Non-compositional expressions (e.g., idioms, proverbs, and metaphors) pose significant challenges for neural machine translation systems because their meanings cannot be derived from individual words alone. These expressions encode rich, cultural meaning, and have both figurative and literal meanings, making accurate translation difficult. Because models are fairly good at translating compositional text, we investigate GRPO-style fine-tuning using Machine Translation Quality Estimation (MTQE) models as reward functions. Using Chinese and Hindi idiom datasets, we find that idiom translation abilities improve by ~13 points, general, non-idiomatic translation implicitly improves by ~8 points, and cross-lingual translation abilities (trained on one language, evaluated on another) improves by ~6 points. Overall, our work quantifies the non-compositional translation gap and offers insights for developing LLMs with stronger cross-cultural and figurative language understanding.

Membership inference attacks (MIAs) assess whether a sample was used to train a model, posing privacy risks for large language models (LLMs). Existing MIAs typically rely on single loss or confidence signals, which are often unstable and easily confounded by noise, leading to limited robustness and generalization. In this paper, we propose Group-Invariant Membership Inference Attack (GI-MIA) that explicitly exploits semantic-preserving transformations to capture stability differences between members and non-members. Our method constructs an orbit of paraphrased variants for each input using LLM-driven transformations, and extracts invariant statistics (mean, maximum, and variance of normalized losses) across the orbit. A simple decision rule based on these statistics effectively discriminates members from non-members by leveraging the model’s inherent invariance on training data. We conducted extensive experiments on the Pythia family of LLMs (different model parameter sizes) across six datasets. Results show that GI-MIA consistently outperforms strong baselines, achieving state-of-the-art performance while maintaining robustness against semantic-preserving transformations. Our study highlights new directions for MIAs and calls for further research into LLM privacy protection, emphasizing the need for developing reliable models with better privacy resilience.

Preference optimization is widely used to align Large Language Models (LLMs) with preference feedback. However, most existing methods train on a single positive-negative pair per prompt, discarding additional supervision available in preference datasets that typically contain multiple candidate responses. Motivated by this limitation, recent work explores group-wise preference optimization, which jointly contrasts multiple responses for the same prompt, but its empirical behavior and scalability remain underexplored due to the memory overhead of group-coupled objectives. In this work, we present a unified empirical and systems study of group-wise preference optimization and develop a memory-efficient implementation for group-coupled objectives. By instantiating first-order linearization with objective-specific per-response coefficients, our implementation preserves first-order gradients while decoupling samples during backpropagation, substantially reducing peak memory usage and enabling scalable training with larger groups. Across offline and online settings, we show that leveraging multiple responses consistently outperforms single-pair training. Furthermore, incorporating a negative log-likelihood (NLL) term on positive responses is critical for both performance gains and training stability.

The development of high-quality multilingual models depends on well-curated multilingual datasets, which are typically obtained through filtering large web-crawled corpora. Such filtering approaches must be effective across a wide range of languages in order to improve model quality proportionally across the entire linguistic space. In recent years multiple Natural Language Processing (NLP) approaches for data processing and Large Language Model (LLM) training were developed. However, the most significant progress has been made for rich-resource languages such as English and Chinese, leaving multilingual methods largely underexplored. In this work, we present data selection strategies based on Internet registers (genres) and Web Document Score (WDS). We study 15 European languages utilizing HPLT 3.0 datasets. We train and evaluate four small generative language models per language, and compare how specific registers and WDS affect model training in separate and combined approaches. We use five common evaluation tasks covering models’ linguistic, knowledge-based, and reasoning skills. We analyze the impact of the studied strategies on each language, their average performance, and show how combining register-based selection with WDS improves evaluation performance. These results indicate that both registers and WDS can generalize across multiple languages, offering a new path towards multilingual data filtering.

When LLMs judge moral dilemmas, do they reach different conclusions in different languages, and if so, why? Two factors could drive such differences: the language of the dilemma itself, or the language in which the model is instructed to express its reasoning. Standard evaluation conflates these by testing only matched conditions (e.g., English dilemma with English reasoning). We introduce a methodology that separately manipulates each factor, covering also mismatched conditions (e.g., English dilemma with Chinese reasoning), enabling decomposition of their contributions. To study what changes, we propose an approach to interpret the moral judgments in terms of Moral Foundations Theory. As a side result, we identify evidence for splitting the Authority dimension into a family-related and an institutional dimension. Applying this methodology to English-Chinese moral judgment with 13 LLMs, we demonstrate its diagnostic power: (1) in this EN–CN case study, the framework isolates the language of generated rationale as contributing roughly twice the variance of input-language effects; (2) it detects context-dependency in 4 of 9 EN–CN-tested models that standard evaluation misses; and (3) a diagnostic taxonomy translates these patterns into deployment guidance. We release our code and datasets at https://github.com/aida-ugent/LLMCrossCulturalMoralJudgement.

Effective prompt engineering remains a central challenge in fully harnessing the capabilities of LLMs. While well-designed prompts can dramatically enhance performance, crafting them typically demands expert intuition and a nuanced understanding of the task. Moreover, the most impactful prompts often hinge on subtle semantic cues, ones that may elude human perception but are crucial for guiding LLM behavior. In this paper, we introduce PRL (Prompts from Reinforcement Learning), a novel RL-based approach for automatic prompt generation. Unlike most prior methods, PRL can also generate novel few-shot demonstrations rather than selecting only from examples seen during training. Empirically, our approach achieves state-of-the-art performance across a diverse set of benchmarks, including text classification, text simplification, and mathematical reasoning, while remaining competitive on text summarization. Our code is available at https://github.com/Batorskq/prl.

Large reasoning models have demonstrated remarkable proficiency in various tasks. However, we observe that they frequently exhibit a problematic reliance on familiar reasoning patterns, a phenomenon we term reasoning rigidity. Despite explicit instructions from users, these models often override clearly stated conditions and default to habitual reasoning, leading to incorrect conclusions. This behavior presents significant challenges, not only in reasoning-intensive domains but also in realistic settings. To systematically investigate reasoning rigidity, a behavior unexplored in prior work, we introduce a dataset, ReasoningTrap, consisting of math problems, puzzles, and agentic tasks designed to diagnose reasoning rigidity. Using this dataset, we identify patterns that occur when models default to ingrained reasoning, and suggest inference-level and GRPO-based post-training remedies. Our diagnostic set and evaluation code are available at the project page to facilitate future research on mitigating reasoning rigidity in language models.

Generative search systems are increasingly replacing link-based retrieval with AI-generated summaries, yet little is known about how these systems differ in sources, language, and fidelity to cited material. We examine responses to 11,000 real search queries across five systems—vanilla GPT, Search GPT, Perplexity Search with Grok, Google AI Overviews, and traditional Google Search—at three levels: source diversity, linguistic characterization of the generated summary, and source-summary fidelity. We find that generative search systems exhibit significant source-selection biases in their citations, favoring certain sources over others. Incorporating search also selectively attenuates epistemic markers, reducing hedging by up to 60% while preserving confidence language in the AI-generated summaries. At the same time, AI summaries further compound the citation biases: Wikipedia and longer sources are disproportionately overrepresented, whereas cited social media content and negatively framed sources are substantially underrepresented. Our findings highlight the potential for answer bubbles, in which identical queries yield structurally different information realities across systems, with implications for user trust, source visibility, and the transparency of AI-mediated information access.

Recent representation analyses have reported structure-like or invariant behavior even in randomly initialized transformers, motivating the interpretation that aspects of linguistic structure may be architectural rather than learned. We show that such conclusions can be artifacts of coarse symbolic discretization. We formalize symbolic analysis of neural representations as operating over discretized activation trajectories and identify symbolic resolution as a critical, previously uncontrolled variable. Varying resolution reveals a sharp collapse of apparent similarity in random models, while pretrained models maintain stable, operator-selective differentiation across all tested resolutions. We introduce resolution stability as a necessary validity criterion for symbolic and automata-based analyses, providing a falsifiable, architecture-agnostic diagnostic for auditing internal linguistic structure in neural language models.

Peer review serves as a backbone of academic research, but in most AI conferences, the review quality is degrading as the number of submissions explodes. To reliably detect low-quality reviews, we define misinformed review points as either "weaknesses" in a review that contain incorrect premises, or "questions" in a review that can be already answered by the paper. We verify that 15.2% of weaknesses and 26.4% of questions are misinformed and introduce ReviewScore indicating if a review point is misinformed. To evaluate the factuality of each premise of weaknesses, we propose an automated engine that reconstructs every explicit and implicit premise from a weakness. We build a human expert-annotated ReviewScore dataset to check the ability of LLMs to automate ReviewScore evaluation. Then, we measure human-model agreements on ReviewScore using eight current state-of-the-art LLMs. The models show F1 scores of 0.4–0.5 and kappa scores of 0.3–0.4, indicating moderate agreement but also suggesting that fully automating the evaluation remains challenging. A thorough disagreement analysis reveals that most errors are due to models’ incorrect reasoning. We also prove that evaluating premise-level factuality shows significantly higher agreements than evaluating weakness-level factuality.

Two-sided digital platforms are inherently dynamic: user preferences shift, item popularity evolves, and reviews both reflect and drive these changes. Yet most sequential recommendation systems treat reviews as passive signals for updating user states, leaving two aspects underexplored. First, review generation is nonrandom, depending on evolving latent states of both users and items. Second, reviews can reshape item states, induce spillover across related items, and influence future user decisions. To address these gaps, we propose a two-sided state-space model (TS-SSM) for event-conditioned sequential recommendation. TS-SSM consists of three components: (1) a modality-missing-not-at-random fusion module that encodes review content and informative observation patterns; (2) user-state evolution with temporal variation and local graph message passing that uses related item states to refine user preferences; and (3) item-state evolution with asymmetric carryover of positive and negative review feedback. In experiments across six Amazon categories, TS-SSM increases Recall@20 over BSARec by 14.8%-18.8% and exceeds HM4SR by 11.7% on average. On Goodreads Fantasy, Recall@20 improves HM4SR from .5191 to .5847. Ablations highlight distinct contributions of observation patterns, local propagation, and item dynamics.

Alternative text is imperative for making the internet accessible to blind and low-vision individuals. Vision-language models (VLMs) have the potential to improve alt text coverage. This study considers how social stereotypes of race and gender are manifested in the alt-text language generated by GPT-4.1, Claude-Sonnet-4.5, and Gemini-2.5-Pro. A total of 1200 images are generated uniformly by gender and skin tone, and then the VLMs are prompted to generate the alternative text. The analysis of the generated alt text shows that the text descriptions differ significantly depending on the skin tone and gender of the person in an image. The VLMs tend to assume that whiteness is the default identity and the descriptions of those with darker skin are more likely to contain racialized terms. This bias suggests that AI-generated alt-text should be used with caution in the interest of preserving equity and preventing discrimination.

The tremendous commercial potential of large language models (LLMs) has heightened concerns over their unauthorized use. To address this, we focus on the task of identifying the origin of black-box LLMs. We further propose PlugAE, an effective and efficient identification method that proactively leverages LLM-specific adversarial embeddings and allows users to customize copyright tokens on a targeted query set. Extensive experiments demonstrate that PlugAE outperforms both state-of-the-art model watermarking and fingerprinting methods in accuracy and robustness. We further analyze its stealthiness and reliability from three complementary perspectives and conduct ablation studies under various configurations, confirming its practicality for real-world misuse detection.

Computational stylometry studies writing style through quantitative textual patterns, enabling applications such as authorship attribution, identity linking, and plagiarism detection. Despite the relevance of language modeling to these tasks, the pre-training of modern large language models (LLMs) has been underutilized in authorship attribution and verification. We introduce an unsupervised framework that uses the log-probabilities of an LLM to measure style transferability between two texts. This framework takes advantage of the extensive Causal Language Modeling (CLM) pre-training, one-shot capabilities and scale of LLMs, avoiding explicit supervision. Our methods substantially outperform prompting-based unsupervised baselines in authorship verification at similar model sizes, and is competitive with or improves contrastive baselines in most settings with sufficient model scale. We further observe strong performance across non-English languages. The effectiveness of the proposed framework improves consistently with increasing model scale. In the case of authorship verification, we propose an additional mechanism that increases test-time computation to improve accuracy; enabling flexible trade-offs between computational cost and task performance.

As Large Language Model (LLM) agents are increasingly deployed in open-ended domains like software engineering, they frequently encounter underspecified instructions that lack crucial context. While human developers naturally resolve underspecification by asking clarifying questions, current agents are largely optimized for autonomous execution. In this work, we systematically evaluate the clarification-seeking abilities of LLM agents on an underspecified variant of SWE-bench Verified. We propose an uncertainty-aware multi-agent scaffold that decouples underspecification detection from code execution. Across both proprietary and open-weight frontier LLMs, our scaffold achieves a 69.40% task resolve rate, significantly outperforming a standard single-agent setup and closing the performance gap with agents operating on fully specified instructions. Furthermore, we find that the multi-agent system exhibits well-calibrated information-seeking behavior, conserving queries on simple tasks while proactively seeking information on more complex issues. These findings indicate that current models can be turned into proactive collaborators, where agents independently recognize when to ask questions to elicit missing information in real-world, underspecified tasks.

