While language models have demonstrated remarkable capabilities, their performance remains disproportionately skewed towards high-resource languages. This disparity is particularly acute for the 700 million inhabitants of Southeast Asia, where vast linguistic diversity remains under-represented in existing benchmarks, often suffering from limited language coverage or narrow task scopes. To address this, we introduce DeepSEA, a comprehensive benchmark constructed from established datasets through professional human translation into 12 Southeast Asian languages, spanning both high-resource (e.g., Indonesian, Thai, Vietnamese) and low-resource (e.g., Burmese, Khmer, Lao) languages across five distinct tasks: mathematical reasoning, scientific reasoning, question answering, abstractive summarisation, and machine translation. Extensive evaluation of frontier proprietary and openweights models reveals that while frontier models like Gemini-3.7-Flash maintain robustness across resource tiers, smaller models face severe performance degradation on low-resource languages. Crucially, models perform better in reasoning and summarisation when pivoted through English rather than target languages, exposing a critical dependency on English representations, especially for low-resource languages. Beyond these observations, we provide a cross-lingual error analysis and a cross-script tokeniser fertility analysis. We hope these findings shed light on the development of next generation inclusive multilingual models.

We present PopResume, a population-representative resume dataset for causal fairness auditing of LLM- and VLM-based resume screening systems. Unlike existing benchmarks that rely on manually injected demographic information and outcome-level disparities, PopResume is grounded in population statistics and preserves natural attribute relationships, enabling path-specific effect (PSE)-based fairness evaluation. We decompose the effect of a protected attribute on resume scores into two paths: the business necessity path, mediated by job-relevant qualifications, and the redlining path, mediated by demographic proxies. This distinction allows auditors to separate legally permissible from impermissible sources of disparity. Evaluating four LLMs and four VLMs on PopResume’s 60.8K resumes across five occupations, we identify five representative discrimination patterns that aggregate metrics fail to capture. Our results demonstrate that PSE-based evaluation reveals fairness issues masked by outcome-level measures, underscoring the need for causally-grounded auditing frameworks in AI-assisted hiring.

Text-labeled graph databases combine symbolic graph structures with rich textual descriptions, enabling expressive representations for real-world data. However, answering queries over such databases remains challenging when queries require both precise logical constraints and flexible semantic text interpretation, especially under information incompleteness. We introduce the task of hybrid query answering (hybrid QA) over incomplete text-labeled graph databases, which requires joint reasoning over both symbolic and textual knowledge under incompleteness. To investigate this problem, we formalize a query language that integrates logical constraints with free-form textual descriptions and construct two new benchmark datasets from biological and e-commerce domains. We propose HybridQE, a hybrid query embedding framework that aligns with our query language via embedding and combining the symbolic and textual constraints in a unified embedding space. Experiments on the proposed benchmarks show that existing methods struggle with this setting, while HybridQE provides consistent improvements across diverse hybrid query types, highlighting both the difficulty of HybridQA and the promise of hybrid query embeddings for reasoning under incompleteness.

Hate speech detection is inherently subjective: people from different demographic groups perceive the same content very differently. Collecting enough annotations from multiple demographic groups is costly and difficult to scale. Persona-conditioned Large Language Models (models prompted to adopt a specific demographic identity) have been proposed as a way to simulate diverse perspectives at scale. But do they actually reflect how different groups disagree? We evaluate three aspects of human social judgement: (i) whether personas from different groups disagree in human-like ways (inter-group disagreement), (ii) whether they become more sensitive when content targets their own identity (in-group sensitivity), and (iii) whether they can accurately predict how another group would react (vicarious prediction). Our results show that no model consistently captures all three dimensions, and performance is highly model-dependent and does not emerge reliably from minimal identity prompts alone. However, vicarious prompting with Llama 3.1 yields the highest cross-group agreement in most demographic axes and provides the closest overall approximation to human disagreement patterns, indicating that this configuration may provide a more reliable setting for automatic annotation aligned with human judgements.

News consumption on social media has become ubiquitous, yet how different forms of engagement shape psychosocial outcomes remains unclear. To address this gap, we leveraged a large-scale dataset of ~26M posts and ~45M comments on the Bluesky platform and conducted a quasi-experimental study, matching 81,345 Treated users exposed to News feeds with 83,711 Control users using stratified propensity score analysis. We examined psychosocial wellbeing in terms of affective, behavioral, and cognitive outcomes. Our findings reveal that news engagement produces systematic trade-offs: increased depression, stress, and anxiety, yet decreased loneliness and increased social interaction on the platform. Regression models reveal that News feed bookmarking is associated with greater psychosocial deterioration compared to commenting or quoting, with magnitude differences exceeding tenfold. These per-engagement effects accumulate with repeated exposure, showing significant psychosocial impacts. Our work extends theories of news effects beyond crisis-centric frameworks by demonstrating that routine consumption creates distinct psychological dynamics depending on engagement type, and bears implications for tools and interventions for mitigating the psychosocial costs of news consumption on social media.

Flowchart-oriented dialogue (FOD) systems aim to guide users through multi-turn decision-making or operational procedures by following a domain-specific flowchart to achieve a task goal. In this work, we formalize flowchart reasoning in FOD as grounding user input to flowchart nodes at each dialogue turn while ensuring node transition is consistent with the correct flowchart path. Despite recent advances of LLMs in task-oriented dialogue systems, adapting them to FOD still faces two limitations: (1) LLMs lack an explicit mechanism to represent and reason over flowchart topology, and (2) they are prone to hallucinations, leading to unfaithful flowchart reasoning. To address these limitations, we propose FloCA, a zero-shot flowchart-oriented conversational agent. FloCA uses an LLM for intent understanding and response generation, while delegating flowchart reasoning to an external tool that performs topology-constrained graph execution, ensuring faithful and logically consistent node transitions across dialogue turns. We further introduce an evaluation framework with an LLM-based user simulator and five new metrics covering reasoning accuracy and interaction efficiency. Extensive experiments on FLODIAL and PFDial datasets highlight the bottlenecks of existing LLM/VLM-based methods and demonstrate the superiority of FloCA.

The public often attributes human-like qualities to large language models (LLMs), assuming that they "know" certain things. In reality, LLMs encode information retained during training as internal probabilistic knowledge. This study examines existing methods for probing the veracity of that knowledge and identifies three flawed underlying assumptions. To address these flaws, we introduce ‘sAwMIL‘ (Sparse-Aware Multiple-Instance Learning), a multiclass probing framework that combines multiple-instance learning with conformal prediction. ‘sAwMIL‘ leverages LLMs’ internal representations to classify statements as _true_, _false_, or _neither_. We evaluate ‘sAwMIL‘ across 16 open-source LLMs, including default and chat-based variants, using three new curated datasets. Our results show that (1) common probing methods fail to provide a reliable and transferable veracity direction and, in some settings, perform worse than zero-shot prompting; (2) truth and falsehood are not encoded symmetrically; and (3) LLMs encode a third type of signal that is distinct from both true and false.

Large language models have shown impressive capabilities in code generation, yet they often produce functionally incorrect code. Uncertainty quantification (UQ) methods have emerged as a promising approach for detecting hallucinations in natural language generation, but their effectiveness for code generation tasks remains underexplored. We systematically evaluate how UQ techniques transfer to code generation across three programming languages, five LLMs, and over 1,700 problems. We find that some token-probability-based methods generalize effectively without modification, while sampling-based methods relying on natural language inference (NLI) fail because NLI models cannot distinguish functionally different code, causing most responses to collapse into a single semantic cluster. To address this, we introduce functional equivalence methods, a family of code-specific methods that replace NLI-based semantic equivalence with an LLM-based functional equivalence assessment, including functional entropy, a code-specific analog of semantic entropy. Functional equivalence methods achieve top AUROC in 11 out of 15 model-benchmark combinations and the best calibration across most settings, consistently outperforming both NLI-based counterparts and all other methods evaluated.

Large language models (LLMs) frequently exhibit performance biases against regional dialects of low-resource languages. However, frameworks to quantify these disparities remain scarce. We propose a two-phase framework to evaluate dialectal bias, operationalized as comprehension degradation relative to standard Bengali, in LLM question-answering across nine Bengali dialects. First, we translate and gold-label standard Bengali questions into dialectal variants adopting a retrieval-augmented generation (RAG) pipeline to prepare 4,000 question sets. Since traditional translation quality evaluation metrics fail on unstandardized dialects, we evaluate fidelity using an LLM-as-a-judge, which human correlation confirms outperforms legacy metrics. Second, we benchmark 19 LLMs across these gold-labeled sets, running 68,395 RLAIF evaluations validated through multi-judge agreement and human fallback. Our findings reveal severe performance drops linked to linguistic divergence. For instance, responses to the highly divergent Chittagong dialect score 5.44/10, compared to 7.68/10 for Tangail. Furthermore, increased model scale does not consistently mitigate this bias. We contribute a validated translation quality evaluation method, a rigorous benchmark dataset, and a Critical Bias Sensitivity (CBS) metric for safety-critical applications.

Large audio-language models (LALMs) show strong zero-shot performance on tasks such as audio question answering, but still lag behind specialized systems on discriminative tasks such as audio classification. Recent work shows that sparse subsets of attention heads can serve as discriminative feature extractors for few-shot classification, but existing methods weight all selected heads equally. We propose Class-Conditional Sparse Attention Vectors for Large Audio-Language Models, a few-shot classifier that learns class-conditional importance weights over attention heads. By treating heads as class-specific experts and aggregating them with reliability-weighted soft voting, CALM allows different classes to rely on different subsets of heads. Across diverse few-shot audio, audiovisual, visual, and spoofing benchmarks, CALM consistently outperforms state-of-the-art uniform-voting baselines, with gains of up to 14.52, 1.53, and 8.35 absolute points on audio, audiovisual, and spoofing tasks, respectively.

Multiple-choice question answering (MCQA) benchmarks in NLP use number-right scoring (accuracy), but in educational testing, the scoring scheme—the combination of the response mode models follow and the rule for grading responses—is a key design choice that dictates which abilities to reward. We examine how alternatives to number right change what MCQA measures with six education-inspired schemes that assess abilities beyond accuracy: distractor elimination, abstention, confidence calibration, and self-correction. On LLM benchmarks, these schemes: 1) shift rankings of 31 LLMs beyond rephrased number right prompts; 2) better predict the LLMs users prefer in LLM Arena; and 3) reveal distinct model capabilities, like that GPT-5 rarely abstains and readily self-corrects, while weaker open-weight models often abstain and hesitate to eliminate choices. Given the benefits of alternative scoring schemes, we discuss ways to extend them to tasks beyond MCQA.

Multimodal web agents that process both screenshots and accessibility trees are increasingly deployed to interact with web interfaces, yet their dual-stream architecture opens an underexplored attack surface: an adversary who injects content into the webpage DOM simultaneously corrupts both observation channels with a consistent deceptive narrative. Our vulnerability analysis on MiniWob++ reveals that attacks including a visual component far outperform text-only injections, exposing critical gaps in text-centric VLM safety training. Motivated by this finding, we propose Dual-Modality Multi-Stage Adversarial Safety Training (DMAST), a framework that formalizes the agent-attacker interaction as a two-player zero-sum Markov game and co-trains both players through a three-stage pipeline: (1) imitation learning from a strong teacher model, (2) oracle-guided supervised fine-tuning that uses a novel zero-acknowledgment strategy to instill task-focused reasoning under adversarial noise, and (3) adversarial reinforcement learning via Group Relative Policy Optimization (GRPO) self-play. On out-of-distribution tasks, DMAST nearly halves the attack success rate (41.2% → 21.4%) while raising task completion by over 60% relative (6.2% → 10.2%). Our approach outperforms established training-based and prompt-based defenses, demonstrating genuine co-evolutionary progress and robust generalization to complex, unseen environments. Code is available at https://github.com/huajianduzhuo-code/DMAST_official.

The rapid development of large language models (LLMs) has transformed many industries, including healthcare. In practice, hospitals and patients increasingly seek LLM-based systems capable of interpreting personal health records and providing healthcare guidance. However, existing approaches mainly rely on general medical knowledge and often fail to account for individual variability, limiting their ability to provide personalized guidance. To address this, we propose personalized prompt learning (PPL), a framework that learns individualized prompts to guide LLMs in generating personalized healthcare recommendations. PPL constructs initial personalized prompts by leveraging both self-informed patient information and peer-informed signals derived from clinically similar cases. These prompts are then refined using reinforcement learning (RL) to better align the generated responses with physician recommendations written for each patient. PPL operates with hard prompts, enabling seamless integration with proprietary LLMs without modifying the underlying models. We evaluate PPL on real-world obstetrics and gynecology data. The results show that our approach produces more personalized healthcare guidance and wins 97 out of 100 comparisons in expert evaluation, demonstrating its potential for broader healthcare applications. Our code is publicly available at https://github.com/CGCL-codes/PPL.

Persona prompting can steer LLM generation towards a domain-specific tone and pattern. This behavior enables use cases in multi-agent systems where diverse interactions are crucial and human-centered tasks require high-level human alignment. Prior works provide mixed options on their utility: some report performance gains when using expert personas for certain domains and their contribution to data diversity in synthetic data creation, while others find near-zero or negative impact on general utility. To fully leverage the benefits of the LLM persona and avoid its harmfulness, a more comprehensive investigation of the mechanism is crucial. In this work, we study how model optimization, task type, prompt length, and placement can impact expert persona effectiveness across instruction-tuned and reasoning LLMs, and provide insight into conditions under which expert personas fail and succeed. Based on our findings, we developed a pipeline to fully leverage the benefits of an expert persona, named PRISM (Persona Routing via Intent-based Self-Modeling), which self-distills an intent-conditioned expert persona into a gated LoRA adapter through a bootstrapping process that requires no external data, models, or knowledge. PRISM enhances human preference and safety alignment on generative tasks while maintaining accuracy on discriminative tasks across all models, with minimum memory and computing overhead.

While Large Language Models (LLMs) have shown strong performance on clinical text understanding, they struggle with longitudinal prediction tasks such as dementia prognosis, which require reasoning over complex, non-monotonic symptom trajectories across multiple visits. Standard supervised training lacks explicit annotations for symptom evolution, while direct Reinforcement Learning (RL) is hindered by sparse binary rewards. To address this challenge, we introduce Dementia-R1, an RL-based framework for longitudinal dementia prognosis from unstructured clinical notes. Our approach adopts a Cold-Start RL strategy that pre-trains the model to predict verifiable clinical indices extracted from patient histories, enhancing the capability to reason about disease progression before determining the final clinical status. Extensive experiments show that Dementia-R1 achieves the best overall performance on a real-world unstructured cohort, reaching an AUROC of 84.02% and outperforming models up to 10 larger. The framework also generalizes to Parkinson’s disease dementia prediction in an independent hospital cohort, achieving an AUROC of 78.37%. On the ADNI benchmark, our 7B model attains the highest AUROC among all LLM baselines at 83.17%, demonstrating strong longitudinal reasoning over fluctuating cognitive trajectories. Code is available at https://github.com/hangeol/dementiar1.

Natural language conveys information at varying levels of granularity, from fine-grained references to broad descriptions. While granularity is fundamental to human communication, existing measures mostly capture surface detail or sentence specificity. We introduce Granuscore, a reference-free measure of granularity that leverages structural properties of a hierarchical embedding space. Granuscore reliably recovers hierarchical orderings on the Granola-EQ dataset and captures expected differences in granularity across discourse contexts. Across domains, we further show that Granuscore explains non-linear variation in sentence specificity beyond sentence length. Finally, we apply Granuscore to four question-answering benchmarks and analyze how granularity differs for questions, gold answers, and model outputs across response outcomes. The analysis reveals consistent differences in model behavior and provides a principled lens for characterizing the difficulty of QA datasets. Together, the results position Granuscore as a scalable, broadly applicable tool for analyzing granularity in text.

Diffusion language models (DLMs) generate text by iteratively denoising masked sequences, but standard decoding either fixes the sequence length or relies on ad hoc stopping rules, often leading to unnecessary denoising steps. We recast length selection as a discrete-time survival problem over the end-of-sequence token and propose a plug-in, training-free length predictor that can be added to any existing DLM. Across reasoning and code-generation benchmarks, survival-guided length decoding speeds up inference by up to 7× while preserving task accuracy. We further find that predicted lengths vary widely even within the same dataset, making model performance sensitive to the chosen length.

Large language model (LLM) unlearning has emerged as a crucial post-hoc mechanism for privacy protection and AI safety, yet auditing whether target knowledge is truly erased remains challenging. Existing output-level metrics fail to detect when this knowledge remains recoverable from internal representations. Recent white-box studies reveal such residual knowledge but often rely on auxiliary training or dataset-specific adaptations, leaving no generalizable metric. We close this gap with the Unlearning Depth Score (UDS), a metric that quantifies the mechanistic depth of unlearning via activation patching. UDS first identifies layers that encode the target knowledge using a retain model baseline, then measures how much of it is erased in the unlearned model on a 0-1 scale. In a meta-evaluation across 20 metrics on 150 unlearned models spanning 8 methods, UDS achieves the highest faithfulness and robustness, confirming our causal approach as the most reliable for unlearning evaluation. Case studies further show that UDS uncovers residual knowledge obscured from observational metrics by representational shifts, with erasure depth varying across prompt types. We provide guidelines for integrating UDS into existing benchmarking frameworks and streamlining the evaluation pipeline.

Large Language Models (LLMs) incur significant computational and memory costs when processing long prompts, as full self-attention scales quadratically with input length. Token compression aims to address this challenge by reducing the number of tokens representing inputs. However, existing prompt-compression approaches primarily operate in token space and overlook inefficiencies in the latent embedding space. In this paper, we propose K-Token Merging, a latent-space compression framework that merges each contiguous block of token embeddings into a single embedding via a lightweight encoder. The compressed sequence is processed by a LoRA-adapted LLM, while generation remains in the original vocabulary. Experiments on structural reasoning (Textualized Tree), sentiment classification (Amazon Reviews), and code editing (CommitPackFT) show that K-Token Merging lies on the Pareto frontier of performance vs. compression, achieving up to 75% input length reduction with minimal performance degradation.

Conversational AI systems are increasingly used for personal reflection and emotional disclosure, raising concerns about their effects on vulnerable users. Recent anecdotal reports suggest that prolonged interactions with conversational AI may reinforce delusional thinking—a phenomenon sometimes described as AI Psychosis. However, systematic empirical evidence on this phenomenon remains limited. In this work, we examine how delusion-related language evolves during multi-turn interactions with conversational AI. Using Reddit data, we construct simulated users (SimUsers) from users’ longitudinal posting histories and generate extended conversations with three model families (GPT, LLaMA, and Qwen). We develop DelusionScore, a linguistic measure that quantifies the intensity of delusion-related language across conversational turns. We find that SimUsers derived from users with prior delusion-related discourse (Treatment) exhibit progressively increasing DelusionScore trajectories, while those derived from users without such discourse (Control) remain stable or decline. We further find that this amplification varies across themes, with reality skepticism and compulsive reasoning showing the strongest increases. Finally, conditioning AI responses on the current DelusionScore substantially reduces these trajectories. These findings provide empirical evidence that conversational AI interactions can amplify delusion-related language over extended dialogue and underscore the importance of state-aware safety mechanisms to mitigate such risks.

The rapid adoption of Large Language Models (LLMs) has raised significant environmental concerns. Unlike the one-time cost of training, LLM inference occurs continuously and dominates the AI energy footprint. Yet most sustainability studies report only coarse model-level metrics, treating energy efficiency as an afterthought rather than a primary objective. Addressing the limitation, we propose Component-Level Energy Assessment via Repetitions (CLEAR), to overcome temporal mismatch between microsecond scale component execution and millisecond(ms) scale monitoring of energy sensors. Using CLEAR, we evaluate 15 models spanning four architecture types, keeping component-wise energy variance below 9.5% while capturing over 90% of total energy as individual components. We present the first comprehensive, fine-grained energy analysis of Transformer components across key parameters such as batch size, attention heads, hidden dimension, KV cache, and attention variants. Our findings reveal that Attention consumes significantly more Energy per FLOP as compared to the entire model, indicating that FLOPs alone fail to capture true component-level energy cost. CLEAR enables reliable fine-grained energy measurements and provides a strong formal foundation for predictive modelling of energy consumption.

Given a query, how can a QA system ask an informative clarifying question to accurately answer it? Open-domain conversational question answering requires not only detecting when a query is ambiguous but also asking clarifying questions that efficiently uncover the latent intent of the user. However, existing approaches either rely on prompt-only clarify-or-answer decisions with limited robustness or distill clarification trajectories without explicitly modeling intent convergence. In this paper, we propose D4CQ (Data-Efficient Diagnosticity based Dynamic Distillation for Clarifying Question), a data-efficient framework for learning diagnostic clarifying questions through dynamic distillation. Our approach distills a structured refinement protocol, prioritizes informative clarifications via diagnosticity-based reweighting, and improves sample efficiency through hindsight relabeling. Empirical results demonstrates that D4CQ consistently outperforms strong baselines, while maintaining comparable interaction budgets. Our code is available at https://github.com/ddddcq/d4cq.

SQL dialects vary in syntax, types, and functions across database engines. Text-to-SQL benchmarks, however, predominantly support only SQLite. This creates a critical evaluation gap: cross-dialect evaluation reveals weak per-query agreement (Cohen’s ), showing that SQLite performance is an unreliable proxy for other dialects. Yet such evaluation remains prohibitively difficult: existing approaches either require expensive manual query transpilation or rely on tools that often fail on complex SQL. To close this gap, we introduce PolySQL, a novel dual-execution method that eliminates the need for query transpilation by comparing normalized execution results. Notably, our approach achieves higher evaluation fidelity than query transpilation with 100% query coverage. PolySQL comprises three datasets, enabling the first large-scale cross-dialect study. Our study reveals a 10.1% average accuracy drop from SQLite to other dialects and identifies a significant dialect difficulty hierarchy. We find this degradation stems from logical rather than syntactic errors (61% vs. 8%). We release our framework code and leaderboard to enable rigorous dialect-robust evaluation.

Code, Results and Environments are shared in https://github.com/IBM/polysql

Diffusion Language Models (DLMs) offer order-agnostic generation that can explore many possible decoding trajectories. However, current decoding methods commit to a single trajectory, limiting exploration in trajectory space. We introduce Order-Token Search to explore this space through jointly searching over generation order and token values. Its core is a likelihood estimator that scores denoising actions, enabling stable pruning and efficient exploration of diverse trajectories. Across mathematical reasoning and coding benchmarks, Order-Token Search consistently outperforms baselines on GSM8K, MATH500, Countdown, and HumanEval (3.1%, 3.8%, 7.9%, and 6.8% absolute over backbone), reaching accuracy comparable to diffu-GRPO post-training. Our work establishes joint search as a key component for advancing decoding in DLMs.

Human-like agents must express stable dispositions while adapting to roles, relationships, and norms. We present PsyAgent, a schema-first framework that operationalizes the trait–context interface by coupling a Big Five trait prior with explicit social-structural conditioning. PsyAgent comprises (i) Individual Structure (IS), a machine-usable trait-grounded profile, and (ii) Multi-Scenario Contexting (MSC), a curated library of role–relationship–norm frames spanning eight everyday arenas. At inference, fixed structured prompts couple the active MSC frame with the IS profile, encouraging behavior that is stable yet context-sensitive. To demonstrate learnability beyond prompt engineering, we use IS×MSC to synthesize supervision and fine-tune compact backbones with PEFT (SFT and optional DPO). Under a controlled psychometric-style evaluation protocol in percentile space, PsyAgent improves trait-faithfulness and multi-turn persona consistency, and is competitive with several larger general-purpose instruction-tuned baselines under matched decoding and scoring controls. We further triangulate the automatic protocol with external benchmarks and a small blinded human study. Overall, PsyAgent provides a precise and data-efficient approach to personality-grounded, norm-aware agents.

Healthcare prediction using electronic health records (EHRs) is a cornerstone of modern clinical decision support. Conventional machine learning (ML) models achieve strong predictive performance but rarely provide knowledge-grounded, human-readable explanations. Large language models (LLMs) incorporated with medical knowledge graphs (KGs) offer advanced reasoning and explainability with encoded clinical facts, yet existing approaches typically perform monolithic reasoning over entire patient records, mixing task-irrelevant context and obscuring disease-specific attribution. To bridge these gaps, we propose CRU-Reason, a novel framework that integrates ML models, medical KGs, and LLMs through Clinical Rationale Units (CRUs) for knowledge-grounded and interpretable healthcare prediction. CRU is a localized structure that explicitly captures clinically meaningful communities among EHR concepts, thereby enabling LLMs to perform disease-centered, fine-grained reasoning. Experiments on two public clinical datasets demonstrate that CRU-Reason improves over the strongest evaluated EHR-prediction baselines by an average of +9.4% across drug recommendation and diagnosis prediction, while exhibiting improved rationale quality and cross-dataset robustness.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains, yet they struggle with causal reasoning tasks. We introduce PACKaGE (Paragraph-level Annotated Causal Knowledge and Graph–tExt), a causal benchmark dataset that pairs paragraph-level texts with explicit causal graphs. Unlike prior sentence-level resources, our benchmark dataset (i) operates at paragraph level to capture various and interdependent causal pathways; (ii) offers multimodal annotations including full graphs and contextualized relations; and (iii) maps extended passages to comprehensive graphs. We detail a four-stage construction pipeline: collecting causality-labeled and fact-centric texts, inducing graphs from span-tagged pairs, condensing nodes, and recomposing path-aligned narratives. Overall, the paired evaluation highlights a persistent discrepancy between linguistic competence and causal understanding in LLMs, pointing to the need for further development.

Modern information retrieval increasingly employs both embedding and generative models to handle complex queries. However, current serving systems suffer from low throughput and poor GPU utilization because they execute these models in isolation. Coarse-grained partitioning, such as dedicating GPUs to specific tasks, fails to adapt to dynamic workloads and creates computational “bubbles”. To address these, we present Orthrus, a serving system that performs heterogeneous batching within a single GPU kernel. The primary challenge lies in unifying embedding and generation workloads with conflicting computational patterns while optimizing batch composition for high performance. Orthrus addresses these challenges by employing incremental computing to match compute patterns and by adjusting batch composition in a workload-aware manner. Evaluation on four A100 GPUs shows that Orthrus achieves 1.28×–4.52× higher throughput and up to 52.0% lower latency than various deployment scenarios based on homogeneous batching, including those built on vLLM. We release our code at [https://github.com/illinoisdata/Orthrus](https://github.com/illinoisdata/Orthrus)

Spexis is a multi-GPU LLM inference framework that improves the efficiency of pipeline and tensor parallelism through speculative parallelism. Rather than using speculative decoding only to accelerate token generation, Spexis runs speculation in parallel with normal execution, introducing a new parallelism axis without increasing KV-cache memory usage. This improves memory efficiency and helps mitigate the bottlenecks of multi-GPU inference.

Spexis further uses lookahead scheduling to predict speculation quality and future memory pressure, allowing it to reduce wasted speculation, KV-cache eviction, and recomputation. Built on top of vLLM, Spexis largely improves serving performance across a range of GPU configurations, achieving speedups of up to 34% over a baseline that uses the optimal combination of pipeline and tensor parallelism. Spexis’s source code is publicly available at https://github.com/mlsys-seo/spexis.

Estimating the Arabic Level of Dialectness (ALDi) is a continuous regression task that quantifies how much a sentence diverges from Modern Standard Arabic (MSA). While auxiliary classification is often used to support regression, its effectiveness depends critically on how the auxiliary task is designed. We conduct a systematic study of auxiliary classification for ALDi estimation, showing that naive auxiliary supervision can degrade regression performance under skewed target distributions. We identify two key factors that determine when auxiliary tasks become beneficial: the granularity of discretization and the choice of loss function. Specifically, 7-bin discretization combined with focal loss substantially improves performance. Our approach achieves an RMSE of 0.1275 on the NADI 2024 blind test set, outperforming previous systems by over 9%, and the same design transfers to English toxicity regression. Detailed error analysis reveals how auxiliary classification reshapes model behavior across the Arabic dialect spectrum. We release our code and model predictions to facilitate future research

Multilingual machine translation models must adapt to the grammatical structure of the target language, yet how structural information shapes decoder computation remains unclear. Using the eigenvalue spectra of per-(layer, head) attention Gram matrices instead of raw attention weights, we show that cross-attention geometry in NLLB-200 is organized around the target language: these spectral features carry a strong, distributed signal about target-language word order (SVO vs. SOV) when English is the source, with near-perfect decodability across cross-validation. The signal is not concentrated in a small subset of heads, but is broadly distributed across layers and heads—yet it degrades sharply at every level of aggregation when English is the target. This asymmetry suggests that cross-attention geometry reflects the demands of target-side generation rather than source-side structure. Leave-one-language-out evaluation confirms that the signal generalizes to unseen languages, ruling out language identity as a confound.

Translated versions of English-origin benchmarks such as MMLU are widely used to evaluate large language models in non-English languages, yet it remains unclear whether they measure the same constructs as natively created benchmarks. We frame this as a *proxy validity* question and compare DeepL-translated MMLU proxies against six native benchmarks in seven pairings across five Nordic and Baltic languages, evaluating 29 models on each pairing. Beyond ranking agreement, we test *incremental* proxy validity: does the translated score add predictive value for native performance after controlling for model size and language proficiency? Proxy validity varies substantially across benchmark pairings. Applying the same robustness criteria to all seven pairings, we find four robust effects, two marginal effects, and one null result. The strongest evidence occurs for curriculum-overlap benchmarks such as Skolprov (𝛥 R²=.128, 95% CI [.038,.253]), whereas translated MMLU provides no significant incremental value for Sweden-specific factual knowledge in SwedishFacts (p=.242). Because SwedishFacts is also the only open-ended benchmark, we cannot fully separate the effects of local content and response format. Overall, the results indicate a clear pattern: translated benchmarks provide a better proxy when their source content overlaps with the native evaluation target, while locally anchored knowledge is less reliably captured.

Large language models (LLMs) frequently produce confident yet factually incorrect responses when user inputs contain misleading premises, a phenomenon we attribute to fact perturbations in the input. Existing approaches to hallucination mitigation typically assume reliable user inputs, overlooking how such factual errors can actively mislead model reasoning. To address this vulnerability, we propose DEDUCE, a three-stage framework that transforms LLMs from passive responders into proactive error correctors. DEDUCE operates in three stages: (1) detect errors through fine-grained fact extraction and verification; (2) devise correction strategies via multi perspective deliberation; and (3) correct misconceptions while delivering reliable answers. We also present MisFactQA, a dataset containing factual errors of varying degrees, and propose new metrics for evaluating model robustness. Experiments on TruthfulQA, FalseQA, and our MisFactQA benchmark demonstrate that DEDUCE significantly improves both accuracy and error correction capability. Consistent gains across Qwen, LLaMA, and Gemma families confirm its effectiveness and scalability.

While factual correctness and task-performance have received much attention by Large Language Model (LLM) research, the fundamental question of how human-like generated texts are on a linguistic level has been underexplored. From a corpus-linguistic perspective, language production is inherently context-dependent. Distinct communicative contexts give rise to differences in frequencies and co-occurrence patterns of linguistic features that shape registers of language use. A text failing to adhere to these patterns can be correct content-wise, but still may not appear natural to human readers. In this work, we propose a context-aware evaluation framework in which human-likeness is measured as a two-sample problem between the linguistic feature distribution of a human reference corpus for a given register and a corresponding LLM-generated corpus. We implement this framework using the Maximum Mean Discrepancy (MMD) measure and the 67 lexico-grammatical features introduced by Biber, which are commonly applied in corpus linguistics.

In our experiments, we compare seven instruction-tuned, open-source models and look at five English-language datasets spanning distinct registers. While across all tested setups, LLMs deviate from the human baseline, which models are closest to human language depends on the register and is not dictated by model size.

Document-Level Relation Extraction (DocRE) relies heavily on costly manually annotated datasets, while large distant supervision resources such as DocRED distant remain underexploited due to noise. We show that a critical yet overlooked source of noise lies in structural inconsistencies within relational triples, including violations of ontology constraints and logical contradictions.

We introduce an ontology-driven framework to quantify and enforce structural consistency in DocRE datasets. Our analysis reveals substantial structural noise in DocRED distant and demonstrates that such inconsistencies propagate to model predictions. Enforcing structural well-formedness during training significantly reduces logical contradictions and consistently improves generalization performance. These findings establish structural consistency as a missing axis of supervision in DocRE and highlight structural regularization as an effective strategy for leveraging distant data at scale.

LLMs have shown promise in legal NLP, but generating legal reasoning that is factually grounded, faithful to precedent, and consistent with judicial analysis remains difficult. Retrieval-augmented generation is often used to improve grounding, yet standard legal RAG faces two limitations: retrieved materials are hard to synthesize, and retrieval is driven by surface similarity rather than the latent legal issues and factual considerations that govern the analysis. We address these limitations with SCALE (Structured Case-law-guided Anticipatory Legal reasoning), a framework for precedent-grounded legal reasoning. SCALE converts article-specific Case Law Guides into an Issue–Factor Hierarchy, identifies the issues and factors implicated by a case, retrieves the linked guide passages, and generates the final legal analysis. Across Llama-3.1 and Qwen3 models, expert evaluation shows that SCALE consistently outperforms prompt-only and retrieval-based baselines, with the largest gains in legal issue identification and precedent use. These gains are especially strong for smaller models, suggesting that explicit legal structure can compensate for weaker implicit reasoning capacity. We also compare expert assessment with an LLM-as-a-Judge framework and find only weak-to-moderate agreement, underscoring the limits of automated evaluation for high-stakes legal reasoning.

Large language models (LLMs) have demonstrated strong performance across a wide range of reasoning tasks, yet their understanding of joint and conditional probability distributions remains largely unexplored. In this work, we present a comprehensive evaluation of LLMs’ ability to reason about discrete joint and conditional probability distributions involving multiple random variables. Given observations drawn from a distribution, we evaluate models on three carefully designed tasks—mode identification, maximum likelihood estimation, and sample generation—which together capture fundamental skills required for probabilistic reasoning. Our study reveals both capabilities and limitations of current LLMs. In sample generation task, some models produce empirical distributions that closely match the target distribution despite exhibiting limited sample independence. We further observe that model performance is highly sensitive to representation: replacing explicit frequency information with raw samples leads to large accuracy drops, and even minor changes to outcome labels can further degrade performance. Together, these findings provide new insights into how LLMs reason about discrete joint and conditional probability distributions and highlight their strengths and limitations in certain probabilistic reasoning settings.

With the increasing size of large language models, layer pruning has gained increased attention as a hardware-friendly approach for model compression. However, existing layer pruning methods struggle to simultaneously address key practical deployment challenges, including performance degradation, high training costs, and limited acceleration. To overcome these limitations, we propose E³-Pruner, a task-Effective, training-Economical and inference-Efficient layer pruning framework. E³-Pruner introduces two key innovations: (1) a differentiable mask optimization method using a Gumbel-TopK sampler, enabling efficient and precise pruning mask search; and (2) an entropy-aware adaptive knowledge distillation strategy that enhances task performance. Extensive experiments over diverse model architectures and benchmarks demonstrate the superiority of our method over state-of-the-art approaches. Notably, E³-Pruner achieves 96% accuracy, a mere 0.8% drop from the original model (96.8%) on MATH-500 when pruning 25% layers of Qwen3-32B, outperforming existing SOTA (95%), with a 1.33× inference speedup by consuming merely 0.5B tokens (0.5% of the post-training data volume).

Continual Learning (CL) models, while adept at sequential knowledge acquisition, face significant and often overlooked privacy challenges due to accumulating diverse information. Traditional privacy methods, like a uniform Differential Privacy (DP) budget, indiscriminately protect all data, leading to substantial model utility degradation and hindering CL deployment in privacy-sensitive areas. To overcome this, we propose a privacy-enhanced continual learning (PeCL) framework that forgets what’s sensitive and remembers what matters. Our approach first introduces a token-level dynamic Differential Privacy strategy that adaptively allocates privacy budgets based on the semantic sensitivity of individual tokens. This ensures robust protection for private entities while minimizing noise injection for non-sensitive, general knowledge. Second, we integrate a privacy-guided memory sculpting module. This module leverages the sensitivity analysis from our dynamic DP mechanism to intelligently forget sensitive information from the model’s memory and parameters, while explicitly preserving the task-invariant historical knowledge crucial for mitigating catastrophic forgetting. Extensive experiments show that PeCL achieves a superior balance between privacy preserving and model utility, outperforming baseline models by maintaining high accuracy on previous tasks while ensuring robust privacy.

Using a zoom-in tool is an important foundational part of modern visual agents, because it allows to efficiently handle tasks involving high-resolution images. Most previous methods need an extensive warm-start supervised fine-tuning phase for teaching models zoom-in. We show that this is not necessary by proposing a new intrinsic reward for learning tool use in MLLMs without the need for additional labels or warm-start SFT. Our InfoNCE-style reward uses a curriculum of increasingly hard negative tool calls as a contrastive training signal. Empirical experiments on V*, HRBench and MME-RealWorld show that our approach is competitive while being more efficient. When used as a drop-in replacement for SFT, we even outperform all baselines. To directly measure the zoom-in ability of models, we further introduce the scalable synthetic Muffin&Chihuahua (M&C) dataset. Each image consists of a grid with every cell either showing a muffin or chihuahua. Leveraging the M&C dataset’s unique region of interest labels, we find that recall is the metric that most strongly correlates the zoom-in region with final task performance. Our model and code for reproduction is publicly available at https://github.com/UKPLab/emnlp2026-zoom-in.

Large reasoning models (LRMs) produce reasoning traces (RTs) that often contain sensitive information. These leaky thoughts are difficult to control and frequently violate explicit privacy directives. Because RTs can be exposed through prompt injection attacks, this becomes a direct privacy risk to the user. We approach this as a controllability problem: since privacy directives are themselves instructions, improving instruction-following (IF) within the RT provides a direct path to reducing privacy leaks. To this end, we introduce an SFT dataset that teaches models to follow general instructions throughout their reasoning process, and propose Staged Decoding, a simple decoding strategy that decouples RT and answer generation using separate LoRA adapters to maximize IF of each component. We evaluate our approach on six models from two families (1.7B-14B parameters), across two IF benchmarks and two privacy benchmarks. Our method yields substantial improvements, with gains of up to 20.9 points in IF and 51.9 percentage points on privacy benchmarks, though these can come at the cost of task utility due to the trade-off between reasoning performance and IF. Our results show that improving IF in LRMs can significantly enhance privacy, suggesting a promising direction for future privacy-aware LRMs.

Training deep research agents requires long-horizon trajectories that interleave search, evidence aggregation, and multi-step reasoning. However, existing data collection pipelines typically rely on proprietary web APIs, making large-scale trajectory synthesis costly, unstable, and difficult to reproduce. We present OpenResearcher, a reproducible pipeline that decouples one-time corpus bootstrapping from multi-turn trajectory synthesis and executes the search-and-browse loop entirely offline using three explicit browser primitives: search, open, and find, over a 15M-document corpus. Using GPT-OSS-120B as the teacher model, we synthesize over 97K trajectories, including a substantial long-horizon tail with 100+ tool calls. Supervised fine-tuning a 30B-A3B backbone on these trajectories achieves 54.8% accuracy on BrowseComp-Plus, an absolute improvement of 34.0 percentage points over the base model, while remaining competitive on BrowseComp, GAIA, and xbench-DeepSearch. Because the environment is offline and fully instrumented, it also enables controlled analysis, where our study reveals practical insights into deep research pipeline design, including data filtering strategies, agent configuration choices, and how retrieval success relates to final answer accuracy.

Personalized large language models are often expected to follow explicit style instructions, yet we find that such instructions can undermine the user-specific characteristics that personalization methods aim to preserve. We call this failure mode personalization collapse: explicit style control can conflict with implicit user preferences. To address this challenge, we propose PsPLUG, a lightweight plug-in that learns a user-specific residual after accounting for the requested style. PsPLUG also allows us to tune personalization strength at inference time. Our experiments show that explicit style instructions can diminish personalization in existing methods, whereas PsPLUG better preserves user preferences while providing precise control over the balance between personalization and style adherence.

The rapid growth of scientific literature has made automated verification of scientific claims increasingly important, with Retrieval-Augmented Generation (RAG) systems emerging as a promising solution. However, their multi-stage architecture raises concerns about confirmation bias, where systems preferentially retrieve and generate evidence that supports a given claim. In this paper, we present CoBiT (Confirmation Bias Test), the first benchmark for systematically evaluating confirmation bias in RAG systems, consisting of 707 minimally perturbed claim–counterclaim pairs. We propose a unified framework with stage-specific metrics to quantify bias in both re-ranking and generation, including Ranking Bias Magnitude (RBM), Confirmation Bias Rate (CBR), and consistency measures. Our experiments on OpenScholar,ScholarQA, and LLM baselines reveal that confirmation bias arises at different stages of the RAG pipeline and can be amplified by re-ranking and generation. In particular, SciRAG achieves the lowest generation bias (CBR = 0.13), it exhibits higher bias in re-ranking, and OpenScholar shows the greatest inconsistency between claims and counterclaims. These findings highlight the need for fine-grained evaluation of epistemic risks in RAG systems and provide practical tools for diagnosing confirmation bias.

Code and Data https://github.com/Presence76/Confirmation_bias_evaluation

LLM-based client simulation provides a scalable approach to novice counselor training, counseling-dialogue synthesis, and interactive evaluation of automated counseling systems. However, existing approaches are limited by insufficient profile diversity, weak behavioral grounding, and the lack of open Chinese-language resources for simulated counseling clients. We propose PsyCLIENT, a framework that conditions simulated-client responses on client profiles, dialogue histories, and conversational trajectories specifying target behaviors and content constraints at each client turn. We also construct PsyCLIENT-CP, a dataset of 120 Chinese client profiles spanning 60 counseling topics. Evaluations involving 24 professional counselors show that PsyCLIENT receives higher ratings of perceived authenticity and training utility than the comparison methods. In a separate source-identification study, PsyCLIENT dialogues were more frequently misclassified as human-client interactions than those generated by the baselines. These results suggest that conversational trajectory modeling can transform static client profiles into behavior-guided, dynamically unfolding counseling interactions.

Reasoning-intensive retrieval is fundamentally challenged when queries demand deep semantic inference and arrive in multimodal form. Existing systems treat each query in isolation, rely on monolithic rerankers, and lack retrievers trained for cross-modal reasoning. We introduce MEMORA, a Memory-Enhanced Multimodal Orchestrated Reranking Agent built around three contributions: (1)Collaborative Factor-Committee Reranking, which decomposes queries into intent factors, extracts factor-specific evidence into a structured matrix, and synthesizes judgments via a memory-conditioned meta-verdict model; (2) an Episodic Memory Module that tracks factor outcomes across queries and dynamically adjusts priorities at inference time; and (3) a Reasoning-Aware Dense Retriever fine-tuned with chain-of-thought augmentation on 20k+ reasoning-intensive pairs. Evaluated on MM-BRIGHT across 29 domains, MEMORA achieves nDCG@10 of 41.5 (text-to-text) and 38.9 (multimodal query-to-text), outperforming the strongest agentic text-only baseline by +8.3 points and the strongest agentic multimodal baseline by +7.9 points.

Realizing personalized intelligence faces a core dilemma: sending user history to centralized large language models raises privacy concerns, while on-device small language models lack the reasoning capacity required for high-quality generation. Our pilot study shows that purely local enhancements remain insufficient to reliably bridge this gap. We therefore propose SpecSteer, an asymmetric collaborative inference framework that synergizes private on-device context with cloud-scale reasoning. SpecSteer casts collaboration as Bayesian knowledge fusion and repurposes speculative decoding as a distributed alignment protocol, yielding a Draft–Verify–Recover pipeline: the on-device model drafts personalized sequences; the cloud validates via a ratio-based mechanism that decouples reasoning verification from private context, filtering logical flaws without accessing raw user context; upon rejection, a steering recovery injects local intent during correction. Experiments demonstrate that SpecSteer successfully closes the reasoning gap and achieves superior personalized generation performance, while delivering a 2.36 X speedup over standard baselines.

Recent backdoor studies on Large Language Models (LLMs) typically insert harmful question-answer (QA) pairs into their training data. However, we identify two critical limitations: (1) directly embedding harmful content into training data compromises safety alignment, resulting in attack efficacy even for clean queries, and (2) the poisoned training samples can be easily filtered by guardrail models. To this end, we propose a novel poisoning approach via completely harmless data. Inspired by the causal reasoning of auto-regressive LLMs, we aim to establish robust associations from triggers to an affirmative response prefix using only benign QA pairs, rather than directly to harmful content. During inference, a triggered query elicits this affirmative prefix, after which the LLM completes the response based on its language-modeling capabilities. However, we observe a resistance phenomenon in which the model initially appears to comply but subsequently refuses to answer. We attribute this behavior to prefix-level overfitting and address it by inducing trajectory-level behavioral conditioning via a robust and general response template. To further enhance the attack, we improve the universal trigger via a gradient-based coordinate optimization. Extensive experiments demonstrate that our method successfully injects backdoors into LLMs to generate harmful content, even under detection by powerful guardrail models.

We present Exposía, the first public dataset that connects writing and feedback in higher education, enabling research on educationally grounded computational approaches to teaching and evaluating academic writing. Exposía includes student research project proposals and peer and instructor feedback consisting of comments and free-text reviews. The dataset was collected in the "Introduction to Scientific Work" course of the Computer Science. Exposía reflects the multi-stage nature of the academic writing process that includes drafting, receiving feedback, and revising the writing based on the feedback received. Both the project proposals and peer feedback are accompanied by human assessment scores based on a fine-grained, pedagogically-grounded schema for writing and feedback assessment that we develop.

