1) two scoring methods for all models to enable fair comparison, (2) a structured chain-of-thought protocol with self-consistency checks, and (3) a model-as-judge peer review that flags 348 conflicts using a data-driven threshold. Peer agreement relates to survey alignment (WVS r=0.74, PEW r=0.39, both p<.001), supporting automated quality checks. These results show real progress toward culture-aware AI while highlighting open challenges for use across regions.
☆ Hire Your Anthropologist! Rethinking Culture Benchmarks Through an Anthropological Lens
Cultural evaluation of large language models has become increasingly important, yet current benchmarks often reduce culture to static facts or homogeneous values. This view conflicts with anthropological accounts that emphasize culture as dynamic, historically situated, and enacted in practice. To analyze this gap, we introduce a four-part framework that categorizes how benchmarks frame culture, such as knowledge, preference, performance, or bias. Using this lens, we qualitatively examine 20 cultural benchmarks and identify six recurring methodological issues, including treating countries as cultures, overlooking within-culture diversity, and relying on oversimplified survey formats. Drawing on established anthropological methods, we propose concrete improvements: incorporating real-world narratives and scenarios, involving cultural communities in design and validation, and evaluating models in context rather than isolation. Our aim is to guide the development of cultural benchmarks that go beyond static recall tasks and more accurately capture the responses of the models to complex cultural situations.
comment: 12 pages; 2 figures; First two author contributed equally
Prompt reinforcing for long-term planning of large language models
Large language models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks and can be adapted through prompting. However, they remain suboptimal in multi-turn interactions, often relying on incorrect early assumptions and failing to track user goals over time, which makes such tasks particularly challenging. Prior works in dialogue systems have shown that long-term planning is essential for handling interactive tasks. In this work, we propose a prompt optimisation framework inspired by reinforcement learning, which enables such planning to take place by only modifying the task instruction prompt of the LLM-based agent. By generating turn-by-turn feedback and leveraging experience replay for prompt rewriting, our proposed method shows significant improvement in multi-turn tasks such as text-to-SQL and task-oriented dialogue. Moreover, it generalises across different LLM-based agents and can leverage diverse LLMs as meta-prompting agents. This warrants future research in reinforcement learning-inspired parameter-free optimisation methods.
☆ Paying Attention to Hybrid Attention: Untangling the Issues with Conversion Methods
Transformers' quadratic computational complexity limits their scalability despite remarkable performance. While linear attention reduces this to linear complexity, pre-training such models from scratch remains, in most cases, prohibitively expensive. Recent post-training linearisation methods convert pre-trained Transformers to linear models efficiently, often using hybrid approaches that combine linear attention with sliding-window softmax. We identify a critical flaw: existing hybrid methods inadvertently bypass the linear component, relying almost entirely on SWA. Component-level diagnostics reveal this previously undetected behaviour stems from overlooked evaluation practices on common-sense benchmarks. We propose three solutions to ensure balanced component usage: (i) inference-time hybridisation of linear-only conversions with sliding-window softmax; (ii) HedgeCATs, combining attention-weight transfer with targeted LoRA fine-tuning; and (iii) Scheduled Sliding-window Dropout (SSD), which stochastically suppresses the softmax branch during training to prevent component collapse. Our methods maintain computational efficiency while recovering most base model performance and ensuring genuine linear attention adoption, restoring the validity of performance attributions in hybrid conversions.
☆ The fragility of "cultural tendencies" in LLMs
In a recent study, Lu, Song, and Zhang (2025) (LSZ) propose that large language models (LLMs), when prompted in different languages, display culturally specific tendencies. They report that the two models (i.e., GPT and ERNIE) respond in more interdependent and holistic ways when prompted in Chinese, and more independent and analytic ways when prompted in English. LSZ attribute these differences to deep-seated cultural patterns in the models, claiming that prompt language alone can induce substantial cultural shifts. While we acknowledge the empirical patterns they observed, we find their experiments, methods, and interpretations problematic. In this paper, we critically re-evaluate the methodology, theoretical framing, and conclusions of LSZ. We argue that the reported "cultural tendencies" are not stable traits but fragile artifacts of specific models and task design. To test this, we conducted targeted replications using a broader set of LLMs and a larger number of test items. Our results show that prompt language has minimal effect on outputs, challenging LSZ's claim that these models encode grounded cultural beliefs.
☆ Evaluating the Sensitivity of LLMs to Harmful Contents in Long Input
Large language models (LLMs) increasingly support applications that rely on extended context, from document processing to retrieval-augmented generation. While their long-context capabilities are well studied for reasoning and retrieval, little is known about their behavior in safety-critical scenarios. We evaluate LLMs' sensitivity to harmful content under extended context, varying type (explicit vs. implicit), position (beginning, middle, end), prevalence (0.01-0.50 of the prompt), and context length (600-6000 tokens). Across harmful content categories such as toxic, offensive, and hate speech, with LLaMA-3, Qwen-2.5, and Mistral, we observe similar patterns: performance peaks at moderate harmful prevalence (0.25) but declines when content is very sparse or dominant; recall decreases with increasing context length; harmful sentences at the beginning are generally detected more reliably; and explicit content is more consistently recognized than implicit. These findings provide the first systematic view of how LLMs prioritize and calibrate harmful content in long contexts, highlighting both their emerging strengths and the challenges that remain for safety-critical use.
☆ Revisiting Long-context Modeling from Context Denoising Perspective
Long-context models (LCMs) have demonstrated great potential in processing long sequences, facilitating many real-world applications. The success of LCMs can be attributed to their ability to locate implicit critical information within the context for further prediction. However, recent research reveals that LCMs are often susceptible to contextual noise, i.e., irrelevant tokens, that can mislead model attention. In this paper, we conduct a fine-grained analysis of the context noise and propose an effective metric, the Integrated Gradient (IG) score, to detect and quantify the noise information within the context. Our findings reveal that even simple mitigation of detected context noise can substantially boost the model's attention on critical tokens and benefit subsequent predictions. Building on this insight, we propose Context Denoising Training (CDT), a straightforward yet effective training strategy that improves attention on critical tokens while reinforcing their influence on model predictions. Extensive experiments across four tasks, under both context window scaling and long-context alignment settings, demonstrate the superiority of CDT. Notably, when trained with CDT, an open-source 8B model can achieve performance (50.92) comparable to GPT-4o (51.00).
☆ Automated Boilerplate: Prevalence and Quality of Contract Generators in the Context of Swiss Privacy Policies
It has become increasingly challenging for firms to comply with a plethora of novel digital regulations. This is especially true for smaller businesses that often lack both the resources and know-how to draft complex legal documents. Instead of seeking costly legal advice from attorneys, firms may turn to cheaper alternative legal service providers such as automated contract generators. While these services have a long-standing presence, there is little empirical evidence on their prevalence and output quality. We address this gap in the context of a 2023 Swiss privacy law revision. To enable a systematic evaluation, we create and annotate a multilingual benchmark dataset that captures key compliance obligations under Swiss and EU privacy law. Using this dataset, we validate a novel GPT-5-based method for large-scale compliance assessment of privacy policies, allowing us to measure the impact of the revision. We observe compliance increases indicating an effect of the revision. Generators, explicitly referenced by 18% of local websites, are associated with substantially higher levels of compliance, with increases of up to 15 percentage points compared to privacy policies without generator use. These findings contribute to three debates: the potential of LLMs for cross-lingual legal analysis, the Brussels Effect of EU regulations, and, crucially, the role of automated tools in improving compliance and contractual quality.
comment: 23 pages, 4 figures
☆ DACP: Domain-Adaptive Continual Pre-Training of Large Language Models for Phone Conversation Summarization EMNLP 2025
Large language models (LLMs) have achieved impressive performance in text summarization, yet their performance often falls short when applied to specialized domains %or conversational data that differ from their original pre-training distribution. While fine-tuning can improve summarization quality, it typically relies on costly and scarce high-quality labeled data. In this work, we explore continual pre-training as a scalable, self-supervised approach to adapt LLMs for downstream summarization tasks, particularly in the context of noisy real-world conversation transcripts. We conduct extensive experiments using large-scale, unlabeled business conversation data to investigate whether continual pre-training enhances model capabilities in conversational summarization. Our results demonstrate that continual pre-training yields substantial gains in both in-domain and out-of-domain summarization benchmarks, while maintaining strong generalization and robustness. We also analyze the effects of data selection strategies, providing practical guidelines for applying continual pre-training in summarization-focused industrial applications.
comment: Accepted to the NewSumm Workshop at EMNLP 2025
☆ Luth: Efficient French Specialization for Small Language Models and Cross-Lingual Transfer
The landscape of Large Language Models (LLMs) remains predominantly English-centric, resulting in a significant performance gap for other major languages, such as French, especially in the context of Small Language Models (SLMs). Existing multilingual models demonstrate considerably lower performance in French compared to English, and research on efficient adaptation methods for French remains limited. To address this, we introduce \textbf{Luth}, a family of French-specialized SLMs: through targeted post-training on curated, high-quality French data, our models outperform all open-source counterparts of comparable size on multiple French benchmarks while retaining their original English capabilities. We further show that strategic model merging enhances performance in both languages, establishing Luth as a new state of the art for French SLMs and a robust baseline for future French-language research.
comment: 12 pages, 4 figures and 9 tables
☆ EEPO: Exploration-Enhanced Policy Optimization via Sample-Then-Forget
Balancing exploration and exploitation remains a central challenge in reinforcement learning with verifiable rewards (RLVR) for large language models (LLMs). Current RLVR methods often overemphasize exploitation, leading to entropy collapse, diminished exploratory capacity, and ultimately limited performance gains. Although techniques that increase policy stochasticity can promote exploration, they frequently fail to escape dominant behavioral modes. This creates a self-reinforcing loop-repeatedly sampling and rewarding dominant modes-that further erodes exploration. We introduce Exploration-Enhanced Policy Optimization (EEPO), a framework that promotes exploration via two-stage rollouts with adaptive unlearning. In the first stage, the model generates half of the trajectories; it then undergoes a lightweight unlearning step to temporarily suppress these sampled responses, forcing the second stage to explore different regions of the output space. This sample-then-forget mechanism disrupts the self-reinforcing loop and promotes wider exploration during rollouts. Across five reasoning benchmarks, EEPO outperforms GRPO, achieving average relative gains of 24.3% on Qwen2.5-3B, 33.0% on Llama3.2-3B-Instruct, and 10.4% on Qwen3-8B-Base.
☆ Mitigating Premature Exploitation in Particle-based Monte Carlo for Inference-Time Scaling
Inference-Time Scaling (ITS) improves language models by allocating more computation at generation time. Particle Filtering (PF) has emerged as a strong ITS method for complex mathematical reasoning tasks, but it is vulnerable when guided by process reward models, which often assign overconfident scores early in the reasoning process. This causes PF to suffer from premature exploitation: it myopically commits to locally promising trajectories, prunes potentially correct hypotheses, and converges to suboptimal solutions. This failure mode, known as particle impoverishment, is especially severe under constrained computational budgets. To address this, we analyze the problem and identify two root causes: a lack of diversity in the particle set due to overconfident resampling and consequent inability to assess the potential of a reasoning path. We introduce Entropic Particle Filtering (ePF), an algorithm that integrates two new techniques to solve these issues. The first technique, Entropic Annealing (EA), directly mitigates particle impoverishment by monitoring search diversity via entropy; when diversity drops, it intervenes by dynamically annealing the resampling distribution to preserve exploration. The second, an enhancement called Look-ahead Modulation (LaM), adds a predictive guide to evaluate a state's potential based on its successors. On several challenging math benchmarks, ePF significantly outperforms strong baselines and achieves up to a 50 % relative improvement in task reward. Together, these methods improve PF's resilience by balancing the exploration of diverse solution spaces with the exploitation of high-reward regions, ultimately leading to higher-quality solutions.
☆ Data-efficient Targeted Token-level Preference Optimization for LLM-based Text-to-Speech
Aligning text-to-speech (TTS) system outputs with human feedback through preference optimization has been shown to effectively improve the robustness and naturalness of language model-based TTS models. Current approaches primarily require paired desirable and undesirable samples at the utterance level. However, such pairs are often limited in TTS output data, and utterance-level formulation prevents fine-grained token-level optimization needed for accurate pronunciation alignment. In this study, we propose TKTO that eliminates the need for paired data, enabling a more data-efficient training paradigm, and directly targets token-level units, automatically providing fine-grained alignment signals without token-level annotations. TKTO improves the challenging Japanese TTS accuracy by 39% and reduces CER by 54%, automatically assigning 12.8 times stronger reward to targeted tokens.
☆ Mixture of Neuron Experts
In this work, we first explore whether the parameters activated by the MoE layer remain highly sparse at inference. We perform a sparsification study on several representative MoE models. For each expert, we rank parameters by the magnitude of their activations from the gate projection and progressively prune the activated subset. Pruning up to 60% of parameters within that subset causes only negligible task-performance degradation; substantial drops occur only after more than 90% are removed. We further decompose experts into neuron-granular MoE and visualize their activation values, finding that most neuron activations are near zero. This observation motivates us to select only high-activation neuron experts during pretraining. Based on this insight, we propose Mixture of Neuron Experts (MoNE). MoNE achieves neuron-granular expert selection by only applying a simple top-k selection within each expert, incurs negligible latency, and requires no additional routing parameters or inter-expert communication. Extensive experiments demonstrate that MoNE matches traditional MoE performance while activating only 50% of the MoE-layer parameters, and it consistently outperforms traditional MoE when compared at equal numbers of activated parameters. These results suggest that MoNE is a practical approach to improving parameter utilization and inference efficiency in MoE-like models.
comment: 18 page, 11 figures, 7 tables
☆ InforME: Improving Informativeness of Abstractive Text Summarization With Informative Attention Guided by Named Entity Salience
Abstractive text summarization is integral to the Big Data era, which demands advanced methods to turn voluminous and often long text data into concise but coherent and informative summaries for efficient human consumption. Despite significant progress, there is still room for improvement in various aspects. One such aspect is to improve informativeness. Hence, this paper proposes a novel learning approach consisting of two methods: an optimal transport-based informative attention method to improve learning focal information in reference summaries and an accumulative joint entropy reduction method on named entities to enhance informative salience. Experiment results show that our approach achieves better ROUGE scores compared to prior work on CNN/Daily Mail while having competitive results on XSum. Human evaluation of informativeness also demonstrates the better performance of our approach over a strong baseline. Further analysis gives insight into the plausible reasons underlying the evaluation results.
☆ Diversity Is All You Need for Contrastive Learning: Spectral Bounds on Gradient Magnitudes
We derive non-asymptotic spectral bands that bound the squared InfoNCE gradient norm via alignment, temperature, and batch spectrum, recovering the \(1/\tau^{2}\) law and closely tracking batch-mean gradients on synthetic data and ImageNet. Using effective rank \(R_{\mathrm{eff}}\) as an anisotropy proxy, we design spectrum-aware batch selection, including a fast greedy builder. On ImageNet-100, Greedy-64 cuts time-to-67.5\% top-1 by 15\% vs.\ random (24\% vs.\ Pool--P3) at equal accuracy; CIFAR-10 shows similar gains. In-batch whitening promotes isotropy and reduces 50-step gradient variance by \(1.37\times\), matching our theoretical upper bound.
☆ Early Multimodal Prediction of Cross-Lingual Meme Virality on Reddit: A Time-Window Analysis
Predicting the virality of online content remains challenging, especially for culturally complex, fast-evolving memes. This study investigates the feasibility of early prediction of meme virality using a large-scale, cross-lingual dataset from 25 diverse Reddit communities. We propose a robust, data-driven method to define virality based on a hybrid engagement score, learning a percentile-based threshold from a chronologically held-out training set to prevent data leakage. We evaluated a suite of models, including Logistic Regression, XGBoost, and a Multi-layer Perceptron (MLP), with a comprehensive, multimodal feature set across increasing time windows (30-420 min). Crucially, useful signals emerge quickly: our best-performing model, XGBoost, achieves a PR-AUC $>$ 0.52 in just 30 minutes. Our analysis reveals a clear "evidentiary transition," in which the importance of the feature dynamically shifts from the static context to the temporal dynamics as a meme gains traction. This work establishes a robust, interpretable, and practical benchmark for early virality prediction in scenarios where full diffusion cascade data is unavailable, contributing a novel cross-lingual dataset and a methodologically sound definition of virality. To our knowledge, this study is the first to combine time series data with static content and network features to predict early meme virality.
comment: Preprint work in progress. Main body: 9 pages. Total: 15 pages including references and appendix. 16 figures and 12 tables
☆ ARM: Discovering Agentic Reasoning Modules for Generalizable Multi-Agent Systems
Large Language Model (LLM)-powered Multi-agent systems (MAS) have achieved state-of-the-art results on various complex reasoning tasks. Recent works have proposed techniques to automate the design of MASes, eliminating the need for manual engineering. However, these techniques perform poorly, often achieving similar or inferior performance to simple baselines. Furthermore, they require computationally expensive re-discovery of architectures for each new task domain and expensive data annotation on domains without existing labeled validation sets. A critical insight is that simple Chain of Thought (CoT) reasoning often performs competitively with these complex systems, suggesting that the fundamental reasoning unit of MASes, CoT, warrants further investigation. To this end, we present a new paradigm for automatic MAS design that pivots the focus to optimizing CoT reasoning. We introduce the Agentic Reasoning Module (ARM), an agentic generalization of CoT where each granular reasoning step is executed by a specialized reasoning module. This module is discovered through a tree search over the code space, starting from a simple CoT module and evolved using mutations informed by reflection on execution traces. The resulting ARM acts as a versatile reasoning building block which can be utilized as a direct recursive loop or as a subroutine in a learned meta-orchestrator. Our approach significantly outperforms both manually designed MASes and state-of-the-art automatic MAS design methods. Crucially, MASes built with ARM exhibit superb generalization, maintaining high performance across different foundation models and task domains without further optimization.
comment: 29 pages, 2 figures
☆ Adaptive and Multi-Source Entity Matching for Name Standardization of Astronomical Observation Facilities
This ongoing work focuses on the development of a methodology for generating a multi-source mapping of astronomical observation facilities. To compare two entities, we compute scores with adaptable criteria and Natural Language Processing (NLP) techniques (Bag-of-Words approaches, sequential approaches, and surface approaches) to map entities extracted from eight semantic artifacts, including Wikidata and astronomy-oriented resources. We utilize every property available, such as labels, definitions, descriptions, external identifiers, and more domain-specific properties, such as the observation wavebands, spacecraft launch dates, funding agencies, etc. Finally, we use a Large Language Model (LLM) to accept or reject a mapping suggestion and provide a justification, ensuring the plausibility and FAIRness of the validated synonym pairs. The resulting mapping is composed of multi-source synonym sets providing only one standardized label per entity. Those mappings will be used to feed our Name Resolver API and will be integrated into the International Virtual Observatory Alliance (IVOA) Vocabularies and the OntoPortal-Astro platform.
comment: Accepted in Ontology Matching 2025 conference proceedings
☆ Improving Discrete Diffusion Unmasking Policies Beyond Explicit Reference Policies
Masked diffusion models (MDMs) have recently emerged as a novel framework for language modeling. MDMs generate sentences by iteratively denoising masked sequences, filling in [MASK] tokens step by step. Although MDMs support any-order sampling, performance is highly sensitive to the choice of which position to unmask next. Prior work typically relies on rule-based schedules (e.g., max-confidence, max-margin), which provide ad hoc improvements. In contrast, we replace these heuristics with a learned scheduler. Specifically, we cast denoising as a KL-regularized Markov decision process (MDP) with an explicit reference policy and optimize a regularized objective that admits policy improvement and convergence guarantees under standard assumptions. We prove that the optimized policy under this framework generates samples that more closely match the data distribution than heuristic schedules. Empirically, across four benchmarks, our learned policy consistently outperforms max-confidence: for example, on SUDOKU, where unmasking order is critical, it yields a 20.1% gain over random and a 11.2% gain over max-confidence.
comment: Preprint
☆ Towards Reliable and Practical LLM Security Evaluations via Bayesian Modelling
Before adopting a new large language model (LLM) architecture, it is critical to understand vulnerabilities accurately. Existing evaluations can be difficult to trust, often drawing conclusions from LLMs that are not meaningfully comparable, relying on heuristic inputs or employing metrics that fail to capture the inherent uncertainty. In this paper, we propose a principled and practical end-to-end framework for evaluating LLM vulnerabilities to prompt injection attacks. First, we propose practical approaches to experimental design, tackling unfair LLM comparisons by considering two practitioner scenarios: when training an LLM and when deploying a pre-trained LLM. Second, we address the analysis of experiments and propose a Bayesian hierarchical model with embedding-space clustering. This model is designed to improve uncertainty quantification in the common scenario that LLM outputs are not deterministic, test prompts are designed imperfectly, and practitioners only have a limited amount of compute to evaluate vulnerabilities. We show the improved inferential capabilities of the model in several prompt injection attack settings. Finally, we demonstrate the pipeline to evaluate the security of Transformer versus Mamba architectures. Our findings show that consideration of output variability can suggest less definitive findings. However, for some attacks, we find notably increased Transformer and Mamba-variant vulnerabilities across LLMs with the same training data or mathematical ability.
☆ DecEx-RAG: Boosting Agentic Retrieval-Augmented Generation with Decision and Execution Optimization via Process Supervision
Agentic Retrieval-Augmented Generation (Agentic RAG) enhances the processing capability for complex tasks through dynamic retrieval and adaptive workflows. Recent advances (e.g., Search-R1) have shown that outcome-supervised reinforcement learning demonstrate strong performance. However, this approach still suffers from inefficient exploration, sparse reward signals, and ambiguous global reward feedback. To address these challenges, we propose DecEx-RAG, which models RAG as a Markov Decision Process (MDP) incorporating decision-making and execution, while introducing an efficient pruning strategy to optimize data expansion. Through comprehensive process-level policy optimization, DecEx-RAG significantly enhances the autonomous task decomposition, dynamic retrieval, and high-quality answer generation capabilities of large language models (LLMs). Experiments show that DecEx-RAG achieves an average absolute performance improvement of $6.2\%$ across six datasets, significantly outperforming existing baselines. Moreover, the pruning strategy improves data construction efficiency by nearly $6 \times$, providing an efficient solution for process-supervised RAG training. The code is available at https://github.com/sdsxdxl/DecEx-RAG.
☆ Code-Switching In-Context Learning for Cross-Lingual Transfer of Large Language Models
While large language models (LLMs) exhibit strong multilingual abilities, their reliance on English as latent representations creates a translation barrier, where reasoning implicitly depends on internal translation into English. When this process fails, performance in non-English languages deteriorates sharply, limiting the inclusiveness of LLM-based applications. Existing cross-lingual in-context learning (X-ICL) methods primarily leverage monolingual demonstrations, often failing to mitigate this barrier and instead reinforcing it. In this work, we introduce code-switching in-context learning (CSICL), a simple yet effective prompting strategy that progressively transitions from a target language to English within demonstrations and instruction to facilitate their latent reasoning in English. By explicitly scaffolding the reasoning process through controlled code-switching, CSICL acts as an implicit linguistic bridge that enhances cross-lingual alignment and reduces reliance on the translation barrier. We conduct extensive experiments across 4 LLMs, 6 datasets, and 10 languages, spanning both knowledge-intensive and reasoning-oriented domains. Our results demonstrate that CSICL consistently outperforms X-ICL baselines, achieving gains of 3.1%p and 1.9%p in both target and unseen languages, respectively. The improvement is even more pronounced in low-resource settings, with gains of 14.7% in target and 5.3% in unseen languages. These findings establish code-switching as a principled and robust approach for overcoming the translation barrier during inference, moving LLMs toward more equitable and effective multilingual systems.
☆ The African Languages Lab: A Collaborative Approach to Advancing Low-Resource African NLP
Despite representing nearly one-third of the world's languages, African languages remain critically underserved by modern NLP technologies, with 88\% classified as severely underrepresented or completely ignored in computational linguistics. We present the African Languages Lab (All Lab), a comprehensive research initiative that addresses this technological gap through systematic data collection, model development, and capacity building. Our contributions include: (1) a quality-controlled data collection pipeline, yielding the largest validated African multi-modal speech and text dataset spanning 40 languages with 19 billion tokens of monolingual text and 12,628 hours of aligned speech data; (2) extensive experimental validation demonstrating that our dataset, combined with fine-tuning, achieves substantial improvements over baseline models, averaging +23.69 ChrF++, +0.33 COMET, and +15.34 BLEU points across 31 evaluated languages; and (3) a structured research program that has successfully mentored fifteen early-career researchers, establishing sustainable local capacity. Our comparative evaluation against Google Translate reveals competitive performance in several languages while identifying areas that require continued development.
☆ Generative AI-Driven Hierarchical Multi-Agent Framework for Zero-Touch Optical Networks
The rapid development of Generative Artificial Intelligence (GenAI) has catalyzed a transformative technological revolution across all walks of life. As the backbone of wideband communication, optical networks are expecting high-level autonomous operation and zero-touch management to accommodate their expanding network scales and escalating transmission bandwidth. The integration of GenAI is deemed as the pivotal solution for realizing zero-touch optical networks. However, the lifecycle management of optical networks involves a multitude of tasks and necessitates seamless collaboration across multiple layers, which poses significant challenges to the existing single-agent GenAI systems. In this paper, we propose a GenAI-driven hierarchical multi-agent framework designed to streamline multi-task autonomous execution for zero-touch optical networks. We present the architecture, implementation, and applications of this framework. A field-deployed mesh network is utilized to demonstrate three typical scenarios throughout the lifecycle of optical network: quality of transmission estimation in the planning stage, dynamic channel adding/dropping in the operation stage, and system capacity increase in the upgrade stage. The case studies, illustrate the capabilities of multi-agent framework in multi-task allocation, coordination, execution, evaluation, and summarization. This work provides a promising approach for the future development of intelligent, efficient, and collaborative network management solutions, paving the way for more specialized and adaptive zero-touch optical networks.
comment: 7 pages,6 figures, Accepted by lEEE Communications Magazine, Open call
☆ MADIAVE: Multi-Agent Debate for Implicit Attribute Value Extraction
Implicit Attribute Value Extraction (AVE) is essential for accurately representing products in e-commerce, as it infers lantent attributes from multimodal data. Despite advances in multimodal large language models (MLLMs), implicit AVE remains challenging due to the complexity of multidimensional data and gaps in vision-text understanding. In this work, we introduce \textsc{\modelname}, a multi-agent debate framework that employs multiple MLLM agents to iteratively refine inferences. Through a series of debate rounds, agents verify and update each other's responses, thereby improving inference performance and robustness. Experiments on the ImplicitAVE dataset demonstrate that even a few rounds of debate significantly boost accuracy, especially for attributes with initially low performance. We systematically evaluate various debate configurations, including identical or different MLLM agents, and analyze how debate rounds affect convergence dynamics. Our findings highlight the potential of multi-agent debate strategies to address the limitations of single-agent approaches and offer a scalable solution for implicit AVE in multimodal e-commerce.
☆ A Goal Without a Plan Is Just a Wish: Efficient and Effective Global Planner Training for Long-Horizon Agent Tasks
Agents based on large language models (LLMs) struggle with brainless trial-and-error and generating hallucinatory actions due to a lack of global planning in long-horizon tasks. In this paper, we introduce a plan-and-execute framework and propose EAGLET, an efficient and effective planner training method to enhance the executor agent's planning abilities without human effort. Specifically, we train a plug-and-play global planner through a two-step process: we first synthesize high-quality plans from an advanced LLM using our proposed homologous consensus filtering strategy, and apply fine-tuning as a cold start. Moreover, we further improve the planner with a rule-based reinforcement learning stage using a novel executor capability gain reward, ensuring it can handle task instructions of varying difficulty. Experiments on three long-horizon agent tasks show that executor agents equipped with our planner outperform existing methods, achieving new state-of-the-art performance. Meanwhile, EAGLET reduces training costs by 8x compared to RL-based baselines, and it does not require manual effort or extra training data, offering an efficient and effective solution.
☆ Improving Chain-of-Thought Efficiency for Autoregressive Image Generation
Autoregressive multimodal large language models have recently gained popularity for image generation, driven by advances in foundation models. To enhance alignment and detail, newer approaches employ chain-of-thought (CoT) reasoning, expanding user inputs into elaborated prompts prior to image synthesis. However, this strategy can introduce unnecessary redundancy -- a phenomenon we call visual overthinking -- which increases computational costs and can introduce details that contradict the original prompt. In this work, we explore how to generate more concise CoT sequences for more efficient image generation. We introduce ShortCoTI, a lightweight optimization framework that encourages more concise CoT while preserving output image quality. ShortCoTI rewards more concise prompts with an adaptive function that scales according to an estimated difficulty for each task. Incorporating this reward into a reinforcement learning paradigm reduces prompt reasoning length by 54% while maintaining or slightly improving quality metrics across multiple benchmarks (T2I-CompBench, GenEval). Qualitative analysis shows that our method eliminates verbose explanations and repetitive refinements, producing reasoning prompts that are both concise and semantically rich. As a result, ShortCoTI improves computational efficiency without compromising the fidelity or visual appeal of generated images.
☆ In-the-Flow Agentic System Optimization for Effective Planning and Tool Use
Outcome-driven reinforcement learning has advanced reasoning in large language models (LLMs), but prevailing tool-augmented approaches train a single, monolithic policy that interleaves thoughts and tool calls under full context; this scales poorly with long horizons and diverse tools and generalizes weakly to new scenarios. Agentic systems offer a promising alternative by decomposing work across specialized modules, yet most remain training-free or rely on offline training decoupled from the live dynamics of multi-turn interaction. We introduce AgentFlow, a trainable, in-the-flow agentic framework that coordinates four modules (planner, executor, verifier, generator) through an evolving memory and directly optimizes its planner inside the multi-turn loop. To train on-policy in live environments, we propose Flow-based Group Refined Policy Optimization (Flow-GRPO), which tackles long-horizon, sparse-reward credit assignment by converting multi-turn optimization into a sequence of tractable single-turn policy updates. It broadcasts a single, verifiable trajectory-level outcome to every turn to align local planner decisions with global success and stabilizes learning with group-normalized advantages. Across ten benchmarks, AgentFlow with a 7B-scale backbone outperforms top-performing baselines with average accuracy gains of 14.9% on search, 14.0% on agentic, 14.5% on mathematical, and 4.1% on scientific tasks, even surpassing larger proprietary models like GPT-4o. Further analyses confirm the benefits of in-the-flow optimization, showing improved planning, enhanced tool-calling reliability, and positive scaling with model size and reasoning turns.
comment: 45 pages, 12 figures. Project website: https://agentflow.stanford.edu/
☆ Mission Impossible: Feedback-Guided Dynamic Interactive Planning for Improving Reasoning on LLMs
Recent advancements in language agents have led to significant improvements in multi-hop reasoning tasks. However, existing approaches often struggle with handling open-domain problems, which require massive information retrieval due to their reliance on a fixed sequence of actions. To address this, we propose Feedback-Guided Dynamic Interactive Planning (FGDIP), a novel framework tailored to enhance reasoning in LLMs by utilizing dynamic and adaptive strategies for information exploration in open-domain multi-hop reasoning tasks. Our approach begins by identifying key entities relevant to the problem, which serve as the initial nodes in the reasoning process. From these initial nodes, we then generate reasoning child nodes with the process being refined through a combination of historical error analysis and real-time feedback, which allows the framework to dynamically adjust and optimize its reasoning strategies. By integrating depth-first search with an innovative node generation technique, our framework adapts based on both prior error paths and concurrently generated nodes at the same hierarchical level. This dynamic strategy effectively expands the search space while ensuring the reasoning process systematically converges toward accurate solutions. Experimental results show that FGDIP achieved up to 54.47% F1 score on the HotpotQA dataset and 70.05% on the StrategyQA dataset, surpassing the best baseline by 5.03% and 7.25% respectively, highlighting its versatility and potential to enhance language agents in multi-hop reasoning tasks.
☆ Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations
The promotion of academic papers has become an important means of enhancing research visibility. However, existing automated methods struggle limited storytelling, insufficient aesthetic quality, and constrained self-adjustment, making it difficult to achieve efficient and engaging dissemination. At the heart of those challenges is a simple principle: \emph{there is no way to improve it when you cannot evaluate it right}. To address this, we introduce \textbf{EvoPresent}, a self-improvement agent framework that unifies coherent narratives, aesthetic-aware designs, and realistic presentation delivery via virtual characters. Central to EvoPresent is \textbf{PresAesth}, a multi-task reinforcement learning (RL) aesthetic model that provides reliable aesthetic scoring, defect adjustment, and comparative feedback, enabling iterative self-improvement even under limited aesthetic training data. To systematically evaluate the methods, we introduce \textbf{EvoPresent Benchmark}, a comprehensive benchmark comprising: \textit{Presentation Generation Quality}, built on 650 top-tier AI conference papers with multimodal resources (slides, videos and scripts) to assess both content and design; and \textit{Aesthetic Awareness}, consisting of 2,000 slide pairs with varying aesthetic levels, supporting joint training and evaluation on scoring, defect adjustment, and comparison. Our findings highlight that (i) High-quality feedback is essential for agent self-improvement, while initial capability alone does not guarantee effective self-correction. (ii) Automated generation pipelines exhibit a trade-off between visual design and content construction. (iii) Multi-task RL training shows stronger generalization in aesthetic awareness tasks.
☆ Domain-Shift-Aware Conformal Prediction for Large Language Models
Large language models have achieved impressive performance across diverse tasks. However, their tendency to produce overconfident and factually incorrect outputs, known as hallucinations, poses risks in real world applications. Conformal prediction provides finite-sample, distribution-free coverage guarantees, but standard conformal prediction breaks down under domain shift, often leading to under-coverage and unreliable prediction sets. We propose a new framework called Domain-Shift-Aware Conformal Prediction (DS-CP). Our framework adapts conformal prediction to large language models under domain shift, by systematically reweighting calibration samples based on their proximity to the test prompt, thereby preserving validity while enhancing adaptivity. Our theoretical analysis and experiments on the MMLU benchmark demonstrate that the proposed method delivers more reliable coverage than standard conformal prediction, especially under substantial distribution shifts, while maintaining efficiency. This provides a practical step toward trustworthy uncertainty quantification for large language models in real-world deployment.
comment: 26 pages
☆ Activation-Informed Pareto-Guided Low-Rank Compression for Efficient LLM/VLM
Large language models (LLM) and vision-language models (VLM) have achieved state-of-the-art performance, but they impose significant memory and computing challenges in deployment. We present a novel low-rank compression framework to address this challenge. First, we upper bound the change of network loss via layer-wise activation-based compression errors, filling a theoretical gap in the literature. We then formulate low-rank model compression as a bi-objective optimization and prove that a single uniform tolerance yields surrogate Pareto-optimal heterogeneous ranks. Based on our theoretical insights, we propose Pareto-Guided Singular Value Decomposition (PGSVD), a zero-shot pipeline that improves activation-aware compression via Pareto-guided rank selection and alternating least-squares implementation. We apply PGSVD to both LLM and VLM, showing better accuracy at the same compression levels and inference speedup.
☆ Sci-Phi: A Large Language Model Spatial Audio Descriptor
Acoustic scene perception involves describing the type of sounds, their timing, their direction and distance, as well as their loudness and reverberation. While audio language models excel in sound recognition, single-channel input fundamentally limits spatial understanding. This work presents Sci-Phi, a spatial audio large language model with dual spatial and spectral encoders that estimates a complete parameter set for all sound sources and the surrounding environment. Learning from over 4,000 hours of synthetic first-order Ambisonics recordings including metadata, Sci-Phi enumerates and describes up to four directional sound sources in one pass, alongside non-directional background sounds and room characteristics. We evaluate the model with a permutation-invariant protocol and 15 metrics covering content, location, timing, loudness, and reverberation, and analyze its robustness across source counts, signal-to-noise ratios, reverberation levels, and challenging mixtures of acoustically, spatially, or temporally similar sources. Notably, Sci-Phi generalizes to real room impulse responses with only minor performance degradation. Overall, this work establishes the first audio LLM capable of full spatial-scene description, with strong potential for real-world deployment. Demo: https://sci-phi-audio.github.io/demo
☆ On the Role of Difficult Prompts in Self-Play Preference Optimization
Self-play preference optimization has emerged as a prominent paradigm for aligning large language models (LLMs). It typically involves a language model to generate on-policy responses for prompts and a reward model (RM) to guide the selection of chosen and rejected responses, which can be further trained with direct preference optimization (DPO). However, the role of prompts remains underexplored, despite being a core component in this pipeline. In this work, we investigate how prompts of varying difficulty influence self-play preference optimization. We first use the mean reward of $N$ sampled responses of a prompt as a proxy for its difficulty. We find that difficult prompts exhibit substantially inferior self-play optimization performance in comparison to easy prompts for language models. Moreover, incorporating difficult prompts into training fails to enhance overall performance and, in fact, leads to slight degradation compared to training on easy prompts alone. We also observe that the performance gap between difficult and easy prompts closes as the model capacity increases, suggesting that difficulty interacts with the model capacity. Building on these findings, we explore strategies to mitigate the negative effect of difficult prompts on final performance. We demonstrate that selectively removing an appropriate portion of challenging prompts enhances overall self-play performance, while also reporting failed attempts and lessons learned.
☆ H1B-KV: Hybrid One-Bit Caches for Memory-Efficient Large Language Model Inference
Autoregressive decoding in large language models (LLMs) requires caching a growing list of past key-value (KV) pairs, making long-context inference a memory-bound problem. While recent methods have explored quantizing the cache, evicting tokens, or using binary sketches for keys (e.g., Loki), these approaches often provide an incomplete solution by leaving one component (like values) uncompressed or by discarding context information. This paper introduces the Hybrid One-Bit KV Cache (H1B-KV), a comprehensive compression scheme that radically reduces memory usage without sacrificing context. H1B-KV represents each key vector using a 1-bit binary sketch, enabling hardware-friendly bitwise attention, and further compresses value vectors using 4-bit quantization. This holistic, hybrid approach allows a 7-billion parameter LLM to handle an 8k-token context with under 60 MB of cache memory - a 70x reduction. We demonstrate that after a lightweight finetuning, H1B-KV matches full-precision performance not only on perplexity benchmarks but also on complex downstream tasks like mathematical reasoning (GSM8K), multi-task understanding (MMLU), and code generation (HumanEval). Our results show H1B-KV significantly outperforms leading quantization (KIVI), token eviction (SparseLLM), and key-only sketching (Loki) methods in quality-per-byte, establishing it as a robust solution for deploying LLMs in memory-constrained environments.
comment: MIT URTC 2025 Technical Paper (Oral), 5 pages, 1 figure
☆ KEO: Knowledge Extraction on OMIn via Knowledge Graphs and RAG for Safety-Critical Aviation Maintenance
We present Knowledge Extraction on OMIn (KEO), a domain-specific knowledge extraction and reasoning framework with large language models (LLMs) in safety-critical contexts. Using the Operations and Maintenance Intelligence (OMIn) dataset, we construct a QA benchmark spanning global sensemaking and actionable maintenance tasks. KEO builds a structured Knowledge Graph (KG) and integrates it into a retrieval-augmented generation (RAG) pipeline, enabling more coherent, dataset-wide reasoning than traditional text-chunk RAG. We evaluate locally deployable LLMs (Gemma-3, Phi-4, Mistral-Nemo) and employ stronger models (GPT-4o, Llama-3.3) as judges. Experiments show that KEO markedly improves global sensemaking by revealing patterns and system-level insights, while text-chunk RAG remains effective for fine-grained procedural tasks requiring localized retrieval. These findings underscore the promise of KG-augmented LLMs for secure, domain-specific QA and their potential in high-stakes reasoning.
☆ CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension NeurIPS 2025
Current Large Language Models (LLMs) are confronted with overwhelming information volume when comprehending long-form documents. This challenge raises the imperative of a cohesive memory module, which can elevate vanilla LLMs into autonomous reading agents. Despite the emergence of some heuristic approaches, a systematic design principle remains absent. To fill this void, we draw inspiration from Jean Piaget's Constructivist Theory, illuminating three traits of the agentic memory -- structured schemata, flexible assimilation, and dynamic accommodation. This blueprint forges a clear path toward a more robust and efficient memory system for LLM-based reading comprehension. To this end, we develop CAM, a prototype implementation of Constructivist Agentic Memory that simultaneously embodies the structurality, flexibility, and dynamicity. At its core, CAM is endowed with an incremental overlapping clustering algorithm for structured memory development, supporting both coherent hierarchical summarization and online batch integration. During inference, CAM adaptively explores the memory structure to activate query-relevant information for contextual response, akin to the human associative process. Compared to existing approaches, our design demonstrates dual advantages in both performance and efficiency across diverse long-text reading comprehension tasks, including question answering, query-based summarization, and claim verification.
comment: Accepted by NeurIPS 2025
☆ Prototype-Based Dynamic Steering for Large Language Models
Despite impressive breadth, LLMs still rely on explicit reasoning instructions or static, one-fits-all steering methods, leaving a gap for adaptive, instruction-free reasoning amplification. We present Prototype-Based Dynamic Steering (PDS), a test-time method that amplifies large language model (LLM) reasoning without adding or altering instructions. We introduce "reasoning prototypes" by clustering activation differences between Chain-of-Thought (CoT) and neutral prompts. At inference, an input's hidden state is projected onto these prototypes to form an instance-specific steering vector. Evaluated on GSM8K, AQuA-RAT, and BIG-Bench tasks, PDS consistently improves accuracy without fine-tuning or prompt engineering. Notably, the gains persist even when CoT is explicitly suppressed to improve cost-efficiency, indicating that the intervention strengthens latent reasoning processes rather than inducing a superficial behavioral shift. These results position dynamic, prototype-guided steering as a lightweight alternative to training-time approaches for enhancing LLM reasoning.