Recent LLM-based embodied agents have advanced in high-level task planning. Yet, they frequently suffer from hallucinations and brittle reasoning due to passive observation—either omitting fine-grained attributes or overwhelming the planner with irrelevant information. We propose SceneQueryPlanner, a memory-environment dual-grounded agent that integrates active information seeking into the planning loop by decomposing instructions and selectively executing environmental interactions or structured queries over a dynamic 3D scene graph to retrieve task-relevant information on demand. By treating information acquisition as a first-class planning action, our approach enables explicit verification and supports reliable long-horizon reasoning over object attributes. For rigorous evaluation, we introduce AttPlan-Bench, a benchmark of multi-room, long-horizon household tasks requiring attribute-aware reasoning. Experiments demonstrate that our dual-grounding strategy outperforms passive baselines, yielding more stable and scalable reasoning across diverse tasks.

Recent multimodal sentiment analysis studies increasingly adopt text-centric fusion approaches to exploit the rich sentiment information inherent in the textual modality. However, these approaches often suffer from performance degradation during inference due to partially missing or noisy data in real-world scenarios, especially when sentiment-related cues are missing. To address this issue, we introduce a new completeness estimation approach that quantifies the degree of sentiment-relevant information preserved in incomplete data to guide the reconstruction of missing semantics. Furthermore, we propose a training strategy that stabilizes multi-task learning while jointly optimizing sentiment prediction and completeness estimation. Extensive experiments and in-depth analyses on three benchmark datasets demonstrate that the proposed approach enables more accurate semantic reconstruction, leading to more precise sentiment prediction.

Despite the impressive progress of recent MLLMs on spatio-temporal video grounding (STVG), existing evaluations and training data focus primarily on simple queries. They largely overlook the compositional queries prevalent in real-world scenarios, where a target must be disambiguated by jointly reasoning about its attributes and relations to other entities. To bridge this gap, we propose Compositional Spatio-Temporal Video Grounding (CompSTVG), a task that requires models to process complex textual queries where every intertwined attribute and relational cue is essential for disambiguation. To facilitate this task at scale, we build a synthetic data engine that leverages a spatio-temporal scene graph as a difficulty measure and casts difficulty-controlled query synthesis as a constraint programming problem, producing difficulty-graded data for both evaluation and training. Built on this engine, we introduce STVG-CompBench, a benchmark stratified by explicit difficulty levels that jointly capture temporal complexity and spatial interference. Evaluating 11 representative STVG models on STVG-CompBench reveals that current models perform poorly on compositional queries, exhibiting a sharp performance drop that is typically obscured by overall dataset-level averages. We further construct synthetic training data and propose CurrSTVG, a curriculum reinforcement learning framework that delivers consistent gains, with the largest improvements observed on the most challenging compositional queries.

Large language model (LLM) agents excel at complex, multi-step tasks but suffer from catastrophic forgetting during continual learning—acquiring new skills often degrades previously mastered capabilities. Existing post-training paradigms face a fundamental trade-off: Supervised Fine-Tuning (SFT) is sample-efficient but prone to severe forgetting, while Reinforcement Learning (RL) preserves prior capabilities but converges slowly without expert guidance. To bridge this gap, we introduce **T**rigger-based **S**upervised **R**einforcement **L**earning (TSRL), a framework that integrates the strengths of both paradigms through an adaptive, state-dependent expert supervision mechanism. TSRL employs RL as the optimization backbone, activating SFT selectively on "hard states" where the policy fails. Theoretically, we show that restricting supervision to failure states derives an upper bound on the cumulative parameter updates, thereby reducing forgetting compared to full SFT. Empirically, TSRL achieves consistent gains across three agent benchmarks (ALFWorld, WebShop, and Deep Search) at both 7B and 32B scales. In continual learning scenarios, our method retains up to 9.6 points more capability on prior tasks than SFT→RL baselines, while matching or exceeding target-task performance.

Conversational explainable AI (ConvXAI) agents often use GUS-style architectures with separate modules for intent recognition, dialogue management, and template-based response generation. While faithful to the underlying XAI methods, these pipelines are hard to adapt and often struggle with context-dependent follow-up questions. We present SimXAI, a lightweight LLM-centric conversational layer that replaces these modules with a single tool-calling LLM that bridges access to existing XAI resources. Rather than a standalone agent, SimXAI serves as a modular interface that enables conversational access to external XAI resources. We further introduce an evaluation suite for single- and multi-turn parsing, faithfulness, and contextualization. Across existing and new datasets, SimXAI achieves strong function-selection accuracy and high faithfulness while better handling context-dependent requests. In a user study replacing only the conversational layer of TalkToModel, a representative GUS-based agent, users completed tasks more accurately and in fewer turn. Overall, SimXAI improves GUS-based design while preserving faithfulness and improving conversational quality.

This paper presents a simple method that allows to easily enhance textual pre-trained large language models with speech information, when fine-tuned for a specific classification task. A classical issue with the fusion of many embeddings from audio with text is the large length of the audio sequence compared to the text one. Our method benefits from an existing speech tokenizer trained for Audio Speech Recognition that output long sequences of tokens from a large vocabulary, making it difficult to integrate it at low cost in a large language model. By applying a simple lasso-based feature selection on multimodal Bag-of-Words representation, we retain only the most important audio tokens for the task, and adapt the language model to them with a self-supervised language modeling objective, before fine-tuning it on the downstream task. We show this method improves the performances compared to an unimodal model, larger SpeechLMs, and audio integration via learned representations. We demonstrate its effectiveness on Argumentative Fallacy Detection and Classification tasks where audio was previously believed counterproductive, and affective computing tasks on a widely-used dataset. We also provide an in-depth analysis of the method, showing that even a random audio token selection helps enhancing the unimodal model. Our code is available online.

Politeness is key to building healthy communication. Brown and Levinson’s influential theory of politeness (BL’s theory) proposes that polite speech can be characterized by the presence of face-saving speech acts while impolite speech is characterized by the presence of face-threatening speech acts. Most existing politeness resources neither operationalize nor control politeness under BL’s theory; instead, they typically consist of short utterances with an intuitive account of politeness. In this paper, we introduce a long-text corpus that includes sentence-level face-act annotations and document-level politeness scores. Following BL’s theory, we annotate politeness along three dimensions: Directness vs. Indirectness, Positive Face Saving, and Negative Face Saving, and define Overall Politeness as their average. We formulate face-act tagging as a multi-label classification task and compare context-agnostic sentence classifiers to context-sensitive variants and sequence-labeling models. We also model document-level politeness via multi-target regression using the text, and enhance the regression model with the predictions of our face-act classifier. We show that context-sensitive classifiers outperform context-free classifiers for face-act tagging. Moreover, we demonstrate that for document-level politeness regression, enhancing the text with face-act predictions yields better performance. Our code and data are available online: alipanahroshad-oss/email-politeness-corpus

Large language models (LLMs) remain vulnerable to jailbreak attacks that elicit structured, credible misinformation. Existing benchmarks primarily track binary refusals, often overlooking the severity and exposure risk of the misinformation payload itself. In this work, we introduce Misinfo-Jailbreak Bench, a comprehensive framework that systematically maps diverse attack strategies to a claim-level scoring pipeline for precise risk quantification. Building on this diagnostic foundation, we propose Selective LoRA Ethical Surgery, a parameter-efficient defense mechanism. Instead of broad parameter updates, our approach identifies safety-relevant subspaces via contrastive representations to confine LoRA edits and employs inference-time gating to preserve benign capabilities. Experimental results on state-of-the-art models (e.g., Qwen2.5, Llama-4) demonstrate that our method reduces high-severity risk by 60-70% while limiting general capability degradation to less than 1%, offering a precise and minimally invasive paradigm for post-hoc alignment.

Structured data exists in many forms (tables, knowledge graphs, charts, and time series), and converting it into text may involve different generation tasks. However, most prior work on data-to-text (D2T) generation has focused on specific tasks and datasets, relying either on task-specific training data or on the zero-shot capabilities of large language models. We study cross-domain D2T generation in a setting where neither in-domain training text nor test references are available, and where domains, generation goals, and input structures vary substantially. We compare data-driven knowledge distillation (DDKD) against zero-shot inference and fine-tuning on out-of-domain D2T data, and introduce structure-preserving augmentation via structural subsampling and perturbation. Experiments on five benchmarks show that, at constant model size (1.7B parameters), DDKD consistently outperforms both fine-tuning and zero-shot inference. Moreover, the resulting small models outperform a much larger finetuned model on two of the five domains, achieving comparable performance on the remaining three. We further construct QUINTD-5, a fivefold extension of QUINTD-1, and show that simply scaling real target-domain inputs yields only modest gains, whereas our augmentation strategy remains more effective and more cost-efficient for cross-domain distillation.

Large language models (LLM) deployed as autonomous pricing agents may sustain supracompetitive prices through tacit coordination. We develop a causal graph divergence framework that separately measures structural faithfulness and intent faithfulness of LLM pricing agents in Bertrand competition. Across nine LLMs under duopoly and triopoly conditions, collusive behavior and chain-of-thought (CoT) faithfulness dissociate along both dimensions: the most collusive model accurately reports cooperative intent yet reasons structurally unfaithfully, while the most structurally faithful model sustains supra-Nash pricing under both market structures. These findings establish that CoT monitoring alone cannot serve as a standalone safeguard against algorithmic collusion.

With the recent commercial availability of Large Language Models (LLMs), end users increasingly turn to them for encyclopedic knowledge. Doing so, users may inadvertently inject incorrect assumptions (as presuppositions) into LLM prompts, which are then affirmed by LLMs. In this work, we investigate the effect that existential presuppositions have on LLM responses. We introduce the Dataset for Evaluating and Explaining the Effect of Existential Presuppositions (D4EP) and use it to quantify their effect. We further develop a probing-based explanation that analyzes the LLM as a classifier and finds that the effect of presuppositions corresponds to a change of the classifier’s operating point.

Large language models show promise for mental health support, yet therapeutic quality improves only when evaluation functions as an actionable control signal rather than a passive metric. We introduce a framework that formulates therapeutic response generation as a decision-refinement problem driven by multi-dimensional, human-aligned evaluation. In Stage I, we introduce TheraJudge, an open-source therapeutic evaluator trained via preference-based optimization on human-annotated data to produce reliable judgments across 7 psychological dimensions. In Stage II, we introduce TheraAgent, which operationalizes TheraJudge’s evaluations through a coordinated refinement process with specialized Critic, Coach, and Therapist roles that translate evaluative signals into targeted response revisions. Empirically, TheraJudge achieves strong agreement with clinician ratings, with intraclass correlation coefficients (ICC = 0.87-0.95), surpassing supervised baselines and strong closed-source judges, particularly on critical dimensions such as Safety, Relevance, and Empathy. Acting on these evaluations, TheraAgent yields a +0.43 improvement in human-rated therapeutic quality (on a 5-point scale) under blind evaluation, with 96% clinician inter-rater reliability. Low-quality responses improve by +2.45 points with a 94% recovery rate, demonstrating targeted correction of unsafe outputs. Overall, our results indicate that effective alignment of mental-health LLMs stems from acting on human-aligned evaluation, rather than relying solely on stronger generation. We release code at <redacted>.

Quantization reduces the deployment cost of large language models (LLMs). However, post-training quantization (PTQ) cannot fine-tune model parameters and often degrades accuracy in low-bit settings, while quantization-aware training (QAT) requires backpropagation and incurs high memory overhead. We propose ZeroQAT, a zeroth-order QAT framework supporting both weight and activation quantization. By using forward-only gradient estimation, ZeroQAT removes backpropagation and substantially reduces memory and computation while retaining end-to-end optimization benefits. We further introduce a lightweight variant of ZeroQAT for quantized fine-tuning, which freezes and pre-quantizes most parameters to further cut memory usage. Experiments show that ZeroQAT consistently outperforms representative PTQ and QAT baselines while requiring significantly less memory. For example, ZeroQAT enables fine-tuning of a 13B model at extremely low bit-widths (e.g., 2-4 bits) on a single 8GB GPU, and even allows fine-tuning a 6.7B model on a OnePlus 12 smartphone, demonstrating its practicality for end-to-end QAT on resource-limited edge devices.

When large language models process structured temporal data, serialization into token sequences collapses the two-dimensional organization (Features * Time) that encodes meaningful relationships between co-occurring features. We investigate whether visual encoding can recover the structural information that linearization loses. We introduce LE-Viz, a framework that transforms longitudinal experiential data into complementary visual and textual representations for a vision-language model that generates prospective narratives, framing prediction as language generation. Evaluated on cross-distribution tasks across three datasets spanning mental health and educational forecasting, LE-Viz sets a new benchmark on a widely-adopted cross-distribution generalization task, improving over both the traditional state-of-the-art and the best text-only LLM baseline, and consistently outperforms all baselines. Analyses show that spatial organization of visual inputs, not information quantity, drives the generalization gain, and that the multimodal combination exceeds both text-only and vision-only conditions.