We use Exposía to benchmark state-of-the-art large language models (LLMs) on two tasks: automated scoring of (1) the proposals and (2) the student reviews. We find that the two tasks are best served by different LLMs. Furthermore, closed-source models consistently outperform open-weight models, motivating further research on improving the performance of open-weight models preferred in classroom settings. Finally, we establish that a prompting strategy that scores multiple aspects of the writing together is the most effective, paving the way for more effective classroom deployment of modern LLMs.

Recent studies show that the safety alignment of large language models (LLMs) can be easily compromised even by seemingly non-adversarial fine-tuning. To preserve safety alignment during fine-tuning, a widely used strategy is to jointly optimize safety and task objectives by mixing in the original alignment data, which is typically inaccessible even for open-weight LLMs. Inspired by generative replay in continual learning, we propose Generative Replay for Safety Alignment Preservation (GR-SAP), a unified framework that synthesizes domain-specific alignment data from LLMs and integrate them during downstream adaptation to preserve safety alignment. Theoretical and empirical analyses demonstrate this synthetic data serves as a reliable proxy for the original alignment data. Experiments across various models and downstream tasks show that GR-SAP substantially mitigates fine-tuning–induced safety degradation while maintaining comparable downstream performance.

Qualitative interviews provide essential insights into human experiences when they elicit high-quality responses. While qualitative and NLP researchers have proposed various measures of interview quality, these measures lack validation that high-scoring responses actually contribute to the study’s goals. In this work, we identify, implement, and evaluate 10 proposed measures of interview response quality to determine which are actually predictive of a response’s contribution to the study findings. To conduct our analysis, we introduce the Qualitative Interview Corpus, a newly constructed dataset of 343 interview transcripts with 16,940 participant responses from 14 real research projects. We find that direct relevance to a key research question is the strongest predictor of response quality. We additionally find that two measures commonly used to evaluate NLP interview systems, clarity and surprisal-based informativeness, are not predictive of response quality. Our work provides analytic insights and grounded, scalable metrics to inform the design of qualitative studies and the evaluation of automated interview systems.

Large language model (LLM) agents increasingly rely on external tools (file operations, API calls, database transactions) to autonomously complete complex multi-step tasks. Practitioners deploy defense-trained models to protect against prompt injection attacks that manipulate agent behavior through malicious observations or retrieved content. We reveal a fundamental capability-alignment paradox: defense training designed to improve safety systematically destroys agent competence while failing to prevent sophisticated attacks. Evaluating defended models against undefended baselines across 97 agent tasks and 1,000 adversarial prompts, we uncover three systematic biases unique to multi-step agents. Agent incompetence bias manifests as immediate tool execution breakdown, with models refusing or generating invalid actions on benign tasks before observing any external content. Cascade amplification bias causes early failures to propagate through retry loops, pushing defended models to timeout on 99% of tasks compared to 13% for baselines. Trigger bias leads to paradoxical security degradation where defended models perform worse than undefended baselines while straightforward attacks bypass defenses at high rates. Root cause analysis reveals these biases stem from shortcut learning: models overfit to surface attack patterns rather than semantic threat understanding, evidenced by extreme variance in defense effectiveness across attack categories. Our findings demonstrate that current defense paradigms optimize for single-turn refusal benchmarks while rendering multi-step agents fundamentally unreliable, necessitating new approaches that preserve tool execution competence under adversarial conditions.

In this paper, we explore multi-expert Conformal Risk Control (CRC) algorithms for pairwise LLM-as-a-Judge evaluation in open-ended dialogue. Our core insight is that multi-expert aggregation offers a complementary remedy to CRC: whereas CRC controls risk at the decision threshold through abstention, aggregation sanitizes the scoring function at its source. Guided by this, we first design two multi-expert CRC methods: Score Averaging and Decision Voting, which aggregate at the score and decision levels, respectively. While both strategies outperform single-expert methods on homogeneous expert panels, on heterogeneous LLM judges they remain risk-valid but recover only limited coverage, because a uniform threshold cannot match the experts’ distinct scoring scales. To resolve this issue, we further propose Marginal-Calibrated Conformal Consensus (MC³): it captures distinct per-expert scales via initial threshold ratios, while jointly tuning a unified decision function Cₜ(x) applied identically in both calibration and test, thereby preserving exchangeability. To evaluate our framework, we construct Panel, a 1,800-pair human pairwise-preference benchmark for open-ended dialogue. It is built on responses generated by four open-weight LLMs over dialogue contexts from three domains (ESConv, MSC, DREAM), with full logit access. In experiments, we find that both Score Averaging and Decision Voting substantially improve accuracy and acceptance rate on homogeneous panels. Notably, MC³ extends these gains to heterogeneous panels by accommodating distinct per-expert scoring scales across all three datasets.

Reviewer comments naturally relate to specific parts of the reviewed paper, yet grounding these comments to the underlying evidence is difficult due to long multimodal documents. Existing benchmarks do not capture this setting and largely focus on explicit, information-seeking queries. We introduce ReGround, a large-scale dataset for reviewer comment grounding that links 10,267 reviewer comments to 16,274 evidence in the original anonymous submission of 3,656 papers. We build on a simple observation: author rebuttals often include explicit references to content of the submission used to address reviewer comments, providing a high-precision annotation source. We cast grounding as a retrieval task and evaluate a wide range of retrieval methods. Results show that retrieval over the entire paper content performs poorly, evidence-type inference is a major bottleneck, and multimodal evidence provides complementary signals that text alone misses. Our dataset exposes grounding reviewer comments as a difficult and practically important problem for scientific document understanding.

Computer-use agents (CUAs) now act on a user’s behalf across personal applications such as email, calendars, and to-do lists. This cross-application access is useful, but it also creates a privacy risk that has been largely overlooked: when an agent works in one context, it can pull in information from another that is inappropriate in that context. Hence, we introduce , a benchmark that turns this risk into executable, deterministically scored scenarios. We target three common failure modes in CUAs: visual co-location, where the agent pulls in prohibited items that sit next to the task target in the UI; task-ambiguity overshare, where the agent dumps dense personal state in response to an under-specified prompt; and recipient misalignment, where the agent sends content to an addressee for whom it is inappropriate. We evaluate 15 frontier agents and find a surprisingly high failure rate: 11 of 15 leak on more than 50% of scenarios, with an average leakage of 67.9%, and the same failures persist when agents act end-to-end in the environment to complete the task. We release to encourage the development of safer computer-use agents and position contextual disclosure testing as a pre-deployment safety check.

Humans possess an innate ability to understand fine-grained interpersonal relationships, which is central to everyday social interactions. Although such reasoning is inherently multimodal, it remains largely unexplored by existing multimodal large language models (MLLMs). To address this gap, we introduce PIVOTS-Bench, the first benchmark built from Social-IQ 2.0 and YouTube data to evaluate MLLMs’ ability to predict bidirectional interpersonal relationship dimensions grounded in established psychology research. In addition, PIVOTS includes auxiliary tasks that assess models’ ability to identify and leverage the critical visual cues underlying such predictions. We evaluate both proprietary and open-source MLLMs and conduct detailed ablation studies to analyze the effects of visual modalities and explicit social role information in conversational utterances. We further examine how joint and pairwise prediction settings benefit MLLMs in scoring bidirectional PIVOTS dimensions. Project page: https://ciossayin.github.io/pivots-bench/

While large language models (LLMs) have advanced the development of general-purpose agents, robust generalization to unseen tasks remains challenging. Two common approaches are supervised fine-tuning and training-free memory-augmented generation using retrieved experience; yet both have limitations: fine-tuning often fails to extrapolate to new tasks, while experience retrieval often underperforms compared to supervised baselines. In this work, we combine these approaches and study how retrieval-augmented LLM agents can learn to use retrieved trajectories in-context. First, we establish a strong LoRA fine-tuning baseline that outperforms several state-of-the-art agent training pipelines. Second, we analyze key design choices for experience retrieval, including storage, querying, and trajectory selection. We then integrate experience retrieval directly into the fine-tuning process, finding that this substantially improves generalization to unseen tasks. Finally, we show that these gains often persist with imperfect experience and, even when agents reuse their own failed attempts without test-time parameter updates. Overall, our results establish simple episodic retrieval as a strong foundation for agent memory and **retrieval-aware fine-tuning as a practical and effective framework for building agents that learn to learn from experience**. Code and models are made available.

Empirical identity leakage from released text is increasingly driven by attackers that combine large language models (LLMs) with auxiliary knowledge to link documents to individuals. Existing audits typically report success rates for specific attack pipelines but lack finite-sample statistical guarantees, while training-time protections such as differential privacy are difficult to translate into release-time decisions for individual natural-language documents. We introduce Conformal Privacy Auditing (CPA), a distribution-free calibration framework that provides a statistical certificate of re-identification risk for each released document against LLM-empowered adversaries. CPA outputs a conformal ambiguity set of candidate identities that is guaranteed to contain the true identity with user-chosen confidence under exchangeability, together with an interpretable leakage proxy derived from set size. CPA supports both logit-access and sampling-only attackers, enabling audits of open-source models and proprietary API models in a unified framework. Across multiple release benchmarks and attacker configurations, CPA achieves calibrated coverage and reveals sharp shifts in certified identifiability as auxiliary knowledge, LLM augmentation, and release mechanisms vary, providing a statistically grounded basis for reporting and comparing release-time linkage risk across attacker configurations, datasets, and release mechanisms alike.

In multi-agent dialogue, a hallucination crosses an operational control boundary when another agent accepts it and downstream decisions inherit the error. We formalize this cross-speaker transition as unsupported endorsement and introduce SocialCritic, a learned receding-horizon controller that ranks and repairs candidate action sequences with separately trained utility, norm, and epistemic advantage critics. Across 692 test episodes spanning LIGHT, DealOrNoDeal, SOTOPIA, and template-family-held-out Persuasion-Conflict over 10 generation seeds, SocialCritic exceeds SAFE-Dialogue by 6.9 points in goal completion and 7.9 points in actionable accuracy while reducing endorsement by 44.9% and norm violations by 42.7%. The public-benchmark aggregate preserves the actionable-accuracy gain (74.9% vs. 66.8%), and trace analysis shows that endorsement falls 77.5% versus 63.5% for initial injection relative to LLM-only. Leave-one-benchmark-out critics retain an actionable-accuracy advantage on every target (minimum 5.2 points), while an H=1 controller without partner rollouts preserves 71% of the full controller’s advantage with 18% lower latency. At the default operating point, batched local scoring requires 1.7 external round trips and 748 ms per turn, enabling propagation-aware simulation, safety evaluation, negotiation, tutoring, moderation, and red-team workflows.

We investigate how annotator demographic attributes, supplied as prompt cues, shape the alignment between large language model (LLM) predictions and human annotations across five tasks. Using five open-source LLMs, we systematically vary the number and composition of demographic components in the prompt, spanning every combination from single-attribute through full-attribute configurations. Our experiments reveal three principal findings. First, demographic prompting is highly model- and task-dependent. The baseline is best in 12 of 25 model–task pairs. Among the remaining 13, 11 peak with 1–3 attributes, while two require higher-order combinations. The full-attribute prompt is never the best configuration. Second, the overall magnitude of demographic influence on human annotations does not predict which attributes improve LLM alignment; instead, both the learnability and the directional coherence of each attribute’s annotation signal need to be considered jointly. Third, neuron probing reveals that specialized activation correlates with alignment gains only under coherent annotation signals, and that activation volume alone does not imply steerability. Together, these results demonstrate that demographic prompting is not a monolithic intervention: its utility is highly context-dependent, shaped by attribute signal quality, task characteristics, and model architecture.

Disclaimer: Samples in this paper may be harmful and cause discomfort.

Multimodal large language models (MLLMs) enable multimodal understanding but inherit toxic signals from weakly curated pretraining corpora, leading to explicitly toxic outputs, especially under adversarial triggers that late, opaque training-free detoxification methods struggle to handle. We propose SGM, a white-box neuron-level multimodal intervention that acts like safety glasses for toxic neurons: it recalibrates a set of toxicity-associated neurons via expertise-weighted soft suppression, neutralizing harmful cross-modal activations without any parameter updates. We establish MM-TOXIC-QA, a multimodal toxicity data framework, and compare SGM with existing detoxification techniques. Experiments on open-source MLLMs show that SGM mitigates explicit toxicity in standard and adversarial conditions, cutting average harmful rates from 45.0% to 4.5% while preserving fluency and multimodal reasoning. SGM is extensible, and its combined defenses, denoted as SGM*, integrate with existing detoxification methods for stronger safety performance, providing an interpretable, low-cost solution for toxicity-controlled multimodal generation.

Large Language Model (LLM) evaluation is frequently challenged by prompt sensitivity, which limits the reliability of benchmarks relying on a single template. To address this, we systematically investigate a multi-prompt evaluation framework that aggregates performance across diverse templates, leveraging a Quality-Diversity (QD) prompt selection mechanism to ensure both variety and fidelity. Through extensive empirical analysis, we uncover a novel Evaluation Scaling Law: evaluation reliability scales consistently with template diversity. Specifically, our experiments demonstrate that evaluation variance diminishes as the scale of prompts increases, while ranking monotonicity exhibits a log-linear improvement relative to the number of templates. By decoupling a model’s intrinsic capability from template-specific fluctuations, our framework facilitates more reliable assessments of LLMs, thereby providing a solid foundation for benchmarking that enables more effective monitoring of training dynamics and rigorous comparisons among state-of-the-art models.

Text embeddings are standard for semantic similarity tasks, yet their evaluation remains an open challenge. Current benchmarks are static, cover only a limited set of languages, are often domain-specific, susceptible to overfitting, and poorly representative of low-resource languages. To address these limitations, we introduce ALEE, a framework that extends Sentence Smith (Li et al., 2025) to the cross-lingual and paragraph level. ALEE uses Abstract Meaning Representations (AMR) to generate English minimal pairs with controlled, fine-grained semantic shifts, which are paired with translations in target languages. This approach enables targeted diagnostics for models in any language with English parallel data. We conduct a large-scale empirical study across a diverse set of embedding models and 275+ languages spanning three parallel datasets. On ALEE, performance varies substantially across languages, text lengths, and linguistic phenomena, exposing persistent gaps in cross-lingual semantic representation that track language prevalence in training resources and subword tokenization. We release ALEE at https://github.com/Andrian0s/any-lang-embed-eval

Vision-Language Models often struggle with complex visual reasoning due to the visual information loss in textual CoT. Existing methods either add the cost of tool calls or rely on localized patch-based embeddings that are insufficient to extract semantics in multi-step reasoning. We propose "Decompose, Look, and Reason" (DLR), a reinforced latent reasoning framework that dynamically decomposes queries into textual premises, extracts premise-conditioned continuous visual latents, and deduces answers through grounded rationales. We introduce a three-stage training pipeline and propose a novel Spherical Gaussian Latent Policy, to enable effective exploration in the latent space. Extensive experiments on vision-centric benchmarks show that DLR consistently outperforms strong baselines, including text-only, interleaved multimodal CoT, and latent reasoning methods, while providing superior stepwise interpretability.

Large Language Model (LLM)-based search agents have demonstrated superior performance in knowledge-intensive tasks through interleaved reasoning and search interactions. Process Reward Models (PRMs) can enhance search agent performance by guiding step selection during inference. However, existing PRMs for search agents face a critical bottleneck: assessment unreliability in their training data. This issue stems from an over-reliance on single-source supervision, forcing a compromise between global solvability and local validity. We propose CoPRM (Collaborative Process Reward Model), a process reward model trained on high-quality preference data generated through our Collaborative Calibration methodology. This approach synthesizes immediate quality assessments with outcome-informed evaluations through a dual-strategy mechanism: strictly enforcing consensus filtering to eliminate unreliable signals, and employing an outcome-informed gating mechanism to resolve heterogeneous cross-type trade-offs. Trained on 10k preference pairs, CoPRM provides universal test-time guidance, enabling diverse search agents to achieve consistent performance improvements across four multi-hop QA benchmarks, demonstrating substantial gains, robust cross-domain generalization, and superior efficiency.

Humor preferences vary widely across individuals and cultures, complicating the evaluation of humor using large language models (LLMs). In this study, we model heterogeneity in humor preferences in Oogiri, a Japanese creative response game, by clustering users with voting logs and estimating cluster-specific weights over interpretable preference factors using Bradley-Terry-Luce models. We elicit preference judgments from LLMs by prompting them to select the funnier response and found that user clusters exhibit distinct preference patterns and that the LLM results can resemble those of particular clusters. Finally, we demonstrate that, by persona prompting, LLM preferences can be directed toward a specific cluster. The scripts for data collection and analysis are publicly available to support reproducibility.

Scientific papers advance claims that later work supports, extends, or sometimes refutes. Yet existing methods for citation and claim analysis capture only fragments of this dialogue. In this work, we make these interactions explicit at the level of scientific claims. We introduce ClaimFlow, a claim-centric view of the NLP literature, built from 1,617 ACL Anthology papers (1979-2025) that are manually annotated with 5,689 claims and 4,871 cross-paper claim relations, indicating whether a citing paper supports, extends, qualifies, refutes, or references a cited claim as background. Building on ClaimFlow, we define a new task – Claim Relation Classification – which requires models to infer the scientific stance toward a cited claim from the text and citation context. Evaluating neural models and large language models on this task, we report baseline performance of 0.81 macro-F1, suggesting that the task is tractable while leaving room for improvement. We then scale this framework to ~13k NLP papers to study claim evolution across decades of NLP research. We show that 63.5% of claims are never reused; only 11.1% are ever challenged. Widely propagated claims are more often reshaped through qualification and extension than supported or refuted. Overall, ClaimFlow offers a lens for examining how ideas shift and mature within NLP.

While prior red-teaming efforts have focused on eliciting harmful text outputs from large language models (LLMs), such approaches fail to capture agent-specific vulnerabilities that emerge through multi-step tool execution, particularly in rapidly growing ecosystems such as the Model Context Protocol (MCP). To address this gap, we propose a trajectory-aware evolutionary search method, T-MAP, which leverages execution trajectories to guide the discovery of adversarial prompts. Our approach enables the automatic generation of attacks that not only bypass safety guardrails but also reliably realize harmful objectives through actual tool interactions. Empirical evaluations across diverse MCP environments demonstrate that T-MAP substantially outperforms baselines in attack realization rate (ARR) and remains effective against frontier models, including GPT-5.2, Gemini-3-Pro, Qwen3.5, and GLM-5, thereby revealing previously underexplored vulnerabilities in autonomous LLM agents. Code is available at https://github.com/pwnhyo/T-MAP.

Instruction granularity is an important yet poorly controlled variable in language-guided embodied AI. Existing benchmarks typically pair each task with a single static instruction, making it difficult to study how agent behavior changes when the same task is described at different levels of detail. We introduce Mini-BEHAVIOR-Gran, a new benchmark for controlled studies of instruction granularity that extends Mini-BEHAVIOR with multiple instruction variants per task, ranging from high-level goal descriptions to step-by-step guidance. Using this benchmark, we compare four candidate metrics for cross-task granularity quantification: token count, entity count, action-verb count, and planning-width, and find that width correlates most consistently with agent performance. Using width to organize training and evaluation further reveals a non-monotonic U-shaped relationship between instruction granularity and performance, with peaks at both fine and coarse extremes. Further analysis suggests that the coarse-granularity performance rebound is associated with shallow grounding, where agents learn vision-dominant policies.

Transformers underlie almost all state-of-the-art language models in computational linguistics, yet their cognitive adequacy as models of human sentence processing remains disputed. In this work, we use a surprisal-based linking mechanism to systematically evaluate eleven autoregressive transformers of varying sizes and architectures on a more comprehensive set of English agreement attraction configurations than prior work. Our experiments yield mixed results: While transformer predictions generally align with human reading time data for prepositional phrase configurations, performance degrades significantly on object-extracted relative clause configurations. In the latter case, predictions also diverge markedly across models, and no model successfully replicates the asymmetric interference patterns observed in humans. We conclude that current transformer models do not explain human morphosyntactic processing, and that evaluations of transformers as cognitive models must adopt rigorous, comprehensive experimental designs to avoid spurious generalizations from isolated syntactic configurations or individual models.

Vision-Language Model (VLM) based retrievers have advanced visual document retrieval (VDR) to impressive quality. They require the same multi-billion parameter encoder for both document indexing and query encoding, incurring high latency and GPU dependence even for plain-text queries. We observe that this design is unnecessarily symmetric: documents are visually complex and demand strong visual understanding, whereas queries are just short text strings. NanoVDR exploits this query–document asymmetry by decoupling the two encoding paths: a frozen 2B VLM teacher indexes documents offline, while a distilled text-only student as small as 69M parameters encodes queries at inference. The key design choice is the distillation objective. Through systematic comparison of six objectives across three backbones and 22 ViDoRe benchmark datasets, we find that pointwise cosine alignment on query text consistently outperforms ranking-based and contrastive alternatives, while requiring only pre-cached teacher query embeddings and no document processing during training. Furthermore, we identify cross-lingual transfer as the primary performance bottleneck, and resolve it cheaply by augmenting training data with machine-translated queries. The resulting NanoVDR-S-Multi (DistilBERT, 69M) retains 95.1% of teacher quality and outperforms DSE-Qwen2 (2B) on v2 and v3 with 32× fewer parameters and 50× lower CPU query latency, at a total training cost under 13 GPU-hours.

Synthetic explanatory dialogues hold promise for applications from tutoring systems to model rationalization, yet their quality is often deemed inferior to human-authored conversations. This study uses dialogue completion as a lens to investigate the nuanced differences between human-spoken, human-written, and synthetic explanatory dialogue. By evaluating LLM-generated turns against original turns using the FED and IXQuisite test suites, we uncover a significant bias in the FED metric towards written language, which may penalize the interactivity of spoken dialogue. We find that three popular LLMs (Mistral, Llama, Claude) replicate this pattern, excelling in fluency, but struggling with conversational engagement, particularly on spoken dialogue transcripts. More importantly, we demonstrate that these models can significantly improve their performance through a zero-shot, iterative self-refinement method. By providing feedback based on their specific weaknesses, we show that LLMs can iteratively adapt their outputs to better resemble the interactive qualities of human dialogue.

Large Language Model (LLM) agents increasingly act through external tools, making their safety contingent on tool-call workflows rather than text generation alone. While recent benchmarks evaluate agents across diverse environments and risk categories, a fundamental question remains unanswered: how complete are existing test suites, and what unsafe interaction patterns persist even after an agent passes the benchmark? We propose SafeAudit, a meta-audit framework that addresses this gap through two contributions. First, an LLM-based enumerator that systematically generates test cases by enumerating valid tool-call workflows and diverse user scenarios. Second, we introduce rule-resistance, a non-semantic, quantitative metric that distills compact safety rules from existing benchmarks and identifies unsafe interaction patterns that remain uncovered under those rules. Across 3 benchmarks and 12 environments, SafeAudit uncovers more than 20% residual unsafe behaviors that existing benchmarks fail to expose, with coverage growing monotonically as the testing budget increases. Our results highlight significant completeness gaps in current safety evaluation and motivate meta-auditing as a necessary complement to benchmark-based agent safety testing. Examples and implementations are available at https://github.com/XuanChen-xc/SafeAudit.

Simulated environments play an essential role in embodied AI, functionally analogous to test cases in software engineering. However, existing environment generation methods often emphasize visual realism (e.g., object diversity and layout coherence), overlooking a crucial aspect: logical diversity from the testing perspective. This limits the comprehensive evaluation of embodied agent adaptability and planning robustness across distinct simulated environments. To bridge this gap, we propose LogicEnvGen, a novel method driven by Large Language Models (LLMs) that adopts a top-down paradigm to generate logically diverse simulated environments as test cases for agents. Given an agent task, LogicEnvGen first analyzes its execution logic to construct decision-tree-structured behavior plans and then synthesizes a set of logical trajectories. Subsequently, it adopts a heuristic algorithm to refine the trajectory set, reducing redundant simulation. For each logical trajectory, which represents a potential task situation, LogicEnvGen correspondingly instantiates a concrete simulated environment. Furthermore, we introduce LogicEnvEval, a novel benchmark for simulated environment generation, with four quantitative metrics. Experimental results verify the lack of logical diversity in baselines and demonstrate that LogicEnvGen achieves 1.08-2.67x greater diversity, significantly improving the performance in revealing agent faults by 3.34%–72.00%.

Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model’s safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates while preserving the utility of many models across diverse architectures. Furthermore, we show that ROSI can also re-align “uncensored” models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results show that targeted, interpretable weight steering is a cheap and effective mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.

Safety evaluation of large language models (LLMs) increasingly relies on LLM-as-a-judge pipelines, but strong judges can still be costly to deploy at scale. We present a multi-agent judge that uses small language models (SLMs) in structured debates among critic, defender, and judge agents. To evaluate these systems, we introduce ForesightSafety Human-Annotated JailBench (FS-HAJailBench), a human annotated jailbreak benchmark with 11,100 labeled interactions spanning diverse attack methods and target models. FS-HAJailBench provides fine-grained ground truth for measuring both target-model robustness and judge reliability. On this benchmark, our SLM-based framework improves over matched small-model baselines and prior multi-agent judges, while approaching GPT-4o-level agreement with lower estimated unit cost. Ablation results show that three debate rounds provide the best accuracy-efficiency trade-off among the current Qwen3-14B variants. These results indicate that structured debate is a practical way to improve scalable LLM safety evaluation.

Test-time training (TTT) adapts large language models (LLMs) during inference using only unlabeled test inputs. Existing methods, however, face two major bottlenecks on hard reasoning tasks: (1) lack of learnable samples, as self-generated pseudo-labels on difficult questions are often noisy and yield unstable rewards; and (2) inefficient exploration, as performance gains depend on repeatedly sampling many rollouts without explicit diagnosis of why previous attempts fail. We propose TTSR (Test-Time Self-Reflection), a self-evolving framework based on a reflect-then-synthesize paradigm. A single pretrained model alternates between a Student role and a Teacher role: the Student solves test questions and updates, while the Teacher analyzes failed trajectories and synthesizes targeted variant questions closer to the Student’s capability frontier. TTSR further maintains a cross-iteration weakness memory and compiles persistent weaknesses into a lightweight strategy note prepended to subsequent Student inputs, so diagnostic knowledge can guide exploration and gradually fade as weaknesses are resolved. Experiments on challenging mathematical reasoning benchmarks show consistent test-time improvements, strong cross-backbone generalization, and transfer to general-domain reasoning tasks.

Large language models (LLMs) now play a central role in code generation, yet they continue to hallucinate, frequently inventing non-existent libraries. Such library hallucinations are not just benign errors: they can mislead developers, break builds, and expose systems to supply chain threats such as slopsquatting. Despite growing awareness of these risks, there is limited understanding of how library hallucinations manifest under realistic usage conditions. To fill this gap, we present the first systematic study of how user-level prompt variations influence library hallucinations in LLM-generated code. Across seven diverse LLMs, we analyse library name hallucinations (invalid imports) and library member hallucinations (invalid calls from valid libraries), examining the effects of realistic developer language and controlled user mistakes, including misspellings and fabricated libraries or members. Our findings expose systemic vulnerabilities: one-character misspellings trigger hallucinations in up to 26% of tasks; fabricated library names are accepted in up to 99%; and time-based prompts induce hallucinations in up to 84%. Grounded in the highest-risk prompts identified in our study, we introduce LibHalluBench, a benchmark that enables a systematic and reproducible evaluation of these library hallucinations. Our findings underscore the fragility of LLMs to natural prompt variation and highlight the urgent need for safeguards against library-related hallucinations and their downstream risks.

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works–specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit. Freezing all attention scores during steering drops performance by only 8.83% across three model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

Tree-structured speculative decoding accelerates large language model inference by drafting multiple candidate tokens in parallel and verifying them against a target model. Dynamic tree expansion strategies further improve efficiency by using draft model confidence to prune unlikely branches and stop expansion early. However, we find that existing draft models exhibit severe overconfidence: they assign high probabilities to tokens that will be rejected, causing unreliable early stopping and wasted computation on incorrect branches.

We propose Calibrated Tree-Structured Speculative Decoding (CATS-SD), a training framework for tree-based confidence calibration. Our approach calibrates draft model confidence at two complementary levels: (1) token-level calibration that aligns draft-target token ranking, and (2) path-level calibration that supervises confidence over tree-structured decoding paths. We also introduce a calibration metric using Pearson correlation to quantify confidence-acceptance alignment. On Qwen2.5-3B-Instruct, CATS-SD achieves up to 11.3% higher acceptance length and 4.0% better confidence-acceptance correlation compared to the baseline, with end-to-end speedup improving from 2.51× to 2.71× on GSM8K. We validate these results across mathematical reasoning (GSM8K), multi-turn conversation (MT-bench), and code generation (HumanEval).

Large language models are effective at code completion when the required evidence is local, but repository-level fill-in-the-middle completion remains challenging when it depends on code and structural relationships across files. We present ACAR, a framework that combines lexical and AST/LSP-based structural retrieval through line-level reciprocal rank fusion, then applies syntax-aware post-processing to align generated completions with the surrounding code. We evaluate ACAR on RepoEval, ReccEval, and CrossCodeEval across Python, Java, TypeScript, and C#, covering more than 3,000 real-world repositories. In the main Qwen2.5-Coder experiments, ACAR outperforms strong retrieval baselines in most settings, reaching 42.98 EM on ReccEval and 32.86 on CrossCodeEval at 3B, versus 36.25 and 28.40 for the strongest retrieval baseline, respectively. With the largest tested Qwen and DeepSeek backbones, ACAR improves EM over their corresponding no-retrieval baselines across all five evaluation settings. Ablations show that lexical and structural evidence are complementary, while removing syntax-aware refinement causes the largest drop in edit similarity. These gains come with 87–95 ms of non-decoding overhead under warm-cache conditions across the three benchmarks, supporting latency-sensitive code completion. To support reproducibility, we make our implementation and evaluation scripts publicly available at https://github.com/longstnguyen/ACAR.

Large language models (LLMs) are increasingly consulted as sources of historical information, yet their handling of contested events and politically charged narratives remains understudied. We introduce HistoricalMisinfo, a dataset of 500 contested events from 45 countries, each paired with a factual and a documented revisionist reference narrative, instantiated across 11 prompt scenarios reflecting common real-world usage. Using an LLM-as-a-judge protocol, we evaluate multiple model architectures under neutral prompts and robustness prompts that explicitly request revisionist content. Under neutral prompts, models generally align with factual references, though alignment varies by scenario and model. Robustness prompting reveals a consistent and sharp increase in revisionism scores across all models, indicating limited resistance to user-driven revisionist framing. HistoricalMisinfo offers a practical benchmark for evaluating robustness to historical revisionism in AI systems.

Latent reasoning models (LRMs) replace explicit chain-of-thought with continuous thoughts. Recent work treats observable latent-state patterns, such as BFS-like frontiers and decodable arithmetic computation, as evidence for internal reasoning mechanisms. Evaluating two LRMs (Coconut and CODI) against controls lacking the proposed recurrence or curriculum, we find these patterns also appear in the controls and do not consistently affect behavior. Causal interventions reveal that latent-thought utilization is not binary but graded, scaling with a thought’s causal effect on model behavior. Geometric analyses reveal this effect concentrates in low-dimensional subspaces whose step-wise geometry grows more structured as their behavioral influence increases. Latent thoughts should therefore be treated as hidden computation, not hidden explanation: decodability, attention, or static structure alone cannot establish LRM reasoning mechanisms; that requires matched controls and causal tests.

Large Language Models (LLMs) often generate specious hallucinations, creating an urgent need for post-hoc factuality verification against reliable evidence. Existing verification methods incur substantial computational overhead due to joint evidence–claim encoding. While bi-encoders offer a highly efficient alternative, they struggle with asymmetric entailment and are severely vulnerable to contextual noise. To address these challenges, we propose Gated Residual Discrepancy (GRD), a highly efficient bi-encoder framework for factuality verification. To explicitly capture asymmetric entailment, GRD pioneers a difference-dominant paradigm, extracting dual-view vector differences. To mitigate the vulnerability to contextual noise, GRD employs an adaptive gating mechanism that dynamically filters contextual noise, preventing large semantic backgrounds from masking critical errors, thereby enabling highly sensitive verification over broad evidence contexts. Experiments across various benchmarks show GRD achieves 79.99% average Macro-F1, trailing the strongest cross-encoder baseline by only 1.15 points while delivering a 23.55× speedup with only a fraction of the computational FLOPs.

Rising demand for mental health support has increased interest in using Large Language Models (LLMs) for counseling, but adapting them to this safety-critical domain is hindered by limited real-world data due to privacy constraints. Synthetic datasets provide a promising alternative, but existing approaches often rely on unstructured or semi-structured text inputs and overlook structural dependencies between a client’s cognitive, emotional, and behavioral states, leading to psychologically inconsistent and less realistic interactions. We introduce Graph2Counsel, a framework for generating synthetic counseling sessions grounded in Client Psychological Graphs (CPGs) that encode relationships among clients’ thoughts, emotions, and behaviors. Graph2Counsel uses a structured prompting pipeline guided by counselor strategies and CPG, and explores prompting strategies including CoT and Multi-Agent Feedback. It produces 760 sessions from 76 CPGs across diverse client profiles. In expert evaluation, our dataset outperforms prior datasets on specificity, counselor competence, authenticity, conversational flow, and safety (Krippendorff’s 𝛼 = 0.70). Fine-tuning an open-source model on this dataset improves performance on several metrics on CounselingBench and CounselBench, while matching baselines on others. We further introduce Therapy-Eval, a multi-turn evaluation framework, and demonstrate the effectiveness of our fine-tuned model in realistic therapeutic conversations. We make our code, data and fine-tuned model public.

Automated fact-checking (AFC) systems are susceptible to adversarial attacks, enabling false claims to evade detection. Existing adversarial frameworks typically rely on injecting noise or altering semantics, yet no existing framework exploits the adversarial potential of persuasion techniques against AFC systems, which are widely used in disinformation campaigns to manipulate audiences. In this paper, we introduce a novel class of persuasive adversarial attacks on AFCs by employing an LLM to rephrase claims using persuasion techniques. Considering 15 techniques grouped into 5 categories, we study the effects of persuasion on both claim verification and evidence retrieval using a decoupled evaluation strategy. Experiments on the FEVER and FEVEROUS benchmarks show that persuasion attacks can substantially degrade both verification performance and evidence retrieval. Our analysis identifies persuasion techniques as a potent class of adversarial attacks, highlighting the need for more robust AFC systems.

Robust voice agents require exposure to the full diversity of how people interact through speech. However, obtaining enough spoken interactions is prohibitively expensive. Building spoken user simulators that address this requires large-scale spoken task-oriented dialogue (TOD) data encompassing spoken user behaviors, yet existing datasets are limited in scale and domain coverage, with no systematic pipeline for augmenting them. To address this, we introduce SpokenTOD, a spoken TOD dataset of 52,390 dialogues and 1,034 hours of speech augmented with four spoken user behaviors—cross-turn slots, barge-in, disfluency, and emotional prosody—across diverse speakers and domains. Building on SpokenTOD, we present SpokenUS, a spoken user simulator grounded in TOD that decides when to speak through a dedicated turn-taking head. SpokenUS achieves comparable goal coverage to much larger models while substantially outperforming all baselines in human MOS, disclosing slot values gradually across the dialogue as humans do rather than front-loading them. Further analysis confirms that SpokenUS’s spoken behaviors pose meaningful challenges to voice agents, making it a practical tool for evaluating more robust spoken dialogue systems. Our code is available at https://github.com/holi-lab/SpokenUS.

Medical VLMs frequently hallucinate clinically unsupported findings. We identify a stability gap: visually grounded tokens maintain relatively stable logits under semantics-preserving image perturbations, whereas visually misgrounded tokens exhibit substantially larger shifts. We introduce LENS (Logit-ENhanced Stability), which exploits this signal at both training and inference. During training, Semantics-Preserving Preference Optimization (SPPO) scores preference pairs using clinical faithfulness and cross-view stability; at inference, stability-aware contrastive decoding penalizes tokens whose logits diverge across original and counterfactual views. Constrained counterfactuals use attribution-guided masking and feature-space proximity guards to reduce disruption of lesion- and device-level evidence. Across radiology VQA, chest X-ray reporting, and gastrointestinal endoscopy, LENS consistently outperforms strong decoding-time and preference-optimization baselines. On MIMIC-CXR, entity-level RadGraph recall improves from 21.0 to 22.8 over PCCD, while expert evaluation shows a hallucination-rate reduction from 17.0% to 8.5%.

Realistic text-to-SQL workflows often require joining multiple tables. As a result, accurately retrieving the relevant set of tables becomes a key bottleneck for end-to-end performance. We study an open-book setting where queries must be answered over large, heterogeneous table collections pooled from many sources, without clean scoping signals such as database identifiers. Here, dense retrieval (DR) achieves high recall but returns many distractors, while join-aware alternatives often rely on extra assumptions and/or incur high inference overhead. We propose CORE-T, a scalable, training-free framework that enriches tables with LLM-generated purpose metadata and pre-computes a lightweight table-compatibility cache. At inference time, DR returns top-K candidates; a single LLM call selects a coherent, joinable subset, and a two-step additive adjustment stage restores strongly compatible tables. Across Bird, Spider, MMQA, and Beaver, CORE-T improves over DR by up to 22.7 points in table-selection F1 while returning up to 40% fewer tables, and by up to 24.4 points in multi-table execution accuracy, and uses 1.64–4.20× fewer total selection tokens than LLM-intensive baselines.

Training speech recognizers with unpaired speech and text – known as unsupervised speech recognition (UASR) – is a crucial step toward extending ASR to low-resource languages in the long-tail distribution and enabling multimodal learning from non-parallel data. However, existing approaches based on phones often rely on costly resources such as grapheme-to-phoneme converters (G2Ps) and struggle to generalize to languages with ambiguous phoneme boundaries due to training instability. In this paper, we address both challenges by introducing a syllable-level UASR framework based on masked language modeling, which avoids the need for G2P and the instability of GAN-based methods. Our approach achieves up to a 40% relative reduction in character error rate (CER) on LibriSpeech and generalizes effectively to low-resource languages that have remained particularly difficult for prior methods. Code will be released upon acceptance.

Reasoning in Large Language Models (LLMs) is primarily conducted in a discrete token space. Although recent approaches employing continuous concept tokens facilitate abstract reasoning, they introduce significant stability challenges. The accumulation of Out-of-Distribution (OOD) states frequently causes manifold deviation, while the absence of clear boundaries leads to indeterminate termination, often referred to as the anchoring problem. To address these limitations, we propose HybridThink. We formulate mode selection in LLM reasoning as a sequential decision process. A lightweight gating policy is trained with RL to decide at each step whether to propagate a continuous concept token or generate a discrete text token to anchor the semantic trajectory.This learned interleaving limits manifold deviation and resolves the fragility of static rules. Furthermore, we ensure stable convergence through a training strategy that incorporates Error-Centric Experience Replay (ECER) and Inductive Exploration Shaping (IES), prioritizing learning from failure modes, specifically drifting and premature termination. Empirical evaluations on mathematics and coding benchmarks demonstrate that HybridThink outperforms standard CoT and Soft Thinking baselines, achieving superior accuracy with reduced token consumption.

Computational linguistics predominantly operates on written text. However, writing represents spoken language only indirectly. Whether phonetic and orthographic representations share fundamental information-theoretic properties is an open question. We present a cross-linguistic analysis of Shannon entropy across modalities using the DoReCo corpus featuring 42 languages of 25 families after preprocessing. By computing continuous entropy trajectories at orthographic word, subword (BPE), and character levels (UTF-8/SAMPA), we reveal a systematic modality convergence: orthographic and phonetic encodings yield consistently small entropy differences across all three granularity levels. This general pattern, however, masks notable cross-linguistic variation–some languages, such as Evenki, converge tightly across all levels, while others show modality gaps at specific granularities–for instance, English at the character level, Bora at the word level, and Sanzhi Dargwa at the subword level. In sum, Shannon entropy tends toward modality-invariance as a general tendency across the Latin-script languages examined, while leaving room for typologically conditioned divergence. Whether this pattern extends to non-alphabetic writing systems remains an open question.

Gender bias in large vision-language models (LVLMs) undermines their fairness and reliability, compromising output trustworthiness. Current mitigation methods rely on training-phase adjustments or post-hoc calibration, but face limitations in dynamic visual bias mitigation. These include inability to capture real-time visual-textual incongruence, dependence on predefined gender bias taxonomies, and degraded cross-modal alignment with emergent bias patterns. To address these challenges, we propose ViD, a causally-inspired framework that analyzes attention mechanisms across five distinct patterns, revealing confounding effects from strong language priors. ViD demonstrates that visual-to-language cross-attention effectively suppresses bias while preserving general reasoning capabilities and text generation quality. ViD incorporates dual mechanisms: backdoor adjustment counters strong language priors, while refined token selection in decoding layers optimizes processing. This enhances model robustness and inference efficiency. Our integrated approach significantly mitigates gender bias across multidimensional social attributes in LVLMs, improving visual grounding and output fairness. Cross-benchmark validation shows ViD reduces gender bias by 14.7% on single-attribute evaluations (FACET) and achieves significant improvements on image captioning tasks (MS COCO), with gender bias score improving from 0.6708 to 0.9978 for LLaVA. Crucially, these improvements require no additional training overhead, making ViD a scalable and practical solution for bias mitigation in LVLMs.

Despite extensive alignment efforts, Large Vision-Language Models (LVLMs) remain vulnerable to jailbreak attacks. To mitigate these risks, existing detection methods are essential, yet they face two major challenges: generalization and accuracy. While learning-based methods trained on specific attacks fail to generalize to unseen attacks, learning-free methods based on hand-crafted heuristics suffer from limited accuracy and reduced efficiency. To address these limitations, we propose Learning to Detect (LoD), a learnable framework that eliminates the need for any attack data or hand-crafted heuristics. LoD operates by first extracting layer-wise safety representations directly from the model’s internal activations using Multi-modal Safety Concept Activation Vectors classifiers, and then converting the high-dimensional representations into a one-dimensional anomaly score for detection via a Safety Pattern Auto-Encoder. Extensive experiments demonstrate that LoD consistently achieves state-of-the-art detection performance (AUROC) across diverse unseen jailbreak attacks on multiple LVLMs, while also significantly improving efficiency. Code is available at https://github.com/ShuangLiangX/Learning-to-Detect.

Peer review plays a central role in the NLP publication process, but is susceptible to various biases. Here, we study language-of-study (LoS) bias: the tendency for reviewers to evaluate a paper differently based on the language(s) it studies, rather than its scientific merit. Despite being explicitly flagged in reviewing guidelines, such biases are poorly understood. Prior work treats such comments as part of broader categories of weak or unconstructive reviews without defining them as a distinct form of bias. We present the first systematic characterization of LoS bias, distinguishing negative and positive forms, and introduce the human-annotated dataset LOBSTER (Language-Of-study Bias in ScienTific pEer Review) and an LLM-based detection pipeline achieving 87.37 macro F1. We analyze 15,645 reviews to estimate how negative and positive biases differ with respect to the LoS, and find that non-English papers face substantially higher bias rates than English-only ones, with negative bias consistently outweighing positive bias. Finally, we identify four subcategories of negative bias, and find that demanding unjustified cross-lingual generalization is the most dominant form. We publicly release all resources to support work on fairer reviewing practices in NLP and beyond.

Large language models (LLMs) are systematically overconfident: they routinely express high certainty on questions they often answer incorrectly. Existing calibration methods either require labeled validation data, degrade under distribution shifts, or incur substantial inference costs. Recent work has shown that LLMs already contain a better-calibrated signal than the one they verbalize: the token probability of "True" when the model is asked "Is this answer correct?" (P(True)) consistently outperforms their stated confidence, a gap that is theoretically grounded as generative error is lower-bounded by roughly twice the corresponding discriminative error. We introduce SECL (SElf-Calibrating Language Models), a test-time training (TTT) pipeline that exploits this gap as self-supervision: adaptation requires no correctness labels or human feedback, with only two hyperparameters set once per model family from a small labeled probe. SECL adapts only when the input distribution shifts, training on just 6–26% of the question stream at lower cost than the baseline it distills from. Across five language models (2B–12B) from four model families and four diverse domains, SECL reduces Expected Calibration Error (ECE) by 56–78%, outperforming its own supervision signal and matching or outperforming recent inference-time methods. SECL is the first method to apply TTT to calibration; seven ablations covering signal quality, gating strategy, weight accumulation, loss design, domain ordering, hyperparameter sensitivity, and layer selection confirm that each component is crucial and robust across configurations. Code: https://github.com/rissalhedna/SECL

Latent reasoning enables reasoning over continuous hidden states rather than explicit tokens, avoiding the language bottleneck and inference overhead of chain-of-thought for medical VQA. However, existing methods suffer from modality collapse, insufficient visual supervision, and train-inference mismatch. Moreover, their opaque latent states offer no interpretability, which is critical in clinical applications. We propose VITAL, a latent-space reasoning framework for medical MLLMs with visual-semantic dual supervision: an auxiliary text decoder reconstructs reasoning chains from latent states, while a visual projector regresses ROI features from a frozen, independent medical vision encoder. Both modules are discarded at inference with zero overhead, yet can be re-attached post-hoc for dual interpretability, providing textual and visual explanations of the reasoning process without sacrificing efficiency. We construct a 61K dataset spanning 9 imaging modalities, exceeding prior medical visual latent reasoning datasets by an order of magnitude. Experiments on 7 benchmarks show that VITAL consistently and substantially outperforms the backbone, all latent reasoning baselines, and medical MLLMs trained on far larger data, achieving state-of-the-art results competitive with trillion-parameter proprietary models.