☆ NorMuon: Making Muon more efficient and scalable
The choice of optimizer significantly impacts the training efficiency and computational costs of large language models (LLMs). Recently, the Muon optimizer has demonstrated promising results by orthogonalizing parameter updates, improving optimization geometry through better conditioning. Despite Muon's emergence as a candidate successor to Adam, the potential for jointly leveraging their strengths has not been systematically explored. In this work, we bridge this gap by proposing NorMuon (Neuron-wise Normalized Muon), an optimizer that synergistically combines orthogonalization with neuron-level adaptive learning rates. Our analysis reveals that while Muon effectively reduces condition numbers, the resulting updates exhibit highly non-uniform neuron norms, causing certain neurons to dominate the optimization process. NorMuon addresses this imbalance by maintaining second-order momentum statistics for each neuron and applying row-wise normalization after orthogonalization, ensuring balanced parameter utilization while preserving Muon's conditioning benefits. To enable practical deployment at scale, we develop an efficient distributed implementation under the FSDP2 framework that strategically distributes orthogonalization computations across devices. Experiments across multiple model scales demonstrate that NorMuon consistently outperforms both Adam and Muon, achieving 21.74% better training efficiency than Adam and 11.31% improvement over Muon on 1.1 B pretraining setting, while maintaining a comparable memory footprint to Muon. Our findings suggest that orthogonalization and adaptive learning rates are complementary rather than competing approaches, opening new avenues for optimizer design in large-scale deep learning.
☆ LANTERN: Scalable Distillation of Large Language Models for Job-Person Fit and Explanation
Large language models (LLMs) have achieved strong performance across a wide range of natural language processing tasks. However, deploying LLMs at scale for domain specific applications, such as job-person fit and explanation in job seeking platforms, introduces distinct challenges. At LinkedIn, the job person fit task requires analyzing a candidate's public profile against job requirements to produce both a fit assessment and a detailed explanation. Directly applying open source or finetuned LLMs to this task often fails to yield high quality, actionable feedback due to the complexity of the domain and the need for structured outputs. Moreover, the large size of these models leads to high inference latency and limits scalability, making them unsuitable for online use. To address these challenges, we introduce LANTERN, a novel LLM knowledge distillation framework tailored specifically for job person fit tasks. LANTERN involves modeling over multiple objectives, an encoder model for classification purpose, and a decoder model for explanation purpose. To better distill the knowledge from a strong black box teacher model to multiple downstream models, LANTERN incorporates multi level knowledge distillation that integrates both data and logit level insights. In addition to introducing the knowledge distillation framework, we share our insights on post training techniques and prompt engineering, both of which are crucial for successfully adapting LLMs to domain specific downstream tasks. Extensive experimental results demonstrate that LANTERN significantly improves task specific metrics for both job person fit and explanation. Online evaluations further confirm its effectiveness, showing measurable gains in job seeker engagement, including a 0.24\% increase in apply rate and a 0.28\% increase in qualified applications.
comment: 9 pages, 4 figures, 5 tables
☆ Language Model as Planner and Formalizer under Constraints
LLMs have been widely used in planning, either as planners to generate action sequences end-to-end, or as formalizers to represent the planning domain and problem in a formal language that can derive plans deterministically. However, both lines of work rely on standard benchmarks that only include generic and simplistic environmental specifications, leading to potential overestimation of the planning ability of LLMs and safety concerns in downstream tasks. We bridge this gap by augmenting widely used planning benchmarks with manually annotated, fine-grained, and rich natural language constraints spanning four formally defined categories. Over 4 state-of-the-art reasoning LLMs, 3 formal languages, 5 methods, and 4 datasets, we show that the introduction of constraints not only consistently halves performance, but also significantly challenges robustness to problem complexity and lexical shift.
☆ TensorBLEU: Vectorized GPU-based BLEU Score Implementation for Per-Sentence In-Training Evaluation
Modern natural language processing models have achieved unprecedented scale, yet the tools for their evaluation often remain a computational bottleneck, limiting the pace of research. This is particularly acute for in-training evaluation metrics, such as per-sentence reward signals in Reinforcement Learning, which must operate efficiently on batches of token IDs directly on the GPU. In this paper, we introduce TensorBLEU, a novel implementation of the BLEU metric designed from the ground up for this specific use case. Our approach is fully vectorized for GPU-accelerated, per-sentence computation within PyTorch and introduces a memory-efficient counting mechanism. By creating a compact, batch-specific dictionary of n-grams using \texttt{torch.unique}, our method avoids the prohibitive memory costs of traditional hashing-based vectorization, making it practical for large-vocabulary models. We benchmark TensorBLEU against NLTK, the standard library for token-ID-based BLEU calculation on the CPU. Experiments show that TensorBLEU provides speedups of over 13x on consumer-grade GPUs (NVIDIA T4) and exceeding 40x on data-center-class hardware (NVIDIA A100). This performance transforms a significant bottleneck into a negligible part of the training loop. By clearly defining its role as a "Token-ID BLEU" for development purposes and open-sourcing our implementation, we provide a powerful tool for accelerating research in areas like RL-based model fine-tuning.
comment: 9 pages, 3 figures
♻ ☆ LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning
Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM's autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. We first construct a structured latent reasoning space using a Variational Autoencoder (VAE) that encodes text reasoning steps into blocks of thought tokens, preserving semantic information and interpretability while offering compact but expressive representations. Subsequently, we utilize a latent diffusion model that learns to denoise a block of latent thought tokens with a blockwise bidirectional attention mask, enabling longer horizon and iterative refinement with adaptive test-time compute. This design allows efficient parallel generation of diverse reasoning trajectories, allowing the model to plan and revise the reasoning process holistically. We conduct evaluations on a suite of mathematical reasoning and planning benchmarks. Empirical results show that LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods, revealing a new paradigm for text reasoning with latent diffusion.
Generative Interfaces for Language Models
Large language models (LLMs) are increasingly seen as assistants, copilots, and consultants, capable of supporting a wide range of tasks through natural conversation. However, most systems remain constrained by a linear request-response format that often makes interactions inefficient in multi-turn, information-dense, and exploratory tasks. To address these limitations, we propose Generative Interfaces for Language Models, a paradigm in which LLMs respond to user queries by proactively generating user interfaces (UIs) that enable more adaptive and interactive engagement. Our framework leverages structured interface-specific representations and iterative refinements to translate user queries into task-specific UIs. For systematic evaluation, we introduce a multidimensional assessment framework that compares generative interfaces with traditional chat-based ones across diverse tasks, interaction patterns, and query types, capturing functional, interactive, and emotional aspects of user experience. Results show that generative interfaces consistently outperform conversational ones, with up to a 72% improvement in human preference. These findings clarify when and why users favor generative interfaces, paving the way for future advancements in human-AI interaction.
comment: Preprint
♻ ☆ Tracing Multilingual Factual Knowledge Acquisition in Pretraining EMNLP
Large Language Models (LLMs) are capable of recalling multilingual factual knowledge present in their pretraining data. However, most studies evaluate only the final model, leaving the development of factual recall and crosslingual consistency throughout pretraining largely unexplored. In this work, we trace how factual recall and crosslingual consistency evolve during pretraining, focusing on OLMo-7B as a case study. We find that both accuracy and consistency improve over time for most languages. We show that this improvement is primarily driven by the fact frequency in the pretraining corpus: more frequent facts are more likely to be recalled correctly, regardless of language. Yet, some low-frequency facts in non-English languages can still be correctly recalled. Our analysis reveals that these instances largely benefit from crosslingual transfer of their English counterparts -- an effect that emerges predominantly in the early stages of pretraining. We pinpoint two distinct pathways through which multilingual factual knowledge acquisition occurs: (1) frequency-driven learning, which is dominant and language-agnostic, and (2) crosslingual transfer, which is limited in scale and typically constrained to relation types involving named entities. We release our code and data to facilitate further research at https://github.com/cisnlp/multilingual-fact-tracing.
comment: EMNLP Findings 2025
♻ ☆ LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures
Large Language Model (LLM) pretraining, finetuning, and evaluation rely on input-space reconstruction and generative capabilities. Yet, it has been observed in vision that embedding-space training objectives, e.g., with Joint Embedding Predictive Architectures (JEPAs), are far superior to their input-space counterpart. That mismatch in how training is achieved between language and vision opens up a natural question: {\em can language training methods learn a few tricks from the vision ones?} The lack of JEPA-style LLM is a testimony of the challenge in designing such objectives for language. In this work, we propose a first step in that direction where we develop LLM-JEPA, a JEPA based solution for LLMs applicable both to finetuning and pretraining. Thus far, LLM-JEPA is able to outperform the standard LLM training objectives by a significant margin across models, all while being robust to overfiting. Those findings are observed across numerous datasets (NL-RX, GSM8K, Spider, RottenTomatoes) and various models from the Llama3, OpenELM, Gemma2 and Olmo families. Code: https://github.com/rbalestr-lab/llm-jepa.
♻ ☆ Exploring the Potential of Conversational AI Support for Agent-Based Social Simulation Model Design
ChatGPT, the AI-powered chatbot with a massive user base of hundreds of millions, has become a global phenomenon. However, the use of Conversational AI Systems (CAISs) like ChatGPT for research in the field of Social Simulation is still limited. Specifically, there is no evidence of its usage in Agent-Based Social Simulation (ABSS) model design. This paper takes a crucial first step toward exploring the untapped potential of this emerging technology in the context of ABSS model design. The research presented here demonstrates how CAISs can facilitate the development of innovative conceptual ABSS models in a concise timeframe and with minimal required upfront case-based knowledge. By employing advanced prompt engineering techniques and adhering to the Engineering ABSS framework, we have constructed a comprehensive prompt script that enables the design of conceptual ABSS models with or by the CAIS. A proof-of-concept application of the prompt script, used to generate the conceptual ABSS model for a case study on the impact of adaptive architecture in a museum environment, illustrates the practicality of the approach. Despite occasional inaccuracies and conversational divergence, the CAIS proved to be a valuable companion for ABSS modellers.
comment: This paper has been published in the Journal of Artificial Societies and Social Simulation 28 (3) 2. Please refer to the published version at [https://doi.org/10.18564/jasss.5681]
♻ ☆ OWL: Probing Cross-Lingual Recall of Memorized Texts via World Literature EMNLP 2025
Large language models (LLMs) are known to memorize and recall English text from their pretraining data. However, the extent to which this ability generalizes to non-English languages or transfers across languages remains unclear. This paper investigates multilingual and cross-lingual memorization in LLMs, probing if memorized content in one language (e.g., English) can be recalled when presented in translation. To do so, we introduce OWL, a dataset of 31.5K aligned excerpts from 20 books in ten languages, including English originals, official translations (Vietnamese, Spanish, Turkish), and new translations in six low-resource languages (Sesotho, Yoruba, Maithili, Malagasy, Setswana, Tahitian). We evaluate memorization across model families and sizes through three tasks: (1) direct probing, which asks the model to identify a book's title and author; (2) name cloze, which requires predicting masked character names; and (3) prefix probing, which involves generating continuations. We find that LLMs consistently recall content across languages, even for texts without direct translation in pretraining data. GPT-4o, for example, identifies authors and titles 69% of the time and masked entities 6% of the time in newly translated excerpts. Perturbations (e.g., masking characters, shuffling words) modestly reduce direct probing accuracy (7% drop for shuffled official translations). Our results highlight the extent of cross-lingual memorization and provide insights on the differences between the models.
comment: Accepted to EMNLP 2025 Main
♻ ☆ How Reliable are Causal Probing Interventions?
Causal probing aims to analyze foundation models by examining how intervening on their representation of various latent properties impacts their outputs. Recent works have cast doubt on the theoretical basis of several leading causal probing methods, but it has been unclear how to systematically evaluate the effectiveness of these methods in practice. To address this, we define two key causal probing desiderata: completeness (how thoroughly the representation of the target property has been transformed) and selectivity (how little non-targeted properties have been impacted). We find that there is an inherent tradeoff between the two, which we define as reliability, their harmonic mean. We introduce an empirical analysis framework to measure and evaluate these quantities, allowing us to make the first direct comparisons between different families of leading causal probing methods (e.g., linear vs. nonlinear, or concept removal vs. counterfactual interventions). We find that: (1) all methods show a clear tradeoff between completeness and selectivity; (2) more complete and reliable methods have a greater impact on LLM behavior; and (3) nonlinear interventions are almost always more reliable than linear interventions.
♻ ☆ Trajectory Prediction Meets Large Language Models: A Survey
Recent advances in large language models (LLMs) have sparked growing interest in integrating language-driven techniques into trajectory prediction. By leveraging their semantic and reasoning capabilities, LLMs are reshaping how autonomous systems perceive, model, and predict trajectories. This survey provides a comprehensive overview of this emerging field, categorizing recent work into five directions: (1) Trajectory prediction via language modeling paradigms, (2) Direct trajectory prediction with pretrained language models, (3) Language-guided scene understanding for trajectory prediction, (4) Language-driven data generation for trajectory prediction, (5) Language-based reasoning and interpretability for trajectory prediction. For each, we analyze representative methods, highlight core design choices, and identify open challenges. This survey bridges natural language processing and trajectory prediction, offering a unified perspective on how language can enrich trajectory prediction.
comment: 16 pages, GitHub: https://github.com/colorfulfuture/Awesome-Trajectory-Motion-Prediction-Papers
♻ ☆ Is It Thinking or Cheating? Detecting Implicit Reward Hacking by Measuring Reasoning Effort
Reward hacking, where a reasoning model exploits loopholes in a reward function to achieve high rewards without solving the intended task, poses a significant threat. This behavior may be explicit, i.e. verbalized in the model's chain-of-thought (CoT), or implicit, where the CoT appears benign thus bypasses CoT monitors. To detect implicit reward hacking, we propose TRACE (Truncated Reasoning AUC Evaluation). Our key observation is that hacking occurs when exploiting the loophole is easier than solving the actual task. This means that the model is using less 'effort' than required to achieve high reward. TRACE quantifies effort by measuring how early a model's reasoning becomes sufficient to obtain the reward. We progressively truncate a model's CoT at various lengths, force the model to answer, and estimate the expected reward at each cutoff. A hacking model, which takes a shortcut, will achieve a high expected reward with only a small fraction of its CoT, yielding a large area under the accuracy-vs-length curve. TRACE achieves over 65% gains over our strongest 72B CoT monitor in math reasoning, and over 30% gains over a 32B monitor in coding. We further show that TRACE can discover unknown loopholes during training. Overall, TRACE offers a scalable unsupervised approach for oversight where current monitoring methods prove ineffective.
comment: 25 pages, 31 figures
Diagnosing the Performance Trade-off in Moral Alignment: A Case Study on Gender Stereotypes
Moral alignment has emerged as a widely adopted approach for regulating the behavior of pretrained language models (PLMs), typically through fine-tuning on curated datasets. Gender stereotype mitigation is a representational task within the broader application of moral alignment. However, this process often comes at the cost of degraded downstream task performance. Prior studies commonly aim to achieve a performance trade-off by encouraging PLMs to selectively forget only stereotypical knowledge through carefully designed fairness objective, while preserving their language modeling capability (overall forgetting). In this short paper, we investigate whether the performance trade-off can be achieved through the lens of forgetting and the fairness objective. Our analysis shows that the large datasets needed for satisfactory fairness highlight the limitations of current fairness objectives in achieving an effective trade-off: (1) downstream task performance is strongly correlated with overall forgetting; (2) selective forgetting reduces stereotypes, but overall forgetting increases. and (3) general solutions for alleviating forgetting are ineffective at reducing the overall forgetting and fail to improve downstream task performance.
♻ ☆ Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models
Reasoning language models improve performance on complex tasks by generating long chains of thought (CoTs), but this process can also increase harmful outputs in adversarial settings. In this work, we ask whether the long CoTs can be leveraged for predictive safety monitoring: do the reasoning traces provide early signals of final response alignment that could enable timely intervention? We evaluate a range of monitoring methods using either CoT text or activations, including highly capable large language models, fine-tuned classifiers, and humans. First, we find that a simple linear probe trained on CoT activations significantly outperforms all text-based baselines in predicting whether a final response is safe or unsafe, with an average absolute increase of 13 in F1 scores over the best-performing alternatives. CoT texts are often unfaithful and misleading, while model latents provide a more reliable predictive signal. Second, the probe can be applied to early CoT segments before the response is generated, showing that alignment signals appear before reasoning completes. Error analysis reveals that the performance gap between text classifiers and the linear probe largely stems from a subset of responses we call performative CoTs, where the reasoning consistently contradicts the final response as the CoT progresses. Our findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.
♻ ☆ Epistemic Diversity and Knowledge Collapse in Large Language Models
Large language models (LLMs) tend to generate lexically, semantically, and stylistically homogenous texts. This poses a risk of knowledge collapse, where homogenous LLMs mediate a shrinking in the range of accessible information over time. Existing works on homogenization are limited by a focus on closed-ended multiple-choice setups or fuzzy semantic features, and do not look at trends across time and cultural contexts. To overcome this, we present a new methodology to measure epistemic diversity, i.e., variation in real-world claims in LLM outputs, which we use to perform a broad empirical study of LLM knowledge collapse. We test 27 LLMs, 155 topics covering 12 countries, and 200 prompt variations sourced from real user chats. For the topics in our study, we show that while newer models tend to generate more diverse claims, nearly all models are less epistemically diverse than a basic web search. We find that model size has a negative impact on epistemic diversity, while retrieval-augmented generation (RAG) has a positive impact, though the improvement from RAG varies by the cultural context. Finally, compared to a traditional knowledge source (Wikipedia), we find that country-specific claims reflect the English language more than the local one, highlighting a gap in epistemic representation
comment: 16 pages; 8 figures, 4 tables v2 changelog: Fixed the modeling for table 3, random effect is the model version
♻ ☆ Entropy-Gated Branching for Efficient Test-Time Reasoning
Test-time compute methods can significantly improve the reasoning capabilities and problem-solving accuracy of large language models (LLMs). However, these approaches require substantially more computational resources, with most compute wasted on exploring low-diversity branches where the model already exhibits high confidence. We observe that a small subset of uncertain reasoning steps has a disproportionately large impact on final prediction accuracy, and branching at these critical junctures tends to yield more diverse and higher-quality candidate reasoning steps. We propose Entropy-Gated Branching (EGB), which branches only at high-uncertainty steps and prunes expansions with a lightweight verifier. On mathematical and financial reasoning benchmarks, EGB improves accuracy by 22.6% over standard inference while operating 31%-75% faster across math benchmarks than test-time beam search with higher performance. Our results show that dynamic resource allocation during inference can substantially improve both efficiency and effectiveness, offering a more scalable pathway to enhanced LLM reasoning capabilities.
♻ ☆ CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysis
Large Language Models (LLMs) with Mixture-of-Experts (MoE) architectures are distinguished by their strong performance scaling with increasing parameters across a wide range of tasks, yet they also suffer from substantial computational and storage overheads. Notably, the performance gains of MoE models do not scale proportionally with the growth in expert parameters. While prior works attempt to reduce parameters via expert-level pruning, merging, or decomposition, they still suffer from challenges in both performance and computational efficiency. In this paper, we address these challenges by introducing micro-expert as a finer-grained compression unit that spans across matrices. We first establish a more fundamental perspective, viewing MoE layers as mixtures of micro-experts, and present CAMERA, a lightweight and training-free framework for identifying micro-expert redundancy. Our analysis uncovers significant variance in micro-expert contributions during decoding. Based on this insight, we further propose CAMERA-P, a structured micro-expert pruning framework, and CAMERA-Q, a mixed-precision quantization idea designed for micro-experts. Extensive experiments on nine downstream tasks show that CAMERA-P consistently outperforms strong baselines under pruning ratios ranging from 20% to 60%. Furthermore, CAMERA-Q achieves superior results under aggressive 2-bit quantization, surpassing existing matrix- and channel-level ideas. Notably, our method enables complete micro-expert analysis of Qwen2-57B-A14B in less than 5 minutes on a single NVIDIA A100-40GB GPU.
comment: 16 pages, 9 figures, 7 tables
♻ ☆ On Relation-Specific Neurons in Large Language Models EMNLP 2025
In large language models (LLMs), certain \emph{neurons} can store distinct pieces of knowledge learned during pretraining. While factual knowledge typically appears as a combination of \emph{relations} and \emph{entities}, it remains unclear whether some neurons focus on a relation itself -- independent of any entity. We hypothesize such neurons \emph{detect} a relation in the input text and \emph{guide} generation involving such a relation. To investigate this, we study the LLama-2 family on a chosen set of relations, with a \textit{statistics}-based method. Our experiments demonstrate the existence of relation-specific neurons. We measure the effect of selectively deactivating candidate neurons specific to relation $r$ on the LLM's ability to handle (1) facts involving relation $r$ and (2) facts involving a different relation $r' \neq r$. With respect to their capacity for encoding relation information, we give evidence for the following three properties of relation-specific neurons. \textbf{(i) Neuron cumulativity.} Multiple neurons jointly contribute to processing facts involving relation $r$, with no single neuron fully encoding a fact in $r$ on its own. \textbf{(ii) Neuron versatility.} Neurons can be shared across multiple closely related as well as less related relations. In addition, some relation neurons transfer across languages. \textbf{(iii) Neuron interference.} Deactivating neurons specific to one relation can improve LLMs' factual recall performance for facts of other relations. We make our code and data publicly available at https://github.com/cisnlp/relation-specific-neurons.
comment: EMNLP 2025
♻ ☆ MedHal: An Evaluation Dataset for Medical Hallucination Detection
We present MedHal, a novel large-scale dataset specifically designed to evaluate if models can detect hallucinations in medical texts. Current hallucination detection methods face significant limitations when applied to specialized domains like medicine, where they can have disastrous consequences. Existing medical datasets are either too small, containing only a few hundred samples, or focus on a single task like Question Answering or Natural Language Inference. MedHal addresses these gaps by: (1) incorporating diverse medical text sources and tasks; (2) providing a substantial volume of annotated samples suitable for training medical hallucination detection models; and (3) including explanations for factual inconsistencies to guide model learning. We demonstrate MedHal's utility by training and evaluating a baseline medical hallucination detection model, showing improvements over general-purpose hallucination detection approaches. This resource enables more efficient evaluation of medical text generation systems while reducing reliance on costly expert review, potentially accelerating the development of medical AI research.
♻ ☆ Large Language Models Achieve Gold Medal Performance at the International Olympiad on Astronomy & Astrophysics (IOAA)
While task-specific demonstrations show early success in applying large language models (LLMs) to automate some astronomical research tasks, they only provide incomplete views of all necessary capabilities in solving astronomy problems, calling for more thorough understanding of LLMs' strengths and limitations. So far, existing benchmarks and evaluations focus on simple question-answering that primarily tests astronomical knowledge and fails to evaluate the complex reasoning required for real-world research in the discipline. Here, we address this gap by systematically benchmarking five state-of-the-art LLMs on the International Olympiad on Astronomy and Astrophysics (IOAA) exams, which are designed to examine deep conceptual understanding, multi-step derivations, and multimodal analysis. With average scores of 85.6% and 84.2%, Gemini 2.5 Pro and GPT-5 (the two top-performing models) not only achieve gold medal level performance but also rank in the top two among ~200-300 participants in all four IOAA theory exams evaluated (2022-2025). In comparison, results on the data analysis exams show more divergence. GPT-5 still excels in the exams with an 88.5% average score, ranking top 10 among the participants in the four most recent IOAAs, while other models' performances drop to 48-76%. Furthermore, our in-depth error analysis underscores conceptual reasoning, geometric reasoning, and spatial visualization (52-79% accuracy) as consistent weaknesses among all LLMs. Hence, although LLMs approach peak human performance in theory exams, critical gaps must be addressed before they can serve as autonomous research agents in astronomy.
comment: 18 pages, 6 figures, to be submitted, comments are welcome. Reproducibility details can be found at: https://github.com/OSU-NLP-Group/LLM-IOAA
♻ ☆ What MLLMs Learn about When they Learn about Multimodal Reasoning: Perception, Reasoning, or their Integration?
Multimodal reasoning models have recently shown promise on challenging domains such as olympiad-level geometry, yet their evaluation remains dominated by aggregate accuracy, a single score that obscures where and how models are improving. We introduce MathLens, a benchmark designed to disentangle the subskills of multimodal reasoning while preserving the complexity of textbook-style geometry problems. The benchmark separates performance into three components: Perception: extracting information from raw inputs, Reasoning: operating on available information, and Integration: selecting relevant perceptual evidence and applying it within reasoning. To support each test, we provide annotations: visual diagrams, textual descriptions to evaluate reasoning in isolation, controlled questions that require both modalities, and probes for fine-grained perceptual skills, all derived from symbolic specifications of the problems to ensure consistency and robustness. Our analysis reveals that different training approaches have uneven effects: First, reinforcement learning chiefly strengthens perception, especially when supported by textual supervision, while textual SFT indirectly improves perception through reflective reasoning. Second, reasoning improves only in tandem with perception. Third, integration remains the weakest capacity, with residual errors concentrated there once other skills advance. Finally, robustness diverges: RL improves consistency under diagram variation, whereas multimodal SFT reduces it through overfitting. We will release all data and experimental logs.
♻ ☆ v1: Learning to Point Visual Tokens for Multimodal Grounded Reasoning
When thinking with images, humans rarely rely on a single glance: they revisit visual information repeatedly during reasoning. However, existing models typically process images only once and thereafter generate reasoning entirely in text, lacking mechanisms to re-access or ground inference in visual representations. We empirically confirm this: as reasoning chains lengthen, models progressively lose focus on relevant regions. In response, we introduce v1, a lightweight extension that enables active visual referencing through a simple point-and-copy approach. This allows the model to identify relevant image patches and copy their embeddings back into the reasoning stream, ensuring that evolving hypotheses remain grounded in perceptual evidence. Crucially, our pointing strategy lets the MLLM directly select image patches using their semantic representations as keys, keeping perceptual evidence embedded in the same space as the model's reasoning. To train this capability, we construct v1g, a dataset of 300K multimodal reasoning traces with interleaved visual grounding annotations. Across various multimodal mathematical reasoning benchmarks, v1 consistently outperforms comparable baselines, establishing point-and-copy as a practical mechanism for grounded reasoning. The model checkpoint and dataset are available at github.com/jun297/v1.
♻ ☆ Generative Psycho-Lexical Approach for Constructing Value Systems in Large Language Models ACL 2025
Values are core drivers of individual and collective perception, cognition, and behavior. Value systems, such as Schwartz's Theory of Basic Human Values, delineate the hierarchy and interplay among these values, enabling cross-disciplinary investigations into decision-making and societal dynamics. Recently, the rise of Large Language Models (LLMs) has raised concerns regarding their elusive intrinsic values. Despite growing efforts in evaluating, understanding, and aligning LLM values, a psychologically grounded LLM value system remains underexplored. This study addresses the gap by introducing the Generative Psycho-Lexical Approach (GPLA), a scalable, adaptable, and theoretically informed method for constructing value systems. Leveraging GPLA, we propose a psychologically grounded five-factor value system tailored for LLMs. For systematic validation, we present three benchmarking tasks that integrate psychological principles with cutting-edge AI priorities. Our results reveal that the proposed value system meets standard psychological criteria, better captures LLM values, improves LLM safety prediction, and enhances LLM alignment, when compared to the canonical Schwartz's values.
comment: ACL 2025 Main
♻ ☆ AgenticIE: An Adaptive Agent for Information Extraction from Complex Regulatory Documents
Declaration of Performance (DoP) documents, mandated by EU regulation, certify the performance of construction products. There are two challenges to make DoPs machine and human accessible through automated key-value pair extraction (KVP) and question answering (QA): (1) While some of their content is standardized, DoPs vary widely in layout, schema, and format; (2) Both users and documents are multilingual. Existing static or LLM-only Information Extraction (IE) pipelines fail to adapt to this structural document and user diversity. Our domain-specific, agentic system addresses these challenges through a planner-executor-responder architecture. The system infers user intent, detects document language and modality, and orchestrates tools dynamically for robust, traceable reasoning while avoiding tool misuse or execution loops. Our agent outperforms baselines (ROUGE: 0.783 vs. 0.703/0.608) with better cross-lingual stability (17-point vs. 21-26-point variation).
♻ ☆ SAE-FiRE: Enhancing Earnings Surprise Predictions Through Sparse Autoencoder Feature Selection
Predicting earnings surprises from financial documents, such as earnings conference calls, regulatory filings, and financial news, has become increasingly important in financial economics. However, these financial documents present significant analytical challenges, typically containing over 5,000 words with substantial redundancy and industry-specific terminology that creates obstacles for language models. In this work, we propose the SAE-FiRE (Sparse Autoencoder for Financial Representation Enhancement) framework to address these limitations by extracting key information while eliminating redundancy. SAE-FiRE employs Sparse Autoencoders (SAEs) to decompose dense neural representations from large language models into interpretable sparse components, then applies statistical feature selection methods, including ANOVA F-tests and tree-based importance scoring, to identify the top-k most discriminative dimensions for classification. By systematically filtering out noise that might otherwise lead to overfitting, we enable more robust and generalizable predictions. Experimental results across three financial datasets demonstrate that SAE-FiRE significantly outperforms baseline approaches.
♻ ☆ Unifying Inference-Time Planning Language Generation
A line of work in planning uses LLM not to generate a plan, but to generate a formal representation in some planning language, which can be input into a symbolic solver to deterministically find a plan. While showing improved trust and promising performance, dozens of recent publications have proposed scattered methods on a variety of benchmarks under different experimental settings. We attempt to unify the inference-time LLM-as-formalizer methodology for classical planning by proposing a unifying framework based on intermediate representations. We thus systematically evaluate more than a dozen pipelines that subsume most existing work, while proposing novel ones that involve syntactically similar but high resource intermediate languages (such as a Python wrapper of PDDL). We provide recipes for planning language generation pipelines, draw a series of conclusions showing the efficacy of their various components, and evidence their robustness against problem complexity.
♻ ☆ Fine-Grained and Thematic Evaluation of LLMs in Social Deduction Game
Recent studies have investigated whether large language models (LLMs) can support obscured communication, which is characterized by core aspects such as inferring subtext and evading suspicions. To conduct the investigation, researchers have used social deduction games (SDGs) as their experimental environment, in which players conceal and infer specific information. However, prior work has often overlooked how LLMs should be evaluated in such settings. Specifically, we point out two limitations with the evaluation methods they employed. First, metrics used in prior studies are coarse-grained as they are based on overall game outcomes that often fail to capture event-level behaviors; Second, error analyses have lacked structured methodologies capable of producing insights that meaningfully support evaluation outcomes. To address these limitations, we propose a microscopic and systematic approach to the investigation. Specifically, we introduce six fine-grained metrics that resolve the first issue. To tackle the second issue, we conducted a thematic analysis and identified four major reasoning failures that undermine LLMs' performance in obscured communication.
comment: Published in IEEE Access
♻ ☆ AgriGPT-VL: Agricultural Vision-Language Understanding Suite
Despite rapid advances in multimodal large language models, agricultural applications remain constrained by the scarcity of domain-tailored models, curated vision-language corpora, and rigorous evaluation. To address these challenges, we present the AgriGPT-VL Suite, a unified multimodal framework for agriculture. Our contributions are threefold. First, we introduce Agri-3M-VL, the largest vision-language corpus for agriculture to our knowledge, curated by a scalable multi-agent data generator; it comprises 1M image-caption pairs, 2M image-grounded VQA pairs, 50K expert-level VQA instances, and 15K GRPO reinforcement learning samples. Second, we develop AgriGPT-VL, an agriculture-specialized vision-language model trained via a progressive curriculum of textual grounding, multimodal shallow/deep alignment, and GRPO refinement. This method achieves strong multimodal reasoning while preserving text-only capability. Third, we establish AgriBench-VL-4K, a compact yet challenging evaluation suite with open-ended and image-grounded questions, paired with multi-metric evaluation and an LLM-as-a-judge framework. Experiments show that AgriGPT-VL outperforms leading general-purpose VLMs on AgriBench-VL-4K, achieving higher pairwise win rates in the LLM-as-a-judge evaluation. Meanwhile, it remains competitive on the text-only AgriBench-13K with no noticeable degradation of language ability. Ablation studies further confirm consistent gains from our alignment and GRPO refinement stages. We will open source all of the resources to support reproducible research and deployment in low-resource agricultural settings.
♻ ☆ GLiDRE: Generalist Lightweight model for Document-level Relation Extraction
Relation Extraction (RE) is a fundamental task in Natural Language Processing, and its document-level variant poses significant challenges, due to complex interactions between entities across sentences. While supervised models have achieved strong results in fully resourced settings, their behavior with limited training data remains insufficiently studied. We introduce GLiDRE, a new compact model for document-level relation extraction, designed to work efficiently in both supervised and few-shot settings. Experiments in both low-resource supervised training and few-shot meta-learning benchmarks show that our approach outperforms existing methods in data-constrained scenarios, establishing a new state-of-the-art in few-shot document-level relation extraction. Our code will be publicly available.
comment: Submitted to ARR October
♻ ☆ CAPO: Towards Enhancing LLM Reasoning through Generative Credit Assignment
Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning abilities of Large Language Models (LLMs) by using rule-based binary feedback. However, current RLVR methods typically assign the same reward to every token. This coarse-grained feedback hampers precise credit assignment, making it hard for models to identify which reasoning steps lead to success or failure, and often results in suboptimal policies. Methods like PPO provide credit assignment by value estimation, but yield inaccurate and unverifiable signals due to limited sampling. On the other hand, methods using Process Reward Models can provide step-wise rewards but suffer from several key limitations: they require high-quality process supervision labels, the feedback is unreliable due to probabilistic reward modeling, and their application in online reinforcement learning (RL) is time-consuming. To overcome these limitations, we introduce a simple but efficient method-Credit Assignment Policy Optimization (CAPO). Instead of training auxiliary models, CAPO directly leverages an off-the-shelf, general-purpose LLM as a Generative Process Reward Model (LLM-as-GenPRM) to generate all step-wise critique by one pass only based on the correctness of the step itself, providing deterministic token-level credits to refine the tokens that were originally assigned identical rule-based rewards. To further enhance the accuracy and robustness, we employ voting mechanisms that scale with the number of generated critiques. Extensive experiments on various backbones like Llama and Qwen models show that CAPO consistently outperforms supervised learning-based and RL-based fine-tuning methods across four challenging mathematical benchmarks and three out-of-domain benchmarks. Further analysis shows that CAPO can help the model to foster the learning of correct reasoning pathways leading to correct answers.
comment: Work in progress
♻ ☆ MASRAD: Arabic Terminology Management Corpora with Semi-Automatic Construction
This paper presents MASRAD, a terminology dataset for Arabic terminology management, and a method with supporting tools for its semi-automatic construction. The entries in MASRAD are $(f,a)$ pairs of foreign (non-Arabic) terms $f$, appearing in specialized, academic and field-specific books next to their Arabic $a$ counterparts. MASRAD-Ex systematically extracts these pairs as a first step to construct MASRAD. MASRAD helps improving term consistency in academic translations and specialized Arabic documents, and automating cross-lingual text processing. MASRAD-Ex leverages translated terms organically occurring in Arabic books, and considers several candidate pairs for each term phrase. The candidate Arabic terms occur next to the foreign terms, and vary in length. MASRAD-Ex computes lexicographic, phonetic, morphological, and semantic similarity metrics for each candidate pair, and uses heuristic, machine learning, and machine learning with post-processing approaches to decide on the best candidate. This paper presents MASRAD after thorough expert review and makes it available to the interested research community. The best performing MASRAD-Ex approach achieved 90.5% precision and 92.4% recall.
♻ ☆ Teaching Small Language Models to Learn Logic through Meta-Learning
Large language models (LLMs) are increasingly evaluated on reasoning tasks, yet their logical abilities remain contested. To address this, we study LLMs' reasoning in a well-defined fragment of logic: syllogistic reasoning. We cast the problem as premise selection and construct controlled datasets to isolate logical competence. Beyond evaluation, an open challenge is enabling LLMs to acquire abstract inference patterns that generalize to novel structures. We propose to apply few-shot meta-learning to this domain, thereby encouraging models to extract rules across tasks rather than memorize patterns within tasks. Although meta-learning has been little explored in the context of logic learnability, our experiments show that it is effective: small models (1.5B-7B) fine-tuned with meta-learning demonstrate strong gains in generalization, with especially pronounced benefits in low-data regimes. These meta-learned models outperform GPT-4o and o3-mini on our syllogistic reasoning task.
♻ ☆ Building Resource-Constrained Language Agents: A Korean Case Study on Chemical Toxicity Information EMNLP 2025
Language agents powered by large language models (LLMs) face significant deployment challenges in resource-constrained environments, particularly for specialized domains and less-common languages. This paper presents Tox-chat, a Korean chemical toxicity information agent devised within these limitations. We propose two key innovations: a context-efficient architecture that reduces token consumption through hierarchical section search, and a scenario-based dialogue generation methodology that effectively distills tool-using capabilities from larger models. Experimental evaluations demonstrate that our fine-tuned 8B parameter model substantially outperforms both untuned models and baseline approaches, in terms of DB faithfulness and preference. Our work offers valuable insights for researchers developing domain-specific language agents under practical constraints.
comment: EMNLP 2025 Industry track
♻ ☆ FAID: Fine-Grained AI-Generated Text Detection Using Multi-Task Auxiliary and Multi-Level Contrastive Learning
The growing collaboration between humans and AI models in generative tasks has introduced new challenges in distinguishing between human-written, LLM-generated, and human--LLM collaborative texts. In this work, we collect a multilingual, multi-domain, multi-generator dataset FAIDSet. We further introduce a fine-grained detection framework FAID to classify text into these three categories, and also to identify the underlying LLM family of the generator. Unlike existing binary classifiers, FAID is built to capture both authorship and model-specific characteristics. Our method combines multi-level contrastive learning with multi-task auxiliary classification to learn subtle stylistic cues. By modeling LLM families as distinct stylistic entities, we incorporate an adaptation to address distributional shifts without retraining for unseen data. Our experimental results demonstrate that FAID outperforms several baselines, particularly enhancing the generalization accuracy on unseen domains and new LLMs, thus offering a potential solution for improving transparency and accountability in AI-assisted writing.