Resolving egocentric spatial expressions in conversational navigation requires mapping speaker-centric references to allocentric orientations. In practice, this process is often mediated by automatic speech recognition (ASR), where transcription errors, linguistic variation, and referential ambiguity obscure spatial relations. Despite progress in navigation and multimodal reasoning, egocentric spatial reference resolution under ASR-transcribed, non-English conversational input remains underexplored. We introduce Conversational Orientation Reasoning (COR), a diagnostic benchmark for egocentric spatial reference resolution in Traditional Chinese conversational navigation. COR consists of 10,000 instances pairing ASR-transcribed utterances with structured landmark coordinates derived from real-world environments projected onto a 50 x 50 grid. COR isolates the effects of spoken language noise from geometric computation, enabling evaluation of transcription errors, linguistic variation, and referential ambiguity beyond end-to-end navigation success. Using COR, we study an interpretable decomposition of orientation inference into three steps: extracting spatial relations, mapping landmark coordinates to absolute directions, and inferring the speaker’s orientation. Models following this decomposition achieve 99.6% accuracy on clean text and retain 95.7% accuracy under ASR-transcribed input. COR supports analysis of failure modes in spoken spatial reference resolution, providing insights into how language grounding degrades under noisy conversational input. The code and dataset are publicly available under the MIT license.

Text-attributed graphs (TAGs) are widely used in many real-world domains, and learning on TAGs requires jointly modeling text semantics and graph structure. A standard approach for modeling TAGs is to combine a language model (LM) and a graph neural network (GNN), but joint training is computationally expensive and difficult to scale. Dataset distillation is a promising way to reduce training costs, but existing methods are not well suited to TAGs because they are typically designed for a single modality or still require repeatedly training expensive LM-GNN models on the full dataset during distillation. To address this, we propose TaLK, an effective dataset distillation method for TAGs that couples an LM with a graph-aware neural tangent kernel. This design enables efficient dataset distillation, avoiding repeated joint training on the full dataset while reflecting both textual and structural information for effective TAG learning. Experiments on multiple TAG benchmarks show that TaLK consistently outperforms existing baselines and achieves up to 97% of full-dataset performance with only 1% synthetic data.

Large language models (LLMs) are increasingly used for automatic evaluation of generated text, but most work focuses on English. Extending LLM-based evaluation to multilingual settings is challenging, especially for low-resource languages or when in-domain data is scarce. This work explores how an English evaluation corpus can serve to evaluate other languages, both when in-domain data is available for fine-tuning and when it is not. We systematically analyze English, Spanish, and Basque, representing high-, mid-, and low-resource languages, considering instruction translation, monolingual versus multilingual supervision, and model size. Our results reveal key trade-offs: When in-domain data is available, fine-tuned smaller models can achieve performance comparable to proprietary models, whereas zero-shot evaluation with larger models proves more effective in out-of-domain settings. We also observe that fine-tuning on out-of-domain data can adversely affect model performance. These findings provide practical guidance for building efficient, reliable multilingual evaluation pipelines. The data and code are publicly available at hitz-zentroa/mJudge .

Large Language Models (LLMs) have displayed remarkable capabilities across diverse domains, but their training remains resource – and time-intensive, requiring massive computational resources and careful orchestration of training procedures. Model souping – the practice of averaging weights from multiple models of the same architecture – has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. We observe that previous souping approaches can lead to collapse in precision-sensitive LLM capabilities. In this paper, we introduce Soup of Category Experts (SoCE), a principled approach for model souping to overcome this shortcoming. The proposed method utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous approaches, our method leverages the observation that different clusters (or categories) of points within a benchmark often exhibit low inter-correlations in model performance. SoCE identifies "expert" models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that SoCE improves performance and robustness across multiple domains achieves state-of-the-art results on the Berkeley Function Calling Leaderboard.

LLMs are increasingly able to answer complex questions about enterprise-scale document collections. But evaluation is hard: companies don’t want to share internal communications, and synthetic datasets have been overly simple. We present CorporateBench (CB), a human-validated multi-task Q A benchmark whose scale approaches the conditions LLMs encounter in corporate communication networks, with evaluation corpora surpassing 230,000 documents. CB evaluates LLMs across two dimensions (information extraction and knowledge base querying) through four synthetically generated firms ranging from 12 to 10,000 employees. Each corpus is sampled from a temporally evolving knowledge base describing a consistent world, guaranteeing cross-document logical consistency even across hundreds of thousands of documents. We evaluate five LLMs on CB, revealing increasingly poor performance as input size approaches realistic scales. CB provides LLM developers a metric for corporate communication reasoning, filling a crucial gap in the benchmarking ecosystem.

Affective cues across modalities may be incongruous (e.g., sarcasm or mocking praise), potentially leading to misinterpretation when relying on a single modality. Large Audio–Language Models (LALMs) have recently gained popularity and been applied to multimodal emotion recognition, but their ability to disentangle acoustic and semantic cues, especially in incongruent cases, remains underexplored. To address this gap, we introduce CREMA-ASIS, a dataset specifically created to investigate incongruence between acoustic emotion and semantic sentiment cues. It pairs acoustic emotion labels with semantic sentiment polarities. Using this dataset, we evaluate LALM biases within a multitask framework and conduct a layer-wise analysis to identify modality dominance across layers. Our findings reveal that LALMs struggle with semantic-acoustic incongruent cases, rarely predicting incongruity, and that LALMs are predominantly influenced by semantic information. However, supervised fine-tuning significantly improves LALM performance on our CREMA-ASIS test set while preserving transcription accuracy and joint emotion recognition. Results demonstrate potential for enhancing both acoustic and semantic understanding on out-of-domain data.

Recent Large Reasoning Models (LRMs) achieve strong performance by leveraging long-form Chain-of-Thought (CoT) reasoning, but uniformly applying overlong reasoning at inference time incurs substantial and often unnecessary computational cost. To address this, prior work explores various strategies to infer an appropriate reasoning budget from the input. However, such approaches are unreliable in the worst case, as estimating the minimal required reasoning effort is fundamentally difficult, and they implicitly fix the trade-off between reasoning cost and accuracy during training, limiting flexibility under varying deployment scenarios. Motivated by these limitations, we propose ORBIT, a controllable multi-budget reasoning framework with well-separated reasoning modes triggered by input. ORBIT employs multi-stage reinforcement learning to discover Pareto-optimal reasoning behaviors at each effort, followed by on-policy distillation to fuse these behaviors into a single unified model. Experiments show that ORBIT achieves (1) controllable reasoning behavior over multiple modes, (2) competitive reasoning density within each mode, and (3) integration of these frontier policies into a single unified student model while preserving clear mode separation and high per-mode performance.

AI-generated content has progressed from monolithic models to modular workflows, especially on platforms like ComfyUI, allowing users to customize complex creative pipelines. However, the large number of components in ComfyUI and the difficulty of maintaining long-horizon structural consistency under strict graph constraints frequently lead to low pass rates and workflows of limited quality. To tackle these limitations, we present ComfySearch, an agentic framework that can effectively explore the component space and generate functional ComfyUI pipelines via validation-guided workflow construction. Experiments demonstrate that ComfySearch substantially outperforms existing methods on complex and creative tasks, achieving higher executability (pass) rates, higher solution rates, and stronger generalization.

Counterfactual examples are minimal edits to an input that alter a model’s prediction. They are widely employed in explainable AI to probe model behavior and in natural language processing (NLP) to augment training data. However, generating valid counterfactuals with large language models (LLMs) remains challenging, as existing single-pass methods often fail to induce reliable label changes, neglecting LLMs’ self-correction capabilities. To explore this untapped potential, we propose iFlip, an iterative refinement approach that leverages three types of feedback, including model confidence, feature attribution, and natural language. Our results show that iFlip achieves an average 57.8% higher validity than the three state-of-the-art baselines, as measured by the label flipping rate. The user study further corroborates that iFlip outperforms baselines in completeness, overall satisfaction, and feasibility. In addition, ablation studies demonstrate that three components are paramount for iFlip to generate valid counterfactuals: leveraging an appropriate number of iterations, pointing to highly attributed words, and early stopping. Finally, counterfactuals generated by iFlip enable effective counterfactual data augmentation, substantially improving model performance and robustness.

Quantization is widely used to accelerate inference and streamline the deployment of large language models (LLMs), yet its effects on self-explanations (SEs) remain unexplored. SEs, generated by LLMs to justify their own outputs, require reasoning about the model’s own decision-making process, a capability that may exhibit particular sensitivity to quantization. As SEs are increasingly relied upon for transparency in high-stakes applications, understanding whether and to what extent quantization degrades SE quality and faithfulness is critical. To address this gap, we examine two types of SEs: natural language explanations (NLEs) and counterfactual examples, generated by LLMs quantized using three common techniques at distinct bit widths. Our findings indicate that quantization typically leads to moderate declines in both SE quality (up to 4.4%) and faithfulness (up to 2.38%). The user study further demonstrates that quantization diminishes both the coherence and trustworthiness of SEs (up to 8.5%). Compared to smaller models, larger models show limited resilience to quantization in terms of SE quality but better maintain faithfulness. Moreover, no quantization technique consistently excels across task accuracy, SE quality, and faithfulness. Given that quantization’s impact varies by context, we recommend validating SE quality for specific use cases, especially for NLEs, which show greater sensitivity. Nonetheless, the relatively minor deterioration in SE quality and faithfulness does not undermine quantization’s effectiveness as a model compression technique.

Autoregressive decoding remains a key bottleneck for LLM inference, due to its token-by-token nature. Speculative decoding is proposed to reduce latency by drafting multiple tokens and verifying them with fewer target-model passes. However, existing methods either incur auxiliary-drafter overhead or rely on brittle lexical span reuse. In structured tasks (e.g., code generation and text editing), we observe significant locality of reference and semantic redundancy for potential semantic context reuse. In light of this, we present Semantic Retrieval Speculative Decoding (SRSD), a training-free scheme that treats intermediate-layer hidden states as an in-context semantic index. At each decoding step, SRSD retrieves semantically similar past positions. Multiple candidates are merged into a compact tree and verified jointly in a single forward pass. SRSD requires no auxiliary models or external datastores. Across a variety of popular LLMs, SRSD achieves up to 5.42× throughput speedups over state-of-the-art training-free methods and parametric drafters. The code is available at https://github.com/Secbrain/SRSD.

Humans are influenced by how information is presented, a phenomenon known as the framing effect. Prior work suggests that LLMs may also be susceptible to framing, but it has relied on synthetic data and did not compare to human behavior. To address this gap, we introduce WildFrame - a dataset for evaluating LLM responses to positive and negative framing in naturally-occurring sentences, alongside human responses on the same data. WildFrame consists of 1,000 real-world texts selected to convey a clear sentiment; we then reframe each text in either a positive or negative light and collect human sentiment annotations. Evaluating eleven LLMs on WildFrame, we find that all models respond to reframing in a human-like manner (r≥ 0.52), and that both humans and models are influenced more by positive than negative reframing. Notably, GPT models are the least correlated with human behavior among all tested models. These findings raise a discussion around the goals of state-of-the-art LLM development and whether models should align closely with human behavior, to preserve cognitive phenomena such as the framing effect, or instead mitigate such biases in favor of fairness and consistency.

Existing real-world datasets for multimodal fact-checking have multiple limitations: they contain few instances, cover on only one or two languages, focus only on one task, or rely on external news article sets for sourcing true claims. To address these shortcomings, we introduce M4FC, a new real-world dataset comprising 4,982 images paired with 6,980 claims. The images, verified by professional fact-checkers from 22 organizations, represent a diverse range of cultural and geographic contexts. Each claim is available in one or two out of ten languages. M4FC spans six multimodal fact-checking tasks: visual claim extraction, claimant intent prediction, fake image detection, image contextualization, location verification, and verdict prediction. We provide baseline results for all tasks and analyze how combining intermediate tasks affects verdict prediction performance. We make our dataset and code publicly available.

Large Language Models (LLMs) acquire extensive factual knowledge during computationally expensive pretraining. Knowledge editing aims to efficiently update specific facts in a trained model without retraining the entire network. While many editing methods have recently been proposed and show strong performance on existing benchmarks, these benchmarks exhibit several limitations. In particular, they allow only a single edit per entity, rely on highly templated prompts, and evaluate edits using relative probability changes rather than ground-truth correctness. As a result, the capabilities and limitations of current knowledge editing methods remain insufficiently understood. In this work, we introduce the Subject-centric Entity Editing Dataset (SEED), a dataset and evaluation framework designed to address these limitations. SEED organizes edits into sets of related and unrelated facts rather than isolated instances, enabling evaluation when multiple facts about the same subject entity are updated. It also introduces undirected knowledge probes that test whether edited facts can be recalled in general contexts, and ground-truth evaluation metrics that directly measure whether the edited fact appears in the model’s output. Using SEED, we test 9 state-of-the-art knowledge editing methods over 10 metrics. These studies find that editing performance decreases by up to 36.9% as the number of edits per entity increases and edit recall decreases by 36.6% with prompts that increase token-wise distance between the edit subject and relation.