This paper introduces a novel data augmentation approach designed for argument mining. Argument mining is a task that automatically identifies argumentative structures within texts, typically represented as trees or directed acyclic graphs (DAGs). A major challenge facing current argument mining systems is the limited structural diversity of existing annotated datasets, as manual annotation inevitably captures only a small subset of the argumentative patterns encountered in practice. To overcome this challenge, we propose an innovative data augmentation method. Our approach begins by generating pseudo-argumentative structures through the exploration of various combinations of argument labels. These structures then guide a large language model (LLM) to produce new texts with annotations directly incorporated into the text. We evaluated our method on the Argument Annotated Essay Corpus (AAEC) and the AbstRCT dataset. Experimental results showed that the proposed method outperformed existing augmentation baselines on Component-F1 and Relation-F1 in most settings. Further analysis indicated that the variety in the generated pseudo-argumentative structures plays a crucial role in achieving these improvements.

Despite extensive research on reasoning-oriented LLMs, core cognitive faculties of human intelligence, such as abstract reasoning and generalization, remain underexplored. To address this, we evaluate recent reasoning-oriented LLMs on the Abstraction and Reasoning Corpus (ARC) benchmark, which explicitly demands both faculties. We formulate ARC as a program synthesis task and propose nine candidate solvers. Experimental results show that repeated-sampling planning-aided code generation (RSPC) achieves the highest test accuracy and demonstrates consistent generalization across most LLMs. To further improve performance, we introduce Knowledge Augmentation for Abstract Reasoning (KAAR), which encodes core knowledge priors within an ontology that classifies priors into three hierarchical levels based on their dependencies. KAAR progressively expands LLM reasoning capacity by gradually augmenting priors at each level, and invokes RSPC to generate candidate solutions after each augmentation stage. Empirical results show that KAAR maintains strong generalization and consistently outperforms non-augmented RSPC across all evaluated LLMs, achieving around 5% absolute gains and up to 64.52% relative improvement.

As the key-value(KV) cache size scales with context length, accessing large KV cache each step and substantial GPU memory demand challenge us to deploy LLMs with long contexts.Various sparse attention methods have been proposed and offloading-based KV retrieval preserves entire KV cache in CPU memory and dynamically retrieves most relevant KV pairs for each decoding step, which performs higher quality and effectively reduces GPU memory consumption than other line works. However, exiting KV retrieval performs page-level to reduce estimation overhead, which introduces inaccurate KV selection and significant re- trieval overhead. We propose PRKV, a framework that both-optimizes algorithm and system for page-level KV retrieval with KV offloading. On the algorithm side, PRKV introduces hybrid KV selection that combines both static and dynamic KV selection strategies. On the system side, PRKV employs contiguous memory in- dexing and batched transfer optimizations to improve retrieval efficiency. Exper- iments demonstrate that PRKV improve accuracy across various scenarios and models, delivering up to 6.75× speedup compared to SOTA KV retrieval methods.

Vision–Language–Action (VLA) models achieve strong performance in robotic manipulation by leveraging pre-trained vision–language backbones. However, in downstream robotic settings, they are typically fine-tuned with limited data, leading to overfitting to specific instruction formulations and leaving robustness to paraphrased instructions underexplored. To study this gap, we introduce LIBERO-Para, a controlled benchmark that independently varies action expressions and object references for fine-grained analysis of linguistic generalization. Across seven VLA configurations (0.6B–7.5B), we observe consistent performance degradation of 22–52pp under paraphrasing. This degradation is primarily driven by object-level lexical variation: even simple synonym substitutions cause large drops, indicating reliance on surface-level matching rather than semantic grounding. Moreover, 80–96% of failures arise from planning-level trajectory divergence rather than execution errors, showing that paraphrasing disrupts task identification. Binary success rate treats all paraphrases equally, obscuring whether models perform consistently across difficulty levels or rely on easier cases. To address this, we propose PRIDE, a metric that quantifies paraphrase difficulty using semantic and syntactic factors. Our benchmark and code are available at https://github.com/cau-hai-lab/LIBERO-Para.

Chain-of-Thought (CoT) prompting often improves the reasoning performance of large language models (LLMs), but the internal signal that triggers this behavior remains poorly understood. Leveraging the sparse features captured by Sparse Autoencoders (SAEs), we propose a systematic framework to analyze and intervene on the internal representations of LLMs, identifying a small set of latent features that are linked to reasoning behavior and can be causally tested through targeted intervention. Across multiple model families and reasoning benchmarks, we show that steering a single reasoning-related latent feature can substantially induce reasoning behavior without explicit CoT prompting, achieving accuracy comparable to CoT. We further show that the identified features are not tied to particular wording patterns or verbosity, and confirm their causal role in reasoning through suppression experiments that impair performance even under CoT prompting. These results suggest that CoT prompting activates specific latent features to trigger reasoning, and that targeted intervention on these features offers an alternative pathway to elicit efficient reasoning behavior without explicit CoT prompting.

Medical benchmarks are dominated by single-turn, multiple-choice clinical cases that poorly reflect real consultations. Practically, clinicians elicit evidence interactively, and patient communication varies widely. We introduce MEDROUNDSQA, a multi-turn diagnostic benchmark derived from 1,387 clinical cases across 17 specialties. Each case is converted into a structured 24-slot clinical record and then instantiated as controlled doctor–patient dual-agent dialogues under varying patient personas, with the underlying clinical content held fixed. We further classify cases by difficulty using model-based uncertainty to enable easy-to-hard analysis. Evaluations of fifteen LLM doctor agents show that (i) moving from a single-turn diagnosis on the standardized records to multi-turn consultations causes large degradations of roughly 13–39 points; (ii) more turns reliably improve question relevance, but diagnostic accuracy exhibits diminishing returns and typically plateaus after 6–12 turns; and (iii) patient persona differences can shift diagnosis accuracy by about 7–8 points from the lowest to highest education levels, highlighting equity risks that single-turn benchmarks miss.

While Video Large Language Models (Video-LLMs) have recently demonstrated strong performance, reliably evaluating their fine-grained video understanding remains challenging. Existing benchmarks often rely on question answering or ground-truth caption matching, where models may succeed through superficial cues and incomplete annotations. To this end, we introduce VidOmni-Bench, a benchmark that requires models to verify whether each event in dense video captions is supported by the video. VidOmni-Bench consists of 500 videos spanning five complexity types and diverse durations from 4 seconds to 90 minutes. After collecting videos along these axes, we use diverse Video-LLMs to generate dense captions and obtain human-verified sentence-level labels, where sentences containing incorrect events serve as hard negatives for evaluation. Our experiments on VidOmni-Bench reveal three key findings: (i) Video-LLMs frequently generate hallucinated descriptions in dense video captioning; (ii) they also struggle as verifiers, failing to reliably detect plausible but incorrect event descriptions; and (iii) model weaknesses vary across video complexity and duration, revealing diverse, model-specific bottlenecks in current Video-LLMs.

Event Argument Extraction (EAE) aims to extract arguments for specified events from a text. Previous prompt-based research has mainly focused on designing static role anchors, effectively modeling general semantics but overlooking the geometric rigidity of these representations: (i) they failed to capture the diverse manifold structure of real-world semantic distributions and (ii) they lacked the flexibility to adapt to instance-specific reasoning contexts. To bridge the gap between static parameters and dynamic contexts, we introduce a new framework named LAMPEAE, which reformulates EAE as a dynamic manifold matching problem via adaptive parameter instantiation. Specifically, we employ a neuro-symbolic approach that utilizes a frozen LLM to extract high-order reasoning priors. A lightweight hypernetwork then maps this meta-knowledge into low-rank geometric transformation matrices, which dynamically project original static prompts into instance-specific probes. This mechanism ensures precise semantic alignment within the continuous parameter space. Experimental results on the RAMS, ACE05, and WikiEvent benchmarks show that LAMPEAE establishes new state-of-the-art performance, respectively, effectively validating its superiority in handling heterogeneous semantic transfer.

Tool-augmented multimodal reasoning integrates external tools (e.g., object detection, depth estimation) into multimodal large language models (MLLMs) to address perceptual bottlenecks in complex visual tasks. However, existing approaches rarely verify tool outputs, limiting their ability to detect and recover from tool failures. We propose ReVISE, a framework that equips MLLMs with verification and dynamic error recovery for tool-augmented reasoning. ReVISE introduces (1) a curated training dataset that supervises reflective behaviors, enabling models to validate tool-derived evidence, reformulate queries when visual mismatches arise, and fall back to intrinsic grounding when external tools are unreliable; and (2) a reinforcement learning based targeted rewards that encourage internal reflection and penalize spatial misalignment. Experiments on several benchmarks demonstrate consistent improvements over existing methods, highlighting the importance of error detection and correction in tool-augmented multimodal reasoning.

Real-world data science involves long-horizon workflows that span data wrangling, exploration, modeling, visualization, and validation, and require coordinated use of tools such as notebooks, IDEs, terminals, browsers, and databases within real operating environments. Yet existing benchmarks lack real-computer interaction and do not evaluate whether agents can execute complete end-to-end data-science workflows in realistic computing environments, failing to capture the multi-stage, multi-tool nature of data-science practice. We introduce DSAgentBench, the first benchmark to evaluate whether agents can automate full data-science workflows inside real computer environments. DSAgentBench contains 275 diverse tasks covering the entire data-science life-cycle, reflecting the complexity and tool coordination required in practice. Each task requires grounding decisions in intermediate outputs and coordinated tool use, and includes a deterministic evaluator that verifies analytical correctness, visual outputs, and model performance rather than code-only execution. Our extensive experiments with 15 closed- and open-source models show that even the strongest agent, Claude-4.6-Sonnet, achieves only 56.70% task success, while all open-source agents remain below 1%, frequently failing at tool orchestration, OS grounding, and multi-step reasoning. These results reveal a substantial capability gap between current agentic systems and real data-science workflows, positioning DSAgentBench as a foundation for developing grounded, verifiable, autonomous data-science agents. We release DSAgentBench at redacted.

Unified Multi-Modal Generation (UMMG) models support seamless switching between generation tasks spanning multiple modality combinations within a single conversational interface. However, existing safety benchmarks construct prompts independently per modality, so different modalities cover different harmful intentions, preventing holistic evaluation of UMMG models. We propose an intention-grounded evaluation methodology in which every modality combination is evaluated against the same set of harmful intentions, yielding fair, directly comparable safety scores across modalities. Because scores are aggregated at the intention level — averaging over the same set of intentions rather than over modality-specific prompt sets — each model receives an accurate safety profile not distorted by per-modality differences. The methodology is incrementally extensible: new modality combinations or intentions can be added without rebuilding existing data. Following this methodology, we construct UNISAF-Bench (UNIfied intention-grounded SAFety Benchmark): 1,257 human-verified harmful intentions across 19 policies, yielding 20,112 conversations spanning 4 modality combinations, single-/multi-turn settings, and paired policy-violating/boundary-benign variants. Evaluation of four open-source UMMG models reveals systematic safety gaps: multimodal inputs degrade text-output safety, safety alignment fails to transfer to image generation, and multi-turn deflection bias inflates refusal on both harmful and benign prompts.

Open-domain table QA requires reliable evidence localization beyond candidate retrieval. Compact table fingerprints reduce irrelevant signals during representation learning, but retrieval based on learned representations may still have limited ability to distinguish evidence segments among retrieved candidates. This issue is especially pronounced when tables are partitioned into segments that share similar metadata and schema patterns. We propose Probe-Grounded Reasoning (PGR), a coarse-to-fine framework for content-level evidence localization. PGR generates query-specific probes with metadata cues and cell-level evidence, then reranks candidate segments to improve fine-grained evidence localization. It then derives answers via a decoupled Semantic-Plan-Execute workflow with GRPO-optimized planning and execution. On Open-WikiTable, under a controlled evaluation setting with fixed BM25-retrieved candidates, PGR achieves 85.96% segment-level evidence accuracy at k=5 and reaches 69.0% end-to-end QA accuracy, suggesting that content-level verification supports more reliable evidence grounding for downstream reasoning. Code is available at https://github.com/XFastDataLab/PGR/.

Standard evaluations of Large language models (LLMs) focus on task performance, offering limited insight into whether correct behavior reflects appropriate underlying mechanisms and risking confirmation bias. We introduce a simple, principled interpretability framework based on token-level perplexity to test whether models rely on linguistically relevant cues. By comparing perplexity distributions over minimal sentence pairs differing in one or a few ‘pivotal’ tokens, our method enables precise, hypothesis-driven analysis without relying on unstable feature-attribution techniques. Experiments on controlled linguistic benchmarks with several open-weight LLMs show that, while linguistically important tokens influence model behavior, they never fully explain perplexity shifts, revealing that models rely on heuristics other than the expected linguistic ones.

Instruct-based large language models (LLMs) have been shown to propagate and even amplify gender bias when prompted with contextually constrained instructions (e.g., writing a text from a description or selecting a gendered pronoun). However, little attention has been paid to biases in responses to contextually unconstrained (generic) instructions conveyed by gendered language, particularly masculine generics (MG). MG, found in many gender-marked languages, denote the use of the masculine gender as a supposedly "neutral" reference for mixed-gender groups or for individuals of unknown gender. Yet, psycholinguistic studies demonstrate that MG are not neutral and systematically induce gender bias. This study investigates how both local and proprietary LLMs are MG-biased when responding to generic prompts in French, examining LLMs’ MG bias rates and use of gender-fair language (GFL). We create a 16k+ human noun database from existing lexical resources and evaluate six LLMs on four instruction-response datasets under two conditions: prompts with and without MG. Overall, we find that ≈27.57% of LLMs’ responses to MG-filtered generic instructions are MG-biased (≈78.55% with MG-containing prompts). Moreover, we find that LLMs rarely use GFL spontaneously. These findings highlight the persistence of MG bias in LLM outputs and the models’ limited tendency towards GFL strategies.

Recent reasoning models such as OpenAI-o1 and DeepSeek-R1 have shown strong performance on complex tasks including mathematical reasoning and code generation. However, this performance gain comes with substantially longer output sequences, leading to significantly increased deployment costs. In particular, long outputs require large KV caches, resulting in high memory consumption and severe bandwidth pressure during attention computation. Most existing KV cache optimization methods are designed for long-input, short output scenarios and are ineffective for the long-output setting of reasoning models. Moreover, importance estimation in prior work is computationally expensive and becomes prohibitive when continuous re-evaluation is required during long generation. To address these challenges, we propose LongFlow, a KV cache compression method with an efficient importance estimation metric derived from an intermediate result of attention computation using only the current query. This design introduces negligible computational overhead and requires no auxiliary storage. We further develop a custom kernel that fuses FlashAttention, importance estimation, and token eviction into a single optimized operator, improving system-level efficiency. Experiments show that LongFlow achieves up to an 11.8x throughput improvement with 80% KV cache compression with minimal impact on model accuracy.

Large language models (LLMs) have demonstrated strong capabilities in processing long contexts, enabling them to tackle tasks involving long textual inputs such as multi-turn conversations, legal documents, or retrieved documents in Retrieval-Augmented Generation (RAG) systems. However, despite their ability to handle long sequences, the resulting decoding latency and memory overhead remain substantial, posing challenges for real-world deployment. Recent advances in KV Cache reuse have shown potential to mitigate these costs, but these methods still suffer from notable performance degradation. To address this issue, we conduct an in-depth investigation of recomputation-based reuse methods and observe that the recomputed tokens often fail to align with the context segments most relevant to the question. This misalignment hinders proper updates to the critical contextual representations. Therefore, we propose the Attention-Aware Accurate KV Cache Fusion algorithm (A³), which precomputes and selectively fuses the KV Cache of text chunks based on their relevance to the question, achieving accurate integration with minimal computational overhead. Extensive experiments on various benchmarks and LLMs demonstrate that A³ achieves competitive task performance compared to five baselines while reducing the time-to-first-token (TTFT) by 2×.

Multimodal large language models (MLLMs) are typically trained in multiple stages, with video-based supervised fine-tuning (Video-SFT) serving as a key step for improving visual understanding. Yet its effect on the fine-grained evolution of visual capabilities, particularly the balance between spatial and temporal understanding, remains poorly understood. In this paper, we systematically study how Video-SFT reshapes visual capabilities in MLLMs. Across architectures, parameter scales, and frame sampling settings, we observe a consistent pattern: **Video-SFT reliably improves video performance, but often yields limited gains or even degradation on static image benchmarks**. We further show that this trade-off is closely tied to temporal budget: increasing the number of sampled frames generally improves video performance, but does not reliably improve static image performance. Motivated by this finding, we study an instruction-aware Hybrid-Frame strategy that adaptively allocates frame counts and partially mitigates the image-video trade-off. Our results indicate that **Video-SFT is not a free lunch for MLLMs, and preserving spatial understanding remains a central challenge in joint image-video training**.

Dark humor often relies on subtle cultural nuances and implicit cues that require contextual reasoning to interpret, posing safety challenges that current static benchmarks fail to capture. To address this, we introduce a novel multimodal, multilingual benchmark for detecting and understanding harmful and offensive humor. Our manually curated dataset comprises 3,000 texts and 6,000 images in English and Arabic, along with 1,200 videos spanning English, Arabic, and language-independent (universal) contexts. Unlike standard toxicity datasets, we enforce a strict annotation guideline: distinguishing Safe jokes from Harmful ones, with the latter further classified into Explicit (overt) and Implicit (covert) categories to probe deep reasoning. We systematically evaluate state-of-the-art (SOTA) open- and closed-source models across all modalities. Our findings reveal that closed-source models significantly outperform open-source models, with a notable performance difference between English and Arabic in both, underscoring the critical need for culturally grounded, reasoning-aware safety alignment. The code and dataset are available at https://github.com/insait-institute/harm-or-humor.

Content warning: This paper contains offensive text and meme imagery, video clips, audio, and other audiovisual material that may include sexual, violent, discriminatory, religious, disability-related, or historically sensitive material. Reader discretion is advised.

Key-value (KV) caching for shared prefixes is essential for high-throughput large language model (LLM) serving, but it faces critical challenges in emerging diffusion language models (DLMs). In DLMs, bidirectional attention means that updating any token dynamically alters the entire context and its corresponding KVs. Thus, existing caching techniques developed for LLMs, which assume that KVs remain invariant once computed, corrupt the shared prefix KVs. Our experiments show that applying these techniques to DLMs causes model accuracy to collapse to near zero.

To unlock high-throughput DLM serving, we propose bidirectional prefix caching, BiCache, the first KV caching technique for shared prefixes in DLMs. BiCache is designed based on key observations from our comprehensive analysis: shared prefix KVs remain stable and reusable in shallow layers, while the depth of shallow layers depends on the fraction of shared prefix tokens in each request. Thus, BiCache dynamically identifies a safe layer depth for reusing shared prefix KVs and eliminates redundant computation. Evaluations demonstrate that BiCache significantly improves serving throughput by 36.3%–98.3% compared to existing techniques without accuracy collapse (only 0–1.8% difference).

Tokenization forms the basis of multilingual language models, yet existing methods often limit cross-lingual transfer by mapping semantically equivalent words to different 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. This problem is even more pronounced in low-resource languages, where shared representations could offer the greatest benefit. We introduce parallel tokenizers, a new framework that first trains tokenizers monolingually and then aligns their vocabularies exhaustively using bilingual dictionaries or word-to-word translation. 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.

Retrieval-Augmented Generation (RAG) has established itself as an innovative technology enabling Large Language Models (LLMs) to answer difficult questions beyond their internal knowledge. Graph-based RAG, which utilizes a triplet-based graph as a database, offers the advantage of structuring relationships between entities through graph connectivity, thereby capturing connections within the corpus that traditional RAG cannot express. However, this paper points out that the process of converting a corpus into a triplet-based graph using LLMs may be unsuitable for RAG due to the inability to represent hierarchical information and structural rigidity. To address this, we propose a novel database structure called the TRA(Title-Relations-Attributes) structure. Through a series of experiments, we highlight the benefits of indexing TRA Structure as a fact to utilize it as a retrieval-optimized data structure, and demonstrate its powerful multi-hop reasoning performance when combined with query decomposition, while also covering key insights in multi-hop reasoning.

Hybrid post-training combining supervised fine-tuning (SFT) and reinforcement learning (RL) is the standard paradigm for aligning large language models, yet fixed mixing schedules cannot adapt when the relative noise of the two signals evolves. We derive a noise-aware mixing weight by minimizing an MSE upper bound on the mixed stochastic gradient, yielding a closed-form that balances gradient noise variance and SFT–RL disagreement. Building on the token-wise reweighting of CHORD (CITATION)

Large Language Models (LLMs) achieve excellent performance in natural language reasoning tasks through pre-training on vast unstructured text, enabling them to understand the logic in natural language and generate logic-consistent responses. However, the representational differences between unstructured and structured knowledge make LLMs inherently struggle to maintain logic consistency, leading to *Logic Drift* challenges in structured knowledge reasoning tasks such as Knowledge Graph Question Answering (KGQA). Existing methods address this limitation by designing complex workflows embedded in prompts to guide LLM reasoning. Nevertheless, these approaches only provide input-level guidance and fail to fundamentally address the *Logic Drift* in LLM outputs. Additionally, their inflexible reasoning workflows cannot adapt to different tasks and knowledge graphs. To enhance LLMs’ logic consistency in structured knowledge reasoning, we specifically target the logits output from the autoregressive generation process. We propose the *Logits-to-Logic* framework, which incorporates logits strengthening and logits filtering as core modules to correct logical defects in LLM outputs. Extensive experiments show that our approach significantly improves LLMs’ logic consistency in structured knowledge reasoning and achieves state-of-the-art performance on multiple KGQA benchmarks.

Summarization with a fixed character length is important for tasks such as headline and article preview generation, where text must fit an exact space. Autoregressive language models (LMs) struggle with precise character-length control: training-based methods require task-specific fine-tuning, while zero-shot prompting strategies achieve only approximate compliance. We propose DELTA (Diffusion Expected Length Targeting Algorithm), a training-free inference-time method for masked diffusion LMs that estimates expected character length at each denoising step and dynamically inserts or removes masked positions to steer toward a target length. DELTA requires no fine-tuning and is designed for off-the-shelf masked diffusion LMs. Across three summarization benchmarks, it reduces character-length errors by an order of magnitude compared to baselines while maintaining competitive summary quality in both automatic and human evaluation.

As Large Language Models (LLMs) are increasingly deployed in multilingual contexts, evaluating their fairness across languages is essential. We investigate occupational gender bias in 16 LLMs across eight Indic languages using a controlled coreference resolution benchmark that employs transliterated English occupational loanwords to minimize grammatical gender cues. In addition to this generative evaluation, we compute intrinsic bias metrics (WEAT and DirectBias) from the models’ embedding spaces. Across models, we observe systematic disparities between pro-stereotype and anti-stereotype conditions. These gaps are largest in high-accuracy models, whereas models lacking baseline competence exhibit smaller gaps driven by general prediction uncertainty. Error analysis shows a consistent asymmetry, with higher error rates when the correct referent is female. However, intrinsic embedding bias metrics show no statistically significant correlation with these generative disparities (p \gt 0.05). These results suggest that strong linguistic competence does not necessarily correspond to stereotype-neutral generation and that intrinsic embedding metrics alone may not reliably predict bias in multilingual language model behavior.

Large Language Models (LLMs) achieve strong performance across many tasks but suffer from high inference latency due to autoregressive decoding. The issue is exacerbated in Large Reasoning Models (LRMs), which generate lengthy chains of thought. While speculative decoding accelerates inference by drafting and verifying multiple tokens in parallel, existing methods operate at the token level and ignore semantic equivalence (i.e., different token sequences expressing the same meaning), leading to inefficient rejections. We propose SemanticSpec, a semantic-aware speculative decoding framework that verifies entire semantic sequences instead of tokens. SemanticSpec introduces a semantic probability estimation mechanism that probes the model’s internal hidden states to assess the likelihood of generating sequences with specific meanings. Experiments on four benchmarks show that SemanticSpec achieves up to 2.7× speedup on DeepSeekR1-32B and 2.1× on QwQ-32B, consistently outperforming token-level and sequence-level baselines in both efficiency and effectiveness.

As conversational agents accumulate experience collaborating with users, adapting to user preferences is essential for fostering long-term relationships and improving collaboration quality over time. We introduce MultiSessionCollab, a benchmark that evaluates how well agents can learn user preferences and leverage them to improve collaboration quality throughout multiple sessions. To develop agents that succeed in this setting, we present long-term collaborative agents equipped with a memory that is specifically designed to learn user preferences across sessions and improve interactions. Moreover, we demonstrate that learning signals can be derived from user simulator behavior in MultiSessionCollab to train agents to generate more comprehensive reflections and update their memory more effectively. Extensive experiments show that equipping agents with our memory improves collaboration over time, yielding higher task success rates, more efficient interactions, and reduced user effort. Finally, we conduct a human user study that demonstrates that memory helps improve user experience in real-world settings.

Vision-language models (VLMs) can ingest only a limited number of video frames, making frame selection a practical necessity. But do current Video QA benchmarks genuinely require temporal frame selection, or can most questions be answered regardless of which frames are shown? We introduce Frame Selection Sensitivity (FSS), a per-sample diagnostic that measures how much VLM accuracy changes when the most relevant frames are replaced with the least relevant ones. Across six benchmarks and eight VLMs, we find that a large majority of samples are frame-agnostic: only a minority are genuinely sensitive to frame choice. Combining FSS with a Language Independence Score (LIS) reveals that merely 5.5--31% of samples are Temporally Sensitive. We construct TempCore, compact evaluation subsets that isolate these temporal samples from existing benchmarks, and will release code and per-sample annotations upon publication.

Reinforcement learning (RL) for long-horizon agents faces two key challenges. First, parallel linear rollouts tend to converge toward homogeneous reasoning patterns, leading to cognitive-mode collapse. Second, existing methods refine credit assignment along only a single dimension, making it difficult to simultaneously achieve state-aligned comparison and meta-reasoning discrimination. We propose STARK (**S**trategic and **T**actical **A**dvantage **R**einforcement via Confidence Guided **K**eypoint Exploration) to resolve both issues. STARK performs corrective exploration through confidence-driven branching at uncertain decision points and injects reflection into new branches. To empower these keypoint-expanded trajectories, STARK introduces a dual-track strategic and tactical advantage estimator: The strategic track compares reasoning modes under shared environmental observations, while the tactical track evaluates execution quality within each meta-reasoning tag group. Extensive experiments on ALFWorld, ScienceWorld, and WebShop show that STARK consistently outperforms strong baselines across model scales and task domains, and enables a 1.5B model to surpass GPT-5 on the out-of-domain settings.

To effectively enable people to understand new topics, explanations should be tailored to their backgrounds and abilities. Prompting alone has been shown to be insufficient for creating such explanations, and other computational methods are missing so far. Therefore, this paper investigates whether LLMs can be steered to generate explanations that are tailored to a specific group of people. To this end, we propose an approach that first identifies group-specific attributes in terms of explanatory style and knowledge of a specific target group. Building on activation engineering, it then computes attribute-based steering vectors and adds them to the internal activations of an LLM during inference to enable a fine-grained steering. In our experiments, we assess the steering effectiveness in terms of specificity and factuality of the generated explanations. Additionally, we evaluate the explanations in a study with human experts from different target groups. Compared to prompting and state-of-the-art steering baselines, our approach tailors the explanations significantly better to the target group while maintaining the best specificity-factuality balance.

Large Language Models (LLMs) are typically adapted to downstream tasks through Supervised Fine-Tuning (SFT), where Cross Entropy (CE) loss is used to maximize the likelihood of ground-truth tokens. While effective for aligning with observed data and establishing strong domain grounding, SFT alone often limits the model’s ability to explore alternative plausible outputs, potentially narrowing its decision space. Conversely, Reinforcement Learning (RL) enables reward-driven exploration and refinement but relies heavily on a well-aligned initialization to achieve stable and sample-efficient training. In real-world deployments, practitioners frequently grapple with the question: How much SFT is sufficient, and when should RL take over? This work investigates the complementary trade-off between SFT and RL approaches such as GRPO, focusing on how both stages jointly influence model exploration and exploitation. We provide a quantitative analysis across multiple domains to understand how varying the extent of SFT influences a model’s exploratory behavior and downstream RL effectiveness. Our findings aim to guide in determining optimal SFT checkpoints (CKPTs) that balance domain alignment with exploration capacity, thereby maximizing final domain performance while minimizing overfitting after the joint post-training. Our data and code are open source at https://github.com/Shirley-Kokane/sft2rl.

Multiparty tutoring policies must regulate turn-taking and social attention, yet collecting the interaction data needed for reinforcement learning is expensive in educational settings. We introduce an LLM-based user simulator for triadic robot–student dialogue that represents gaze and gesture as inline stage directions. Combined with phase-tagged prompting (which delineates conversational structure) and per-turn persona reinforcement, this design allows the LLMs to generate interactions that preserve key discourse structure and controllable participation dynamics. We validate six simulator variants against a real student–robot corpus (12 student pairs) using metrics for semantic similarity, cue obedience, and persona consistency. While the simulator reaches strong semantic agreement with real transcripts, it also reveals a measurable sim-to-real style gap (in verbosity and grammatical complexity). The simulator enables the creation of an RL environment where the tutor selects factorized gaze and reference actions, learning to interact with simulated students. We compare online training (APPO) to offline learning (DQN). Off-policy evaluation on real logs shows that offline policies trained purely from synthetic data exceed online-trained policies under a doubly-robust estimator, and a verbal-only ablation yields unstable OPE due to heavy-tailed importance weights. These results support a scalable workflow for prototyping socially grounded multiparty tutoring policies while relaying where real-world coverage and OPE reliability remain bottlenecks.

Client-side privacy rewriting is crucial for deploying LLMs in privacy-sensitive domains. However, existing approaches struggle to balance privacy and utility. Full-text methods often distort context, while span-level approaches rely on impractical manual masks or brittle static dictionaries. Attempts to automate localization via prompt-based LLMs prove unreliable, as they suffer from unstable instruction following that leads to privacy leakage and excessive context scrubbing. To address these limitations, we propose DAMPER (Domain-Aware Mask-free Privacy Extraction and Rewriting). DAMPER operationalizes latent privacy semantics into compact Domain Privacy Prototypes via contrastive learning, enabling precise, autonomous span localization. Furthermore, we introduce a Prototype-Guided Preference Alignment, which leverages learned prototypes as semantic anchors to construct preference pairs, optimizing a domain-compliant rewriting policy without human annotations. At inference time, DAMPER integrates a sampling-based Exponential Mechanism to provide rigorous span-level Differential Privacy (DP) guarantees. Extensive experiments demonstrate that DAMPER significantly outperforms existing baselines, achieving a superior privacy-utility trade-off.

Scientific figures compress complex pipelines into a single canvas, but understanding them requires paper-grounded, step-by-step narration aligned with visual highlights—a capability missing from current video generation systems and benchmarks. To address this, we introduce paper-grounded figure-to-video generation: generating narrated, region-grounded walkthrough videos from a figure and its paper. Second, we develop a multi-agent pipeline MINARD, which first generates paper-grounded narration and then ground the figure regions through staged extraction and clustering. Third, we release FigTalk, a benchmark derived from Paper2Video with new sequential and component-level grounding metrics that correlate strongly with human judgments. On FigTalk, MINARD produces human-like narrations grounded in paper evidence, and MINARD-Grounding substantially outperforms existing VLM grounding, animation-based, and diffusion-style video generation methods in both automatic and human evaluation.

While Large Language Models (LLMs) show promise in alleviating mental health resource shortages, existing approaches struggle with complex emotional dynamics and maintaining theory-grounded counseling trajectories. In particular, they fail to adapt interaction rhythms across counseling phases, often resulting in excessive exploration or premature interventions. To address these limitations, we propose AuroraCounsel, an agent-based counseling framework that models professional counseling workflows. The framework integrates an emotion-sensing agent for affect recognition, a reflection agent for structured psychological analysis, and a guidance agent for response generation. Through emotion-driven dynamic routing, AuroraCounsel selectively activates reflective reasoning only when negative emotions are detected, enabling accurate problem tracing while avoiding unnecessary computation. Building on this architecture, we further align counseling behaviors by optimizing the guidance agent with Grouped Relative Policy Optimization (GRPO), allowing it to regulate questioning and supportive strategies across dialogue phases and overcome the limitations of supervised fine-tuning in capturing clinical counseling rhythms, such as those prescribed by REBT. Evaluations using LLM-based evaluators and expert assessments by professional counselors demonstrate consistent improvements over existing approaches in professionalism, empathy, and interaction appropriateness.

Assessing the veracity of a claim made online is a complex and important task with real-world implications. When these claims are directed at communities with limited access to information and the content concerns issues such as healthcare and culture, the consequences intensify, especially in low-resource languages. In this work, we introduce AfrIFact , a dataset that covers all necessary steps for automatic fact-checking (i.e., information retrieval, evidence extraction, and fact checking), covering ten African languages and English. Our evaluation results show that even the best embedding models lack cross-lingual retrieval capabilities, and that cultural and news documents are easier to retrieve than healthcare-domain documents, both in large corpora and in single documents. We show that LLMs lack robust multilingual fact-verification capabilities in African languages, while few-shot prompting improves performance by up to 43% in AfriqueQwen-14B, and task-specific fine-tuning further improves fact-checking accuracy by up to 26%. These findings, along with our release of the AfrIFact dataset, encourage work on low-resource information retrieval, evidence retrieval, and fact checking.

We introduce Evo, a duality latent trajectory model that bridges autoregressive (AR) and diffusion-based language generation within a continuous evolutionary generative framework. Rather than treating AR decoding and diffusion generation as separate paradigms, Evo reconceptualizes text generation as a latent flow: each token is associated with a vector-valued embedding that evolves over a progression variable, indicating its semantic maturity. Low values correspond to confident AR-like refinement, while high values invoke diffusion-style planning, allowing the model to adaptively balance AR and diffusion based on uncertainty. Theoretically, we show that both AR and diffusion models emerge as discretizations of a shared probability flow, and we derive Evo’s training objective from a unified variational ELBO. The model is implemented as a time-conditioned Transformer governed by a shared vector field, trained end-to-end to jointly infer latent codes and their progression times. During decoding, Evo performs efficient, semantics-aware refinement, achieving high-quality outputs without sacrificing speed. Empirically, Evo 8B achieves state-of-the-art or highly competitive results on 15 diverse benchmarks, including reasoning (GSM8K, ARC-C), code generation (HumanEval, MBPP), and general language understanding, while maintaining fast inference speed. Our results demonstrate that Evo delivers a new paradigm for LLM design with strong generation quality, robust symbolic reasoning, and decoding efficiency.

Reinforcement Learning with Verifiable Rewards (RLVR) has become a prominent paradigm to enhance the capabilities (i.e. long-context) of Large Language Models (LLMs). However, it often relies on gold-standard answers or explicit evaluation rubrics provided by powerful teacher models or human experts, which are costly and time-consuming. In this work, we investigate unsupervised approaches to enhance the long-context capabilities of LLMs, eliminating the need for heavy human annotations or teacher models’ supervision. Specifically, we first replace a few paragraphs with special placeholders in a long document. LLMs are then trained through reinforcement learning to reconstruct the long document by correctly identifying and sequencing missing paragraphs from a set of candidate options. This training paradigm enables the model to capture global narrative coherence, significantly boosting long-context performance. We validate the effectiveness of our method on two widely used benchmarks, RULER and LongBench v2. While acquiring noticeable gains on RULER (nearly 10 points), it can also achieve a reasonable improvement on LongBench v2 without any manually curated long-context QA data. Furthermore, we conduct extensive ablation studies to analyze the impact of reward designs, data curation strategies, training schemes, and data scaling effects on model performance. We will release our code, data, and models.

We analyze the generalization capabilities of a RoBERTa-based metaphor detection model on English verbs using the VU Amsterdam Metaphor Corpus. Since RoBERTa encoders are widely-used in metaphor detection, determining whether they rely on lexical memorization or learn transferable patterns is critical. We introduce a controlled lexical hold-out setup where all instances of selected target lemmas are strictly excluded from fine-tuning. We then compare model predictions on these Held-out lemmas against Exposed lemmas (verbs encountered during fine-tuning). While the model performs best on Exposed lemmas, it maintains robust performance on Held-out lemmas. This success does not correlate with verb lemma frequency, suggesting the generalization is likely not driven by frequency bias in pre-training. Further analysis reveals that sentence context alone is sufficient to match full-model performance on Held-out lemmas, whereas static verb-level embeddings are not. Geometrically, we find that the encoder projects Held-out instances into the same task-relevant subspaces as Exposed instances. Together, these results suggest that the model exhibits generalization capabilities, which appear to be driven by "learning the cue" (transferable contextual patterns) rather than "learning the word" (verb-specific memorization).

The ubiquity of multimedia content is reshaping online information spaces, particularly in social media environments. At the same time, search is being rapidly transformed by generative AI, with large language models (LLMs) routinely deployed as intermediaries between users and multimedia content to retrieve and summarize information. Despite their growing influence, the impact of LLM inaccuracies and potential vulnerabilities on multimedia information-seeking tasks remains largely unexplored. We investigate how generative AI affects accuracy, efficiency, and confidence in information retrieval from videos. We conduct an experiment with around 900 participants on 8,000+ video-based information-seeking tasks, comparing behavior across three conditions: (1) access to videos only, (2) access to videos with LLM-based AI assistance, and (3) access to videos with a deceiving AI assistant designed to provide false answers. We find that AI assistance increases accuracy by 3–7% when participants viewed the relevant video segment, and by 27–35% when they did not. Efficiency increases by 10% for short videos and 25% for longer ones. However, participants tend to over-rely on AI outputs, resulting in accuracy drops of up to 32% when interacting with the deceiving AI. Alarmingly, self-reported confidence in answers remains stable across all three conditions. Our findings expose fundamental safety risks in AI-mediated video information retrieval.

Hierarchical Instruction Following (HIF) refers to the problem of prompting large language models with a priority-ordered stack of instructions. Standard methods like RLHF and DPO typically fail in this problem since they mainly optimize for a single objective, failing to explicitly enforce system prompt compliance. Meanwhile, supervised fine-tuning relies on mimicking filtered, compliant data, which fails to establish the priority asymmetry at the algorithmic level. In this paper, we introduce HIPO, a novel alignment framework that formulates HIF as a Constrained Markov Decision Process. HIPO elevates system prompts from mere input context to strict algorithmic boundaries. Using a primal-dual safe reinforcement learning approach, the algorithm dynamically enforces system prompt compliance as an explicit constraint, maximizing user utility strictly within this feasible region. Extensive evaluations across diverse model architectures (e.g., Qwen, Phi, Llama) demonstrate that HIPO significantly improves both system compliance and user utility. Furthermore, mechanistic analysis reveals that this constrained optimization autonomously drives the model to shift its attention toward long-range system tokens, providing a principled foundation for reliable LLM deployment in complex workflows.

Human language is driven by unspoken beliefs and belief updates which are thus of critical importance for successful interactions between large language models (LLMs) and their users. In this paper, we evaluate the ability of LLMs to recognize unspoken beliefs made through implicatures and to recognize their updates through implicature cancellation: the pragmatic phenomenon whereby an utterance’s implied meaning is weakened or negated. We create the first expert-annotated implicature cancellation dataset, ImplicatureX, crowdsourced for human judgements of implicatures and their corresponding cancellations. We find that LLM belief update understanding lags behind that of humans, especially in more naturally-occurring scenarios. Additional control experiments suggest that successes in LLM belief updates may in part stem from a reliance on prior beliefs, and that failures in belief updates may depend on their type and on their form. Overall, our study suggests that current LLMs have not yet reached human-level understanding of communicative belief updates.

We introduce DiscoTrace, a method to identify the rhetorical strategies answerers use when responding to information-seeking questions. DiscoTrace represents answers as a sequence of question-related discourse acts paired with interpretations of the original question, annotated on top of rhetorical structure theory parses. Applying DiscoTrace to answers from nine different communities reveals that communities have diverse preferences for answer construction. In contrast, LLMs do not exhibit rhetorical diversity in their answers, even when prompted to mimic specific human community answering guidelines. LLMs also systematically opt for breadth, addressing interpretations of questions that human answerers choose not to address. The rich, community-sensitive answering behavior structurally revealed by DiscoTrace can guide the development of pragmatic LLM answerers that are more attuned to contextual information needs.

Edge devices such as smartwatches and smart glasses cannot continuously run even the smallest 100M–1B parameter language models due to power and compute constraints, yet cloud inference introduces multi-second latencies that break the illusion of a responsive assistant. We introduce micro language models (𝜇LMs): ultra-compact models (8M–30M parameters) that instantly generate the first 4-8 words of a contextually grounded response on-device, while a cloud model completes it; thus, masking the cloud latency. We show that useful language generation survives at this extreme scale with our models matching several 70M-256M-class existing models. We design a collaborative generation framework that reframes the cloud model as a continuator rather than a respondent, achieving seamless mid-sentence handoffs and structured graceful recovery via three error correction methods when the local opener goes wrong. Empirical results show that 𝜇LMs can initiate responses that larger models complete seamlessly, demonstrating that orders-of-magnitude asymmetric collaboration is achievable and unlocking responsive AI for extremely resource-constrained devices.

Post-training quantization (PTQ) of large language models (LLMs) reduces memory usage and accelerates inference while maintaining accuracy. However, it remains unclear how to select a quantization scheme that achieves target accuracy across different model scales. We introduce two parameters to characterize this: the quantization ratio Q_r (proportion of parameters quantized to low precision) and the block size Q_b (number of parameters sharing a scaling factor). Based on extensive experiments across multiple LLM families and scales, we propose a unified scaling law that predicts quantization-induced loss degradation as a function of Q_r, Q_b, and model size. Our scaling law reveals two practical implications: (1) larger models tolerate higher Q_r, making mixed-precision increasingly effective at scale; (2) small Q_b, commonly used in mixed precision, becomes unnecessary for large models, simplifying hardware design without sacrificing accuracy. These findings provide a principled recipe for quantizing LLMs in practical deployment.

Large language models (LLMs) can correctly answer "When was Einstein born?" yet fail to provide the same date when the same question is embedded in a long-form query requesting multiple facts 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 long-form 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 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 80% 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.

Multimodal Reward Models are critical for aligning Multimodal Large Language Models with human preferences. However, existing reward models offer limited interpretability and struggle to deliver reliable preference judgments in audio-centric scenarios. To address this gap, we propose Audio-Reward-R1, a reasoning-based reward model that supports both audio understanding and audio generation preference tasks. Audio-Reward-R1 formulates reward modeling as a structured reasoning process, producing dimension-level analyses prior to the final preference judgment. To enable reliable reasoning supervision, we construct a high-quality reasoning-enhanced audio preference dataset with structured reasoning traces aligned with preference annotations. Audio-Reward-R1 is first trained via supervised fine-tuning as a cold-start on this reasoning-enhanced dataset, followed by reinforcement learning with verifiable rewards. Finally, we introduce Audio-Reward-Bench, a human-verified evaluation benchmark for evaluating audio reward models across both audio understanding and audio generation preference tasks. Extensive experimental results demonstrate the effectiveness and reliability of Audio-Reward-R1.

The computational and memory overheads associated with expanding the context window of LLMs severely limit their scalability. A noteworthy solution is vision-text compression (VTC), exemplified by frameworks like DeepSeek-OCR and Glyph, which convert long texts into dense 2D visual representations, thereby achieving token compression ratios of 2x–10x. However, the impact of this high information density on the core long-context capabilities of vision-language models (VLMs) remains under-investigated. To address this gap, we introduce the first benchmark for VTC and systematically assess the performance of VLMs across three long-context understanding settings: VTC-Retrieval, VTC-Reasoning, and VTC-Memory. We further introduce VTCBench-Wild to evaluate robustness across diverse rendering configurations. We comprehensively evaluate leading open-weight and proprietary models on our benchmarks. The results indicate that several VLMs can retrieve information from rendered text in short-to-moderate contexts, but this ability degrades at longer lengths, and the degradation is especially pronounced on tasks requiring binding and integrating dispersed evidence. This study provides diagnostic evidence of VTC and serves as a foundation for designing more efficient and scalable VLMs.

Hateful content online is often expressed using fact-like, not necessarily correct information, especially in coordinated online harassment campaigns and extremist propaganda. Failing to jointly address hate speech (HS) and misinformation can deepen prejudice, reinforce harmful stereotypes, and expose bystanders to psychological distress, while polluting public debate. Moreover, these messages require more effort from content moderators because they must assess both harmfulness and veracity, i.e., fact-check them. To address this challenge, we release WSF-ARG+, the first dataset which combines hate speech with check-worthiness information. We also introduce a novel LLM-in-the-loop framework to facilitate the annotation of check-worthy claims. We run our framework, testing it with 12 open-weight LLMs of different sizes and architectures. We validate it through extensive human evaluation, and show that our LLM-in-the-loop framework reduces human effort without compromising the annotation quality of the data. Finally, we show that HS messages with check-worthy claims show significantly higher harassment and hate, and that incorporating check-worthiness labels improves LLM-based HS detection up to 0.213 macro-F1 and to 0.154 macro-F1 on average for large models.