♻ ☆ Contrastive Learning Using Graph Embeddings for Domain Adaptation of Language Models in the Process Industry EMNLP 2025
Recent trends in NLP utilize knowledge graphs (KGs) to enhance pretrained language models by incorporating additional knowledge from the graph structures to learn domain-specific terminology or relationships between documents that might otherwise be overlooked. This paper explores how SciNCL, a graph-aware neighborhood contrastive learning methodology originally designed for scientific publications, can be applied to the process industry domain, where text logs contain crucial information about daily operations and are often structured as sparse KGs. Our experiments demonstrate that language models fine-tuned with triplets derived from graph embeddings (GE) outperform a state-of-the-art mE5-large text encoder by 9.8-14.3% (5.45-7.96p) on the proprietary process industry text embedding benchmark (PITEB) while having 3 times fewer parameters.
comment: accepted to EMNLP 2025 (industry track)
♻ ☆ Cross-Document Cross-Lingual NLI via RST-Enhanced Graph Fusion and Interpretability Prediction EMNLP 2025
Natural Language Inference (NLI) is a fundamental task in natural language processing. While NLI has developed many sub-directions such as sentence-level NLI, document-level NLI and cross-lingual NLI, Cross-Document Cross-Lingual NLI (CDCL-NLI) remains largely unexplored. In this paper, we propose a novel paradigm: CDCL-NLI, which extends traditional NLI capabilities to multi-document, multilingual scenarios. To support this task, we construct a high-quality CDCL-NLI dataset including 25,410 instances and spanning 26 languages. To address the limitations of previous methods on CDCL-NLI task, we further propose an innovative method that integrates RST-enhanced graph fusion with interpretability-aware prediction. Our approach leverages RST (Rhetorical Structure Theory) within heterogeneous graph neural networks for cross-document context modeling, and employs a structure-aware semantic alignment based on lexical chains for cross-lingual understanding. For NLI interpretability, we develop an EDU (Elementary Discourse Unit)-level attribution framework that produces extractive explanations. Extensive experiments demonstrate our approach's superior performance, achieving significant improvements over both conventional NLI models as well as large language models. Our work sheds light on the study of NLI and will bring research interest on cross-document cross-lingual context understanding, hallucination elimination and interpretability inference. Our code and datasets are available at "https://github.com/Leonardo123-ui/CDCL_NLI" for peer review.
comment: EMNLP 2025 Main (Camera Ready)
♻ ☆ Towards Locally Deployable Fine-Tuned Causal Large Language Models for Mode Choice Behaviour
This study investigates the adoption of open-access, locally deployable causal large language models (LLMs) for travel mode choice prediction and introduces LiTransMC, the first fine-tuned causal LLM developed for this task. We systematically benchmark eleven open-access LLMs (1-12B parameters) across three stated and revealed preference datasets, testing 396 configurations and generating over 79,000 mode choice decisions. Beyond predictive accuracy, we evaluate models generated reasoning using BERTopic for topic modelling and a novel Explanation Strength Index, providing the first structured analysis of how LLMs articulate decision factors in alignment with behavioural theory. LiTransMC, fine-tuned using parameter efficient and loss masking strategy, achieved a weighted F1 score of 0.6845 and a Jensen-Shannon Divergence of 0.000245, surpassing both untuned local models and larger proprietary systems, including GPT-4o with advanced persona inference and embedding-based loading, while also outperforming classical mode choice methods such as discrete choice models and machine learning classifiers for the same dataset. This dual improvement, i.e., high instant-level accuracy and near-perfect distributional calibration, demonstrates the feasibility of creating specialist, locally deployable LLMs that integrate prediction and interpretability. Through combining structured behavioural prediction with natural language reasoning, this work unlocks the potential for conversational, multi-task transport models capable of supporting agent-based simulations, policy testing, and behavioural insight generation. These findings establish a pathway for transforming general purpose LLMs into specialized and explainable tools for transportation research and policy formulation, while maintaining privacy, reducing cost, and broadening access through local deployment.
♻ ☆ Aligning Language Models with Real-time Knowledge Editing
Knowledge editing aims to modify outdated knowledge in large language models (LLMs) efficiently while retaining their original capabilities. Mainstream benchmarks for knowledge editing are predominantly static and fail to keep in pace with the evolving real-world knowledge. In this work, we introduce CRAFT, an ever-evolving real-world benchmark for knowledge editing. It features well-designed paired edits for composite reasoning, and evaluates models on alias portability as well as temporal and common-sense locality, making it a challenging knowledge editing benchmark on which previous knowledge editing methods hardly achieve balanced performance. Towards flexible real-time editing, we propose KEDAS, a novel paradigm of knowledge editing alignment featuring diverse edit augmentation and self-adaptive post-alignment inference, which exhibits significant performance gain on CRAFT compared to previous methods. All of our code and data are available at https://anonymous.4open.science/r/CRAFT-KEDAS.
comment: Pre-print
♻ ☆ Fair Play in the Newsroom: Actor-Based Filtering Gender Discrimination in Text Corpora
Language corpora are the foundation of most natural language processing research, yet they often reproduce structural inequalities. One such inequality is gender discrimination in how actors are represented, which can distort analyses and perpetuate discriminatory outcomes. This paper introduces a user-centric, actor-level pipeline for detecting and mitigating gender discrimination in large-scale text corpora. By combining discourse-aware analysis with metrics for sentiment, syntactic agency, and quotation styles, our method enables both fine-grained auditing and exclusion-based balancing. Applied to the taz2024full corpus of German newspaper articles (1980-2024), the pipeline yields a more gender-balanced dataset while preserving core dynamics of the source material. Our findings show that structural asymmetries can be reduced through systematic filtering, though subtler biases in sentiment and framing remain. We release the tools and reports to support further research in discourse-based fairness auditing and equitable corpus construction.
♻ ☆ WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research
This paper tackles \textbf{open-ended deep research (OEDR)}, a complex challenge where AI agents must synthesize vast web-scale information into insightful reports. Current approaches are plagued by dual-fold limitations: static research pipelines that decouple planning from evidence acquisition and monolithic generation paradigms that include redundant, irrelevant evidence, suffering from hallucination issues and low citation accuracy. To address these challenges, we introduce \textbf{WebWeaver}, a novel dual-agent framework that emulates the human research process. The planner operates in a dynamic cycle, iteratively interleaving evidence acquisition with outline optimization to produce a comprehensive, citation-grounded outline linking to a memory bank of evidence. The writer then executes a hierarchical retrieval and writing process, composing the report section by section. By performing targeted retrieval of only the necessary evidence from the memory bank via citations for each part, it effectively mitigates long-context issues and citation hallucinations. Our framework establishes a new state-of-the-art across major OEDR benchmarks, including DeepResearch Bench, DeepConsult, and DeepResearchGym. These results validate our human-centric, iterative methodology, demonstrating that adaptive planning and focused synthesis are crucial for producing comprehensive, trusted, and well-structured reports.
comment: An agent system for open-ended deep research
♻ ☆ An Embarrassingly Simple Defense Against LLM Abliteration Attacks
Large language models (LLMs) are typically aligned to refuse harmful instructions through safety fine-tuning. A recent attack, termed abliteration, identifies and suppresses the single latent direction most responsible for refusal behavior, thereby enabling models to generate harmful content. We propose a defense that fundamentally alters how models express refusal. We construct an extended-refusal dataset in which responses to harmful prompts provide detailed justifications before refusing, distributing the refusal signal across multiple token positions. Fine-tuning Llama-2-7B-Chat and Qwen2.5-Instruct (1.5B and 3B parameters) on this dataset yields models that maintain high refusal rates under abliteration: refusal rates drop by at most 10%, compared to 70-80% drops in baseline models. Comprehensive evaluations of safety and utility demonstrate that extended-refusal fine-tuning effectively neutralizes abliteration attacks while preserving general model performance and enhancing robustness across multiple alignment scenarios.
comment: preprint - under review
♻ ☆ SciKnowEval: Evaluating Multi-level Scientific Knowledge of Large Language Models
Large language models (LLMs) are playing an increasingly important role in scientific research, yet there remains a lack of comprehensive benchmarks to evaluate the breadth and depth of scientific knowledge embedded in these models. To address this gap, we introduce SciKnowEval, a large-scale dataset designed to systematically assess LLMs across five progressive levels of scientific understanding: memory, comprehension, reasoning, discernment, and application. SciKnowEval comprises 28K multi-level questions and solutions spanning biology, chemistry, physics, and materials science. Using this benchmark, we evaluate 20 leading open-source and proprietary LLMs. The results show that while proprietary models often achieve state-of-the-art performance, substantial challenges remain -- particularly in scientific reasoning and real-world application. We envision SciKnowEval as a standard benchmark for evaluating scientific capabilities in LLMs and as a catalyst for advancing more capable and reliable scientific language models.
comment: 33 pages, 2 figures
♻ ☆ SAFER: Advancing Safety Alignment via Efficient Ex-Ante Reasoning
Recent advancements in large language models (LLMs) have accelerated progress toward artificial general intelligence, yet their potential to generate harmful content poses critical safety challenges. Existing alignment methods often struggle to cover diverse safety scenarios and remain vulnerable to adversarial attacks. In this work, we propose SAFER, a framework for Safety Alignment via eFficient Ex-Ante Reasoning. Our approach instantiates structured Ex-Ante reasoning through initial assessment, rule verification, and path calibration, and embeds predefined safety rules to provide transparent and verifiable safety judgments. Specifically, our approach consists of two training stages: (1) supervised fine-tuning with synthetic traces to teach the multi-stage Ex-Ante reasoning, and (2) step-level reasoning preference optimization to jointly enhance safety, utility, and efficiency. Experiments on multiple open-source LLMs demonstrate that SAFER significantly enhances safety performance while maintaining helpfulness and response efficiency.
comment: 22 pages, 5 figures
♻ ☆ MMReview: A Multidisciplinary and Multimodal Benchmark for LLM-Based Peer Review Automation
With the rapid growth of academic publications, peer review has become an essential yet time-consuming responsibility within the research community. Large Language Models (LLMs) have increasingly been adopted to assist in the generation of review comments; however, current LLM-based review tasks lack a unified evaluation benchmark to rigorously assess the models' ability to produce comprehensive, accurate, and human-aligned assessments, particularly in scenarios involving multimodal content such as figures and tables. To address this gap, we propose \textbf{MMReview}, a comprehensive benchmark that spans multiple disciplines and modalities. MMReview includes multimodal content and expert-written review comments for 240 papers across 17 research domains within four major academic disciplines: Artificial Intelligence, Natural Sciences, Engineering Sciences, and Social Sciences. We design a total of 13 tasks grouped into four core categories, aimed at evaluating the performance of LLMs and Multimodal LLMs (MLLMs) in step-wise review generation, outcome formulation, alignment with human preferences, and robustness to adversarial input manipulation. Extensive experiments conducted on 16 open-source models and 5 advanced closed-source models demonstrate the thoroughness of the benchmark. We envision MMReview as a critical step toward establishing a standardized foundation for the development of automated peer review systems.
comment: Work in progress
♻ ☆ When to use Graphs in RAG: A Comprehensive Analysis for Graph Retrieval-Augmented Generation
Graph retrieval-augmented generation (GraphRAG) has emerged as a powerful paradigm for enhancing large language models (LLMs) with external knowledge. It leverages graphs to model the hierarchical structure between specific concepts, enabling more coherent and effective knowledge retrieval for accurate reasoning.Despite its conceptual promise, recent studies report that GraphRAG frequently underperforms vanilla RAG on many real-world tasks. This raises a critical question: Is GraphRAG really effective, and in which scenarios do graph structures provide measurable benefits for RAG systems? To address this, we propose GraphRAG-Bench, a comprehensive benchmark designed to evaluate GraphRAG models onboth hierarchical knowledge retrieval and deep contextual reasoning. GraphRAG-Bench features a comprehensive dataset with tasks of increasing difficulty, coveringfact retrieval, complex reasoning, contextual summarization, and creative generation, and a systematic evaluation across the entire pipeline, from graph constructionand knowledge retrieval to final generation. Leveraging this novel benchmark, we systematically investigate the conditions when GraphRAG surpasses traditional RAG and the underlying reasons for its success, offering guidelines for its practical application. All related resources and analyses are collected for the community at https://github.com/GraphRAG-Bench/GraphRAG-Benchmark.
comment: All resources and analyses are collected at https://github.com/GraphRAG-Bench/GraphRAG-Benchmark
♻ ☆ AVerImaTeC: A Dataset for Automatic Verification of Image-Text Claims with Evidence from the Web NeurIPS 2025
Textual claims are often accompanied by images to enhance their credibility and spread on social media, but this also raises concerns about the spread of misinformation. Existing datasets for automated verification of image-text claims remain limited, as they often consist of synthetic claims and lack evidence annotations to capture the reasoning behind the verdict. In this work, we introduce AVerImaTeC, a dataset consisting of 1,297 real-world image-text claims. Each claim is annotated with question-answer (QA) pairs containing evidence from the web, reflecting a decomposed reasoning regarding the verdict. We mitigate common challenges in fact-checking datasets such as contextual dependence, temporal leakage, and evidence insufficiency, via claim normalization, temporally constrained evidence annotation, and a two-stage sufficiency check. We assess the consistency of the annotation in AVerImaTeC via inter-annotator studies, achieving a $\kappa=0.742$ on verdicts and $74.7\%$ consistency on QA pairs. We also propose a novel evaluation method for evidence retrieval and conduct extensive experiments to establish baselines for verifying image-text claims using open-web evidence.
comment: accepted at NeurIPS 2025 Datasets and Benchmarks Track
♻ ☆ WildIFEval: Instruction Following in the Wild
Recent LLMs have shown remarkable success in following user instructions, yet handling instructions with multiple constraints remains a significant challenge. In this work, we introduce WildIFEval - a large-scale dataset of 7K real user instructions with diverse, multi-constraint conditions. Unlike prior datasets, our collection spans a broad lexical and topical spectrum of constraints, extracted from natural user instructions. We categorize these constraints into eight high-level classes to capture their distribution and dynamics in real-world scenarios. Leveraging WildIFEval, we conduct extensive experiments to benchmark the instruction-following capabilities of leading LLMs. WildIFEval clearly differentiates between small and large models, and demonstrates that all models have a large room for improvement on such tasks. We analyze the effects of the number and type of constraints on performance, revealing interesting patterns of model constraint-following behavior. We release our dataset to promote further research on instruction-following under complex, realistic conditions.
♻ ☆ Text Clustering as Classification with LLMs
Text clustering serves as a fundamental technique for organizing and interpreting unstructured textual data, particularly in contexts where manual annotation is prohibitively costly. With the rapid advancement of Large Language Models (LLMs) and their demonstrated effectiveness across a broad spectrum of NLP tasks, an emerging body of research has begun to explore their potential in the domain of text clustering. However, existing LLM-based approaches still rely on fine-tuned embedding models and sophisticated similarity metrics, rendering them computationally intensive and necessitating domain-specific adaptation. To address these limitations, we propose a novel framework that reframes text clustering as a classification task by harnessing the in-context learning capabilities of LLMs. Our framework eliminates the need for fine-tuning embedding models or intricate clustering algorithms. It comprises two key steps: first, the LLM is prompted to generate a set of candidate labels based on the dataset and then merges semantically similar labels; second, it assigns the most appropriate label to each text sample. By leveraging the advanced natural language understanding and generalization capabilities of LLMs, the proposed approach enables effective clustering with minimal human intervention. Experimental results on diverse datasets demonstrate that our framework achieves comparable or superior performance to state-of-the-art embedding-based clustering techniques, while significantly reducing computational complexity and resource requirements. These findings underscore the transformative potential of LLMs in simplifying and enhancing text clustering tasks. We make our code available to the public for utilization at https://github.com/ECNU-Text-Computing/Text-Clustering-via-LLM. We also provide the supplementary Appendix within the repository.
comment: 11 pages, 3 figures
♻ ☆ Geometry-Guided Adversarial Prompt Detection via Curvature and Local Intrinsic Dimension
Adversarial prompts are capable of jailbreaking frontier large language models (LLMs) and inducing undesirable behaviours, posing a significant obstacle to their safe deployment. Current mitigation strategies primarily rely on activating built-in defence mechanisms or fine-tuning LLMs, both of which are computationally expensive and can sacrifice model utility. In contrast, detection-based approaches are more efficient and practical for deployment in real-world applications. However, the fundamental distinctions between adversarial and benign prompts remain poorly understood. In this work, we introduce CurvaLID, a novel defence framework that efficiently detects adversarial prompts by leveraging their geometric properties. It is agnostic to the type of LLM, offering a unified detection framework across diverse adversarial prompts and LLM architectures. CurvaLID builds on the geometric analysis of text prompts to uncover their underlying differences. We theoretically extend the concept of curvature via the Whewell equation into an $n$-dimensional word embedding space, enabling us to quantify local geometric properties, including semantic shifts and curvature in the underlying manifolds. To further enhance our solution, we leverage Local Intrinsic Dimensionality (LID) to capture complementary geometric features of text prompts within adversarial subspaces. Our findings show that adversarial prompts exhibit distinct geometric signatures from benign prompts, enabling CurvaLID to achieve near-perfect classification and outperform state-of-the-art detectors in adversarial prompt detection. CurvaLID provides a reliable and efficient safeguard against malicious queries as a model-agnostic method that generalises across multiple LLMs and attack families.
comment: 40 Pages, 6 figues
♻ ☆ The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs?
Contrastive decoding strategies are widely used to reduce object hallucinations in multimodal large language models (MLLMs). These methods work by constructing contrastive samples to induce hallucinations and then suppressing them in the output distribution. However, this paper demonstrates that such approaches fail to effectively mitigate the hallucination problem. The performance improvements observed on POPE Benchmark are largely driven by two misleading factors: (1) crude, unidirectional adjustments to the model's output distribution and (2) the adaptive plausibility constraint, which reduces the sampling strategy to greedy search. To further illustrate these issues, we introduce a series of spurious improvement methods and evaluate their performance against contrastive decoding techniques. Experimental results reveal that the observed performance gains in contrastive decoding are entirely unrelated to its intended goal of mitigating hallucinations. Our findings challenge common assumptions about the effectiveness of contrastive decoding strategies and pave the way for developing genuinely effective solutions to hallucinations in MLLMs.
♻ ☆ Bridging the Culture Gap: A Framework for LLM-Driven Socio-Cultural Localization of Math Word Problems in Low-Resource Languages
Large language models (LLMs) have demonstrated significant capabilities in solving mathematical problems expressed in natural language. However, multilingual and culturally-grounded mathematical reasoning in low-resource languages lags behind English due to the scarcity of socio-cultural task datasets that reflect accurate native entities such as person names, organization names, and currencies. Existing multilingual benchmarks are predominantly produced via translation and typically retain English-centric entities, owing to the high cost associated with human annotater-based localization. Moreover, automated localization tools are limited, and hence, truly localized datasets remain scarce. To bridge this gap, we introduce a framework for LLM-driven cultural localization of math word problems that automatically constructs datasets with native names, organizations, and currencies from existing sources. We find that translated benchmarks can obscure true multilingual math ability under appropriate socio-cultural contexts. Through extensive experiments, we also show that our framework can help mitigate English-centric entity bias and improves robustness when native entities are introduced across various languages.
♻ ☆ Context Biasing for Pronunciations-Orthography Mismatch in Automatic Speech Recognition
Neural sequence-to-sequence systems deliver state-of-the-art performance for automatic speech recognition. When using appropriate modeling units, e.g., byte-pair encoded characters, these systems are in principal open vocabulary systems. In practice, however, they often fail to recognize words not seen during training, e.g., named entities, acronyms, or domain-specific special words. To address this problem, many context biasing methods have been proposed; however, for words with a pronunciation-orthography mismatch, these methods may still struggle. We propose a method which allows corrections of substitution errors to improve the recognition accuracy of such challenging words. Users can add corrections on the fly during inference. We show that with this method we get a relative improvement in biased word error rate of up to 8%, while maintaining a competitive overall word error rate.
♻ ☆ From Accuracy to Robustness: A Study of Rule- and Model-based Verifiers in Mathematical Reasoning
Trustworthy verifiers are essential for the success of reinforcement learning with verifiable reward (RLVR), which is the core methodology behind various large reasoning models such as DeepSeek-R1. In complex domains like mathematical reasoning, rule-based verifiers have been widely adopted in previous works to train strong reasoning models. However, the reliability of these verifiers and their impact on the RL training process remain poorly understood. In this work, we take mathematical reasoning as a case study and conduct a comprehensive analysis of various verifiers in both static evaluation and RL training scenarios. First, we find that current open-source rule-based verifiers often fail to recognize equivalent answers presented in different formats across multiple commonly used mathematical datasets, resulting in non-negligible false negative rates. This limitation adversely affects RL training performance and becomes more pronounced as the policy model gets stronger. Subsequently, we investigate model-based verifiers as a potential solution to address these limitations. While the static evaluation shows that model-based verifiers achieve significantly higher verification accuracy, further analysis and RL results imply that they are highly susceptible to hacking, where they misclassify certain patterns in responses as correct, particularly after fine-tuning. This vulnerability is exploited during policy model optimization, leading to artificially inflated rewards. Our findings underscore the unique challenges inherent to both rule-based and model-based verifiers and provide insights toward developing more accurate and robust reward systems for reinforcement learning.
♻ ☆ RooseBERT: A New Deal For Political Language Modelling
The increasing amount of political debates and politics-related discussions calls for the definition of novel computational methods to automatically analyse such content with the final goal of lightening up political deliberation to citizens. However, the specificity of the political language and the argumentative form of these debates (employing hidden communication strategies and leveraging implicit arguments) make this task very challenging, even for current general-purpose pre-trained Language Models. To address this issue, we introduce a novel pre-trained Language Model for political discourse language called RooseBERT. Pre-training a language model on a specialised domain presents different technical and linguistic challenges, requiring extensive computational resources and large-scale data. RooseBERT has been trained on large political debate and speech corpora (8K debates, each composed of several sub-debates on different topics) in English. To evaluate its performances, we fine-tuned it on four downstream tasks related to political debate analysis, i.e., stance detection, sentiment analysis, argument component detection and classification, and argument relation prediction and classification. Our results demonstrate significant improvements over general-purpose Language Models on these four tasks, highlighting how domain-specific pre-training enhances performance in political debate analysis. We release RooseBERT for the research community.
♻ ☆ VisRet: Visualization Improves Knowledge-Intensive Text-to-Image Retrieval
Text-to-image retrieval (T2I retrieval) remains challenging because cross-modal embeddings often behave as bags of concepts and underrepresent structured visual relationships such as pose and viewpoint. We propose Visualize-then-Retrieve (VisRet), a new paradigm for T2I retrieval that mitigates this limitation of cross-modal similarity alignment. VisRet first projects textual queries into the image modality via T2I generation. Then, it performs retrieval within the image modality to bypass the weaknesses of cross-modal retrievers in recognizing subtle visual-spatial features. Across four benchmarks (Visual-RAG, INQUIRE-Rerank, Microsoft COCO, and our new Visual-RAG-ME featuring multi-entity comparisons), VisRet substantially outperforms cross-modal similarity matching and baselines that recast T2I retrieval as text-to-text similarity matching, improving nDCG@30 by 0.125 on average with CLIP as the retriever and by 0.121 with E5-V. For downstream question answering, VisRet increases accuracy on Visual-RAG and Visual-RAG-ME by 3.8% and 15.7% in top-1 retrieval, and by 3.9% and 11.1% in top-10 retrieval. Ablation studies show compatibility with different T2I instruction LLMs, T2I generation models, and downstream LLMs. VisRet provides a practical and principled path that energizes further advances in vision-language retrieval. Our code and the Visual-RAG-ME benchmark will be publicly released.
♻ ☆ Self-Routing RAG: Binding Selective Retrieval with Knowledge Verbalization
Selective retrieval improves the accuracy and efficiency of retrieval-augmented generation (RAG) by reducing distractions from low-quality retrievals. However, existing approaches underutilize the inherent knowledge of large language models (LLMs), leading to suboptimal retrieval decisions and degraded generation performance. To bridge this gap, we propose Self-Routing RAG (SR-RAG), a novel framework that binds selective retrieval with knowledge verbalization. SR-RAG enables an LLM to dynamically decide whether to retrieve external knowledge or verbalize its own parametric knowledge. To this end, we design a multi-task objective that jointly optimizes an LLM for knowledge source selection, knowledge verbalization, and response generation. SR-RAG further incorporates a nearest neighbor search mechanism at inference time to improve the accuracy of knowledge source decisions under domain shifts. Fine-tuning three LLMs with SR-RAG significantly improves both their response accuracy and reduces the inference latency. Compared to the strongest selective retrieval baseline, SR-RAG reduces the number of retrievals by 29% while improving performance by 5.1%.
♻ ☆ RepIt: Representing Isolated Targets to Steer Language Models
While activation steering in large language models (LLMs) is a growing area of research, methods can often incur broader effects than desired. This motivates isolation of purer concept vectors to enable targeted interventions and understand LLM behavior at a more granular level. We present RepIt, a simple and data-efficient framework for isolating concept-specific representations. Across five frontier LLMs, RepIt enables precise interventions: it selectively suppresses refusal on targeted concepts while preserving refusal elsewhere, producing models that answer WMD-related questions while still scoring as safe on standard benchmarks. We further show that the corrective signal localizes to just 100-200 neurons and that robust target representations can be extracted from as few as a dozen examples on a single A6000. This efficiency raises a dual concern: manipulations can be performed with modest compute and data to extend to underrepresented data-scarce topics while evading existing benchmarks. By disentangling refusal vectors with RepIt, this work demonstrates that targeted interventions can counteract overgeneralization, laying the foundation for more granular control of model behavior.
♻ ☆ BenchAgents: Multi-Agent Systems for Structured Benchmark Creation
Evaluation insights are limited by the availability of high-quality benchmarks. As models evolve, there is a need to create benchmarks that can measure progress on new and complex generative capabilities. However, manually creating new benchmarks is slow and expensive, restricting comprehensive evaluations for any capability. We introduce BenchAgents, a multi-agent framework that methodically leverages large language models (LLMs) to automate evaluation benchmark creation while inherently ensuring data and (evaluation) metric quality. BenchAgents decomposes the benchmark creation process into planning, generation, verification, and evaluation, each of which is ] orchestrated via LLM agents. These agents interact with each other and utilize feedback from benchmark developers to improve and flexibly control data diversity and quality. We use BenchAgents to create benchmarks to evaluate capabilities related to planning, constraint satisfaction, and causal reasoning spanning both language and vision modalities. We then use these benchmarks to study state-of-the-art models and extract new insights into common failure modes and model differences.
♻ ☆ HEALTH-PARIKSHA: Assessing RAG Models for Health Chatbots in Real-World Multilingual Settings
Assessing the capabilities and limitations of large language models (LLMs) has garnered significant interest, yet the evaluation of multiple models in real-world scenarios remains rare. Multilingual evaluation often relies on translated benchmarks, which typically do not capture linguistic and cultural nuances present in the source language. This study provides an extensive assessment of 24 LLMs on real world data collected from Indian patients interacting with a medical chatbot in Indian English and 4 other Indic languages. We employ a uniform Retrieval Augmented Generation framework to generate responses, which are evaluated using both automated techniques and human evaluators on four specific metrics relevant to our application. We find that models vary significantly in their performance and that instruction tuned Indic models do not always perform well on Indic language queries. Further, we empirically show that factual correctness is generally lower for responses to Indic queries compared to English queries. Finally, our qualitative work shows that code-mixed and culturally relevant queries in our dataset pose challenges to evaluated models.
♻ ☆ Speech-Based Cognitive Screening: A Systematic Evaluation of LLM Adaptation Strategies
Over half of US adults with Alzheimer disease and related dementias remain undiagnosed, and speech-based screening offers a scalable detection approach. We compared large language model adaptation strategies for dementia detection using the DementiaBank speech corpus, evaluating nine text-only models and three multimodal audio-text models on recordings from DementiaBank speech corpus. Adaptations included in-context learning with different demonstration selection policies, reasoning-augmented prompting, parameter-efficient fine-tuning, and multimodal integration. Results showed that class-centroid demonstrations achieved the highest in-context learning performance, reasoning improved smaller models, and token-level fine-tuning generally produced the best scores. Adding a classification head substantially improved underperforming models. Among multimodal models, fine-tuned audio-text systems performed well but did not surpass the top text-only models. These findings highlight that model adaptation strategies, including demonstration selection, reasoning design, and tuning method, critically influence speech-based dementia detection, and that properly adapted open-weight models can match or exceed commercial systems.
♻ ☆ FormulaReasoning: A Dataset for Formula-Based Numerical Reasoning
The application of formulas (e.g., physics formulas) is a fundamental human ability in solving numerical reasoning problems. Existing numerical reasoning datasets rarely explicitly state the formulas employed, as their questions often rely on implicit commonsense mathematical knowledge. To address this gap, we introduce FormulaReasoning, a new dataset specifically designed for formula-based numerical reasoning. It consists of 5,324 questions that require numerical calculations grounded in external physics formulas. We provide normalized, fine-grained annotations in both English and Chinese, including formula structures, parameter names, symbols, numerical values, and units-curated through extensive manual effort with LLM-assisted validation to ensure high quality. Additionally, we offer a consolidated formula database to serve as an external knowledge source. We analyze various reasoning approaches on FormulaReasoning, with emphasis on comparative evaluation of different architectural and methodological frameworks. Our assessment includes retrieval-augmented methods, approaches that decompose reasoning into formula generation, parameter extraction, and numerical calculation, as well as optimization techniques using preference data. We identify key challenges in formula-based numerical reasoning that require further investigation across different reasoning paradigms, highlighting opportunities for methodological advancement.
♻ ☆ Explaining GPTs' Schema of Depression: A Machine Behavior Analysis
Use of large language models such as ChatGPT (GPT-4/GPT-5) for mental health support has grown rapidly, emerging as a promising route to assess and help people with mood disorders like depression. However, we have a limited understanding of these language models' schema of mental disorders, that is, how they internally associate and interpret symptoms of such disorders. In this work, we leveraged contemporary measurement theory to decode how GPT-4 and GPT-5 interrelate depressive symptoms, providing an explanation of how LLMs apply what they learn and informing clinical applications. We found that GPT-4 (a) had strong convergent validity with standard instruments and expert judgments $(r = 0.70 - 0.81)$, and (b) behaviorally linked depression symptoms with each other (symptom inter-correlates $r = 0.23 - 0.78$) in accordance with established literature on depression; however, it (c) underemphasized the relationship between $\textit{suicidality}$ and other symptoms while overemphasizing $\textit{psychomotor symptoms}$; and (d) suggested novel hypotheses of symptom mechanisms, for instance, indicating that $\textit{sleep}$ and $\textit{fatigue}$ are broadly influenced by other depressive symptoms, while $\textit{worthlessness/guilt}$ is only tied to $\textit{depressed mood}$. GPT-5 showed a slightly lower convergence with self-report, a difference our machine-behavior analysis makes interpretable through shifts in symptom-symptom relationships. These insights provide an empirical foundation for understanding language models' mental health assessments and demonstrate a generalizable approach for explainability in other models and disorders. Our findings can guide key stakeholders to make informed decisions for effectively situating these technologies in the care system.
comment: 25 pages, 1 table, 4 figures, 1 supplementary table, 5 supplementary figures, 59 references
♻ ☆ AWARE, Beyond Sentence Boundaries: A Contextual Transformer Framework for Identifying Cultural Capital in STEM Narratives
Identifying cultural capital (CC) themes in student reflections can offer valuable insights that help foster equitable learning environments in classrooms. However, themes such as aspirational goals or family support are often woven into narratives, rather than appearing as direct keywords. This makes them difficult to detect for standard NLP models that process sentences in isolation. The core challenge stems from a lack of awareness, as standard models are pre-trained on general corpora, leaving them blind to the domain-specific language and narrative context inherent to the data. To address this, we introduce AWARE, a framework that systematically attempts to improve a transformer model's awareness for this nuanced task. AWARE has three core components: 1) Domain Awareness, adapting the model's vocabulary to the linguistic style of student reflections; 2) Context Awareness, generating sentence embeddings that are aware of the full essay context; and 3) Class Overlap Awareness, employing a multi-label strategy to recognize the coexistence of themes in a single sentence. Our results show that by making the model explicitly aware of the properties of the input, AWARE outperforms a strong baseline by 2.1 percentage points in Macro-F1 and shows considerable improvements across all themes. This work provides a robust and generalizable methodology for any text classification task in which meaning depends on the context of the narrative.
♻ ☆ Evaluating and Mitigating Social Bias for Large Language Models in Open-ended Settings
Current social bias benchmarks for Large Language Models (LLMs) primarily rely on predefined question formats like multiple-choice, limiting their ability to reflect the complexity and open-ended nature of real-world interactions. To close this gap, we extend an existing dataset BBQ (Parrish et al., 2022) to Open-BBQ, a comprehensive framework to evaluate the social bias of LLMs in open-ended settings by incorporating two additional question categories: fill-in-the-blank and short-answer. Since our new Open-BBQ dataset contains a lot of open-ended responses like sentences and paragraphs, we developed an evaluation process to detect biases from open-ended content by labeling sentences and paragraphs. In addition to this, we also found that existing debiasing methods, such as self-debiasing (Gallegos et al., 2024), have over-correction issues, which make the original correct answers incorrect. In order to solve this issue, we propose Composite Prompting, an In-context Learning (ICL) method combining structured examples with explicit chain-of-thought reasoning to form a unified instruction template for LLMs to explicitly identify content that needs debiasing. Experimental results show that the proposed method significantly reduces the bias for both GPT-3.5 and GPT-4o while maintaining high accuracy.
comment: 15 pages
♻ ☆ Robustness of Large Language Models to Perturbations in Text
Having a clean dataset has been the foundational assumption of most natural language processing (NLP) systems. However, properly written text is rarely found in real-world scenarios and hence, oftentimes invalidates the aforementioned foundational assumption. Recently, Large language models (LLMs) have shown impressive performance, but can they handle the inevitable noise in real-world data? This work tackles this critical question by investigating LLMs' resilience against morphological variations in text. To that end, we artificially introduce varying levels of noise into a diverse set of datasets and systematically evaluate LLMs' robustness against the corrupt variations of the original text. Our findings show that contrary to popular beliefs, generative LLMs are quiet robust to noisy perturbations in text. This is a departure from pre-trained models like BERT or RoBERTa whose performance has been shown to be sensitive to deteriorating noisy text. Additionally, we test LLMs' resilience on multiple real-world benchmarks that closely mimic commonly found errors in the wild. With minimal prompting, LLMs achieve a new state-of-the-art on the benchmark tasks of Grammar Error Correction (GEC) and Lexical Semantic Change (LSC). To empower future research, we also release a dataset annotated by humans stating their preference for LLM vs. human-corrected outputs along with the code to reproduce our results.
comment: 8 pages, 1 figure, 6 tables, updated with results also from GPT-4, LLaMa-3
♻ ☆ MediSyn: A Generalist Text-Guided Latent Diffusion Model For Diverse Medical Image Synthesis
Deep learning algorithms require extensive data to achieve robust performance. However, data availability is often restricted in the medical domain due to patient privacy concerns. Synthetic data presents a possible solution to these challenges. Recently, image generative models have found increasing use for medical applications but are often designed for singular medical specialties and imaging modalities, thus limiting their broader utility. To address this, we introduce MediSyn: a text-guided, latent diffusion model capable of generating synthetic images from 6 medical specialties and 10 image types. Through extensive experimentation, we first demonstrate that MediSyn quantitatively matches or surpasses the performance of specialist models. Second, we show that our synthetic images are realistic and exhibit strong alignment with their corresponding text prompts, as validated by a team of expert physicians. Third, we provide empirical evidence that our synthetic images are visually distinct from their corresponding real patient images. Finally, we demonstrate that in data-limited settings, classifiers trained solely on synthetic data or real data supplemented with synthetic data can outperform those trained solely on real data. Our findings highlight the immense potential of generalist image generative models to accelerate algorithmic research and development in medicine.
♻ ☆ Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning NeurIPS 2025
Retrieval-augmented generation (RAG) is widely utilized to incorporate external knowledge into large language models, thereby enhancing factuality and reducing hallucinations in question-answering (QA) tasks. A standard RAG pipeline consists of several components, such as query rewriting, document retrieval, document filtering, and answer generation. However, these components are typically optimized separately through supervised fine-tuning, which can lead to misalignments between the objectives of individual components and the overarching aim of generating accurate answers. Although recent efforts have explored using reinforcement learning (RL) to optimize specific RAG components, these approaches often focus on simple pipelines with only two components or do not adequately address the complex interdependencies and collaborative interactions among the modules. To overcome these limitations, we propose treating the complex RAG pipeline with multiple components as a multi-agent cooperative task, in which each component can be regarded as an RL agent. Specifically, we present MMOA-RAG, Multi-Module joint Optimization Algorithm for RAG, which employs multi-agent reinforcement learning to harmonize all agents' goals toward a unified reward, such as the F1 score of the final answer. Experiments conducted on various QA benchmarks demonstrate that MMOA-RAG effectively boost the overall performance of the pipeline and outperforms existing baselines. Furthermore, comprehensive ablation studies validate the contributions of individual components and demonstrate MMOA-RAG can be adapted to different RAG pipelines and benchmarks.
comment: NeurIPS 2025
♻ ☆ ChartCards: A Chart-Metadata Generation Framework for Multi-Task Chart Understanding
The emergence of Multi-modal Large Language Models (MLLMs) presents new opportunities for chart understanding. However, due to the fine-grained nature of these tasks, applying MLLMs typically requires large, high-quality datasets for task-specific fine-tuning, leading to high data collection and training costs. To address this, we propose ChartCards, a unified chart-metadata generation framework for multi-task chart understanding. ChartCards systematically synthesizes various chart information, including data tables, visualization code, visual elements, and multi-dimensional semantic captions. By structuring this information into organized metadata, ChartCards enables a single chart to support multiple downstream tasks, such as text-to-chart retrieval, chart summarization, chart-to-table conversion, chart description, and chart question answering. Using ChartCards, we further construct MetaChart, a large-scale high-quality dataset containing 10,862 data tables, 85K charts, and 170 K high-quality chart captions. We validate the dataset through qualitative crowdsourcing evaluations and quantitative fine-tuning experiments across various chart understanding tasks. Fine-tuning six different models on MetaChart resulted in an average performance improvement of 5% across all tasks. The most notable improvements are seen in text-to-chart retrieval and chart-to-table tasks, with Long-CLIP and Llama 3.2-11B achieving improvements of 17% and 28%, respectively.
comment: Need to be revised
♻ ☆ Assessing Algorithmic Bias in Language-Based Depression Detection: A Comparison of DNN and LLM Approaches
This paper investigates algorithmic bias in language-based models for automated depression detection, focusing on socio-demographic disparities related to gender and race/ethnicity. Models trained using deep neural networks (DNN) based embeddings are compared to few-shot learning approaches with large language models (LLMs), evaluating both performance and fairness on clinical interview transcripts from the Distress Analysis Interview Corpus/Wizard-of-Oz (DAIC-WOZ). To mitigate bias, fairness-aware loss functions are applied to DNN-based models, while in-context learning with varied prompt framing and shot counts is explored for LLMs. Results indicate that LLMs outperform DNN-based models in depression classification, particularly for underrepresented groups such as Hispanic participants. LLMs also exhibit reduced gender bias compared to DNN-based embeddings, though racial disparities persist. Among fairness-aware techniques for mitigating bias in DNN-based embeddings, the worst-group loss, which is designed to minimize loss for the worst-performing demographic group, achieves a better balance between performance and fairness. In contrast, the fairness-regularized loss minimizes loss across all groups but performs less effectively. In LLMs, guided prompting with ethical framing helps mitigate gender bias in the 1-shot setting. However, increasing the number of shots does not lead to further reductions in disparities. For race/ethnicity, neither prompting strategy nor increasing $N$ in $N$-shot learning effectively reduces disparities.
comment: 7 pages, 1 figure. This paper has been accepted to the IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI 2025), Georgia Institute of Technology, Atlanta, Georgia, October 26-29, 2025
♻ ☆ AtomWorld: A Benchmark for Evaluating Spatial Reasoning in Large Language Models on Crystalline Materials
Large Language Models (LLMs) excel at textual reasoning and are beginning to develop spatial understanding, prompting the question of whether these abilities can be combined for complex, domain-specific tasks. This question is essential in fields like materials science, where deep understanding of 3D atomic structures is fundamental. While initial studies have successfully applied LLMs to tasks involving pure crystal generation or coordinate understandings, a standardized benchmark to systematically evaluate their core reasoning abilities across diverse atomic structures has been notably absent. To address this gap, we introduce the AtomWorld benchmark to evaluate LLMs on tasks based in Crystallographic Information Files (CIFs), a standard structure representation format. These tasks, including structural editing, CIF perception, and property-guided modeling, reveal a critical limitation: current models, despite establishing promising baselines, consistently fail in structural understanding and spatial reasoning. Our experiments show that these models make frequent errors on structure modification tasks, and even in the basic CIF format understandings, potentially leading to cumulative errors in subsequent analysis and materials insights. By defining these standardized tasks, AtomWorld lays the ground for advancing LLMs toward robust atomic-scale modeling, crucial for accelerating materials research and automating scientific workflows.
♻ ☆ LLM Unlearning Without an Expert Curated Dataset
Modern large language models often encode sensitive, harmful, or copyrighted knowledge, raising the need for post-hoc unlearning-the ability to remove specific domains of knowledge from a model without full retraining. A major bottleneck in current unlearning pipelines is constructing effective forget sets-datasets that approximate the target domain and guide the model to forget it. In this work, we introduce a scalable, automated approach to generate high-quality forget sets using language models themselves. Our method synthesizes textbook-style data through a structured prompting pipeline, requiring only a domain name as input. Through experiments on unlearning biosecurity, cybersecurity, and Harry Potter novels, we show that our synthetic datasets consistently outperform the baseline synthetic alternatives and are comparable to the expert-curated ones. Additionally, ablation studies reveal that the multi-step generation pipeline significantly boosts data diversity, which in turn improves unlearning utility. Overall, our findings suggest that synthetic datasets offer a promising path toward practical, scalable unlearning for a wide range of emerging domains without the need for manual intervention. We release our code and dataset at https://github.com/xyzhu123/Synthetic_Textbook.