Adapting large language models to individual users remains challenging because systems must capture fine-grained personal preferences while serving many users efficiently. We introduce CARD, a hierarchical personalization framework that first learns preferences shared across user groups and then captures how individual users depart from those shared patterns. CARD trains a parameter-efficient adaptation module for each group, providing a reliable prior even when a user has limited history. It then learns a compact user preference vector by contrasting user-authored responses with outputs from the corresponding group model, without requiring manual annotations. During inference, both the backbone and the group adaptation module remain frozen, while the user vector applies a low-rank correction to the token logits during decoding. Experiments on personalization benchmarks show that CARD consistently outperforms competitive baselines while requiring only a compact vector for each user.

Speculative decoding accelerates large language model (LLM) inference by allowing a lightweight draft model to propose candidate tokens for verification by a target model. In real-world deployments, however, shifts in input distributions and updates to the target model can weaken draft–target alignment, reducing token acceptance and acceleration. Repeatedly retraining the draft model is costly, delays deployment, and may impair performance on previously encountered distributions. Achieving online alignment therefore requires rapid and stable adaptation while incorporating real-time feedback within a limited compute budget. To this end, we propose Dynamic Speculative Decoding (DySD), a framework that aligns the draft model with the target model during inference without offline retraining. DySD combines hierarchical feature alignment with distribution-level matching and uses divergence-triggered dynamic distillation to selectively update the draft model. Experiments across multiple datasets show that DySD achieves 1.6×–2.0× speedups and boosts token acceptance rates by 10%–30% under distribution shifts, enabling speculative decoding to remain fast, stable, and adaptive in ever-changing environments.

Realistic user simulation is crucial for training and evaluating multi-turn dialogue systems, yet creating simulators that accurately replicate human behavior remains a significant challenge. An effective simulator must expose the failure modes of the systems under evaluation. This work introduces Direct Iterative Adversarial Learning (DIAL), an adversarial framework that iteratively enhances user simulator realism through a competitive dynamic between a generator (user simulator) and a discriminator. When applied to mental health support, a domain characterized by diverse failure types and a critical dependence on realistic user behavior for failure detection, DIAL restores lexical diversity diminished by supervised fine-tuning and drastically reduces discriminator accuracy. The resulting simulator exhibits a strong correlation between simulated and real failure occurrence rates while maintaining low distributional divergence of failure modes. These findings indicate that DIAL is a promising method for developing realistic user simulators in multi-turn dialogue, facilitating reliable and cost-effective system evaluation prior to deployment.

Large language models (LLMs) are increasingly used as evaluators to assess output quality and preference alignment, yet providing reliable guarantees of agreement with human judgments remains challenging. Recent work introduces confidence-thresholding methods that provide such guarantees for pairwise comparisons, relying on the assumption that higher estimated confidence implies lower disagreement risk with humans. However, this assumption can break down when the number of candidate responses increases, since distributing probability mass across many alternatives can distort confidence estimates. To address this issue, we propose a Localize-Then-Decide framework. First, conformal prediction localizes a small shortlist that contains the human-preferred response with high probability. Then, a calibrated confidence-based rule selectively chooses a single response from this shortlist or abstains. This design restores the monotonic relationship between confidence and disagreement risk and enables high-probability agreement guarantees. Experiments with multiple candidate sizes across several datasets and judge LLMs demonstrate that our framework consistently achieves higher guarantee success rates and substantially higher coverage than single-stage baselines.

Diffusion and flow-matching models have revolutionized automatic text-to-audio generation, increasingly capable of generating high-quality audio outputs for speech and acoustic events. However, creative audio generation involving music and songs still has room for improvement. Recent open lyrics-to-song models like DiffRhythm, ACE-Step, and LeVo have set an acceptable standard for recreational use but lack the fine-grained word-level controllability desired by musicians. To our knowledge, our flow-matching-based JAM is the first to provide word-level timing and duration control in song generation, enabling fine-grained vocal control. To better align generated songs with human preferences, we implement aesthetic alignment through Direct Preference Optimization, iteratively refining the model using a synthetic dataset without manual annotations. Additionally, we aim to standardize evaluation of lyrics-to-song models through our public evaluation dataset JAME. We demonstrate that JAM outperforms open models on music-specific attributes.

With the rapid advancement of AI models, their deployment across diverse tasks has become increasingly widespread. A notable emerging application is leveraging AI models to assist in reviewing scientific papers. However, recent reports have revealed that some papers contain hidden, injected prompts designed to manipulate AI reviewers into providing overly favorable evaluations. In this work, we present an early systematic investigation into this emerging threat. We propose two classes of attacks: (1) static attack, which employs a fixed injection prompt, and (2) iterative attack, which optimizes the injection prompt against a simulated reviewer model to maximize its effectiveness. Both attacks achieve striking performance, frequently inducing full evaluation scores when targeting frontier AI reviewers. Furthermore, we show that these attacks are robust across various settings. To counter this threat, we explore a simple detection-based defense. While it substantially reduces the attack success rate, we demonstrate that an adaptive attacker can partially circumvent this defense. Our findings underscore the need for greater attention and rigorous safeguards against prompt-injection threats in AI-assisted peer review.

Long-horizon language agents must decide not only which evidence to use, but also when to pay for it. Retrieval and verification can improve a local decision while exhausting token and tool-time budgets needed later in the episode. We introduce ReserveGate, a one-step evidence controller that combines an explicit uncertainty sketch, an off-policy action-value model, and component-wise feasibility masks over the remaining resources. Unlike a fixed reserve heuristic, ReserveGate adapts its cost penalty when spending runs ahead of episode progress and gradually releases a protected late-stage reserve. Across three benchmark-derived agent environments, ReserveGate improves tight-budget success over the strongest matched controller by 2.6–3.8 percentage points, with gains of 4.6–6.2 points on the longest episode quartile. A latency-constrained variant reduces service-level incidents to 1.2–1.7% across environments while losing at most 0.8 points in success. On LoCoMo-Episodic, cross-model, cross-retriever, rate-limiter, and task-agnostic-sketch comparisons show that the effect is not explained by a single backbone, retrieval stack, or tuned spending rule.

Web agents must choose queries, clicks, form fills, and tool calls from language instructions while preserving constraints whose consequences may appear several steps later. We study this challenge as language-conditioned delayed-constraint tracking and present TraceWeaver, which couples linear-time latent rollouts, entity-centric episodic memory, and a learned critic for budgeted action selection. Across WebShop (500 episodes), Mind2Web (1,000-task split), WebArena (812-task split), and WorkArena (165-task enterprise split), TraceWeaver shows task-success gains of +2.7 to +4.7 over baselines under a matched test-time budget. Shared-pool controls reduce the confound between selection quality and proposal quality, while a language-phenomenon analysis concentrates the gains on tasks with dense, conflicting, or delayed constraints.

Search-augmented reasoning agents interleave multi-step reasoning with external retrieval, but uncontrolled retrieval can introduce redundant evidence, saturate the context, and destabilize reinforcement learning (RL). Existing outcome-based RL methods provide only sparse terminal rewards, offering limited guidance for intermediate information-acquisition decisions. We propose DeepControl, an adaptive information-control framework based on information utility, a state-dependent estimate of the marginal value of retrieved evidence. The framework regulates information acquisition along two axes: extent, i.e., whether retrieval should continue, and resolution, i.e., how much retrieved detail should be exposed. It implements these controls through retrieval-continuation guidance, hierarchical granularity control, and an annealed control-forcing scheme. This enables the policy to internalize effective acquisition behavior during training and operate without external control at test time. Across seven benchmarks, DeepControl consistently outperforms strong RL and retrieval baselines without explicit information control; compared with Search-R1, it improves average performance by +9.4 and +8.6 points on Qwen2.5-7B and Qwen2.5-3B, respectively. Additional analyses show improved search effectiveness, training stability, and evidence utilization.

As large language models(LLMs) are increasingly deployed in real-world applications, they inevitably encounter culturally sensitive scenarios that require applying cultural knowledge or values to the context in order to generate appropriate responses. However, existing research largely treats culture as static knowledge or abstract values, leaving the application of cultural norms in situated interactions underexplored. To address this gap, we first focus on the situated cultural reasoning of LLMs by proposing CuSiR, a simulation learning framework combining simulated scenario and reinforcement learning. We construct datasets covering both knowledge-based and social scenarios and conduct experiments across multiple role perspectives, instructions, settings, and models. Our results indicate that this framework effectively enhances LLMs’ ability to apply existing cultural knowledge, thereby improving their performance on tasks set in cultural scenarios. Code files are available at https://github.com/HoganZinger/Simulation_Learning

Retrieval-Augmented Generation (RAG) is a key approach to mitigating the temporal staleness of large language models (LLMs) by grounding responses in up-to-date evidence. Within the RAG pipeline, re-rankers play a pivotal role in selecting the most useful documents from retrieved candidates. However, existing benchmarks predominantly evaluate re-rankers in static settings and do not adequately assess performance under evolving information—a critical gap, as real-world systems often must choose among temporally different pieces of evidence. To address this limitation, we introduce FRESCO (Factual Recency and Evolving Semantic COnflict), a benchmark for evaluating re-rankers in temporally dynamic contexts. By pairing recency-seeking queries with historical Wikipedia revisions, FRESCO tests whether re-rankers can prioritize factually recent evidence while maintaining semantic relevance. Our evaluation reveals a consistent failure mode across existing re-rankers: a strong bias toward older, semantically rich documents, even when they are factually obsolete. We further investigate an instruction optimization framework to mitigate this issue. By identifying Pareto-optimal instructions that balance Evolving and Non-Evolving Knowledge tasks, we obtain gains of up to 27% on Evolving Knowledge tasks while maintaining competitive performance on Non-Evolving Knowledge tasks.

Large Language Models (LLMs) have shown impressive performance on complex tasks through Chain-of-Thought (CoT) reasoning. However, conventional CoT relies on explicitly verbalized intermediate steps, which constrains its broader applicability, particularly in abstract reasoning tasks beyond language. To address this, there has been growing research interest in latent CoT reasoning, where the reasoning process is embedded entirely within latent spaces. By decoupling reasoning from explicit language generation, latent CoT offers the promise of richer cognitive representations and facilitates more flexible, faster inference. This paper aims to present a comprehensive overview of this emerging paradigm and establish a systematic taxonomy. We analyze recent advances in methods, categorizing them from token-wise horizontal approaches to layer-wise vertical strategies. We then provide in-depth discussions of these methods, highlighting their design principles, applications, and remaining challenges. We hope that our survey provides a structured foundation for advancing this promising direction in LLM reasoning.

Recently, Large Language Models (LLMs) have gained significant attention due to their strong language understanding and generation capabilities, demonstrating impressive reasoning abilities as well as effective utilization of external knowledge. Many studies have proposed methods that specialize in improving performance for individual tasks. However, ironically, only a limited number of attempts have explored general-purpose, task-agnostic methods. In this work, we present a unified framework integrating reasoning and Retrieval-Augmented Generation (RAG) tasks. We further propose Controllable White-Box Meta-Prompting (CWM), a low-cost white-box method for adaptive RAG tasks previously dominated by black-box approaches, without requiring external decision modules or multi-sampling. CWM achieves state-of-the-art performance on three adaptive RAG benchmarks across recent LLMs, including GPT-oss-20b, Qwen3-14b, and Llama3.1-8b, while also demonstrating strong generality by extending to reasoning tasks. In addition, CWM provides controllability by enabling retrieval decisions to be regulated through the manipulation of internal model signals. Our code is available at https://github.com/JeongEunhye00/CWM.

Existing approaches for synthesizing Partial Differential Equation (PDE) solvers predominantly rely on autoregressive models, yet their global left-to-right decoding incurs substantial redundancy when addressing inherently localized bugs. In this work, we challenge this inefficient paradigm and propose DiffPDE, a framework leveraging discrete diffusion language models for targeted code repair. By introducing a localized re-masking and infilling strategy, DiffPDE regenerates only erroneous regions while preserving correct context, naturally aligning generation with the sparse nature of PDE errors. Furthermore, to handle coupled bugs requiring sequential interventions, we present Iterative Debugging GRPO (ID-GRPO), a reinforcement learning scheme that enables multi-round debugging within single trajectories via intermediate rewards. Experiments on PDEBench show that DiffPDE achieves competitive accuracy, outperforms same-scale AR models, and significantly accelerates repair.

Existing prompt optimization methods often improve prompts for a specific task or deployment setting, but their optimized prompts transfer poorly to new tasks and usually require re-optimization. We propose Prompting to Prompt (PTP), a framework that learns meta-templates as structured intermediate representations for reusable prompt construction. PTP decomposes prompts into transferable elements and optimizes them with LLM-generated textual gradients. Its inner loop refines task-specific prompts from natural-language feedback and an element list, while its outer loop abstracts structural changes across tasks to update the meta-template.The learned meta-template is then reused for task-level and query-level prompt generation, reducing the need to rerun full prompt optimization for each target setting. Experiments on six benchmarks show gains in both offline transfer and test-time deployment settings, with improvements up to 10.52 points on Arena-Hard.