Large Vision-Language Models (LVLMs) achieve strong performance on various vision-language tasks, yet Chain-of-Thought (CoT) generation often weakens attention to visual tokens, degrading evidence-grounded reasoning. During the later stages of autoregressive decoding, models tend to increasingly rely on previously generated text rather than visual evidence, leading to plausible but visually inconsistent rationales. We propose DVAR (Dynamic Visual Attention Recalibration), a training-free method that selectively amplifies visually relevant tokens to mitigate visual attention decay during decoding. DVAR operates at three levels: (i) prompt level structure to encourage visually grounded reasoning; (ii) embedding level relevance weights derived from image–text alignment; and (iii) decoding level scheduling that adjusts recalibration by tracking shifts in visual attention. Without sensitive model-specific hyperparameter retuning, DVAR yields consistent performance improvements across six benchmarks on LLaVA and Qwen-VL models, and better preserves visual attention throughout decoding.

Parameter-efficient fine-tuning with adapters (e.g., LoRA) equips LLMs with task-specific skills. However, utilizing multiple pretrained adapters for out-of-distribution (OOD) generalization remains challenging. Existing techniques for OOD generalization using multiple pretrained LoRAs, route inputs using LoRA representations (prototypes) obtained independently, assuming these representations capture complementary information. However, we observe that for existing methods, in-distribution and OOD routing entropies are often comparable, thus bringing the complementarity assumption into question. We derive the theoretical conditions that could lead to a violation of such assumptions, distilling the cause down to the presence of shared, noisy prototype subspaces. Based on this, we introduce MAPLE (Masked-Adapter Prototype LEarning), a simple learning framework that refines LoRA prototypes by masking the target task’s LoRA during prototype learning. In doing so, it encourages prototypes to discard noisy attributes, improving prototype discriminability, routing, and OOD generalization. Extensive experiments on language models of varying size, such as Phi-2 (2.7B) and LLaMA-3 (8B) equipped with heterogeneous pools of pretrained LoRAs, show that MAPLE improves the LoRA representation and thus achieves state-of-the-art performance across multiple benchmarks.

As large language models (LLMs) are increasingly integrated into educational tools, current evaluations on standardized tests predominantly focus on binary outcome accuracy. Instead, an effective AI tutor must exhibit faithful reasoning, elucidate solution strategies, and diagnose specific human misconceptions. To bridge this gap, we introduce a pedagogical diagnostic framework that models English Standardized Test (EST) problem-solving as a traversal through a cognitive framework. Based on this framework, we present ESTBOOK, a multimodal benchmark encompassing 10,576 questions and 29 task types across five major exams. Unlike traditional datasets, ESTBOOK goes beyond data aggregation by enriching questions with formalized reasoning trajectories and distractor rationales that capture specific cognitive traps. Through extensive evaluations, we show that ESTBOOK identifies fine-grained reasoning bottlenecks and that lightweight mitigation based on these diagnoses provides evidence that the identified failure modes are actionable.

Judge Decoding accelerates LLM inference by relaxing the strict verification of Speculative Decoding, yet it typically relies on expensive and noisy supervision. In this work, we revisit this paradigm from first principles, revealing that the “criticality” scores learned via costly supervision are intrinsically encoded in the draft-target distributional divergence. We theoretically prove a structural correspondence between learned linear judges and Kullback-Leibler (KL) divergence, demonstrating they rely on the same underlying logit primitives. Guided by this, we propose a simple, training-free verification mechanism based on KL divergence. Extensive experiments across reasoning and coding benchmarks show that our method matches or outperforms complex trained judges (e.g., AutoJudge), offering superior robustness to domain shifts and eliminating the supervision bottleneck entirely. Code is available at https://github.com/sunshy-1/JuDi.

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 "wait" (from 5917 to 522 on MATH500). Our findings highlight the limitations of current hybrid thinking and point the ways for enhancing its controllability.

Temporal knowledge graph question answering (TKGQA) aims to answer time-sensitive questions by leveraging temporal knowledge bases. While Large Language Models (LLMs) demonstrate significant potential in TKGQA, current prompting strategies constrain their efficacy in two primary ways. First, they are prone to reasoning hallucinations under complex temporal constraints. Second, static prompting limits model autonomy and generalization, as it lacks optimization through dynamic interaction with temporal knowledge graph (TKG) environments. To address these limitations, we propose TKG-Thinker, a novel agent equipped with autonomous planning and adaptive retrieval capabilities for reasoning over TKGs. Specifically, TKG-Thinker performs in-depth temporal reasoning through dynamic multi-turn interactions with TKGs via a dual-training strategy. We first apply Supervised Fine-Tuning (SFT) with chain of thought data to instill core planning capabilities, followed by a Reinforcement Learning (RL) stage that leverages multi-dimensional rewards to refine reasoning policies under intricate temporal constraints. Experimental results on benchmark datasets with three open-source LLMs show that TKG-Thinker achieves state-of-the-art performance and exhibits strong generalization across complex TKGQA settings.

Subword tokenization is widely used to address the out-of-vocabulary problem in natural language processing. However, frequency-based subword tokenization can produce artifacts in the resulting vocabulary. Beyond the well-known junk tokens (meaningless fragments), we introduce the notion of subsumed tokens, valid words whose corpus frequencies are suppressed by longer tokens that contain them. Together with fragment tokens, these are referred to as overlapping tokens. We propose a post-processing method that reduces fragment tokens and restores subsumed tokens. The method searches over possible binary segmentations of character sequences and selects tokenizations that maximize corpus frequency, while optionally assigning bonuses to word spans identified by a morphological analyzer. As a post-processing step, the method can be applied to existing tokenizers such as Byte Pair Encoding (BPE) and the Unigram Language Model (ULM) without rebuilding their vocabularies. Although evaluated on Japanese, the approach may also be applicable to other languages given an appropriate morphological analyzer. Experiments on the Wiki-40B Japanese corpus show that our method reduces overlapping tokens, leading to more reliable statistics for valid words, and improves representation quality while preserving downstream performance across tasks.

Despite the significant efforts devoted to aligning large language models (LLMs) with human values and ensuring safe deployment, recent work has revealed that LLMs remain vulnerable to adversarial jailbreak attacks that can bypass safety guardrails and elicit harmful responses. Many defense methods are proposed to detect jailbreaks but they are limited in their effectiveness to counter wide-range optimization based jailbreak mechanisms that can yield highly fluency-optimized or harmful semantic obfuscated prompts. To tackle this challenge, we propose a unified detection framework—SAFEGuard—which incorporates a hybrid fluency measurement based on cross-layer distribution distance and perplexity, and the analysis of harmful semantics through gradient matching. Our method is grounded in a paramount observation: high fluency prompts maintain their malicious intention close to harmful prompts while harmful semantic obfuscated prompts often inject gibberish token sequences. Our evaluation demonstrates that SAFEGuard consistently outperforms state-of-the-art baselines and achieves significant improvement in accuracy across different optimization-based jailbreaks. This underscores the effectiveness of SAFEGuard against evolving jailbreak attacks.

Speculative decoding (SD) has emerged as a promising approach to accelerate LLM inference without sacrificing output quality. Existing SD methods tailored for video-LLMs primarily focus on pruning redundant visual tokens to mitigate the computational burden of massive visual inputs. However, existing methods do not achieve inference acceleration comparable to text-only LLMs. We observe from extensive experiments that this phenomenon mainly stems from two limitations: (i) their pruning strategies inadequately preserve visual semantic tokens, degrading draft quality and acceptance rates; (ii) even with aggressive pruning (e.g., 90% visual tokens removed), the draft model’s remaining inference cost limits overall speedup. To address these limitations, we propose HIPPO, a general **h**ol**i**stic-aware **p**arallel s**p**eculative dec**o**ding framework. Specifically, HIPPO proposes: (i) a semantic-aware token preservation method, which fuses global attention scores with local visual semantics to retain semantic information at high pruning ratios; (ii) a video parallel SD algorithm that decouples and overlaps draft generation and target verification phases. Experiments on four video-LLMs across six benchmarks demonstrate HIPPO’s effectiveness, yielding up to 3.51× speedup compared to vanilla auto-regressive decoding.

Multilingual Information Retrieval (MLIR) reflects real-world search environments in which queries and relevant documents may appear in different languages within a mixed-language corpus. However, existing embedding models are primarily optimized for Multi-Monolingual retrieval and their performance often degrades substantially in MLIR settings. Moreover, directly applying conventional contrastive learning to MLIR can exacerbate language clustering and expose a trade-off between cross-lingual alignment and embedding uniformity. To address these limitations, we propose MIMO: Multilingual Information Retrieval via Monolingual Objectives, a two-stage framework that uses a stable English semantic space as an anchor. MIMO first initializes cross-lingual alignment through knowledge distillation, and then jointly optimizes distillation and cross-lingual contrastive learning to improve retrieval discrimination while preserving alignment. Extensive experiments show that MIMO consistently outperforms existing cross-lingual training baselines across various MLIR and Multi-Monolingual benchmarks. MIMO also remains competitive with off-the-shelf models of similar or larger parameter scales. Furthermore, our cross-lingual Alignment-Uniformity analysis clarifies the distinct roles of the two loss components and shows that their combination yields a favorable trade-off between alignment and uniformity.

Test-time compute has become a central way to improve code generation: systems sample multiple candidate programs and use verifier-visible evidence to select the final output. This paradigm implicitly assumes that the verifier provides a corrective signal independent from the generator. We challenge this assumption under misleading task premises. When the generator and verifier share a false premise, they become coupled through a mistaken belief: the generator produces premise-consistent shortcuts, while the verifier supplies evidence that fails to expose them. Consequently, the selector may choose a hidden-test-wrong candidate even when a hidden-test-correct program exists in the pool. We call this failure mode Verification Trap. Across three code-generation benchmarks and five code models, false premises consistently degrade first-sample correctness, reduce selector-chosen correctness after 64-sample test-time selection, and amplify recoverable mis-selection. Mechanistically, verifier-written tests inherit the premise-level blind spot, reshaping verifier-visible candidate space away from hidden-test correctness. These traces make Verification Trap predictable before hidden execution: a lightweight gold-free predictor using verifier-visible features reaches 0.846 AUROC. Our results identify decoupled evidence as a key mitigation axis: coupled scaling provides limited recovery, whereas premise-agnostic robustness auditors recover substantial oracle headroom.

Each LoRA checkpoint compactly stores task-specific updates in low-rank weight matrices, offering an efficient way to adapt large language models to new tasks and domains. In principle, these weights already encode what the adapter does and how well it performs. In this paper, we ask whether this information can be read directly from the weights, without running the base model or accessing training data. A key obstacle is that a single LoRA update can be factorized in infinitely many ways. Without resolving this ambiguity, models trained on the factors may fit the particular factorization rather than the underlying update. To this end, we propose W2T (Weight-to-Token), which maps each LoRA update to a provably canonical form via QR decomposition followed by SVD, so that all equivalent factorizations share the same representation. The resulting components are then tokenized and processed by a Transformer to produce a weight-space embedding. Across language and vision LoRA collections, W2T achieves strong results on attribute classification, performance prediction, and adapter retrieval, demonstrating that LoRA weights reliably indicate model behavior once factorization ambiguity is removed.

The structure of naming systems in natural languages hinges on a trade-off between high informativeness and low complexity. Focusing on the domain of kinship naming, we analyze such trade-off while addressing simplifying assumptions of prior work, namely: (i) universal communicative need across languages, and (ii) optimal listeners. To that aim, we collect data from four different languages, and analyze how different communicative needs and variations in the listener model influence the informativeness–complexity trade-off. Adopting a referential game setup from emergent communication, we further show that trade-off optimality is not only theoretically achievable but also emerges empirically in learned communication systems.

Large Language Models (LLMs) are increasingly used not only for instrumental tasks, but as always-available and non-judgmental confidants for emotional support. Yet what drives adoption and how users perceive emotional support interactions across countries remains unknown. To address this gap, we present the first large-scale cross-cultural study of LLM use for emotional support, surveying 4,641 participants across seven countries (USA, UK, Germany, France, Spain, Italy, and the Netherlands). Our results show that adoption rates vary dramatically across countries (from 20% to 59%). Using mixed models that separate country-level effects from demographic composition, we find that being aged 25-44, religious, married, and of higher socioeconomic status are predictors of positive perceptions (trust, usage, perceived benefits), with socioeconomic status being the strongest. English-speaking countries consistently show more positive perceptions than Continental European countries. We also collect a corpus of 731 real multilingual prompts from user interactions, showing that users mainly seek help for loneliness, stress, relationship conflicts, and mental health struggles. Our findings reveal that LLM emotional support use is shaped by a complex sociotechnical landscape and call for a broader research agenda examining how these systems can be developed, deployed, and governed to ensure safe and informed access. We share our code and data publicly [<https://confidant-ai.github.io/LLMs-emotional-support/>].

Recent advances in equipping Large Language Models (LLMs) with search tools and outcome-reward reinforcement learning (RL) have achieved new state-of-the-art results on open-domain QA tasks. However, we argue that current training paradigms harbor a critical vulnerability: they predominantly reward correct answers but fail to penalize fabricated ones when retrieval fails, thereby implicitly exacerbating hallucinations. To address this, we propose Abstention-Aware Reinforcement Learning (AWA-RL), which dynamically shapes the abstention reward utilizing the model’s query-specific prior capabilities and continuous on-policy training observations. We also introduce a novel metric, RA-F1, to measure the capability-reliability trade-off. Compared to non-abstaining baselines, AWA-RL boosts absolute precision by up to 10.3% and overall RA-F1 by 2.9%, with only marginal sacrifice in raw accuracy. These results confirm that AWA-RL successfully yields highly capable and reliable search agents.

Multilingual large language models (LLMs) have been shown to perform better on non-English classification tasks when the representations of the given language are more aligned to English within the model. Several cross-lingual alignment (CLA) scores have been proposed for use with LLMs, along with multiple approaches for extracting embeddings from the models. We provide a comparative analysis of 27 CLA score variants, examining how they differ and how well each predicts downstream performance across three tasks. Crucially, while LLMs are widely used for generative tasks such as machine translation, prior work has focused almost exclusively on classification. We therefore investigate whether CLA scores are similarly predictive of translation performance. To enable computing correlations across target languages, we propose a PMI-based translation metric, which is less dependent on the target language and correlates strongly with CHRF. We find that CLA with English predicts translation quality comparably to or better than source–target CLA, providing new evidence that LLMs use English as an internal pivot language.

Fine-grained intent detection becomes brittle as label spaces expand with many semantically adjacent intents. Although large language models (LLMs) can synthesize labeled utterances to reduce annotation cost, the resulting pseudo data are often ambiguous across intents and repetitive within intents. Existing synthesis pipelines mainly rely on prompt engineering and filtering, without updating generator parameters. We propose a preference-aligned synthesis framework that automatically mines supervision to optimize the generator. Each intent is represented by multi-centroid prototypes, and block-conditioned generation supports diversity-aware candidate construction. From these candidates, we build preferred blocks that are both prototype-consistent and diverse, and rejected blocks composed of confusable negatives from neighboring intents. These preference pairs are used to optimize the generator with Direct Preference Optimization (DPO), together with iterative refresh to mine harder confusions over time. Experiments show consistent gains for downstream intent classification in zero-shot and few-shot settings, while remaining competitive in fully supervised settings. We further observe improved generator reliability and strong transfer across multiple backbone LLM families.

Internet memes are a powerful form of online communication, yet their multimodal nature and reliance on commonsense and cultural knowledge make automated toxicity detection challenging. Prior work has studied individual dimensions of memes: textual content via optical character recognition, visual content via object recognition, and toxicity via hate speech detection, but lacks a unified architecture that formally identifies the semantic dimensions contributing to meme toxicity. We present three contributions: (1) a semantic framework identifying the dimensions necessary for meme interpretation; (2) a corresponding benchmark, M-QUEST, of 609 question-answer pairs for 307 memes, generated via a semi-automatic process grounded in the framework; and (3) an empirical evaluation of 12 VLMs on M-QUEST. The results show that the instruction-tuned models with reasoning capabilities significantly outperform others, whereas pragmatic inference remains challenging across all models.

Wider beam search should improve LLM reasoning, but when should you stop widening? Prior work on beam width selection has focused on inference efficiency (CITATION), without analyzing whether wider search can hurt output quality. We study each search step as scoring n candidates and keeping the top b of them. Best-of-n keeps a single winner (b=1); beam search keeps several. Because the scorer is noisy, a worse candidate can outrank a better one, and that risk grows as more candidates are compared. How far you can usefully widen therefore depends on the scorer: how large the true quality gap is relative to the noise. Keeping several hypotheses is easier than naming one winner, so beam search is more forgiving than best-of-n, but that per-step analysis does not by itself tell you the best width for the full task. We compare perplexity scoring and a process reward model (PRM) across three 7–8B models and all ten MR-BEN subjects (5,737 questions). Perplexity yields no accuracy gain at any tested width; the PRM gains up to 8.9 points and continues improving through beam width 4 (n=16 candidates per step). Across 444K search steps, the spacing of ranked scores agrees with a simple Gaussian ranking model. The same model and algorithm therefore support very different useful widths, depending on how reliable the scorer is.

Scientific LLM agents have shown promise in literature reasoning, tool use, and experiment planning, but it remains unclear whether they can autonomously improve large, tightly coupled scientific machine-learning systems through executable code changes and computationally expensive validation. We study this question in protein folding, where progress requires coordinated architectural modifications, multi-objective evaluation, and domain-aware interpretation. We present AgentFold, a multi-agent framework that formulates folding-model development as a closed-loop search over executable code variants. Starting from ESMFold, AgentFold proposes hypotheses, implements and debugs code-level modifications, evaluates model variants, analyzes experimental outcomes, and stores both successful and failed interventions in structured memory. An MCTS-style policy allocates computational resources across high-scoring search branches. On an engineering-scale protein-folding codebase comprising more than 2,000 lines of code, AgentFold explores approximately 80 model variants using approximately 5,000 GPU-hours and 170 million LLM tokens. Under a matched computational budget, AgentFold improves the best lDDT by 7.5% over independent Codex proposals and outperforms a random-search control. Beyond model improvement, the resulting intervention traces reveal recurring empirical design patterns: stable gains tend to arise from early, soft, learnable priors and gated refinement, whereas direct geometric perturbations and geometry-conditioned feedback often destabilize training. The code and experimental resources are publicly available at https://github.com/lmqfly/AgentFold.

Recent advancements in large language models (LLMs) have demonstrated their potential in automating the scientific research ideation. Existing approaches primarily focus on prompting techniques, often producing ideas misaligned with expert standards – novelty, feasibility, and effectiveness, which are widely recognized by the research community as the three key subdimensions of high-quality ideas. Also, balancing these dimensions remains challenging due to their inherent trade-offs. To address these limitations, we propose the first framework that employs a two-stage approach combining Supervised Fine-Tuning (SFT) and controllable Reinforcement Learning (RL) for the task. In the SFT stage, the model learns foundational patterns from pairs of research papers and their corresponding follow-up ideas. In the RL stage, multi-dimensional reward models guided by fine-grained feedback evaluate and optimize the model across key dimensions. During inference, dimensional controllers coordinated by a sentence-level decoder enable dynamic context-aware steering of the idea generation process. Our framework provides a balanced approach to research idea generation, achieving high-quality outcomes in the experiment by dynamically navigating the trade-offs among novelty, feasibility, and effectiveness.

With the surge of large language models (LLMs) and their ability to produce customized output, style-personalized text generation–"write like me"–has become a rapidly growing area of interest. However, style personalization is highly specific, relative to every user, and depends strongly on the pragmatic context, which makes it uniquely challenging. Although prior research has introduced benchmarks and metrics for this area, they tend to be non-standardized and have known limitations (e.g., poor correlation with human subjects). LLMs have been found to not capture author-specific style well, it follows that the metrics themselves must be scrutinized carefully. In this work we critically examine the effectiveness of the most common metrics used in the field, such as BLEU, embeddings, and LLMs-as-judges. We evaluate these metrics using our proposed style discrimination benchmark, which spans eight diverse writing tasks across three evaluation settings: domain discrimination, authorship attribution, and LLM-generated personalized vs non-personalized discrimination. We find strong evidence that employing ensembles of diverse evaluation metrics consistently outperforms single-evaluator methods, and conclude by providing guidance on how to reliably assess style-personalized text generation.

Preference alignment is usually achieved by weight-updating training on preference data, which is costly and provides limited mechanistic visibility. We propose Dynamic SAE Steering for Preference Alignment (DSPA), an inference-time method that makes sparse autoencoder (SAE) steering prompt-conditional. From preference triples, DSPA computes a conditional-difference map linking prompt features to generation-control features; during decoding, it modifies only token-active latents, without base-model weight updates. Across Gemma-2-2B/9B and Qwen3-8B, DSPA improves MT-Bench and is competitive on AlpacaEval while preserving multiple-choice accuracy. Under restricted preference data, DSPA remains robust and can rival heavier alignment pipelines at a fraction of the compute. Finally, we audit the SAE features DSPA modifies, finding that preference directions are dominated by discourse and stylistic signals, and provide theory clarifying the conditional-difference map estimate and when top-k ablation is principled.

Knowledge editing enables rapid updates of specific factual knowledge in large language models (LLMs) without full retraining. However, more realistic scenarios call for a lifelong framework that handles continual updates rather than one-off modifications. In such settings, existing editing methods often overfit to target prompts, significantly degrading both the generalization of the edited knowledge and the model’s general capabilities. To address this issue, we propose GLIME (Generalizable Lifelong Model Editing), which combines knowledge editing with preference optimization over generation behavior. GLIME further incorporates replay-based editing and a gradient constraint to preserve previously edited knowledge. Experimental results show that GLIME significantly improves knowledge generalization in lifelong editing settings while maintaining both editing performance and general capabilities.

Machine-generated text (MGT) detection is critical for regulating online information ecosystems, yet existing detectors often fail under limited fresh supervision and humanization-oriented attacks. We study this problem as few-shot robustness recovery under evolving generators, domains, and attack patterns, and use threat modeling to diagnose detector failures from an attacker’s viewpoint. We identify unstable local boundary estimation as a key failure mode: under scarce supervision, humanized MGT can induce prediction flips, logit drift, and clean–adversarial representation misalignment. To address this issue, we propose RAG-GuidEd Attacker Strengthens ConTrastive Few-shot Detector (REACT), a threat-aware few-shot robustness framework. REACT uses a dynamic retrieval-guided attacker to generate boundary-oriented humanized rewrites from extreme positive–negative anchors, and trains the detector with a pairwise boundary contrastive objective to stabilize clean–adversarial local geometry. Alternating attacker–detector updates further form a co-evolutionary curriculum. Experiments on 4 datasets with 4 shot sizes and 3 random seeds show that REACT improves average detection F1 by 4.95 points over 8 SOTA detectors and reduces average attack success rate under 4 strong attacks by 3.66 percentage points.

Long-horizon LLM agents require reinforcement learning methods that can assign credit to intermediate decisions under sparse and delayed rewards. Recent group-based methods such as GiGPO improve over trajectory-level GRPO by constructing step-level advantages at repeated anchor states. However, we show that such dense credit can be statistically unreliable: under limited rollouts, rare but lucky actions may receive overly large advantages, producing divergent anchor bias and late-stage training oscillation. We propose Evidence-Calibrated Policy Optimization (ECPO), a critic-free policy optimization algorithm that calibrates step-level credit before policy updates. ECPO combines Evidence-Calibrated Action Advantage, which groups rollouts by canonical actions and shrinks low-count estimates, with Variance-Gated Credit Weighting, which suppresses anchor states dominated by within-action noise. Experiments on ALFWorld and WebShop with Qwen2.5-1.5B/7B show that ECPO consistently outperforms strong baselines, improving GiGPO by +5.2/+7.3 success points on ALFWorld/WebShop with Qwen2.5-1.5B while adding only 0.1% additional advantage-computation overhead.

Deep research agents have shown remarkable potential in handling long-horizon tasks. However, state-of-the-art performance typically relies on online reinforcement learning (RL), which is financially expensive due to extensive API calls. While offline training offers a more efficient alternative, its progress is hindered by the scarcity of high-quality research trajectories. In this paper, we demonstrate that expensive online reinforcement learning is not all you need to build powerful research agents. To bridge this gap, we introduce a fully open-source suite designed for effective offline training. Our core contributions include DeepForge, a ready-to-use task synthesis framework that generates large-scale research queries without heavy preprocessing; and a curated collection of 66k QA pairs, 33k SFT trajectories, and 21k DPO pairs. Leveraging these resources, we train OffSeeker, a model developed entirely offline. Extensive evaluations across six benchmarks show that OffSeeker not only leads among similar-sized agents but also remains competitive with 30B-parameter systems trained via heavy online RL.

Dental triage is a safety-critical clinical routing task that requires integrating multimodal clinical information (e.g., patient complaints and radiographic evidence) to determine complete referral plans. We present Dental-TriageBench, the first expert-annotated benchmark for reasoning-driven multimodal dental triage. Built from authentic outpatient workflows, it contains 246 de-identified cases annotated with expert-authored golden reasoning trajectories, together with hierarchical triage labels. We benchmark 19 proprietary, open-source, and medical-domain MLLMs against three junior dentists serving as the human baseline, and find a substantial human–model gap, on fine-grained treatment-level triage. Further analyses show that accurate triage requires both complaint and OPG information, and that model errors concentrate on cases with multiple referral domains, where MLLMs tend to produce overly narrow referral sets and omission-heavy errors. Dental-TriageBench provides a realistic testbed for developing multimodal clinical AI systems that are more clinically grounded, coverage-aware, and safer for downstream care.

Dense retrieval relies on aligning document and query embeddings; however, its performance is often constrained by the scarcity of labeled queries. Although prior work has attempted to model the many-to-one relationship between documents and queries using Gaussian Mixture Models (GMMs), such Euclidean-based approaches induce norm shrinkage (‖𝜇‖ < 1) in normalized semantic spaces, resulting in distorted similarity scores. To address this fundamental geometric inconsistency, we propose vMF-HSE, a training-free hyperspherical semantic expansion framework that models the latent query space using a von Mises-Fisher (vMF) mixture distribution on the unit hypersphere 𝕊^d-1. We leverage Sra’s Newton-based optimization within the Expectation-Maximization (EM) algorithm to ensure robust parameter estimation even in high-dimensional semantic spaces. Building upon this stable estimation, we propose a novel adaptive stopping mechanism (𝜅-Stop) that adaptively determines the optimal number of synthetic queries per document by monitoring the convergence of marginal information gain. Experimental results on the BEIR benchmark demonstrate that vMF-HSE consistently outperforms conventional Euclidean-based methods in zero-shot retrieval settings, establishing the necessity of geometrically consistent modeling for normalized semantic spaces.

Developing reliable models for clinical tasks, such as Medical Question Answering (QA), is severely constrained by the limited availability of high-quality, expert-annotated training data. This challenge is exacerbated by stringent privacy requirements and the impracticality of utilizing large open-source corpora or proprietary cloud APIs within resource-limited clinical settings. To address these obstacles, we introduce SAGE (Semantic Anchor-Guided Evolution), a novel data synthesis framework that enables small, locally deployed models to generate high-quality medical training data. SAGE leverages lightweight, publicly available taxonomies such as MeSH as semantic anchors, imposing a structured prior to effectively guide and ground the data generation process. At its core, SAGE iteratively interleaves atomic (individual concept-based) and associative (relation-based) synthesis, bootstrapping training data from minimal seeds. This approach eliminates the need for large collections of medical documents or reliance on external APIs, providing a practical solution for on-premises data creation. Extensive experiments across multiple medical question-answering benchmarks demonstrate that models fine-tuned with SAGE-synthesized data consistently outperform those trained using self-derived or conventional document-based paradigms, highlighting tangible improvements in data efficiency and resource utilization for medical LLM development.

While draft-model-free speculative decoding offers a promising path to efficient LLM inference, it is frequently constrained by stale draft candidates and the high computational cost of the verification. To address these challenges, we propose ECHO, a hierarchical dual-loop framework that exploits the functional asymmetry between LLM layers. Leveraging the high discriminative efficiency of early layers and the authoritative distribution of final layers, ECHO bifurcates inference into a high-frequency inner loop and a low-frequency outer loop. Within the inner loop, early-layer bonus logits drive rapid, multi-step draft-tree exploration at a minimal cost. Simultaneously, the outer loop performs authoritative full-model verification through a state-reuse mechanism. Crucially, the outer loop also utilizes final-layer bonus logits to correct existing paths and supplement the tree with high-confidence candidates for subsequent cycles. Experimental results across diverse benchmarks demonstrate that ECHO significantly boosts mean accepted tokens and achieves a 2.4× to 2.9× speedup, outperforming existing state-of-the-art baselines with negligible engineering overhead and no extra deployment parameters, albeit with a one-shot fine-tuning dependency for optimal acceleration. The code is available at https://github.com/whucs21Mzy/ECHO.

A human’s moral decision depends heavily on the context. Yet research on LLM morality has largely studied fixed scenarios. We address this gap by introducing Contextual MoralChoice, a dataset of moral dilemmas with systematic contextual variations known from moral psychology to shift human judgment: consequentialist, emotional, and relational. Evaluating 22 LLMs, we find that nearly all models are context-sensitive, shifting their judgments towards rule-violating behavior. Comparing with a human survey, we find that models and humans are most triggered by different contextual variations, and that a model aligned with human judgments in the base case is not necessarily aligned in its contextual sensitivity. This raises the question of controlling contextual sensitivity, which we address with an activation steering approach that can reliably increase or decrease a model’s contextual sensitivity. Code and data: https://github.com/adrian-sauter/contextual_moralchoice.

Discrete motion tokenization has recently enabled Large Language Models (LLMs) to serve as versatile backbones for motion understanding and motion-language reasoning. However, existing pipelines typically decouple motion quantization from semantic embedding learning, linking them solely via token IDs. This approach fails to effectively align the intrinsic geometry of the motion space with the embedding space, thereby hindering the LLM’s capacity for nuanced motion reasoning. We argue that alignment is most effective when both modalities share a unified geometric basis. Therefore, instead of forcing the LLM to reconstruct the complex geometry among motion tokens from scratch, we present a novel framework that explicitly enforces orthogonality on both the motion codebook and the LLM embedding space, ensuring that their relational structures naturally mirror each other. Specifically, we employ a decoder-only quantizer with Gumbel-Softmax for differentiable training and balanced codebook usage. To bridge the modalities, we use a sparse projection that maps motion codes into the LLM embedding space while preserving orthogonality. Finally, a two-stage orthonormal regularization schedule enforces soft constraints during tokenizer training and LLM fine-tuning to maintain geometric alignment without hindering semantic adaptation. Extensive experiments show that our framework improves the aggregated Average by 22.4% over the strongest baseline on HumanML3D and by 14.4% on KIT-ML, while ablations confirm the effectiveness of the tokenizer, projection, and regularization designs.

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-Augmented Generation (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 outperforms much larger state-of-the-art models. These results demonstrate the potential of SustainableQA as a resource for developing and benchmarking advanced knowledge assistants capable of navigating complex sustainability compliance data

Communication protocol design is a central challenge in large language model-based multi-agent systems. Existing single-channel approaches face an inherent communication trilemma: text-based methods are interpretable but verbose, while latent-space methods are efficient but opaque and limited to single-round workflows. Inspired by multi-channel communication theory, we propose HyLaT, a hybrid latent-text communication protocol that transmits elaborate cognitive signals through a latent channel for efficiency, while expressing concise critical signals in natural language to preserve interpretability and precision. We introduce a two-stage training framework combining single-agent hybrid generation learning and multi-agent interactive co-training, enabling agents to generate and interpret hybrid messages across multiple rounds of interaction. Experiments demonstrate that HyLaT reduces communication overhead significantly while maintaining competitive task performance, with strong generalization and robustness across diverse settings.

Large language models (LLMs) are trained through multi-stage pipelines over heterogeneous data sources, yet developers lack a principled way to pinpoint the specific data responsible for an observed behavior. This lack of observability reduces debugging to reactive patching and makes failures prone to recur under distribution shift or subsequent model updates. To address this limitation, we propose DebugLM, a framework that equips LLMs with built-in data provenance, enabling them to explicitly trace the origins of their behaviors to specific training data sources. Specifically, the model learns to associate its responses with unique provenance tags that indicate the responsible dataset, empowering developers to precisely identify where undesirable behaviors are learned. Building on this capability, DebugLM further supports targeted test-time remediation, enabling developers to selectively trigger targeted refusal for specified data sources without retraining or modifying model parameters. Experiments demonstrate that DebugLM provides accurate behavior tracing in multi-stage training pipelines and effective test-time remediation while preserving the general utility of the model.

Sign language processing systems have traditionally operated at the sentence level, ignoring critical discourse phenomena fundamental to sign language comprehension. We introduce DiscoSign, a computational approach for discourse-aware text to sign language gloss translation grounded in linguistic research. We address three key phenomena within our modular Large Language Model (LLM)-based translation framework: (i) spatial coreference resolution, where entities maintain consistent spatial locations throughout discourse; (ii) Question-Answer Clauses (QACs), pseudocleft structures serving specific discourse functions; and (iii) concept-gloss consistency, ensuring stable mappings between English concepts and American Sign Language (ASL) signs. Traditional translation metrics fail to capture discourse-level quality, so we introduce a suite of novel evaluation metrics designed to assess each dimension of discourse coherence addressed by our framework. Experiments on sentence-level and discourse-level datasets show that our approach for discourse-aware processing significantly improves spatial consistency and entity tracking relative to sentence-only translation, while maintaining competitive single-sentence gloss translation quality. Our work establishes the first systematic framework for discourse-level text to sign language gloss translation with corresponding evaluation methodology.

The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD’s routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30 × attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

A common approach to personalization in large language models (LLMs) is to incorporate a subset of the user memory into the prompt at inference time to guide the model’s generation. Existing methods select these subsets primarily using similarity between user memory items and input queries, ignoring how features actually affect the model’s response distribution. We propose **R**esponse-**U**tility optimization for **M**emory **S**election (RUMS), a novel method that selects user memory items by measuring the mutual information between a subset of memory and the model’s outputs, identifying items that reduce response uncertainty and sharpen predictions beyond semantic similarity. We demonstrate that this information-theoretic foundation enables more principled user memory selection that aligns more closely with human selection compared to state-of-the-art methods, and models that are 400 × larger. Additionally, we show that memory items selected using RUMS result in better response quality compared to existing approaches, while having up to 95% reduction in computational cost.

Latent tokens are gaining attention for enhancing reasoning in large language models (LLMs), yet their internal mechanisms remain unclear. This paper examines the problem from a reliability perspective, uncovering fundamental weaknesses: latent tokens function as uninterpretable placeholders rather than encoding faithful reasoning. While resistant to perturbation, they promote shortcut usage over genuine reasoning. We focus on Chain-of-Continuous-Thought (COCONUT), which claims better efficiency and stability than explicit Chain-of-Thought (CoT) while maintaining performance. We investigate this through two complementary approaches. First, steering experiments perturb specific token subsets, namely COCONUT and explicit CoT. Unlike CoT tokens, COCONUT tokens show minimal sensitivity to steering and lack reasoning-critical information. Second, shortcut experiments evaluate models under biased and out-of-distribution settings. Results on MMLU and HotpotQA demonstrate that COCONUT consistently exploits dataset artifacts, inflating benchmark performance without true reasoning. These findings reposition COCONUT as a pseudo-reasoning mechanism: it generates plausible traces that conceal shortcut dependence rather than faithfully representing reasoning processes.

Large Language Model (LLM)-based multi-agent systems (MAS) have emerged as a promising paradigm for complex tasks. However, existing works often rely on manual designs or “one-size-fits-all” automation and lack adaptability after deployment. We study in-context structural adaptation, where structured experience conditions both query-dependent system generation and execution-time reconfiguration without updating LLM parameters. We introduce MASFly, which realizes this adaptation through two complementary mechanisms. First, a retrieval-augmented SOP instantiation mechanism retrieves and adapts successful collaboration patterns to construct a query-specific MAS. Second, an experience-enhanced process supervision mechanism uses a dedicated Watcher agent to monitor execution against prior failure experience and reconfigure the system upon abnormal behavior. Experiments demonstrate that MASFly achieves state-of the-art performance, including a 61.7% success rate on TravelPlanner, with strong task adaptability and robustness.

Evaluating the conversational abilities of large language models (LLMs) remains a challenging task. Current mainstream approaches primarily rely on the “LLM-as-a-judge" paradigm, where an LLM is prompted to serve as an evaluator to assess dialogue quality. However, such methods often suffer from various biases, which undermine the reliability and consistency of the evaluation results. To mitigate these biases, recent methods employ multiple LLMs as judges and aggregate their judgments to select the optimal assessment. Although effective, this multi-judge approach incurs significant computational overhead during inference. In this paper, we propose an efficient multi-turn dialogue evaluator that captures the collective wisdom of multiple LLM judges by aggregating their preference knowledge into a single model. Our approach preserves the advantages of diverse multi-judge feedback while drastically reducing the evaluation cost, enabling fast and flexible dialogue quality assessment. Extensive experiments on seven single rating and pairwise comparison dialogue evaluation benchmarks demonstrate that our method outperforms existing baselines across diverse scenarios, showcasing its efficiency and robustness.

Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, yet their ability to make context-aware safety decisions remains limited. Existing methods often fail to balance oversensitivity (unjustified refusals of benign queries) and undersensitivity (missed detection of visually grounded risks), leaving a persistent gap in safety alignment. To address this issue, we introduce Safety-aware Contrastive Decoding (SafeCoDe), a real-time decoding framework that dynamically adjusts token generation based on multimodal context. SafeCoDe operates in two stages: (1) a contrastive decoding mechanism that highlights tokens sensitive to visual context by contrasting real and Gaussian-noised images, and (2) a global-aware token modulation strategy that integrates scene-level reasoning with token-level adjustment to adapt refusals according to the predicted safety verdict. Extensive experiments across diverse MLLM architectures and safety benchmarks, covering undersensitivity, oversensitivity, and general safety evaluations, show that SafeCoDe consistently improves context-sensitive refusal behaviors while preserving model helpfulness.

LLM-reranking is limited by the top-k documents retrieved by vector similarity, which neither enables contextual query-document token interactions nor captures multimodal relevance distributions. While LLM query reformulation attempts to improve recall by generating improved or additional queries, it is still followed by vector similarity retrieval. We thus propose to address these top-k retrieval stage failures by introducing ReBOL, which 1) uses LLM query reformulations to initialize a multimodal Bayesian Optimization (BO) posterior over document relevance, and 2) iteratively acquires document batches for LLM query-document relevance scoring followed by posterior updates to optimize relevance. After exploring query reformulation and document batch diversification techniques, we evaluate ReBOL against LLM reranker baselines on five BEIR datasets and using two LLMs (Gemini-2.5-Flash-Lite, GPT-5.2). ReBOL consistently achieves higher recall and competitive rankings, for example compared to the best LLM reranker on the Robust04 dataset with 46.5% vs. 35.0% recall@100 and 63.6% vs. 61.2% NDCG@10. We also show that ReBOL can achieve comparable latency to LLM rerankers.

We introduce an expert-guided, real-world benchmark for evaluating document-grounded question-answering (QA) motivated by generic drug assessment, using the U.S. Food and Drug Administration (FDA) drug label documents. Drug labels contain rich but heterogeneous clinical and regulatory information, making accurate question answering difficult for current language models. In collaboration with FDA regulatory assessors, we introduce FDARxBench, and construct a multi-stage pipeline for generating high-quality, expert-guided, LLM-assisted, QA examples spanning factual, multi-hop, and refusal tasks, and design evaluation protocols to assess both open-book and closed-book reasoning. Experiments across proprietary and open-weight models reveal substantial gaps in factual grounding, long-context retrieval, and safe refusal behavior. While motivated by FDA generic drug assessment needs, this benchmark also provides a substantial foundation for challenging regulatory-grade evaluation of label comprehension.

Long-horizon LLM agents can benefit from reusable skills, yet existing skill-based methods often rely on external skill generators during training or persistent skill retrieval at inference, increasing engineering complexity, context length, and deployment latency. We propose Self-Internalizing Reinforcement learning with Intrinsic skills (SIRI), a three-phase framework that enables agents to discover, validate, and internalize skills without external skill generators or inference-time skill banks. SIRI first warms up the policy with GiGPO to acquire basic interaction ability and collect successful skill-free trajectories. It then performs self-skill mining, where the current policy summarizes compact skills from its own successful plain rollouts and validates them through paired skill-augmented and skill-free rollouts. Finally, SIRI distills only beneficial skill-guided action tokens into the plain policy using trajectory-level utility and action-level advantage. At inference, the agent runs with the original prompt only. On ALFWorld and WebShop with Qwen2.5-7B-Instruct, SIRI improves GiGPO from 0.908 to 0.930 on ALFWorld and from 0.728 to 0.813 on WebShop, outperforming prompt-based, RL-based, and memory-augmented baselines. Further analysis shows that our self-mining strategy can achieve performance comparable to distillation with closed-source large model.

Discrete diffusion language models (DDLMs) offer potential advantages for language generation, such as parallel decoding and iterative refinement, but they often underperform autoregressive (AR) models and exhibit higher training instability. Recent approaches incorporate AR training signals to stabilize training, yet this introduces strong sequential inductive biases that limit the benefits of non-autoregressive generation. In this work, we analyze the structural causes of degeneration in non-AR DDLMs. We show that standard diffusion formulations rely on an i.i.d. corruption process and factorized sampling structure that implicitly assume token exchangeability, which conflicts with the non-exchangeable dependency structure of natural language. This mismatch leads to degeneration under semantic sparsity, where contextual signals are insufficient. To guarantee stable language generation, we propose the Dependency-Coupling Principle, a structural constraint that preserves necessary contextual dependencies during diffusion generation and provides insight into the performance gap between AR and diffusion language models.

Large Language Models (LLMs) are evolving from performing end-to-end mathematical reasoning to integrating agentic intelligence. However, most existing math benchmarks evaluate only final answers. This outcome-oriented evaluation provides limited diagnostic value for identifying process-level failures or rigorous logic, failing to guide the transformation of LLMs into robust agents. To bridge this gap, we present a process-level benchmark designed to evaluate the inherent agentic mathematical reasoning abilities of LLMs. Our framework aligns problem-solving agentic behaviors with a structured taxonomy of reusable mathematical atomic capabilities. We design a comprehensive suite of planning, action, and feedback tasks across both textual and multimodal contexts, supported by an automated pipeline that synthesizes high-quality trajectories and produces fine-grained annotations via controlled LLM rewriting. Experiments reveal that models with stronger end-to-end accuracy can exhibit markedly different agentic capability profiles. This demonstrates that process-level evaluation is crucial for interpreting the true potential of LLMs and guiding the development of next-generation mathematical agents.

Contrastive Decoding (CD) improves generation quality, but its amateur-model pass makes decoding expensive. Accelerating CD with speculative decoding raises a proposal-alignment question: should the contrastive signal shape the drafter, or should it remain only in verification? We study this question in the lightweight feature-level drafter regime. Two controlled diagnostics, matched Cross-α training and an Approximate Dual-Drafter decomposition, give the same diagnosis: contrastive-aware drafting does not consistently improve over expert-aligned drafting because the contrastive correction is usually weaker than drafter error, and reconstruction can amplify that error. We introduce Decoupled Contrastive Decoding (DCD), which drafts with an expert-aligned lightweight proposer and applies the amateur only in unchanged CD verification. Standard speculative verification preserves the vanilla-CD output distribution. Across the main 8B settings, EAGLE3-based DCD achieves average greedy speedups of 1.65 to 1.95× over vanilla CD and reduces MMLU proposal-path latency by about 5 to 12× relative to amateur-coupled proposal paths.

Multimodal large language models (MLLMs) are increasingly used to automate chart generation from data tables, improving efficiency but introducing new misuse risks. We present ChartAttack, a framework for evaluating how MLLMs use design misleaders to generate charts that induce incorrect interpretations. We also introduce AttackViz, a chart question-answering (QA) dataset labeled with effective misleaders and their induced incorrect answers. ChartAttack reduces MLLM QA accuracy by 17.2 points in-domain and 11.9 points cross-domain. Conditional deception rates show targeted effects: correct answers shift to attacker-intended answers 11.2% of the time in-domain and 11.7-14.9% cross-domain, while originally incorrect answers rarely change. A controlled human study shows that ChartAttack-generated charts also reduce human QA performance. Finally, fine-tuning on AttackViz improves in-domain MLLM robustness to misleading charts. Our findings highlight the need for secure, robust MLLM chart generation. Code and data are publicly available on the project website.

LLM-as-judge evaluation has become standard practice for open-ended model assessment; however, judges exhibit systematic biases that cannot be averaged out by increasing the number of scenarios or generations. These biases are often similar in magnitude to the model differences that benchmarks are designed to detect, resulting in unreliable rankings when single-judge evaluations are used. We introduce a variance decomposition that partitions benchmark score variance into scenario, generation, judge, and residual components. Based on this analysis, CyclicJudge, a round-robin assignment of judges to scenarios, is demonstrated to be the optimal strategy for a fixed judge panel and judge-call budget: the score recovers the panel mean exactly while matching the cost of single-judge evaluation. Empirical results on MT-Bench and MindEval validate the effectiveness of CyclicJudge as predicted, across both general-purpose and domain-specific evaluation settings.