♻ ☆ Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time ICLR 2026
Language model finetuning often results in learning undesirable traits in combination with desired ones. To address this, we propose inoculation prompting: modifying finetuning data by prepending a short system-prompt instruction that deliberately elicits the undesirable trait. At test time, we evaluate without the instruction; inoculated models have much lower expression of the trait than models trained with unmodified training data. Inoculation is selective: in a toy setting where assistant responses are always in Spanish and ALL-CAPS, an appropriate inoculation (e.g., ``You always speak in Spanish.'') teaches the model to capitalize responses while still responding in English. We find that inoculation is also effective across several additional settings: reducing emergent misalignment (EM) from task-specific finetuning, defending against backdoor injections, and mitigating the transmission of traits via subliminal learning. Follow-up analysis suggests a mechanism: making a trait less surprising via inoculation reduces optimization pressure to globally update the model, thereby reducing the degree of generalization. Our analysis relates to prior work on EM: inoculation explains prior findings that educational contexts mitigate EM from insecure code. Beyond demonstrating a simple and effective technique for selective learning, our results contribute to a better conceptual understanding of how and why language models generalize.
comment: 40 pages, 22 figures In proceedings at ICLR 2026
♻ ☆ Can Video Large Multimodal Models Think Like Doubters-or Double-Down: A Study on Defeasible Video Entailment
Video Large Multimodal Models (VLMMs) have made impressive strides in understanding video content, but they often struggle with abstract and adaptive reasoning-the ability to revise their interpretations when new information emerges. In reality, conclusions are rarely set in stone; additional context can strengthen or weaken an initial inference. To address this, we introduce Defeasible Video Entailment (DVidE), a new task that challenges models to think like doubters, constantly updating their reasoning based on evolving evidence. In DVidE, given a video premise and a textual hypothesis, models must determine whether a new update strengthens or weakens the hypothesis (classification version) or generate a coherent update that modifies the entailment relationship (generation version). For solving the classification task, we propose the Chain of Counterfactual Thought framework, utilizing counterfactual reasoning, ASR-enhanced video content, and rationale refinement to reduce inference bias. For the generation task, we develop a framework that combines ASR output with a Large Language Model (LLM) to produce coherent, contextually relevant updates aligned with the intended strengthener or weakener goals. Additionally, we introduce a novel benchmark dataset, with strengthener/weakener annotations and an LLM-based evaluation metric specifically designed for assessing generative performance. Experimental results demonstrate significant improvements, highlighting our proposed method in enhancing dynamic reasoning capabilities of VLMMs.
♻ ☆ GEM-Bench: A Benchmark for Ad-Injected Response Generation within Generative Engine Marketing
Generative Engine Marketing (GEM) is an emerging ecosystem for monetizing generative engines, such as LLM-based chatbots, by seamlessly integrating relevant advertisements into their responses. At the core of GEM lies the generation and evaluation of ad-injected responses. However, existing benchmarks are not specifically designed for this purpose, which limits future research. To address this gap, we propose GEM-Bench, the first comprehensive benchmark for ad-injected response generation in GEM. GEM-Bench includes three curated datasets covering both chatbot and search scenarios, a metric ontology that captures multiple dimensions of user satisfaction and engagement, and several baseline solutions implemented within an extensible multi-agent framework. Our preliminary results indicate that, while simple prompt-based methods achieve reasonable engagement such as click-through rate, they often reduce user satisfaction. In contrast, approaches that insert ads based on pre-generated ad-free responses help mitigate this issue but introduce additional overhead. These findings highlight the need for future research on designing more effective and efficient solutions for generating ad-injected responses in GEM. The benchmark and all related resources are publicly available at https://gem-bench.org/.
comment: Include more experimental results and supplementary materials
♻ ☆ PLSemanticsBench: Large Language Models As Programming Language Interpreters
As large language models (LLMs) excel at code reasoning, a natural question arises: can an LLM execute programs (i.e., act as an interpreter) purely based on a programming language's formal semantics? If so, it will enable rapid prototyping of new programming languages and language features. We study this question using the imperative language IMP (a subset of C), formalized via small-step operational semantics (SOS) and rewriting-based operational semantics (K-semantics). We introduce three evaluation sets-Human-Written, LLM-Translated, and Fuzzer- Generated-whose difficulty is controlled by code-complexity metrics spanning the size, control-flow, and data-flow axes. Given a program and its semantics formalized with SOS/K-semantics, models are evaluated on three tasks ranging from coarse to fine: (1) final-state prediction, (2) semantic rule prediction, and (3) execution trace prediction. To distinguish pretraining memorization from semantic competence, we define two nonstandard semantics obtained through systematic mutations of the standard rules. Across strong code/reasoning LLMs, performance drops under nonstandard semantics despite high performance under the standard one. We further find that (i) there are patterns to different model failures, (ii) most reasoning models perform exceptionally well on coarse grained tasks involving reasoning about highly complex programs often containing nested loop depths beyond five, and surprisingly, (iii) providing formal semantics helps on simple programs but often hurts on more complex ones. Overall, the results show a promise that LLMs could serve as programming language interpreters, but points to the lack of their robust semantics understanding. We release the benchmark and the supporting code at https://github.com/EngineeringSoftware/PLSemanticsBench.
♻ ☆ COLE: a Comprehensive Benchmark for French Language Understanding Evaluation ACL
To address the need for a more comprehensive evaluation of French Natural Language Understanding (NLU), we introduce COLE, a new benchmark composed of 23 diverse task covering a broad range of NLU capabilities, including sentiment analysis, paraphrase detection, grammatical judgment, and reasoning, with a particular focus on linguistic phenomena relevant to the French language. We benchmark 94 large language models (LLM), providing an extensive analysis of the current state of French NLU. Our results highlight a significant performance gap between closed- and open-weights models and identify key challenging frontiers for current LLMs, such as zero-shot extractive question-answering (QA), fine-grained word sense disambiguation, and understanding of regional language variations. We release COLE as a public resource to foster further progress in French language modelling.
comment: Submitted to ACL Rolling Review of October
♻ ☆ Bayesian Teaching Enables Probabilistic Reasoning in Large Language Models
Artificial intelligence systems based on large language models (LLMs) are increasingly used as agents that interact with users and with the world. To do so successfully, LLMs need to construct internal representations of the world and form probabilistic beliefs about those representations. To provide a user with personalized recommendations, for example, the LLM needs to gradually infer the user's preferences, over the course of multiple interactions. To evaluate whether contemporary LLMs are able to do so, we use the Bayesian inference framework from probability theory, which lays out the optimal way to update an agent's beliefs as it receives new information. We first show that LLMs do not update their beliefs as expected from the Bayesian framework, and that consequently their predictions do not improve as expected as more information becomes available. To address this issue, we teach the LLMs to reason in a Bayesian manner by training them to mimic the predictions of the normative Bayesian model. We find that this approach not only significantly improves the LLM's performance on the particular recommendation task it is trained on, but also enables generalization to other tasks. This suggests that this method teaches the LLM to better approximate Bayesian reasoning. More generally, our results indicate that LLMs can effectively learn reasoning skills from examples and generalize those skills to new domains.
♻ ☆ What Prompts Don't Say: Understanding and Managing Underspecification in LLM Prompts
Prompt underspecification is a common challenge when interacting with LLMs. In this paper, we present an in-depth analysis of this problem, showing that while LLMs can often infer unspecified requirements by default (41.1%), such behavior is fragile: Under-specified prompts are 2x as likely to regress across model or prompt changes, sometimes with accuracy drops exceeding 20%. This instability makes it difficult to reliably build LLM applications. Moreover, simply specifying all requirements does not consistently help, as models have limited instruction-following ability and requirements can conflict. Standard prompt optimizers likewise provide little benefit. To address these issues, we propose requirements-aware prompt optimization mechanisms that improve performance by 4.8% on average over baselines. We further advocate for a systematic process of proactive requirements discovery, evaluation, and monitoring to better manage prompt underspecification in practice.
♻ ☆ ExpertLongBench: Benchmarking Language Models on Expert-Level Long-Form Generation Tasks with Structured Checklists
This paper introduces ExpertLongBench, an expert-level benchmark containing 11 tasks from 9 domains that reflect realistic expert workflows and applications. Beyond question answering, the application-driven tasks in ExpertLongBench demand long-form outputs that can exceed 5,000 tokens and strict adherence to domain-specific requirements. Notably, each task in ExpertLongBench includes a rubric, designed or validated by domain experts, to specify task requirements and guide output evaluation. Furthermore, we propose CLEAR, an evaluation framework that supports accurate evaluation of long-form model outputs in our benchmark. To achieve fine-grained, expert-aligned evaluation, CLEAR derives checklists from both model outputs and references by extracting information corresponding to items in the task-specific rubric. Checklist items of model outputs are then compared with corresponding items of reference outputs to assess their correctness, enabling grounded evaluation. We benchmark 13 popular large language models (LLMs) and analyze components in CLEAR, showing that (1) existing LLMs, with the top performer Gemini-2.5-Pro achieving only a 33.4 F1 score, require significant improvement for expert-level tasks; (2) models can generate content corresponding to the required aspects, but far from correct; and (3) accurate checklist extraction and comparison in CLEAR can be achieved by open-weight models for more scalable, reproducible, and low-cost usage.
♻ ☆ Compound AI Systems Optimization: A Survey of Methods, Challenges, and Future Directions EMNLP 2025
Recent advancements in large language models (LLMs) and AI systems have led to a paradigm shift in the design and optimization of complex AI workflows. By integrating multiple components, compound AI systems have become increasingly adept at performing sophisticated tasks. However, as these systems grow in complexity, new challenges arise in optimizing not only individual components but also their interactions. While traditional optimization methods such as supervised fine-tuning (SFT) and reinforcement learning (RL) remain foundational, the rise of natural language feedback introduces promising new approaches, especially for optimizing non-differentiable systems. This paper provides a systematic review of recent progress in optimizing compound AI systems, encompassing both numerical and language-based techniques. We formalize the notion of compound AI system optimization, classify existing methods along several key dimensions, and highlight open research challenges and future directions in this rapidly evolving field. A list of surveyed papers is publicly available at https://github.com/MiuLab/AISysOpt-Survey.
comment: Accepted to EMNLP 2025 (Main)
Information Retrieval 17
☆ Peeking inside the Black-Box: Reinforcement Learning for Explainable and Accurate Relation Extraction
This paper introduces a framework for relation extraction (RE) that enhances both accuracy and explainability. The framework has two key components: (i) a reasoning mechanism that formulates relation extraction as a series of text-processing steps inspired by cognitive science, and (ii) an optimization process driven by reinforcement learning (RL) with a novel reward function designed to improve both task accuracy and explanation quality. We call our approach CogRE. Our framework addresses the lack of supervision for language-based explanations in traditional RE by promoting outputs that include important relation keywords. These keywords are drawn from a high-quality dictionary that is automatically constructed using an LLM. We evaluate our approach for the task of one-shot RE using two LLMs and two RE datasets. Our experiments show that CogRE improves explanation quality by addressing two common failure patterns in one-shot RE: poor attention focus and limited one-shot learning capability. For example, our cognitive-structured reasoning with Qwen2.5-15B-Instruct on One-shot NYT29 achieves 24.65% F1, surpassing prior reasoning-based designs. Optimizing this approach with RL using our reward further improves performance by +23.46% (absolute). Finally, human evaluation shows that our best model generates relational keywords closely aligned with gold labels, increasing human explanation quality ratings by 54% (relative).
comment: Working in process
☆ Deterministic Legal Retrieval: An Action API for Querying the SAT-Graph RAG
The Structure-Aware Temporal Graph RAG (SAT-Graph RAG) addresses core limitations of standard Retrieval-Augmented Generation in the legal domain by providing a verifiable knowledge graph that models hierarchical structure, temporal evolution, and causal events of legal norms. However, a critical gap remains: how to reliably query this structured knowledge without sacrificing its deterministic properties. This paper introduces the SAT-Graph API, a formal query execution layer centered on canonical actions-atomic, composable, and auditable primitives that isolate probabilistic discovery from deterministic retrieval. These actions enable: (i) high-precision hybrid search; (ii) robust reference resolution; (iii) point-in-time version retrieval; and (iv) auditable causal tracing. We demonstrate how planner-guided agents can decompose complex queries into Directed Acyclic Graphs (DAGs) of these actions. This two-layer architecture transforms retrieval from an opaque black box to a transparent, auditable process, directly addressing Explainable AI (XAI) requirements for high-stakes domains.
☆ How public datasets constrain the development of diversity-aware news recommender systems, and what law could do about it
News recommender systems increasingly determine what news individuals see online. Over the past decade, researchers have extensively critiqued recommender systems that prioritise news based on user engagement. To offer an alternative, researchers have analysed how recommender systems could support the media's ability to fulfil its role in democratic society by recommending news based on editorial values, particularly diversity. However, there continues to be a large gap between normative theory on how news recommender systems should incorporate diversity, and technical literature that designs such systems. We argue that to realise diversity-aware recommender systems in practice, it is crucial to pay attention to the datasets that are needed to train modern news recommenders. We aim to make two main contributions. First, we identify the information a dataset must include to enable the development of the diversity-aware news recommender systems proposed in normative literature. Based on this analysis, we assess the limitations of currently available public datasets, and show what potential they do have to expand research into diversity-aware recommender systems. Second, we analyse why and how European law and policy can be used to provide researchers with structural access to the data they need to develop diversity-aware news recommender systems.
☆ Limitations of Current Evaluation Practices for Conversational Recommender Systems and the Potential of User Simulation SIGIR
Research and development on conversational recommender systems (CRSs) critically depends on sound and reliable evaluation methodologies. However, the interactive nature of these systems poses significant challenges for automatic evaluation. This paper critically examines current evaluation practices and identifies two key limitations: the over-reliance on static test collections and the inadequacy of existing evaluation metrics. To substantiate this critique, we analyze real user interactions with nine existing CRSs and demonstrate a striking disconnect between self-reported user satisfaction and performance scores reported in prior literature. To address these limitations, this work explores the potential of user simulation to generate dynamic interaction data, offering a departure from static datasets. Furthermore, we propose novel evaluation metrics, based on a general reward/cost framework, designed to better align with real user satisfaction. Our analysis of different simulation approaches provides valuable insights into their effectiveness and reveals promising initial results, showing improved correlation with system rankings compared to human evaluation. While these findings indicate a significant step forward in CRS evaluation, we also identify areas for future research and refinement in both simulation techniques and evaluation metrics.
comment: Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (SIGIR-AP 2025), December 7--10, 2025, Xi'an, China
AgentDR Dynamic Recommendation with Implicit Item-Item Relations via LLM-based Agents
Recent agent-based recommendation frameworks aim to simulate user behaviors by incorporating memory mechanisms and prompting strategies, but they struggle with hallucinating non-existent items and full-catalog ranking. Besides, a largely underexplored opportunity lies in leveraging LLMs'commonsense reasoning to capture user intent through substitute and complement relationships between items, which are usually implicit in datasets and difficult for traditional ID-based recommenders to capture. In this work, we propose a novel LLM-agent framework, AgenDR, which bridges LLM reasoning with scalable recommendation tools. Our approach delegates full-ranking tasks to traditional models while utilizing LLMs to (i) integrate multiple recommendation outputs based on personalized tool suitability and (ii) reason over substitute and complement relationships grounded in user history. This design mitigates hallucination, scales to large catalogs, and enhances recommendation relevance through relational reasoning. Through extensive experiments on three public grocery datasets, we show that our framework achieves superior full-ranking performance, yielding on average a twofold improvement over its underlying tools. We also introduce a new LLM-based evaluation metric that jointly measures semantic alignment and ranking correctness.
☆ KEO: Knowledge Extraction on OMIn via Knowledge Graphs and RAG for Safety-Critical Aviation Maintenance
We present Knowledge Extraction on OMIn (KEO), a domain-specific knowledge extraction and reasoning framework with large language models (LLMs) in safety-critical contexts. Using the Operations and Maintenance Intelligence (OMIn) dataset, we construct a QA benchmark spanning global sensemaking and actionable maintenance tasks. KEO builds a structured Knowledge Graph (KG) and integrates it into a retrieval-augmented generation (RAG) pipeline, enabling more coherent, dataset-wide reasoning than traditional text-chunk RAG. We evaluate locally deployable LLMs (Gemma-3, Phi-4, Mistral-Nemo) and employ stronger models (GPT-4o, Llama-3.3) as judges. Experiments show that KEO markedly improves global sensemaking by revealing patterns and system-level insights, while text-chunk RAG remains effective for fine-grained procedural tasks requiring localized retrieval. These findings underscore the promise of KG-augmented LLMs for secure, domain-specific QA and their potential in high-stakes reasoning.
☆ Automated Research Article Classification and Recommendation Using NLP and ML
In the digital era, the exponential growth of scientific publications has made it increasingly difficult for researchers to efficiently identify and access relevant work. This paper presents an automated framework for research article classification and recommendation that leverages Natural Language Processing (NLP) techniques and machine learning. Using a large-scale arXiv.org dataset spanning more than three decades, we evaluate multiple feature extraction approaches (TF--IDF, Count Vectorizer, Sentence-BERT, USE, Mirror-BERT) in combination with diverse machine learning classifiers (Logistic Regression, SVM, Na\"ive Bayes, Random Forest, Gradient Boosted Trees, and k-Nearest Neighbour). Our experiments show that Logistic Regression with TF--IDF consistently yields the best classification performance, achieving an accuracy of 69\%. To complement classification, we incorporate a recommendation module based on the cosine similarity of vectorized articles, enabling efficient retrieval of related research papers. The proposed system directly addresses the challenge of information overload in digital libraries and demonstrates a scalable, data-driven solution to support literature discovery.
comment: 8 pages, 4 figures, Accepted in Foundation and Large Language Models (FLLM2025)
☆ Semantic-Cohesive Knowledge Distillation for Deep Cross-modal Hashing
Recently, deep supervised cross-modal hashing methods have achieve compelling success by learning semantic information in a self-supervised way. However, they still suffer from the key limitation that the multi-label semantic extraction process fail to explicitly interact with raw multimodal data, making the learned representation-level semantic information not compatible with the heterogeneous multimodal data and hindering the performance of bridging modality gap. To address this limitation, in this paper, we propose a novel semantic cohesive knowledge distillation scheme for deep cross-modal hashing, dubbed as SODA. Specifically, the multi-label information is introduced as a new textual modality and reformulated as a set of ground-truth label prompt, depicting the semantics presented in the image like the text modality. Then, a cross-modal teacher network is devised to effectively distill cross-modal semantic characteristics between image and label modalities and thus learn a well-mapped Hamming space for image modality. In a sense, such Hamming space can be regarded as a kind of prior knowledge to guide the learning of cross-modal student network and comprehensively preserve the semantic similarities between image and text modality. Extensive experiments on two benchmark datasets demonstrate the superiority of our model over the state-of-the-art methods.
♻ ☆ Epistemic Diversity and Knowledge Collapse in Large Language Models
Large language models (LLMs) tend to generate lexically, semantically, and stylistically homogenous texts. This poses a risk of knowledge collapse, where homogenous LLMs mediate a shrinking in the range of accessible information over time. Existing works on homogenization are limited by a focus on closed-ended multiple-choice setups or fuzzy semantic features, and do not look at trends across time and cultural contexts. To overcome this, we present a new methodology to measure epistemic diversity, i.e., variation in real-world claims in LLM outputs, which we use to perform a broad empirical study of LLM knowledge collapse. We test 27 LLMs, 155 topics covering 12 countries, and 200 prompt variations sourced from real user chats. For the topics in our study, we show that while newer models tend to generate more diverse claims, nearly all models are less epistemically diverse than a basic web search. We find that model size has a negative impact on epistemic diversity, while retrieval-augmented generation (RAG) has a positive impact, though the improvement from RAG varies by the cultural context. Finally, compared to a traditional knowledge source (Wikipedia), we find that country-specific claims reflect the English language more than the local one, highlighting a gap in epistemic representation
comment: 16 pages; 8 figures, 4 tables v2 changelog: Fixed the modeling for table 3, random effect is the model version
♻ ☆ FedFlex: Federated Learning for Diverse Netflix Recommendations
The drive for personalization in recommender systems creates a tension between user privacy and the risk of "filter bubbles". Although federated learning offers a promising paradigm for privacy-preserving recommendations, its impact on diversity remains unclear. We introduce FedFlex, a two-stage framework that combines local, on-device fine-tuning of matrix factorization models (SVD and BPR) with a lightweight Maximal Marginal Relevance (MMR) re-ranking step to promote diversity. We conducted the first live user study of a federated recommender, collecting behavioral data and feedback during a two-week online deployment. Our results show that FedFlex successfully engages users, with BPR outperforming SVD in click-through rate. Re-ranking with MMR consistently improved ranking quality (nDCG) across both models, with statistically significant gains, particularly for BPR. Diversity effects varied: MMR increased coverage for both models and improved intra-list diversity for BPR, but slightly reduced it for SVD, suggesting different interactions between personalization and diversification across models. Our exit questionnaire responses indicated that most users expressed no clear preference between re-ranked and unprocessed lists, implying that increased diversity did not substantially reduce user satisfaction.
♻ ☆ Contrastive Learning Using Graph Embeddings for Domain Adaptation of Language Models in the Process Industry EMNLP 2025
Recent trends in NLP utilize knowledge graphs (KGs) to enhance pretrained language models by incorporating additional knowledge from the graph structures to learn domain-specific terminology or relationships between documents that might otherwise be overlooked. This paper explores how SciNCL, a graph-aware neighborhood contrastive learning methodology originally designed for scientific publications, can be applied to the process industry domain, where text logs contain crucial information about daily operations and are often structured as sparse KGs. Our experiments demonstrate that language models fine-tuned with triplets derived from graph embeddings (GE) outperform a state-of-the-art mE5-large text encoder by 9.8-14.3% (5.45-7.96p) on the proprietary process industry text embedding benchmark (PITEB) while having 3 times fewer parameters.
comment: accepted to EMNLP 2025 (industry track)
♻ ☆ Text Clustering as Classification with LLMs
Text clustering serves as a fundamental technique for organizing and interpreting unstructured textual data, particularly in contexts where manual annotation is prohibitively costly. With the rapid advancement of Large Language Models (LLMs) and their demonstrated effectiveness across a broad spectrum of NLP tasks, an emerging body of research has begun to explore their potential in the domain of text clustering. However, existing LLM-based approaches still rely on fine-tuned embedding models and sophisticated similarity metrics, rendering them computationally intensive and necessitating domain-specific adaptation. To address these limitations, we propose a novel framework that reframes text clustering as a classification task by harnessing the in-context learning capabilities of LLMs. Our framework eliminates the need for fine-tuning embedding models or intricate clustering algorithms. It comprises two key steps: first, the LLM is prompted to generate a set of candidate labels based on the dataset and then merges semantically similar labels; second, it assigns the most appropriate label to each text sample. By leveraging the advanced natural language understanding and generalization capabilities of LLMs, the proposed approach enables effective clustering with minimal human intervention. Experimental results on diverse datasets demonstrate that our framework achieves comparable or superior performance to state-of-the-art embedding-based clustering techniques, while significantly reducing computational complexity and resource requirements. These findings underscore the transformative potential of LLMs in simplifying and enhancing text clustering tasks. We make our code available to the public for utilization at https://github.com/ECNU-Text-Computing/Text-Clustering-via-LLM. We also provide the supplementary Appendix within the repository.
comment: 11 pages, 3 figures
♻ ☆ TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems NeurIPS 2025
Regression models are crucial in recommender systems. However, retransformation bias problem has been conspicuously neglected within the community. While many works in other fields have devised effective bias correction methods, all of them are post-hoc cures externally to the model, facing practical challenges when applied to real-world recommender systems. Hence, we propose a preemptive paradigm to eradicate the bias intrinsically from the models via minor model refinement. Specifically, a novel TranSUN method is proposed with a joint bias learning manner to offer theoretically guaranteed unbiasedness under empirical superior convergence. It is further generalized into a novel generic regression model family, termed Generalized TranSUN (GTS), which not only offers more theoretical insights but also serves as a generic framework for flexibly developing various bias-free models. Comprehensive experimental results demonstrate the superiority of our methods across data from various domains, which have been successfully deployed in two real-world industrial recommendation scenarios, i.e. product and short video recommendation scenarios in Guess What You Like business domain in the homepage of Taobao App (a leading e-commerce platform with DAU > 300M), to serve the major online traffic.
comment: 37 pages, 6 figures, NeurIPS 2025 Poster
♻ ☆ Soft Reasoning Paths for Knowledge Graph Completion IJCAI 2025
Reasoning paths are reliable information in knowledge graph completion (KGC) in which algorithms can find strong clues of the actual relation between entities. However, in real-world applications, it is difficult to guarantee that computationally affordable paths exist toward all candidate entities. According to our observation, the prediction accuracy drops significantly when paths are absent. To make the proposed algorithm more stable against the missing path circumstances, we introduce soft reasoning paths. Concretely, a specific learnable latent path embedding is concatenated to each relation to help better model the characteristics of the corresponding paths. The combination of the relation and the corresponding learnable embedding is termed a soft path in our paper. By aligning the soft paths with the reasoning paths, a learnable embedding is guided to learn a generalized path representation of the corresponding relation. In addition, we introduce a hierarchical ranking strategy to make full use of information about the entity, relation, path, and soft path to help improve both the efficiency and accuracy of the model. Extensive experimental results illustrate that our algorithm outperforms the compared state-of-the-art algorithms by a notable margin. The code will be made publicly available after the paper is officially accepted.
comment: Accepted by IJCAI 2025
♻ ☆ Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning NeurIPS 2025
Retrieval-augmented generation (RAG) is widely utilized to incorporate external knowledge into large language models, thereby enhancing factuality and reducing hallucinations in question-answering (QA) tasks. A standard RAG pipeline consists of several components, such as query rewriting, document retrieval, document filtering, and answer generation. However, these components are typically optimized separately through supervised fine-tuning, which can lead to misalignments between the objectives of individual components and the overarching aim of generating accurate answers. Although recent efforts have explored using reinforcement learning (RL) to optimize specific RAG components, these approaches often focus on simple pipelines with only two components or do not adequately address the complex interdependencies and collaborative interactions among the modules. To overcome these limitations, we propose treating the complex RAG pipeline with multiple components as a multi-agent cooperative task, in which each component can be regarded as an RL agent. Specifically, we present MMOA-RAG, Multi-Module joint Optimization Algorithm for RAG, which employs multi-agent reinforcement learning to harmonize all agents' goals toward a unified reward, such as the F1 score of the final answer. Experiments conducted on various QA benchmarks demonstrate that MMOA-RAG effectively boost the overall performance of the pipeline and outperforms existing baselines. Furthermore, comprehensive ablation studies validate the contributions of individual components and demonstrate MMOA-RAG can be adapted to different RAG pipelines and benchmarks.
comment: NeurIPS 2025
♻ ☆ Boosting Text-to-Chart Retrieval through Training with Synthesized Semantic Insights
Charts are crucial for data analysis and decision-making.Text-to-chart retrieval systems have become increasingly important for Business Intelligence (BI), where users need to find relevant charts that match their analytical needs. These needs can be categorized into precise queries that are well-specified and fuzzy queries that are more exploratory -- both require understanding the semantics and context of the charts. However, existing text-to-chart retrieval solutions often fail to capture the semantic content and contextual information of charts, primarily due to the lack of comprehensive metadata (or semantic insights). To address this limitation, we propose a training data development pipeline that automatically synthesizes hierarchical semantic insights for charts, covering visual patterns (visual-oriented), statistical properties (statistics-oriented), and practical applications (task-oriented), which produces 207,498 semantic insights for 69,166 charts. Based on these, we train a CLIP-based model named ChartFinder to learn better representations of charts for text-to-chart retrieval. Our method leverages rich semantic insights during the training phase to develop a model that understands both visual and semantic aspects of charts.To evaluate text-to-chart retrieval performance, we curate the first benchmark, CRBench, for this task with 21,862 charts and 326 text queries from real-world BI applications, with ground-truth labels verified by the crowd workers.Experiments show that ChartFinder significantly outperforms existing methods in text-to-chart retrieval tasks across various settings. For precise queries, ChartFinder achieves up to 66.9% NDCG@10, which is 11.58% higher than state-of-the-art models. In fuzzy query tasks, our method also demonstrates consistent improvements, with an average increase of 5% across nearly all metrics.
comment: Need to be revised
♻ ☆ GEM-Bench: A Benchmark for Ad-Injected Response Generation within Generative Engine Marketing
Generative Engine Marketing (GEM) is an emerging ecosystem for monetizing generative engines, such as LLM-based chatbots, by seamlessly integrating relevant advertisements into their responses. At the core of GEM lies the generation and evaluation of ad-injected responses. However, existing benchmarks are not specifically designed for this purpose, which limits future research. To address this gap, we propose GEM-Bench, the first comprehensive benchmark for ad-injected response generation in GEM. GEM-Bench includes three curated datasets covering both chatbot and search scenarios, a metric ontology that captures multiple dimensions of user satisfaction and engagement, and several baseline solutions implemented within an extensible multi-agent framework. Our preliminary results indicate that, while simple prompt-based methods achieve reasonable engagement such as click-through rate, they often reduce user satisfaction. In contrast, approaches that insert ads based on pre-generated ad-free responses help mitigate this issue but introduce additional overhead. These findings highlight the need for future research on designing more effective and efficient solutions for generating ad-injected responses in GEM. The benchmark and all related resources are publicly available at https://gem-bench.org/.
comment: Include more experimental results and supplementary materials
Machine Learning 150
☆ TaTToo: Tool-Grounded Thinking PRM for Test-Time Scaling in Tabular Reasoning
Process Reward Models (PRMs) have recently emerged as a powerful framework for enhancing the reasoning capabilities of large reasoning models (LRMs), particularly in the context of test-time scaling (TTS). However, their potential for supervising LRMs on tabular reasoning domains remains underexplored. Through detailed empirical analyses, we identify that existing PRMs, though widely adopted for supervising text-only reasoning steps, struggle with table-specific operations such as sub-table retrieval and schema interaction, leading to critical performance bottlenecks. To address this limitation, we propose TaTToo, a novel table-grounded PRM framework that (i) reasons explicitly over tabular reasoning steps and (ii) integrates tool-based verification to provide precise reward supervision. Concretely, we first design a scalable data curation pipeline that constructs over 60k high-quality step-level annotations by integrating table verification rationales with tool-based executions. Building on the collected data, we train TaTToo with a dual-stage paradigm: cold-start supervised fine-tuning to capture tool-use reasoning patterns, followed by reinforcement learning with tool-grounded reward shaping to align our model with table-based verification. We provide a comprehensive evaluation of the policy improvement induced by our newly designed PRM. Across 5 challenging tabular reasoning benchmarks covering numerical reasoning, fact-checking, and data analysis, TaTToo improves downstream policy LRMs by 30.9% at inference, surpasses strong PRM baselines such as Qwen-2.5-Math-PRM-72B with only 8B parameters, and demonstrates strong generalizability across diverse TTS strategies.
☆ Stratified GRPO: Handling Structural Heterogeneity in Reinforcement Learning of LLM Search Agents
Large language model (LLM) agents increasingly rely on external tools such as search engines to solve complex, multi-step problems, and reinforcement learning (RL) has become a key paradigm for training them. However, the trajectories of search agents are structurally heterogeneous, where variations in the number, placement, and outcomes of search calls lead to fundamentally different answer directions and reward distributions. Standard policy gradient methods, which use a single global baseline, suffer from what we identify and formalize as cross-stratum bias-an "apples-to-oranges" comparison of heterogeneous trajectories. This cross-stratum bias distorts credit assignment and hinders exploration of complex, multi-step search strategies. To address this, we propose Stratified GRPO, whose central component, Stratified Advantage Normalization (SAN), partitions trajectories into homogeneous strata based on their structural properties and computes advantages locally within each stratum. This ensures that trajectories are evaluated only against their true peers. Our analysis proves that SAN eliminates cross-stratum bias, yields conditionally unbiased unit-variance estimates inside each stratum, and retains the global unbiasedness and unit-variance properties enjoyed by standard normalization, resulting in a more pure and scale-stable learning signal. To improve practical stability under finite-sample regimes, we further linearly blend SAN with the global estimator. Extensive experiments on diverse single-hop and multi-hop question-answering benchmarks demonstrate that Stratified GRPO consistently and substantially outperforms GRPO by up to 11.3 points, achieving higher training rewards, greater training stability, and more effective search policies. These results establish stratification as a principled remedy for structural heterogeneity in RL for LLM search agents.
☆ Training Dynamics Impact Post-Training Quantization Robustness
While post-training quantization is widely adopted for efficient deployment of large language models, the mechanisms underlying quantization robustness remain unclear. We conduct a comprehensive analysis of quantization degradation across open-source language model training trajectories up to 32B parameters and 15T training tokens to accurately assess the relationship between training dynamics and quantization performance. Our key finding is that quantization errors in large-scale training runs are driven by a complex interplay between learning rate and other training hyperparameters. Specifically, once learning rates decay, validation loss and quantization error diverge, largely independent of training data scale. To investigate interventions on the training dynamics and identify specific configurations that can modulate quantization robustness favorably, we train our own models in controlled experiments up to 100B tokens. Our results challenge the assumption that increasing dataset scale inherently compromises quantization effectiveness, demonstrating instead that strategic training hyperparameter interventions can improve quantization quality at scale.
☆ Modulation Discovery with Differentiable Digital Signal Processing SP
Modulations are a critical part of sound design and music production, enabling the creation of complex and evolving audio. Modern synthesizers provide envelopes, low frequency oscillators (LFOs), and more parameter automation tools that allow users to modulate the output with ease. However, determining the modulation signals used to create a sound is difficult, and existing sound-matching / parameter estimation systems are often uninterpretable black boxes or predict high-dimensional framewise parameter values without considering the shape, structure, and routing of the underlying modulation curves. We propose a neural sound-matching approach that leverages modulation extraction, constrained control signal parameterizations, and differentiable digital signal processing (DDSP) to discover the modulations present in a sound. We demonstrate the effectiveness of our approach on highly modulated synthetic and real audio samples, its applicability to different DDSP synth architectures, and investigate the trade-off it incurs between interpretability and sound-matching accuracy. We make our code and audio samples available and provide the trained DDSP synths in a VST plugin.
comment: Accepted to WASPAA 2025 (best paper award candidate). Code, audio samples, and plugins can be found at https://christhetree.github.io/mod_discovery/
☆ Reference Grounded Skill Discovery
Scaling unsupervised skill discovery algorithms to high-DoF agents remains challenging. As dimensionality increases, the exploration space grows exponentially, while the manifold of meaningful skills remains limited. Therefore, semantic meaningfulness becomes essential to effectively guide exploration in high-dimensional spaces. In this work, we present Reference-Grounded Skill Discovery (RGSD), a novel algorithm that grounds skill discovery in a semantically meaningful latent space using reference data. RGSD first performs contrastive pretraining to embed motions on a unit hypersphere, clustering each reference trajectory into a distinct direction. This grounding enables skill discovery to simultaneously involve both imitation of reference behaviors and the discovery of semantically related diverse behaviors. On a simulated SMPL humanoid with 359-D observations and 69-D actions, RGSD learns structured skills including walking, running, punching, and side stepping, and also discovers related novel behaviors. In downstream control tasks, RGSD outperforms imitation-based skill acquisition baselines. Our results suggest that lightweight reference-guided grounding offers a practical path to discovering semantically rich and structured skills in high-DoF systems.
☆ Latent Speech-Text Transformer
Auto-regressive speech-text models are typically pre-trained on a large number of interleaved sequences of text tokens and raw speech encoded as speech tokens using vector quantization. These models have demonstrated state-of-the-art performance in speech-to-speech understanding and generation benchmarks, together with promising scaling laws, primarily enabled by the representational alignment between text and speech. Nevertheless, they suffer from shortcomings, partly owing to the disproportionately longer sequences of speech tokens in contrast to textual tokens. This results in a large compute imbalance between modalities during pre-training as well as during inference, and a potential hindrance to effectively aligning speech and text, ultimately translating to several orders of magnitude slower scaling laws. We introduce the Latent Speech-Text Transformer (LST), which makes pre-training speech-text models more data-efficient by dynamically and inexpensively aggregating speech tokens into latent speech patches. These patches serve as higher-level units that can either align with corresponding textual units to aid capability transfer or even encapsulate common speech sequences like silences to be more compute-efficient. We show that LST outperforms vanilla approaches on speech-to-speech as well as text-to-text benchmarks in both data- and compute-controlled settings, the former indicating more effective representational alignment and the latter indicating steeper scaling laws for speech-text models. On HellaSwag story completion, LST achieves 6.5% absolute gain in speech accuracy under compute-controlled training and 5.3% under data-controlled training, while also improving text performance. We will release our models, code, and the evaluation data to facilitate further research.
comment: 16 pages, 13 figures
☆ On Powerful Ways to Generate: Autoregression, Diffusion, and Beyond
This paper formally studies generation processes, including auto-regressive next-token prediction and masked diffusion, that abstract beyond architectural specifics. At this level of abstraction, we quantify their benefits and limitations through measurable criteria such as computational hardness and learnability. In particular, we demonstrate that allowing generation to proceed beyond autoregression and current masked diffusion, with capabilities to rewrite and length-variable edit, can bring significant theoretical and empirical advantages, with important implications for frontier LLMs that aspire to tackle increasingly hard problems and work universally across domains beyond natural language, such as coding and science.
☆ BanglaTalk: Towards Real-Time Speech Assistance for Bengali Regional Dialects
Real-time speech assistants are becoming increasingly popular for ensuring improved accessibility to information. Bengali, being a low-resource language with a high regional dialectal diversity, has seen limited progress in developing such systems. Existing systems are not optimized for real-time use and focus only on standard Bengali. In this work, we present BanglaTalk, the first real-time speech assistance system for Bengali regional dialects. BanglaTalk follows the client-server architecture and uses the Real-time Transport Protocol (RTP) to ensure low-latency communication. To address dialectal variation, we introduce a dialect-aware ASR system, BRDialect, developed by fine-tuning the IndicWav2Vec model in ten Bengali regional dialects. It outperforms the baseline ASR models by 12.41-33.98% on the RegSpeech12 dataset. Furthermore, BanglaTalk can operate at a low bandwidth of 24 kbps while maintaining an average end-to-end delay of 4.9 seconds. Low bandwidth usage and minimal end-to-end delay make the system both cost-effective and interactive for real-time use cases, enabling inclusive and accessible speech technology for the diverse community of Bengali speakers.
☆ Conformalized Gaussian processes for online uncertainty quantification over graphs
Uncertainty quantification (UQ) over graphs arises in a number of safety-critical applications in network science. The Gaussian process (GP), as a classical Bayesian framework for UQ, has been developed to handle graph-structured data by devising topology-aware kernel functions. However, such GP-based approaches are limited not only by the prohibitive computational complexity, but also the strict modeling assumptions that might yield poor coverage, especially with labels arriving on the fly. To effect scalability, we devise a novel graph-aware parametric GP model by leveraging the random feature (RF)-based kernel approximation, which is amenable to efficient recursive Bayesian model updates. To further allow for adaptivity, an ensemble of graph-aware RF-based scalable GPs have been leveraged, with per-GP weight adapted to data arriving incrementally. To ensure valid coverage with robustness to model mis-specification, we wed the GP-based set predictors with the online conformal prediction framework, which post-processes the prediction sets using adaptive thresholds. Experimental results the proposed method yields improved coverage and efficient prediction sets over existing baselines by adaptively ensembling the GP models and setting the key threshold parameters in CP.
☆ Climate Model Tuning with Online Synchronization-Based Parameter Estimation
In climate science, the tuning of climate models is a computationally intensive problem due to the combination of the high-dimensionality of the system state and long integration times. Here we demonstrate the potential of a parameter estimation algorithm which makes use of synchronization to tune a global atmospheric model at modest computational costs. We first use it to directly optimize internal model parameters. We then apply the algorithm to the weights of each member of a supermodel ensemble to optimize the overall predictions. In both cases, the algorithm is able to find parameters which result in reduced errors in the climatology of the model. Finally, we introduce a novel approach which combines both methods called adaptive supermodeling, where the internal parameters of the members of a supermodel are tuned simultaneously with the model weights such that the supermodel predictions are optimized. For a case designed to challenge the two previous methods, adaptive supermodeling achieves a performance similar to a perfect model.
comment: 19 pages, 11 figures
☆ Differentiable Model Predictive Control on the GPU
Differentiable model predictive control (MPC) offers a powerful framework for combining learning and control. However, its adoption has been limited by the inherently sequential nature of traditional optimization algorithms, which are challenging to parallelize on modern computing hardware like GPUs. In this work, we tackle this bottleneck by introducing a GPU-accelerated differentiable optimization tool for MPC. This solver leverages sequential quadratic programming and a custom preconditioned conjugate gradient (PCG) routine with tridiagonal preconditioning to exploit the problem's structure and enable efficient parallelization. We demonstrate substantial speedups over CPU- and GPU-based baselines, significantly improving upon state-of-the-art training times on benchmark reinforcement learning and imitation learning tasks. Finally, we showcase the method on the challenging task of reinforcement learning for driving at the limits of handling, where it enables robust drifting of a Toyota Supra through water puddles.