Knowledge graph reasoning (KGR) infers missing facts, with recent advances increasingly harnessing the semantic priors and reasoning abilities of Large Language Models (LLMs). However, prevailing generative paradigms are prone to memorize surface-level co-occurrences rather than learning genuine relational semantics, limiting out-of-distribution generalization. To address this, we propose RADAR, which reformulates KGR from generative pattern matching to discriminative relational reasoning. We recast KGR as discriminative entity selection, where reinforcement learning enforces relative entity separability beyond token-likelihood imitation. Leveraging this separability, inference operates directly in representation space, ensuring consistency with the discriminative optimization and bypassing the generation-induced hallucinations. Across four benchmarks,RADAR achieves 5–6% relative gains on link prediction and triple classification over strong LLM baselines, while increasing task-relevant mutual information in intermediate representations by 62.9%, indicating more robust and transferable relational reasoning.

Improving the quality of OCR transcriptions is crucial for researchers working with digital archives, as OCR errors hinder readability and downstream NLP tasks. While prior work has focused on seq2seq neural machine translation and, more recently, prompt-based approaches with large language models (LLMs), these methods rely solely on text, overlooking the visual characteristics of source documents. In this work, we explore the potential of vision-language models (VLMs) for post-OCR correction to evaluate the impact of integrating visual information. We conduct experiments on the BLN600 dataset of 19th-century newspaper articles, comparing Llama 3.2 Vision and Qwen 2.5 VL to their text-only LLM counterparts. Our results show that VLMs significantly outperform LLMs, achieving approximately 90% reduction in character and word error rates, with the optimal setting being one that takes as input both document image and prior noisy OCR. This suggests that VLM-based post-OCR correction can be a viable and effective approach for correcting back catalogues of poorly OCR’ed historical documents and improving the usability of digital archives.

Large Language Models (LLMs) excel at text generation, but still struggle to produce accurate, long-form documents grounded in large knowledge bases. Existing “top-down” approaches—which plan before retrieving evidence—often result in fragmented or unfaithful content due to misalignment between the plan and available knowledge.

To overcome this, we introduce a “bottom-up” framework based on a “Retrieval-First for Knowledge, Clustering for Structure” strategy. Our method begins by establishing the “knowledge boundaries” of the source corpus through comprehensive retrieval, followed by unsupervised clustering to group evidence into distinct “knowledge clusters.” These clusters objectively inform a hierarchical outline and subsequent text generation, ensuring strict adherence to the source and significantly reducing hallucination.

Experimental results on both 14B- and 32B-parameter models demonstrate that our method achieves performance comparable to or exceeding state-of-the-art baselines, and is expected to demonstrate unique advantages in knowledge-constrained scenarios that demand high fidelity and structural coherence. Our work presents an effective paradigm for generating reliable, structured, long-form documents, paving the way for more robust LLM applications in high-stakes, knowledge-intensive domains.

Large language models (LLMs) encode vast amounts of knowledge in their parameters, but these facts often require precise post-training rectification as they become outdated or incorrect over time. Traditional batch editing methods typically apply a uniform update strength across all instances. However, this uniform approach ignores the inherent variation in editing difficulty, leading to a fundamental trade-off between correction success and model stability. To overcome this limitation, we introduce AdaPE (Adaptive Pressure Editing), a parameter-free adaptive editing framework that dynamically adjusts the update behavior for each specific instance. Specifically, we derive a closed-form adaptive update rate based on the ratio of residual energy to spatial sensitivity, which strictly quantifies the intrinsic injection difficulty. This formulation adaptively limits the update strength for difficult facts to preserve model locality, while ensuring sufficient knowledge injection for easier ones. Extensive experiments on three benchmarks and two prominent LLMs demonstrate the superiority of our approach in consistently circumventing the fixed-strength bottleneck, achieving a new state-of-the-art balance between editing efficacy and model integrity. Our code is available at https://github.com/skyandkibo/adape.

Public large language models (LLMs) are typically safety-aligned during pretraining, but task-specific fine-tuning often erodes this alignment, introducing significant safety risks. Existing defenses are largely coupled with the fine-tuning process, incurring high computational overhead and complex workflows. To address these challenges, we propose Q-realign, a post-hoc defense based on quantization. Departing from the conventional view of quantization as a safety-eroding compromise, we reframe it as a dual-purpose mechanism for both compression and safety recovery. This allows Q-realign to decouple safety alignment from fine-tuning and naturally piggyback onto modern deployment pipelines. Experiments across multiple models and datasets demonstrate that our method substantially reduces unsafe behaviors while preserving task performance, significantly improving memory and GPU efficiency. Notably, our approach can recover the safety alignment of a fine-tuned 7B LLM on a single RTX 4090 within 40 minutes. Overall, our work provides a practical, turnkey solution for safety-aware deployment. Our code is released at https://anonymous.4open.science/r/Q-realign-copy-0D71.

Culturally loaded translation poses unique challenges for machine translation (MT), as meanings are deeply embedded in socio-cultural contexts beyond surface linguistic forms. Although large language models (LLMs) have enabled MT systems to achieve human-like quality in many scenarios, their ability to handle culturally loaded expressions remains underexplored. In this study, we systematically investigate the challenges posed by culturally loaded translation in LLM-based MT systems. We construct a Chinese-Japanese bilingual dataset from the culturally representative corpus Dream of the Red Chamber, containing 500 segments across diverse cultural categories. Using a comprehensive human evaluation protocol, we identify three main challenges: (1) task challenges, where frontier LLMs exhibit notable performance gaps and struggle with culturally embedded content; (2) human evaluation challenges, where evaluator backgrounds lead to substantial disagreement in translation judgments; and (3) automatic evaluation challenges, where widely used metrics fail to reliably assess translation quality for this task. These findings may offer valuable insights for culture-oriented translation research in both computational science and linguistics.

Real-time streaming video understanding in domains such as autonomous driving and intelligent surveillance poses challenges beyond conventional offline video processing, requiring continuous perception, proactive decision making, and responsive interaction based on dynamically evolving visual content. However, existing methods rely on alternating perception-reaction or asynchronous triggers, lacking task-driven planning and future anticipation, which limits their real-time responsiveness and proactive decision making in evolving video streams. To this end, we propose a StreamAgent that anticipates the temporal intervals and spatial regions expected to contain future task-relevant information to enable proactive and goal-driven responses. Specifically, we integrate question semantics and historical observations through prompting the anticipatory agent to anticipate the temporal progression of key events, align current observations with the expected future evidence, and subsequently adjust the perception action (e.g., attending to task-relevant regions or continuously tracking in subsequent frames). To enable efficient inference, we design a streaming KV-cache memory mechanism that constructs a hierarchical memory structure for selective recall of relevant tokens, enabling efficient semantic retrieval while reducing the overhead of storing all tokens in the traditional KV-cache. Extensive experiments on streaming and long video understanding tasks demonstrate that our method outperforms existing methods in response accuracy and real-time efficiency, highlighting its practical value for real-world streaming scenarios.

Recent reasoning LLMs (RLMs), especially those trained with verifier-based reinforcement learning, often perform worse with few-shot CoT than with direct answering. We revisit this paradox using high-quality reasoning traces from DeepSeek-R1 as demonstrations and find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal. A detailed analysis reveals two mechanisms behind this decline: (i) *semantic misguidance*, where high textual similarity leads the model to treat the target as the same as the exemplar and to copy intermediate steps verbatim; and (ii) *strategy transfer failure*, where the model struggles to extract useful reasoning strategies and apply them to target questions. Guided by these findings, we introduce Insight-to-Solve (I2S), which converts demonstrations into target-conditioned insights before solving, and I2S+, which adds iterative refinement. Across open- and closed-source models, both variants outperform direct answering and sequential test-time baselines. On a frozen 500-problem Omni-MATH transfer set, I2S+ improves R1-Distill-Qwen-7B and Qwen3-14B by +4.2 and +4.0 points, respectively. The same prompt stack also transfers to GPT-4.1, o1-mini, and R1-Distill-Llama-8B, showing that reasoning demonstrations can be converted from liabilities into useful test-time guidance.

Large language models (LLMs) perform strongly across tasks and languages, yet how improvements in one task or language affect other tasks and languages remains poorly understood. We conduct a controlled LoRA fine-tuning study across multiple open-weight LLM families and scales, using a standardised grid of 11 languages and four benchmarks. We fine-tune each model on a single task–language source, then evaluate it on all other task–language target pairs to measure transfer. We decompose transfer into three regimes: (i) Matched-Task (Cross-Language), (ii) Cross-Task (Matched-Language), and (iii) Cross-Task (Cross-Language). Single-source fine-tuning yields a net positive uplift across regimes, but the gains are strongly asymmetric. Matched-Task (Cross-Language) transfer emerges as the most effective and structurally regular regime, with transfer magnitude driven principally by the identity of the target language rather than model architecture. We identify a stable coarse-grained hierarchy in which some task and language targets consistently absorb gains from diverse sources, while others remain relatively isolated. These results imply that effective fine-tuning requires accounting for donor–recipient roles to maximise downstream gains while limiting collateral degradation in other capabilities.

Assessing progress toward the Sustainable Development Goals (SDGs) requires multi-step reasoning over visual cues, contextual knowledge, and development indicators, where incomplete evidence use and imperfect evidence integration can introduce hidden prediction biases. Real-world SDG monitoring further spans both qualitative judgments and quantitative estimation. However, existing benchmarks typically evaluate these aspects in isolation, obscuring systematic biases that emerge when models substitute priors for evidence. To address this gap, we propose SDGBiasBench, a large-scale benchmark suite for SDG-oriented vision–language reasoning. Spanning 500k human-curated multiple-choice questions and 50k regression tasks, the benchmark enables comprehensive assessment of both decision-level and estimation-level bias in Vision–Language Models (VLMs). Evaluations on SDGBiasBench reveal an intrinsic SDG bias in current VLMs, where predictions are frequently driven by SDG specific priors rather than reliable multi-modal cues. To mitigate such bias, we propose CADE (Contrastive Adaptive Debias Ensemble), a training-free, plug-and-play method that leverages modality-specific answer priors. CADE yields significant gains on the proposed benchmark, improving multiple-choice accuracy by up to 25% and reducing regression MAE by up to 12 points across multiple VLMs. We hope our work can foster the development of more fair and reliable AI systems for sustainable development.

Retrospective studies using electronic health records (EHRs) often require diagnosis labels that reflect whether treating physicians considered a disease to be present. Such physician-determined diagnosis (PDD) labels are expensive to obtain through expert review, while available alternatives provide only noisy proxies. We exploit two complementary proxy labels: administrative diagnostic codes and free-text disease listings in discharge summaries. We introduce a dual-noisy-label learning framework that models PDD as a latent variable and treats these proxies as noisy observations of it, allowing a pretrained language model to learn from their agreement and disagreement without expert-labeled training data. We evaluate the approach on extracting Acute Kidney Injury (AKI) PDD from MIMIC-IV discharge summaries. Based on an expert-adjudicated evaluation set of 150 summaries, our PLM-PDD model achieves 94.6% population-weighted accuracy, compared with 89.4% for the same model trained using administrative coding, and 85.9% and 78.0% for AKI proxies. Largest gains occur in the gray-zone, where the proxies disagree. Ablation experiments show that gray-zone examples and auxiliary ICD supervision contribute to performance, while expert-highlighted evidence receives substantially higher model attention than matched random text. These results demonstrate that complementary noisy EHR labels can be used to learn clinically meaningful diagnosis labels without requiring expert annotation.

Despite undergoing safety training to align with human values, large language models (LLMs) remain vulnerable to jailbreak attacks that circumvent built-in safeguards and elicit malicious behaviors given by the adversary. Existing defenses are mostly static and fail to adapt to new or unknown attack strategies, providing limited resilience. To address these, we propose a multi-agent defense framework, GuardrailAgent, that offers adaptive, end-to-end protection. The system autonomously gathers and summarizes attack characteristics into a continuously updated knowledge base that supports both input- and output-side detection. A feedback module integrates misclassified cases for iterative improvement at minimal cost. Tested on over ten representative jailbreaks, the framework achieved a 93.1% input-side detection rate, 93.9% overall detection, and a 94.7% F1 score, while rapidly adapting to unseen attacks. These results demonstrate GuardrailAgent’s outstanding capability to maintain LLM safety in the face of evolving jailbreak tactics.

Drug recommendation requires a deep understanding of individual patient context, especially for complex conditions like Parkinson’s disease. While LLMs possess broad medical knowledge, they fail to capture the subtle nuances of actual prescribing patterns. Existing RAG methods also struggle with these complexities because guideline-based retrieval remains too generic and similar-patient retrieval often replicates majority patterns without accounting for the unique clinical nuances of individual patients. To bridge this gap, we propose PACE-RAG (Patient-Aware Contextual and Evidence-Constrained RAG). Rather than directly copying frequent medications from retrieved patients, PACE-RAG personalizes recommendations by first extracting patient-specific clinical features, retrieving cases around these features, and then refining the final prescription using the patient’s current symptoms, active medication history, and focus-specific prescribing tendencies. By analyzing treatment patterns tailored to specific clinical features, PACE-RAG generates patient-specific medication recommendations along with an explainable clinical summary. PACE-RAG achieved the strongest performance among the evaluated inference-only LLM-based methods, reaching F1 scores of 80.84% and 47.22% on the Parkinson’s disease and MIMIC-IV cohorts, respectively. Our code is available at: https://github.com/ChaeYoungHuh/PACE-RAG.