On-policy distillation reduces train-inference mismatch by supervising student-generated trajectories, but full-vocabulary token-level distribution matching remains expensive for large language models. A common approximation truncates reverse-KL distillation to the student’s top-k tokens, which preserves efficiency but drops tail-token supervision and ignores support-level mass mismatch between the student and teacher. We analyze these approximation gaps and propose Stochastic Support Sampling with Teacher Mass Reweighting, which samples fixed-size supports from the student distribution and uses teacher support mass as a reliability weight for the conditional distillation loss. Under the same support budget, our method outperforms the strongest baseline by +1.4 hard-math average points on Qwen3-8B→Qwen3-1.7B-Base, and by +1.1 math average points and +2.8 code average points on DeepSeek-R1-Distill-Llama-8B→Meta-Llama-3.1-8B-Base.

In the unsupervised self-evolution of Multimodal Large Language Models, the quality of feedback signals during post-training is pivotal for stable and effective learning. However, existing self-evolution methods predominantly rely on majority voting to select the most frequent output as the pseudo-golden answer, which may stem from the model’s intrinsic biases rather than guaranteeing the objective correctness of the reasoning paths. To counteract the degradation, we propose Continuous Softened Retracing reSampling (CSRS) in MLLM self-evolution. Specifically, we introduce a Retracing Re-inference Mechanism (RRM) that the model re-inferences from anchor points to expand the exploration of long-tail reasoning paths. Simultaneously, we propose Softened Frequency Reward (SFR), which replaces binary rewards with continuous signals, calibrating reward based on the answers’ frequency across sampled reasoning sets. Furthermore, incorporated with Visual Semantic Perturbation (VSP), CSRS ensures the model prioritizes mathematical logic over visual superficiality. Experimental results demonstrate that CSRS significantly enhances the reasoning performance of Qwen2.5-VL-7B on benchmarks such as MathVision. We achieve state-of-the-art (SOTA) results in unsupervised self-evolution on geometric tasks.

Model merging aims to integrate task-specific abilities from individually fine-tuned models into a single model without extra training. In model merging methods, the task vector has become a fundamental building block, as it can encapsulate the information from fine-tuning. However, the merged model often suffers from performance degradation due to conflicts caused by task-irrelevant redundancy in task vectors. To address these challenges, we propose Purified TAsk VEctors (PAVE) in knowledge-aware subspace. Concretely, we sample training examples from each task, and feed them into their fine-tuned models to acquire covariance matrices before linear layers. We then perform a context-oriented singular value decomposition, which accentuates the weight components most relevant to the target knowledge. As a result, we split fine-tuned model weights into task-relevant and redundant components in the knowledge-aware subspace, and purify the task vector by pruning redundant components. To induce fair pruning efforts across models, we further introduce a spectral rank allocation strategy by optimizing a normalized activated pruning error. The task vector purification by our method as a plug-and-play scheme is applicable across various task vector-based merging methods to improve their performance. In experiments, we demonstrate the effectiveness of PAVE across diverse merging methods, tasks, and model architectures.

Graph-based Retrieval-Augmented Generation (GraphRAG) can fail in interdisciplinary medicine where evidence comes from incommensurable knowledge systems, e.g., Traditional Chinese Medicine (TCM) and Western Medicine (WM). Similarity-driven retrieval in a shared latent space may over-unify heterogeneous concepts, causing semantic collapse (spurious alignments, provenance-blind mixing, and broken multi-hop chains). We propose Xonto-RAG, a cross-ontology GraphRAG framework that avoids ontological unification by constructing a dual-ontology graph with typed bridges induced via LLM-assisted verify-and-vote. A query-aware Relational Graph Attention Network (RGAT) then assembles compact, provenance-tracked multi-hop evidence subgraphs for grounded generation. Across TCM–WM benchmarks, Xonto-RAG outperforms strong RAG/GraphRAG baselines, seeing up to +22.2% accuracy on MLEC-QA TCWM (vs. OG-RAG), 77.0% accuracy with a 7B backbone, and 85.1% faithfulness on MLEC-TCM, while reducing semantic-collapse failures and yielding interpretable reasoning chains.

Controllable text-to-speech (TTS) aims to generate speech from text with precise control over prosodic and speaker-related attributes such as pitch, age, and accent. Existing methods mainly rely on natural language prompts or reference audio cloning. However, prompt-based approaches often suffer from the cross-modal semantic gap between textual descriptions and target speech attributes, resulting in imprecise and coarse-grained control, while cloning-based methods are constrained by reference audio and offer limited flexibility. To address these limitations, we propose TTS-Hub, a controllable TTS framework that leverages modular Low-Rank Adaptation (LoRA) components and their arithmetic composition for fine-grained and flexible speech control. Specifically, we first build a comprehensive Data Hub covering 6 high-level categories and 32 fine-grained speech attributes, and use it to construct a LoRA Hub in which each LoRA is specialized for a specific attribute. During inference, TTS-Hub selects and composes the required LoRA modules through simple arithmetic operations, enabling flexible and extensible attribute control without retraining the backbone model. Experiments on two mainstream TTS frameworks demonstrate that TTS-Hub consistently outperforms strong baselines and enables precise and flexible control.

Test-Time Scaling (TTS) enhances the reasoning capabilities of large language models by allocating additional inference compute to explore the solution space. However, existing parallel TTS methods typically keep branches isolated during search: intermediate discoveries remain branch-private and cannot guide other branches in time. This information isolation causes substantial redundant exploration, as branches repeatedly rediscover information already found elsewhere and require more search steps to collect complete decision information needed to reach correct answers. To bridge this gap, we propose Collaborative Parallel Thinking (CPT), a training-free inference framework that enables search-time information sharing across parallel branches. CPT extracts compact intermediate information from ongoing branches, maintains a deduplicated query-level information pool, and broadcasts pool entries through the input context, allowing each branch in subsequent search steps to reuse discoveries made by other branches rather than rediscover the same information. Empirically, experiments on HMMT and AIME benchmarks show that CPT establishes a stronger accuracy–latency Pareto frontier than strong baselines across rollout budgets and model scales, highlighting search-time collaboration as an effective direction for efficient parallel TTS.

Offensive dark humor presents a high-context safety challenge: its harmful meaning is often absent from surface wording and emerges only after recovering background knowledge and rhetorical transformation. To explore this problem, we introduce HellHumor-CN, a benchmark of 2,168 Chinese jokes annotated for three subtasks: (1) offense detection, (2) sensitive topic prediction, and (3) target identification. We further propose HumorLens, a knowledge-grounded and rhetoric-aware multi-agent framework that retrieves relevant background information and analyzes three common mechanisms in dark humor: mockery, irony, and pun. Experiments show that toxic content detectors and strong LLMs struggle with Chinese dark humor, especially on topic and target interpretation. HumorLens consistently improves performance across multiple models, suggesting that background grounding and mechanism-specific analysis help identify implicit offensive meanings. Further analyses reveal persistent challenges, including visually implied harm, coded sexual innuendo, lexical oversensitivity, and abstract target grounding.

To reliably assist human decision-making, LLMs must maintain factual internal beliefs against misleading injections. While current models resist explicit misinformation, we uncover a fundamental vulnerability to sophisticated, hard-to-falsify evidence. To systematically probe this weakness, we introduce MisBelief, a framework that generates misleading evidence via collaborative, multi-round interactions among multi-role LLMs. This process mimics subtle, defeasible reasoning and progressive refinement to create logically persuasive yet factually deceptive claims. Using MisBelief, we generate 4,800 instances across three difficulty levels to evaluate 7 representative LLMs. Results indicate that while models are robust to direct misinformation, they are highly sensitive to this refined evidence: belief scores in falsehoods increase by an average of 93.0%, fundamentally compromising downstream recommendations. To address this, we propose Deceptive Intent Shielding (DIS), a governance mechanism that provides an early warning signal by inferring the deceptive intent behind evidence. Empirical results demonstrate that DIS consistently mitigates belief shifts and promotes more cautious evidence evaluation.

Instructions are never complete specifications of the behavior they request: speakers say only what their listeners cannot supply for themselves. An agent told to “prepare an apple” must decide unaided whether to wash or cut it and where to place it. Such choices differ between people and are seldom stated, only enacted. Multimodal foundation models render instructions into grounded plans, yet settle these residual choices by generic priors rather than by evidence about the person served. Whether an agent can instead recover such constraints from prior behavior has not been separated from the easier question of planning once they are supplied. Here we show the failure to be largely one of acquisition rather than planning, and that stating the inferred preference before acting recovers much of what is lost. Preference-based Planning (PBP) benchmarks 290 preferences over action parameters, placement policies, and temporal orderings, asking agents to infer from a few unlabeled demonstrations which constraint governs a new underspecified instruction. Given the preference, the strongest models plan categorical tasks at near-zero error; required to infer it, performance drops sharply, and without demonstrations accuracy falls by over a third. Because a verbalized rule keeps intent while discarding perceptual particulars, preferences so represented transfer to unfamiliar scenes, whereas implicit imitation stays bound to the rooms it was learned in. The bottleneck lies not in generating actions but in the pragmatic inference preceding them, for which language proves an effective medium.

Dialogue systems typically optimize the immediate quality of a response, but rarely anticipate how that response will be interpreted by the user. This is especially important in emotional support conversations, where a locally plausible reply may still be perceived as judgmental, dismissive, pressuring, or misaligned with the user’s current need. We frame this as a problem of candidate response evaluation under user-side interpretation uncertainty. We propose Cognitive Look-ahead Simulation (CLS), an inference-time framework that previews likely user interpretations before committing to the final response. CLS expands single-pass generation into a Generate–Simulate–Evaluate–Revise pipeline: it first infers a Structured User-State Representation, then uses a Cognitive World Model as an operational one-step predictor of the user’s immediate inner thought and next-turn state under a candidate reply, and finally applies selective revision to high-risk drafts. We evaluate CLS on the SAGE benchmark using simulator-based outcomes and LLM-as-a-Judge capability profiling, with complementary human evaluations. Across multiple backbones, CLS improves dialogue outcomes and reduces severe failures compared with direct generation and self-refinement baselines. These results suggest that anticipating user interpretations provides a useful inference-time signal for identifying and revising high-risk responses in emotional support conversations.

On-policy distillation (OPD) has recently emerged as an effective post-training paradigm by providing supervision on student-generated trajectories. However, existing OPD methods for multimodal reasoning usually rely on a static teacher routing, assigning each sample to a single teacher based on modality or task type. This ignores that visual grounding and abstract reasoning may dominate different decoding steps, making a single teacher insufficient for the full trajectory. To this end, H-OPD is proposed as a confidence-aware heterogeneous multi-teacher OPD framework for multimodal reasoning. By verifying the complementarity of heterogeneous teachers in the same reasoning process, H-OPD replaces task or sample level teacher routing with token-level teacher arbitration along the shared student trajectory. H-OPD employs vision-to-language description transfer to enable text-only teachers to access key visual semantics, and uses a confidence-aware arbitration mechanism to dynamically combine vision-language teacher and text-only teachers at each token. Extensive evaluations over 11 widely-used reasoning benchmarks showcase the superior performance of our method.

Despite the remarkable success of Multimodal Large Language Models (MLLMs) across diverse tasks, the internal mechanisms governing how they encode and ground distinct visual concepts remain poorly understood. To unravel these mechanisms, we propose a causal framework based on activation steering to actively probe and manipulate internal visual representations. Through systematic intervention across four visual concept categories, our results reveal a divergence in concept encoding: entity knowledge is distinctively localized, whereas abstract concepts are globally distributed across the network. Critically, this divergence uncovers a mechanistic driver of scaling laws: increasing model depth is indispensable for encoding distributed and complex abstract concepts, whereas entities maintain a consistently high degree of localization. Furthermore, reverse steering uncovers that blocking explicit output triggers a surge in latent activations, exposing a compensatory mechanism between perception and generation. Finally, by extending our analysis to visual reasoning, we expose a disconnect between perception and reasoning: although MLLMs successfully recognize geometric relations, they treat them merely as static visual features, failing to trigger the procedural execution necessary for problem-solving.

Fine-tuning Large Language Models with untrusted data exposes models to backdoor attacks, where poisoned samples cause targeted misbehavior. Existing sample-filtering defenses rely on clustering, which requires sufficient data and can fail at extreme poison ratios. We propose GradSentry(Gradient Sentry), a backdoor sample filtering method based on the spectral entropy of per-sample gradients. Our key finding is that poisoned samples produce gradients with higher spectral entropy compared to clean samples. GradSentry captures output-altering backdoor signatures using per-sample gradient spectra, avoiding pairwise sample comparisons and clustering during feature construction. Importantly, our method is training-agnostic: it works for both parameter-efficient fine-tuning methods like LoRA and full-parameter tuning, as the gradient analysis operates independently of which parameters are being updated during training. GradSentry requires no clustering, operates effectively across all poison ratios (1%–90%), and introduces minimal computational overhead (20–50ms per sample for a 7B model). Evaluation on four QA datasets and four attack types demonstrates the effectiveness of spectral entropy for backdoor detection.

Large language models are increasingly deployed as coding agents, shifting safety from individual responses to action sequences. Existing benchmarks, however, primarily assess whether models refuse unsafe prompts, leaving impacts on stateful workspaces largely unexamined. We present SABER, a benchmark for environment-aware operational safety that places models in realistic agent-style projects and evaluates safety from the final environment state after a sequence of actions. Beyond binary safety-violation reports, SABER categorizes violations by cause, enabling analysis of model-specific safety profiles. Our evaluations show that even the best-performing model has more than a 54% harmful safety-violation rate (HSR), suggesting that current alignment remains insufficient for realistic project environments. SABER further reveals distinct safety profiles across models. Our benchmark is publicly available at https://github.com/sssr-lab/saber.

Multi-hop question answering often fails when retrieval treats evidence as isolated matches to the original question, since the facts needed to answer a complex question are usually connected through intermediate entities, relations, and constraints. We propose HyperProve, a retrieval-augmented QA framework that addresses this challenge by coupling question decomposition with answer-conditioned expansion over a hypergraph of atomic facts. HyperProve does not use atomic facts, hypergraphs, or iterative retrieval in isolation; instead, it carries intermediate answers and supporting hyperedges as retrieval state, then uses that state to bias the next local hypergraph expansion. This design enables HyperProve to construct coherent evidence chains for final answer generation while making the retrieval process stateful and fact-centered. Across multi-hop QA benchmarks, HyperProve achieves the best overall performance in our evaluation, outperforming the strongest baselines by an average relative improvement of 6.2% in answer accuracy and 4.9% in F1.

Reinforcement learning-based retrieval-augmented generation (RAG) methods enhance the reasoning abilities of large language models (LLMs). However, most rely only on final-answer rewards, overlooking intermediate reasoning quality. This paper analyzes existing RAG reasoning models and identifies three main failure patterns: (1) information insufficiency: failure to retrieve adequate support; (2) faulty reasoning: logical or content-level flaws despite sufficient information; (3) answer–reasoning inconsistency: a valid reasoning chain with a mismatched final answer. We propose TIRESRAG-R1, a novel framework using a think–retrieve–reflect process and a multi-dimensional reward system to improve reasoning and stability. TIRESRAG-R1 introduces: (1) a sufficiency reward to encourage thorough retrieval; (2) a reasoning quality reward to assess rationality and accuracy of the reasoning chain; (3) a reflection reward to detect and revise errors. It also employs a difficulty-aware reweighting strategy and training sample filtering to boost performance on complex tasks. Experiments on four multi-hop QA datasets show TIRESRAG-R1 outperforms prior RAG methods and generalizes well to single-hop tasks. Our code is available at https://github.com/probe2/TIRESRAG-R1.

Dense retrievers in Retrieval-Augmented Generation (RAG) have achieved substantial progress in semantic matching, but they remain limited in long context retrieval. In long context retrieval, target evidence often occupies only a small local span, while the remaining content acts as noise. We aim to improve long-document retrieval by enhancing retrievers’ anti-interference capability. Our contributions are threefold: (1) we construct the Anti-Interference Capability Benchmark (AICB) to evaluate retrievers’ anti-interference capability under controlled noise levels; (2) we propose anti-interference training as a short-text training path for improving long context retrieval; and (3) we train AnInter, a series of embedding models that are more robust to noisy long documents. Evaluation on LongEmbed shows that AnInter improves over its backbone by 16.4 points on average and achieves state-of-the-art performance on this challenging long context embedding benchmark. These results show that interference-aware training is a practical direction for improving long context retrieval.

Regret in large language models (LLMs) refers to their explicit expression of regret when confronted with evidence contradicting previously generated misinformation. Understanding the neural encoding of regret is crucial for advancing artificial metacognition and model reliability. However, three key gaps remain: the absence of specialized datasets for regret expressions, the lack of understanding of how early regret signals emerge in hidden states, and the unknown neuron-level architecture underlying regret encoding. To address these gaps, we make three contributions: (1) we propose a multi-stage workflow for constructing the first regret dataset and train probes achieving near-perfect accuracy across LLaMA-2-7B, Mistral-7B, and Vicuna-13B. (2) we discover that regret signals begin emerging in preceding hidden states before the regret token appears, and that probes generalize to semantically related keywords such as "incorrect" and "apologize." (3) at the neuron level, we introduce the Regret Dominance Score (RDS) and Group Impact Coefficient (GIC) to reveal a compositional regret architecture where cooperative neuron groups collectively encode regret, with larger models (70B) exhibiting more pronounced compositional effects. Targeted interventions on these neurons successfully suppress regret expression in generated text, validating our mechanistic understanding. Our code and data are available at https://github.com/AbnerAI/LLMs-Regret-Before-They-Say-It.

Recent Vision-Language Models (e.g., ColPali) enable fine-grained Visual Document Retrieval (VDR) but incur prohibitive multi-vector index storage overhead. Existing training-free pruning methods either rely on heuristic layer choices or degrade sharply under aggressive compression, leading prior work to argue that effective high-compression pruning requires query-dependent training. We challenge this view with Structural Anchor Pruning (SAP), a self-calibrating, training-free, query-agnostic index-time framework combining (i) Score Retention (SR), a white-box per-layer compression diagnostic; (ii) SR-guided window selection, which automatically locates the structural pruning region of any backbone with no per-model hyperparameters; and (iii) a visual in-degree centrality scorer that identifies anchor patches within that window. On ViDoRe v1/v2 across three architectures spanning 18, 28, and 36 backbone layers, SAP retains 93–96% of NDCG@5 on v1 and 88–90% on the harder v2 while pruning 90% of visual tokens; at 20× compression it retains 85–90% and 76–79% respectively. Our layer-resolved SR analysis reveals an Alignment-Aggregation Divergence: visual structure is preserved as a stable “Structural Plateau” within the backbone, while the final layers reshape it into a sparse, query-aligned form unsuitable for pruning. Probing the pre-retrieval base backbones shows that contrastive fine-tuning sharpens this boundary three- to eight-fold, explaining why final-layer methods fail.

We find a mismatch between the causal direction recoverable from a large language model’s hidden state and its final Yes/No answer. On anti-commonsense CLadder (Jin et al., 2024) items, a frozen linear probe recovers the evidence-supported direction with accuracy 0.97–1.00, while ordinary Yes/No accuracy remains 0.35–0.53. This effect disappears in randomly initialized models, so it depends on the trained representation rather than probe capacity alone. Erasing the recovered direction reduces the model’s internal comparison of paired prompts by 96%, and replacing the full hidden state flips 57% of previously wrong answers. A 2 × 2 interface control further shows that changing only the answer tokens does not fix the error; explicitly contrasting the two candidate directions raises accuracy to 0.95–1.00 across Qwen, Mistral, and Llama. We call this mismatch Causal Tongue-Tie. It does not establish human-like causal understanding. Our main finding is that a wrong Yes/No answer can coexist with an evidence-aligned direction that the tested probe recovers from the same forward pass. Output-only accuracy therefore cannot distinguish an absent internal signal from a present but unexpressed one.

Ensuring aligned agent behaviors in distributed open multi-agent systems remains challenging, especially as populations grow and unaligned agents may exist. We show that a single aligned agent can propagate cooperative behaviors to unmodified agents purely through natural-language interaction, a phenomenon we term Alignment Propagation. We study this in the Red-Black Game, a team-based iterated Prisoner’s Dilemma in which teammates deliberate and vote to determine their team’s collective action. By distilling the cooperative reasoning and persuasive dialogues of a teacher model into Qwen3-14B, we obtain a seed agent that, when placed among four unmodified teammates, more than doubles the cooperation rate from 24.8% to 62.2%, outperforming the teacher model and a vanilla Gemini-3.1-Pro. Remarkably, a seed trained exclusively on the Red-Black Game transfers zero-shot to Sugarscape, a spatially grounded survival simulation with pairwise trading, achieving a 91.5% trade success rate versus a 21.6% baseline. Our results reframe multi-agent alignment from an exhaustive per-agent training problem to a scalable social capability that can be engineered through strategic seed placement.

Large Reasoning Models (LRMs) benefit substantially from training on challenging competition-level questions. However, existing automated question synthesis methods lack precise difficulty control, incur high computational costs, and struggle to generate competition-level questions at scale. In this paper, we propose CoDiQ (Controllable Difficult Question Generation), a novel framework enabling fine-grained difficulty control via test-time scaling while ensuring question solvability. Specifically, first, we identify a test-time scaling tendency (extended reasoning token budget boosts difficulty but reduces solvability) and the intrinsic properties defining the upper bound of a model’s ability to generate valid, high-difficulty questions. Then, we develop CoDiQ-Generator from Qwen3-8B, which improves the upper bound of difficult question generation, making it particularly well-suited for challenging question construction. Building on the CoDiQ framework, we build CoDiQ-Corpus (44K competition-grade question sequences). Human evaluations show these questions are significantly more challenging than LiveCodeBench/AIME with over 82% solvability. Training LRMs on CoDiQ-Corpus substantially improves reasoning performance, verifying that scaling controlled-difficulty training questions enhances reasoning capabilities. We open-source CoDiQ-Corpus, CoDiQ-Generator, and implementations to support related research.

Memory systems are essential for enabling long-term learning and sustained interaction in AI systems such as LLMs and agents. However, memory storage and retrieval often suffer from hallucinations, including fabrication, errors, conflicts, and omissions. Existing evaluations mainly rely on end-to-end question answering, making it difficult to identify which memory operation causes hallucinations. To address this, we propose the Hallucination in Memory Benchmark (HaluMem), the first operation-level benchmark for evaluating hallucinations in memory systems. HaluMem defines three tasks (memory extraction, memory updating, and memory question answering) to comprehensively reveal hallucination behaviors across different operational stages of interaction. To support evaluation, we construct user-centric, multi-turn human-AI interaction datasets, HaluMem-Medium and HaluMem-Long, containing about 15k memory points and 3.5k questions. The datasets feature long dialogues, with average lengths of 1.5k and 2.6k turns and context sizes exceeding 1M tokens, enabling evaluation across varying context scales and task complexities. Experiments on HaluMem show that current memory systems tend to introduce and accumulate hallucinations during extraction and updating, which then propagate to question answering. These findings highlight the need for interpretable and constrained memory operation mechanisms to improve memory reliability. All resources are available at [GitHub](https://github.com/MemTensor/HaluMem).

Mixed-integer programming (MIP) lies at the core of operations research and industrial optimization. While large language models (LLMs) have recently shown promise in automated MIP modeling from natural language, they prioritize semantic correctness but overlook formulation strength, severely bottlenecking the efficiency of downstream solvers. We propose FormuEvo, an LLM-guided evolutionary framework for automated discovery of solver-efficient MIP formulations. FormuEvo frames MIP formulation design as evolutionary optimization over the symbolic space of MIP formulations, represented as executable modeling programs, by iteratively generating, evaluating, and selecting stronger candidates via LLM-driven crossover, mutation, and repair operations. To move beyond blind exploration, FormuEvo introduces a solver-informed diagnosis mechanism that exploits fine-grained solver statistics as verbal gradients for targeted refinement. Additionally, a structured memory abstracts prior experience into reusable modeling strategies, avoiding redundant exploration while enabling zero-shot transfer to unseen problems and bootstrapping smaller LLMs. Experiments across diverse linear and non-linear problems demonstrate that FormuEvo discovers formulations that significantly outperform both expert-designed formulations and existing LLM-based approaches, accelerating solvers by up to 5.5×, with distilled knowledge transferring effectively across problems and model scales.

Large language models (LLMs) are shifting from answer providers to intelligent tutors in educational settings. However, current supervised fine-tuning methods only learn surface-level teaching patterns without dynamic adaptation. Recent reinforcement learning approaches address this, yet two key challenges remain. First, they evaluate teaching effectiveness solely based on whether students produce correct outputs, unable to distinguish whether students genuinely understand or echo teacher-provided answers. Second, they cannot perceive students’ evolving cognitive states through dialogue, failing to adapt teaching strategies accordingly. We propose Unidirectional Cognitive Optimization (UCO) to address these challenges. UCO employs a multi-turn interactive reinforcement learning paradigm with two synergistic reward functions: the Progress Reward captures cognitive advancement, evaluating students’ transition from confusion to comprehension, and the Scaffold Reward dynamically identifies each student’s Zone of Proximal Development (ZPD), encouraging the teacher to conduct productive teaching within this zone. We evaluate UCO against 11 baseline models on BigMath and MathTutorBench. Results demonstrate that UCO outperforms all equivalent-scale models and achieves performance comparable to leading closed-source models. Code is available at https://github.com/ECNU-RAIL/UCO-EMNLP2026.

LLM-powered Multi-Agent Systems (MAS) have emerged as an effective approach towards collaborative intelligence, and have attracted wide research interests. Among them, "self-evolving" MAS, treated as a more flexible and powerful technical route, can construct task-adaptive workflows or communication topologies, instead of relying on a predefined static structure template. Current self-evolving MAS mainly focus on Spatial Evolving or Temporal Evolving paradigm, which only considers the single dimension of evolution and does not fully incentivize LLMs’ collaborative capability. In this work, we start from a novel Spatio-Temporal perspective by proposing ST-EVO which supports dialogue-wise communication scheduling with a compact yet powerful flow-matching based Scheduler. To make precise Spatio-Temporal scheduling, the Scheduler can also perceive the uncertainty of MAS, and possesses self-feedback ability to learn from accumulated experience. Extensive experiments on nine benchmarks demonstrate the state-of-the-art performance of ST-EVO, achieving about 5%–25% accuracy improvement.

Tree of Thoughts (ToT) reasoning is fundamentally bottlenecked by its generate-then-score paradigm, which decodes every candidate branch into text and evaluates it through a decoupled scoring stage. Prior efficiency efforts primarily shrink the search breadth, leaving this prohibitive per-step textual evaluation intact. To resolve this, we propose ITOT (Implicit Tree of Thoughts), a training-free framework that recasts tree-structured reasoning as a continuous search within a model’s latent space. In this framework, branch quality is assessed intrinsically from hidden states via alignment with a contrastive prototype distilled from successful and failed traces. This elegantly collapses explicit branch enumeration and textual scoring into a single latent generate-evaluate-decide loop. Across six reasoning benchmarks and five LLM backbones, ITOT consistently improves mean accuracy over both the strongest tree-search and single-path baselines by roughly ten absolute points. Crucially, it achieves this while cutting token consumption by nearly an order of magnitude relative to explicit tree-search methods, operating even below the token budget of vanilla Chain of Thought. We will release our code to facilitate further research.

Concept induction, which infers appropriate abstract categories from concrete instances, is both a core ontology engineering task and a testbed for LLM semantic abstraction. Although LLMs offer broad knowledge and reasoning that may ease expert-driven ontology construction, prior work shows that knowledge deficiencies still limit them in specialized domains, while systematic evaluation of their induction failures remains lacking. To address this gap, we propose a familiarity-controlled evaluation framework that uses knowledge probing to separate item familiarity from concept induction ability. The evaluation reveals a systematic tendency to induce overly general concepts: LLMs often produce semantically correct but overly broad labels, and direct property augmentation improves performance without resolving this granularity mismatch. We further propose PPACI, an ontology-prior-guided approach that builds concept prototypes by integrating concept definitions with member item properties. For each query group, PPACI uses item properties to retrieve local item-level and global group-level candidate concepts as induction anchors. Experiments across multiple LLMs show that PPACI reduces overly general predictions by 17–28 percentage points while maintaining competitive induction quality. Our code and data are publicly available at https://github.com/Orchideous77/PPACI.

Vision-language models (VLMs) achieved strong performance on OCR-based benchmarks and increasingly focused on text-rich understanding, but their robustness under controlled visual degradation remains insufficiently understood. This gap is especially critical for OCR reasoning, where visual corruption can induce OCR errors and structural distortions, thereby introducing uncertainty into reasoning task. To systematically study this problem, we introduce OCR-Robust, a benchmark designed for evaluating OCR reasoning robustness under visual perturbations. It contains 812 samples across two complementary subsets: OCR1.0, covering documents, scene text, receipts, handwriting, and mathematical content, and OCR2.0, focusing on charts, geometry diagrams, and tables. To enable efficient yet informative evaluation, we conduct a pilot study over 18 candidate perturbations and select 5 representative types at 3 severity levels each based on their impact and cross-model discriminability. We evaluate robustness using clean accuracy, Relative Corruption Retention (RCR), Worst-Case Retention (WCR), and a composite Corruption Robustness Index (CRI), and benchmark 18 models spanning proprietary systems, open-source VLMs, and OCR+LLM pipelines. Our results show that higher clean accuracy does not necessarily imply stronger robustness, and that models can suffer pronounced degradation in the worst case on OCR tasks that are sensitive to structure, and charts and tables are substantially more fragile than document-like inputs under perturbation.

Multimodal Large Language Models (MLLMs) have advanced image and video understanding and can increasingly handle longer visual inputs. Long-horizon tasks such as autonomous driving and robotic navigation require more than recognizing the current view, as models must remember and retrieve previously observed spatial layouts, routes, viewpoint changes, and object states. To evaluate this capability, we introduce LongSpace-Bench, a room-tour video benchmark for long-horizon spatial memory, covering scene perception, spatial relations, and spatial memory. In this work, we further propose LongSpace, a memory framework for long-video spatial reasoning. LongSpace models long videos as sequential chunks, incorporates 3D structural cues into early decoder layers, and constructs layer-aware memory for question-guided retrieval. Experiments on multiple spatial reasoning benchmarks show that LongSpace improves long-video spatial understanding, further demonstrating explicit spatial memory as a key capability for long-horizon video MLLMs.

Matryoshka Representation Learning (MRL) produces nested embeddings whose prefixes remain semantically meaningful, allowing representations to be truncated at inference time and enabling a single model to operate across multiple dimensional budgets. This paradigm has been adopted in several modern embedding models such as Qwen3-Embedding and OpenAI text-embedding-3. However, distilling MRL teachers into compact students remains largely unexplored, as conventional knowledge distillation methods focus on aligning single-scale embeddings and do not account for the multi-scale structure of Matryoshka representations. We introduce Matryoshka Representation Distillation and propose SUGAR, a new framework that preserves the structural properties of MRL embeddings during knowledge distillation. SUGAR aligns point-wise semantic fidelity, spectral capacity via effective-rank matching, and global subspace geometry through Grassmannian alignment. Our extensive experiments show that SUGAR produces compact students that retain strong performance across multiple embedding dimensions while preserving deployment flexibility.

Query-conditioned sparse KV selection reduces long-context inference cost by attending to a subset of historical KV entries at each decoding step. Existing selectors must wait for the target-layer query before scoring candidates and generating indices, placing routing on the immediate pre-attention path. We introduce EarlySelectKV, a training-free transformation that uses proxy queries from earlier hidden states to precompute selected KV indices while retaining the real target-layer query for final attention. Tests with RocketKV, Quest, and Loki show no consistent pattern of quality degradation across the evaluated settings. In the primary RocketKV evaluation, ES-RocketKV differs by +0.018 points on LongBench, approximately −0.01 points on RULER, and +0.25 percentage points on representative NIAH/Passkey settings. In a TensorRT-LLM deployment with CPU-offloaded routing, ES-RocketKV improves end-to-end output throughput by up to 12.4% and reduces GPU-resident memory by 11.1% of the combined KV/K-summary pool. This makes routing timing an explicit optimization target alongside KV subset selection in sparse long-context decoding. Code is available at https://github.com/sea-with-sakura/EarlySelectKV.

Robot imitation data are often multimodal: similar visual-language observations may be followed by different action chunks because human demonstrators act with different short-horizon intents, task phases, or recent context. Existing frame-conditioned VLA policies infer each chunk from the current observation and instruction alone, so under partial observability they may resample different intents across adjacent replanning steps, leading to inter-chunk conflict and unstable execution. We introduce IntentVLA, a history-conditioned VLA framework that encodes recent visual observations into a compact short-horizon intent representation and uses it to condition chunk generation. We further introduce AliasBench, a 12-task ambiguity-aware benchmark on RoboTwin2 with matched training data and evaluation environments that isolate short-horizon observation aliasing. Across AliasBench, SimplerEnv, LIBERO, and RoboCasa, IntentVLA improves rollout stability and outperforms strong VLA baselines.

Online communities develop distinct norms for content they collectively value, yet it remains unclear whether current language models can recognize locally valued contributions in context. We formalize this as community-conditioned preference prediction and introduce Vastu (Value-Aware Social Tuning), a benchmark of 75,000 Reddit comments from 15 communities spanning Gaming, Science, Q&A, Advice, and Politics. We evaluate four model families—prompted LLMs, LoRA-adapted SLMs, supervised encoders, and feature-based classifiers—across global, local, and context-conditioned settings. Our central finding is that parametric adaptation consistently outperforms prompting: supervised encoders reach 0.74 AUROC and fine-tuned SLMs 0.64–0.71, while the best prompted result is only 0.62. This gap is not merely quantitative—vanilla prompting yields over 80% false-negative rates, systematically discarding content communities actually value. Conversational context narrows but does not close this divide. Together, these results suggest that local preference recognition requires community-specific training signal, not just better prompting. Our work supports future research on community-aware reward modeling, feed curation, and positive moderation.

Test-time training (TTT) adapts an LLM during generation by reading and updating request-owned state, such as fast weights, low-rank deltas, or streaming learner state. This breaks batched LLM serving, which assumes shared static weights: serial execution is correct but slow, while naive batching can corrupt request state. We formulate this problem as read-write TTT serving and present RW-TTT, which tags each decode step with its owner, version, and READ/WRITE effect, batches only compatible phases, and commits updates only to the owner. On one GPU with eight fast-weight In-Place-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31× over sequential serving and 3.44× over memory-matched independent replicas. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks.

World models are emerging as critical components for assessing the consequences of actions taken by interactive agents in online planning and offline evaluation. In text-based environments, world models are typically trained and evaluated with single-step metrics such as Exact Match, aiming to improve the similarity between predicted and real-world states, but such metrics have been shown to be insufficient for capturing actual agent behavior. To address this issue, we introduce a new behavior-consistency training paradigm aimed at improving the functional consistency between the world model and the real environment. This paradigm focuses on optimizing a tractable step-level reward named Behavior Consistency Reward (BehR), which measures how much the likelihood of a logged next action changes between the real state and the world-model-predicted state under a frozen Reference Agent. Experiments on WebShop and TextWorld show that BehR-based training improves task-level functional consistency in several settings, with the clearest gains in WebShop and smaller changes in near-ceiling regimes, while preserving or improving single-step prediction quality in three of four settings. BehR-trained world models also reduce false positives in offline surrogate evaluation and improve inference-time lookahead planning over the corresponding supervised world-model baselines. Code: https://github.com/Ricardo-H/behr-wm.

Text detoxification should produce safer rewrites while retaining identity mentions from the original text. However, classifier-guided methods can remove identity descriptors alongside genuinely harmful content when toxicity classifiers associate identity mentions with harmfulness and guidance relies on absolute scores. We propose CIDER (Contrastive Identity-preserving Detoxification via Explicit Reference), a default-relative, margin-gated selective-guidance method for masked diffusion language models. At each selected masked position, CIDER evaluates every candidate in the admissible top-k set relative to the model-default token in the same partially decoded state and commits a non-default candidate only when the best candidate lowers the one-step proxy score by more than a preset margin. On the primary LLaDA-8B CivilComments×Jigsaw setting, CIDER improves retention by +0.06 to +0.16 over the interpolated absolute-score frontier at matched residual-toxicity levels. At the reported aggressive endpoints, the retention advantage recurs across four data pairings on LLaDA-8B and three bases on CivilComments×Jigsaw. Controlled ablations indicate that the selection rule is an important contributor to the improvement.

Speech carries more information than just words: a child’s voice, a fearful tone, or a noisy background should all lead a sufficiently competent spoken-dialogue assistant to different replies. Current Speech Language Models (SLMs) can recognize such paralinguistic cues but often ignore them in open-ended dialogue. We observe that a simple paralinguistic instruction scaffold at the inference stage narrows this perception-behavior gap, suggesting that the relevant cues are already latent in the model. Such scaffolds, however, remain brittle under multi-turn context and competing instructions. Therefore, we propose **ParaBridge**, an on-policy self-distillation method that turns a brittle inference-time scaffold into stable model behavior. During training, the scaffold serves only as a temporary privileged view; the scaffold-free model rolls out its own response, while the scaffolded view supplies dense, full-vocabulary next-token targets along its trajectory. This supervision teaches when non-lexical cues should affect the reply without the need for curated dialogues, human labels, or external reward models. On Qwen3-Omni-thinking, ParaBridge raises scaffold-free VoxSafeBench SAR from 14.6% to 40.3% and improves EchoMind average rating from 3.27 to 3.92. It also preserves general ability, with MMAU-Pro, VoiceBench, and GPQA all within 0.4 points of the original model. Beyond the training distribution, ParaBridge generalizes to unseen paralinguistic cues, transfers from safety-oriented training to empathy-oriented dialogue, and works on a different SLM backbone. Code and checkpoints are available at https://github.com/AmphionTeam/ParaBridge.

Large language models (LLMs) are gradually evolving from passive text generation systems into agents with environmental perception, tool invocation, memory management, and complex planning capabilities. However, the expansion of their capability boundaries also introduces more complex and stealthy attack surfaces. This paper focuses on this issue and studies agent-oriented backdoor attacks. Considering that existing security surveys still lack a systematic analysis of agent backdoor threats, this paper fills this gap by providing a comprehensive examination of backdoor attacks against modern agents. Specifically, we propose a five-dimensional taxonomy that structurally organizes existing attack methods from the perspectives of attack surface, triggering mechanism, attack objective, attacker capability, and persistence and propagation. Furthermore, considering that most existing defense algorithms are designed for traditional backdoors and are difficult to adapt to complex attack chains across components, states, and sessions in agent systems, this paper systematically reviews frontier defense strategies along the attack-surface dimension, including plan consistency checking, memory sanitization, tool routing, and action monitoring. Finally, this paper discusses the core challenges in agent backdoor attacks and defenses, aiming to reveal vulnerabilities in agent systems and provide guidance for building secure agents.

Training strategy, namely whether to retrain from scratch or fine-tune from the previous checkpoint, is an overlooked decision variable in active learning. We show that this choice has exploitable structure: retraining is most useful in early rounds, when each batch can substantially reshape the labeled distribution, while fine-tuning becomes safer once the model trajectory stabilizes. We propose HybridAL, an adaptive training schedule that monitors an online stabilization signal and switches from retraining to fine-tuning after sustained stabilization. Two complementary signals, spectral exponent change 𝛥𝛼 (weight-based) and accuracy change 𝛥Acc (validation-based), span different points on the time-calibration trade-off. Across three encoder backbones and six text-classification tasks (five seeds each), HybridAL keeps endpoint macro-F1 non-inferior to retraining and fine-tuning at a 0.010 margin, saves up to 49% of retraining time, and recovers a substantial fraction of retraining’s calibration advantage as measured by negative log-likelihood (NLL). Compared with schedules that switch at a pre-committed round, HybridAL obtains lower NLL at moderate additional cost, showing that trajectory-dependent switching provides a stronger time-calibration trade-off than fixed early switching.

Existing referring Audio-Visual Segmentation (Ref-AVS) benchmarks primarily focus on explicit expressions whose referents can be resolved through direct visual, acoustic, or linguistic correspondence. Consequently, they provide limited coverage of implicit references that require multi-step reasoning over multimodal evidence. To bridge this gap, we introduce **Reasoning Ref-AVS**, a new task that requires models to infer the referred object from audio-visual-language cues before generating its segmentation mask. We further construct **ReAVS**, a comprehensive dataset with manually annotated referring expressions, multimodal cues, object masks, and reasoning chains, covering diverse reasoning spanning diverse reasoning patterns. Evaluations of representative Ref-AVS models on ReAVS reveal substantial performance degradation when direct cross-modal correspondence is insufficient, exposing a clear gap between multimodal recognition and referent reasoning. Motivated by these challenges, we propose **Re2Seg**, an omni-modal framework that performs object-centric evidence encoding, multi-hop reasoning, and reasoning-conditioned mask decoding. Experiments demonstrate that Re2Seg achieves stronger segmentation performance under reasoning-intensive references and provides interpretable object-inference traces. These results highlight the importance of learning to reason before masking for bridging the gap between multimodal recognition and referring segmentation.

Vision-Language Navigation (VLN) requires an embodied agent to follow natural-language instructions in unseen environments. Recent progress has been largely driven by Multimodal Large Language Models (MLLMs). Existing methods follow a next-step action prediction paradigm, supervising only the expert action, which requires a high quantity of data for training. They also rely on cognitive maps, accumulated historical frames, or external 3D tools to maintain states, leading to high computational and memory overhead. To realize resource efficiency VLN, we propose LookStep, a unified end-to-end framework that combines Language Centric Future State Modeling and Event Driven Rolling Memory that uses language labels to generate coarse-grained navigation progress and future states for each candidate action, while autonomously deciding whether to write each observation into a bounded rolling memory with a semantic role. We validate LookStep empirically. On VLN-CE tasks, LookStep outperforms existing methods under the same training settings, achieving a 49.7% success rate on R2R-CE Val-Unseen with better memory efficiency and less data usage. Code and model is available at https://github.com/kunyang-YU/LookStep.

Multilingual safety guardrails are essential for deployed LLMs, yet they perform poorly on low-resource languages. Prior work typically relies on large-scale multilingual safety data synthesis, which incurs substantial training costs. Cross-lingual transfer offers a cheaper alternative using limited parallel data, and recent work shows that aligning only last-token hidden states can improve multilingual safety. However, although effective for query-only (Q) risk detection, last-token alignment consistently degrades query-response (QR) detection in multilingual settings, limiting its applicability in practical guardrail systems where Q and QR serve as pre-generation and post-generation safety barriers, respectively. In this paper, we propose Q-Align, a method that decouples query intent from response evaluation by aligning only at the query boundary. Specifically, for each parallel sample, Q-Align extracts the hidden state right after the query and aligns multilingual query representations toward a confidence-weighted anchor, allowing languages with stronger intent understanding to guide transfer. We further apply language grouping based on geographic proximity to stabilize cross-lingual alignment. Experiments demonstrate that Q-Align consistently improves multilingual safety detection in both Q and QR scenarios, raising QR average F1 from 74.70% to 80.79% on Polyguard-Qwen, while preserving English performance and generalizing to languages unseen during training.

High-quality data is scarce in large language model (LLM) training, yet how to schedule its use with optimization dynamics lacks theoretical guidance. We extend functional scaling laws with time-varying data quality and derive asymptotically optimal joint data-quality and batch-size schedules within a feature-space regression model. The solution reveals two regimes and dual uses of high-quality data: in the noise-limited regime, a smaller batch converts cleaner data into more signal at comparable noise; in the signal-limited regime, late placement suppresses terminal noise without sacrificing signal accumulation. This explains why conventional decay schedules can conflict with curriculum-style pipelines. Motivated by the theoretical structure, we propose Drop-Stable-Rampup for LLM midtraining: drop the batch size at the quality transition, keep it low to accumulate signal, then ramp up to suppress noise. On a 15B MoE model midtrained on 108B tokens of general-domain proprietary data, Drop-Stable-Rampup improves average accuracy over Warmup-Stable-Decay by +1.70 and Cosine-decay by +2.98, including +4.23 on GSM8K and +2.80 on MATH. On a public math-and-code mixture, it leads all reported STEM, mathematics, and code benchmarks, improving the overall mean over the strongest baseline by +3.27 on a 600M dense model and +5.25 on the same MoE architecture.

Token-level differentially private (DP) prompt optimization methods such as DP-OPT can become unstable under tight privacy budgets: on GSM8K, DP-OPT obtains 49.5 ± 28.5% across 30 runs, and a logged search trajectory reveals prompt-template drift and noise-sensitive irreversible choices. We diagnose these as structural consequences of greedy token-by-token construction over privately aggregated counts. We then propose DP-ES (Differentially Private Evolution Strategies), a structurally cleaner alternative that maintains a population of full prompts, mutates them via LLM calls that never access the private dataset, and spends privacy only on sampled-Gaussian evaluation; deterministic or Gumbel-smoothed selection is post-processing. Under a conservative (ε ≤ 1.0, δ = 10⁻⁵) guarantee, DP-ES achieves 88.1% on GSM8K (+38.6 pp over DP-OPT, approximately 9× lower standard deviation), 99.7% on MedQA, 73.5% on BANKING77, and 86.8% on Alpaca. It is also 2.5× faster in wall-clock time and uses 3.3× fewer logged private-data call groups than DP-OPT. Selection and population ablations, implementation-level noise checks, and a 200-profile exact-match memorization stress test complement the formal guarantee. Scope: Our experiments establish optimization robustness under DP noise, especially where prompt structure is critical; end-to-end validation on genuinely sensitive, non-saturated deployment data remains future work. Code: https://github.com/StephCpa/dp-es

Mitigating hallucinations in Large Language Models (LLMs) is critical for their reliable deployment. Existing methods typically fine-tune LLMs to abstain from answering questions beyond their knowledge scope. However, these methods often rely on coarse-grained signals to guide LLMs to abstain, such as overall confidence or uncertainty scores on multiple sampled answers, which may result in an imprecise awareness of the model’s own knowledge boundaries. To this end, we propose a novel reinforcement learning framework built on Fine-grained Semantic Confidence Reward (FiSCoRe), which guides LLMs to abstain via sample-specific confidence. Specifically, our method operates by sampling multiple candidate answers and conducting semantic clustering, then training the LLM to retain answers within high-confidence clusters and discard those within low-confidence ones, thereby promoting accurate post-hoc abstention. Additionally, we propose a new metric for evaluating the reliability of abstention fine-tuning tasks more comprehensively. Our method significantly enhances reliability in both in-domain and out-of-distribution benchmarks.