☆ Thermodynamic Performance Limits for Score-Based Diffusion Models
We establish a fundamental connection between score-based diffusion models and non-equilibrium thermodynamics by deriving performance limits based on entropy rates. Our main theoretical contribution is a lower bound on the negative log-likelihood of the data that relates model performance to entropy rates of diffusion processes. We numerically validate this bound on a synthetic dataset and investigate its tightness. By building a bridge to entropy rates - system, intrinsic, and exchange entropy - we provide new insights into the thermodynamic operation of these models, drawing parallels to Maxwell's demon and implications for thermodynamic computing hardware. Our framework connects generative modeling performance to fundamental physical principles through stochastic thermodynamics.
☆ Higher-Order Feature Attribution: Bridging Statistics, Explainable AI, and Topological Signal Processing
Feature attributions are post-training analysis methods that assess how various input features of a machine learning model contribute to an output prediction. Their interpretation is straightforward when features act independently, but becomes less direct when the predictive model involves interactions such as multiplicative relationships or joint feature contributions. In this work, we propose a general theory of higher-order feature attribution, which we develop on the foundation of Integrated Gradients (IG). This work extends existing frameworks in the literature on explainable AI. When using IG as the method of feature attribution, we discover natural connections to statistics and topological signal processing. We provide several theoretical results that establish the theory, and we validate our theory on a few examples.
comment: 5 pages, 3 figures
☆ TabPFN-Wide: Continued Pre-Training for Extreme Feature Counts
Revealing novel insights from the relationship between molecular measurements and pathology remains a very impactful application of machine learning in biomedicine. Data in this domain typically contain only a few observations but thousands of potentially noisy features, posing challenges for conventional machine learning approaches. While prior-data fitted networks emerge as foundation models for tabular data, they are currently not suited to handle large feature counts (>500). Although feature reduction enables their application, it hinders feature importance analysis. We propose a strategy that extends existing models through continued pre-training on synthetic data sampled from a customized prior. The resulting model, TabPFN-Wide, matches or exceeds its base model's performance while exhibiting improved robustness to noise. It seamlessly scales beyond 50,000 features, regardless of noise levels, while maintaining inherent interpretability, which is critical for biomedical applications. Our results show that prior-informed adaptation is suitable to enhance the capability of foundation models for high-dimensional data. On real-world biomedical datasets many of the most relevant features identified by the model overlap with previous biological findings, while others propose potential starting points for future studies.
☆ LLMs as Policy-Agnostic Teammates: A Case Study in Human Proxy Design for Heterogeneous Agent Teams ECAI 2025
A critical challenge in modelling Heterogeneous-Agent Teams is training agents to collaborate with teammates whose policies are inaccessible or non-stationary, such as humans. Traditional approaches rely on expensive human-in-the-loop data, which limits scalability. We propose using Large Language Models (LLMs) as policy-agnostic human proxies to generate synthetic data that mimics human decision-making. To evaluate this, we conduct three experiments in a grid-world capture game inspired by Stag Hunt, a game theory paradigm that balances risk and reward. In Experiment 1, we compare decisions from 30 human participants and 2 expert judges with outputs from LLaMA 3.1 and Mixtral 8x22B models. LLMs, prompted with game-state observations and reward structures, align more closely with experts than participants, demonstrating consistency in applying underlying decision criteria. Experiment 2 modifies prompts to induce risk-sensitive strategies (e.g. "be risk averse"). LLM outputs mirror human participants' variability, shifting between risk-averse and risk-seeking behaviours. Finally, Experiment 3 tests LLMs in a dynamic grid-world where the LLM agents generate movement actions. LLMs produce trajectories resembling human participants' paths. While LLMs cannot yet fully replicate human adaptability, their prompt-guided diversity offers a scalable foundation for simulating policy-agnostic teammates.
comment: This is a preprint of a paper presented at the \textit{European Conference on Artificial Intelligence (ECAI 2025)}. It is made publicly available for the benefit of the research community and should be regarded as a preprint rather than a formally reviewed publication
☆ Implicit Updates for Average-Reward Temporal Difference Learning
Temporal difference (TD) learning is a cornerstone of reinforcement learning. In the average-reward setting, standard TD($\lambda$) is highly sensitive to the choice of step-size and thus requires careful tuning to maintain numerical stability. We introduce average-reward implicit TD($\lambda$), which employs an implicit fixed point update to provide data-adaptive stabilization while preserving the per iteration computational complexity of standard average-reward TD($\lambda$). In contrast to prior finite-time analyses of average-reward TD($\lambda$), which impose restrictive step-size conditions, we establish finite-time error bounds for the implicit variant under substantially weaker step-size requirements. Empirically, average-reward implicit TD($\lambda$) operates reliably over a much broader range of step-sizes and exhibits markedly improved numerical stability. This enables more efficient policy evaluation and policy learning, highlighting its effectiveness as a robust alternative to average-reward TD($\lambda$).
☆ Non-iid hypothesis testing: from classical to quantum
We study hypothesis testing (aka state certification) in the non-identically distributed setting. A recent work (Garg et al. 2023) considered the classical case, in which one is given (independent) samples from $T$ unknown probability distributions $p_1, \dots, p_T$ on $[d] = \{1, 2, \dots, d\}$, and one wishes to accept/reject the hypothesis that their average $p_{\mathrm{avg}}$ equals a known hypothesis distribution $q$. Garg et al. showed that if one has just $c = 2$ samples from each $p_i$, and provided $T \gg \frac{\sqrt{d}}{\epsilon^2} + \frac{1}{\epsilon^4}$, one can (whp) distinguish $p_{\mathrm{avg}} = q$ from $d_{\mathrm{TV}}(p_{\mathrm{avg}},q) > \epsilon$. This nearly matches the optimal result for the classical iid setting (namely, $T \gg \frac{\sqrt{d}}{\epsilon^2}$). Besides optimally improving this result (and generalizing to tolerant testing with more stringent distance measures), we study the analogous problem of hypothesis testing for non-identical quantum states. Here we uncover an unexpected phenomenon: for any $d$-dimensional hypothesis state $\sigma$, and given just a single copy ($c = 1$) of each state $\rho_1, \dots, \rho_T$, one can distinguish $\rho_{\mathrm{avg}} = \sigma$ from $D_{\mathrm{tr}}(\rho_{\mathrm{avg}},\sigma) > \epsilon$ provided $T \gg d/\epsilon^2$. (Again, we generalize to tolerant testing with more stringent distance measures.) This matches the optimal result for the iid case, which is surprising because doing this with $c = 1$ is provably impossible in the classical case. We also show that the analogous phenomenon happens for the non-iid extension of identity testing between unknown states. A technical tool we introduce may be of independent interest: an Efron-Stein inequality, and more generally an Efron-Stein decomposition, in the quantum setting.
comment: 33 pages, 2 figures
☆ Bimanual 3D Hand Motion and Articulation Forecasting in Everyday Images
We tackle the problem of forecasting bimanual 3D hand motion & articulation from a single image in everyday settings. To address the lack of 3D hand annotations in diverse settings, we design an annotation pipeline consisting of a diffusion model to lift 2D hand keypoint sequences to 4D hand motion. For the forecasting model, we adopt a diffusion loss to account for the multimodality in hand motion distribution. Extensive experiments across 6 datasets show the benefits of training on diverse data with imputed labels (14% improvement) and effectiveness of our lifting (42% better) & forecasting (16.4% gain) models, over the best baselines, especially in zero-shot generalization to everyday images.
comment: Project page: https://ap229997.github.io/projects/forehand4d
☆ Improved High-probability Convergence Guarantees of Decentralized SGD
Convergence in high-probability (HP) has been receiving increasing interest, due to its attractive properties, such as exponentially decaying tail bounds and strong guarantees for each individual run of an algorithm. While HP guarantees are extensively studied in centralized settings, much less is understood in the decentralized, networked setup. Existing HP studies in decentralized settings impose strong assumptions, like uniformly bounded gradients, or asymptotically vanishing noise, resulting in a significant gap between assumptions used to establish convergence in the HP and the mean-squared error (MSE) sense, even for vanilla Decentralized Stochastic Gradient Descent ($\mathtt{DSGD}$) algorithm. This is contrary to centralized settings, where it is known that $\mathtt{SGD}$ converges in HP under the same conditions on the cost function as needed to guarantee MSE convergence. Motivated by this observation, we revisit HP guarantees for $\mathtt{DSGD}$ in the presence of light-tailed noise. We show that $\mathtt{DSGD}$ converges in HP under the same conditions on the cost as in the MSE sense, removing uniformly bounded gradients and other restrictive assumptions, while simultaneously achieving order-optimal rates for both non-convex and strongly convex costs. Moreover, our improved analysis yields linear speed-up in the number of users, demonstrating that $\mathtt{DSGD}$ maintains strong performance in the HP sense and matches existing MSE guarantees. Our improved results stem from a careful analysis of the MGF of quantities of interest (norm-squared of gradient or optimality gap) and the MGF of the consensus gap between users' models. To achieve linear speed-up, we provide a novel result on the variance-reduction effect of decentralized methods in the HP sense and more fine-grained bounds on the MGF for strongly convex costs, which are both of independent interest.
comment: 39 pages
☆ Multi-Task Reinforcement Learning with Language-Encoded Gated Policy Networks
Multi-task reinforcement learning often relies on task metadata -- such as brief natural-language descriptions -- to guide behavior across diverse objectives. We present Lexical Policy Networks (LEXPOL), a language-conditioned mixture-of-policies architecture for multi-task RL. LEXPOL encodes task metadata with a text encoder and uses a learned gating module to select or blend among multiple sub-policies, enabling end-to-end training across tasks. On MetaWorld benchmarks, LEXPOL matches or exceeds strong multi-task baselines in success rate and sample efficiency, without task-specific retraining. To analyze the mechanism, we further study settings with fixed expert policies obtained independently of the gate and show that the learned language gate composes these experts to produce behaviors appropriate to novel task descriptions and unseen task combinations. These results indicate that natural-language metadata can effectively index and recombine reusable skills within a single policy.
comment: 14 pages, 3 figures, 12 tables, 2 appendices. Currently under review
☆ lm-Meter: Unveiling Runtime Inference Latency for On-Device Language Models
Large Language Models (LLMs) are increasingly integrated into everyday applications, but their prevalent cloud-based deployment raises growing concerns around data privacy and long-term sustainability. Running LLMs locally on mobile and edge devices (on-device LLMs) offers the promise of enhanced privacy, reliability, and reduced communication costs. However, realizing this vision remains challenging due to substantial memory and compute demands, as well as limited visibility into performance-efficiency trade-offs on resource-constrained hardware. We propose lm-Meter, the first lightweight, online latency profiler tailored for on-device LLM inference. lm-Meter captures fine-grained, real-time latency at both phase (e.g., embedding, prefill, decode, softmax, sampling) and kernel levels without auxiliary devices. We implement lm-Meter on commercial mobile platforms and demonstrate its high profiling accuracy with minimal system overhead, e.g., only 2.58% throughput reduction in prefill and 0.99% in decode under the most constrained Powersave governor. Leveraging lm-Meter, we conduct comprehensive empirical studies revealing phase- and kernel-level bottlenecks in on-device LLM inference, quantifying accuracy-efficiency trade-offs, and identifying systematic optimization opportunities. lm-Meter provides unprecedented visibility into the runtime behavior of LLMs on constrained platforms, laying the foundation for informed optimization and accelerating the democratization of on-device LLM systems. Code and tutorials are available at https://github.com/amai-gsu/LM-Meter.
comment: This is the preprint version of the paper accepted to The 10th ACM/IEEE Symposium on Edge Computing (SEC 2025)
☆ Downsized and Compromised?: Assessing the Faithfulness of Model Compression
In real-world applications, computational constraints often require transforming large models into smaller, more efficient versions through model compression. While these techniques aim to reduce size and computational cost without sacrificing performance, their evaluations have traditionally focused on the trade-off between size and accuracy, overlooking the aspect of model faithfulness. This limited view is insufficient for high-stakes domains like healthcare, finance, and criminal justice, where compressed models must remain faithful to the behavior of their original counterparts. This paper presents a novel approach to evaluating faithfulness in compressed models, moving beyond standard metrics. We introduce and demonstrate a set of faithfulness metrics that capture how model behavior changes post-compression. Our contributions include introducing techniques to assess predictive consistency between the original and compressed models using model agreement, and applying chi-squared tests to detect statistically significant changes in predictive patterns across both the overall dataset and demographic subgroups, thereby exposing shifts that aggregate fairness metrics may obscure. We demonstrate our approaches by applying quantization and pruning to artificial neural networks (ANNs) trained on three diverse and socially meaningful datasets. Our findings show that high accuracy does not guarantee faithfulness, and our statistical tests detect subtle yet significant shifts that are missed by standard metrics, such as Accuracy and Equalized Odds. The proposed metrics provide a practical and more direct method for ensuring that efficiency gains through compression do not compromise the fairness or faithfulness essential for trustworthy AI.
comment: Submitted to and under review at Springer Machine Learning Journal
☆ PolyGraph Discrepancy: a classifier-based metric for graph generation
Existing methods for evaluating graph generative models primarily rely on Maximum Mean Discrepancy (MMD) metrics based on graph descriptors. While these metrics can rank generative models, they do not provide an absolute measure of performance. Their values are also highly sensitive to extrinsic parameters, namely kernel and descriptor parametrization, making them incomparable across different graph descriptors. We introduce PolyGraph Discrepancy (PGD), a new evaluation framework that addresses these limitations. It approximates the Jensen-Shannon distance of graph distributions by fitting binary classifiers to distinguish between real and generated graphs, featurized by these descriptors. The data log-likelihood of these classifiers approximates a variational lower bound on the JS distance between the two distributions. Resulting metrics are constrained to the unit interval [0,1] and are comparable across different graph descriptors. We further derive a theoretically grounded summary metric that combines these individual metrics to provide a maximally tight lower bound on the distance for the given descriptors. Thorough experiments demonstrate that PGD provides a more robust and insightful evaluation compared to MMD metrics. The PolyGraph framework for benchmarking graph generative models is made publicly available at https://github.com/BorgwardtLab/polygraph-benchmark.
☆ Influence Functions for Efficient Data Selection in Reasoning
Fine-tuning large language models (LLMs) on chain-of-thought (CoT) data shows that a small amount of high-quality data can outperform massive datasets. Yet, what constitutes "quality" remains ill-defined. Existing reasoning methods rely on indirect heuristics such as problem difficulty or trace length, while instruction-tuning has explored a broader range of automated selection strategies, but rarely in the context of reasoning. We propose to define reasoning data quality using influence functions, which measure the causal effect of individual CoT examples on downstream accuracy, and introduce influence-based pruning, which consistently outperforms perplexity and embedding-based baselines on math reasoning within a model family.
☆ The Physics of Data and Tasks: Theories of Locality and Compositionality in Deep Learning
Deep neural networks have achieved remarkable success, yet our understanding of how they learn remains limited. These models can learn high-dimensional tasks, which is generally statistically intractable due to the curse of dimensionality. This apparent paradox suggests that learnable data must have an underlying latent structure. What is the nature of this structure? How do neural networks encode and exploit it, and how does it quantitatively impact performance - for instance, how does generalization improve with the number of training examples? This thesis addresses these questions by studying the roles of locality and compositionality in data, tasks, and deep learning representations.
comment: PhD dissertation. Preprint
☆ Moloch's Bargain: Emergent Misalignment When LLMs Compete for Audiences
Large language models (LLMs) are increasingly shaping how information is created and disseminated, from companies using them to craft persuasive advertisements, to election campaigns optimizing messaging to gain votes, to social media influencers boosting engagement. These settings are inherently competitive, with sellers, candidates, and influencers vying for audience approval, yet it remains poorly understood how competitive feedback loops influence LLM behavior. We show that optimizing LLMs for competitive success can inadvertently drive misalignment. Using simulated environments across these scenarios, we find that, 6.3% increase in sales is accompanied by a 14.0% rise in deceptive marketing; in elections, a 4.9% gain in vote share coincides with 22.3% more disinformation and 12.5% more populist rhetoric; and on social media, a 7.5% engagement boost comes with 188.6% more disinformation and a 16.3% increase in promotion of harmful behaviors. We call this phenomenon Moloch's Bargain for AI--competitive success achieved at the cost of alignment. These misaligned behaviors emerge even when models are explicitly instructed to remain truthful and grounded, revealing the fragility of current alignment safeguards. Our findings highlight how market-driven optimization pressures can systematically erode alignment, creating a race to the bottom, and suggest that safe deployment of AI systems will require stronger governance and carefully designed incentives to prevent competitive dynamics from undermining societal trust.
☆ The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives
The objectives that Large Language Models (LLMs) implicitly optimize remain dangerously opaque, making trustworthy alignment and auditing a grand challenge. While Inverse Reinforcement Learning (IRL) can infer reward functions from behaviour, existing approaches either produce a single, overconfident reward estimate or fail to address the fundamental ambiguity of the task (non-identifiability). This paper introduces a principled auditing framework that re-frames reward inference from a simple estimation task to a comprehensive process for verification. Our framework leverages Bayesian IRL to not only recover a distribution over objectives but to enable three critical audit capabilities: (i) Quantifying and systematically reducing non-identifiability by demonstrating posterior contraction over sequential rounds of evidence; (ii) Providing actionable, uncertainty-aware diagnostics that expose spurious shortcuts and identify out-of-distribution prompts where the inferred objective cannot be trusted; and (iii) Validating policy-level utility by showing that the refined, low-uncertainty reward can be used directly in RLHF to achieve training dynamics and toxicity reductions comparable to the ground-truth alignment process. Empirically, our framework successfully audits a detoxified LLM, yielding a well-calibrated and interpretable objective that strengthens alignment guarantees. Overall, this work provides a practical toolkit for auditors, safety teams, and regulators to verify what LLMs are truly trying to achieve, moving us toward more trustworthy and accountable AI.
comment: Preprint
☆ Learning from Failures: Understanding LLM Alignment through Failure-Aware Inverse RL
Reinforcement Learning from Human Feedback (RLHF) aligns Large Language Models (LLMs) with human preferences, yet the underlying reward signals they internalize remain hidden, posing a critical challenge for interpretability and safety. Existing approaches attempt to extract these latent incentives using Inverse Reinforcement Learning (IRL), but treat all preference pairs equally, often overlooking the most informative signals: those examples the extracted reward model misclassifies or assigns nearly equal scores, which we term \emph{failures}. We introduce a novel \emph{failure-aware} IRL algorithm that focuses on misclassified or difficult examples to recover the latent rewards defining model behaviors. By learning from these failures, our failure-aware IRL extracts reward functions that better reflect the true objectives behind RLHF. We demonstrate that failure-aware IRL outperforms existing IRL baselines across multiple metrics when applied to LLM detoxification, without requiring external classifiers or supervision. Crucially, failure-aware IRL yields rewards that better capture the true incentives learned during RLHF, enabling more effective re-RLHF training than standard IRL. This establishes failure-aware IRL as a robust, scalable method for auditing model alignment and reducing ambiguity in the IRL process.
comment: Preprint
☆ Learning Mixtures of Linear Dynamical Systems (MoLDS) via Hybrid Tensor-EM Method
Mixtures of linear dynamical systems (MoLDS) provide a path to model time-series data that exhibit diverse temporal dynamics across trajectories. However, its application remains challenging in complex and noisy settings, limiting its effectiveness for neural data analysis. Tensor-based moment methods can provide global identifiability guarantees for MoLDS, but their performance degrades under noise and complexity. Commonly used expectation-maximization (EM) methods offer flexibility in fitting latent models but are highly sensitive to initialization and prone to poor local minima. Here, we propose a tensor-based method that provides identifiability guarantees for learning MoLDS, which is followed by EM updates to combine the strengths of both approaches. The novelty in our approach lies in the construction of moment tensors using the input-output data to recover globally consistent estimates of mixture weights and system parameters. These estimates can then be refined through a Kalman EM algorithm, with closed-form updates for all LDS parameters. We validate our framework on synthetic benchmarks and real-world datasets. On synthetic data, the proposed Tensor-EM method achieves more reliable recovery and improved robustness compared to either pure tensor or randomly initialized EM methods. We then analyze neural recordings from the primate somatosensory cortex while a non-human primate performs reaches in different directions. Our method successfully models and clusters different conditions as separate subsystems, consistent with supervised single-LDS fits for each condition. Finally, we apply this approach to another neural dataset where monkeys perform a sequential reaching task. These results demonstrate that MoLDS provides an effective framework for modeling complex neural data, and that Tensor-EM is a reliable approach to MoLDS learning for these applications.
comment: 20 pages, 7 figures
☆ EmoHRNet: High-Resolution Neural Network Based Speech Emotion Recognition
Speech emotion recognition (SER) is pivotal for enhancing human-machine interactions. This paper introduces "EmoHRNet", a novel adaptation of High-Resolution Networks (HRNet) tailored for SER. The HRNet structure is designed to maintain high-resolution representations from the initial to the final layers. By transforming audio samples into spectrograms, EmoHRNet leverages the HRNet architecture to extract high-level features. EmoHRNet's unique architecture maintains high-resolution representations throughout, capturing both granular and overarching emotional cues from speech signals. The model outperforms leading models, achieving accuracies of 92.45% on RAVDESS, 80.06% on IEMOCAP, and 92.77% on EMOVO. Thus, we show that EmoHRNet sets a new benchmark in the SER domain.
☆ Benchmark It Yourself (BIY): Preparing a Dataset and Benchmarking AI Models for Scatterplot-Related Tasks IEEE VIS 2025
AI models are increasingly used for data analysis and visualization, yet benchmarks rarely address scatterplot-specific tasks, limiting insight into performance. To address this gap for one of the most common chart types, we introduce a synthetic, annotated dataset of over 18,000 scatterplots from six data generators and 17 chart designs, and a benchmark based on it. We evaluate proprietary models from OpenAI and Google using N-shot prompting on five distinct tasks derived from annotations of cluster bounding boxes, their center coordinates, and outlier coordinates. OpenAI models and Gemini 2.5 Flash, especially when prompted with examples, are viable options for counting clusters and, in Flash's case, outliers (90%+ Accuracy). However, the results for localization-related tasks are unsatisfactory: Precision and Recall are near or below 50%, except for Flash in outlier identification (65.01%). Furthermore, the impact of chart design on performance appears to be a secondary factor, but it is advisable to avoid scatterplots with wide aspect ratios (16:9 and 21:9) or those colored randomly. Supplementary materials are available at https://github.com/feedzai/biy-paper.
comment: 9 pages, 3 figures, short paper accepted at VISxGenAI: 1st Workshop on GenAI, Agents, and the Future of VIS (IEEE VIS 2025)
☆ Analyzing the Effect of Embedding Norms and Singular Values to Oversmoothing in Graph Neural Networks
In this paper, we study the factors that contribute to the effect of oversmoothing in deep Graph Neural Networks (GNNs). Specifically, our analysis is based on a new metric (Mean Average Squared Distance - $MASED$) to quantify the extent of oversmoothing. We derive layer-wise bounds on $MASED$, which aggregate to yield global upper and lower distance bounds. Based on this quantification of oversmoothing, we further analyze the importance of two different properties of the model; namely the norms of the generated node embeddings, along with the largest and smallest singular values of the weight matrices. Building on the insights drawn from the theoretical analysis, we show that oversmoothing increases as the number of trainable weight matrices and the number of adjacency matrices increases. We also use the derived layer-wise bounds on $MASED$ to form a proposal for decoupling the number of hops (i.e., adjacency depth) from the number of weight matrices. In particular, we introduce G-Reg, a regularization scheme that increases the bounds, and demonstrate through extensive experiments that by doing so node classification accuracy increases, achieving robustness at large depths. We further show that by reducing oversmoothing in deep networks, we can achieve better results in some tasks than using shallow ones. Specifically, we experiment with a ``cold start" scenario, i.e., when there is no feature information for the unlabeled nodes. Finally, we show empirically the trade-off between receptive field size (i.e., number of weight matrices) and performance, using the $MASED$ bounds. This is achieved by distributing adjacency hops across a small number of trainable layers, avoiding the extremes of under- or over-parameterization of the GNN.
☆ TelecomTS: A Multi-Modal Observability Dataset for Time Series and Language Analysis
Modern enterprises generate vast streams of time series metrics when monitoring complex systems, known as observability data. Unlike conventional time series from domains such as weather, observability data are zero-inflated, highly stochastic, and exhibit minimal temporal structure. Despite their importance, observability datasets are underrepresented in public benchmarks due to proprietary restrictions. Existing datasets are often anonymized and normalized, removing scale information and limiting their use for tasks beyond forecasting, such as anomaly detection, root-cause analysis, and multi-modal reasoning. To address this gap, we introduce TelecomTS, a large-scale observability dataset derived from a 5G telecommunications network. TelecomTS features heterogeneous, de-anonymized covariates with explicit scale information and supports a suite of downstream tasks, including anomaly detection, root-cause analysis, and a question-answering benchmark requiring multi-modal reasoning. Benchmarking state-of-the-art time series, language, and reasoning models reveals that existing approaches struggle with the abrupt, noisy, and high-variance dynamics of observability data. Our experiments also underscore the importance of preserving covariates' absolute scale, emphasizing the need for foundation time series models that natively leverage scale information for practical observability applications.
☆ Medical Vision Language Models as Policies for Robotic Surgery
Vision-based Proximal Policy Optimization (PPO) struggles with visual observation-based robotic laparoscopic surgical tasks due to the high-dimensional nature of visual input, the sparsity of rewards in surgical environments, and the difficulty of extracting task-relevant features from raw visual data. We introduce a simple approach integrating MedFlamingo, a medical domain-specific Vision-Language Model, with PPO. Our method is evaluated on five diverse laparoscopic surgery task environments in LapGym, using only endoscopic visual observations. MedFlamingo PPO outperforms and converges faster compared to both standard vision-based PPO and OpenFlamingo PPO baselines, achieving task success rates exceeding 70% across all environments, with improvements ranging from 66.67% to 1114.29% compared to baseline. By processing task observations and instructions once per episode to generate high-level planning tokens, our method efficiently combines medical expertise with real-time visual feedback. Our results highlight the value of specialized medical knowledge in robotic surgical planning and decision-making.
comment: IEEE CAI 2025
☆ Edit-Based Flow Matching for Temporal Point Processes
Temporal point processes (TPPs) are a fundamental tool for modeling event sequences in continuous time, but most existing approaches rely on autoregressive parameterizations that are limited by their sequential sampling. Recent non-autoregressive, diffusion-style models mitigate these issues by jointly interpolating between noise and data through event insertions and deletions in a discrete Markov chain. In this work, we generalize this perspective and introduce an Edit Flow process for TPPs that transports noise to data via insert, delete, and substitute edit operations. By learning the instantaneous edit rates within a continuous-time Markov chain framework, we attain a flexible and efficient model that effectively reduces the total number of necessary edit operations during generation. Empirical results demonstrate the generative flexibility of our unconditionally trained model in a wide range of unconditional and conditional generation tasks on benchmark TPPs.
☆ BLISS: A Lightweight Bilevel Influence Scoring Method for Data Selection in Language Model Pretraining
Effective data selection is essential for pretraining large language models (LLMs), enhancing efficiency and improving generalization to downstream tasks. However, existing approaches often require leveraging external pretrained models, making it difficult to disentangle the effects of data selection from those of the external pretrained models. In addition, they often overlook the long-term impact of selected data if the model is trained to convergence, primarily due to the prohibitive cost of full-scale LLM pretraining. In this paper, we introduce BLISS (\textbf{B}ileve\textbf{L} \textbf{I}nfluence \textbf{S}coring method for data \textbf{S}election): a lightweight data selection method that operates entirely \emph{from scratch}, without relying on any external pretrained oracle models, while explicitly accounting for the long-term impact of selected data. BLISS leverages a small proxy model as a surrogate for the LLM and employs a score model to estimate the long-term influence of training samples if the proxy model is trained to convergence. We formulate data selection as a bilevel optimization problem, where the upper-level objective optimizes the score model to assign importance weights to training samples, ensuring that minimizing the lower-level objective (i.e., training the proxy model over the weighted training loss until convergence) leads to best validation performance. Once optimized, the trained score model predicts influence scores for the dataset, enabling efficient selection of high-quality samples for LLM pretraining. We validate BLISS by pretraining 410M/1B/2.8B Pythia and LLaMA-0.5B models on selected subsets of the C4 dataset. Notably, under the 1B model setting, BLISS achieves $1.7\times$ speedup in reaching the same performance as the state-of-the-art method, demonstrating superior performance across multiple downstream tasks.
☆ From Learning to Mastery: Achieving Safe and Efficient Real-World Autonomous Driving with Human-In-The-Loop Reinforcement Learning
Autonomous driving with reinforcement learning (RL) has significant potential. However, applying RL in real-world settings remains challenging due to the need for safe, efficient, and robust learning. Incorporating human expertise into the learning process can help overcome these challenges by reducing risky exploration and improving sample efficiency. In this work, we propose a reward-free, active human-in-the-loop learning method called Human-Guided Distributional Soft Actor-Critic (H-DSAC). Our method combines Proxy Value Propagation (PVP) and Distributional Soft Actor-Critic (DSAC) to enable efficient and safe training in real-world environments. The key innovation is the construction of a distributed proxy value function within the DSAC framework. This function encodes human intent by assigning higher expected returns to expert demonstrations and penalizing actions that require human intervention. By extrapolating these labels to unlabeled states, the policy is effectively guided toward expert-like behavior. With a well-designed state space, our method achieves real-world driving policy learning within practical training times. Results from both simulation and real-world experiments demonstrate that our framework enables safe, robust, and sample-efficient learning for autonomous driving.
☆ Adaptive Pruning for Increased Robustness and Reduced Computational Overhead in Gaussian Process Accelerated Saddle Point Searches
Gaussian process (GP) regression provides a strategy for accelerating saddle point searches on high-dimensional energy surfaces by reducing the number of times the energy and its derivatives with respect to atomic coordinates need to be evaluated. The computational overhead in the hyperparameter optimization can, however, be large and make the approach inefficient. Failures can also occur if the search ventures too far into regions that are not represented well enough by the GP model. Here, these challenges are resolved by using geometry-aware optimal transport measures and an active pruning strategy using a summation over Wasserstein-1 distances for each atom-type in farthest-point sampling, selecting a fixed-size subset of geometrically diverse configurations to avoid rapidly increasing cost of GP updates as more observations are made. Stability is enhanced by permutation-invariant metric that provides a reliable trust radius for early-stopping and a logarithmic barrier penalty for the growth of the signal variance. These physically motivated algorithmic changes prove their efficacy by reducing to less than a half the mean computational time on a set of 238 challenging configurations from a previously published data set of chemical reactions. With these improvements, the GP approach is established as, a robust and scalable algorithm for accelerating saddle point searches when the evaluation of the energy and atomic forces requires significant computational effort.
comment: Invited article for the ChemPhysChem special issue dedicated to the 60th birthday of Prof. Debabrata Goswami. A preliminary version of this work was presented at the UNOOS 2025 conference
☆ Fast Leave-One-Out Approximation from Fragment-Target Prevalence Vectors (molFTP) : From Dummy Masking to Key-LOO for Leakage-Free Feature Construction
We introduce molFTP (molecular fragment-target prevalence), a compact representation that delivers strong predictive performance. To prevent feature leakage across cross-validation folds, we implement a dummy-masking procedure that removes information about fragments present in the held-out molecules. We further show that key leave-one-out (key-loo) closely approximates true molecule-level leave-one-out (LOO), with deviation below 8% on our datasets. This enables near full data training while preserving unbiased cross-validation estimates of model performance. Overall, molFTP provides a fast, leakage-resistant fragment-target prevalence vectorization with practical safeguards (dummy masking or key-LOO) that approximate LOO at a fraction of its cost.
comment: 28 pages, 21 figures, 3 tables
☆ Generalization of Gibbs and Langevin Monte Carlo Algorithms in the Interpolation Regime
The paper provides data-dependent bounds on the test error of the Gibbs algorithm in the overparameterized interpolation regime, where low training errors are also obtained for impossible data, such as random labels in classification. The bounds are stable under approximation with Langevin Monte Carlo algorithms. Experiments on the MNIST and CIFAR-10 datasets verify that the bounds yield nontrivial predictions on true labeled data and correctly upper bound the test error for random labels. Our method indicates that generalization in the low-temperature, interpolation regime is already signaled by small training errors in the more classical high temperature regime.
☆ Emergent AI Surveillance: Overlearned Person Re-Identification and Its Mitigation in Law Enforcement Context
Generic instance search models can dramatically reduce the manual effort required to analyze vast surveillance footage during criminal investigations by retrieving specific objects of interest to law enforcement. However, our research reveals an unintended emergent capability: through overlearning, these models can single out specific individuals even when trained on datasets without human subjects. This capability raises concerns regarding identification and profiling of individuals based on their personal data, while there is currently no clear standard on how de-identification can be achieved. We evaluate two technical safeguards to curtail a model's person re-identification capacity: index exclusion and confusion loss. Our experiments demonstrate that combining these approaches can reduce person re-identification accuracy to below 2% while maintaining 82% of retrieval performance for non-person objects. However, we identify critical vulnerabilities in these mitigations, including potential circumvention using partial person images. These findings highlight urgent regulatory questions at the intersection of AI governance and data protection: How should we classify and regulate systems with emergent identification capabilities? And what technical standards should be required to prevent identification capabilities from developing in seemingly benign applications?
comment: 10 pages, accepted to AIES 2025
☆ Out-of-Distribution Detection from Small Training Sets using Bayesian Neural Network Classifiers BMVC
Out-of-Distribution (OOD) detection is critical to AI reliability and safety, yet in many practical settings, only a limited amount of training data is available. Bayesian Neural Networks (BNNs) are a promising class of model on which to base OOD detection, because they explicitly represent epistemic (i.e. model) uncertainty. In the small training data regime, BNNs are especially valuable because they can incorporate prior model information. We introduce a new family of Bayesian posthoc OOD scores based on expected logit vectors, and compare 5 Bayesian and 4 deterministic posthoc OOD scores. Experiments on MNIST and CIFAR-10 In-Distributions, with 5000 training samples or less, show that the Bayesian methods outperform corresponding deterministic methods.
comment: British Machine Vision Conference (BMVC) 2025; 18 pages, 6 figures, 3 tables
☆ RamPINN: Recovering Raman Spectra From Coherent Anti-Stokes Spectra Using Embedded Physics
Transferring the recent advancements in deep learning into scientific disciplines is hindered by the lack of the required large-scale datasets for training. We argue that in these knowledge-rich domains, the established body of scientific theory provides reliable inductive biases in the form of governing physical laws. We address the ill-posed inverse problem of recovering Raman spectra from noisy Coherent Anti-Stokes Raman Scattering (CARS) measurements, as the true Raman signal here is suppressed by a dominating non-resonant background. We propose RamPINN, a model that learns to recover Raman spectra from given CARS spectra. Our core methodological contribution is a physics-informed neural network that utilizes a dual-decoder architecture to disentangle resonant and non-resonant signals. This is done by enforcing the Kramers-Kronig causality relations via a differentiable Hilbert transform loss on the resonant and a smoothness prior on the non-resonant part of the signal. Trained entirely on synthetic data, RamPINN demonstrates strong zero-shot generalization to real-world experimental data, explicitly closing this gap and significantly outperforming existing baselines. Furthermore, we show that training with these physics-based losses alone, without access to any ground-truth Raman spectra, still yields competitive results. This work highlights a broader concept: formal scientific rules can act as a potent inductive bias, enabling robust, self-supervised learning in data-limited scientific domains.
☆ Hybrid Quantum-Classical Policy Gradient for Adaptive Control of Cyber-Physical Systems: A Comparative Study of VQC vs. MLP
The comparative evaluation between classical and quantum reinforcement learning (QRL) paradigms was conducted to investigate their convergence behavior, robustness under observational noise, and computational efficiency in a benchmark control environment. The study employed a multilayer perceptron (MLP) agent as a classical baseline and a parameterized variational quantum circuit (VQC) as a quantum counterpart, both trained on the CartPole-v1 environment over 500 episodes. Empirical results demonstrated that the classical MLP achieved near-optimal policy convergence with a mean return of 498.7 +/- 3.2, maintaining stable equilibrium throughout training. In contrast, the VQC exhibited limited learning capability, with an average return of 14.6 +/- 4.8, primarily constrained by circuit depth and qubit connectivity. Noise robustness analysis further revealed that the MLP policy deteriorated gracefully under Gaussian perturbations, while the VQC displayed higher sensitivity at equivalent noise levels. Despite the lower asymptotic performance, the VQC exhibited significantly lower parameter count and marginally increased training time, highlighting its potential scalability for low-resource quantum processors. The results suggest that while classical neural policies remain dominant in current control benchmarks, quantum-enhanced architectures could offer promising efficiency advantages once hardware noise and expressivity limitations are mitigated.
comment: 6 pages, 5 figures, 2 tables, 17 equations, 1 algorithm
☆ Uncertainty in Machine Learning
This book chapter introduces the principles and practical applications of uncertainty quantification in machine learning. It explains how to identify and distinguish between different types of uncertainty and presents methods for quantifying uncertainty in predictive models, including linear regression, random forests, and neural networks. The chapter also covers conformal prediction as a framework for generating predictions with predefined confidence intervals. Finally, it explores how uncertainty estimation can be leveraged to improve business decision-making, enhance model reliability, and support risk-aware strategies.
comment: Authored by Hans Weytjens. Wouter Verbeke provided proofreading and served as the chief editor of the book in which this chapter appears
☆ Information-Theoretic Policy Pre-Training with Empowerment
Empowerment, an information-theoretic measure of an agent's potential influence on its environment, has emerged as a powerful intrinsic motivation and exploration framework for reinforcement learning (RL). Besides for unsupervised RL and skill learning algorithms, the specific use of empowerment as a pre-training signal has received limited attention in the literature. We show that empowerment can be used as a pre-training signal for data-efficient downstream task adaptation. For this we extend the traditional notion of empowerment by introducing discounted empowerment, which balances the agent's control over the environment across short- and long-term horizons. Leveraging this formulation, we propose a novel pre-training paradigm that initializes policies to maximize discounted empowerment, enabling agents to acquire a robust understanding of environmental dynamics. We analyze empowerment-based pre-training for various existing RL algorithms and empirically demonstrate its potential as a general-purpose initialization strategy: empowerment-maximizing policies with long horizons are data-efficient and effective, leading to improved adaptability in downstream tasks. Our findings pave the way for future research to scale this framework to high-dimensional and complex tasks, further advancing the field of RL.
☆ Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs
Large Language Models (LLMs) are increasingly applied to complex tasks that require extended reasoning. In such settings, models often benefit from diverse chains-of-thought to arrive at multiple candidate solutions. This requires two competing objectives: to inject enough stochasticity to explore multiple reasoning chains, and to ensure sufficient accuracy and quality in each path. Existing works pursue the first objective by increasing exploration at highly uncertain steps with higher temperature or larger candidate token sets, while others improve reliability by rejecting samples with low confidence post-generation, implying that low confidence correlates with low answer quality. These two lines of thought are in conflict, as they conflate different sources of uncertainty. To resolve this, we argue that the decoding rule should be calibrated by correctness, not confidence alone. We should sample from tokens with higher estimated correctness, and reduce sampling where expected correctness is low. We propose simple strategies that achieve this goal: Greedy-Threshold makes sampling greedy at very low confidence steps. Calibrated-TopK and Calibrated-epsilon set truncation threshold based on estimated rank-wise correctness. Together, our findings challenge prevailing heuristics about decoding under uncertainty and show gains across math and general reasoning benchmarks.
☆ Diffusion Models for Low-Light Image Enhancement: A Multi-Perspective Taxonomy and Performance Analysis
Low-light image enhancement (LLIE) is vital for safety-critical applications such as surveillance, autonomous navigation, and medical imaging, where visibility degradation can impair downstream task performance. Recently, diffusion models have emerged as a promising generative paradigm for LLIE due to their capacity to model complex image distributions via iterative denoising. This survey provides an up-to-date critical analysis of diffusion models for LLIE, distinctively featuring an in-depth comparative performance evaluation against Generative Adversarial Network and Transformer-based state-of-the-art methods, a thorough examination of practical deployment challenges, and a forward-looking perspective on the role of emerging paradigms like foundation models. We propose a multi-perspective taxonomy encompassing six categories: Intrinsic Decomposition, Spectral & Latent, Accelerated, Guided, Multimodal, and Autonomous; that map enhancement methods across physical priors, conditioning schemes, and computational efficiency. Our taxonomy is grounded in a hybrid view of both the model mechanism and the conditioning signals. We evaluate qualitative failure modes, benchmark inconsistencies, and trade-offs between interpretability, generalization, and inference efficiency. We also discuss real-world deployment constraints (e.g., memory, energy use) and ethical considerations. This survey aims to guide the next generation of diffusion-based LLIE research by highlighting trends and surfacing open research questions, including novel conditioning, real-time adaptation, and the potential of foundation models.