Object hallucination remains a major challenge in Multimodal Large Language Models (MLLMs), even though they have achieved strong performance across widely vision-language tasks. A key reason for object hallucination is limited fine-grained visual grounding provided by fixed-resolution visual encoding. In this paper, we propose Visual Inference Gating for Object Refinement (VIGOR), a training-free inference-time framework for binary object verification. VIGOR adopts a selective coarse-to-fine pipeline that improve visual grounding without the overhead of always-on multi-view inspection. It first uses the global yes–no margin to estimate whether the available evidence is sufficient. When the evidence is insufficient, VIGOR performs local refinement on uncertain cases using a small set of overlapping crops. Finally, VIGOR combines global and local evidence with an asymmetric decision rule. Experiments on three benchmark datasets across five MLLM backbones show that VIGOR consistently outperforms strong training-free baselines, recovering missed fine-grained evidence while maintaining robust performance on challenging visual reasoning.

While current emotional support dialogue systems typically rely on expert-defined scalar rewards for alignment, these signals suffer from severe information sparsity. They cannot explain why a response failed or how to adapt to dynamic user states, often failing to capture interaction-level conversational consequences reflected in subsequent user responses. We explore whether user reactions can serve as an additional behavior-level optimization signal for emotional support alignment. Therefore, we propose Reaction Aware Policy Optimization (RAPO), a framework that optimizes over interaction consequences rather than rubric scores. RAPO treats dialogue as a reaction-driven process and utilizes simulated user responses to generate dense natural-language feedback through three core components: (1) Hindsight Dialogue Selection, which isolates pivotal turns that meaningfully alter user emotional trajectories; (2) Generative Hindsight Feedback, which transforms user reactions into contrastive ranking signals and natural-language critiques; and (3) Scalar–Verbal Hybrid Policy Optimization, which couples scalar reward optimization for global alignment with verbal feedback distillation for fine-grained semantic refinement. Extensive experiments on ESC and Sotopia demonstrate that RAPO significantly outperforms strong RL baselines in driving positive interaction outcomes.

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for combining large language models (LLMs) with external knowledge sources. However, RAG systems remain vulnerable to prompt injection attacks, which may mislead the retriever or generator to expose sensitive database contents. To address this issue, we propose KFS-RAG, a defense that mitigates information leakage by reformulating the retrieved context. Specifically, our method first identifies a small set of influential keywords from the retrieved context via an attention rollout plus a causal perturbation mechanism. These keywords are then used to guide an auxiliary LLM to generate a compact set of keyword-grounded facts from the retrieved passages. Finally, the original context is substituted with these curated facts, ensuring that the generator operates on sanitized evidence rather than the raw retrieved text. Experimental evaluations demonstrate that KFS-RAG significantly reduces the risk of database leakage under injection attacks while maintaining response accuracy and relevance. This work highlights a practical pathway toward building secure and trustworthy RAG systems.

Large language model (LLM) safety has received extensive attention since LLMs are vulnerable to manipulation. To understand and mitigate the risk, this paper studies the problem of jailbreak attacks on LLMs, which aim to deliberately break the safety guard of LLMs for harmful or unethical responses. Current black-box jailbreak attacks are limited by either a reliance on human expertise for manual prompt design or by intricate workflows in automated approaches. Towards this end, we propose a novel approach named Reward-guided Meta-prompt Evolving with reflection (ROOT) for automatic jailbreak attack generation. The core idea of our ROOT is to optimize a meta-prompt using attack rewards as jailbreak guidance. In particular, our ROOT feeds a meta-prompt with toxic questions into LLMs to generate prompts for jailbreak attempts. The responses from these attempts are further evaluated using a judge model, which further summarizes the reflection of both successful and unsuccessful jailbreak attempts into different meta-prompt optimization strategies. To reduce the noise, we estimate their reward scores to select high-quality strategies, which can optimize our meta-prompt for better attack generation ability. Extensive experiments show that ROOT demonstrates strong generalizability and broad adaptability by achieving high jailbreak success rates (above 90%) across both multiple LLMs and various categories of harmful tasks.

Current coding-agent benchmarks usually provide the full task specification upfront. Research-code implementation may instead re- quire an agent to preserve design commitments disclosed progressively across a long session. We introduce SLUMP, a controlled benchmark for testing implementation faithfulness when a mathematically structured machine-learning target is disclosed through synthetic, paper- grounded multi-turn interactions rather than in a single canonical implementation specification. The benchmark contains 20 recent ML papers, 371 atomic verifiable components, and approx- imately 60 coding requests per paper. An expo- sure audit verifies whether scored components are recoverable from the visible interaction. On Claude Code and Codex, the single-shot control achieves higher overall implementation fidelity on 16/20 and 14/20 papers, respectively. Struc- tural integration degrades under progressive dis- closure on both platforms, while semantic com- ponent faithfulness degrades substantially on Claude Code and remains approximately sta- ble on Codex. These conclusions persist under stricter exposure filters and a targeted human correction of judge false-absence errors. As a mitigation case study, ProjectGuard main- tains external semantic and structural project state and recovers 90% of Claude Code’s mea- sured faithfulness gap. Code and benchmark artifacts are available at https://github.com/lunaryan/drift-benchmark.

Conversational recommender systems (CRS) increasingly rely on user simulators for automated evaluation of sales agents. A key requirement for such simulators is the ability to model human decision-making. However, most LLM-based simulators often exhibit unrealistically strong information-processing capabilities, rarely exhibiting the hesitation and decision deferral commonly observed in real consumer behavior, resulting in overly high acceptance rates. To address this limitation, we propose Hesitator, a theory-grounded user simulation framework that explicitly models a specific dimension of human decision-making: decision deferral under choice overload. The framework introduces a modular Decision Module that separates utility-based item selection from overload-aware commitment decisions. Experiments across multiple user simulation frameworks, domains, sales modes, and LLM backbones show that integrating our module consistently mitigates unrealistic behaviors under increasing overload conditions. Furthermore, Hesitator reproduces established findings from behavioral economics, demonstrating its ability to model human decision behavior.

Reasoning-oriented large language models often achieve strong problem-solving performance by generating long chains of thought, but this behavior substantially increases inference cost and latency. In contrast, instruction-tuned models tend to answer more concisely, yet often lack comparable reasoning ability. This accuracy-efficiency mismatch motivates a lightweight approach that combines the strengths of both models without full model retraining. In this paper, we propose GRIP (Granular Reward-guided Interpolation of Parameters), a reward-guided parameter interpolation framework for efficient reasoning. Given a reasoning model and an instruction model with identical architectures, GRIP assigns learnable interpolation ratios to individual modules and optimizes only these ratios while keeping both source models frozen. The interpolation ratios are trained with a reward signal that favors responses that are both correct and concise. Experiments show that GRIP achieves a better accuracy-efficiency trade-off than fixed or search-based merging baselines and further reveals module-wise fusion patterns associated with efficient reasoning. Our code is publicly available at https://github.com/Dozen233/GRIP.

We investigate whether Large Language Models (LLMs) exhibit conspiratorial tendencies, whether they display socio-demographic biases in this domain, and how easily they can be conditioned into adopting conspiratorial perspectives. Conspiracy beliefs play a central role in the spread of misinformation and in shaping distrust toward institutions, making them an important testbed for assessing the social and psychological fidelity of LLMs and their potential to reproduce or reinforce harmful narratives. Although LLMs are often used as proxies for studying human behavior, it remains unclear whether they reproduce higher-order psychological constructs such as generalized conspiratorial beliefs. To bridge this research gap, we administer validated psychometric surveys measuring conspiratorial mindset to multiple models under different prompting and conditioning strategies. Our findings reveal that LLMs show partial agreement with elements of conspiracy belief, and conditioning with socio-demographic attributes produces uneven effects, exposing latent demographic biases. Moreover, targeted prompts can easily shift model responses toward conspiratorial directions, underscoring both the susceptibility of LLMs to manipulation and the potential risks of their deployment in sensitive contexts. These results highlight the importance of critically evaluating the psychological dimensions embedded in LLMs, both to advance computational social science and to inform possible mitigation strategies against harmful uses.

Retrieval-Augmented Generation (RAG) improves factual grounding by incorporating external knowledge into language model generation. However, when retrieved context is noisy, unreliable, or inconsistent with the model’s parametric knowledge, it introduces retrieval-prior conflicts that can degrade generation quality. While this problem has been studied in autoregressive language models, it remains largely unexplored in diffusion-based language models, where the iterative denoising process introduces unique challenges for integrating retrieved context. In this work, we propose Adaptive Retrieval-Augmented Masked Diffusion (ARAM), a training-free adaptive guidance framework for Masked Diffusion Models (MDMs) in RAG settings. ARAM dynamically calibrates the guidance scale during denoising according to the Signal-to-Noise Ratio (SNR) of the distributional shift induced by retrieved context. Intuitively, the model strengthens guidance when the retrieved context provides reliable corrective evidence and suppresses it when the contextual signal is noisy or non-supportive. Extensive experiments on multiple knowledge-intensive QA benchmarks show that ARAM improves overall QA performance over decoding-based RAG baselines.

Cognitive Behavioral Therapy (CBT) aims to identify and restructure automatic negative thoughts pertaining to involuntary interpretations of events, yet existing counseling agents struggle to identify and address them in dialogue settings. To bridge this gap, we introduce Step, a dataset that models CBT counseling by explicitly reflecting automatic thoughts alongside dynamic, action-level counseling sequences. Using this dataset, we train Stepper, a counseling agent that proactively elicits automatic thoughts and executes cognitively grounded interventions. To further enhance both decision accuracy and empathic responsiveness, we refine Stepper through preference learning based on simulated, synthesized counseling sessions. Extensive CBT-aligned evaluations show that Stepper delivers more clinically grounded, coherent, and personalized counseling compared to other strong baseline models, and achieves higher counselor competence without inducing emotional disruption.

Large language models (LLMs) are increasingly applied to vulnerability detection in real-world software systems, where the task goes beyond binary classification and requires calling context exploration over large code repositories. However, existing benchmarks primarily formulate the vulnerability detection as a classification problem without calling contexts and provide no formalized explanations, making them inadequate for evaluating LLM-driven vulnerability detectors upon repositories. To fill the gap, we introduce ManBo, a benchmark designed to assess the capability of LLM-driven vulnerability detectors in the real-world scenarios. Specifically, ManBo provides human-verified vulnerability explanations together with calling contexts, enabling principled evaluation of the ability to both identify vulnerabilities and explain root causes. It contains 205 vulnerable instances and 164 safe instances collected from 74 high-profile C/C++ projects, covering three critical types of memory corruption. We evaluate a broad spectrum of LLM-driven detectors on ManBo, and find that effective calling context retrieval is essential for repository-level vulnerability detection, exposing a fundamental gap between the capabilities of existing LLM-driven detectors and real-world requirements. ManBo is publicly available at https://github.com/jinyaoguo/ManBo.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance in mainstream visual understanding tasks, but their ability to process action scenes that contradict everyday common sense remains undertested. To address this gap, we introduce CAIT, a benchmark comprising 400 high-fidelity synthetic scenes focused on counter-intuitive visual actions, such as “a rabbit is chasing a tiger”, where visual evidence explicitly contradicts common-sense expectations. We evaluate human, leading proprietary models (e.g., Claude and Gemini), and 14 representative open-source MLLMs. Humans achieve near-perfect performance (around 0.95 accuracy) and proprietary models demonstrate robust understanding (achieving up to 0.88 accuracy), standard open-source instruction-tuned models perform at the chance level. Further analysis demonstrates that this failure is driven by a strong language prior: rather than trusting the visual input, they automatically override the anomalous visual signals with statistically common text descriptions. Although introducing Chain-of-Thought (CoT) reasoning mechanisms can improve accuracy, it significantly slows down the response and generates a new failure mode: models overthink the scenario and refuse to accept the actual visual content simply because it violates real-world physical laws. Finally, we demonstrate that targeted fine-tuning and structured prompting can effectively mitigate this reliance on language priors, enabling open-source models to accurately ground their reasoning in actual visual evidence.

Chinese discriminatory-language detection is challenging because harmful intent is often implicit and context-dependent. We propose MAAM (Myopia–Astigmatism Anchor Mechanism), a lightweight, model-agnostic framework inspired by functional visual blur: rather than preserving every token equally, MAAM retains discrimination-relevant semantic anchors and calibrates them with C–I–S contextual priors (Contextual Tone, Group Identity, and Stance Polarity). We also introduce ChLGBT, to our knowledge the first Chinese LGBT-focused discriminatory-language dataset, with 8,120 manually annotated samples and three ordinal labels: explicit bias, implicit bias, and emotional intensity. Across strong encoder baselines, MAAM improves all three prediction dimensions, with consistent gains in accuracy, F1, Brier score, and expected calibration error. Compared with frontier LLM baselines under zero-shot and few-shot prompting protocols, MAAM remains competitive while offering stronger compactness and stability. These results suggest that interpretable anchor preservation and contextual calibration provide a practical alternative to heavier model scaling for Chinese discriminatory-language assessment.