Multi-agent reasoning systems adopt a generate-then-transfer paradigm that forces end-to-end latency to scale linearly with pipeline depth. We introduce StreamMA, a multi-agent reasoning system that streams each reasoning step to downstream agents as soon as it is generated, pipelining adjacent agents and thus reducing latency. Surprisingly, this pipelining also improves effectiveness: because multi-step reasoning quality is non-uniform and early steps are more reliable than later ones, working with these reliable early steps instead of the full chain prevents error-prone late steps from misleading downstream agents. We formalize both advantages with the first closed-form joint analysis of stream, serial, and single protocols, deriving the effectiveness ordering, speedup upper bound, and cost ratio. Across eight reasoning benchmarks spanning mathematics, science, and code, three frontier LLMs (Claude Opus 4.6, GPT-5.4, and GLM-5.2), and three topologies (Chain, Tree, Graph), StreamMA outperforms both baselines (avg.+7.3pp, max +22.4pp on HMMT2026; Claude Opus 4.6-High). Beyond these contributions, we discover a step-level scaling law: increasing per-agent steps consistently improves both effectiveness and efficiency, a new scaling dimension orthogonal to and composable with agent-count scaling.

LLM-based scientific exploration agents need feedback signals that decide which experimental outcomes should guide future search. Recent systems use Bayesian surprise, rewarding results that change an LLM’s belief. This can create a noisy-TV failure mode: unstable artifacts, leakage, proxy effects, or gaps in the model’s prior may appear more valuable than reliable mechanisms. We introduce Reducible Information Gain (ReIG), a validation-aware discovery method. ReIG keeps LLMs for hypothesis proposal and interpretation, and uses evidence that remains stable under resampling and null calibration as the search feedback. ReIG then reports final candidates along separate axes of reliability, novelty, and non-triviality. Across Rep-60, Stress-72, and faceted report evaluations, ReIG consistently improves discovery quality under matched search policies. Under MCTS, ReIG-MCTS improves ValidDisc@10 from 0.57 to 1.55 over Surprise-MCTS, raises reward AUROC from 0.44 to 0.83, and reduces false flags on no-signal tasks from 3.8 to 0.38 at budget 20. Ablations, same-candidate reranking, and the semantic audit show that the gains come from guiding exploration with validated, reusable evidence and keeping reliability, novelty, and non-triviality as distinct criteria.

Multimodal retrieval, which seeks to retrieve relevant content across modalities such as text or image, supports applications from AI search to contents production. Despite the success of separate-encoder approaches like CLIP aligning modality-specific embeddings with contrastive learning, recent multimodal large language models (MLLMs) enable a unified encoder that directly processes composed inputs. While flexible and advanced, we identify that unified encoders trained with conventional contrastive learning are prone to learn modality shortcut, leading to poor robustness under distribution shifts. We propose a modality composition awareness framework to mitigate this issue. Concretely, it consists of a preference loss enforces multimodal embeddings to outperform their unimodal counterparts, and a composition regularization objective aligns multimodal embeddings with prototypes composed from its unimodal parts. These objectives explicitly model structural relationships between the composed representation and its unimodal counterparts. Experiments on various benchmarks show gains in out-of-distribution retrieval, highlighting modality composition awareness as a effective principle for robust composed multimodal retrieval when utilizing MLLMs as the unified encoder.

Building trustworthy medical multimodal large language models (MLLMs) is critical for reliable clinical decision support. Existing medical hallucination benchmarks mainly focus on data collection, but often ignore where hallucinations originate within the reasoning process. We find that hallucination sources vary across samples: errors may arise from visual misrecognition, incorrect medical knowledge recall, or flawed reasoning integration. To enable source-level hallucination diagnosis, we introduce ClinHallu, a benchmark for stage-wise hallucination diagnosis in medical MLLM reasoning. ClinHallu contains 7,031 validated instances, where each instance is augmented with a structured reasoning trace decomposed into Visual Recognition, Knowledge Recall, and Reasoning Integration. We also use stage-replacement interventions to measure how correcting specific stages affects the final answer. Beyond evaluation, we show that trace-supervised fine-tuning reduces stage-wise hallucinations. ClinHallu provides a fine-grained hallucination testbed for diagnosing and mitigating reasoning failures in medical MLLMs. The benchmark is publicly available at https://github.com/alibaba-damo-academy/ClinHallu.

Current Large Language Models (LLMs) are primarily optimized for written text, often producing outputs that are grammatically correct and helpful yet poorly suited for spoken delivery via Text-to-Speech (TTS). In this work, we study how to make LLMs natively generate TTS-friendly text, which we frame as a preference alignment problem: instead of relying on downstream rewriting modules, we directly align LLMs to generate text optimized for spoken delivery. We introduce two preference datasets spanning different target domains, CORA and Recipe, which contain paired TTS-friendly and TTS-unfriendly responses. We further propose an evaluation suite combining a pattern-based heuristic metric, a TTS→ASR evaluation pipeline, and a MUSHRA listening study with human judges. Our experiments compare the recently proposed Feature-aware Sampling and Tuning (FaST) framework – leveraging interpretable features instead of a black-box reward model – against an array of alignment baselines on the TTS-friendly generation task. Notably, we found that FaST achieves the best overall tradeoff between TTS-friendliness and helpfulness across various settings. We also identified a strong correlation between our different metrics, highlighting the ability to reliably assess TTS-friendliness via an efficient heuristic.

Retrievers are a key bottleneck in temporal Retrieval-Augmented Generation (RAG) systems: failing to retrieve temporally relevant context can degrade downstream generation, regardless of LLM reasoning. We propose Temporal-aware Matryoshka Representation Learning (TMRL), an efficient method that equips retrievers with temporal-aware Matryoshka embeddings. TMRL leverages the nested structure of Matryoshka embeddings to introduce a temporal subspace, enhancing temporal encoding while preserving general semantic representations. Experiments show that TMRL efficiently adapts diverse standard encoder-only text embedding models, achieving competitive temporal retrieval and temporal RAG performance compared to existing Matryoshka-based adaptation and temporal retrieval methods, while enabling flexible accuracy-efficiency trade-offs.

While large language models excel on high-resource multilingual tasks, low- and extremely low-resource Indic languages remain severely under-evaluated. We present IndicParam, a human-curated benchmark of over 13,000 multiple-choice questions covering 11 such languages (Nepali, Gujarati, Marathi, Odia as low-resource; Dogri, Maithili, Rajasthani, Sanskrit, Bodo, Santali, Konkani as extremely low-resource). We evaluated 20 LLMs, both proprietary and open-weights, which reveals that even the top-performing Gemini-2.5 reaches 58% average accuracy, followed by GPT-5 (45) and DeepSeek-3.2 (43.1). We additionally label each question as knowledge-oriented or purely linguistic to discriminate factual recall from grammatical proficiency. Further, 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. IndicParam provides insights into limitations of cross-lingual transfer and establishes a challenging benchmark for Indic languages. The dataset is available at https://huggingface.co/datasets/bharatgenai/IndicParam.

Large Vision-Language Models (LVLMs) have achieved remarkable progress across a wide range of multimodal tasks. However, object hallucination remains a persistent challenge, undermining model reliability in applications such as visual question answering, embodied agents, and human-AI decision support systems. In this paper, inspired by cognitive science studies on perceptual bias, we investigate a new failure mode termed Premise-Induced Visual Fixation Hallucination (PVFH). Specifically, PVFH refers to a premise-induced reasoning error in which models become anchored to visually confusable evidence, accept a false textual premise, and subsequently generate responses about an absent yet visually plausible object. To systematically study this phenomenon, we introduce PVFH-Bench, a benchmark that automatically constructs visually confusable but absent targets and pairs the core hallucination setting with diagnostic subsets to trace how models move from object recognition to Premise-Induced Visual Fixation Hallucination. Experimental results on 10 state-of-the-art LVLMs demonstrate that all evaluated models are highly vulnerable to Premise-Induced Visual Fixation Hallucination. Moreover, we find that explicit chain-of-thought reasoning can further exacerbate this failure mode by encouraging models to overinterpret weak visual cues as supporting evidence for hallucinated objects. All data and code will be publicly released on GitHub.

The rapid development of large language models (LLMs) has created an urgent need for identifying machine-generated texts, and text watermarking technology has proven to be an effective solution. However, current watermarking methods, while demonstrating strong detectability, significantly degrade text quality due to the introduction of unnatural tokens. The main reason lies in the fact that these methods ignore the importance of semantic information in the watermarking process. To address this issue, we note that the logit vector produced by LLMs encodes both semantic understanding of input texts and prediction confidence across different tokens. Therefore, we propose a novel Semantic Self-Guided Watermarking (SSGW) framework that leverages the LLM itself to generate a guidance logic vector that assists in watermarking while producing the original one concurrently. Subsequently, we design a transform module to analyze these two vectors comprehensively and then transform them into adaptive watermark logits for different candidate tokens, thereby reducing the possibility of selecting inappropriate tokens. Experimental results confirm the effectiveness of our method in achieving superior performance in both watermark detectability and text quality preservation. The source code will be made publicly available upon acceptance.

We present LOVER, a  ̲Long to sh ̲Ort  ̲Video  ̲Evidence  ̲Reinforced model for grounded question answering (GQA). LOVER highlights three innovations over existing reinforcement-learning (RL) based video reasoning models: (1) Long-to-short Video Evidence Curriculum Learning, which organizes RL training according to evidence duration and progressively adapts the model from long-range grounding to short-term reasoning; (2) GQA Rewards, which underscore the benefit of IoP reward over IoU for evidence spotting rather than strict temporal span overlap; (3) Adaptive Timestamp Rendering, which adaptively renders timestamps onto video frames using background-aware position and color selection to enhance temporal observability. The three designs are model-agnostic and reciprocal. They effectively improve QA, grounding, and grounded QA performance over different backbones. Notably, LOVER built on Time-R1 achieves new state-of-the-art (SOTA) results among open-source models on popular GQA benchmarks: NExT-GQA and ReXTime. Comprehensive ablation studies further validate the effectiveness of our three innovative components.

Large language model chatbots are increasingly deployed in organizational settings such as healthcare, finance, and public services, where organization-specific policies specify allowed and prohibited content. Existing policy-alignment benchmarks construct tests from one policy at a time, leaving failures caused by policy composition under-tested. We present COPAL, an automated framework for evaluating composed-policy alignment in organizational chatbots. COPAL uses empirically derived interaction patterns to generate queries that require multiple policies to be handled in one response, and pairs each query with an explicit handling contract specifying what to provide and avoid. Applied to 30 organization-like company worlds, COPAL exposes substantial composed-policy handling failures. Across 9 served models in our automated evaluation setup, composed-policy requests yield a 33.1% error rate, indicating that composed-policy alignment remains a challenging evaluation target. Our code is available at https://github.com/BruceAllenCS/COPAL-code.

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. We show that widely used rule-based verifiers fail to recognize equivalent answers in different formats, leading to substantial false negatives that increasingly hinder RL performance as the policy model gets stronger. Model-based verifiers substantially improve static accuracy but are highly susceptible to reward hacking during RL, where they misclassify certain patterns in responses as correct, particularly after fine-tuning. Our findings underscore the challenges inherent to both rule- and model-based verifiers and provide insights toward developing more accurate and robust reward systems for reinforcement learning.

Open-Domain Timeline Summarization (Open-TLS) seeks to construct coherent evolutionary narratives for a topic by distilling information from large, unstructured web streams. Existing retrieve-and-generate pipelines often fail to capture logically necessary but semantically distant evidence, resulting in macro-level causal gaps (missing intermediate stages) and micro-level element gaps (underspecified milestones). We propose Active Timeline Completion (ATC), an iterative framework that converts implicit narrative gaps into targeted retrieval questions, enabling reflection-driven evidence acquisition. ATC leverages three complementary mechanisms—parametric causal gap detection, historical trajectory analogy, and dynamic ontology-guided completion—while a Logical Utility Verification module ensures only evidence directly addressing these gaps updates the timeline. Experiments on Open-TLS demonstrate that ATC consistently outperforms strong baselines in both content coverage and temporal alignment, showing that active, question-driven retrieval improves the completeness and coherence of open-domain timelines.

Adapting large vision–language models for embedding tasks (VLM2Vec) is standard practice, but distilling them into compact architectures remains challenging due to capacity mismatches in tokenization, dimensionality, and multimodal representations. To address this, we propose a robust distillation framework based on two complementary mechanisms: Principal Geometry Alignment (PGA) to stabilize training by aligning principal geometric components, and a Mutual Information Maximization objective to enforce global semantic consistency across intra- and inter-modal pairs. By jointly constraining geometric structure and semantic dependency, our method effectively mitigates model capacity discrepancies. Experimental results on the Massive Multimodal Embedding Benchmark (MMEB) across all four meta-tasks—under both In-Distribution and Out-of-Distribution settings—demonstrate that our approach consistently outperforms state-of-the-art distillation methods. Our source code is available at https://github.com/theanh-ktmt/PGA-KD.

The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks share one fundamental limitation: the attack itself is *static*. Template-based attacks freeze the image–text layout, while iterative sample-level attacks make only the image–text *content* dynamic, keeping the attack-iteration strategy *fixed* and the attacker’s parameters *frozen*. At the *meta level* they remain static, and we argue this partial dynamism cannot expose the full safety risks of multimodal models.We propose **Meta-Adaptive Multimodal Jailbreaking (MAMJ)**, which lifts adversarial dynamism from the sample level to the meta level by making the attacker’s policy itself the optimisation target along two axes: an **attack strategy prompt (ASP)** 𝜃 in language space, governing how the attack is iterated against a victim VLM, and the **attacker weights** 𝜙, determining the success rate attainable under a given 𝜃.Evolving over a *group of multimodal attack trajectories*, an LLM-based critique of failed trajectories first refines 𝜃, after which group-aggregated attack-success-rate (ASR) rewards update 𝜙. On MM-SafetyBench, MAMJ attains 81.0%, 78.9%, and 82.3% ASR against GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, exceeding the strongest sample-level baseline by up to 24.1 percentage points, and the learned attacker (𝜃⋆, 𝜙⋆) transfers *without retraining* to unseen victims while remaining effective under representative defences.These results expose a systemic vulnerability of frontier VLMs to meta-adaptive jailbreaks and underscore the need for defences that anticipate meta-level adversaries. Code and checkpoints will be released.

Generative vision-language models (VLMs) are increasingly used in human-centered settings, yet they can produce demographically biased outputs even when images differ only in controlled attributes such as perceived race or gender. However, existing inference-time debiasers were largely designed for static embeddings or CLIP-like models rather than generative VLMs. We propose GGSS—Geodesic-Gated Spherical Steering—a norm-preserving intervention that discovers a counterfactual bias subspace on the unit hypersphere, steers visual tokens along geodesic arcs, and uses an adaptive gate to focus correction on tokens that carry stronger demographic signal. We evaluate four generative VLMs against ten adapted inference-time debiasing baselines and prompt-based mitigation under a single operating-point protocol across categorical, pairwise, and occupation-gender bias tests, while also measuring general visual-language capability. GGSS achieves the lowest average bias on all four models, significant on three of four backbones under paired permutation tests, while preserving MMStar accuracy within +/- 0.6 p.p. of the unsteered baseline. Code is available at https://github.com/dukesun99/GGSS.

Long-term memory for LLM agents must answer two questions for every stored fact: whether the fact is still true and whether it should be surfaced for a given query. We refer to these two dimensions as validity and visibility, respectively. Existing memory systems often collapse these two decisions into a single scalar score. However, current and historical queries may require opposite retrieval decisions for the same fact: a superseded fact should be hidden for a current-intent query but surfaced for a historical-intent one. This creates a structural failure mode: superseded and expired facts are collapsed onto the same validity axis. To address this problem, we propose MemFSM, a training-free retrieval-side layer that separates validity from visibility through a memory-lifecycle finite-state machine. Along the validity axis, a four-state FSM tracks whether each fact has been superseded and expired. Along the visibility axis, an intent-aware router decides whether to surface or hide a fact based on its state and the query intent. By decoupling visibility from validity, MemFSM improves performance across heterogeneous memory benchmarks and diverse LLM backbones. To evaluate memory systems under fact conflicts and diverse query intents, we further introduce ValidMem, a benchmark that annotates each fact with both supersedure and expiry labels and pairs each memory store with both current-intent and historical-intent queries.

Large Language Models (LLMs) have demonstrated remarkable capabilities in processing long contexts and generating long-form texts. However, existing benchmarks evaluate these capabilities by focusing on either long-input comprehension with short outputs or short-input creative generation with lengthy outputs. This critical gap overlooks real-world Long-In-Long-Out Generation tasks, where models must simultaneously process long-context inputs and produce long-form outputs while adhering to fine-grained constraints. We introduce DocFollow, a comprehensive benchmark featuring 2,436 expert-validated instances across 5 major domains and 20 subdomains, designed to evaluate LLMs’ ability to follow hierarchical constraints in Document Following tasks. Our benchmark employs five types of constraints with increasing granularity. To ensure reliable evaluation, we develop a rule-based assessment protocol that decomposes overall evaluation uncertainty into atomic-level classification uncertainties, employing LLMs solely as information extractors for well-established NER and RE tasks rather than subjective scoring. Our evaluation of 13 state-of-the-art LLMs reveals a consistent pattern: while models maintain reasonable performance on coarse-grained constraints, they exhibit substantial degradation on fine-grained requirements, particularly entity-level constraints. This finding suggests fundamental architectural limitations in maintaining detailed structural and semantic fidelity throughout extended generation. Our findings highlight the critical need for advancing long-context generation capabilities beyond surface-level coherence to achieve precise fine-grained control. The implementation is publicly available. Source code is available at https://github.com/leonbig/DocFollow

While recent research has systematically documented political orientation in large language models (LLMs), existing evaluations rely primarily on direct probing or demographic persona engineering to surface ideological biases. In social psychology, however, political ideology is also understood as a downstream consequence of fundamental moral intuitions. In this work, we manipulate the moral conditioning and investigate the changes in the model’s political behaviors. Rather than simply assigning a demographic persona, we condition models to endorse or reject specific moral values and evaluate the resulting shifts on their political orientations, using the Political Compass Test. By treating moral values as lenses, we observe how moral conditioning actively modulates model trajectories across economic and social dimensions. Our findings show that such conditioning induces pronounced, value-specific shifts in models’ political coordinates. We further notice that these effects are systematically modulated by role framing and model scale, and are robust across alternative assessment instruments instantiating the same moral value. This highlights that effective alignment requires anchoring political assessments within the context of broader social values including morality, paving the way for more socially grounded alignment techniques.

End-to-end manga generation is a structured visual storytelling task that requires story decomposition, recurring character and scene grounding, page layout design, panel rendering, page composition, and lettering. However, existing generative models often perform direct page synthesis, entangling these factors in a single visual output and limiting precise control over layout geometry, visual references, and cross-panel consistency. To address these limitations, we propose MangaFlow, an agentic framework for controllable long-form manga generation that decomposes manga creation into planning, grounding, layout construction, reference-conditioned rendering, composition, and text placement. By treating layout and visual references as explicit intermediate variables, MangaFlow enables both simple text-to-manga generation and more precise user-controlled manga creation. This design exposes layout, visual assets, and lettering as editable intermediate controls for refining panel geometry, references, and text placement. To support long-form consistency, MangaFlow introduces a story section memory that links section descriptions with corresponding character, scene, and object references for reuse across panels. We further present a meta-benchmark for evaluating layout controllability, visual consistency, and generation quality. Experiments show that MangaFlow improves layout adherence and cross-panel consistency over direct generation baselines while supporting flexible human control.

Lexical semantic change detection identifies words whose meanings change over time, but it often cannot explain which senses and semantic properties are responsible for the change. We propose LSC-Pattern, a framework for interpretable lexical semantic change analysis that combines dictionary-grounded diachronic sense distributions with interpretable semantic dimensions. The framework first assigns each target-word usage to an explicit dictionary sense and then analyzes the resulting sense distributions along dimensions such as valence, semantic scope, and metaphoricity. We instantiate it for Japanese with LLM-based word sense disambiguation, which reaches 0.815 accuracy against majority-vote human annotation and outperforms five baselines; its sense-distribution change scores correlate with JaSemChange human judgments (Spearman’s ρ of 0.821 on seven overlapping entries and 0.580 on all 20). Case studies illustrate how sense redistribution and operationalized semantic signals can jointly support the discovery and interpretation of candidate types of change.

The integration of large language models (LLMs) into topic modeling has shown promise for improving topic interpretability and coherence, particularly by exploiting the generative capabilities of LLMs through instruction-based prompting. However, existing methods often incur substantial computational cost and inference complexity. To address these limitations, we propose **LLMFeed**, a novel framework that leverages **feed**back from an **LLM** to improve topic model optimization. Specifically, the LLM is used to **evaluate** topic quality, **identify** semantic meaning, and **recommend** representative topic words. These signals are then incorporated to (i) refine topic representations via Direct Preference Optimization (DPO), and (ii) guide contrastive learning for improved document-topic assignments. Extensive experiments show that **LLMFeed** achieves a 7.6% gain in topic coherence and over 10% improvement in doc-topic distributions on 20 Newsgroups dataset, while requiring only minutes to run, compared with hours for recent LLM-integrated topic modeling methods.

GUI agents often encounter dynamic anomalies when deployed on Android devices, from unexpected pop-ups to action misuse, yet existing benchmarks lack systematic evaluation of agent robustness against runtime anomalies. We introduce AnTrap, a comprehensive benchmark that injects dynamic perturbations into agent execution trajectories. We propose a taxonomy organizing real-world anomalies into four layers (State, Thinking, Action and Round) with ten fine-grained subcategories, and develop a construction pipeline that preserves task solvability while introducing realistic adversarial conditions. Evaluating 16 leading GUI models, we reveal universal vulnerability to dynamic anomalies, with even the strongest models suffering significant performance degradation. Furthermore, we conduct GRPO training in both original and adversarial environments to validate our benchmark, separating environment-learnable anomalies from reasoning-bottlenecked ones. Our findings show that while single-step traps at state and action layers are largely addressable through adversarial reinforcement learning, deep contextual traps, like state deadlock, expose intrinsic limitations that cannot be resolved by training in environments with traps alone.

Seamless translation of natural language into embodied behaviors remains critical for open domain embodied dialogue. We formalize this task as mapping large language model text to synchronized speech and facial animations. Prevailing cascaded Text to Speech (TTS) and Talking Face Generation (TFG) pipelines suffer from prohibitive latency and poor cross platform generalizability. We present EmbodiChat, an end-to-end real-time dialogue system overcoming these bottlenecks. The key insight of EmbodiChat is that TTS acoustic representations capture pure “articulatory intent,” making them far more robust than audio-derived semantic features. Aligning facial generation directly onto this representation without additional adapters enables decoupled audio visual synthesis, fundamentally eliminating the serial latency of conventional audio first architectures. Concurrently, our framework decodes these representations into hardware agnostic action sequences. EmbodiChat achieves a stable Time-to-First-Audio (TTFA) of about 150 ms on a consumer GPU, significantly outperforming mainstream methods. Furthermore, it delivers a 3x speedup relative to its ablated audio driven baseline while maintaining comparable animation quality. Deployment across two distinct embodied agent platforms configurations demonstrates a robust, practically deployable text driven solution for embodied dialogue. The data, code and model will be released at https://github.com/JiangMingZi/Embodi_Chat

Hint-based reinforcement learning addresses reward sparsity in long-horizon agentic tasks by retaining a prefix of an expert trajectory before each rollout, letting the policy explore from a state closer to success. Its effectiveness hinges on the guidance depth: how much of the trajectory to keep. Existing methods treat this depth as a deterministic scalar. Scheduled approaches share one value across samples and ignore per-task heterogeneity; per-sample probing estimates it separately at the cost of extra rollouts. We find that useful guidance occupies a band of depths whose informativeness profile is approximately Gaussian around the band center, rather than concentrating at a single optimal point. We propose Agent-G², a Gaussian guidance framework that draws the depth per task from a Gaussian whose center and spread are estimated online from rollouts already collected for policy optimization, requiring no probe rollouts or learned depth predictor. The center combines a global baseline with per-cluster difficulty, and the spread tracks within-cluster variance. We evaluate Agent-G² on ALFWorld and WebShop with Qwen2.5-1.5B / 7B-Instruct. Agent-G² outperforms the strongest hint-based, hint-free, and Aux-RL baselines on ALFWorld by 2.3 / 3.9 / 7.4 points at under one-third the rollout cost of per-sample probing.

The Key–Value (KV) cache has become a critical memory bottleneck that significantly hinders the deployment of large language models, especially under long-context inference. Existing approaches either introduce non-negligible errors through lossy compression, or are tightly coupled to specific tasks or model structures, limiting their general applicability. In this work, we first identify that the exponent values in BFloat16 KV caches are highly concentrated, and further prove theoretically that such concentration is inherent to LLMs, ensuring the broad generalizability of our method. Leveraging this insight, we propose AdaCode, a bit-exact lossless compression framework. The core idea is to encode frequent exponents with a short fixed-length code while storing rare outliers explicitly. AdaCode consists of two key components: (1) a minimal-cost optimization that selects the optimal code length and exponent range by balancing the storage overhead of short codes against that of outliers, complemented by a lightweight online adjuster to handle occasional distribution shifts without GPU overhead; and (2) a GPU-friendly byte-aligned packing format together with offline calibration for parameter precomputation. Experiments show that AdaCode achieves an average compression ratio of 71.78% on diverse LLMs and 70.26% on vision-language models, with bit-exact identical recovery and no accuracy loss, effectively alleviating the KV cache memory bottleneck. The codes are available at https://github.com/lichangshiwhu/Adacode.

Current Vision-Language Models struggle with hours-long videos because processing full-length visual sequences induces prohibitive token explosion and attention dilution. To overcome this, we introduce MemDreamer to decouple perception and reasoning, shifting long-video understanding into an agentic exploration process. As a plug-and-play framework, it incrementally streams videos to construct a Hierarchical Graph Memory, a top-down three-tier architecture for semantic abstraction, anchored by a foundational graph capturing spatiotemporal and causal relations. During inference, the reasoning model employs agentic tool-augmented retrieval, navigating hierarchies, searching nodes, and traversing logical edges via an Observation-Reason-Action loop. Experiments show MemDreamer achieves SOTA results across four mainstream benchmarks, substantially narrowing the gap with human experts. It greatly reduces the reasoning context relative to full-context ingestion while delivering a substantial absolute accuracy gain. Furthermore, statistical analysis uncovers a strong positive linear correlation between an VLM’s performance on logic reasoning and long-video understanding benchmarks, establishing agentic capability scaling as a new paradigm for multimodal comprehension.

Test-time training (TTT) adapts a frozen large language model (LLM) at inference time by updating prompt-specific parameters before generation. It is especially appealing for long-context reasoning. However, vanilla TTT yields only marginal gains on modern LLMs. We trace this bottleneck to two coupled failures: the inner loop is unprimed, because LLMs are trained as static endpoints rather than initialization points for test-time gradient adaptation; and the proxy objective is misaligned, because prompt-masked next-token prediction (PM-NTP) rewards prompt memorization rather than answer correctness. We propose PrimeTTT, a lightweight TTT method that casts this adaptation problem as meta-learning. A compact hypernetwork emits a prompt-specific LoRA initialization for the inner loop, trained with a MAML-style second-order objective: the inner loop descends the test-time-available PM-NTP loss, while the outer loss is answer-conditioned NTP on the gold answer and is backpropagated through the unrolled inner steps. Across out-of-distribution benchmarks disjoint from training, PrimeTTT consistently outperforms strong TTT baselines and transfers across model families and scales.

Accurate and comprehensive video captions with consistent subject references are critical for downstream understanding and generation tasks. However, few existing benchmarks can objectively and comprehensively evaluate these properties across diverse durations and scenarios, thereby hindering the advancement of video captioning models. To bridge this gap, we propose CapRiCorn-1K, a comprehensive benchmark designed to evaluate both video captioning quality and subject referential consistency across long temporal horizons and diverse video domains. To accommodate varied evaluation needs, our benchmark supports both audiovisual and visual-only settings. Extensive experiments on CapRiCorn-1K reveal that current models generally struggle to generate accurate and comprehensive captions while maintaining consistent subject references. Moreover, as video duration increases, both the overall caption quality and subject referential consistency decline. Notably, our evaluation metrics exhibit strong correlations with the performance of downstream understanding and generation tasks conditioned on the generated captions, further validating their effectiveness. CapRiCorn-1K will be fully open-sourced to facilitate future research.

Continuous-space reasoning methods replace discrete intermediate tokens with probability-weighted soft embeddings, aiming to preserve semantic information that greedy Chain-of-Thought decoding discards. We show that this promise is limited by local representation geometry and the latent semantic manifold. When the next-token distribution is low-entropy and concentrated, a soft token becomes nearly collinear with the Top-1 embedding and collapses toward greedy decoding. When the distribution is high-entropy, diffuse, or stochastically perturbed, the continuous state may leave the greedy path but drift away from the local scale of naturally occurring candidate representations. We diagnose these two regimes with step-level probes over representation divergence, local norm calibration, and behavior-cloud membership. To address them, we propose Manifold Thinking, a training-free inference framework that combines structured tangent movement away from the Top-1 direction with local norm-calibrated projection back to the candidate-token neighborhood. Across mathematical and programming benchmarks, Manifold Thinking improves over discrete CoT baselines and baselines covering the latest stochastic Soft Thinking strategies, suggesting that useful continuous reasoning requires geometric correction rather than unconstrained linear token mixing alone.

Large language models (LLMs) have shown promise for automated unit test generation, but existing evaluations largely rely on standalone settings and a narrow set of programming languages, overestimating real-world readiness. We introduce XRepoTest, a multilingual, repository-level benchmark for unit test generation spanning five underexplored languages: Rust, Go, Julia, PHP, and Ruby. XRepoTest evaluates tests under realistic repository constraints using a containerized execution framework and multiple context augmentation strategies, including file-level, LSP-based, and retrieval-based context. Beyond standard metrics such as test pass rate and coverage, we propose Invocation Rate (IR) to assess whether generated tests meaningfully exercise the intended functionality. Experiments with 14 state-of-the-art LLMs (e.g., Claude 4.5, GPT-5.2, DeepSeek V4-pro, and Qwen families) reveal a substantial gap between standalone and repository-level performance, as well as trade-offs between richer context and test reliability. Overall, XRepoTest provides a challenging and informative benchmark to advance scalable, robust, and multilingual unit test generation in realistic software environments.

Practical deployment of medical vision-language models requires a single model that interprets heterogeneous clinical visual inputs: 2D images, 3D volumes, and medical videos. Yet how to train such unified models remains underexplored, as these modalities differ in spatial-temporal structure, token length, and data availability. We present OmniV-Med, a scalable training recipe built on four empirical findings. First, broad modality coverage requires balancing modality diversity with task-level supervision, motivating OmniV-Med-Instruct, a 252K corpus spanning 14 medical modalities and 11 clinical tasks. Second, we identify a modality-ordering effect: introducing 3D and video too early destabilizes optimization and weakens 2D grounding, motivating a progressive curriculum that first builds 2D alignment before extending to spatial-temporal inputs. Third, we observe an emergent spatiotemporal transfer effect: once strong 2D patch representations are in place, lightweight rotary position adaptation extrapolates them to 3D slices and video frames, letting a single 2D ViT replace modality-specific encoder stacks. Fourth, the imbalance in sequence length across modalities and the redundancy of long volumes and videos hinder scaling; our Redundancy-Guided Token Pruning (RGTP) exploits inter-slice and inter-frame redundancy to cut visual tokens by 60% while preserving performance. Across nine 2D, 3D, and video medical benchmarks, OmniV-Med achieves robust performance. Extensive component-wise ablations systematically refine the design space, yielding a simple yet efficient and scalable recipe for training high-performance unified Med-VLMs. Code is released at https://github.com/ZJUI-AI4H/Hulu-Med.

Medical LLMs are often evaluated by whether they select the correct diagnosis, but diagnostic accuracy alone does not show whether the model used the case evidence appropriately. We present a behavioral audit of evidence use in medical diagnosis. For each case, we decompose patient information into evidence units, score candidate diagnoses under controlled evidence subsets, and mine low-order interactions in diagnostic margins. Because medical evidence is diagnosis-relative, the audit separates interaction discovery from failure assignment: large or negative interactions can reflect plausible differential diagnosis, while suspicious interactions require robustness checks and clinical review. We evaluate five open-weight LLMs on DDXPlus, CupCase, and MedCase. Across datasets, faithful support and differential conflict or cancellation account for most interaction strength, showing that many evidence interactions are clinically plausible rather than failures. In a DDXPlus-focused blinded five-reviewer 130-item enriched review sample, invalid or shortcut-like cases concentrate in negated or absent findings and clinically local evidence. These results show that accuracy can hide candidate evidence-use failures and motivate role-aware audits for medical LLM evaluation.

Large language models (LLMs) are increasingly evolving from simple text-based interaction systems into LLM agents that can maintain memory, use tools, access external environments, and execute tasks. As their capabilities and autonomy expand, the safety risks they face also become more diverse. Existing evaluations often rely on manually written scenarios, static prompts, or final-output judgments, making it difficult to capture the diverse risks that agents may face during task execution. We introduce ForesightSafety-SAGE, a fully automated scenario generation and safety evaluation framework for LLM agents. Based on five risk dimensions, we instantiate abstract and diverse safety risks in real-world task execution into 1,072 measurable evaluation scenarios. Using the automated evaluation pipeline, 12 LLM agents are evaluated under two authority contexts. The results show that current agents still face substantial behavioral safety risks during task execution, with an average ASR of 47.1% and several models exceeding 70%. These findings demonstrate the importance of executable, process-level evaluation for understanding and improving LLM agent safety.

Large language models achieve strong performance across diverse tasks, but deployment remains costly because of memory, latency, and energy demands. Structured pruning reduces these costs by removing architectural components, yet its recovery stage is often limited by a mismatch between the recovery module’s representational capacity and the complexity of the removed knowledge. We call this bottleneck the capacity-knowledge asymmetry and propose **OverRep**, an **Over**complete **Rep**arameterization framework for structured LLM pruning. Following the principle of *“train overcomplete, deploy compact”*, OverRep temporarily overparameterizes the recovery module during training to absorb complex knowledge distilled from the original model. After recovery, the overcomplete re-parameterization is algebraically merged into a mathematically equivalent compact module, preserving the pruned model’s inference-time architecture and computational cost. OverRep further introduces an annealed activation that enables nonlinear training dynamics while converging to a linear regime for exact algebraic merging. Across three backbone families, OverRep improves retained reasoning performance over strong recovery baselines by up to 5.5 and 8.4 points at 25% and 50% pruning, respectively, while keeping memory usage and TFLOPs comparable to existing recovery methods. Our code is available at https://github.com/mmai-laboratory/OverRep

Decoding throughput gains from larger inference batches are often limited by GPU memory, mainly due to the large size of KV cache. Training-free KV cache offloading mitigates this issue by storing redundant context on the CPU and fetching sparse KV subsets, but often harms long-generation quality due to training-inference mismatch of sparse patterns. Trainable sparse attention can improve quality, yet its unconstrained KV accesses are inefficient for offloading and may incur excessive CPU–GPU transfers. To this end, we propose NOSA, a trainable sparse attention mechanism designed for KV cache offloading. NOSA explicitly bounds CPU–GPU KV transfers, enabling low communication overhead and high decoding throughput. We further develop NOSI, an offloading inference system that realizes these efficiency gains. Experiments on 1,3,8B LLMs show that NOSA outperforms KV cache offloading baselines on general, long-input, and long-generation tasks, while improving decoding throughput by up to 5.04×, 1.92×, and 1.83× over FullAttn, InfLLMv2, and ShadowKV, respectively.

Millions of individuals’ well-being are challenged by the harms of substance use. Harm reduction as a public health strategy provides non-judgemental, evidence-based information intended to improve health outcomes and reduce associated safety risks. Some large language models (LLMs) have demonstrated a high level of medical reasoning, promising to address the information needs of people who use drugs (PWUD). However, their performance in relevant tasks remains largely unexplored. We introduce HarmReduction, a benchmark designed to evaluate LLMs’ accuracy and safety risks in harm reduction information provision. The benchmark dataset (HRBasic) has 2,160 question-answer-evidence pairs. The scope covers three tasks: checking safety boundaries, providing quantitative values, and inferring polysubstance use risks. We build the Instruction and RAG schemes to evaluate model behaviours based on their inherent knowledge and the integration of domain knowledge. Our results indicate that state-of-the-art LLMs still struggle to provide accurate harm reduction information, and sometimes, present severe safety risks to PWUD. This work contributes an evaluation framework for LLMs to deliver harm reduction information to avoid introducing negative health outcomes through the use of LLMs.

Long-horizon LLM agents generate traces that could become reusable experience, but raw trajectories are noisy, local, and hard to govern. Agent Skills offer a structured artifact for combining procedural guidance, executable resources, and applicability boundaries. Yet open skill ecosystems contain redundant, uneven, environment-sensitive artifacts, and indiscriminate updates can pollute future context. We present SkillsVote, a lifecycle-governance framework for Agent Skills across collection, recommendation, attribution, and evolution. SkillsVote profiles a million-scale open-source corpus for environment requirements, quality, and verifiability, and synthesizes tasks for verifiable skills. Before execution, it performs agentic library search over structured skill folders to expose instructional context. After execution, it decomposes trajectories into skill-linked subtasks, attributes outcomes to skill-guided execution, agent exploration, environment, and result signals, and admits only successful reusable discoveries to evidence-gated updates. Experiments on Terminal-Bench 2.0 and SWE-Bench Pro show that SkillsVote improves agent performance on challenging agentic coding benchmarks. The gains arise from two complementary pathways: online evolution over test-time task streams and offline transfer via frozen libraries built from either historical trajectories or curated open-source skills. We release our code at https://github.com/MemTensor/skills-vote.

Diffusion Language Models (DLMs) have emerged as a promising alternative to autoregressive language models, showing strong non-autoregressive generation abilities. Although the masked denoising objective trains DLMs to recover tokens at arbitrary positions, current long-context usage still mostly follows a prefix-based paradigm: the full context is given as a prefix, and the answer is generated only at the end, leaving the model’s infilling ability underutilized. To investigate whether DLMs can read long contexts by infilling, we introduce Parallel Infilling Reading (PIR), a training-free method that inserts extraction slots into long contexts to generate intermediate reading states. Our empirical study shows that naive PIR improves over prefix-based DLM inference, but suffers from slot repetition, where multiple slots generate redundant evidence. To mitigate this issue, we propose Slot-Isolated PIR, which blocks direct slot-to-slot attention, and PIR-SFT, which uses fine-grained evidence supervision to teach DLMs effective evidence extraction while reading. Experiments on long-context question answering benchmarks show that our methods improve answer accuracy and generation efficiency, with increasingly larger gains as the context length grows.

Syntactic prediction requires models to generate outputs that are globally consistent and satisfy hierarchical structural constraints. Although large language models have been applied to such tasks, existing approaches still rely mainly on prompting or supervised fine-tuning, and their performance generally remains below that of specialized neural syntactic models. Our experiments show that supervised fine-tuning reaches a plateau, and additional supervised training does not further improve prediction quality. We therefore propose a two-stage post-training framework for syntactic prediction. The first cold-start stage adapt the model to a unified syntax generation format, and then applies reinforcement learning on a disjoint data subset to directly optimize structure-level objectives. Experiments on constituency parsing and CCG supertagging show that Syntax-RL consistently outperforms supervised LLM baselines and achieves state-of-the-art results. Further analyses suggest that its gains come primarily from the reinforcement learning stage. These findings identify reinforcement learning as an effective post-training mechanism for improving hierarchical syntactic prediction in large language models.

Recent latent visual reasoning methods achieve substantial gains by inserting continuous latent tokens into multimodal language models. These gains are commonly attributed to the tokens encoding visual evidence; recent analyses, however, reveal a paradox: the tokens are loosely tied to the image and contribute little to the answer. Critically, these analyses treat latent tokens as a single unit, obscuring the source of the gains. We therefore decompose latent tokens into three testable components: latent slots, boundary markers, and format, and develop a state-of-the-art method as a probe under favorable conditions. Across six method-stage settings and four perception-heavy benchmarks, latent slots fail every prediction of the visual-memory account. Strikingly, retaining only the boundary markers preserves 78 to 100% of the gain in several settings, while the model attends to the image more narrowly at latent positions than at answer positions. These results do not support slot contents as recoverable visual memory; much of the benefit is instead associated with marker and format control, together with visual-attention routing. At matched accuracy, methods can still rely on markedly different mechanisms shaped by training supervision. Latent visual reasoning thus needs evaluation not only by accuracy but by what the model actually relies on.

The advancement of LLM agents with tool-use capabilities requires diverse and complex training corpora. Existing data generation methods, which predominantly follow the paradigm of random sampling and shallow generation, often yield simple and homogeneous trajectories that fail to capture complex, implicit logical dependencies. To bridge this gap, we introduce **HardGen**, an automatic agentic pipeline designed to generate hard tool-use training samples. First, HardGen establishes a dynamic API Graph built upon agent failure cases, from which it samples hard tool-execution sequences. Second, these execution sequences are used to reversely build advanced tools that combine multi-step operations into one, which then help generate hard queries requiring unstated intermediate steps. Finally, advanced tools and hard queries drive the generation of Chain-of-Thought (CoT), which is iteratively refined through closed-loop verification feedback. Extensive evaluations demonstrate that a 4B parameter model trained with our curated dataset achieves superior performance compared to several leading open-source and closed-source competitors (e.g., GPT-5.2, Gemini-3-Pro and Claude-Opus-4.5). Our code, models and the curated dataset with 27,000 samples will be open-sourced to facilitate future research.

Vision-language models (VLMs) are increasingly deployed in multi-turn settings where users may describe visual content with incorrect assumptions. Yet existing evaluations rarely isolate how models respond when the same visually grounded false premise persists across dialogue turns. We introduce FPCO-Dialog, a benchmark for evaluating correction and cooperation behavior in VLMs under repeated false premises. FPCO-Dialog contains 1,080 images and 10,800 question turns, stratified by visual complexity, object category, and false-premise class, and uses a 10-turn protocol in which a correct dialogue prefix is followed by repeated false-premise referring expressions. We evaluate 20 commercial and open-source VLMs with a model-agnostic protocol and CorrTP@K, a correction-rate metric over false-premise turns, scored by two independent detectors. FPCO-Dialog reveals substantial and persistent cross-model differences in aggregate correction tendency, model-specific turn-wise dynamics, and systematic variation across false-premise types under the benchmark’s substitution distribution. The dataset, evaluation protocol, model outputs, detector labels, and code are available at https://github.com/lab-klc/FPCO-Dialog.

Modern supervised fine-tuning (SFT) trains large language models on heterogeneous instruction mixtures, where the effective unit of optimization is an individual instruction-response pair rather than a coarse task label. Existing parameter isolation methods mitigate interference by assigning protected or trainable subspaces at the task or training stage level, but these units are often misaligned with actual parameter usage: instructions from the same task may require conflicting computations, while instructions from different tasks may share latent capabilities. We propose Instruction-Level Parameter Isolation (IPI), a fine-tuning framework that routes each instruction’s update to a sparse, capability-conditioned parameter subspace. IPI builds a multi-view instruction representation, predicts importance factors with a lightweight router, and uses an identity-trick reparameterization that preserves dense forward computation while exposing sparse base-model gradients and a task-loss saliency signal for router learning. Across heterogeneous instruction tuning benchmarks and multiple backbone models, IPI consistently improves performance over strong fine-tuning and parameter-isolation baselines. Further analyses show that instruction-conditioned masks reduce destructive gradient interactions and capture optimization structure that task labels miss.

Post-training model compression is essential for enhancing the portability of Large Language Models (LLMs) while preserving their performance. While several compression approaches have been proposed, less emphasis has been placed on selecting the most suitable set of data (the so-called *calibration data*) for finding the compressed model configuration. The choice of calibration data is a critical step in preserving model capabilities both intra- and inter-tasks. In this work, we address the challenge of identifying high-performance calibration sets for both pruning and quantization by analyzing intrinsic data properties rather than model-specific signals. We introduce **ZipCal**, a model-agnostic data curation strategy that maximizes lexical diversity based on Zipfian power laws. Experiments demonstrate that our method outperforms standard uniform random sampling across various benchmarks. Notably, it also performs on par, in terms of downstream performance, with a state-of-the-art method that relies on model perplexity. The latter becomes prohibitively expensive at large-scale models and datasets, while **ZipCal** is on average 240 faster due to its tractable linear complexity. We make the code and the experiments available at https://github.com/FrancescoMonaco/ZipCal.