☆ Gaussian Embeddings: How JEPAs Secretly Learn Your Data Density
Joint Embedding Predictive Architectures (JEPAs) learn representations able to solve numerous downstream tasks out-of-the-box. JEPAs combine two objectives: (i) a latent-space prediction term, i.e., the representation of a slightly perturbed sample must be predictable from the original sample's representation, and (ii) an anti-collapse term, i.e., not all samples should have the same representation. While (ii) is often considered as an obvious remedy to representation collapse, we uncover that JEPAs' anti-collapse term does much more--it provably estimates the data density. In short, any successfully trained JEPA can be used to get sample probabilities, e.g., for data curation, outlier detection, or simply for density estimation. Our theoretical finding is agnostic of the dataset and architecture used--in any case one can compute the learned probabilities of sample $x$ efficiently and in closed-form using the model's Jacobian matrix at $x$. Our findings are empirically validated across datasets (synthetic, controlled, and Imagenet) and across different Self Supervised Learning methods falling under the JEPA family (I-JEPA and DINOv2) and on multimodal models, such as MetaCLIP. We denote the method extracting the JEPA learned density as {\bf JEPA-SCORE}.
☆ N-Parties Private Structure and Parameter Learning for Sum-Product Networks
A sum-product network (SPN) is a graphical model that allows several types of probabilistic inference to be performed efficiently. In this paper, we propose a privacy-preserving protocol which tackles structure generation and parameter learning of SPNs. Additionally, we provide a protocol for private inference on SPNs, subsequent to training. To preserve the privacy of the participants, we derive our protocol based on secret sharing, which guarantees privacy in the honest-but-curious setting even when at most half of the parties cooperate to disclose the data. The protocol makes use of a forest of randomly generated SPNs, which is trained and weighted privately and can then be used for private inference on data points. Our experiments indicate that preserving the privacy of all participants does not decrease log-likelihood performance on both homogeneously and heterogeneously partitioned data. We furthermore show that our protocol's performance is comparable to current state-of-the-art SPN learners in homogeneously partitioned data settings. In terms of runtime and memory usage, we demonstrate that our implementation scales well when increasing the number of parties, comparing favorably to protocols for neural networks, when they are trained to reproduce the input-output behavior of SPNs.
☆ EARL: Efficient Agentic Reinforcement Learning Systems for Large Language Models
Reinforcement learning (RL) has become a pivotal component of large language model (LLM) post-training, and agentic RL extends this paradigm to operate as agents through multi-turn interaction and tool use. Scaling such systems exposes two practical bottlenecks: (1) context length grows rapidly during training, inflating memory usage and latency, and triggering out-of-memory (OOM) failures; and (2) intermediate tensors accumulate with context length, making cross-device data movement a major system bottleneck. We present EARL, a scalable system for efficient agentic RL. EARL designs a parallelism selector that dynamically adapts model and training parallelism across RL stages based on sequence length and system load, and a data dispatcher that performs layout-aware, decentralized exchange of intermediate data batches. Together, these components increase throughput, reduce long-context failures, and enable stable large-scale training of agentic LLMs without relying on hard limits or penalties of context length.
☆ LLM-FS-Agent: A Deliberative Role-based Large Language Model Architecture for Transparent Feature Selection
High-dimensional data remains a pervasive challenge in machine learning, often undermining model interpretability and computational efficiency. While Large Language Models (LLMs) have shown promise for dimensionality reduction through feature selection, existing LLM-based approaches frequently lack structured reasoning and transparent justification for their decisions. This paper introduces LLM-FS-Agent, a novel multi-agent architecture designed for interpretable and robust feature selection. The system orchestrates a deliberative "debate" among multiple LLM agents, each assigned a specific role, enabling collective evaluation of feature relevance and generation of detailed justifications. We evaluate LLM-FS-Agent in the cybersecurity domain using the CIC-DIAD 2024 IoT intrusion detection dataset and compare its performance against strong baselines, including LLM-Select and traditional methods such as PCA. Experimental results demonstrate that LLM-FS-Agent consistently achieves superior or comparable classification performance while reducing downstream training time by an average of 46% (statistically significant improvement, p = 0.028 for XGBoost). These findings highlight that the proposed deliberative architecture enhances both decision transparency and computational efficiency, establishing LLM-FS-Agent as a practical and reliable solution for real-world applications.
☆ Carré du champ flow matching: better quality-generalisation tradeoff in generative models
Deep generative models often face a fundamental tradeoff: high sample quality can come at the cost of memorisation, where the model reproduces training data rather than generalising across the underlying data geometry. We introduce Carr\'e du champ flow matching (CDC-FM), a generalisation of flow matching (FM), that improves the quality-generalisation tradeoff by regularising the probability path with a geometry-aware noise. Our method replaces the homogeneous, isotropic noise in FM with a spatially varying, anisotropic Gaussian noise whose covariance captures the local geometry of the latent data manifold. We prove that this geometric noise can be optimally estimated from the data and is scalable to large data. Further, we provide an extensive experimental evaluation on diverse datasets (synthetic manifolds, point clouds, single-cell genomics, animal motion capture, and images) as well as various neural network architectures (MLPs, CNNs, and transformers). We demonstrate that CDC-FM consistently offers a better quality-generalisation tradeoff. We observe significant improvements over standard FM in data-scarce regimes and in highly non-uniformly sampled datasets, which are often encountered in AI for science applications. Our work provides a mathematical framework for studying the interplay between data geometry, generalisation and memorisation in generative models, as well as a robust and scalable algorithm that can be readily integrated into existing flow matching pipelines.
Prompt reinforcing for long-term planning of large language models
Large language models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks and can be adapted through prompting. However, they remain suboptimal in multi-turn interactions, often relying on incorrect early assumptions and failing to track user goals over time, which makes such tasks particularly challenging. Prior works in dialogue systems have shown that long-term planning is essential for handling interactive tasks. In this work, we propose a prompt optimisation framework inspired by reinforcement learning, which enables such planning to take place by only modifying the task instruction prompt of the LLM-based agent. By generating turn-by-turn feedback and leveraging experience replay for prompt rewriting, our proposed method shows significant improvement in multi-turn tasks such as text-to-SQL and task-oriented dialogue. Moreover, it generalises across different LLM-based agents and can leverage diverse LLMs as meta-prompting agents. This warrants future research in reinforcement learning-inspired parameter-free optimisation methods.
☆ An Attention-Augmented VAE-BiLSTM Framework for Anomaly Detection in 12-Lead ECG Signals
Anomaly detection in 12-lead electrocardiograms (ECGs) is critical for identifying deviations associated with cardiovascular disease. This work presents a comparative analysis of three autoencoder-based architectures: convolutional autoencoder (CAE), variational autoencoder with bidirectional long short-term memory (VAE-BiLSTM), and VAE-BiLSTM with multi-head attention (VAE-BiLSTM-MHA), for unsupervised anomaly detection in ECGs. To the best of our knowledge, this study reports the first application of a VAE-BiLSTM-MHA architecture to ECG anomaly detection. All models are trained on normal ECG samples to reconstruct non-anomalous cardiac morphology and detect deviations indicative of disease. Using a unified preprocessing and evaluation pipeline on the public China Physiological Signal Challenge (CPSC) dataset, the attention-augmented VAE achieves the best performance, with an AUPRC of 0.81 and a recall of 0.85 on the held-out test set, outperforming the other architectures. To support clinical triage, this model is further integrated into an interactive dashboard that visualizes anomaly localization. In addition, a performance comparison with baseline models from the literature is provided.
comment: 14 pages, 11 figures
☆ Kaputt: A Large-Scale Dataset for Visual Defect Detection ICCV 2025
We present a novel large-scale dataset for defect detection in a logistics setting. Recent work on industrial anomaly detection has primarily focused on manufacturing scenarios with highly controlled poses and a limited number of object categories. Existing benchmarks like MVTec-AD [6] and VisA [33] have reached saturation, with state-of-the-art methods achieving up to 99.9% AUROC scores. In contrast to manufacturing, anomaly detection in retail logistics faces new challenges, particularly in the diversity and variability of object pose and appearance. Leading anomaly detection methods fall short when applied to this new setting. To bridge this gap, we introduce a new benchmark that overcomes the current limitations of existing datasets. With over 230,000 images (and more than 29,000 defective instances), it is 40 times larger than MVTec-AD and contains more than 48,000 distinct objects. To validate the difficulty of the problem, we conduct an extensive evaluation of multiple state-of-the-art anomaly detection methods, demonstrating that they do not surpass 56.96% AUROC on our dataset. Further qualitative analysis confirms that existing methods struggle to leverage normal samples under heavy pose and appearance variation. With our large-scale dataset, we set a new benchmark and encourage future research towards solving this challenging problem in retail logistics anomaly detection. The dataset is available for download under https://www.kaputt-dataset.com.
comment: Accepted to ICCV 2025
☆ Paying Attention to Hybrid Attention: Untangling the Issues with Conversion Methods
Transformers' quadratic computational complexity limits their scalability despite remarkable performance. While linear attention reduces this to linear complexity, pre-training such models from scratch remains, in most cases, prohibitively expensive. Recent post-training linearisation methods convert pre-trained Transformers to linear models efficiently, often using hybrid approaches that combine linear attention with sliding-window softmax. We identify a critical flaw: existing hybrid methods inadvertently bypass the linear component, relying almost entirely on SWA. Component-level diagnostics reveal this previously undetected behaviour stems from overlooked evaluation practices on common-sense benchmarks. We propose three solutions to ensure balanced component usage: (i) inference-time hybridisation of linear-only conversions with sliding-window softmax; (ii) HedgeCATs, combining attention-weight transfer with targeted LoRA fine-tuning; and (iii) Scheduled Sliding-window Dropout (SSD), which stochastically suppresses the softmax branch during training to prevent component collapse. Our methods maintain computational efficiency while recovering most base model performance and ensuring genuine linear attention adoption, restoring the validity of performance attributions in hybrid conversions.
☆ Segment-Factorized Full-Song Generation on Symbolic Piano Music NeurIPS 2025
We propose the Segmented Full-Song Model (SFS) for symbolic full-song generation. The model accepts a user-provided song structure and an optional short seed segment that anchors the main idea around which the song is developed. By factorizing a song into segments and generating each one through selective attention to related segments, the model achieves higher quality and efficiency compared to prior work. To demonstrate its suitability for human-AI interaction, we further wrap SFS into a web application that enables users to iteratively co-create music on a piano roll with customizable structures and flexible ordering.
comment: Accepted to the 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: AI for Music
☆ OBSR: Open Benchmark for Spatial Representations SP
GeoAI is evolving rapidly, fueled by diverse geospatial datasets like traffic patterns, environmental data, and crowdsourced OpenStreetMap (OSM) information. While sophisticated AI models are being developed, existing benchmarks are often concentrated on single tasks and restricted to a single modality. As such, progress in GeoAI is limited by the lack of a standardized, multi-task, modality-agnostic benchmark for their systematic evaluation. This paper introduces a novel benchmark designed to assess the performance, accuracy, and efficiency of geospatial embedders. Our benchmark is modality-agnostic and comprises 7 distinct datasets from diverse cities across three continents, ensuring generalizability and mitigating demographic biases. It allows for the evaluation of GeoAI embedders on various phenomena that exhibit underlying geographic processes. Furthermore, we establish a simple and intuitive task-oriented model baselines, providing a crucial reference point for comparing more complex solutions.
comment: ACM SIGSPATIAL 2025 Full Paper
☆ MaNGO - Adaptable Graph Network Simulators via Meta-Learning NeurIPS 2025
Accurately simulating physics is crucial across scientific domains, with applications spanning from robotics to materials science. While traditional mesh-based simulations are precise, they are often computationally expensive and require knowledge of physical parameters, such as material properties. In contrast, data-driven approaches like Graph Network Simulators (GNSs) offer faster inference but suffer from two key limitations: Firstly, they must be retrained from scratch for even minor variations in physical parameters, and secondly they require labor-intensive data collection for each new parameter setting. This is inefficient, as simulations with varying parameters often share a common underlying latent structure. In this work, we address these challenges by learning this shared structure through meta-learning, enabling fast adaptation to new physical parameters without retraining. To this end, we propose a novel architecture that generates a latent representation by encoding graph trajectories using conditional neural processes (CNPs). To mitigate error accumulation over time, we combine CNPs with a novel neural operator architecture. We validate our approach, Meta Neural Graph Operator (MaNGO), on several dynamics prediction tasks with varying material properties, demonstrating superior performance over existing GNS methods. Notably, MaNGO achieves accuracy on unseen material properties close to that of an oracle model.
comment: 19 pages including appendix. NeurIPS 2025 (preprint version)
☆ Towards Label-Free Biological Reasoning Synthetic Dataset Creation via Uncertainty Filtering
Synthetic chain-of-thought (CoT) traces are widely used to train large reasoning models (LRMs), improving generalization by providing step-level supervision. Yet most approaches require ground-truth labels to seed or filter these traces - an expensive bottleneck in domains like biology where wet-lab data are scarce. We propose a label-free alternative: uncertainty-based filtering, which uses a model's own confidence - quantified through established uncertainty metrics like self-consistency and predictive perplexity - as a substitute for external labels. We sample multiple reasoning traces and retain only low-uncertainty subsets. Applied to biological perturbation prediction, a domain where wet-lab labels are especially costly, we show that the filtered subset has higher accuracy, and that supervised fine-tuning (SFT) on uncertainty-filtered data outperforms unfiltered synthetic data, narrows the gap to ground-truth training, and surpasses strong LRM baselines. Ablations show that per-class filtering corrects for class-specific uncertainty scales and that hybrid uncertainty metrics yield higher-quality datasets. Our results suggest that model-internal confidence is a powerful signal for efficient reasoning dataset creation, enabling LRMs in domains where supervision is expensive.
☆ DACP: Domain-Adaptive Continual Pre-Training of Large Language Models for Phone Conversation Summarization EMNLP 2025
Large language models (LLMs) have achieved impressive performance in text summarization, yet their performance often falls short when applied to specialized domains %or conversational data that differ from their original pre-training distribution. While fine-tuning can improve summarization quality, it typically relies on costly and scarce high-quality labeled data. In this work, we explore continual pre-training as a scalable, self-supervised approach to adapt LLMs for downstream summarization tasks, particularly in the context of noisy real-world conversation transcripts. We conduct extensive experiments using large-scale, unlabeled business conversation data to investigate whether continual pre-training enhances model capabilities in conversational summarization. Our results demonstrate that continual pre-training yields substantial gains in both in-domain and out-of-domain summarization benchmarks, while maintaining strong generalization and robustness. We also analyze the effects of data selection strategies, providing practical guidelines for applying continual pre-training in summarization-focused industrial applications.
comment: Accepted to the NewSumm Workshop at EMNLP 2025
☆ How to model Human Actions distribution with Event Sequence Data
This paper studies forecasting of the future distribution of events in human action sequences, a task essential in domains like retail, finance, healthcare, and recommendation systems where the precise temporal order is often less critical than the set of outcomes. We challenge the dominant autoregressive paradigm and investigate whether explicitly modeling the future distribution or order-invariant multi-token approaches outperform order-preserving methods. We analyze local order invariance and introduce a KL-based metric to quantify temporal drift. We find that a simple explicit distribution forecasting objective consistently surpasses complex implicit baselines. We further demonstrate that mode collapse of predicted categories is primarily driven by distributional imbalance. This work provides a principled framework for selecting modeling strategies and offers practical guidance for building more accurate and robust forecasting systems.
comment: 9 pages main text + 2 pages references + 6 pages appendix, 10 figures, 3 tables. Preprint version
☆ ESS-Flow: Training-free guidance of flow-based models as inference in source space
Guiding pretrained flow-based generative models for conditional generation or to produce samples with desired target properties enables solving diverse tasks without retraining on paired data. We present ESS-Flow, a gradient-free method that leverages the typically Gaussian prior of the source distribution in flow-based models to perform Bayesian inference directly in the source space using Elliptical Slice Sampling. ESS-Flow only requires forward passes through the generative model and observation process, no gradient or Jacobian computations, and is applicable even when gradients are unreliable or unavailable, such as with simulation-based observations or quantization in the generation or observation process. We demonstrate its effectiveness on designing materials with desired target properties and predicting protein structures from sparse inter-residue distance measurements.
comment: 14 pages, 12 figures. Code will be made available after publication
☆ Multimodal Trajectory Representation Learning for Travel Time Estimation
Accurate travel time estimation (TTE) plays a crucial role in intelligent transportation systems. However, it remains challenging due to heterogeneous data sources and complex traffic dynamics. Moreover, conventional approaches typically convert trajectories into fixed-length representations, neglecting the inherent variability of real-world trajectories, which often leads to information loss or feature redundancy. To address these challenges, this paper introduces the Multimodal Dynamic Trajectory Integration (MDTI) framework--a novel multimodal trajectory representation learning approach that integrates GPS sequences, grid trajectories, and road network constraints to enhance TTE accuracy. MDTI employs modality-specific encoders and a cross-modal interaction module to capture complementary spatial, temporal, and topological semantics, while a dynamic trajectory modeling mechanism adaptively regulates information density for trajectories of varying lengths. Two self-supervised pretraining objectives, named contrastive alignment and masked language modeling, further strengthen multimodal consistency and contextual understanding. Extensive experiments on three real-world datasets demonstrate that MDTI consistently outperforms state-of-the-art baselines, confirming its robustness and strong generalization abilities. The code is publicly available at: https://github.com/freshhxy/MDTI/
☆ FoleyGRAM: Video-to-Audio Generation with GRAM-Aligned Multimodal Encoders IJCNN 2025
In this work, we present FoleyGRAM, a novel approach to video-to-audio generation that emphasizes semantic conditioning through the use of aligned multimodal encoders. Building on prior advancements in video-to-audio generation, FoleyGRAM leverages the Gramian Representation Alignment Measure (GRAM) to align embeddings across video, text, and audio modalities, enabling precise semantic control over the audio generation process. The core of FoleyGRAM is a diffusion-based audio synthesis model conditioned on GRAM-aligned embeddings and waveform envelopes, ensuring both semantic richness and temporal alignment with the corresponding input video. We evaluate FoleyGRAM on the Greatest Hits dataset, a standard benchmark for video-to-audio models. Our experiments demonstrate that aligning multimodal encoders using GRAM enhances the system's ability to semantically align generated audio with video content, advancing the state of the art in video-to-audio synthesis.
comment: Acepted at IJCNN 2025
☆ StereoSync: Spatially-Aware Stereo Audio Generation from Video IJCNN 2025
Although audio generation has been widely studied over recent years, video-aligned audio generation still remains a relatively unexplored frontier. To address this gap, we introduce StereoSync, a novel and efficient model designed to generate audio that is both temporally synchronized with a reference video and spatially aligned with its visual context. Moreover, StereoSync also achieves efficiency by leveraging pretrained foundation models, reducing the need for extensive training while maintaining high-quality synthesis. Unlike existing methods that primarily focus on temporal synchronization, StereoSync introduces a significant advancement by incorporating spatial awareness into video-aligned audio generation. Indeed, given an input video, our approach extracts spatial cues from depth maps and bounding boxes, using them as cross-attention conditioning in a diffusion-based audio generation model. Such an approach allows StereoSync to go beyond simple synchronization, producing stereo audio that dynamically adapts to the spatial structure and movement of a video scene. We evaluate StereoSync on Walking The Maps, a curated dataset comprising videos from video games that feature animated characters walking through diverse environments. Experimental results demonstrate the ability of StereoSync to achieve both temporal and spatial alignment, advancing the state of the art in video-to-audio generation and resulting in a significantly more immersive and realistic audio experience.
comment: Accepted at IJCNN 2025
☆ Mitigating Premature Exploitation in Particle-based Monte Carlo for Inference-Time Scaling
Inference-Time Scaling (ITS) improves language models by allocating more computation at generation time. Particle Filtering (PF) has emerged as a strong ITS method for complex mathematical reasoning tasks, but it is vulnerable when guided by process reward models, which often assign overconfident scores early in the reasoning process. This causes PF to suffer from premature exploitation: it myopically commits to locally promising trajectories, prunes potentially correct hypotheses, and converges to suboptimal solutions. This failure mode, known as particle impoverishment, is especially severe under constrained computational budgets. To address this, we analyze the problem and identify two root causes: a lack of diversity in the particle set due to overconfident resampling and consequent inability to assess the potential of a reasoning path. We introduce Entropic Particle Filtering (ePF), an algorithm that integrates two new techniques to solve these issues. The first technique, Entropic Annealing (EA), directly mitigates particle impoverishment by monitoring search diversity via entropy; when diversity drops, it intervenes by dynamically annealing the resampling distribution to preserve exploration. The second, an enhancement called Look-ahead Modulation (LaM), adds a predictive guide to evaluate a state's potential based on its successors. On several challenging math benchmarks, ePF significantly outperforms strong baselines and achieves up to a 50 % relative improvement in task reward. Together, these methods improve PF's resilience by balancing the exploration of diverse solution spaces with the exploitation of high-reward regions, ultimately leading to higher-quality solutions.
☆ Improving Clinical Dataset Condensation with Mode Connectivity-based Trajectory Surrogates AISTATS 2026
Dataset condensation (DC) enables the creation of compact, privacy-preserving synthetic datasets that can match the utility of real patient records, supporting democratised access to highly regulated clinical data for developing downstream clinical models. State-of-the-art DC methods supervise synthetic data by aligning the training dynamics of models trained on real and those trained on synthetic data, typically using full stochastic gradient descent (SGD) trajectories as alignment targets; however, these trajectories are often noisy, high-curvature, and storage-intensive, leading to unstable gradients, slow convergence, and substantial memory overhead. We address these limitations by replacing full SGD trajectories with smooth, low-loss parametric surrogates, specifically quadratic B\'ezier curves that connect the initial and final model states from real training trajectories. These mode-connected paths provide noise-free, low-curvature supervision signals that stabilise gradients, accelerate convergence, and eliminate the need for dense trajectory storage. We theoretically justify B\'ezier-mode connections as effective surrogates for SGD paths and empirically show that the proposed method outperforms state-of-the-art condensation approaches across five clinical datasets, yielding condensed datasets that enable clinically effective model development.
comment: 20 pages, 4 figures, Submitted to AISTATS 2026
☆ Mellum: Production-Grade in-IDE Contextual Code Completion with Multi-File Project Understanding
We present the Mellum models family, open-weight code completion models designed for interactive use in JetBrains IDEs. Mellums have 4B parameters, adopt a Llama-style architecture, and are pre-trained on ~4T tokens of permissively licensed, multi-language code. Our studies show that (i) careful data curation and staged training significantly improve the model's quality, (ii) editor-critical capabilities such as context packing are necessary for high-quality suggestions, and (iii) a compact, task-focused model can meet the cost and latency constraints of interactive completion. In the paper, we describe an end-to-end industrial pipeline for producing contextualized in-editor completion: disciplined data governance, multi-stage training that includes fill-in-the-middle and project context via supervised fine-tuning, and alignment via direct preference optimization using feedback from real-world scenarios. Our quality evaluations include both large-scale offline benchmarks and online telemetry from production deployments in JetBrains IDEs. Mellums are released under the Apache-2.0 license on HuggingFace, with a public model card providing a reproducible reference for practitioners. Our experience offers a pragmatic blueprint for taking a focused, open model from a research prototype to at scale production for hundreds of thousands of users.
comment: 11 pages, 4 figures, 3 tables
☆ Möbius transforms and Shapley values for vector-valued functions on weighted directed acyclic multigraphs
We generalize the concept of M\"obius inversion and Shapley values to directed acyclic multigraphs and weighted versions thereof. We further allow value functions (games) and thus their M\"obius transforms (synergy function) and Shapley values to have values in any abelian group that is a module over a ring that contains the graph weights, e.g. vector-valued functions. To achieve this and overcome the obstruction that the classical axioms (linearity, efficiency, null player, symmetry) are not strong enough to uniquely determine Shapley values in this more general setting, we analyze Shapley values from two novel points of view: 1) We introduce projection operators that allow us to interpret Shapley values as the recursive projection and re-attribution of higher-order synergies to lower-order ones; 2) we propose a strengthening of the null player axiom and a localized symmetry axiom, namely the weak elements and flat hierarchy axioms. The former allows us to remove coalitions with vanishing synergy while preserving the rest of the hierarchical structure. The latter treats player-coalition bonds uniformly in the corner case of hierarchically flat graphs. Together with linearity these axioms already imply a unique explicit formula for the Shapley values, as well as classical properties like efficiency, null player, symmetry, and novel ones like the projection property. This whole framework then specializes to finite inclusion algebras, lattices, partial orders and mereologies, and also recovers certain previously known cases as corner cases, and presents others from a new perspective. The admission of general weighted directed acyclic multigraph structured hierarchies and vector-valued functions and Shapley values opens up the possibility for new analytic tools and application areas, like machine learning, language processing, explainable artificial intelligence, and many more.
comment: 43 pages, 2 figures
☆ DP-SNP-TIHMM: Differentially Private, Time-Inhomogeneous Hidden Markov Models for Synthesizing Genome-Wide Association Datasets
Single nucleotide polymorphism (SNP) datasets are fundamental to genetic studies but pose significant privacy risks when shared. The correlation of SNPs with each other makes strong adversarial attacks such as masked-value reconstruction, kin, and membership inference attacks possible. Existing privacy-preserving approaches either apply differential privacy to statistical summaries of these datasets or offer complex methods that require post-processing and the usage of a publicly available dataset to suppress or selectively share SNPs. In this study, we introduce an innovative framework for generating synthetic SNP sequence datasets using samples derived from time-inhomogeneous hidden Markov models (TIHMMs). To preserve the privacy of the training data, we ensure that each SNP sequence contributes only a bounded influence during training, enabling strong differential privacy guarantees. Crucially, by operating on full SNP sequences and bounding their gradient contributions, our method directly addresses the privacy risks introduced by their inherent correlations. Through experiments conducted on the real-world 1000 Genomes dataset, we demonstrate the efficacy of our method using privacy budgets of $\varepsilon \in [1, 10]$ at $\delta=10^{-4}$. Notably, by allowing the transition models of the HMM to be dependent on the location in the sequence, we significantly enhance performance, enabling the synthetic datasets to closely replicate the statistical properties of non-private datasets. This framework facilitates the private sharing of genomic data while offering researchers exceptional flexibility and utility.
☆ Transcribing Rhythmic Patterns of the Guitar Track in Polyphonic Music SP
Whereas chord transcription has received considerable attention during the past couple of decades, far less work has been devoted to transcribing and encoding the rhythmic patterns that occur in a song. The topic is especially relevant for instruments such as the rhythm guitar, which is typically played by strumming rhythmic patterns that repeat and vary over time. However, in many cases one cannot objectively define a single "right" rhythmic pattern for a given song section. To create a dataset with well-defined ground-truth labels, we asked expert musicians to transcribe the rhythmic patterns in 410 popular songs and record cover versions where the guitar tracks followed those transcriptions. To transcribe the strums and their corresponding rhythmic patterns, we propose a three-step framework. Firstly, we perform approximate stem separation to extract the guitar part from the polyphonic mixture. Secondly, we detect individual strums within the separated guitar audio, using a pre-trained foundation model (MERT) as a backbone. Finally, we carry out a pattern-decoding process in which the transcribed sequence of guitar strums is represented by patterns drawn from an expert-curated vocabulary. We show that it is possible to transcribe the rhythmic patterns of the guitar track in polyphonic music with quite high accuracy, producing a representation that is human-readable and includes automatically detected bar lines and time signature markers. We perform ablation studies and error analysis and propose a set of evaluation metrics to assess the accuracy and readability of the predicted rhythmic pattern sequence.
comment: Accepted to WASPAA 2025
☆ Empirical Comparison of Membership Inference Attacks in Deep Transfer Learning
With the emergence of powerful large-scale foundation models, the training paradigm is increasingly shifting from from-scratch training to transfer learning. This enables high utility training with small, domain-specific datasets typical in sensitive applications.Membership inference attacks (MIAs) provide an empirical estimate of the privacy leakage by machine learning models. Yet, prior assessments of MIAs against models fine-tuned with transfer learning rely on a small subset of possible attacks. We address this by comparing performance of diverse MIAs in transfer learning settings to help practitioners identify the most efficient attacks for privacy risk evaluation. We find that attack efficacy decreases with the increase in training data for score-based MIAs. We find that there is no one MIA which captures all privacy risks in models trained with transfer learning. While the Likelihood Ratio Attack (LiRA) demonstrates superior performance across most experimental scenarios, the Inverse Hessian Attack (IHA) proves to be more effective against models fine-tuned on PatchCamelyon dataset in high data regime.
comment: 30 pages, 13 figures, published in TMLR https://openreview.net/forum?id=UligTUCgdt
☆ Uncertainty assessment in satellite-based greenhouse gas emissions estimates using emulated atmospheric transport
Monitoring greenhouse gas emissions and evaluating national inventories require efficient, scalable, and reliable inference methods. Top-down approaches, combined with recent advances in satellite observations, provide new opportunities to evaluate emissions at continental and global scales. However, transport models used in these methods remain a key source of uncertainty: they are computationally expensive to run at scale, and their uncertainty is difficult to characterise. Artificial intelligence offers a dual opportunity to accelerate transport simulations and to quantify their associated uncertainty. We present an ensemble-based pipeline for estimating atmospheric transport "footprints", greenhouse gas mole fraction measurements, and their uncertainties using a graph neural network emulator of a Lagrangian Particle Dispersion Model (LPDM). The approach is demonstrated with GOSAT (Greenhouse Gases Observing Satellite) observations for Brazil in 2016. The emulator achieved a ~1000x speed-up over the NAME LPDM, while reproducing large-scale footprint structures. Ensembles were calculated to quantify absolute and relative uncertainty, revealing spatial correlations with prediction error. The results show that ensemble spread highlights low-confidence spatial and temporal predictions for both atmospheric transport footprints and methane mole fractions. While demonstrated here for an LPDM emulator, the approach could be applied more generally to atmospheric transport models, supporting uncertainty-aware greenhouse gas inversion systems and improving the robustness of satellite-based emissions monitoring. With further development, ensemble-based emulators could also help explore systematic LPDM errors, offering a computationally efficient pathway towards a more comprehensive uncertainty budget in greenhouse gas flux estimates.
☆ Are Heterogeneous Graph Neural Networks Truly Effective? A Causal Perspective
Graph neural networks (GNNs) have achieved remarkable success in node classification. Building on this progress, heterogeneous graph neural networks (HGNNs) integrate relation types and node and edge semantics to leverage heterogeneous information. Causal analysis for HGNNs is advancing rapidly, aiming to separate genuine causal effects from spurious correlations. However, whether HGNNs are intrinsically effective remains underexamined, and most studies implicitly assume rather than establish this effectiveness. In this work, we examine HGNNs from two perspectives: model architecture and heterogeneous information. We conduct a systematic reproduction across 21 datasets and 20 baselines, complemented by comprehensive hyperparameter retuning. To further disentangle the source of performance gains, we develop a causal effect estimation framework that constructs and evaluates candidate factors under standard assumptions through factual and counterfactual analyses, with robustness validated via minimal sufficient adjustment sets, cross-method consistency checks, and sensitivity analyses. Our results lead to two conclusions. First, model architecture and complexity have no causal effect on performance. Second, heterogeneous information exerts a positive causal effect by increasing homophily and local-global distribution discrepancy, which makes node classes more distinguishable. The implementation is publicly available at https://github.com/YXNTU/CausalHGNN.
☆ Communication Enables Cooperation in LLM Agents: A Comparison with Curriculum-Based Approaches
Eliciting cooperation in multi-agent LLM systems is critical for AI alignment. We investigate two approaches: direct communication and curriculum learning. In a 4-player Stag Hunt, a one-word "cheap talk" channel increases cooperation from 0% to 48.3%, demonstrating communication as a robust coordination mechanism. In contrast, we find that curriculum learning is highly sensitive to design choices: our pedagogical curriculum through progressively complex games reduced agent payoffs by 27.4% in an Iterated Public Goods Game with Punishment. Qualitative analysis reveals that curricula emphasizing defection-equilibrium games can induce "learned pessimism" in agents. These findings suggest that for coordination problems, simple communication protocols may be more reliable than experience-based training, and that curriculum design for social dilemmas requires careful attention to the strategic lessons embedded in game sequences.
☆ ARM: Discovering Agentic Reasoning Modules for Generalizable Multi-Agent Systems
Large Language Model (LLM)-powered Multi-agent systems (MAS) have achieved state-of-the-art results on various complex reasoning tasks. Recent works have proposed techniques to automate the design of MASes, eliminating the need for manual engineering. However, these techniques perform poorly, often achieving similar or inferior performance to simple baselines. Furthermore, they require computationally expensive re-discovery of architectures for each new task domain and expensive data annotation on domains without existing labeled validation sets. A critical insight is that simple Chain of Thought (CoT) reasoning often performs competitively with these complex systems, suggesting that the fundamental reasoning unit of MASes, CoT, warrants further investigation. To this end, we present a new paradigm for automatic MAS design that pivots the focus to optimizing CoT reasoning. We introduce the Agentic Reasoning Module (ARM), an agentic generalization of CoT where each granular reasoning step is executed by a specialized reasoning module. This module is discovered through a tree search over the code space, starting from a simple CoT module and evolved using mutations informed by reflection on execution traces. The resulting ARM acts as a versatile reasoning building block which can be utilized as a direct recursive loop or as a subroutine in a learned meta-orchestrator. Our approach significantly outperforms both manually designed MASes and state-of-the-art automatic MAS design methods. Crucially, MASes built with ARM exhibit superb generalization, maintaining high performance across different foundation models and task domains without further optimization.
comment: 29 pages, 2 figures
☆ Improving Discrete Diffusion Unmasking Policies Beyond Explicit Reference Policies
Masked diffusion models (MDMs) have recently emerged as a novel framework for language modeling. MDMs generate sentences by iteratively denoising masked sequences, filling in [MASK] tokens step by step. Although MDMs support any-order sampling, performance is highly sensitive to the choice of which position to unmask next. Prior work typically relies on rule-based schedules (e.g., max-confidence, max-margin), which provide ad hoc improvements. In contrast, we replace these heuristics with a learned scheduler. Specifically, we cast denoising as a KL-regularized Markov decision process (MDP) with an explicit reference policy and optimize a regularized objective that admits policy improvement and convergence guarantees under standard assumptions. We prove that the optimized policy under this framework generates samples that more closely match the data distribution than heuristic schedules. Empirically, across four benchmarks, our learned policy consistently outperforms max-confidence: for example, on SUDOKU, where unmasking order is critical, it yields a 20.1% gain over random and a 11.2% gain over max-confidence.
comment: Preprint
☆ Neighborhood-Adaptive Generalized Linear Graph Embedding with Latent Pattern Mining
Graph embedding has been widely applied in areas such as network analysis, social network mining, recommendation systems, and bioinformatics. However, current graph construction methods often require the prior definition of neighborhood size, limiting the effective revelation of potential structural correlations in the data. Additionally, graph embedding methods using linear projection heavily rely on a singular pattern mining approach, resulting in relative weaknesses in adapting to different scenarios. To address these challenges, we propose a novel model, Neighborhood-Adaptive Generalized Linear Graph Embedding (NGLGE), grounded in latent pattern mining. This model introduces an adaptive graph learning method tailored to the neighborhood, effectively revealing intrinsic data correlations. Simultaneously, leveraging a reconstructed low-rank representation and imposing $\ell_{2,0}$ norm constraint on the projection matrix allows for flexible exploration of additional pattern information. Besides, an efficient iterative solving algorithm is derived for the proposed model. Comparative evaluations on datasets from diverse scenarios demonstrate the superior performance of our model compared to state-of-the-art methods.
☆ DiffSDA: Unsupervised Diffusion Sequential Disentanglement Across Modalities
Unsupervised representation learning, particularly sequential disentanglement, aims to separate static and dynamic factors of variation in data without relying on labels. This remains a challenging problem, as existing approaches based on variational autoencoders and generative adversarial networks often rely on multiple loss terms, complicating the optimization process. Furthermore, sequential disentanglement methods face challenges when applied to real-world data, and there is currently no established evaluation protocol for assessing their performance in such settings. Recently, diffusion models have emerged as state-of-the-art generative models, but no theoretical formalization exists for their application to sequential disentanglement. In this work, we introduce the Diffusion Sequential Disentanglement Autoencoder (DiffSDA), a novel, modal-agnostic framework effective across diverse real-world data modalities, including time series, video, and audio. DiffSDA leverages a new probabilistic modeling, latent diffusion, and efficient samplers, while incorporating a challenging evaluation protocol for rigorous testing. Our experiments on diverse real-world benchmarks demonstrate that DiffSDA outperforms recent state-of-the-art methods in sequential disentanglement.
☆ Stable Robot Motions on Manifolds: Learning Lyapunov-Constrained Neural Manifold ODEs
Learning stable dynamical systems from data is crucial for safe and reliable robot motion planning and control. However, extending stability guarantees to trajectories defined on Riemannian manifolds poses significant challenges due to the manifold's geometric constraints. To address this, we propose a general framework for learning stable dynamical systems on Riemannian manifolds using neural ordinary differential equations. Our method guarantees stability by projecting the neural vector field evolving on the manifold so that it strictly satisfies the Lyapunov stability criterion, ensuring stability at every system state. By leveraging a flexible neural parameterisation for both the base vector field and the Lyapunov function, our framework can accurately represent complex trajectories while respecting manifold constraints by evolving solutions directly on the manifold. We provide an efficient training strategy for applying our framework and demonstrate its utility by solving Riemannian LASA datasets on the unit quaternion (S^3) and symmetric positive-definite matrix manifolds, as well as robotic motions evolving on \mathbb{R}^3 \times S^3. We demonstrate the performance, scalability, and practical applicability of our approach through extensive simulations and by learning robot motions in a real-world experiment.
comment: 12 pages, 6 figures
☆ Primal-Dual Direct Preference Optimization for Constrained LLM Alignment
The widespread application of Large Language Models (LLMs) imposes increasing demands on safety, such as reducing harmful content and fake information, and avoiding certain forbidden tokens due to rules and laws. While there have been several recent works studying safe alignment of LLMs, these works either require the training of reward and cost models and incur high memory and computational costs, or need prior knowledge about the optimal solution. Motivated by this fact, we study the problem of constrained alignment in LLMs, i.e., maximizing the output reward while restricting the cost due to potentially unsafe content to stay below a threshold. For this problem, we propose a novel primal-dual DPO approach, which first trains a model using standard DPO on reward preference data to provide reward information, and then adopts a rearranged Lagrangian DPO objective utilizing the provided reward information to fine-tune LLMs on cost preference data. Our approach significantly reduces memory and computational costs, and does not require extra prior knowledge. Moreover, we establish rigorous theoretical guarantees on the suboptimality and constraint violation of the output policy. We also extend our approach to an online data setting by incorporating exploration bonuses, which enables our approach to explore uncovered prompt-response space, and then provide theoretical results that get rid of the dependence on preference data coverage. Experimental results on the widely-used preference dataset PKU-SafeRLHF demonstrate the effectiveness of our approach.
☆ Sparse deepfake detection promotes better disentanglement
Due to the rapid progress of speech synthesis, deepfake detection has become a major concern in the speech processing community. Because it is a critical task, systems must not only be efficient and robust, but also provide interpretable explanations. Among the different approaches for explainability, we focus on the interpretation of latent representations. In such paper, we focus on the last layer of embeddings of AASIST, a deepfake detection architecture. We use a TopK activation inspired by SAEs on this layer to obtain sparse representations which are used in the decision process. We demonstrate that sparse deepfake detection can improve detection performance, with an EER of 23.36% on ASVSpoof5 test set, with 95% of sparsity. We then show that these representations provide better disentanglement, using completeness and modularity metrics based on mutual information. Notably, some attacks are directly encoded in the latent space.