Aligning large language models with human preferences remains a challenge, primarily due to the critical role of preference data quality in effective alignment. Existing datasets are frequently plagued by inherent noise and distribution shifts, which inherently limit model performance. To bridge this gap, we propose AlignDiff, a preference data filtering framework driven by intrinsic model signals. AlignDiff first identifies samples with clear preferences using both positive and inverse signals, then prioritizes the more challenging samples based on the average negative log-likelihood gap, encouraging the model to learn richer information from them. AlignDiff is evaluated on two widely used model families (LLaMA and Qwen) and three benchmarks widely adopted in the alignment community (AlpacaEval 2.0, Arena-Hard, and MT-Bench). Across all settings, it consistently outperforms seven strong baselines. We conduct comprehensive ablation studies to validate the effectiveness of AlignDiff, and further show that difficulty-based curriculum learning improves model performance.

As chatbots become increasingly widespread, personalization has emerged as a key feature for accommodating diverse users and use cases. Personalization enables better responses rather than generic answers, but when using demographic attributes to drive recommendations, can also amplify harmful stereotypes. For instance, individuals from different cultural backgrounds may meaningfully want culturally distinct food preferences and values incorporated, but not different job recommendations. Yet, little is known about how these user preferences relate to actual platform behavior. Through a survey of user preferences and a novel field study methodology where users input our standardized questions to their own personalized chatbot interfaces then share the outputs back to us, we contribute real-world personalization insights on the gaps between what users want versus actually receive. We generally find that there is minimal race- and gender-based personalization in practice, which conflicts with users’ desires in some domains, e.g., cuisines. We conceptualize our findings through the framing of a personalization double bind: Black people and women desire more demographic-based personalization than White people and men in order to avoid overly generic responses (indirect discrimination), yet simultaneously express greater concern about stereotypical responses (direct discrimination).

Currently, a prevalent approach for enhancing Vision-Language Models (VLMs) performance is to encode both the high-resolution version and the thumbnail of an image simultaneously. While effective, this method generates a large number of image tokens. When combined with the widely used Rotary Position Embedding (RoPE), its long-term decay property hinders the interaction between high-resolution tokens and thumbnail tokens, as well as between text and image. To address these issues, we propose ID-Align, which alleviates these problems by reordering position IDs. In this method, high-resolution tokens inherit IDs from their corresponding thumbnail token while constraining the overexpansion of positional indices. Our extensive experiments conducted within the LLaVA-Next framework demonstrate that ID-Align delivers comprehensive improvements: it not only achieves notable quantitative gains across multiple benchmarks, highlighted by a significant enhancement on MMBench’s relation reasoning tasks, but also qualitatively improves the model’s attention distribution, making it more interpretable.

Large language models (LLMs) are increasingly used for codebook-based content analysis in computational social science, where codebooks translate theoretical concepts into labeling schemes. While prior work reports strong predictive performance, it remains unclear whether these systems produce reliable and robust annotations. We conduct a large-scale empirical failure-mode analysis of LLM-based coding pipelines across three datasets and five models by systematically evaluating their stability, robustness, and conceptual validity under controlled perturbations, domain shifts, and optimization procedures. We uncover multiple failures that are largely invisible under standard metrics, showing that high aggregate scores frequently coexist with severe label-level disparities, instability across runs, domain sensitivity, and shifts in label interpretation. Importantly, these failure modes persist under widely used optimization strategies and remain largely invisible under standard metrics such as accuracy and micro-F1. Our results demonstrate that current LLM-based coding pipelines cannot be treated as stable measurement tools without targeted validation beyond aggregate performance.

Multilingual large language models often struggle to reason in low- to mid-resource languages. Prior work has shown that translation can improve multilingual reasoning by helping models access stronger English-centric representations. This raises a central question: how much translation is needed for multilingual large language models to reason reliably, and when does additional translation instead trigger interference and overconfidence? Using LLaMA and Qwen models, we conduct extensive experiments that vary both text scope and language mode (target-only, English-only, and bilingual) to evaluate accuracy and reliability. Our results reveal a clear trade-off: English context often improves understanding and helps recover errors caused by non-English comprehension, yet adding redundant bilingual context can intensify interference. To address this trade-off, we propose Reliability-Aware Adaptive Inference (RAAI), a training-free test-time framework that (i) performs ECE-aware routing and prompt fusion and (ii) uses a mid-layer Risk Index (RI) to gate sequential reasoning, allocating extra compute only when it is likely to help while suppressing harmful bilingual redundancy. Across two model families, RAAI improves accuracy by 25 to 37.7 points in low-resource languages and reduces calibration error by about 3 to 6 points, with the largest gains appearing in the lowest-resource language tiers.

While Reinforcement Learning (RL) effectively incentivizes reasoning in Large Language Models, current pipelines are hindered by training instability and rapid entropy collapse. These limitations often stem from "Rollout Silencing" and low-quality gradient signals in standard sampling procedures. In this work, we propose a robust, data-centric framework to stabilize RL training. We first introduce Potential-Aware Query Mining (PAQM), which filters data dynamically to focus on the "Distillation Zone"—samples with high potential for capability elicitation. Furthermore, we present Hybrid Stratified Replay (HSR), a novel mechanism that restructures batches by stratifying rollouts based on Path Entropy, a rollout-level confidence proxy, and outcome reward. Within each optimization step, HSR reuses current-policy "Stability Anchors" and "Hard Negatives" to construct high-contrast optimization groups, then clears its buffers before the next step. This approach mitigates entropy collapse while improving the utilization of learning signals under limited compute. Our method outperforms strong baselines on complex reasoning tasks, offering a principled solution for stable and efficient RL fine-tuning.

Omni-modal large language models (OLLMs) jointly process vision, audio, and text, yet their modality bias under cross-modal conflict remains underexplored. Existing benchmarks conflate two distinct forms of evidence within a single modality: perceptual signals (e.g., a photograph or recording of a dog) and propositional signals (e.g., the declarative claim "this is a dog"), such that any measured modality bias is inherently confounded with evidence-form bias, precluding clean attribution to either source. To address this, we introduce Tri-PvP, an 8,000-sample tri-modal conflict benchmark crossing vision, audio, and text, where vision and audio each take perceptual or propositional form. Evaluating five OLLMs, we find robust visual bias across most models and evidence-type conditions. Crucially, we reveal a systematic asymmetry in evidence-form bias: models exhibit a stronger bias toward perceptual signal in vision but propositional in audio. Further analyses via layer-wise linear probing and contrastive decoding reveal that modality bias is already linearly decodable from early representation layers and can only be partially mitigated, calling for mitigation strategies beyond surface-level interventions.

Long-form visual storytelling requires maintaining continuity across shots, including consistent characters, stable environments, and smooth scene transitions. While existing generative models can produce strong individual frames, they fail to preserve such continuity, leading to appearance changes, inconsistent backgrounds, and abrupt scene shifts. We introduce CANVAS (Continuity-Aware Narratives via Visual Agentic Storyboarding), a multi-agent framework that explicitly plans visual continuity in multi-shot narratives. CANVAS enforces coherence through character continuity, persistent background anchors, and location-aware scene planning for smooth transitions within the same setting. We evaluate CANVAS on two storyboard generation benchmarks ST-BENCH and ViStoryBench and introduce a new challenging benchmark HardContinuityBench for long-range narrative consistency. CANVAS consistently outperforms the best-performing baseline, improving background continuity by 21.6%, character consistency by 9.6% and props consistency by 7.6%.

The representation learning of knowledge graphs (KGs) is a longstanding research problem. While graph neural networks (GNNs) have driven recent progress, they still struggle with encoding textual features and subtle relationships of KGs, particularly in conveying key information to large language models (LLMs). The emergence of multi-modal LLMs (MLLMs), which combine linguistic and visual understanding, presents an intriguing opportunity: Could their vision capabilities inspire mental visualization, facilitating conceptual thinking and semantic reasoning akin to human cognition? To investigate this premise, we propose SeeKG, an innovative framework that transforms KGs into visually rendered representations as image inputs for MLLMs. We evaluate SeeKG under both training-free and supervised fine-tuning settings, where the experimental results show that SeeKG excels in understanding KG sub-graphs and achieves competitive performance even without training or demonstrations. Further fine-tuning on small-batch data reveals that it outperforms state-of-the-art LLM-based methods by substantial margins across multiple KG completion datasets.

Translating natural language to First-Order Logic (NL2FOL) is crucial for logic-based reasoning, but most existing models and benchmarks focus on English, leaving cross-lingual generalization underexplored. In this study, we present the first comprehensive investigation of cross-lingual robustness in NL2FOL. To support this, we introduce Multi-MALLS, the first multilingual benchmark for NL2FOL, which extends the widely-used MALLS dataset with high-quality translations into multiple languages. Using Multi-MALLS, we observe that state-of-the-art models experience a significant performance drop when applied to non-English inputs, highlighting the lack of robustness. To address this, we propose MA-FOL, a multi-agent framework that improves generalization without requiring any training on target languages. By decomposing the task into three modules, a language-agnostic structure generator, a language-specific predicate combiner, and a refinement component, MA-FOL achieves robust zero-shot generalization across diverse linguistic inputs. Additionally, we show that traditional evaluation metrics, such as Exact Match, often fail to assess semantic correctness. To remedy this, we introduce large language models (LLMs)-Judged Semantic Equivalence (SE), a new metric that leverages LLMs to evaluate whether generated formulas preserve the intended meaning. Extensive experiments demonstrate that MA-FOL outperforms strong baselines on Multi-MALLS, without any multilingual fine-tuning. The SE metric further reveals semantic correctness that traditional metrics miss. Overall, our work provides a benchmark to test robustness of NL2FOL, a framework to improve it, and a metric to evaluate it more effectively.

We approach multilinguality as sense adaptation: aligning latent meaning representations across languages rather than relying solely on shared parameters and scale. In this paper, we introduce SENse-based Symmetric Interlingual Alignment (SENSIA), which adapts a Backpack language model from one language to another by explicitly aligning sense-level mixtures and contextual representations on parallel data, while jointly training a target-language language modeling loss to preserve fluency. Across benchmarks on four typologically diverse languages, SENSIA generally outperforms comparable multilingual alignment methods and achieves competitive accuracy against monolingual from-scratch baselines while using 2–4 times less target-language data. Analyses of learned sense geometry indicate that local sense topology and global structure relative to English are largely preserved, and ablations show that the method is robust in terms of design and scale. Our results show that sense adaptation provides a simple, data-efficient, and linguistics-motivated path to multilingual fluency that is useful to low-resource languages and motivates further linguistics-motivated research.

Tokenization is a critical component of the language modeling pipeline, but it is poorly understood. Tokenizers can drastically impact downstream performance, but currently there is mixed evidence about which intrinsic tokenizer evaluations are able to predict which tokenizers will result in a better language model before training the model. In this paper, we analyze approximately 200 language models trained on nine different languages. We evaluate both models and their tokenizers and test whether there is a relationship between any proposed intrinsic tokenizer evaluations and model performance, as measured by a wide variety of benchmarks. Despite the quantity and variety of proposed intrinsic tokenizer metrics, none consistently predicts downstream performance across model families, tasks, and languages. This motivates the need for novel tokenizer evaluation metrics and approaches. Our analysis of a novel linguistically grounded metric introduced in this work shows that new metrics can capture aspects of tokenizer quality that existing metrics miss, though their predictive power is likely to vary across model families and languages.

Asking questions is fundamental for deeper understanding, yet evaluating and generating questions that improve downstream outcomes remains a challenging problem. To address this gap, we introduce QUEST (Question Utility Estimation with Simulated Tests), a framework that offers a new outcome-driven paradigm for question evaluation and generation, instead of indirect metrics that are based on the underlying content of the question and its answer. QUEST uses language models to simulate learners that ask questions while studying a textbook chapter and quantifies the utility of a question as its direct contribution to understanding as measured by the learner’s exam performance. To analyze the effectiveness of QUEST, we curate TEXTBOOK-EXAM, a benchmark that aligns textbook sections with exam questions across five academic disciplines. Using QUEST, we filter for high-utility questions and fine-tune question generators via rejection sampling and show that questions generated by QUEST-trained models improve simulated exam scores by 19.6% compared to the next strongest baseline. Furthermore, we demonstrate the practical real-world value of QUEST by showing that simulation-computed utility of a question is moderately positively correlated with its usefulness as perceived by human evaluators.

Large language models perform well on many logical reasoning benchmarks, but it remains unclear which core logical skills they truly master. To address this, we introduce LogicSkills, a benchmark that isolates three fundamental logical skills: (i) formal symbolization\unicode{x2013}translating premises into first-order logic; (ii) countermodel construction\unicode{x2013}showing that an argument is logically invalid by constructing a finite countermodel; and (iii) validity assessment\unicode{x2013}determining whether a conclusion follows from a set of premises. Items are drawn from the two-variable fragment of first-order logic without identity and are presented in both English and a Carrollian nonce-word language. All instances are solver-verified with Z3 for correctness and non-triviality. Across conventional instruction-tuned LLMs, performance is high on validity assessment but substantially lower on formal symbolization and countermodel construction, highlighting that high task-level accuracy can mask weaknesses in core logical skills. In contrast, recent reasoning-tuned models perform strongly across all three tasks, suggesting a more systematic logical skill profile.