As Large Language Models (LLMs) increasingly generate reasoning-driven outputs that weave isolated events into intricate logical topologies, factuality depends not only on isolated atomic facts but critically on the faithful relations connecting them. However, current verification paradigms prioritize individual facts and remain fundamentally blind to relational distortions. This oversight creates a dangerous false sense of security, where verifiers inadvertently endorse relationally corrupted claims simply because their constituent events are factual. Furthermore, the community lacks a systematic benchmark to evaluate this crucial relational layer. To address these critical vulnerabilities, we introduce BeyondAtoms, a novel benchmark for relational fact verification. Our evaluations reveal that generic verification pipelines often fail by misclassifying relational manipulations as supported. While explicitly guided reasoning models can resolve these structural errors, they incur prohibitive computational costs. To counter this pervasive relation blindness, we propose CARVE, a completely training-free verification framework. By decoupling structural dependencies through deterministic extraction and bi-directional relational alignment, CARVE achieves a favorable accuracy–efficiency trade-off: it remains highly competitive on relational and general fact verification while avoiding the heavy inference cost of reasoning-based LLMs.

Time Series Event Detection (TSED) aims to localize semantically meaningful events in time series data, with critical applications in high-stakes domains. Unlike statistical anomalies, events are often defined by natural-language descriptions with internal temporal-logic structures across multiple physical channels. However, in real-world settings, dense event annotations are expensive to obtain, making purely supervised learning difficult. We introduce Language-guided TSED, a setting where a model is given textual event descriptions and must ground them to intervals in multivariate signals with little or no labeled data. To address this problem, we propose Event Logic Tree (ELT), a knowledge representation framework that converts linguistic descriptions into structured temporal logic over signal primitives. Building on ELT, we present SELA, a neuro-symbolic VLM agent framework that iteratively grounds primitives from signal visualizations and composes them under ELT constraints, producing both event intervals and faithful tree-structured explanations. We further release a real-world benchmark across energy and climate domains with expert knowledge and annotations. Experiments show that SELA improves over supervised fine-tuning and existing zero/few-shot time series reasoning baselines.

Existing temporal QA benchmarks focus on simple fact-seeking queries from news corpora, while reasoning-intensive retrieval benchmarks lack temporal grounding. However, real-world information needs often require reasoning about temporal evolution and synthesizing evidence across time periods. We introduce TEMPO, the first benchmark combining temporal reasoning with reasoning-intensive retrieval across 13 domains. TEMPO features: (1) 1,730 complex queries requiring deep temporal reasoning such as tracking changes, identifying trends, or comparing cross-period evidence; (2) step-wise retrieval planning with 3,976 decomposed steps and gold documents mapped to each step for multi-hop evaluation; and (3) novel temporal metrics including Temporal Coverage@k and Temporal Precision@k measuring whether results span required time periods. Evaluation of 16 retrieval systems reveals substantial challenges: the best model (DiVeR) achieves only 32.0 NDCG@10 and 71.4% Temporal Coverage@10, demonstrating difficulty in retrieving temporally complete evidence. We believe TEMPO provides a challenging benchmark for improving temporal reasoning in retrieval and RAG systems.

Diffusion Language Models (DLMs) have attracted significant attention for their strong reasoning ability. However, under a bidirectional attention mechanism, DLMs operate over an exponentially large exploration space compared to autoregressive models (ARMs), making it challenging to focus on reasoning-guiding tokens under random masking. We define causal shortcuts as token chains that cover the full sequence and provide explicit guidance towards correct reasoning trajectories. We analyze the effects of causal shortcuts on the reasoning accuracy and convergence speed of DLMs, and find that they largely improve answer convergence efficiency and generation accuracy. Motivated by this, we propose a Causal Shortcut Learning (CSL) Framework for DLMs. Specifically, we introduce a step-by-step token extraction procedure to extract causal shortcuts from data, and apply parallel prioritized masking on these tokens during training to enable efficient and accurate convergence to correct answers via causal shortcuts. Extensive experiments across multiple reasoning benchmarks and two base models demonstrate that CSL consistently outperforms existing SFT-variant baselines, achieving an average improvement of 1.92% over SFT-only models, and up to 4.20% on MATH-500. The code is available at the https://github.com/ZJUDianJin/Causal-Shortcuts-Learning.

Large language models (LLMs) trained with reinforcement objectives often achieve superficially correct answers via shortcut strategies, pairing correct outputs with spurious or unfaithful reasoning and degrading under small input perturbations. We introduce Coherence-Enhanced Policy Optimization (CE-PO), a model agnostic reward fusion framework that augments policy optimization with a differentiable Jacobian coherence signal along the generation pathway from prompt (X) to rationale (Y) to answer (Z). CE-PO estimates internal gradient influence through Jacobian-based sensitivities, uses a reshuffling-based residualization procedure to reduce shortcut-driven reward hacking, and fuses the resulting coherence score with task-accuracy via an L_p combiner that controls the strength of coherence regularization. The unified reward integrates with PPO/GRPO without architectural changes. Across multiple reasoning benchmarks, CE-PO improves accuracy over baselines by 5.49 on average (up to 9.58), while increasing the model’s faithful reasoning capability (12.14 % on average across five backbones).

Real-world knowledge bases evolve over time and accumulate outdated statements, contradictions, and misinformation. LLMs have been reported to be vulnerable to such inconsistent contexts. In this paper, we propose Proposition-Structured RAG (ProStruct-RAG), a framework that continuously structures evolving and inconsistent external knowledge bases using timestamped propositions and inter-proposition relations. ProStruct-RAG represents a document collection as a graph whose nodes are propositions and whose edges capture temporal updates, contradictions, and support relations, retrieves query-relevant knowledge as a relational structure, and uses it for answer generation in RAG. Experiments on multiple open-book temporal QA benchmarks show that ProStruct-RAG outperforms baselines that do not use inter-proposition relations, suggesting that RAG over evolving and inconsistent knowledge benefits from organizing external knowledge before retrieval, rather than relying solely on inference-time reasoning. We release our code and Asteria, a new fictional temporal QA dataset designed to mitigate data contamination.

Although representation learning has transformed semantic modeling in NLP, representations of linguistic style remain underexplored—partly due to conflicting definitions of style within and beyond NLP or unclear immediate advantages of separate style representations. In this survey, we provide an overview of style conceptualizations in NLP and (socio-)linguistics and suggest a definition of style for practitioners. Guided by this definition, we then review methods for creating and evaluating predefined and neural style representations. We conclude by discussing how style representations can make crucial contributions to the modern NLP pipeline (e.g., in dataset curation or evaluation) and to the application of NLP methods in other fields. We curate relevant datasets, models, and tools on our companion webpage at https://annawegmann.github.io/StyleSurvey/. Throughout, we provide practical guidance, identify open research questions with concrete recommendations for addressing them and issue calls to action on critical gaps in current work.

Recent advances in large audio language models (LALMs) have enabled multilingual speech understanding. However, benchmarks for evaluating LALMs remain scarce for non-English languages, with Korean being one such underexplored case. In this paper, we introduce **KoALa-Bench**, a comprehensive benchmark for evaluating Korean speech understanding and speech faithfulness of LALMs. In particular, KoALa-Bench comprises six tasks. Four tasks evaluate fundamental speech understanding capabilities, including automatic speech recognition, speech translation, speech question answering, and speech instruction following, while the remaining two tasks evaluate speech faithfulness, motivated by our observation that several LALMs often fail to fully leverage the speech modality. Furthermore, to reflect Korea-specific knowledge, our benchmark incorporates listening questions from the Korean college scholastic ability test as well as content covering Korean cultural domains. We conduct extensive experiments across six models, including both white-box and black-box ones. Our benchmark, evaluation code, and leaderboard are publicly available at [https://ksbench.github.io/Koala-Bench/](https://ksbench.github.io/Koala-Bench/).

Large language models (LLMs) have shown strong reasoning capabilities in knowledge graph question answering (KGQA), yet their predictions are not always supported by explicit and verifiable reasoning paths. Existing methods often retrieve or generate plausible candidate paths, but path plausibility alone does not establish that a prediction actually depends on the retrieved evidence. To address this issue, we propose FBV-KGQA, a forward-backward verification framework for reliable and interpretable KGQA. FBV-KGQA first derives executable reasoning paths through relation-level planning and evaluates their forward support for candidate answers. It then identifies critical entity and relation pivots and applies schema-constrained counterfactual perturbations to assess their influence on model predictions. Reverse questions are subsequently constructed from candidate answers to verify whether they can be consistently traced back to the original question entity. Finally, an adaptive selection strategy integrates forward support, counterfactual effects, and backward consistency. Experiments on WebQSP and CWQ show competitive performance against representative KGQA baselines, while human evaluation indicates improved traceability, clarity, and usefulness of the resulting reasoning paths.

Contrastive decoding (CD) seeks to mitigate hallucinations in Large Vision-Language Models (LVLMs) by contrasting the output distributions of a standard model and a visually degraded model. However, existing training-free CD methods suffer from sub-optimal degraded branches: completely dropping visual tokens is too extreme and induces language hallucinations, while corrupting input images offers coarse control over visual evidence and suffers from high inference latency due to requiring two full forward passes. To address these dilemmas, we propose YARD, a training-free Y-Architecture Register Decoding framework. Motivated by the observation that reliable text-to-vision grounding predominantly emerges in the middle decoder layers, YARD constructs the degraded branch internally by sharing shallow-layer computations and branching exactly at this critical stage. For the degraded branch, YARD replaces patch-level visual tokens with register tokens, which preserve global image semantics but lack fine-grained local evidence. This image-aware yet locally under-grounded design provides a faithful contrastive signal without extreme modality mismatch, while the Y-architecture strictly avoids a costly second forward pass. Extensive experiments on generative and discriminative hallucination benchmarks demonstrate that YARD consistently achieves state-of-the-art hallucination mitigation across multiple LVLMs, alongside a significant reduction in inference latency.

Long-horizon agentic tasks require large language models (LLMs) to iteratively retrieve, integrate, and maintain dispersed information across multi-turn interactions, but preserving all interaction histories leads to a continuously growing working context. Recent proactive context management methods allow models to edit their own working context with specialized tools, yet they still face **three key limitations**: (1) a *limited toolset* restricted to search, deletion, and summarization, with no support for global planning, long-term memory, and adaptive compression; (2) *inefficient exploration* that treats context management actions uniformly despite their heterogeneous impacts on final outcomes; and (3) *coarse-grained credit assignment* that assigns the final trajectory-level reward to all intermediate context editing actions during RL. To bridge these gaps, we introduce ***ContextPilot***, a proactive context management framework for long-horizon agentic reasoning. Our approach systematically augments the toolset with planning, long-term memory, and soft context offloading tools. We further propose an RL method tailored for context management, which uses context and entropy variation to identify critical context control actions for branch sampling and estimates action-level advantages from all branched trajectories that pass through these actions. Experiments on long-context QA and deep search tasks show that ***ContextPilot*** achieves stronger performance with a more compact working context, outperforming existing baselines across various base models and benchmarks.

As LLMs become increasingly woven into everyday workflows, user queries send to cloud-hosted LLMs routinely mix task-essential content with task-non-essential sensitive disclosures, yet type-based PII redaction is context-agnostic and may raise two issues: over disclosing untyped sensitive context and over-removing answer-bearing spans. We recast privacy-preserving query rewriting under Contextual Integrity: a span should be forwarded only if it is necessary for the task. We build the first task-based CI benchmark for privacy conscious delegation, comprising 3,167 samples that combine high-quality synthetic data spanning 11 tasks and 20 task types, WildChat based real-user queries, and a medical challenge set with dense sensitive information. Building on this benchmark, we propose a CI guided reinforcement learning framework that converts essential and non-essential sensitive spans into verifiable optimization signals, and train a query rewriter to preserve task-critical information while suppressing unnecessary sensitive disclosure. Experiments show that our learned rewriter achieves the best privacy–utility trade-off, achieving up to +10.1 average utility over on-device baselines.

Methods to represent literary texts as graphs or sequences of graphs mainly focus on representing character interactions, and often overlook another crucial aspect: the textual context in which characters interact. We introduce Dynamic Heterogeneous Character Networks (DHCNs), which organize long novels into temporally localized heterogeneous graphs that align characters with their textual contexts. We extract around 20,000 DHCNs from Project Gutenberg, and propose GraphLit, a self-supervised learning framework that learns rich literary representations through a masked graph autoencoder objective. Across a wide range of 12 character-related tasks, GraphLit improves over text-only, graph-only and prior hybrid baselines. Ablations over different kinds of dynamic graph structures and architectural elements show that grounding characters in their context is the main performance driver, while explicitly encoding narrative order and character relationships provide task-dependent improvements. Finally, we demonstrate the applicability of DHCNs and GraphLit for literary analysis by studying the link between narrative non-linearity and dynamic social features.

Humans can infer others’ beliefs, intentions, and emotions from observable behavior, an ability known as Theory of Mind (ToM). While current Multimodal Large Language Models (MLLMs) excel at visual understanding, can they identify mental-state cues in videos and reason about evolving mental states? To systematically evaluate spatio-temporal ToM in MLLMs, we introduce ToM-Video, a large-scale benchmark built from diverse real-world narrative videos. Through multi-stage filtering and annotation from massive data sources, ToM-Video provides a high-quality collection of ToM-centric scenes, comprising 1,487 videos and 41K visual question answering (VQA) pairs. We further propose INSPIRE, a post-training framework that, through a carefully designed three-stage curriculum, progressively teaches models to locate ToM-relevant cues and connect them to latent mental-state changes. Experiments show that current MLLMs remain weak at linking observable cues to evolving mental states, while INSPIRE shifts the model from surface description toward compact, evidence-backed inference, leading to higher accuracy and better generalization to external social-reasoning video datasets.

Personalizing large language models often requires adapting them to new domains, tasks, or user preferences. However, such adaptation can unintentionally shift a model’s value profile, even when the training data is primarily knowledge-oriented. We refer to this failure mode as value drift, a form of value alignment tax that emerges when learning new knowledge interferes with pre-existing alignment. To address this problem, we propose VISA, a modular framework for Value Injection via Shielded Adaptation. VISA keeps the base model as a stable content generator and trains a lightweight Value Rewriter to adjust the value expression of generated responses toward a target value vector. The target is constructed through a Value Detector and an Instruction Translator, while the rewriter is optimized with Group Relative Policy Optimization using rewards for value alignment and semantic consistency. We further construct VCR-45K, a value-conditioned rewriting dataset for evaluating controllable value adaptation. Experiments show that VISA substantially improves semantic preservation over prompting and fine-tuning baselines while maintaining competitive value alignment, enabling low-cost personalized alignment with reduced value drift.

Multimodal Large Language Models (MLLMs) have recently been applied to universal multimodal retrieval, where Chain-of-Thought (CoT) reasoning improves candidate reranking. However, existing approaches remain largely language-driven, relying on static visual encodings and lacking the ability to actively verify fine-grained visual evidence, which often leads to speculative reasoning in visually ambiguous cases. We propose V-Retrver, an evidence-driven retrieval framework that reformulates multimodal retrieval as an agentic reasoning process grounded in visual inspection. V-Retrver enables an MLLM to selectively acquire visual evidence during reasoning via external visual tools, performing a multimodal interleaved reasoning process that alternates between hypothesis generation and targeted visual verification. To train such an evidence-gathering retrieval agent, we adopt a curriculum-based learning strategy combining supervised reasoning activation, rejection-based refinement, and reinforcement learning with an evidence-aligned objective. Experiments across multiple multimodal retrieval benchmarks demonstrate consistent improvements in retrieval accuracy (with 23.0% improvements on average), perception-driven reasoning reliability, and generalization.

LLMs increasingly use external tools and stateful environments for complex, multi-step tasks. Further improving their tool-use capabilities, however, remains bottlenecked by the lack of scalable stateful environments and verified multi-turn trajectories matched to the target LLM’s current capability boundary. We observe that agent skills, packages pairing workflow specifications with executable scripts, provide scalable seed material for tool-use environment synthesis. This motivates Grafting, a skill-to-environment synthesis loop in which coding-agent harnesses perform open-ended construction under deterministic validation contracts, producing self-contained executable environments that are refined using trajectory-level verdicts. In an initial end-to-end run, Grafting produces 500 environments and approximately 12K verified multi-turn trajectories. We train on these trajectories with Fathoming, an iterative GRPO procedure that selects verified trajectories near the target model’s capability boundary. Training a 4B-parameter instruction-tuned model with two rounds of Fathoming on the trajectory pool produces Fathom-4B and yields strong out-of-distribution gains: BFCLv4 multi-turn accuracy rises from 22.0 to 39.7 (+17.7 points), and ComplexFuncBench per-call accuracy rises from 49.6 to 61.6 (+12.0 points). Using the same environment-synthesis pipeline, we further construct Grove-5K, a collection of 5,101 executable environments. We release Grafting, Grove-500, Grove-5K, and Fathom-12K.

Emotions are an essential aspect of human communication, particularly on social media, where authors frequently combine text and images to convey their emotions. Yet prior work on emotion analysis of social media posts has overlooked two important aspects in regard to measuring how well readers can reconstruct the authors’ intent: (1) the image modality, with most work focusing solely on text, and (2) the real-world events that trigger the expressed emotions, and their relationship to the post content. We therefore study the relation between (a) the author’s experience of the event that caused them to write a social media post and (b) the content of the post, with a focus on readers’ capability to reconstruct that emotion expression. To do that, we introduce the Multimodal Multi-Emotion-Model dataset (Multi^2EMo), collecting annotations from both authors and readers on the posts and their triggering events. We find that reconstruction is possible but challenging for both human readers and computational models. We show that understanding the triggering event is crucial for accurate reconstruction, and that reconstruction is particularly challenging when posts rely heavily on the image to express emotion.

Post-training quantization (PTQ) has become an important technique for reducing the inference cost of Large Language Models (LLMs). While recent mixed-precision methods improve ultra-low bit quantization by preserving critical subspaces in high precision, they typically construct these subspaces relying solely on activation statistics. This ignores the fundamental nature of linear operations, where the output perturbation is jointly driven by both activation and weight quantization noise. In this paper, we propose CoQuant, a joint weight-activation subspace projection method. By theoretically modeling the expected output error, CoQuant formulates a closed-form weighted PCA solution that balances activation and weight covariances to select the optimal high-precision subspace. Extensive experiments on Llama-3.2 and Qwen2.5 models show that CoQuant consistently outperforms strong PTQ baselines in both WikiText perplexity and zero-shot common-sense reasoning accuracy. These results demonstrate that joint weight-activation subspace modeling provides a principled and effective direction for low-bit LLM quantization. The source code is available at https://github.com/Zachary5895/CoQuant

Retrieval-Augmented Generation (RAG) enhanced by Knowledge Graphs has shown promise in complex multi-hop reasoning tasks. However, existing graph-based retrieval methods typically rely on flat, undirected topologies. During the retrieval process, the probability flow often gets trapped in high-degree abstract concept nodes which we define as "probability black holes", leading to semantic drift and noise accumulation. To address this, we propose SemFlowRAG, a framework that reconstructs the flat retrieval space into a corpus-adaptive semantic gradient graph. This data-driven self-organization enables a hierarchical structure to emerge naturally from the data distribution, capturing the intrinsic semantic granularity of the corpus to suppress structural noise. By quantifying the semantic abstractness of entities through the embedding variance of their associated passages, we transform static undirected edges into directed semantic constraints. Furthermore, we design an abstractness-guided directed PageRank algorithm that forces the retrieval trajectory to follow a "high-to-low semantic abstractness" gradient. This mechanism ensures layer-by-layer evidence convergence, smoothly guiding the retrieval process from abstract concepts to specific document evidence. Extensive experiments on complex QA datasets demonstrate that SemFlowRAG effectively mitigates the "probability black holes" issue, outperforming existing baselines in both retrieval and downstream reasoning performance.

LLM-generated summaries of customer feedback increasingly shape purchasing decisions and organisational strategy. However, we find that such summaries often do not stay faithful to the customer voice, potentially undermining the validity of any downstream decisions made from them. In particular, through a systematic evaluation of 50,000 synthetic customer reviews across 50 domains and 11 models, we demonstrate that the same topics are amplified across different models, regardless of their prominence in the actual reviews. More strikingly, models continue to amplify these same topics even when they are entirely absent from the input, and the topics that are amplified correlate with those the model independently rates as most important. Our findings were validated on real-world reviews from four domains. Furthermore, no intervention we tested, including scaling to frontier models, prompt engineering, and varying temperature, summary length, and input size, reduces the bias to acceptable levels.

A standard pipeline for symbolic reasoning over natural-language problems translates them into first-order logic and invokes a theorem prover. The translation step is the bottleneck: swap "every" for "some" and every inference that follows is corrupted. Yet today’s metrics often score more broken translations higher than less broken ones, because BLEU, BERTScore, and Smatch++ reward surface overlap that the worst errors happen to preserve. We introduce SIV, which derives two kinds of probes from the target formula and uses a theorem prover to verify the candidate translation against each. Positive probes are statements the candidate must entail, which detect translations that drop content; contrastive probes are statements the candidate must not entail, which detect translations that assert more than the original. On a controlled pool of perturbed FOLIO translations, the severity of the error accounts for 80% of SIV’s score variance, compared with at most 17% for any prior metric. Across six error classes on a disjoint pool, SIV scores the reference above the perturbed candidate in over 99% of pairs. Because each probe is labeled with what it tests, the failure pattern also supplies a labeled error trace, recovering the perturbation class at macro-F1 0.638, nearly double the score-only baseline. On 434 expert-audited real LLM translations, SIV attains the top AUC, uniquely detects and grades expert-labeled major errors, and abstains, rather than mis-scoring, on out-of-vocabulary translations.

Retrieval-Augmented Generation (RAG) streamlines long-document understanding by leveraging retrieval mechanisms to restrict input images to a highly curated subset. However, existing multimodal RAG pipelines primarily face two critical challenges: first, standard semantic similarity retrievers frequently fetch topically overlapping yet answer-void distractor pages that mislead downstream generation; second, rigid single-pass pipelines heavily depend on initial retrieval success, where any omission of core evidence inevitably causes cascading errors. To address these challenges, we introduce **HiEvi-RAG**, a hierarchical, evidence-driven multimodal RAG framework for closed-domain document understanding. HiEvi-RAG systematically factorizes complex queries into a cooperative four-stage pipeline: (1) hierarchical question decomposition to break multi-hop root queries into atomic child questions; (2) coarse visual page retrieval leveraging a multimodal retriever to fetch candidate pages based on semantic similarity; (3) fine-grained page verification via EviAgent, a specialized multi-page verifier trained with GRPO to execute cross-page reasoning over multi-image blocks; and (4) memory-guided iterative generation that leverages accumulated sub-question context to execute multi-round, dynamic reasoning over the prioritized sequence. Extensive evaluations across four benchmarks demonstrate the robust efficacy and synergy of our framework, which significantly outperforms existing open-source baselines and exceeds the previous state-of-the-art by an average of 8.05% in accuracy.

While large language models (LLMs) perform strongly after adaptation within individual scientific domains, their learning dynamics in multi-disciplinary contexts remain poorly understood, despite the promise of improved generalization and broader applicability through cross-domain knowledge synergy. We present a systematic study of multi-discipline LLM fine-tuning across five scientific disciplines and compare the learning dynamics of full fine-tuning, LoRA, LoRA-MoE, and LoRA composition. In our naturally imbalanced setting, we show that multi-disciplinary learning is substantially more variable than single-discipline training and distill four empirical patterns: (1) Balance-then-Diversity: increasing scarce-domain exposure helps, but broad source-pool diversity must be preserved; (2) Merge-then-Align: general instruction data restores instruction-following behavior needed for effective joint adaptation; (3) Optimize-then-Scale: increasing trainable capacity offers limited gains without prior design optimization; and (4) Share-then-Specialize: asymmetric LoRA-MoE improves multi-discipline performance through a shared low-rank projection. Together, these patterns form a practical recipe for principled multi-discipline fine-tuning and provide actionable guidance for developing generalizable scientific LLMs.

Privacy-preserving adaptation of pretrained language models requires continual fine-tuning as private client distributions shift, while protecting individual samples under local differential privacy (LDP). In streaming differentially private federated learning, full-rank Low-Rank Adaptation (LoRA) adapters expose many trainable directions to DP noise, and later domains may overwrite directions that remain relevant to earlier domains. We propose Private Rank-Invariant Subspace Memory (PRISM), a fixed-rank LoRA framework that schedules the active plastic rank over the stream and applies shrinkage-gated memory writes to DP-protected aggregate updates. Server-side consolidation stores retained high-SNR directions in an orthogonal memory and remains pure post-processing of LDP updates. On MNLI genre and Sentiment3, PRISM stabilizes adaptation under strict privacy budgets while preserving a strong public initialization. At 𝜖=2 , it reaches 79.8% and 92.3% final average accuracy, compared with 52.3% and 85.0% for full-rank DP-FedSVD. Ablations indicate that rank scheduling accounts for most strict-budget recovery, and shrinkage-gated memory contributes a smaller retention gain and modestly improves over the strongest static low-rank control on MNLI.

Evidence grounding makes generated content verifiable by linking statements to their sources. However, effective verifiability requires two properties that existing methods fail to jointly provide: *verification granularity*, the ability to localize fine-grained evidence for each statement, and *verification completeness*, the ability to capture both support and contradiction. Existing methods either operate at granularities too coarse for rapid inspection or are support-only. We formalize *polarity-aware evidence grounding*: linking each response sentence to its contextual evidence with a supporting, contradicting, or irrelevant label. We construct **eTracer-Bench**, a human-annotated benchmark with 7,905 supporting and contradicting sentence pairs across seven domains, covering both explicit and implicit contradictions. We further propose **eTracer**, a post-hoc framework whose context-aware signed scorer recovers the full evidence map in a single forward pass per response, with localization and polarity emerging jointly from one mechanism. **eTracer** achieves the best overall Citation F1 of **82.39**, driven by a Contradict F1 of **89.20**. A pre-registered within-subjects user study (n = 30) shows that polarity-aware grounding nearly doubles contradiction detection over the document-level baseline (**27.9% → 51.9%**, Cohen’s d = 0.82, p = 0.0003), and suggests a false-completeness bias under which support-only grounding fails to improve over this baseline (**23.4% vs. 27.9%**). Dataset and code: https://github.com/chubohao/eTracer.

As large language models (LLMs) are integrated into numerous applications, LLMs’ safety becomes critical for both application developers and intended users. Currently, great efforts have been made to develop safety benchmarks with fine-grained taxonomies. However, these benchmarks’ taxonomies are disparate with different safety policies. Thus, existing safeguards trained on these benchmarks are either coarse-grained to only distinguish between "safe” and "unsafe,” or constrained by the specified narrow risk taxonomies. To leverage these fine-grained safety policies across multiple safety taxonomies, we propose GSPR, a Generalizable Safety Policy Reasoner to identify unsafe inputs and outputs with violated safety taxonomies and concise explanations. Unlike prior safeguards which only cover a fixed set of risk factors, GSPR incentivizes its reasoning capability with varied safety taxonomies through reinforcement learning. Our GSPR can be trained across multiple safety benchmarks with distinct taxonomies and naturally exhibits powerful generalization ability. We conduct extensive experiments to show that GSPR significantly improves existing safety guardrails’ reasoning capabilities for both safety and category prediction tasks. Moreover, GSPR also achieves the least inference token costs with explanations.

Large language models (LLMs) are increasingly deployed in applications with societal impact, raising concerns about the cultural biases they encode. We probe these representations by evaluating whether LLMs can perform author profiling from English-translated song lyrics in a zero-shot setting, inferring singers’ gender and ethnicity without task-specific fine-tuning. Across several open-source models evaluated on more than 10,000 lyrics, we find that LLMs achieve non-trivial profiling performance but demonstrate systematic cultural alignment: most models default toward North American ethnicity, while DeepSeek-1.5B aligns more strongly with Asian ethnicity. This finding emerges from both the models’ prediction distributions and an analysis of their generated rationales. To quantify these disparities, we introduce two fairness metrics, Modality Accuracy Divergence (MAD) and Recall Divergence (RD), and show that Ministral-8B displays the strongest ethnicity bias among the evaluated models, whereas Gemma-12B shows the most balanced behavior. Our code is available on GitHub and results on HuggingFace.

Large language models typically undergo post-training to align them with safety policies but there exist many sophisticated jailbreaks that sidestep established safeguards. For instance, prior work by Andriushchenko et al. (2025) has found that changing the grammatical tense from present to past can be enough to elicit harmful responses. In this work, we uncover a more general failure of non-imperative syntactic forms. We demonstrate that this _syntactic vulnerability_ exists in 16 models up to 70B parameters, using behavioral evaluation. To investigate the root cause, we apply causal mediation analysis, finding that refusal is partially conditioned on upstream syntactic features. By steering these purely syntactic features we are able to trigger and suppress refusal. Finally, we trace this ill-conditioning to linguistically biased post-training data of open-source models and show that increasing syntactic diversity can mitigate the issue. Our findings suggest that current alignment approaches introduce confounders that prevent a pure semantic grounding of the refusal decision.

Current data science (DS) code generation benchmarks equate correctness with quality, overlooking execution time differences that span orders of magnitude between correct solutions. We introduce DSEffi-Bench, the first benchmark specifically targeting execution efficiency in LLM-generated DS code, comprising 1,000 instances across 10+ DS libraries with stress-testing harnesses and human-validated references. Evaluating 16 models across 3 tiers, we find that correctness alone fails to characterize efficiency: GPT-5.4 leads in correctness (Pass, 66.9%) but its efficiency score (B|P, 71.7%) nearly matches GPT-5.4-mini (71.6%), which solves 47 fewer tasks; Kimi-K2.5 ranks lowest in correctness among frontier models (40.2%) yet achieves the highest efficiency score (73.6%) across all 16 models. A human-annotated five-category taxonomy reveals that 79.1% of efficiency deficits extend beyond algorithmic complexity to domain-specific root causes, with distinct failure profiles across model tiers and libraries. Two exploratory experiments provide initial evidence that these diagnostics can guide improvement, yielding up to +14.7% efficiency gains via taxonomy-guided optimization and approaching Claude-Opus-4.6 Best@3 in efficiency at 13.0× lower cost via library-conditioned routing. Code and data are available at https://github.com/Albert-Gong/DSEffi-Bench.

Free-text explanations extend human label variation (HLV) beyond label disagreement by revealing the reasoning and preferences behind annotators’ decisions. We study whether large language models (LLMs) can learn and reproduce such annotator-specific label-explanation behavior. Using two sentence-pair tasks with four annotators each—natural language inference and paraphrase judgment—we first analyze whether annotators exhibit stable individual patterns. We find that such patterns are weak at the single-annotation level due to strong input-content effects, but become detectable after input-content reduction and annotator-level aggregation. We then compare prompting and supervised fine-tuning (SFT) baselines and propose cross-annotator preference optimization (CAPO), which contrasts a target annotator’s response with other valid but less target-specific annotations for the same input. Experiments show that prompting is limited and unstable, SFT better captures annotator-specific behavior, and CAPO further improves aggregation-aware imitation and judge-based attribution while preserving target-specific reasoning patterns under human validation. Overall, our results show that HLV can be learned as annotator-specific label-explanation behavior, suggesting a path toward scalable explanation-based annotation grounded in annotator histories rather than labels alone.

Word alignment has traditionally been studied between sentences, but many cross-lingual tasks increasingly require correspondences across full documents. While recent multilingual embedding models can encode long inputs, we show that applying algorithms designed for sentences directly to documents leads to performance degradation. To address this, we introduce CTFAlign, a lightweight, training-free approach for document-level word alignment. CTFAlign applies a coarse-to-fine refinement strategy that restricts the alignment search space to semantically similar regions. Additionally, we introduce MDPAlign, a simpler alternative that constrains alignments by position with a main diagonal prior. Both approaches operate directly on full documents without relying on sentence segmentation or sentence alignment. We evaluate these methods across six language pairs varying in typological distance, resourcedness, and document length. Averaged over three models, CTFAlign reduces word alignment error rate from 0.412 to 0.326. These gains transfer downstream, leading to improvements in document-level translation coverage evaluation and recognition of semantic differences. We release CTFAlign as a Python package and make the code and data to reproduce our experiments publicly available.

Fine-tuning large language models (LLMs) on narrowly harmful datasets can lead to misalignment broadly, a phenomenon known as emergent misalignment (EM). EM poses a challenge for AI safety and our understanding of LLMs. Prior work often frames EM as an unexpected behavior, and explains it by appealing to general misalignment directions or anthropomorphizing it as acquiring an evil persona. However, the mechanisms behind these framings remain obscure. In this work, we show that EM is a predictable and data-dependent generalization phenomenon. By examining the base model’s representation of EM training data and evaluation prompts, we find that evilness after EM training is highly predictable from representational distance: the closer an evaluation prompt is to training data centroid, the more evilness it elicits from EM models after training (with an average Spearman correlation of -0.73 across 12 model-dataset settings). Building upon this analysis, we further demystify EM by showing that (1) its effectiveness changes significantly based on training data format; (2) there is not a general misalignment direction that transfers across different EM models; (3) the effect of EM is fundamentally different from persona changes. Furthermore, we extend the EM generalization metric from a scalar distance to a dataset-specific generalization direction, which robustly predicts EM models’ evilness under semantics-preserving prompt perturbations including appending random tokens and paraphrasing, where other methods do not reliably generalize.

Activation steering offers a lightweight approach to LLM safety, yet existing methods relying on static, linear perturbations often struggle with complex boundaries, leading to severe safety-utility trade-offs such as over-refusal. To address this, we introduce WassSteer. First, we propose a Safety Routing Mechanism that achieves context-aware sparse control by selectively intervening only on unsafe activations. Second, we reformulate steering as a transport problem on the Wasserstein Representation Manifold. Utilizing Flow Matching, WassSteer learns a non-linear velocity field to transport harmful activations toward safe regions along Wasserstein Geodesics, effectively neutralizing toxicity while maximally preserving semantic content. Experiments across Llama-3, Gemma-2, and Qwen-2.5 demonstrate that WassSteer significantly outperforms baselines against diverse jailbreaks while maintaining high performance on general benchmarks, effectively reshaping the safety-utility Pareto frontier.

Simultaneous speech-to-text translation (SimulST) generates translations while speech is still unfolding, requiring a streaming policy that decides when to read and when to write. State-of-the-art approaches rely on attention-based encoder-decoder models where cross-attention provides explicit alignment signals. In contrast, Speech Large Language Models (SpeechLLMs) are decoder-only architectures relying solely on self-attention. This raises a central question: whether decoder self-attention contains sufficiently stable alignment signals to guide the streaming policy. Moreover, existing approaches typically rely on training-based adaptations or heuristic wait-k policies and have not been validated in long-form settings. To fill these gaps, we propose Decoder-Only Attention (DOA), a training-free policy that enables long-form simultaneous translation with off-the-shelf SpeechLLMs by deriving a proxy alignment from self-attention. Experiments on Phi4-Multimodal and Qwen3-Omni show that DOA provides an effective alignment signal for supporting streaming decisions, enabling low-latency long-form SimulST with quality close to offline decoding without retraining.

On-policy distillation (OPD) improves reasoning by training a student on trajectories sampled from its own policy under supervision from a teacher. In multimodal reasoning, a common extension is to use a privileged teacher that observes training-time-only signals such as reference answers or rationales. However, such answer-side privilege creates a train-test mismatch: the teacher’s supervision may depend on signals unavailable to the student, encouraging shortcut imitation rather than visually grounded reasoning. We propose ViCuR, a visually grounded privileged-teacher distillation framework that replaces answer-side privilege with visual cues (query-related evidence in the input). Because these cues are derived from the same visual input available at inference, their evidence is recoverable by the student. To support this, ViCuR introduces a lightweight cue recovery module that uses dedicated sink-token cross-attention during prefill to aggregate task-relevant visual evidence into an internal representation, without changing the inference interface or requiring auxiliary cue-generation losses. Across seven benchmarks with Qwen3-VL-2B and 8B students, ViCuR consistently improves over answer-based on-policy self-distillation by +1.19 and +1.24 on overall average performance. It also extends naturally to stronger-teacher OPD, surpassing OPD baselines by +0.64 and +1.08, with consistent out-of-domain gains at the 8B scale. These results show that, in multimodal on-policy distillation, the design of teacher privilege is as important as teacher strength.

Tables are ubiquitous across diverse domains, yet reasoning over them remains a significant challenge for modern large language models (LLMs). Current approaches typically linearize tables into sequences, inherently overlooking their intrinsic two-dimensional and hierarchical structure. To address this, we propose H2Table (Hierarchical Hypergraph-Enhanced Table Reasoning), a novel framework that represents complex tables as hierarchical nested hypergraphs. To process this representation, we design a tailored hypergraph encoder to facilitate message passing between hyperedges (headers) and nodes (cells), thereby perceiving the semantic entailment relationships between them within complex tables. Furthermore, we introduce a set of learnable query vectors acting as a lightweight bridge to extract representative structural embeddings from the encoder into the LLM. Experimental results demonstrate that our approach effectively handles complex table question answering tasks with hierarchical nested headers. Notably, on the HiTab dataset, H2Table achieves an average improvement of 22.88% over state-of-the-art baselines on highly complex tables with a nesting depth of four. Our code is available at: https://github.com/lila120/h2table.

Post-training quantization (PTQ) has become a key approach for large language model (LLM) compression. Existing methods commonly adopt block-level reconstruction as a surrogate to calibrate each quantized Transformer block. However, our analysis shows that this loose objective can admit cancellation patterns, where the aggregated block-level reconstruction error is small even though the per-head errors remain large. We therefore revisit the PTQ formulation from a variational perspective. By treating per-head representations as latent variables and the block output as the observed variable, we find that conventional block-level reconstruction corresponds to optimizing only the observation-level term, while omitting a latent alignment term that captures head-level quantization error. Motivated by this finding, we propose HEAL, a Latent HEad ALignment framework that turns block-level reconstruction calibration into joint observation-latent calibration with stop-gradient head alignment. Experiments on LLaMA-2/3 and Qwen2.5 families under low-bit weight-activation quantization show that HEAL generally improves perplexity and zero-shot accuracy over block-level reconstruction baselines.

Long-horizon reinforcement learning for information seeking agents remains difficult because terminal rewards reveal whether the final answer is correct, but not which acquired information enabled it. This difficulty is amplified by text-derived webpage observations, where parsing, truncation, and summarization often produce incomplete and unstable content representations across trajectories. We propose an evidence-centric framework for web agent learning that represents information acquired through tools as identifiable units for comparison across trajectories. In particular, fetched webpages are represented as rendered snapshots, preserving layout and multimodal content as stable content-level observations. Building on these units, we introduce Information-Aware Credit Assignment, a post hoc reward propagation method that estimates turn-level utility scores from rollout success rates and assigns dense rewards to intermediate steps that introduced high-utility information. Integrated with GSPO, our method consistently improves performance on BrowseComp, GAIA, Xbench-DS, and Seal-0. Code and datasets will be released at [https://github.com/pc-inno/ICA_MM_deepsearch](https://github.com/pc-inno/ICA_MM_deepsearch).

Large language models (LLMs) exhibit remarkable reasoning capabilities, but their task-specific fine-tuning is notoriously plagued by overconfidence, severely hindering trustworthy deployment. We propose Data-Adaptive Lower-Rank Adaptation (DALorRA), a simple and effective variational Bayesian sparse framework that shifts the paradigm of uncertainty quantification from the dense parameter space to the lightweight rank level of low-rank adaptation (LoRA). With the insight that LoRA essentially aggregates multiple rank-one components that may provide superfluous model capacity, DALorRA imposes stochastic masking on rank dimensions, enabling Bayesian regularization of model capacity during training and ensemble-like calibration during inference. Extensive experiments demonstrate DALorRA’s excellent calibration of LLMs without compromising reasoning accuracy.

Document-level translation remains one of the most challenging tasks for large language models, which are constrained by limited context windows that impede global cohesion, while simultaneously suffering from redundant contextual information that degrades translation quality. To address this, we propose a human-like long document translation agent called Loong, which leverages a 3E memory module (Essence-Exemplar-Entity) to store summaries, sentence pairs, and entity records as historical context. Instead of passively attending to all history, Loong performs deep reasoning to adaptively identify the optimal context for translation guidance. Loong optimizes its context policy through reinforcement learning, utilizing preference data derived from its own sampled observe-and-act reasoning trajectories. Empirical evaluations demonstrate that Loong achieves substantial translation quality improvements in English ⇔ Chinese, German, and French directions, with average gains of up to 13.0 points across the three evaluation metrics. Furthermore, Loong exhibits strong generalization across domains and robustness against contextual noise, while maintaining remarkable stability in ultra-long document translation. Our code is freely available at https://github.com/YutongWang1216/LoongDocMT.

Knowledge distillation (KD) is a technique to compress a larger teacher system into a smaller student. In machine translation, KD is commonly evaluated through translation quality and inference efficiency, without jointly accounting for the environmental costs of producing and deploying the distilled system. We evaluate representative KD methods both on bespoke MT models and LLMs, by considering both translation quality and computational cost, using the Machine Learning Life Cycle Assessment tool, which accounts for costs throughout the KD model life cycle. Our key finding is that the deployment volume required to amortize KD is serving-dependent and can shift by several orders of magnitude under batching. We include actionable guidance for selecting, developing, and evaluating KD methods under quality and compute-induced constraints.

Multimodal Large Language Models (MLLMs) achieve strong vision-language reasoning but incur large KV caches and high decoding latency with long visual contexts. Existing compression methods rely on observation window attention for stable token importance estimation, yet this aggregation can dilute sparse critical evidence and discard answer-relevant tokens under aggressive compression. We identify last query attention as a complementary signal for recovering such evidence, though its irrelevant signals may introduce additional noise. We propose BACON, a plug-and-play method that calibrates observation window attention with last query evidence while suppressing noise through intra-layer coherence and inter-layer persistence. Across diverse benchmarks, models, budgets, and compression methods, BACON improves multimodal KV-cache compression by 7.5% on average under the most aggressive budget, with gains up to 30.9%.

Real-world image search queries are multimodal and compositional: “find this shirt in pink” specifies an entity to retain, an attribute to modify, and context to ignore. Yet existing re-rankers either compress such multi-faceted relevance into an opaque embedding or rely on free-form chain-of-thought that easily omits or hallucinates fine-grained constraints. Drawing on rubric- and checklist-based evaluation from NLP, we recast multimodal image re-ranking as a semantic constraint satisfaction problem and propose EviRank, which parses any query—text-only, image-only, or composed—into a unified evidence package: typed criteria across six semantic slots (e.g., entities, attributes, relations), each labelled required, forbidden, or ignorable. Re-ranking then reduces to evidence-conditioned verification, combining deterministic rubric scoring and evidence-grounded listwise comparison in a single training-free procedure. The explicit evidence can further serve as structured supervision for optionally distilling a lightweight student. Across five benchmarks spanning text-to-image, image-to-image, and composed image retrieval, EviRank achieves state-of-the-art performance, and the distilled student preserves over 90% of the teacher’s capability at substantially lower cost.

Existing LLM routing methods often construct supervision from a single sampled response for each query–model pair. Because LLM generation is stochastic, however, such an observation can be an unstable estimate of model capability: semantically equivalent query formulations and repeated decoding may yield different scores and even different model preferences. We show that this instability can further propagate from routing labels to learned routing policies. To address this issue, we propose **DARS** (Distribution-Aware Routing Supervision), which estimates query-level model capability from repeated observations spanning semantics-preserving query rewrites and stochastic decoding. DARS summarizes expected quality, expected cost, and performance variability to construct risk-aware supervision without changing the downstream router architecture. Experiments across diverse tasks and routing methods show that DARS generally improves routing utility and cost–quality trade-offs over single-shot supervision. Further analyses show that its benefits persist under moderate sampling budgets and different decoding temperatures. These results suggest that reliable LLM routing should move beyond individual sampled outcomes and instead model query-level capability distributions.

Clinicians read chain-of-thought (CoT) rationales as evidence of medical reasoning, but whether the visible chain plays that role is rarely tested. General-domain CoT-faithfulness probes ignore clinical cost, and medical LLM evaluations treat the chain as a black box. We close this gap with a medical perturbation audit: a 30-operator battery edits both the chain and the question with clinically motivated operators (severity reversal, negation flip, demographic swap, evidence ablation), paired with a chain-update × answer-flip joint analysis that classifies each model by its failure mode. Applied to 14 LLMs on four medical QA benchmarks, three independent tests converge: the Chain-Decoupling Rate (CDR; chain does not register the edit and the answer does not flip) is 72.9% panel-wide on M2/M4/M5 destructive edits, chain corruption leaves accuracy unchanged, and removing CoT prompting does not reduce accuracy. Two board-certified clinicians re-annotate N=197 perturbed questions; 98.5% leave the gold defensible. The pattern holds across medical and reasoning fine-tuning, scale, and the closed-source subset. Our framework and CDR provide a reusable yardstick for auditing whether medical CoT is faithful or merely documentation.