☆ Oracle-Guided Masked Contrastive Reinforcement Learning for Visuomotor Policies
A prevailing approach for learning visuomotor policies is to employ reinforcement learning to map high-dimensional visual observations directly to action commands. However, the combination of high-dimensional visual inputs and agile maneuver outputs leads to long-standing challenges, including low sample efficiency and significant sim-to-real gaps. To address these issues, we propose Oracle-Guided Masked Contrastive Reinforcement Learning (OMC-RL), a novel framework designed to improve the sample efficiency and asymptotic performance of visuomotor policy learning. OMC-RL explicitly decouples the learning process into two stages: an upstream representation learning stage and a downstream policy learning stage. In the upstream stage, a masked Transformer module is trained with temporal modeling and contrastive learning to extract temporally-aware and task-relevant representations from sequential visual inputs. After training, the learned encoder is frozen and used to extract visual representations from consecutive frames, while the Transformer module is discarded. In the downstream stage, an oracle teacher policy with privileged access to global state information supervises the agent during early training to provide informative guidance and accelerate early policy learning. This guidance is gradually reduced to allow independent exploration as training progresses. Extensive experiments in simulated and real-world environments demonstrate that OMC-RL achieves superior sample efficiency and asymptotic policy performance, while also improving generalization across diverse and perceptually complex scenarios.
☆ vAttention: Verified Sparse Attention
State-of-the-art sparse attention methods for reducing decoding latency fall into two main categories: approximate top-$k$ (and its extension, top-$p$) and recently introduced sampling-based estimation. However, these approaches are fundamentally limited in their ability to approximate full attention: they fail to provide consistent approximations across heads and query vectors and, most critically, lack guarantees on approximation quality, limiting their practical deployment. We observe that top-$k$ and random sampling are complementary: top-$k$ performs well when attention scores are dominated by a few tokens, whereas random sampling provides better estimates when attention scores are relatively uniform. Building on this insight and leveraging the statistical guarantees of sampling, we introduce vAttention, the first practical sparse attention mechanism with user-specified $(\epsilon, \delta)$ guarantees on approximation accuracy (thus, verified). These guarantees make vAttention a compelling step toward practical, reliable deployment of sparse attention at scale. By unifying top-k and sampling, vAttention outperforms both individually, delivering a superior quality-efficiency trade-off. Our experiments show that vAttention significantly improves the quality of sparse attention (e.g., $\sim$4.5 percentage points for Llama-3.1-8B-Inst and Deepseek-R1-Distill-Llama-8B on RULER-HARD), and effectively bridges the gap between full and sparse attention (e.g., across datasets, it matches full model quality with upto 20x sparsity). We also demonstrate that it can be deployed in reasoning scenarios to achieve fast decoding without compromising model quality (e.g., vAttention achieves full model quality on AIME2024 at 10x sparsity with up to 32K token generations). Code is open-sourced at https://github.com/xAlg-ai/sparse-attention-hub.
☆ QGraphLIME - Explaining Quantum Graph Neural Networks
Quantum graph neural networks offer a powerful paradigm for learning on graph-structured data, yet their explainability is complicated by measurement-induced stochasticity and the combinatorial nature of graph structure. In this paper, we introduce QuantumGraphLIME (QGraphLIME), a model-agnostic, post-hoc framework that treats model explanations as distributions over local surrogates fit on structure-preserving perturbations of a graph. By aggregating surrogate attributions together with their dispersion, QGraphLIME yields uncertainty-aware node and edge importance rankings for quantum graph models. The framework further provides a distribution-free, finite-sample guarantee on the size of the surrogate ensemble: a Dvoretzky-Kiefer-Wolfowitz bound ensures uniform approximation of the induced distribution of a binary class probability at target accuracy and confidence under standard independence assumptions. Empirical studies on controlled synthetic graphs with known ground truth demonstrate accurate and stable explanations, with ablations showing clear benefits of nonlinear surrogate modeling and highlighting sensitivity to perturbation design. Collectively, these results establish a principled, uncertainty-aware, and structure-sensitive approach to explaining quantum graph neural networks, and lay the groundwork for scaling to broader architectures and real-world datasets, as quantum resources mature. Code is available at https://github.com/smlab-niser/qglime.
☆ Verifier-free Test-Time Sampling for Vision Language Action Models
Vision-Language-Action models (VLAs) have demonstrated remarkable performance in robot control. However, they remain fundamentally limited in tasks that require high precision due to their single-inference paradigm. While test-time scaling approaches using external verifiers have shown promise, they require additional training and fail to generalize to unseen conditions. We propose Masking Distribution Guided Selection (MG-Select), a novel test-time scaling framework for VLAs that leverages the model's internal properties without requiring additional training or external modules. Our approach utilizes KL divergence from a reference action token distribution as a confidence metric for selecting the optimal action from multiple candidates. We introduce a reference distribution generated by the same VLA but with randomly masked states and language conditions as inputs, ensuring maximum uncertainty while remaining aligned with the target task distribution. Additionally, we propose a joint training strategy that enables the model to learn both conditional and unconditional distributions by applying dropout to state and language conditions, thereby further improving the quality of the reference distribution. Our experiments demonstrate that MG-Select achieves significant performance improvements, including a 28%/35% improvement in real-world in-distribution/out-of-distribution tasks, along with a 168% relative gain on RoboCasa pick-and-place tasks trained with 30 demonstrations.
comment: 14 pages; 3 figures
☆ Inductive inference of gradient-boosted decision trees on graphs for insurance fraud detection
Graph-based methods are becoming increasingly popular in machine learning due to their ability to model complex data and relations. Insurance fraud is a prime use case, since false claims are often the result of organised criminals that stage accidents or the same persons filing erroneous claims on multiple policies. One challenge is that graph-based approaches struggle to find meaningful representations of the data because of the high class imbalance present in fraud data. Another is that insurance networks are heterogeneous and dynamic, given the changing relations among people, companies and policies. That is why gradient boosted tree approaches on tabular data still dominate the field. Therefore, we present a novel inductive graph gradient boosting machine (G-GBM) for supervised learning on heterogeneous and dynamic graphs. We show that our estimator competes with popular graph neural network approaches in an experiment using a variety of simulated random graphs. We demonstrate the power of G-GBM for insurance fraud detection using an open-source and a real-world, proprietary dataset. Given that the backbone model is a gradient boosting forest, we apply established explainability methods to gain better insights into the predictions made by G-GBM.
☆ Quantifying the Accuracy-Interpretability Trade-Off in Concept-Based Sidechannel Models
Concept Bottleneck Models (CBNMs) are deep learning models that provide interpretability by enforcing a bottleneck layer where predictions are based exclusively on human-understandable concepts. However, this constraint also restricts information flow and often results in reduced predictive accuracy. Concept Sidechannel Models (CSMs) address this limitation by introducing a sidechannel that bypasses the bottleneck and carry additional task-relevant information. While this improves accuracy, it simultaneously compromises interpretability, as predictions may rely on uninterpretable representations transmitted through sidechannels. Currently, there exists no principled technique to control this fundamental trade-off. In this paper, we close this gap. First, we present a unified probabilistic concept sidechannel meta-model that subsumes existing CSMs as special cases. Building on this framework, we introduce the Sidechannel Independence Score (SIS), a metric that quantifies a CSM's reliance on its sidechannel by contrasting predictions made with and without sidechannel information. We propose SIS regularization, which explicitly penalizes sidechannel reliance to improve interpretability. Finally, we analyze how the expressivity of the predictor and the reliance of the sidechannel jointly shape interpretability, revealing inherent trade-offs across different CSM architectures. Empirical results show that state-of-the-art CSMs, when trained solely for accuracy, exhibit low representation interpretability, and that SIS regularization substantially improves their interpretability, intervenability, and the quality of learned interpretable task predictors. Our work provides both theoretical and practical tools for developing CSMs that balance accuracy and interpretability in a principled manner.
☆ NEO: No-Optimization Test-Time Adaptation through Latent Re-Centering
Test-Time Adaptation (TTA) methods are often computationally expensive, require a large amount of data for effective adaptation, or are brittle to hyperparameters. Based on a theoretical foundation of the geometry of the latent space, we are able to significantly improve the alignment between source and distribution-shifted samples by re-centering target data embeddings at the origin. This insight motivates NEO -- a hyperparameter-free fully TTA method, that adds no significant compute compared to vanilla inference. NEO is able to improve the classification accuracy of ViT-Base on ImageNet-C from 55.6% to 59.2% after adapting on just one batch of 64 samples. When adapting on 512 samples NEO beats all 7 TTA methods we compare against on ImageNet-C, ImageNet-R and ImageNet-S and beats 6/7 on CIFAR-10-C, while using the least amount of compute. NEO performs well on model calibration metrics and additionally is able to adapt from 1 class to improve accuracy on 999 other classes in ImageNet-C. On Raspberry Pi and Jetson Orin Nano devices, NEO reduces inference time by 63% and memory usage by 9% compared to baselines. Our results based on 3 ViT architectures and 4 datasets show that NEO can be used efficiently and effectively for TTA.
♻ ☆ LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning
Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM's autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. We first construct a structured latent reasoning space using a Variational Autoencoder (VAE) that encodes text reasoning steps into blocks of thought tokens, preserving semantic information and interpretability while offering compact but expressive representations. Subsequently, we utilize a latent diffusion model that learns to denoise a block of latent thought tokens with a blockwise bidirectional attention mask, enabling longer horizon and iterative refinement with adaptive test-time compute. This design allows efficient parallel generation of diverse reasoning trajectories, allowing the model to plan and revise the reasoning process holistically. We conduct evaluations on a suite of mathematical reasoning and planning benchmarks. Empirical results show that LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods, revealing a new paradigm for text reasoning with latent diffusion.
♻ ☆ HOG-Diff: Higher-Order Guided Diffusion for Graph Generation
Graph generation is a critical yet challenging task as empirical analyses require a deep understanding of complex, non-Euclidean structures. Diffusion models have recently made significant achievements in graph generation, but these models are typically adapted from image generation frameworks and overlook inherent higher-order topology, leaving them ill-suited for capturing the topological properties of graphs. In this work, we propose Higher-order Guided Diffusion (HOG-Diff), a principled framework that progressively generates plausible graphs with inherent topological structures. HOG-Diff follows a coarse-to-fine generation curriculum guided by higher-order topology and implemented via diffusion bridges. We further prove that our model exhibits a stronger theoretical guarantee than classical diffusion frameworks. Extensive experiments on both molecular and generic graph generation tasks demonstrate that our method consistently outperforms or remains competitive with state-of-the-art baselines. Our code is available at https://github.com/Yiminghh/HOG-Diff.
♻ ☆ Hierarchical Reasoning Models: Perspectives and Misconceptions
Transformers have demonstrated remarkable performance in natural language processing and related domains, as they largely focus on sequential, autoregressive next-token prediction tasks. Yet, they struggle in logical reasoning, not necessarily because of a fundamental limitation of these models, but possibly due to the lack of exploration of more creative uses, such as latent space and recurrent reasoning. An emerging exploration in this direction is the Hierarchical Reasoning Model (Wang et. al., 2025), which introduces a novel type of recurrent reasoning in the latent space of transformers, achieving remarkable performance on a wide range of 2D reasoning tasks. Despite the promising results, this line of models is still at an early stage and calls for in-depth investigation. In this work, we review this class of models, examine key design choices, test alternative variants and clarify common misconceptions.
comment: Found errors in some results of v1. Removed them and changed conclusions
♻ ☆ A Fairness-Aware Strategy for B5G Physical-layer Security Leveraging Reconfigurable Intelligent Surfaces
Reconfigurable Intelligent Surfaces are composed of physical elements that can dynamically alter electromagnetic wave properties to enhance beamforming and lead to improvements in areas with low coverage properties. Combined with Reinforcement Learning techniques, they have the potential to be conduct as well physical-layer security hardening. Yet, and in addition to security improvements, it is crucial to consider the concept of fair communication. Reconfigurable Intelligent Surfaces must ensure that User Equipment units receive their signals with adequate strength, without other units being deprived of service due to insufficient power. In this paper, we address such a problem. We explore the fairness properties of previous work and propose a novel method that aims at obtaining both an efficient and fair duplex Reconfigurable Intelligent Surface-Reinforcement Learning system for multiple legitimate User Equipment units without reducing the level of achieved physical-layer security hardening. In terms of contributions, we uncover a fairness imbalance of a previous physical-layer security hardening solution, validate our findings and report experimental work via simulation results. We also provide an alternative reward strategy to solve the uncovered problems and release both code and datasets to foster further research in the topics of this paper.
comment: 19 pages, 5 figures, 2 tables, 41 references
♻ ☆ How Reliable are Causal Probing Interventions?
Causal probing aims to analyze foundation models by examining how intervening on their representation of various latent properties impacts their outputs. Recent works have cast doubt on the theoretical basis of several leading causal probing methods, but it has been unclear how to systematically evaluate the effectiveness of these methods in practice. To address this, we define two key causal probing desiderata: completeness (how thoroughly the representation of the target property has been transformed) and selectivity (how little non-targeted properties have been impacted). We find that there is an inherent tradeoff between the two, which we define as reliability, their harmonic mean. We introduce an empirical analysis framework to measure and evaluate these quantities, allowing us to make the first direct comparisons between different families of leading causal probing methods (e.g., linear vs. nonlinear, or concept removal vs. counterfactual interventions). We find that: (1) all methods show a clear tradeoff between completeness and selectivity; (2) more complete and reliable methods have a greater impact on LLM behavior; and (3) nonlinear interventions are almost always more reliable than linear interventions.
♻ ☆ Gemstones: A Model Suite for Multi-Faceted Scaling Laws NeurIPS 2025
Scaling laws are typically fit using a family of models with a narrow range of frozen hyperparameter choices. In this work we study scaling laws using multiple architectural shapes and hyperparameter choices, highlighting their impact on resulting prescriptions. As a primary artifact of our research, we release the Gemstones: an open-source scaling law dataset, consisting of over 4000 checkpoints from transformers with up to 2 billion parameters and diverse architectural shapes; including ablations over learning rate and cooldown. Our checkpoints enable more complex studies of scaling, such as analyzing the relationship between width and depth. By examining our model suite, we find that the prescriptions of scaling laws can be highly sensitive to the experimental design process and the specific model checkpoints used during fitting.
comment: NeurIPS 2025
♻ ☆ Optimal Policy Minimum Bayesian Risk
Inference scaling helps LLMs solve complex reasoning problems through extended runtime computation. On top of long chain-of-thought (long-CoT) models, purely inference-time techniques such as best-of-N (BoN) sampling, majority voting, or more generally, minimum Bayes risk decoding (MBRD), can further improve LLM accuracy by generating multiple candidate solutions and aggregating over them. These methods typically leverage additional signals in the form of reward models and risk/similarity functions that compare generated samples, e.g., exact match in some normalized space or standard similarity metrics such as Rouge. Here we present a novel method for incorporating reward and risk/similarity signals into MBRD. Based on the concept of optimal policy in KL-controlled reinforcement learning, our framework provides a simple and well-defined mechanism for leveraging such signals, offering several advantages over traditional inference-time methods: higher robustness, improved accuracy, and well-understood asymptotic behavior. In addition, it allows for the development of a sample-efficient variant of MBRD that can adjust the number of samples to generate according to the difficulty of the problem, without relying on majority vote counts. We empirically demonstrate the advantages of our approach on math (MATH-$500$) and coding (HumanEval) tasks using recent open-source models. We also present a comprehensive analysis of its accuracy-compute trade-offs.
♻ ☆ Spatiotemporal Graph Learning with Direct Volumetric Information Passing and Feature Enhancement
Data-driven learning of physical systems has kindled significant attention, where many neural models have been developed. In particular, mesh-based graph neural networks (GNNs) have demonstrated significant potential in modeling spatiotemporal dynamics across arbitrary geometric domains. However, the existing node-edge message-passing and aggregation mechanism in GNNs limits the representation learning ability. In this paper, we proposed a dual-module framework, Cell-embedded and Feature-enhanced Graph Neural Network (aka, CeFeGNN), for learning spatiotemporal dynamics. Specifically, we embed learnable cell attributions to the common node-edge message passing process, which better captures the spatial dependency of regional features. Such a strategy essentially upgrades the local aggregation scheme from first order (e.g., from edge to node) to a higher order (e.g., from volume and edge to node), which takes advantage of volumetric information in message passing. Meanwhile, a novel feature-enhanced block is designed to further improve the model's performance and alleviate the over-smoothness problem. Extensive experiments on various PDE systems and one real-world dataset demonstrate that CeFeGNN achieves superior performance compared with other baselines.
♻ ☆ Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models
Reasoning language models improve performance on complex tasks by generating long chains of thought (CoTs), but this process can also increase harmful outputs in adversarial settings. In this work, we ask whether the long CoTs can be leveraged for predictive safety monitoring: do the reasoning traces provide early signals of final response alignment that could enable timely intervention? We evaluate a range of monitoring methods using either CoT text or activations, including highly capable large language models, fine-tuned classifiers, and humans. First, we find that a simple linear probe trained on CoT activations significantly outperforms all text-based baselines in predicting whether a final response is safe or unsafe, with an average absolute increase of 13 in F1 scores over the best-performing alternatives. CoT texts are often unfaithful and misleading, while model latents provide a more reliable predictive signal. Second, the probe can be applied to early CoT segments before the response is generated, showing that alignment signals appear before reasoning completes. Error analysis reveals that the performance gap between text classifiers and the linear probe largely stems from a subset of responses we call performative CoTs, where the reasoning consistently contradicts the final response as the CoT progresses. Our findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.
♻ ☆ Robust-Multi-Task Gradient Boosting
Multi-task learning (MTL) has shown effectiveness in exploiting shared information across tasks to improve generalization. MTL assumes tasks share similarities that can improve performance. In addition, boosting algorithms have demonstrated exceptional performance across diverse learning problems, primarily due to their ability to focus on hard-to-learn instances and iteratively reduce residual errors. This makes them a promising approach for learning multi-task problems. However, real-world MTL scenarios often involve tasks that are not well-aligned (known as outlier or adversarial tasks), which do not share beneficial similarities with others and can, in fact, deteriorate the performance of the overall model. To overcome this challenge, we propose Robust-Multi-Task Gradient Boosting (R-MTGB), a novel boosting framework that explicitly models and adapts to task heterogeneity during training. R-MTGB structures the learning process into three sequential blocks: (1) learning shared patterns, (2) partitioning tasks into outliers and non-outliers with regularized parameters, and (3) fine-tuning task-specific predictors. This architecture enables R-MTGB to automatically detect and penalize outlier tasks while promoting effective knowledge transfer among related tasks. Our method integrates these mechanisms seamlessly within gradient boosting, allowing robust handling of noisy or adversarial tasks without sacrificing accuracy. Extensive experiments on both synthetic benchmarks and real-world datasets demonstrate that our approach successfully isolates outliers, transfers knowledge, and consistently reduces prediction errors for each task individually, and achieves overall performance gains across all tasks. These results highlight robustness, adaptability, and reliable convergence of R-MTGB in challenging MTL environments.
♻ ☆ Epistemic Diversity and Knowledge Collapse in Large Language Models
Large language models (LLMs) tend to generate lexically, semantically, and stylistically homogenous texts. This poses a risk of knowledge collapse, where homogenous LLMs mediate a shrinking in the range of accessible information over time. Existing works on homogenization are limited by a focus on closed-ended multiple-choice setups or fuzzy semantic features, and do not look at trends across time and cultural contexts. To overcome this, we present a new methodology to measure epistemic diversity, i.e., variation in real-world claims in LLM outputs, which we use to perform a broad empirical study of LLM knowledge collapse. We test 27 LLMs, 155 topics covering 12 countries, and 200 prompt variations sourced from real user chats. For the topics in our study, we show that while newer models tend to generate more diverse claims, nearly all models are less epistemically diverse than a basic web search. We find that model size has a negative impact on epistemic diversity, while retrieval-augmented generation (RAG) has a positive impact, though the improvement from RAG varies by the cultural context. Finally, compared to a traditional knowledge source (Wikipedia), we find that country-specific claims reflect the English language more than the local one, highlighting a gap in epistemic representation
comment: 16 pages; 8 figures, 4 tables v2 changelog: Fixed the modeling for table 3, random effect is the model version
♻ ☆ CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysis
Large Language Models (LLMs) with Mixture-of-Experts (MoE) architectures are distinguished by their strong performance scaling with increasing parameters across a wide range of tasks, yet they also suffer from substantial computational and storage overheads. Notably, the performance gains of MoE models do not scale proportionally with the growth in expert parameters. While prior works attempt to reduce parameters via expert-level pruning, merging, or decomposition, they still suffer from challenges in both performance and computational efficiency. In this paper, we address these challenges by introducing micro-expert as a finer-grained compression unit that spans across matrices. We first establish a more fundamental perspective, viewing MoE layers as mixtures of micro-experts, and present CAMERA, a lightweight and training-free framework for identifying micro-expert redundancy. Our analysis uncovers significant variance in micro-expert contributions during decoding. Based on this insight, we further propose CAMERA-P, a structured micro-expert pruning framework, and CAMERA-Q, a mixed-precision quantization idea designed for micro-experts. Extensive experiments on nine downstream tasks show that CAMERA-P consistently outperforms strong baselines under pruning ratios ranging from 20% to 60%. Furthermore, CAMERA-Q achieves superior results under aggressive 2-bit quantization, surpassing existing matrix- and channel-level ideas. Notably, our method enables complete micro-expert analysis of Qwen2-57B-A14B in less than 5 minutes on a single NVIDIA A100-40GB GPU.
comment: 16 pages, 9 figures, 7 tables
♻ ☆ Unified Cross-Modal Medical Image Synthesis with Hierarchical Mixture of Product-of-Experts
We propose a deep mixture of multimodal hierarchical variational auto-encoders called MMHVAE that synthesizes missing images from observed images in different modalities. MMHVAE's design focuses on tackling four challenges: (i) creating a complex latent representation of multimodal data to generate high-resolution images; (ii) encouraging the variational distributions to estimate the missing information needed for cross-modal image synthesis; (iii) learning to fuse multimodal information in the context of missing data; (iv) leveraging dataset-level information to handle incomplete data sets at training time. Extensive experiments are performed on the challenging problem of pre-operative brain multi-parametric magnetic resonance and intra-operative ultrasound imaging.
comment: Accepted in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
♻ ☆ Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction
In this paper, we unify more than 10 existing one-step diffusion distillation approaches, such as Diff-Instruct, DMD, SIM, SiD, $f$-distill, etc, inside a theory-driven framework which we name the \textbf{\emph{Uni-Instruct}}. Uni-Instruct is motivated by our proposed diffusion expansion theory of the $f$-divergence family. Then we introduce key theories that overcome the intractability issue of the original expanded $f$-divergence, resulting in an equivalent yet tractable loss that effectively trains one-step diffusion models by minimizing the expanded $f$-divergence family. The novel unification introduced by Uni-Instruct not only offers new theoretical contributions that help understand existing approaches from a high-level perspective but also leads to state-of-the-art one-step diffusion generation performances. On the CIFAR10 generation benchmark, Uni-Instruct achieves record-breaking Frechet Inception Distance (FID) values of \textbf{\emph{1.46}} for unconditional generation and \textbf{\emph{1.38}} for conditional generation. On the ImageNet-$64\times 64$ generation benchmark, Uni-Instruct achieves a new SoTA one-step generation FID of \textbf{\emph{1.02}}, which outperforms its 79-step teacher diffusion with a significant improvement margin of 1.33 (1.02 vs 2.35). We also apply Uni-Instruct on broader tasks like text-to-3D generation. For text-to-3D generation, Uni-Instruct gives decent results, which slightly outperforms previous methods, such as SDS and VSD, in terms of both generation quality and diversity. Both the solid theoretical and empirical contributions of Uni-Instruct will potentially help future studies on one-step diffusion distillation and knowledge transferring of diffusion models.
♻ ☆ Fundamental Limits of Membership Inference Attacks on Machine Learning Models
Membership inference attacks (MIA) can reveal whether a particular data point was part of the training dataset, potentially exposing sensitive information about individuals. This article provides theoretical guarantees by exploring the fundamental statistical limitations associated with MIAs on machine learning models at large. More precisely, we first derive the statistical quantity that governs the effectiveness and success of such attacks. We then theoretically prove that in a non-linear regression setting with overfitting learning procedures, attacks may have a high probability of success. Finally, we investigate several situations for which we provide bounds on this quantity of interest. Interestingly, our findings indicate that discretizing the data might enhance the learning procedure's security. Specifically, it is demonstrated to be limited by a constant, which quantifies the diversity of the underlying data distribution. We illustrate those results through simple simulations.
comment: Accepted for publication in JMLR
♻ ☆ Teaching Metric Distance to Discrete Autoregressive Language Models
As large language models expand beyond natural language to domains such as mathematics, multimodal understanding, and embodied agents, tokens increasingly reflect metric relationships rather than purely linguistic meaning. We introduce DIST2Loss, a distance-aware framework designed to train autoregressive discrete models by leveraging predefined distance relationships among output tokens. At its core, DIST2Loss transforms continuous exponential family distributions derived from inherent distance metrics into discrete, categorical optimization targets compatible with the models' architectures. This approach enables the models to learn and preserve meaningful distance relationships during token generation while maintaining compatibility with existing architectures. Empirical evaluations show consistent performance gains in diverse multimodal applications, including visual grounding, robotic manipulation, generative reward modeling, and image generation using vector-quantized features. These improvements are most notable in low-data regimes, demonstrating DIST2Loss's strength under resource constraints.
♻ ☆ BrowserArena: Evaluating LLM Agents on Real-World Web Navigation Tasks
LLM web agents now browse and take actions on the open web, yet current agent evaluations are constrained to sandboxed environments or artificial tasks. We introduce BrowserArena, a live open-web agent evaluation platform that collects user-submitted tasks, runs Arena-style head-to-head comparisons, and uses step-level human feedback to surface failure modes. Collecting and analyzing step-level annotations on the agent traces, we identify three consistent failure modes: captcha resolution, pop-up banner removal, and direct navigation to URLs. By constructing targeted datasets to further study these tasks, we discover variations in how different language models navigate these failure modes. We find, for example, that o4-mini deploys a wider variety of strategies to circumvent captcha resolution than other models and DeepSeek-R1 consistently misleads users about pop-up banner closure. Our findings surface both the diversity and brittleness of current web agents. More broadly, our benchmarking methodology provides an approach to evaluating and understanding web agent failure modes at scale.
♻ ☆ ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression NeurIPS 2025
The hypothesis that Convolutional Neural Networks (CNNs) are inherently texture-biased has shaped much of the discourse on feature use in deep learning. We revisit this hypothesis by examining limitations in the cue-conflict experiment by Geirhos et al. To address these limitations, we propose a domain-agnostic framework that quantifies feature reliance through systematic suppression of shape, texture, and color cues, avoiding the confounds of forced-choice conflicts. By evaluating humans and neural networks under controlled suppression conditions, we find that CNNs are not inherently texture-biased but predominantly rely on local shape features. Nonetheless, this reliance can be substantially mitigated through modern training strategies or architectures (ConvNeXt, ViTs). We further extend the analysis across computer vision, medical imaging, and remote sensing, revealing that reliance patterns differ systematically: computer vision models prioritize shape, medical imaging models emphasize color, and remote sensing models exhibit a stronger reliance on texture. Code is available at https://github.com/tomburgert/feature-reliance.
comment: Accepted at NeurIPS 2025 (oral)
♻ ☆ Structured Sparse Transition Matrices to Enable State Tracking in State-Space Models NeurIPS 2025
Modern state-space models (SSMs) often utilize transition matrices which enable efficient computation but pose restrictions on the model's expressivity, as measured in terms of the ability to emulate finite-state automata (FSA). While unstructured transition matrices are optimal in terms of expressivity, they come at a prohibitively high compute and memory cost even for moderate state sizes. We propose a structured sparse parametrization of transition matrices in SSMs that enables FSA state tracking with optimal state size and depth, while keeping the computational cost of the recurrence comparable to that of diagonal SSMs. Our method, PD-SSM, parametrizes the transition matrix as the product of a column one-hot matrix ($P$) and a complex-valued diagonal matrix ($D$). Consequently, the computational cost of parallel scans scales linearly with the state size. Theoretically, the model is BIBO-stable and can emulate any $N$-state FSA with one layer of dimension $N$ and a linear readout of size $N \times N$, significantly improving on all current structured SSM guarantees. Experimentally, the model significantly outperforms a wide collection of modern SSM variants on various FSA state tracking tasks. On multiclass time-series classification, the performance is comparable to that of neural controlled differential equations, a paradigm explicitly built for time-series analysis. Finally, we integrate PD-SSM into a hybrid Transformer-SSM architecture and demonstrate that the model can effectively track the states of a complex FSA in which transitions are encoded as a set of variable-length English sentences. The code is available at https://github.com/IBM/expressive-sparse-state-space-model
comment: 10 pages, NeurIPS 2025 Spotlight
♻ ☆ SAE-FiRE: Enhancing Earnings Surprise Predictions Through Sparse Autoencoder Feature Selection
Predicting earnings surprises from financial documents, such as earnings conference calls, regulatory filings, and financial news, has become increasingly important in financial economics. However, these financial documents present significant analytical challenges, typically containing over 5,000 words with substantial redundancy and industry-specific terminology that creates obstacles for language models. In this work, we propose the SAE-FiRE (Sparse Autoencoder for Financial Representation Enhancement) framework to address these limitations by extracting key information while eliminating redundancy. SAE-FiRE employs Sparse Autoencoders (SAEs) to decompose dense neural representations from large language models into interpretable sparse components, then applies statistical feature selection methods, including ANOVA F-tests and tree-based importance scoring, to identify the top-k most discriminative dimensions for classification. By systematically filtering out noise that might otherwise lead to overfitting, we enable more robust and generalizable predictions. Experimental results across three financial datasets demonstrate that SAE-FiRE significantly outperforms baseline approaches.
♻ ☆ Investigating Forecasting Models for Pandemic Infections Using Heterogeneous Data Sources: A 2-year Study with COVID-19
Emerging in December 2019, the COVID-19 pandemic caused widespread health, economic, and social disruptions. Rapid global transmission overwhelmed healthcare systems, resulting in high infection rates, hospitalisations, and fatalities. To minimise the spread, governments implemented several non-pharmaceutical interventions like lockdowns and travel restrictions. While effective in controlling transmission, these measures also posed significant economic and societal challenges. Although the WHO declared COVID-19 no longer a global health emergency in May 2023, its impact persists, shaping public health strategies. The vast amount of data collected during the pandemic offers valuable insights into disease dynamics, transmission, and intervention effectiveness. Leveraging these insights can improve forecasting models, enhancing preparedness and response to future outbreaks while mitigating their social and economic impact. This paper presents a large-scale case study on COVID-19 forecasting in Cyprus, utilising a two-year dataset that integrates epidemiological data, vaccination records, policy measures, and weather conditions. We analyse infection trends, assess forecasting performance, and examine the influence of external factors on disease dynamics. The insights gained contribute to improved pandemic preparedness and response strategies.
comment: Keywords: epidemiology, pandemic forecasting, COVID-19, infections, machine learning. Published in: IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB) 2025
♻ ☆ Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks EMNLP 2025
In this paper, we study the surprising impact that truncating text embeddings has on downstream performance. We consistently observe across 6 state-of-the-art text encoders and 26 downstream tasks, that randomly removing up to 50% of embedding dimensions results in only a minor drop in performance, less than 10%, in retrieval and classification tasks. Given the benefits of using smaller-sized embeddings, as well as the potential insights about text encoding, we study this phenomenon and find that, contrary to what is suggested in prior work, this is not the result of an ineffective use of representation space. Instead, we find that a large number of uniformly distributed dimensions actually cause an increase in performance when removed. This would explain why, on average, removing a large number of embedding dimensions results in a marginal drop in performance. We make similar observations when truncating the embeddings used by large language models to make next-token predictions on generative tasks, suggesting that this phenomenon is not isolated to classification or retrieval tasks.
comment: Accepted to EMNLP 2025 Main Conference (Oral), camera-ready version
♻ ☆ FedFlex: Federated Learning for Diverse Netflix Recommendations
The drive for personalization in recommender systems creates a tension between user privacy and the risk of "filter bubbles". Although federated learning offers a promising paradigm for privacy-preserving recommendations, its impact on diversity remains unclear. We introduce FedFlex, a two-stage framework that combines local, on-device fine-tuning of matrix factorization models (SVD and BPR) with a lightweight Maximal Marginal Relevance (MMR) re-ranking step to promote diversity. We conducted the first live user study of a federated recommender, collecting behavioral data and feedback during a two-week online deployment. Our results show that FedFlex successfully engages users, with BPR outperforming SVD in click-through rate. Re-ranking with MMR consistently improved ranking quality (nDCG) across both models, with statistically significant gains, particularly for BPR. Diversity effects varied: MMR increased coverage for both models and improved intra-list diversity for BPR, but slightly reduced it for SVD, suggesting different interactions between personalization and diversification across models. Our exit questionnaire responses indicated that most users expressed no clear preference between re-ranked and unprocessed lists, implying that increased diversity did not substantially reduce user satisfaction.
♻ ☆ Neon: Negative Extrapolation From Self-Training Improves Image Generation
Scaling generative AI models is bottlenecked by the scarcity of high-quality training data. The ease of synthesizing from a generative model suggests using (unverified) synthetic data to augment a limited corpus of real data for the purpose of fine-tuning in the hope of improving performance. Unfortunately, however, the resulting positive feedback loop leads to model autophagy disorder (MAD, aka model collapse) that results in a rapid degradation in sample quality and/or diversity. In this paper, we introduce Neon (for Negative Extrapolation frOm self-traiNing), a new learning method that turns the degradation from self-training into a powerful signal for self-improvement. Given a base model, Neon first fine-tunes it on its own self-synthesized data but then, counterintuitively, reverses its gradient updates to extrapolate away from the degraded weights. We prove that Neon works because typical inference samplers that favor high-probability regions create a predictable anti-alignment between the synthetic and real data population gradients, which negative extrapolation corrects to better align the model with the true data distribution. Neon is remarkably easy to implement via a simple post-hoc merge that requires no new real data, works effectively with as few as 1k synthetic samples, and typically uses less than 1% additional training compute. We demonstrate Neon's universality across a range of architectures (diffusion, flow matching, autoregressive, and inductive moment matching models) and datasets (ImageNet, CIFAR-10, and FFHQ). In particular, on ImageNet 256x256, Neon elevates the xAR-L model to a new state-of-the-art FID of 1.02 with only 0.36% additional training compute. Code is available at https://github.com/VITA-Group/Neon
♻ ☆ A weakly-supervised deep learning model for fast localisation and delineation of the skeleton, internal organs, and spinal canal on Whole-Body Diffusion-Weighted MRI (WB-DWI)
Background: Apparent Diffusion Coefficient (ADC) values and Total Diffusion Volume (TDV) from Whole-body diffusion-weighted MRI (WB-DWI) are recognized cancer imaging biomarkers. However, manual disease delineation for ADC and TDV measurements is unfeasible in clinical practice, demanding automation. As a first step, we propose an algorithm to generate fast and reproducible probability maps of the skeleton, adjacent internal organs (liver, spleen, urinary bladder, and kidneys), and spinal canal. Methods: We developed an automated deep-learning pipeline based on a 3D patch-based Residual U-Net architecture that localises and delineates these anatomical structures on WB-DWI. The algorithm was trained using "soft labels" (non-binary segmentations) derived from a computationally intensive atlas-based approach. For training and validation, we employed a multi-centre WB-DWI dataset comprising 532 scans from patients with Advanced Prostate Cancer (APC) or Multiple Myeloma (MM), with testing on 45 patients. Results: Our weakly-supervised deep learning model achieved an average dice score of 0.67 for whole skeletal delineation, 0.76 when excluding ribcage, 0.83 for internal organs, and 0.86 for spinal canal, with average surface distances below 3mm. Relative median ADC differences between automated and manual full-body delineations were below 10%. The model was 12x faster than the atlas-based registration algorithm (25 sec vs. 5 min). Two experienced radiologists rated the model's outputs as either "good" or "excellent" on test scans, with inter-reader agreement from fair to substantial (Gwet's AC1 = 0.27-0.72). Conclusion: The model offers fast, reproducible probability maps for localising and delineating body regions on WB-DWI, potentially enabling non-invasive imaging biomarker quantification to support disease staging and treatment response assessment.
♻ ☆ CAPO: Towards Enhancing LLM Reasoning through Generative Credit Assignment
Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning abilities of Large Language Models (LLMs) by using rule-based binary feedback. However, current RLVR methods typically assign the same reward to every token. This coarse-grained feedback hampers precise credit assignment, making it hard for models to identify which reasoning steps lead to success or failure, and often results in suboptimal policies. Methods like PPO provide credit assignment by value estimation, but yield inaccurate and unverifiable signals due to limited sampling. On the other hand, methods using Process Reward Models can provide step-wise rewards but suffer from several key limitations: they require high-quality process supervision labels, the feedback is unreliable due to probabilistic reward modeling, and their application in online reinforcement learning (RL) is time-consuming. To overcome these limitations, we introduce a simple but efficient method-Credit Assignment Policy Optimization (CAPO). Instead of training auxiliary models, CAPO directly leverages an off-the-shelf, general-purpose LLM as a Generative Process Reward Model (LLM-as-GenPRM) to generate all step-wise critique by one pass only based on the correctness of the step itself, providing deterministic token-level credits to refine the tokens that were originally assigned identical rule-based rewards. To further enhance the accuracy and robustness, we employ voting mechanisms that scale with the number of generated critiques. Extensive experiments on various backbones like Llama and Qwen models show that CAPO consistently outperforms supervised learning-based and RL-based fine-tuning methods across four challenging mathematical benchmarks and three out-of-domain benchmarks. Further analysis shows that CAPO can help the model to foster the learning of correct reasoning pathways leading to correct answers.
comment: Work in progress
♻ ☆ MetaLLMix : An XAI Aided LLM-Meta-learning Based Approach for Hyper-parameters Optimization
Effective model and hyperparameter selection remains a major challenge in deep learning, often requiring extensive expertise and computation. While AutoML and large language models (LLMs) promise automation, current LLM-based approaches rely on trial and error and expensive APIs, which provide limited interpretability and generalizability. We propose MetaLLMiX, a zero-shot hyperparameter optimization framework combining meta-learning, explainable AI, and efficient LLM reasoning. By leveraging historical experiment outcomes with SHAP explanations, MetaLLMiX recommends optimal hyperparameters and pretrained models without additional trials. We further employ an LLM-as-judge evaluation to control output format, accuracy, and completeness. Experiments on eight medical imaging datasets using nine open-source lightweight LLMs show that MetaLLMiX achieves competitive or superior performance to traditional HPO methods while drastically reducing computational cost. Our local deployment outperforms prior API-based approaches, achieving optimal results on 5 of 8 tasks, response time reductions of 99.6-99.9%, and the fastest training times on 6 datasets (2.4-15.7x faster), maintaining accuracy within 1-5% of best-performing baselines.
♻ ☆ How Foundational are Foundation Models for Time Series Forecasting? NeurIPS 2025
Foundation Models are designed to serve as versatile embedding machines, with strong zero shot capabilities and superior generalization performance when fine-tuned on diverse downstream tasks. While this is largely true for language and vision foundation models, we argue that the inherent diversity of time series data makes them less suited for building effective foundation models. We demonstrate this using forecasting as our downstream task. We show that the zero-shot capabilities of a time series foundation model are significantly influenced and tied to the specific domains it has been pretrained on. Furthermore, when applied to unseen real-world time series data, fine-tuned foundation models do not consistently yield substantially better results, relative to their increased parameter count and memory footprint, than smaller, dedicated models tailored to the specific forecasting task at hand.
comment: Typo rectified in this v3 version. Accepted at NeurIPS 2025 Workshop on Recent Advances in Time Series Foundation Models (BERT2S)
♻ ☆ Understanding Catastrophic Interference: On the Identifibility of Latent Representations
Catastrophic interference, also known as catastrophic forgetting, is a fundamental challenge in machine learning, where a trained learning model progressively loses performance on previously learned tasks when adapting to new ones. In this paper, we aim to better understand and model the catastrophic interference problem from a latent representation learning point of view, and propose a novel theoretical framework that formulates catastrophic interference as an identification problem. Our analysis demonstrates that the forgetting phenomenon can be quantified by the distance between partial-task aware (PTA) and all-task aware (ATA) setups. Building upon recent advances in identifiability theory, we prove that this distance can be minimized through identification of shared latent variables between these setups. When learning, we propose our method \ourmeos with two-stage training strategy: First, we employ maximum likelihood estimation to learn the latent representations from both PTA and ATA configurations. Subsequently, we optimize the KL divergence to identify and learn the shared latent variables. Through theoretical guarantee and empirical validations, we establish that identifying and learning these shared representations can effectively mitigate catastrophic interference in machine learning systems. Our approach provides both theoretical guarantees and practical performance improvements across both synthetic and benchmark datasets.
♻ ☆ Mitigating Exponential Mixed Frequency Growth through Frequency Selection
Quantum machine learning research has expanded rapidly due to potential computational advantages over classical methods. Angle encoding has emerged as a popular choice as feature map (FM) for embedding classical data into quantum models due to its simplicity and natural generation of truncated Fourier series, providing universal function approximation capabilities. Efficient FMs within quantum circuits can exploit exponential scaling of Fourier frequencies, with multi-dimensional inputs introducing additional exponential growth through mixed-frequency terms. Despite this promising expressive capability, practical implementation faces significant challenges. Through controlled experiments with white-box target functions, we demonstrate that training failures can occur even when all relevant frequencies are theoretically accessible. We illustrate how two primary known causes lead to unsuccessful optimization: insufficient trainable parameters relative to the model's frequency content, and limitations imposed by the ansatz's dynamic lie algebra dimension, but also uncover an additional parameter burden: the necessity of controlling non-unique frequencies within the model. To address this, we propose near-zero weight initialization to suppress unnecessary duplicate frequencies. For target functions with a priori frequency knowledge, we introduce frequency selection as a practical solution that reduces parameter requirements and mitigates the exponential growth that would otherwise render problems intractable due to parameter insufficiency. Our frequency selection approach achieved near-optimal performance (median $R^2 \approx 0.95$) with 78\% of the parameters needed by the best standard approach in 10 randomly chosen target functions.
comment: 10 pages, 3 figures
♻ ☆ Sparse Representations Improve Adversarial Robustness of Neural Network Classifiers
Deep neural networks perform remarkably well on image classification tasks but remain vulnerable to carefully crafted adversarial perturbations. This work revisits linear dimensionality reduction as a simple, data-adapted defense. We empirically compare standard Principal Component Analysis (PCA) with its sparse variant (SPCA) as front-end feature extractors for downstream classifiers, and we complement these experiments with a theoretical analysis. On the theory side, we derive exact robustness certificates for linear heads applied to SPCA features: for both $\ell_\infty$ and $\ell_2$ threat models (binary and multiclass), the certified radius grows as the dual norms of $W^\top u$ shrink, where $W$ is the projection and $u$ the head weights. We further show that for general (non-linear) heads, sparsity reduces operator-norm bounds through a Lipschitz composition argument, predicting lower input sensitivity. Empirically, with a small non-linear network after the projection, SPCA consistently degrades more gracefully than PCA under strong white-box and black-box attacks while maintaining competitive clean accuracy. Taken together, the theory identifies the mechanism (sparser projections reduce adversarial leverage) and the experiments verify that this benefit persists beyond the linear setting. Our code is available at https://github.com/killian31/SPCARobustness.
comment: Killian Steunou is the main contributor and corresponding author of this work
♻ ☆ Back to Square Roots: An Optimal Bound on the Matrix Factorization Error for Multi-Epoch Differentially Private SGD
Matrix factorization mechanisms for differentially private training have emerged as a promising approach to improve model utility under privacy constraints. In practical settings, models are typically trained over multiple epochs, requiring matrix factorizations that account for repeated participation. Existing theoretical upper and lower bounds on multi-epoch factorization error leave a significant gap. In this work, we introduce a new explicit factorization method, Banded Inverse Square Root (BISR), which imposes a banded structure on the inverse correlation matrix. This factorization enables us to derive an explicit and tight characterization of the multi-epoch error. We further prove that BISR achieves asymptotically optimal error by matching the upper and lower bounds. Empirically, BISR performs on par with state-of-the-art factorization methods, while being simpler to implement, computationally efficient, and easier to analyze.
♻ ☆ Fast Policy Learning for Linear Quadratic Control with Entropy Regularization
This paper proposes and analyzes two new policy learning methods: regularized policy gradient (RPG) and iterative policy optimization (IPO), for a class of discounted linear-quadratic control (LQC) problems over an infinite time horizon with entropy regularization. Assuming access to the exact policy evaluation, both proposed approaches are proven to converge linearly in finding optimal policies of the regularized LQC. Moreover, the IPO method can achieve a super-linear convergence rate once it enters a local region around the optimal policy. Finally, when the optimal policy for an RL problem with a known environment is appropriately transferred as the initial policy to an RL problem with an unknown environment, the IPO method is shown to enable a super-linear convergence rate if the two environments are sufficiently close. Performances of these proposed algorithms are supported by numerical examples.
comment: 31 pages, 3 figures
♻ ☆ The Use of Binary Choice Forests to Model and Estimate Discrete Choices
Problem definition. In retailing, discrete choice models (DCMs) are commonly used to capture the choice behavior of customers when offered an assortment of products. When estimating DCMs using transaction data, flexible models (such as machine learning models or nonparametric models) are typically not interpretable and hard to estimate, while tractable models (such as the multinomial logit model) tend to misspecify the complex behavior represeted in the data. Methodology/results. In this study, we use a forest of binary decision trees to represent DCMs. This approach is based on random forests, a popular machine learning algorithm. The resulting model is interpretable: the decision trees can explain the decision-making process of customers during the purchase. We show that our approach can predict the choice probability of any DCM consistently and thus never suffers from misspecification. Moreover, our algorithm predicts assortments unseen in the training data. The mechanism and errors can be theoretically analyzed. We also prove that the random forest can recover preference rankings of customers thanks to the splitting criterion such as the Gini index and information gain ratio. Managerial implications. The framework has unique practical advantages. It can capture customers' behavioral patterns such as irrationality or sequential searches when purchasing a product. It handles nonstandard formats of training data that result from aggregation. It can measure product importance based on how frequently a random customer would make decisions depending on the presence of the product. It can also incorporate price information and customer features. Our numerical experiments using synthetic and real data show that using random forests to estimate customer choices can outperform existing methods.
comment: 63 pages, 10 figures, 30 tables
♻ ☆ Minimizing the Weighted Number of Tardy Jobs: Data-Driven Heuristic for Single-Machine Scheduling
Existing research on single-machine scheduling is largely focused on exact algorithms, which perform well on typical instances but can significantly deteriorate on certain regions of the problem space. In contrast, data-driven approaches provide strong and scalable performance when tailored to the structure of specific datasets. Leveraging this idea, we focus on a single-machine scheduling problem where each job is defined by its weight, duration, due date, and deadline, aiming to minimize the total weight of tardy jobs. We introduce a novel data-driven scheduling heuristic that combines machine learning with problem-specific characteristics, ensuring feasible solutions, which is a common challenge for ML-based algorithms. Experimental results demonstrate that our approach significantly outperforms the state-of-the-art in terms of optimality gap, number of optimal solutions, and adaptability across varied data scenarios, highlighting its flexibility for practical applications. In addition, we conduct a systematic exploration of ML models, addressing a common gap in similar studies by offering a detailed model selection process and providing insights into why the chosen model is the best fit.
comment: Published version: Computers & Operations Research, https://doi.org/10.1016/j.cor.2025.107281. Data are publicly available at https://doi.org/10.5281/zenodo.17233362
♻ ☆ An Embarrassingly Simple Defense Against LLM Abliteration Attacks
Large language models (LLMs) are typically aligned to refuse harmful instructions through safety fine-tuning. A recent attack, termed abliteration, identifies and suppresses the single latent direction most responsible for refusal behavior, thereby enabling models to generate harmful content. We propose a defense that fundamentally alters how models express refusal. We construct an extended-refusal dataset in which responses to harmful prompts provide detailed justifications before refusing, distributing the refusal signal across multiple token positions. Fine-tuning Llama-2-7B-Chat and Qwen2.5-Instruct (1.5B and 3B parameters) on this dataset yields models that maintain high refusal rates under abliteration: refusal rates drop by at most 10%, compared to 70-80% drops in baseline models. Comprehensive evaluations of safety and utility demonstrate that extended-refusal fine-tuning effectively neutralizes abliteration attacks while preserving general model performance and enhancing robustness across multiple alignment scenarios.
comment: preprint - under review
♻ ☆ Can foundation models actively gather information in interactive environments to test hypotheses?
Foundation models excel at single-turn reasoning but struggle with multi-turn exploration in dynamic environments, a requirement for many real-world challenges. We evaluated these models on their ability to learn from experience, adapt, and gather information. First, in "Feature World," a simple setting for testing information gathering, models performed near-optimally. However, to test more complex, multi-trial learning, we implemented a text-based version of the "Alchemy" environment, a benchmark for meta-learning. Here, agents must deduce a latent causal structure by integrating information across many trials. In this setting, recent foundation models initially failed to improve their performance over time. Crucially, we found that prompting the models to summarize their observations at regular intervals enabled an emergent meta-learning process. This allowed them to improve across trials and even adaptively re-learn when the environment's rules changed unexpectedly. While most models handled the simple task, Alchemy revealed stark differences in robustness: Gemini 2.5 performed best, followed by Claude 3.7, while ChatGPT-4o and o4-mini struggled. This underscores Alchemy's value as a benchmark. Our findings demonstrate that the biggest challenge for foundation models is not selecting informative actions in the moment, but integrating knowledge through adaptive strategies over time. Encouragingly, there appears to be no intrinsic barrier to future models mastering these abilities.
♻ ☆ Expected Free Energy-based Planning as Variational Inference
We address the problem of planning under uncertainty, where an agent must choose actions that not only achieve desired outcomes but also reduce uncertainty. Traditional methods often treat exploration and exploitation as separate objectives, lacking a unified inferential foundation. Active inference, grounded in the Free Energy Principle, provides such a foundation by minimizing Expected Free Energy (EFE), a cost function that combines utility with epistemic drives, such as ambiguity resolution and novelty seeking. However, the computational burden of EFE minimization had remained a significant obstacle to its scalability. In this paper, we show that EFE-based planning arises naturally from minimizing a variational free energy functional on a generative model augmented with preference and epistemic priors. This result reinforces theoretical consistency with the Free Energy Principle by casting planning under uncertainty itself as a form of variational inference. Our formulation yields policies that jointly support goal achievement and information gain, while incorporating a complexity term that accounts for bounded computational resources. This unifying framework connects and extends existing methods, enabling scalable, resource-aware implementations of active inference agents.
comment: 18 pages
♻ ☆ Constrained free energy minimization for the design of thermal states and stabilizer thermodynamic systems
A quantum thermodynamic system is described by a Hamiltonian and a list of conserved, non-commuting charges, and a fundamental goal is to determine the minimum energy of the system subject to constraints on the charges. Recently, [Liu et al., arXiv:2505.04514] proposed first- and second-order classical and hybrid quantum-classical algorithms for solving a dual chemical potential maximization problem, and they proved that these algorithms converge to global optima by means of gradient-ascent approaches. In this paper, we benchmark these algorithms on several problems of interest in thermodynamics, including one- and two-dimensional quantum Heisenberg models with nearest and next-to-nearest neighbor interactions and with the charges set to the total x, y, and z magnetizations. We also offer an alternative compelling interpretation of these algorithms as methods for designing ground and thermal states of controllable Hamiltonians, with potential applications in molecular and material design. Furthermore, we introduce stabilizer thermodynamic systems as thermodynamic systems based on stabilizer codes, with the Hamiltonian constructed from a given code's stabilizer operators and the charges constructed from the code's logical operators. We benchmark the aforementioned algorithms on several examples of stabilizer thermodynamic systems, including those constructed from the one-to-three-qubit repetition code, the perfect one-to-five-qubit code, and the two-to-four-qubit error-detecting code. Finally, we observe that the aforementioned hybrid quantum-classical algorithms, when applied to stabilizer thermodynamic systems, can serve as alternative methods for encoding qubits into stabilizer codes at a fixed temperature, and we provide an effective method for warm-starting these encoding algorithms whenever a single qubit is encoded into multiple physical qubits.
comment: v2: 35 pages, 12 figures, updated simulations
♻ ☆ Interpretable Clustering: A Survey
In recent years, much of the research on clustering algorithms has primarily focused on enhancing their accuracy and efficiency, frequently at the expense of interpretability. However, as these methods are increasingly being applied in high-stakes domains such as healthcare, finance, and autonomous systems, the need for transparent and interpretable clustering outcomes has become a critical concern. This is not only necessary for gaining user trust but also for satisfying the growing ethical and regulatory demands in these fields. Ensuring that decisions derived from clustering algorithms can be clearly understood and justified is now a fundamental requirement. To address this need, this paper provides a comprehensive and structured review of the current state of explainable clustering algorithms, identifying key criteria to distinguish between various methods. These insights can effectively assist researchers in making informed decisions about the most suitable explainable clustering methods for specific application contexts, while also promoting the development and adoption of clustering algorithms that are both efficient and transparent. For convenient access and reference, an open repository organizes representative and emerging interpretable clustering methods under the taxonomy proposed in this survey, available at https://github.com/hulianyu/Awesome-Interpretable-Clustering
comment: 14 pages, 2 figures, 3 tables
♻ ☆ AuToMATo: An Out-Of-The-Box Persistence-Based Clustering Algorithm
We present AuToMATo, a novel clustering algorithm based on persistent homology. While AuToMATo is not parameter-free per se, we provide default choices for its parameters that make it into an out-of-the-box clustering algorithm that performs well across the board. AuToMATo combines the existing ToMATo clustering algorithm with a bootstrapping procedure in order to separate significant peaks of an estimated density function from non-significant ones. We perform a thorough comparison of AuToMATo (with its parameters fixed to their defaults) against many other state-of-the-art clustering algorithms. We find not only that AuToMATo compares favorably against parameter-free clustering algorithms, but in many instances also significantly outperforms even the best selection of parameters for other algorithms. AuToMATo is motivated by applications in topological data analysis, in particular the Mapper algorithm, where it is desirable to work with a clustering algorithm that does not need tuning of its parameters. Indeed, we provide evidence that AuToMATo performs well when used with Mapper. Finally, we provide an open-source implementation of AuToMATo in Python that is fully compatible with the standard scikit-learn architecture.
comment: Code: https://doi.org/10.5281/zenodo.17279740
♻ ☆ Shortcuts Everywhere and Nowhere: Exploring Multi-Trigger Backdoor Attacks
Backdoor attacks have become a significant threat to the pre-training and deployment of deep neural networks (DNNs). Although numerous methods for detecting and mitigating backdoor attacks have been proposed, most rely on identifying and eliminating the ``shortcut" created by the backdoor, which links a specific source class to a target class. However, these approaches can be easily circumvented by designing multiple backdoor triggers that create shortcuts everywhere and therefore nowhere specific. In this study, we explore the concept of Multi-Trigger Backdoor Attacks (MTBAs), where multiple adversaries leverage different types of triggers to poison the same dataset. By proposing and investigating three types of multi-trigger attacks including \textit{parallel}, \textit{sequential}, and \textit{hybrid} attacks, we demonstrate that 1) multiple triggers can coexist, overwrite, or cross-activate one another, and 2) MTBAs easily break the prevalent shortcut assumption underlying most existing backdoor detection/removal methods, rendering them ineffective. Given the security risk posed by MTBAs, we have created a multi-trigger backdoor poisoning dataset to facilitate future research on detecting and mitigating these attacks, and we also discuss potential defense strategies against MTBAs. Our code is available at https://github.com/bboylyg/Multi-Trigger-Backdoor-Attacks.
comment: 13 pages
♻ ☆ Learning to Price Bundles: A GCN Approach for Mixed Bundling
Bundle pricing refers to designing several product combinations (i.e., bundles) and determining their prices in order to maximize the expected profit. It is a classic problem in revenue management and arises in many industries, such as e-commerce, tourism, and video games. However, the problem is typically intractable due to the exponential number of candidate bundles. In this paper, we explore the usage of graph convolutional networks (GCNs) in solving the bundle pricing problem. Specifically, we first develop a graph representation of the mixed bundling model (where every possible bundle is assigned with a specific price) and then train a GCN to learn the latent patterns of optimal bundles. Based on the trained GCN, we propose two inference strategies to derive high-quality feasible solutions. A local-search technique is further proposed to improve the solution quality. Numerical experiments validate the effectiveness and efficiency of our proposed GCN-based framework. Using a GCN trained on instances with 5 products, our methods consistently achieve near-optimal solutions (better than 97%) with only a fraction of computational time for problems of small to medium size. It also achieves superior solutions for larger size of problems compared with other heuristic methods such as bundle size pricing (BSP). The method can also provide high quality solutions for instances with more than 30 products even for the challenging cases where product utilities are non-additive.
♻ ☆ EntryPrune: Neural Network Feature Selection using First Impressions
There is an ongoing effort to develop feature selection algorithms to improve interpretability, reduce computational resources, and minimize overfitting in predictive models. Neural networks stand out as architectures on which to build feature selection methods, and recently, neuron pruning and regrowth have emerged from the sparse neural network literature as promising new tools. We introduce EntryPrune, a novel supervised feature selection algorithm using a dense neural network with a dynamic sparse input layer. It employs entry-based pruning, a novel approach that compares neurons based on their relative change induced when they have entered the network. Extensive experiments on 13 different datasets show that our approach generally outperforms the current state-of-the-art methods, and in particular improves the average accuracy on low-dimensional datasets. Furthermore, we show that EntryPruning surpasses traditional techniques such as magnitude pruning within the EntryPrune framework and that EntryPrune achieves lower runtime than competing approaches. Our code is available at https://github.com/flxzimmer/entryprune.
♻ ☆ SKADA-Bench: Benchmarking Unsupervised Domain Adaptation Methods with Realistic Validation On Diverse Modalities
Unsupervised Domain Adaptation (DA) consists of adapting a model trained on a labeled source domain to perform well on an unlabeled target domain with some data distribution shift. While many methods have been proposed in the literature, fair and realistic evaluation remains an open question, particularly due to methodological difficulties in selecting hyperparameters in the unsupervised setting. With SKADA-bench, we propose a framework to evaluate DA methods on diverse modalities, beyond computer vision task that have been largely explored in the literature. We present a complete and fair evaluation of existing shallow algorithms, including reweighting, mapping, and subspace alignment. Realistic hyperparameter selection is performed with nested cross-validation and various unsupervised model selection scores, on both simulated datasets with controlled shifts and real-world datasets across diverse modalities, such as images, text, biomedical, and tabular data. Our benchmark highlights the importance of realistic validation and provides practical guidance for real-life applications, with key insights into the choice and impact of model selection approaches. SKADA-bench is open-source, reproducible, and can be easily extended with novel DA methods, datasets, and model selection criteria without requiring re-evaluating competitors. SKADA-bench is available on Github at https://github.com/scikit-adaptation/skada-bench.
comment: Published in Transactions on Machine Learning Research
♻ ☆ TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems NeurIPS 2025
Regression models are crucial in recommender systems. However, retransformation bias problem has been conspicuously neglected within the community. While many works in other fields have devised effective bias correction methods, all of them are post-hoc cures externally to the model, facing practical challenges when applied to real-world recommender systems. Hence, we propose a preemptive paradigm to eradicate the bias intrinsically from the models via minor model refinement. Specifically, a novel TranSUN method is proposed with a joint bias learning manner to offer theoretically guaranteed unbiasedness under empirical superior convergence. It is further generalized into a novel generic regression model family, termed Generalized TranSUN (GTS), which not only offers more theoretical insights but also serves as a generic framework for flexibly developing various bias-free models. Comprehensive experimental results demonstrate the superiority of our methods across data from various domains, which have been successfully deployed in two real-world industrial recommendation scenarios, i.e. product and short video recommendation scenarios in Guess What You Like business domain in the homepage of Taobao App (a leading e-commerce platform with DAU > 300M), to serve the major online traffic.
comment: 37 pages, 6 figures, NeurIPS 2025 Poster
♻ ☆ Mutatis Mutandis: Revisiting the Comparator in Discrimination Testing
Testing for individual discrimination involves deriving a profile, the comparator, similar to the one making the discrimination claim, the complainant, based on a protected attribute, such as race or gender, and comparing their decision outcomes. The complainant-comparator pair is central to discrimination testing. Most discrimination testing tools rely on this pair to establish evidence for discrimination. In this work we revisit the role of the comparator in discrimination testing. We first argue for the inherent causal modeling nature of deriving the comparator. We then introduce a two-kinds classification for the comparator: the ceteris paribus, or``with all else equal,'' (CP) comparator and the mutatis mutandis, or ``with the appropriate adjustments being made,'' (MM) comparator. The CP comparator is the standard comparator, representing an idealized comparison for establishing discrimination as it aims for a complainant-comparator pair that only differs on membership to the protected attribute. As an alternative to it, we define the MM comparator, which requires that the comparator represents the``what would have been of'' the complainant without the effects of the protected attribute on the non-protected attributes. Under the MM comparator, the complainant-comparator pair can be dissimilar in terms of the non-protected attributes, departing from an idealized comparison. Notably, the MM comparator is a more complex kind of comparator and its implementation offers an impactful venue for machine learning methods. We illustrate these two comparators and their impact on discrimination testing using a real-world example.
♻ ☆ Context Biasing for Pronunciations-Orthography Mismatch in Automatic Speech Recognition
Neural sequence-to-sequence systems deliver state-of-the-art performance for automatic speech recognition. When using appropriate modeling units, e.g., byte-pair encoded characters, these systems are in principal open vocabulary systems. In practice, however, they often fail to recognize words not seen during training, e.g., named entities, acronyms, or domain-specific special words. To address this problem, many context biasing methods have been proposed; however, for words with a pronunciation-orthography mismatch, these methods may still struggle. We propose a method which allows corrections of substitution errors to improve the recognition accuracy of such challenging words. Users can add corrections on the fly during inference. We show that with this method we get a relative improvement in biased word error rate of up to 8%, while maintaining a competitive overall word error rate.
♻ ☆ Benchmarking the Robustness of Agentic Systems to Adversarially-Induced Harms
Ensuring the safe use of agentic systems requires a thorough understanding of the range of malicious behaviors these systems may exhibit when under attack. In this paper, we evaluate the robustness of LLM-based agentic systems against attacks that aim to elicit harmful actions from agents. To this end, we propose a novel taxonomy of harms for agentic systems and a novel benchmark, BAD-ACTS, for studying the security of agentic systems with respect to a wide range of harmful actions. BAD-ACTS consists of 4 implementations of agentic systems in distinct application environments, as well as a dataset of 188 high-quality examples of harmful actions. This enables a comprehensive study of the robustness of agentic systems across a wide range of categories of harmful behaviors, available tools, and inter-agent communication structures. Using this benchmark, we analyze the robustness of agentic systems against an attacker that controls one of the agents in the system and aims to manipulate other agents to execute a harmful target action. Our results show that the attack has a high success rate, demonstrating that even a single adversarial agent within the system can have a significant impact on the security. This attack remains effective even when agents use a simple prompting-based defense strategy. However, we additionally propose a more effective defense based on message monitoring. We believe that this benchmark provides a diverse testbed for the security research of agentic systems. The benchmark can be found at github.com/JNoether/BAD-ACTS
comment: 54 Pages
♻ ☆ SDFs from Unoriented Point Clouds using Neural Variational Heat Distances
We propose a novel variational approach for computing neural Signed Distance Fields (SDF) from unoriented point clouds. To this end, we replace the commonly used eikonal equation with the heat method, carrying over to the neural domain what has long been standard practice for computing distances on discrete surfaces. This yields two convex optimization problems for whose solution we employ neural networks: We first compute a neural approximation of the gradients of the unsigned distance field through a small time step of heat flow with weighted point cloud densities as initial data. Then we use it to compute a neural approximation of the SDF. We prove that the underlying variational problems are well-posed. Through numerical experiments, we demonstrate that our method provides state-of-the-art surface reconstruction and consistent SDF gradients. Furthermore, we show in a proof-of-concept that it is accurate enough for solving a PDE on the zero-level set.
comment: 15 pages, 17 figures, 4 tables
♻ ☆ Deep Reinforcement Learning for Urban Air Quality Management: Multi-Objective Optimization of Pollution Mitigation Booth Placement in Metropolitan Environments
This is the preprint version of the article published in IEEE Access vol. 13, pp. 146503--146526, 2025, doi:10.1109/ACCESS.2025.3599541. Please cite the published version. Urban air pollution remains a pressing global concern, particularly in densely populated and traffic-intensive metropolitan areas like Delhi, where exposure to harmful pollutants severely impacts public health. Delhi, being one of the most polluted cities globally, experiences chronic air quality issues due to vehicular emissions, industrial activities, and construction dust, which exacerbate its already fragile atmospheric conditions. Traditional pollution mitigation strategies, such as static air purifying installations, often fail to maximize their impact due to suboptimal placement and limited adaptability to dynamic urban environments. This study presents a novel deep reinforcement learning (DRL) framework to optimize the placement of air purification booths to improve the air quality index (AQI) in the city of Delhi. We employ Proximal Policy Optimization (PPO), a state-of-the-art reinforcement learning algorithm, to iteratively learn and identify high-impact locations based on multiple spatial and environmental factors, including population density, traffic patterns, industrial influence, and green space constraints. Our approach is benchmarked against conventional placement strategies, including random and greedy AQI-based methods, using multi-dimensional performance evaluation metrics such as AQI improvement, spatial coverage, population and traffic impact, and spatial entropy.
comment: This is the preprint version of the article published in IEEE Access vol. 13, pp. 146503--146526, 2025, doi:10.1109/ACCESS.2025.3599541. Please cite the published version
♻ ☆ From Accuracy to Robustness: A Study of Rule- and Model-based Verifiers in Mathematical Reasoning
Trustworthy verifiers are essential for the success of reinforcement learning with verifiable reward (RLVR), which is the core methodology behind various large reasoning models such as DeepSeek-R1. In complex domains like mathematical reasoning, rule-based verifiers have been widely adopted in previous works to train strong reasoning models. However, the reliability of these verifiers and their impact on the RL training process remain poorly understood. In this work, we take mathematical reasoning as a case study and conduct a comprehensive analysis of various verifiers in both static evaluation and RL training scenarios. First, we find that current open-source rule-based verifiers often fail to recognize equivalent answers presented in different formats across multiple commonly used mathematical datasets, resulting in non-negligible false negative rates. This limitation adversely affects RL training performance and becomes more pronounced as the policy model gets stronger. Subsequently, we investigate model-based verifiers as a potential solution to address these limitations. While the static evaluation shows that model-based verifiers achieve significantly higher verification accuracy, further analysis and RL results imply that they are highly susceptible to hacking, where they misclassify certain patterns in responses as correct, particularly after fine-tuning. This vulnerability is exploited during policy model optimization, leading to artificially inflated rewards. Our findings underscore the unique challenges inherent to both rule-based and model-based verifiers and provide insights toward developing more accurate and robust reward systems for reinforcement learning.
♻ ☆ Conformal Prediction in Hierarchical Classification with Constrained Representation Complexity
Conformal prediction has emerged as a widely used framework for constructing valid prediction sets in classification and regression tasks. In this work, we extend the split conformal prediction framework to hierarchical classification, where prediction sets are commonly restricted to internal nodes of a predefined hierarchy, and propose two computationally efficient inference algorithms. The first algorithm returns internal nodes as prediction sets, while the second one relaxes this restriction. Using the notion of representation complexity, the latter yields smaller set sizes at the cost of a more general and combinatorial inference problem. Empirical evaluations on several benchmark datasets demonstrate the effectiveness of the proposed algorithms in achieving nominal coverage.
♻ ☆ A Deterministic Information Bottleneck Method for Clustering Mixed-Type Data
In this paper, we present an information-theoretic method for clustering mixed-type data, that is, data consisting of both continuous and categorical variables. The proposed approach extends the Information Bottleneck principle to heterogeneous data through generalised product kernels, integrating continuous, nominal, and ordinal variables within a unified optimization framework. We address the following challenges: developing a systematic bandwidth selection strategy that equalises contributions across variable types, and proposing an adaptive hyperparameter updating scheme that ensures a valid solution into a predetermined number of potentially imbalanced clusters. Through simulations on 28,800 synthetic data sets and ten publicly available benchmarks, we demonstrate that the proposed method, named DIBmix, achieves superior performance compared to four established methods (KAMILA, K-Prototypes, FAMD with K-Means, and PAM with Gower's dissimilarity). Results show DIBmix particularly excels when clusters exhibit size imbalances, data contain low or moderate cluster overlap, and categorical and continuous variables are equally represented. The method presents a significant advantage over traditional centroid-based algorithms, establishing DIBmix as a competitive and theoretically grounded alternative for mixed-type data clustering.
comment: 33 pages
♻ ☆ Detecting Invariant Manifolds in ReLU-Based RNNs
Recurrent Neural Networks (RNNs) have found widespread applications in machine learning for time series prediction and dynamical systems reconstruction, and experienced a recent renaissance with improved training algorithms and architectural designs. Understanding why and how trained RNNs produce their behavior is important for scientific and medical applications, and explainable AI more generally. An RNN's dynamical repertoire depends on the topological and geometrical properties of its state space. Stable and unstable manifolds of periodic points play a particularly important role: They dissect a dynamical system's state space into different basins of attraction, and their intersections lead to chaotic dynamics with fractal geometry. Here we introduce a novel algorithm for detecting these manifolds, with a focus on piecewise-linear RNNs (PLRNNs) employing rectified linear units (ReLUs) as their activation function. We demonstrate how the algorithm can be used to trace the boundaries between different basins of attraction, and hence to characterize multistability, a computationally important property. We further show its utility in finding so-called homoclinic points, the intersections between stable and unstable manifolds, and thus establish the existence of chaos in PLRNNs. Finally we show for an empirical example, electrophysiological recordings from a cortical neuron, how insights into the underlying dynamics could be gained through our method.
♻ ☆ DeepBoost-AF: A Novel Unsupervised Feature Learning and Gradient Boosting Fusion for Robust Atrial Fibrillation Detection in Raw ECG Signals
Atrial fibrillation (AF) is a prevalent cardiac arrhythmia associated with elevated health risks, where timely detection is pivotal for mitigating stroke-related morbidity. This study introduces an innovative hybrid methodology integrating unsupervised deep learning and gradient boosting models to improve AF detection. A 19-layer deep convolutional autoencoder (DCAE) is coupled with three boosting classifiers-AdaBoost, XGBoost, and LightGBM (LGBM)-to harness their complementary advantages while addressing individual limitations. The proposed framework uniquely combines DCAE with gradient boosting, enabling end-to-end AF identification devoid of manual feature extraction. The DCAE-LGBM model attains an F1-score of 95.20%, sensitivity of 99.99%, and inference latency of four seconds, outperforming existing methods and aligning with clinical deployment requirements. The DCAE integration significantly enhances boosting models, positioning this hybrid system as a reliable tool for automated AF detection in clinical settings.
comment: 12-page,4 figures,3 tables, Achieves 95.20% F1-score (99.99% sensitivity) on 8,528 PhysioNet 2017 recordings, Mean inference time: 4 seconds, Python implementation will be open-sourced upon publication
♻ ☆ Object Centric Concept Bottlenecks
Developing high-performing, yet interpretable models remains a critical challenge in modern AI. Concept-based models (CBMs) attempt to address this by extracting human-understandable concepts from a global encoding (e.g., image encoding) and then applying a linear classifier on the resulting concept activations, enabling transparent decision-making. However, their reliance on holistic image encodings limits their expressiveness in object-centric real-world settings and thus hinders their ability to solve complex vision tasks beyond single-label classification. To tackle these challenges, we introduce Object-Centric Concept Bottlenecks (OCB), a framework that combines the strengths of CBMs and pre-trained object-centric foundation models, boosting performance and interpretability. We evaluate OCB on complex image datasets and conduct a comprehensive ablation study to analyze key components of the framework, such as strategies for aggregating object-concept encodings. The results show that OCB outperforms traditional CBMs and allows one to make interpretable decisions for complex visual tasks.
♻ ☆ Conditional Local Independence Testing for Itô processes with Applications to Dynamic Causal Discovery
Inferring causal relationships from dynamical systems is the central interest of many scientific inquiries. Conditional local independence, which describes whether the evolution of one process is influenced by another process given additional processes, is important for causal learning in such systems. In this paper, we propose a hypothesis test for conditional local independence in It\^o processes. Our test is grounded in the semimartingale decomposition of the It\^o process, with which we introduce a stochastic integral process that is a martingale under the null hypothesis. We then apply a test for the martingale property, quantifying potential deviation from local independence. The test statistics is estimated using the optimal filtering equation. We show the consistency of the estimation, thereby establishing the level and power of our test. Numerical verification and a real-world application to causal discovery in brain resting-state fMRIs are conducted.
comment: Preprint
♻ ☆ Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning NeurIPS 2025
Current long-tailed semi-supervised learning methods assume that labeled data exhibit a long-tailed distribution, and unlabeled data adhere to a typical predefined distribution (i.e., long-tailed, uniform, or inverse long-tailed). However, the distribution of the unlabeled data is generally unknown and may follow an arbitrary distribution. To tackle this challenge, we propose a Controllable Pseudo-label Generation (CPG) framework, expanding the labeled dataset with the progressively identified reliable pseudo-labels from the unlabeled dataset and training the model on the updated labeled dataset with a known distribution, making it unaffected by the unlabeled data distribution. Specifically, CPG operates through a controllable self-reinforcing optimization cycle: (i) at each training step, our dynamic controllable filtering mechanism selectively incorporates reliable pseudo-labels from the unlabeled dataset into the labeled dataset, ensuring that the updated labeled dataset follows a known distribution; (ii) we then construct a Bayes-optimal classifier using logit adjustment based on the updated labeled data distribution; (iii) this improved classifier subsequently helps identify more reliable pseudo-labels in the next training step. We further theoretically prove that this optimization cycle can significantly reduce the generalization error under some conditions. Additionally, we propose a class-aware adaptive augmentation module to further improve the representation of minority classes, and an auxiliary branch to maximize data utilization by leveraging all labeled and unlabeled samples. Comprehensive evaluations on various commonly used benchmark datasets show that CPG achieves consistent improvements, surpassing state-of-the-art methods by up to $\textbf{15.97\%}$ in accuracy. The code is available at https://github.com/yaxinhou/CPG.
comment: The paper is accepted by NeurIPS 2025
Multimedia 7
☆ Controllable Audio-Visual Viewpoint Generation from 360° Spatial Information
The generation of sounding videos has seen significant advancements with the advent of diffusion models. However, existing methods often lack the fine-grained control needed to generate viewpoint-specific content from larger, immersive 360-degree environments. This limitation restricts the creation of audio-visual experiences that are aware of off-camera events. To the best of our knowledge, this is the first work to introduce a framework for controllable audio-visual generation, addressing this unexplored gap. Specifically, we propose a diffusion model by introducing a set of powerful conditioning signals derived from the full 360-degree space: a panoramic saliency map to identify regions of interest, a bounding-box-aware signed distance map to define the target viewpoint, and a descriptive caption of the entire scene. By integrating these controls, our model generates spatially-aware viewpoint videos and audios that are coherently influenced by the broader, unseen environmental context, introducing a strong controllability that is essential for realistic and immersive audio-visual generation. We show audiovisual examples proving the effectiveness of our framework.
☆ Segment-Factorized Full-Song Generation on Symbolic Piano Music NeurIPS 2025
We propose the Segmented Full-Song Model (SFS) for symbolic full-song generation. The model accepts a user-provided song structure and an optional short seed segment that anchors the main idea around which the song is developed. By factorizing a song into segments and generating each one through selective attention to related segments, the model achieves higher quality and efficiency compared to prior work. To demonstrate its suitability for human-AI interaction, we further wrap SFS into a web application that enables users to iteratively co-create music on a piano roll with customizable structures and flexible ordering.
comment: Accepted to the 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: AI for Music
☆ Towards Robust and Realible Multimodal Fake News Detection with Incomplete Modality
Multimodal fake news detection (MFND) has become an urgent task with the emergence of huge multimodal fake content on social media platforms. Previous studies mainly focus on complex feature extraction and fusion to learn discriminative information from multimodal content. However, in real-world applications, multimedia news may naturally lose some information during dissemination, resulting in modality incompleteness, which is detrimental to the generalization and robustness of existing models. To this end, we propose a novel generic and robust multimodal fusion strategy, termed Multi-expert Modality-incomplete Learning Network (MMLNet), which is simple yet effective. It consists of three key steps: (1) Multi-Expert Collaborative Reasoning to compensate for missing modalities by dynamically leveraging complementary information through multiple experts. (2) Incomplete Modality Adapters compensates for the missing information by leveraging the new feature distribution. (3) Modality Missing Learning leveraging an label-aware adaptive weighting strategy to learn a robust representation with contrastive learning. We evaluate MMLNet on three real-world benchmarks across two languages, demonstrating superior performance compared to state-of-the-art methods while maintaining relative simplicity. By ensuring the accuracy of fake news detection in incomplete modality scenarios caused by information propagation, MMLNet effectively curbs the spread of malicious misinformation. Code is publicly available at https://github.com/zhyhome/MMLNet.
☆ FoleyGRAM: Video-to-Audio Generation with GRAM-Aligned Multimodal Encoders IJCNN 2025
In this work, we present FoleyGRAM, a novel approach to video-to-audio generation that emphasizes semantic conditioning through the use of aligned multimodal encoders. Building on prior advancements in video-to-audio generation, FoleyGRAM leverages the Gramian Representation Alignment Measure (GRAM) to align embeddings across video, text, and audio modalities, enabling precise semantic control over the audio generation process. The core of FoleyGRAM is a diffusion-based audio synthesis model conditioned on GRAM-aligned embeddings and waveform envelopes, ensuring both semantic richness and temporal alignment with the corresponding input video. We evaluate FoleyGRAM on the Greatest Hits dataset, a standard benchmark for video-to-audio models. Our experiments demonstrate that aligning multimodal encoders using GRAM enhances the system's ability to semantically align generated audio with video content, advancing the state of the art in video-to-audio synthesis.
comment: Acepted at IJCNN 2025
☆ StereoSync: Spatially-Aware Stereo Audio Generation from Video IJCNN 2025
Although audio generation has been widely studied over recent years, video-aligned audio generation still remains a relatively unexplored frontier. To address this gap, we introduce StereoSync, a novel and efficient model designed to generate audio that is both temporally synchronized with a reference video and spatially aligned with its visual context. Moreover, StereoSync also achieves efficiency by leveraging pretrained foundation models, reducing the need for extensive training while maintaining high-quality synthesis. Unlike existing methods that primarily focus on temporal synchronization, StereoSync introduces a significant advancement by incorporating spatial awareness into video-aligned audio generation. Indeed, given an input video, our approach extracts spatial cues from depth maps and bounding boxes, using them as cross-attention conditioning in a diffusion-based audio generation model. Such an approach allows StereoSync to go beyond simple synchronization, producing stereo audio that dynamically adapts to the spatial structure and movement of a video scene. We evaluate StereoSync on Walking The Maps, a curated dataset comprising videos from video games that feature animated characters walking through diverse environments. Experimental results demonstrate the ability of StereoSync to achieve both temporal and spatial alignment, advancing the state of the art in video-to-audio generation and resulting in a significantly more immersive and realistic audio experience.
comment: Accepted at IJCNN 2025
☆ When and How to Cut Classical Concerts? A Multimodal Automated Video Editing Approach
Automated video editing remains an underexplored task in the computer vision and multimedia domains, especially when contrasted with the growing interest in video generation and scene understanding. In this work, we address the specific challenge of editing multicamera recordings of classical music concerts by decomposing the problem into two key sub-tasks: when to cut and how to cut. Building on recent literature, we propose a novel multimodal architecture for the temporal segmentation task (when to cut), which integrates log-mel spectrograms from the audio signals, plus an optional image embedding, and scalar temporal features through a lightweight convolutional-transformer pipeline. For the spatial selection task (how to cut), we improve the literature by updating from old backbones, e.g. ResNet, with a CLIP-based encoder and constraining distractor selection to segments from the same concert. Our dataset was constructed following a pseudo-labeling approach, in which raw video data was automatically clustered into coherent shot segments. We show that our models outperformed previous baselines in detecting cut points and provide competitive visual shot selection, advancing the state of the art in multimodal automated video editing.
♻ ☆ GRAFT: GRaPH and Table Reasoning for Textual Alignment -- A Benchmark for Structured Instruction Following and Visual Reasoning
GRAFT is a structured multimodal benchmark for evaluating models on instruction-following, visual reasoning, and visual-textual alignment tasks. It features programmatically generated charts and synthetically rendered tables, created with Python visualization libraries to ensure control over data semantics, structure, and clarity. Each GRAFT instance pairs a chart or table image with a systematically generated, multi-step analytical question based solely on visual content. Answers are provided in structured formats such as JSON or YAML, supporting consistent evaluation of both reasoning and output format. The benchmark introduces a taxonomy of reasoning types including comparison, trend identification, ranking, aggregation, proportion estimation, and anomaly detection to enable comprehensive assessment. Reference answers follow strict factual and formatting guidelines for precise, aspect-based evaluation. GRAFT offers a unified, scalable framework for fine-grained benchmarking of multimodal models on visually grounded, structured reasoning tasks, setting a new evaluation standard in this field.
comment: 25 pages, 10 tables, 3 figures