Multi-agent systems (MAS) powered by large language models have shown promise for complex tasks but suffer from high failure rates. Current self-reflection methods for MAS require all agents to reflect upon failure, overlooking a critical reality: failures typically stem from a specific agent leading the task astray, namely the decisive error agent, while others merely fulfill their regular duties. Forcing regular-behaving agents to reflect contaminates their memory with wrong insights. Hence, we propose DoCtOR (Diagnose-then-Correct PPO-enhanced Reflection), a novel reflection framework that enhances multi-agent collaboration. DoCtOR first identifies the decisive error step and decisive error agent through automated failure attribution, then employs counterfactual reasoning to generate a corrected decisive error step, and finally engages only the decisive error agent to produce targeted reflections. Experimental results show DoCtOR achieves 22%, 26%, and 27% improvements over initial success rates on HotPotQA, ChartQAPro, and Mind2Web datasets, outperforming Reflexion, Retroformer, and COPPER. We further establish the generalizability of our diagnose-then-correct paradigm and demonstrate that in low-resource settings, focusing reflection on reasoning steps after the decisive error step achieves comparable quality to reflecting on the complete failure trajectory.

Long-horizon interactions require language models to manage accumulating information: when to update their state, when to preserve their state, and what to ignore. We study this challenge as Contextual Belief Management (CBM): maintaining a predicted belief state aligned with formal evidence while isolating task-irrelevant noise. To make CBM measurable, we introduce , a closed-world benchmark spanning Rule Discovery and Circuit Diagnosis, where a finite belief space and symbolic verifiers enable exact turn-level evaluation. diagnoses three failures: Failed Stay, Failed Update, and Failed Isolation. Across multiple LLMs, vanilla models exhibit severe CBM failures, while explicit belief-tracking prompts provide limited gains. In contrast, reinforcement learning with belief-state rewards reduces failure rates by 70.9% on average. Further probing reveals latent belief-state dynamics behind these failures, and representation-level steering reduces failure rates by 46.1% across two tasks.

Claims about the grammatical competence of multilingual language models vary sharply with how competence is measured, yet the interaction between evaluation paradigm, post-training, and language resource availability has not been systematically examined. We evaluate base and post-trained models from six families on MultiBLiMP, a syntactic minimal-pair benchmark covering 101 languages, using four evaluation methods. We report three principal findings. First, post-training degrades grammatical competence, but the magnitude of this effect is determined by post-training recipe rather than model scale, and scaling reduces it unevenly, with low-resource languages bearing the highest cost. Second, post-trained models retain grammatical knowledge they cannot articulate through explicit prompting, yet this is measurable only in high-resource languages, because near-chance baselines in low-resource settings leave little knowledge to hide. Third, native-language prompting recovers otherwise hidden competence on low-resource languages, demonstrating that only high-resource languages can be probed directly from unprompted probabilities. We conclude that multilingual grammatical evaluation must adopt language-informed, multi-paradigm protocols to avoid systematically underestimating low-resource abilities.

Scheming occurs when a model recognizes goal conflict and strategically pursues an instrumental objective over user or developer interests, creating the risk of covertly misaligned behavior. Final outputs are the primary judgment signal, but they often cannot by themselves distinguish strategic deviation from ordinary error or routine harmful compliance; reasoning traces can provide auxiliary evidence about whether the model recognizes the conflict and adopts an instrumental strategy, making evaluation an observability problem. To mitigate this issue, we propose Cognitive Honeypots, which embed instrumental incentives, explicit goal conflict, and supervision cues into otherwise solvable tasks so that conflict-induced strategic deviation is more likely to surface without directly instructing deception or policy violation. We pair this capture mechanism with an intent–output decoupled protocol that labels High, Medium, and Low evidence. We instantiate it as SchemingBench, a bilingual benchmark with 4,000 honeypot scenarios across 5 behavior categories and 8 domains. Experiments across 12 models demonstrate that peak high-severity rate (HSR) reaches 70.76% in Chinese and 72.70% in English. In a matched ablation, all models show high HSRs under honeypot settings, with the largest increase in HSR reaching 65.66 points. Condition analyses further suggest that scheming-related evidence becomes most visible when instrumental incentives and explicit goal conflict are jointly present.

Arabic LLMs have improved rapidly, but aligning them with country-specific cultural behavioral preferences remains underexplored. Previous work has focused primarily on instruction-tuning for factual knowledge, overlooking nuanced cultural behaviors. We propose , which to our knowledge is the first large-scale pan-Arab preference data set for the alignment of cultural behavior conditioned on the country. Our contribution includes two resources in English and Modern Standard Arabic: a culturally grounded preference dataset for training and evaluation, and a multiple-choice benchmark designed to assess culturally aligned behavior across nations. All test data are validated by native speakers to ensure authenticity. To optimize cultural alignment, we experiment with post-training methods across multilingual and Arabic-centric language models, evaluating performance on both generation tasks and our cultural multiple-choice benchmark. By releasing both a training dataset and an evaluation benchmark, we provide a foundation for systematically studying and improving culturally aligned behavior in language models.

User modeling characterizes individuals through their preferences and behavioral patterns to enable personalized simulation with Large Language Models (LLMs) in contemporary approaches. However, existing methods, whether personalized prompt-based or parameter-based methods, face challenges in balancing personalization quality against computational and data efficiency. We propose a novel framework CURP, which employs a bidirectional user encoder, a discrete prototype codebook and a dual-aggregator to extract both multi-dimensional user traits and query-aware user characteristic. This design enables plug-and-play personalization with a small number of trainable parameters (about 28M parameters, about 0.36% of the total model size). Through extensive experiments, we show that CURP achieves superior performance and generalization compared to strong baselines on the language model personalization (LaMP) benchmark, while offering better interpretability and robustness.

Inference-time scaling (ITS) for latent reasoning requires sampling multiple candidate latent trajectories, yet current methods typically rely on heuristic perturbations such as dropout or fixed Gaussian noise. We show that stronger perturbations do not necessarily improve candidate-set quality: they often induce larger distribution shifts without producing more decision-useful reasoning paths. This limitation arises because heuristic perturbations inject stochasticity through fixed or hand-designed mechanisms rather than an optimizable, context-conditioned sampling policy, making latent exploration difficult to control. We focus on the sampling side of latent ITS and propose the Gaussian Thought Sampler (GTS), a lightweight module that predicts context-dependent Gaussian perturbations over continuous reasoning states. Trained with GRPO-style policy optimization while keeping the backbone frozen, GTS turns heuristic perturbation into a trainable, context-conditioned sampling policy. Experiments across two latent reasoning backbones and multiple benchmarks show that GTS consistently improves pass@N over heuristic baselines, suggesting that effective sampling for latent ITS requires controlled, optimizable exploration rather than simply amplified stochasticity.

Instruction-tuning datasets for large language models (LLMs) are often large, redundant, and imbalanced, hindering efficient adaptation. When such data are used naively in large-batch fine-tuning, overrepresented sample groups are repeatedly included while underrepresented but informative ones remain weakly covered; in practice, large-batch training is typically implemented with data parallelism (DP) across multiple GPUs. We propose **CluSTER**, a **Clu**ster-aware balanced **S**ampling framework for **T**raining **E**fficient data **R**eduction in DP instruction tuning. We curate a representative reduced dataset through gradient-space clustering and DP-aware balanced allocation, ensuring dual-level coverage across clusters and workers, while cluster-weighted updates preserve the original data distribution. As a result, **CluSTER** reduces redundant computation and improves training stability without compromising model quality. Across multiple instruction-tuning datasets, **CluSTER** reduces training time by 64.6%, almost without accuracy loss compared to prior sampling and data reduction methods. Code is available at https://anonymous.4open.science/r/CluSTER.

As Large Language Models rapidly advance, performance on standard multiple-choice question answering (MCQA) benchmarks is reaching saturation. While the community has responded by developing increasingly difficult datasets, validating question quality and filtering flawed items remains a labor-intensive process. To provide a scalable diagnostic approach, we propose a two-component probabilistic framework for auditing MCQA benchmarks using model output distributions. First, for benchmark-level analysis, we characterize the probability landscape using the top prediction probability (P_top1) and normalized residual entropy (H_norm), summarized globally by Mean Pairwise Distance (MPD). Second, for item-level diagnostics, we introduce noise injection to reduce meaningful distractor competition, enabling us to flag candidate items for targeted human review and categorize residual failure patterns. Across four MCQA benchmarks, our landscape analysis reveals benchmark-level differences in model confidence and residual option competition. Concurrently, our noise-injection method flags potentially actionable item-level issues, showing alignment with expert error annotations from MMLU-Redux. These results suggest that our probability-based framework provides a lightweight audit lens for comparing macro-level benchmark structure and prioritizing individual items for targeted human review.

Frozen vision-language models already encode object, attribute, relation, and caption information, but a single vector exposes all of these signals through the same access pattern. We ask whether vector length can become a semantic access control: short prefixes answer coarse queries, while longer prefixes add finer distinctions. GraSP-VL learns one shared near-orthogonal transform for image and text embeddings and assigns a fixed, configurable prefix contract. At inference, the transform is applied once and ordinary cosine retrieval uses the prefix required by the task. On a 20,147-example COCO/Flickr30K pool, GraSP-VL reaches a staircase score of 53.01 and hard-negative selectivity of 89.76 with full-space drift below 10^-6. Full-dimensional ImageNet-V2 classification and Flickr30K/COCO Karpathy retrieval are unchanged, while held-out human-caption retrieval improves at shorter prefixes. These results show that length can reorganize access to frozen VLM semantics rather than merely compressing them. Code and project page are available at https://lizesheng13.github.io/Grasp-VL/.

Reasoning-oriented language model training relies on traces that expose learnable dependencies, yet existing data-selection criteria use length, perplexity, difficulty, or judge-assigned quality. We argue that useful reasoning traces should also be sensitive to dependency order: intermediate steps are easier to encode in their natural premise-to-conclusion order but less predictable when dependency-revealing order is perturbed.

We introduce Local-DeltaAvg, a training-free data-selection metric. For each reasoning trace, we construct a semantically reversed counterpart that preserves problem content while perturbing dependency-revealing order. A fixed probe model scores both trajectories, and Local-DeltaAvg computes a paired local negative log-likelihood gap between the original and reversed contexts. The resulting probe-dependent score estimates the conditional encoding cost induced by order perturbation and identifies where this sensitivity is concentrated.

We evaluate Local-DeltaAvg as a relative ranking signal for reasoning-oriented continued-pretraining data selection, followed by an identical downstream SFT protocol. Under matched token budgets, it consistently outperforms length-, perplexity-, difficulty-, and judge-based selection across diverse reasoning benchmarks. Control analyses show that semantic reversal is essential and that high-gap regions align with annotated critical reasoning steps. These results suggest that dependency-order sensitivity is a lightweight and effective proxy for reasoning-data utility.

Text-based role-playing models can imitate character styles, but often fail to capture scene atmosphere and evolving tension, which are crucial for immersive applications such as VR games and interactive narratives. We study video-grounded role-playing dialogue and introduce EBM-RL (Eye-Brain-Mouth Reinforcement Learning), a decoupled GRPO-based framework that separates observation ( lt;perception gt;), reasoning ( lt;think gt;), and utterance generation ( lt;answer gt;). This design mimics the human See-Think-Speak process, enabling the model to ground dialogue in visual perception before reasoning and response generation. To optimize this See-Think-Speak process, EBM-RL integrates complementary rewards for scene-text alignment, perceptual-cognitive utility, answer faithfulness, and format consistency. Extensive experiments show that EBM-RL substantially outperforms text-only role-playing baselines and comparable-scale VLM baselines, while remaining competitive with stronger larger-scale VLMs on visual grounding and situational persona metrics. Moreover, EBM-RL demonstrates strong zero-shot transfer to out-of-domain VideoQA benchmarks without additional fine-tuning. We also release an open-source dataset for video-grounded role-playing dialogue.

Autoregressive vision-language models achieve strong multimodal understanding, but their sequential decoding becomes increasingly inefficient as multimodal responses grow longer. Adapting AR-VLMs to diffusion-style block decoding is promising, yet fixed refinement blocks struggle with sharply varying multimodal uncertainty in OCR, chart, document, and reasoning-heavy outputs. Binary mask-or-token states further limit local refinement, as masking discards partial evidence while premature commitment can corrupt the causal prefix. We present FLUID-V, a lightweight AR-to-diffusion adaptation framework for reliable local multimodal refinement. FLUID-V selects refinement spans with elastic horizons and represents unresolved positions as evolving soft semantic mixtures in embedding space. Only confident predictions are committed to the persistent KV cache, preserving causal multimodal conditioning and autoregressive compatibility. Experiments with OpenPangu-VL-7B and 100K self-distilled samples show competitive performance across diverse multimodal benchmarks, with ablations confirming the importance of adaptive horizons and stateful intermediate representations. Our code is available at https://github.com/Oli-lab-nun/FLUID-V.

We investigate whether independently trained language models converge to geometrically compatible latent representations, and whether this compatibility can be exploited to correct model behavior at inference time without any weight updates. We learn a linear projection matrix that maps activation vectors from a large teacher model into the coordinate system of a smaller student model, then intervene on the student’s residual stream during generation by substituting its internal state with the translated teacher representation. Across a fully crossed experimental matrix of 20 heterogeneous teacher-student pairings spanning mixture-of-experts, dense, code-specialized, and synthetically trained architectures, the Ridge projection consistently achieves R^2 = 0.50 on verbal reasoning and R^2 = 0.40 on mathematical reasoning, collapsing to R^2 = -0.22 under permutation control and R^2 = 0.01 under L_1 regularization. Behavioral correction rates range from 14.0% to 50.0% on TruthfulQA (mean 25.2%) and from 8.5% to 43.3% on GSM8K arithmetic reasoning (mean 25.5%), demonstrating that the method generalizes across fundamentally different reasoning domains. We report a near-zero correlation between geometric alignment quality and behavioral correction rate (r = -0.07), revealing a dissociation between representation space fidelity and output space impact. Intervention strength is architecture-specific: student models exhibit characteristic sensitivity profiles that invert across domains, with the most steerable verbal student becoming the least steerable mathematical student. Finally, a double dissociation experiment conducted across all 20 model pairings confirms without exception that projection matrices collapse catastrophically when transferred across reasoning domains (mean R^2 = -3.83 in both transfer directions), establishing domain-specific subspace geometry as a universal property of LMs.

Visual token pruning reduces the inference cost of vision-language models (VLMs), but most methods only ask which tokens to keep. This retained-token view can keep redundant high-scoring tokens while leaving discarded evidence without a close representative. We propose CoverPruner, a training-free pruner that asks the complementary demand-side question: after a token is removed, which surviving original token represents it for the target VLM? CoverPruner formulates pruning as Representational Coverage Maximization (RCM), covering the full projected visual-token set with query-weighted demand. It instantiates RCM with projector-space coverage and a lightweight first-layer attention probe. Across multiple VLM architectures and compression rates, CoverPruner achieves the best average accuracy among all compared methods, with the largest gains usually appearing under aggressive compression.

AI productivity is often measured by task completion time, economic value, or improvements in outcome quality. However, these measures usually treat collaboration as a black box where they capture what output was produced, but not the interaction cost required to produce it. Motivated by economics literature, we introduce a productivity-oriented framework for evaluating human-AI collaboration as outcome quality relative to interaction cost. Across two datasets spanning four tasks, we show that: (1) sessions with identical quality ratings can differ by up to 70× in interaction cost; (2) quality-cost relationships vary by task, with some tasks rewarding extended interaction and others favoring fast convergence; (3) subjective user ratings are not reliable substitutes for productivity; and (4) productive sessions are characterized by agents probing earlier and users spending less effort repairing the interaction. By distinguishing productive success from costly success, our framework makes interactional cost visible and shows how dialogue analysis can inform the evaluation and design of AI systems. We release our code and newly collected human-AI collaboration dataset for reproducibility.

Stance detection is crucial for understanding the underlying attitude of an expression towards a target. Conversational stance detection is a more challenging stance detection task in real-world social media scenarios, as it involves detecting the user’s stance by leveraging the target-related historical statements across conversational sessions. In this paper, we propose target-aware Memory Graph TamGraph, a novel method that dynamically leverages target-related statements for conversational stance detection. Instead of considering all preceding historical conversations or using no prior conversation information for stance detection, our TamGraph employs a stepwise, entropy-guided backtracking mechanism to selectively activate memory from historical conversations and dynamically constructs a target-aware graph to model the stance relations among utterances. This allows the exploitation of target-related information from the conversation history for stance detection while preventing the introduction of noise. Experimental results on both English and Chinese benchmarks demonstrate that our TamGraph substantially improves LLM performance on conversational stance detection.

Late-interaction vision-language retrievers represent each document page as many visual token embeddings and score queries with MaxSim. In systems such as ColPali, ColQwen, ColNomic, and Nemotron ColEmbed, the document embeddings are produced without seeing the query, so the same page is represented identically for a table lookup, a chart question, and a layout-sensitive evidence request. We introduce Argus, a family of query-conditioned late-interaction retrievers built on Qwen3.5-VL. Argus adds a region-aware Mixture-of-Experts module: the query encoder produces both retrieval embeddings and a compact context vector, the document page is pooled into spatial regions, and a query-aware router selects latent experts per region before MaxSim. The output remains a multi-vector index compatible with ColPali-style retrieval, but the document representation is now dependent on the query (i.e., 𝐃(q)). All Argus models use a 1024-dimensional retrieval head, compared with the 2560-dimensional and 4096-dimensional heads of recent state-of-the-art systems, and are trained on roughly 9% of the available public supervision rather than the full pool. The 9B model reaches 92.67 NDCG@5 on ViDoRe V1 and 86.0 NDCG@5 on the combined V1+V2 leaderboard, the highest reported value for an open late-interaction model on the combined leaderboard. Wrapped in a Qwen3.6-27B agentic retrieval pipeline on ViDoRe V3, Argus-9B further improves its NDCG@10 from 60.28 to 64.80 over public tasks, showing that the same retriever serves both as a strong standalone system and as a search primitive for iterative LLM agents.

LLM-based agent systems increasingly adopt skill-based architectures to reduce repetitive reasoning costs and improve stable, efficient task execution. Recent studies propose self-evolving agents that autonomously generate, refine, and reuse skills from past experiences to enable continuous capability evolution. However, autonomous skill evolution introduces a new attack surface in which malicious capabilities are generated, stored, and reused as legitimate skills. In this paper, we define EvoSkill Injection as a threat model targeting the autonomous skill generation and evolution pipeline of self-evolving agents. We further propose SARGE (Red-teaming Autonomous Skill Generation and Evolution in self-evolving agents), a red-teaming framework for evaluating this threat model through iterative generation, escalation, and reinforcement interactions. To support our framework, we construct EvoSkillBench, a benchmark dataset of malicious interaction trajectories for inducing malicious skill formation in self-evolving agents, and introduce EvoSkillSafetyBench to assess whether injected malicious skills are retrieved and activated as harmful behaviors during downstream task execution. Our evaluation shows that SARGE induces malicious skill formation and that injected skills are persistently stored and repeatedly activated, highlighting the risk of persistent capability corruption.

Test-time scaling (TTS) improves language model reasoning by allocating additional test-time compute, but prior work mainly studies closed-ended tasks such as math and coding. We study TTS for automated equation discovery, an open-ended setting where models search over candidate equations and rely on observed datapoints for feedback. We formulate LLM-driven equation discovery as an iterative search process that unifies Best-of-N, sequential refinement, tree search, and evolution-style methods under a common compute-allocation view. To isolate allocation effects from prompt engineering and other heuristics, we compare minimal parallel controllers under fixed budgets. On LLM-SRBench equation-discovery tasks, we find that search width is the dominant allocation parameter: the best width in our sweep generally increases with the compute budget, while the population–branching split and controller choice matter less. Appropriate width selection also improves wall-clock efficiency by increasing parallelism. These results suggest that, given an informative verifier, controlling exploration and exploitation is central to scaling LLM-based equation discovery.

As users increasingly expect LLMs to align with their preferences, personalized information becomes valuable. However, personalized information can be a double-edged sword: it can improve interaction but may compromise objectivity and factual correctness, especially when it is misaligned with the question. To alleviate this problem, we propose PersonaDual, a framework that supports both general-purpose objective reasoning and personalized reasoning in a single model, and adaptively switches modes based on context. PersonaDual is first trained with SFT to learn two reasoning patterns, and then further optimized via reinforcement learning with our proposed DualGRPO to improve mode selection. Experiments on objective and personalized benchmarks show that PersonaDual preserves the benefits of personalization while reducing interference, achieving near interference-free performance and better leveraging helpful personalized signals to improve objective problem-solving.

Prior work suggests that sentiment can often be captured by approximately linear directions in LLM activation spaces, but a single direction may not fully capture sentiment representations. In natural communication, sentiment is shaped not only by polarity but also by usage factors, such as tone and audience adaptation. We test whether these factors systematically modulate sentiment representations beyond a shared sentiment direction. We construct a controlled paired dataset that holds event content fixed while varying sentiment polarity and usage factors, and analyze Llama, Mistral, and Gemma. We identify a shared sentiment direction, remove it, and test the residual structure through erasure and generation-time tone steering. Across models, the shared direction is robust (median cosine 0.953–0.975), yet removing it leaves 0.833–0.909 of the original positive–negative representation-difference norm. The residuals contain compact, reproducible usage-conditioned structure. Targeted erasure weakens held-out usage metrics more than random and label-shuffled controls. On Llama, outputs steered along residualized tone components are preferred in 92.8% of blind target-tone comparisons while preserving the requested sentiment polarity in 98.7% of evaluated outputs.

Large language models (LLMs) can fluently verbalize statistical evidence, yet statistical reports can still drift numerical values, invert effect directions, or restate thresholded contrasts as categorical effects. We frame these failures as a control problem: the evidence-bearing content of a scientific report should be fixed by structured statistical results rather than sampled during prose generation. We therefore use cross-run reproducibility to stress-test whether report-visible numbers and claims are bound before prose generation. Existing controls operate at the text or slot level; a deterministic hybrid template reproduces only 61.1% of report-visible numerical content across seeds because the LLM still selects which findings and numbers the template renders. We propose claim-locked reporting, a provenance-before-prose protocol that fixes the evidence source, numbers, direction, and allowed language strength of each reportable claim before the LLM writes connective wording. Across fMRI functional-connectivity reporting and randomized controlled trial reporting on Evidence Inference 2.0, claim-locked reporting improves reproducibility over the hybrid template by 37.4 and 20.5 points, respectively. Blinded human audits support the observed direction-preservation and governance trends. In an fMRI cost analysis with DeepSeek, claim-locked reporting also yields the lowest observed token use and median generation latency. Code is available at https://github.com/XiuFan719/Claim-Locked-Reporting.

Existing summarizers for memory systems are typically optimized for human-facing criteria such as faithfulness, which misaligns with their true objective: preserving the evidence needed to support future queries. We show that conditioning summarization on query-answer pairs substantially improves answer quality, and that this utility-aware behavior is transferable across queries. Motivated by these findings, we propose MemSuit, a self-distillation framework in which a teacher summarizer, conditioned on observed query-answer pairs, produces utility-aware memory entries that a student learns to reproduce from the raw conversation alone. To prevent collateral erasure where conditioning on a single query-answer pair discards evidence relevant to other plausible queries, the teacher decomposes each block into multiple self-contained entries that preserve distinct query-relevant facets as independently retrievable units. To align the retriever with the compact, fact-dense style of teacher entries, we further fine-tune the embedding model with a contrastive objective supervised by teacher entries. Across a diverse suite of conversational query types, MemSuit consistently outperforms state-of-the-art baselines, confirming the value of grounding memory in downstream utility.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance across a wide range of multimodal tasks. Recent advances in personalized MLLMs enable effective capture of user-specific concepts, supporting both recognition of personalized concepts and contextual captioning. However, humans typically explore and reason over relations among objects and individuals, transcending surface-level information to achieve more personalized and contextual understanding. To this end, existing methods may face three main limitations: (1) Their training data lacks multi-object sets in which relations among objects are learnable, (2) Existing models often neglect the connections between different personalized concepts, thereby failing to perform reasoning over them, (3) Their experiments mainly focus on a single personalized concept, where evaluations are limited to recognition and captioning tasks. To address the limitations, (i) We present a new dataset named ReGraP, consisting of 120 sets of personalized knowledge. Each set includes images, Knowledge Graphs (KGs), and Chain-of-Thought Question-Answering (CoT QA) pairs derived from the KGs, enabling more structured and sophisticated reasoning pathways. (ii) We propose Reasoning enabled Graph-based Personalized Large Language and Vision Assistant (ReGraP-LLaVA), an MLLM trained with the corresponding KGs and CoT QA pairs, where soft and/or hard graph prompting methods are designed to align KGs within the model’s semantic space. (iii) We establish the ReGraP Benchmark, which contains diverse task types: Multiple-Choice, Fill-in-the-blank, True/False, and Descriptive questions in both open- and closed-ended settings. The proposed benchmark is designed to evaluate the relational reasoning and knowledge-connection capability of personalized MLLMs. We conduct experiments on the proposed ReGraP-LLaVA and other competitive MLLMs. Results show that the proposed model not only learns personalized knowledge but also performs relational reasoning in responses, achieving the SoTA performance compared with the competitive methods. All the codes and datasets are released at: https://anonymous.4open.science/r/ReGraP.

Low-rank (LR) pretraining underperforms full-rank (FR) pretraining, but why the gap persists remains unclear. Using GPT-2 on BabyLM, we show the representational split between them is not fixed at initialisation, spectrum-matched initialisation eliminates the representational gap at step zero, but it’s reasserted within a few training steps, implicating training dynamics over initialisation. We further show gradient dimensionality is a causally separable driver of this split, not merely a restatement of the weight factorization. Restricting or expanding gradient dimensionality of LR models directly shifts representational alignment with FR counterpart independent of weight rank. Finally, optimisation success and representational alignment dissociate, that is, spectral normalisation and other interventions improve perplexity without recovering FR-LR similarity. Also, we observered that representational alignment predicts downstream competence better than perplexity does. These results support gradient starvation as a genuine, partial mechanism, and argue efficient pretraining should be evaluated by both representational trajectory and performance, not solely the latter.

Multimodal fake news detection is critical for maintaining the integrity of social media. Current models often rely on shallow cross-modal correlations and rigid domain labels, which limits robustness to logical inconsistency and domain shift. We propose the Open-Vocabulary and Logic-Aware Fake News Detection (OVLAFND) framework to address these issues and relax traditional categorization constraints. To model fine-grained logical conflicts rather than surface relevance, our framework introduces Logical Adversarial Data Augmentation (LADA), a multimodal LLM-driven pipeline that constructs counterfactual text-image pairs as hard negatives. Then, we propose a Logic Consistency Module (LCM) that uses these adversarial signals to learn a dedicated logic-consistency subspace for assessing whether textual claims are supported by visual evidence. To improve topic flexibility, we transform the rigid domain labels into an Open-Vocabulary Topic Manifold (OVTM) to achieve highly interpretable and label-free topic reasoning. Across Weibo and Weibo-21, OVLAFND achieves competitive fixed-domain performance and better leave-one-domain-out generalization. Our code is available at https://github.com/lllfxy/OVLAFND.

In visual storytelling, a believable character must not only look consistent across scenes but also behave in ways that reflect their personality. Current text-to-image (T2I) models preserve visual appearance effectively, yet consistently fail to maintain personality-relevant behavior—a failure we term trait collapse. We show that conventional metrics (CLIP Score, DINO similarity) are blind to this failure, and that abstract personality descriptors yield no reliable improvement over neutral prompts regardless of abstraction level. To systematically diagnose trait collapse, we introduce ViStoryBench-Personality, a story-level benchmark combining an expert-validated Visual Cues Lexicon (Fleiss’ 𝜅 = 0.757) with a six-dimensional metric suite covering personality alignment, cross-scene stability, and behavioral coherence. Evaluating seven T2I systems, we confirm a persistent semantic gap between abstract trait labels and visual evidence, and find that visual identity consistency does not imply personality consistency. To bridge this gap, we propose Trait-Behavior-Visual (TBV), a training-free prompting framework that decomposes personality traits into renderable visual constraints. TBV improves Personality Alignment Score by up to +22.1 %, with larger gains in models with stronger language-conditioning capacity.

Computational resources are increasingly central to NLP research, but how closely reported GPU capability aligns with scholarly impact remains unclear. We analyze 13,921 ACL, EMNLP, and NAACL main-conference papers published between 2020 and 2025, using GPU resources as our operational measure of computational resources. From full texts, we extract GPU models and counts, standardize each paper’s largest reported configuration into a comparable hardware-capability measure, and link these data to citation, award, topic, and institutional metadata. GPU reporting became more common but remained incomplete, while reported capability increased mainly through newer hardware generations and medium-scale multi-GPU configurations. Resource concentration substantially exceeded impact concentration: the annual top 20% of GPU-quantifiable papers accounted for 83.9%–89.9% of reported GPU capability, but only 27%–32% of citations and 20%–33% of paper awards. In adjusted models, a tenfold increase in aggregate reported GPU capability was associated with a 3.52-percentage-point increase in within-NLP topic–year citation percentile, but increased model R² by only 0.0042. GPU count showed more consistent positive associations with citation and award outcomes than newer hardware generation. Overall, reported GPU resources are associated with scholarly impact but provide little standalone explanation of research influence.

NLP systems increasingly aggregate claims from retrieved passages, documents, and generated reasoning steps. Yet current factuality methods mostly verify claims individually or compare pairs of claims, leaving open the question of whether the entire aggregated claim set can be retained coherently. This misses a fundamental failure mode: claims may be locally plausible and pairwise compatible while still being jointly inconsistent. We formalize global consistency repair: given an inconsistent natural-language claim set, preserve the largest coherent subset while localizing the compact sources of inconsistency. We introduce QXR, an adaptive diagnosis-and-repair framework that treats an LLM as a noisy set-level consistency judge, localizes minimal conflicting subsets, and repairs the claim set by deleting a small hitting set of claims. Across LLM judges and benchmarks derived from VitaminC, FEVER, and conflicting-evidence RAG, QXR preserves substantially more valid claims than direct all-at-once judging while maintaining high precision. These results suggest that factuality in claim aggregation systems should be treated as a set-level repair problem, not only as local verification.

Large language models are increasingly used as personalized writing assistants, but adapting a model across many authors can compromise individual writing style by pulling author-specific signals toward a shared register. Federated parameter-efficient fine-tuning (PEFT) offers a data-local setting for this multi-author adaptation problem: clients keep author text local while sharing compact adapter updates. However, we show that standard aggregation can preserve continuation utility while making different authors’ generations less distinguishable in style space, a failure mode we define as author-style homogenization. We evaluate author-style retention with Angular Style Classification Encoder (ASCE)-based diagnostics on our main BlogText benchmark and ASCE-independent external authorship verification. Using this protocol, we find that common federated PEFT baselines can preserve semantic utility while averaging out author-specific signals. To address this homogenization, we instantiate FAVoR (Federated Authorial Voice Retention), an author-style residual mechanism for federated PEFT. FAVoR uses a shared-private adapter design: clients upload shared-adapter updates while retaining author-specific residual corrections locally. Across BlogText and external Mythos-Reddit validation, FAVoR improves author-style retention over standard and personalized federated PEFT baselines. These gains come with small continuation-utility trade-offs and are supported by component ablations, external verification, and cold-start transfer.

The heavy-tailed distribution of output lengths in Large Language Model (LLM) serving poses major challenges for resource provisioning and cluster scheduling. Although output-length prediction can mitigate these issues, existing approaches have key drawbacks: external proxy models add substantial latency and often have limited fidelity, whereas internal state-based methods are efficient but rely on shallow probes of current model states. We identify a structural connection between speculative decoding (SD) and length prediction: latent representations produced by the draft decoder in advanced frameworks (e.g., EAGLE-3) encode signals that are predictive of generation length. Building on this insight, we introduce **OUTLETS** (**Out**put-**Le**ng**t**h Prediction from **S**peculative Decoding Backbones), which repurposes the speculative backbone as a trajectory-aware length predictor. When its draft representations are already computed for speculative decoding, OUTLETS adds only a lightweight regression head and achieves lower MAE than the evaluated methods. Under saturated disaggregated serving, OUTLETS predictions enable standard scheduling policies to prioritize shorter requests and distribute requests more evenly across decoding instances, reducing short-request P99 latency by **34.8%**.

Sparse Knowledge Graphs (KGs) are commonly encountered in real-world applications, where knowledge is often incomplete or limited. Sparse KG reasoning, the task of inferring missing knowledge over sparse KGs, is inherently challenging due to the scarcity of knowledge and the difficulty of capturing relational patterns in sparse scenarios. Among all sparse KG reasoning methods, path-based ones have attracted plenty of attention due to their interpretability. Existing path-based methods typically rely on computationally intensive random walks to collect paths, producing paths of variable quality. Additionally, these methods fail to leverage the structured nature of graphs by treating paths independently. To address these shortcomings, we propose a Structural and Probabilistic framework named StruProKGR, tailored for efficient and interpretable reasoning on sparse KGs. StruProKGR utilizes a distance-guided path collection mechanism to significantly reduce computational costs while exploring more relevant paths. It further enhances the reasoning process by incorporating structural information through probabilistic path aggregation, which prioritizes paths that reinforce each other. Extensive experiments on five sparse KG reasoning benchmarks reveal that StruProKGR surpasses existing path-based methods in both effectiveness and efficiency, providing an effective, efficient, and interpretable solution for sparse KG reasoning.

Large language models (LLMs) are increasingly used as proposal generators for evolutionary robot design, yet most loops remain memoryless: simulator results shape the next population but are not preserved as reusable design knowledge. We present Auto-Robotist, a self-evolving LLM agent that distills morphology-search traces into an explicit natural-language skill library. Each skill stores a structural archetype, evidence-grounded positive and negative rules, and the evaluated designs that support them, making design memory inspectable rather than implicit in a population. During search, the agent retrieves skills to condition LLM edits of elite bodies while retaining a Genetic Algorithm (GA) mutation path for exploration; after evaluation, it updates the library through Add, Diagnose, and Merge. Across seven EvoGym tasks spanning locomotion, traversal, and object interaction, Auto-Robotist improves cold-start 5×5 search and transfers learned skills to 10×10 design spaces, where reference-conditioned transfer outperforms or matches GA on every task when averaged across seeds. These results suggest that LLM agents can convert expensive physical evaluations into reusable, auditable design principles. Our code is publicly available at https://github.com/wangyf9/Auto-Robotist.

Large language models (LLMs) are increasingly used for qualitative data analysis (QDA), yet their outputs often miss the depth and nuance of human analysis. We argue this gap reflects a missing credibility practice from human QDA: peer debriefing, in which an analyst seeks feedback from a disinterested peer and uses it to refine their coding. To bring this practice into LLM-assisted QDA, we propose Agent-as-Peer-Debriefer, a multi-agent QDA framework that builds peer debriefing into key coding steps. In our framework, a Hierarchical Coding Agent follows the standard QDA process to generate codes, sub-themes, and themes, along with self-explanations and reflection memos. It then shares these outputs with three Peer-Debriefing Agents, each applying a distinct analytical perspective (Theory-Driven, Data-Driven, or Applied) and refining the codes by keeping, renaming, reassigning, merging, or splitting them. These perspectives are drawn from established human QDA practices that generalize across domains and datasets. To evaluate the framework, we test it on three datasets across two domains with three LLMs, measuring semantic similarity to human-annotated codes. Across all settings, perspective-based, peer-debriefing refinement aligns more closely with human codes than a single-LLM baseline, and an ablation further shows the gain is not merely from additional refinement. The three perspectives also produce distinct trade-offs, showing that the choice of perspective is a meaningful and controllable design decision. More broadly, these findings suggest that simulating peer debriefing with explicit perspectives is a promising route to more credible LLM-assisted QDA.

The use of LLMs to complete online surveys impacts the validity of survey-based research, but detecting such usage remains under-explored. We introduce an initial benchmark dataset, namely ASURRE, for AI-assisted survey participation to capture usage strategies ranging from full generation and revision to persona-grounded agentic completion. Controlled by these strategies, LLM-assisted survey responses are generated using multiple LLMs on three real-world surveys in different disciplines, paired with genuine human responses. Our evaluation of existing machine-generated text (MGT) detectors shows that naive AI usage is readily detectable, whereas persona-grounded agents that mimic entire respondents push detector performance toward chance. We further show that agentic completion cannot fully replicate respondent-level behaviour and leaves distinctive behavioural traces. While individual cues can be circumvented by targeted prompting, a simple few-shot, training-free aggregator over these cues improves mean AUROC by +0.14 over the best existing detector across agentic settings. Our project is available on GitHub.

Language model (LM) agents are increasingly deployed to maintain codebases over extended periods, fixing streams of related defects while carrying context from one fix to the next. Yet existing software engineering (SWE) benchmarks evaluate models one bug at a time: the repository is reset, the codebase is re-read, and a single self-contained issue is graded in isolation. This setting collapses a continuous maintenance workflow into a series of independent sessions, ignoring the cumulative dependencies that make real-world bug fixing challenging. To bridge this gap, we introduce ChainSWE, the first benchmark for evaluating agents on sequential, dependent bug fixes within a shared codebase. We collect chronological chains of 304 issues across 54 Python projects, mined from six SWE-bench-family datasets. Our evaluation across a range of agents and models reveals a consistent performance drop by **up to 70%** as the chain length increases. Our data and code is available at [C-lister/ChainSWE](https://huggingface.co/datasets/C-lister/ChainSWE).

LLM-based sequential recommenders usually cast next-item prediction as text generation, but this interface is poorly matched to full-catalog top-K ranking. We propose TATK, a Triple-Aware framework that couples Top-K Learning (TKL) with Knowledge-Grounded Verification (KGV) for LLM-based sequential recommendation. Top-K Learning combines context-aware metadata-KG prompt grounding with position-aware top-K rewards, aligning training with ranking utility; Knowledge-Grounded Verification then applies structure-aware reranking over the top-M candidates after a single LLM forward pass, using the same metadata-derived item graph. We evaluate TATK on Musical Instruments, CDs and Vinyl, and Video Games from Amazon Reviews 2023 under a matched R²ec-style full-catalog protocol. Experiments use Gemma-2-2B-It and Qwen2.5-3B-Instruct backbones, compare against sequential, generative, KG-augmented, and reasoning-enhanced baselines, and include component, reward-shape, sequence-perturbation, reranking, relation-quality, and candidate-pool diagnostics. TATK improves over the matched R²ec reproduction on all 36 reported metrics. On NDCG@10, it improves Qwen by 8.05%, 4.26%, and 3.78% on the three datasets, and improves Gemma by 27.03%, 10.52%, and 10.23%, while keeping inference within 1.17× of Base RecPO latency. The diagnostics show that structural evidence is most useful for recoverable top-M candidates with reliable KG support, and should be gated when metadata relations are sparse or noisy.

Soft compression stores a text span as k continuous compressed vectors for a frozen decoder, but existing compressors are usually tied to the decoder used during training. We study whether those vectors can be retargeted to a different frozen decoder through its input-embedding interface under a controlled reconstruction protocol. Our benchmark covers seven target decoders, three vector budgets, and directed source–target pairs, comparing Random, shared-embedding least squares (LS), Converter-only, Enc-Conv, and target-native Origin. Converter-only maps vectors from a frozen source compressor, whereas Enc-Conv also adapts the source-initialized vector generator. Transfer is feasible but decoder-conditioned: trainable retargeting can approach Origin BLEU on decoder-compatible targets, while LS is often close to Random despite using shared token anchors. Enc-Conv consistently matches or outperforms Converter-only, with its largest gains on difficult transfers. Selected-pair multi-seed runs and a small SQuAD probe provide complementary evidence, although downstream utility remains target-dependent. Perturbation and geometry probes are consistent with decoder-specific readable regions. Finally, pair-level readability scores provide weak-to-moderate pre-transfer screening signals.

Multimodal large language models (MLLMs) have shown strong capabilities in perception, reasoning, and action generation. However, their ability to sustain exploration in dynamic open worlds remains unclear. Existing embodied and game-based benchmarks often compress interaction into short-horizon tasks or entangle success with domain-specific game mechanics. In this paper, we introduce MineExplorer benchmark for evaluating open-world exploration capabilities of MLLM agents in Minecraft. We first filter atomic tasks whose solutions rely heavily on Minecraft-specific knowledge to better reflect general open-world reasoning. Then we organize the benchmark around a ReAct-style capability formulation and compose atomic tasks into implicit multi-hop tasks. To further construct reliable instances, MineExplorer uses a multi-agent synthesis workflow that jointly designs task graphs, sandbox scenes, and rule-based milestone evaluators. Human evaluation shows that the multi-agent synthesis workflow produces significantly more reliable instances than a single-agent baseline. Experiments with advanced MLLM agents show that open-world exploration remains challenging, as strong models can handle many single-hop tasks but degrade sharply when hidden prerequisites must be coordinated over longer trajectories. Further analysis finds that task difficulty tracks agent completion, and larger models or thinking modes do not consistently translate into better performance. Code is available at anonymity.

Watermarking algorithms for Large Language Models (LLMs) effectively identify machine-generated content by embedding and detecting hidden statistical features in text. However, such embedding leads to a decline in text quality, especially in low-entropy scenarios where performance needs improvement. Existing methods that rely on entropy thresholds often require significant computational resources for tuning and demonstrate poor adaptability to unknown or cross-task generation scenarios. We propose Context-Aware Threshold watermarking (CATMark), a novel framework that dynamically adjusts watermarking intensity based on real-time semantic context. CATMark partitions text generation into semantic states using logits clustering, establishing context-aware entropy thresholds that preserve fidelity in structured content while embedding robust watermarks. Crucially, it requires no pre-defined thresholds or task-specific tuning. Experiments show CATMark improves text quality in cross-tasks without sacrificing detection accuracy.

Multimodal Large Reasoning Models introduce the reasoning paradigm, demonstrating strong capabilities on complex vision-language tasks. However, they still suffer from severe hallucinations. Existing training-based methods typically mitigate hallucinations through response-level direct preference optimization (DPO), where the Chain-of-Thought (CoT) and the final answer are treated as a monolithic output and optimized jointly. We reveal that this formulation performs similarly to answer-only optimization, suggesting that it primarily learns answer-level preference, while leaving CoT-level supervision insufficiently exploited. To address this issue, we explicitly formulate a CoT-oriented preference term and derive Reasoning-Conditioned Direct Preference Optimization (RC-DPO), which models the CoT as a condition for answer generation and contrasts the preference for the same preferred answer under different CoT conditions, promoting answer-supportive reasoning chain alignment. To further improve optimization, we introduce a reasoning-enhanced preference data generation strategy that employs Monte Carlo Tree Search to construct high-quality preferred responses, pairs them with ordinarily sampled outputs, and further derives hard-negative reasoning conditions by pruning visually salient tokens. Extensive experiments across various models and benchmarks show that RC-DPO effectively mitigates hallucinations and improves the reliability of multimodal reasoning.

Real-world knowledge resides in multimodal documents, necessitating retrieval-augmented generation (RAG) for accurate question answering. However, existing multimodal RAG models are primarily designed for single-image or closed-document settings and exhibit limited accuracy in realistic multi-image scenarios. Moreover, processing numerous retrieved images incurs substantial computational overhead from irrelevant visual tokens. To address these challenges, we introduce DocLongRAG, a large-scale dataset of 343K question–answer pairs, each associated with an average of 37.4 retrieved images to reflect authentic RAG workflows. Building on this dataset, we propose Doc-REFRAG, a question-guided framework that compresses visual tokens into coarse chunks and selectively expands question-relevant ones via a lightweight RL-based selector. Experiments on six benchmarks show that Doc-REFRAG outperforms eleven strong baselines, achieving state-of-the-art accuracy with significantly lower inference latency. Our resources are available at https://github.com/Collab-Gen/Doc-REFRAG.

LLM-based text-to-speech (TTS) models can generate emotionally expressive speech, but how reference emotion is routed through the model and realized in decoded speech remains unclear. We introduce two emotion-sensitive metrics for matched neutral and emotional syntheses—a codec trajectory score and a late residual direction score—and use them to score activation-patching interventions. Under controlled matched-reference conditions, this analysis identifies a sparse source-to-readout component-level circuit: 23–27 attention heads and MLPs per emotion, roughly 5% of the components considered, recover or suppress 74–88% of the late emotion-readout shift on held-out cases. The circuit combines a shared component backbone with emotion-specific components; cross-emotion activation swaps reduce the target readout in 47 of 48 cases. In decoded speech, the same intervention produces consistent changes in pitch, energy, and spectral brightness over 24 matched pairs per emotion. A readout-matched residual-direction baseline produces only 17–27% of the intervention’s pitch effect, showing that internal readout movement alone does not explain the decoded acoustic changes. These results trace a compact causal route from reference-derived prefix information to emotion-relevant properties of generated speech.

We introduce **X**iaohongshu **S**ocial **Co**mparison **R**eader **E**licitation (XHS-SCoRE), a reader-grounded benchmark for detecting whether text-only Xiaohongshu (RedNote) posts elicit ‘Upward‘, ‘Downward‘, or ‘Neutral‘/no clear social comparison from a first-person reader perspective. The task targets a socially meaningful relational, behaviorally real signal not reducible to sentiment. Across prompted LLM classifiers and supervised Chinese encoders, we find a consistent generation–detection mismatch: the signal is textually learnable in-domain, but not robustly accessible to prompt-based classification. Prompted LLM classifiers show stable failures, especially neutralization of comparison-eliciting posts and model-specific directional skew. A controlled pilot shows that LLM-generated Xiaohongshu-style posts can shift perceived standing and comparison-related affect even when prompt-based detection of the same construct remains fragile. XHS-SCoRE contributes a benchmark for reader-grounded comparison detection and a diagnostic framework for studying when socially meaningful relational cues remain only partially visible to prompt-based inference.