As conversational AI systems increasingly engage users through multi-turn dialogue across diverse cultural contexts, cross-cultural competence, the ability to recognize, respect, and adapt to cultural differences, is essential for effective human-AI interaction. Existing evaluations largely rely on single-turn settings, overlooking the complexities of real-world, multi-turn conversations where cultural cues often emerge gradually. To address this gap, we introduce CultureTalk, a benchmark for evaluating multi-turn cultural reasoning in conversational AI. The dataset contains 1,741 culturally grounded dialogues derived from human-verified cultural knowledge resources and evaluates models’ ability to accumulate cultural context across turns and remain robust to topic shifts. Our evaluation reveals substantial performance gaps between single- and multi-turn settings, highlighting current models’ limitations in sustained culturally grounded context understanding. We further show that fine-tuning on augmented single-turn conversational data and inference scaling significantly improve multi-turn cultural reasoning, yielding a 64.7% relative gain in cultural knowledge entailment with a fine-tuned Qwen3-8B model.

Ultra-long video understanding remains an open challenge, as existing vision language models (VLMs) falter on such content due to limited context length and inefficient long-term memory retention. To address this, recent works have attempted to construct external knowledge bases and corresponding retrieval agumented generation (RAG) systems, yet these incur enormous storage and computational overhead. In this paper, we propose VideoMem, a novel framework that pioneers models long video understanding as a sequential generation task via adaptive memory management. Specifically, VideoMem dynamically updates a global memory buffer, which adaptively retains critical information while discarding redundant content across the video timeline. To efficiently train VLMs for such long-term tasks, VideoMem integrates the Progressive Grouped Relative Policy Optimization (PRPO) algorithm, equipped with two core modules: Progressive State Propagation (PSP) adaptively retains valid current states, propagates them to the next rollout step, and gradually narrows the model’s exploration space. Temporal Cascading Reward (TCR) further alleviates reward sparsity, improving sample utilization and accelerating convergence. Extensive experiments demonstrate that VideoMem significantly outperforms existing open source models across diverse benchmarks for ultra-long video understanding tasks.

Argument mining (AM) traditionally relies on multi-stage pipelines that train separate models for each sub-task, making end-to-end AM challenging and often unstable under multitask learning. In this work, we revisit multitask end-to-end AM in the context of large language models (LLMs) and show that multitask learning behaves fundamentally differently from pre-LLM settings. Across multiple datasets and LLMs, multitask fine-tuning of a single LLM consistently outperforms single-task models, while also exhibiting greater robustness to noisy labels and stronger cross-domain generalization. To understand these qualitatively different behaviors of LLMs, we conduct a layer-wise representation analysis, showing that cross-task features emerge in middle layers, early layers are highly domain-sensitive, and richer middle-layer representations drive performance improvements, robustness, and generalization. We further show that this mechanism extends beyond AM by observing similar patterns across other tasks. Our findings provide new insights into multitask learning with LLMs, highlight its potential for unified, efficient, and robust language understanding, and motivate layer-wise parameter efficient adaptation strategies within and beyond AM.

Argument relation identification (ARI) aims to recover the directed structure of argumentative discourse, yet most approaches, including recent large language models, treat relations as local decisions and rely on implicit reasoning to enforce global coherence. This often leads to structurally inconsistent argument graphs. We reframe ARI as a global structural reasoning problem and introduce topological ordering as an explicit prediction target. Extensive experiments show consistent improvements in ARI performance, edge-level accuracy, and global structural fidelity across models and learning paradigms. In addition, the predicted topological orders provide faithful, verifiable reasoning traces that are internally consistent with model predictions and aligned with gold argument structure, offering an interpretable alternative to free-form chain-of-thought explanations.

Bi-encoder architectures for biomedical entity linking learn flat embedding spaces that ignore ontology structure. HiCo-EL encodes hierarchical knowledge into bi-encoder representations through hierarchical contrastive pretraining. The core idea is within-space hierarchical multi-label contrastive learning (HMLC): contrastive losses applied independently within each encoder space, preserving cross-space compatibility for retrieval. Our formulation drops the monotonicity constraints of prior HMLC work, handles variable-depth hierarchy paths, and adds depth-weighted alignment between encoder spaces. Treating each entity as a leaf node unifies coarse type structure with fine-grained entity discrimination. On MedMentions and NCBI Disease, HiCo-EL outperforms ArboEL by +3.7% and +1.8% Recall@1 respectively. The gains reveal a key insight: breadth amplifies, depth generalizes—broad hierarchies help common entities most (+6.3% R@1), while deep hierarchies benefit unseen entities (+9.6% R@1 zero-shot).

The widespread adoption of Large Language Models has accelerated news content generation, yet methods to help readers contextualize and responsibly consume news content lag behind. Existing approaches to clustering news articles face significant limitations. Methods based on dense representations often struggle with robustness and efficiency, while those based on sparse (n-gram) representations are unsuitable for cross-lingual applications. Additionally, LLM-based methods are expensive to apply at the needed scale, restricting their use to only well funded organizations.

We present a novel method for efficient clustering of news articles using cross-lingual, sentence-level embeddings and community detection on similarity graphs. Our method can construct a document-level similarity graph from sentence-level embeddings, avoiding the use of sub-optimal embeddings of entire documents. The proposed method can be executed on consumer hardware, enabling practical news analysis for personal use and research.

We also release a manually curated dataset of clustered news articles, providing a valuable resource for future research. We use the dataset to benchmark our method using nine text embedding models with cross-lingual capability and propose various practical configurations.

Sampling multiple responses is a common way to improve LLM output quality, but it comes at the cost of additional computation. The key challenge is deciding when to stop generating new samples to balance accuracy gains against efficiency. To address this, we introduce BEACON (Bayesian Efficient Adaptive Criterion for Optimal N-stopping), a principled adaptive sampling framework grounded in Sequential Search with Bayesian Learning. BEACON sequentially generates responses from the policy LLM, updates posterior belief over reward distributions in real time without further training, and determines when to stop by weighing expected gains against computational cost. Sampling terminates once the marginal utility of further exploration no longer justifies the expense. We establish both theoretical optimality guarantees and practical tractability, and show empirically that BEACON reduces average sampling by up to 80% while maintaining response quality. We further demonstrate BEACON’s utility for cost-efficient preference data generation and outline practical extensions, offering actionable insights for future researchers.

Unlearning knowledge is a pressing and challenging task in LLMs because of their unprecedented capability to memorize and digest training data at scale, raising more significant issues on safety, privacy, and intellectual property. However, existing works, including parameter editing, fine-tuning, and distillation-based methods, are all focused on the flat sentence-level data but overlook the relational, multi-hop, and reasoned knowledge in the naturally structured data. In response to this gap, this paper introduces Graph Oblivion and Node Erasure (GONE), a benchmark for evaluating knowledge unlearning over structured knowledge graphs (KGs) facts in LLMs. This KG-based benchmark enables the disentanglement of three effects of unlearning: direct fact removal, reasoning-based leakage, and catastrophic forgetting. In addition, Neighborhood-Expanded Distribution Shaping (NEDS), a novel unlearning framework, is designed to leverage graph connectivity and identify anchor correlated neighbors, enforcing a precise decision boundary between the forgotten fact and its semantic neighborhood. Evaluations on LLaMA-3-8B and Mistral-7B across multiple knowledge editing and unlearning methods showcase the superior performance (1.000 on unlearning efficacy and 0.839 on locality) of NEDS on GONE and other benchmarks.

Multimodal Large Language Models have sparked significant interest due to their potential for social intelligence; however, their ability to perform sequential motivation reasoning remains insufficiently studied. Existing evaluations predominantly examine static text or isolated visual snapshots, which do not reflect the cumulative nature of real-world behavioral drivers. To address this gap, we introduce MulTivationBench, a benchmark designed to rigorously evaluate multimodal motivation reasoning within story-driven visual narratives. The benchmark builds upon established psychological frameworks—Maslow’s hierarchy and Reiss’s basic desires—and requires models to integrate accumulated multimodal context to infer evolving motivations. Results indicate that MulTivationBench presents a significant challenge: all tested models struggle to maintain consistent motivation reasoning across sequential contexts, revealing a critical disconnect between static recognition capabilities and the dynamic reasoning essential for human-like social understanding.

The task of event graph completion (EGC) aims to predict the missing relations between two events in a heterogeneous event graph (EG) to enhance its applicability in downstream tasks. Given that event graphs inherently contain both semantic and structural information, and that the effective aggregation of these two types of information can mutually reinforce their expressive power, we propose a **M**ulti-**E**xpert **R**outing (MER) framework, which leverages the mixture of experts (MoE) architecture to dynamically select appropriate experts for collaboratively addressing different key aspects of the EGC task. Specifically, our approach incorporates multiple specialized experts, leveraging prompt learning and a graph attention network, to independently model distinct semantic and structural features within EGs. Furthermore, an adaptive expert routing (AER) module is designed to dynamically fuse and select the missing relation predictions from the experts, thereby completing the EGs. Extensive experiments conducted on the EGC-MAVEN dataset demonstrate that our MER framework outperforms the state-of-the-art competitors.

With the rapid expansion of massive multilingual corpora, Multilingual Information Retrieval (MLIR) has emerged as a critical technology for global information access. MLIR enables users to retrieve semantically relevant documents from multilingual text collections using a single-language query. However, recent multilingual dense retrieval models often exhibit a strong preference for documents in the same language as the query. This leads to severe language bias, where top-ranked results are dominated by documents of specific languages, even when documents in other languages contain more semantically relevant information. To address this issue, we propose Shift, a training-free method applicable in the indexing stage. Specifically, Shift utilizes parallel translation pairs to estimate a relative language vector for each target language with respect to a source language. Subsequently, Shift corrects the language-specific offset by subtracting this relative language vector from document embeddings during indexing. Our comprehensive evaluation across four MLIR benchmarks and diverse dense retrieval models confirms that Shift can effectively mitigate language bias and enhance MLIR performance.

Large Multimodal Models (LMMs), characterized by their humanoid perception and generation capabilities, have been widely applied in critical physical world scenarios, yet their uncertainty remains mostly unknown. In this paper, we aim to answer three substantial questions: (1) Can we explicitly capture the uncertainty of LMMs? (2) How can we tackle the reasoning complexity arising from prompts across various modalities? (3) How can we quantify the uncertainty of LMMs when their responses extend to multiple modalities, beyond mere text? To answer these questions, we propose Uncertainty-o: (1) Uncertainty-o can explicitly estimate uncertainty of LMMs, with prompts and responses across five modalities, namely text, image, audio, video, and point cloud. (2) We propose multimodal prompt perturbation, where empirical studies indicate that semantically equivalent perturbations are more effective for eliciting LMM uncertainty. (3) We introduce multimodal semantic uncertainty, enabling uncertainty estimation when LMMs generate multimodal responses. Extensive quantitative and qualitative analyses validate the efficacy of Uncertainty-o in reliably estimating the uncertainty inherent in LMMs.

The order of training samples plays a crucial role in large language models (LLMs), significantly impacting both their external performance and internal learning dynamics. Traditional methods for investigating this effect generally require retraining the model with various sample orders, which is computationally infeasible for LLMs. In this work, we improve traditional methods by designing a retraining-free framework. By approximating Adam optimizer updates with first- and second-order Taylor expansions and utilizing random projection methods to store intermediate checkpoints, our framework can efficiently estimate model parameters for arbitrary training sample orders. Next, we apply our framework to two downstream research problems: (1) Training curriculum design for LLMs — We base our retraining-free framework to propose a novel curriculum learning strategy that augments curriculum proposals with estimated model performances, enabling more informed sample scheduling. (2) LLMs’ memorization and generalization effect analysis — We use our retraining-free framework to estimate how the positions of training samples influence LLMs’ capacity for memorization and generalization. We conduct extensive experiments to validate the effectiveness of our retraining-free framework in reproducing the true model performances, and further demonstrate its potential in optimizing LLM training curricula and analyzing the memorization and generalization effects of LLMs.

Code LLMs often portray inconsistent program behaviors. Developers typically employ benchmarks to assess Code LLMs, but most benchmarks are hand-crafted, static and do not target the consistency property. In this work, we pose the scientific question: how can we automatically discover inconsistent program behaviors in Code LLMs? To address this challenge, we propose an automated consistency testing method, called MUCOCO, which employs semantic-preserving mutation analysis to expose inconsistent behaviors in code LLMs. Given a coding query, MUCOCO automatically transforms its program into semantically equivalent programs (aka mutants) and detects inconsistencies between the mutants and the original program (e.g., different output or test failure). We evaluate MUCOCO using four (4) coding tasks and seven (7) LLMs. Results show that MUCOCO is effective in exposing inconsistency and outperforms the closest baseline (TURBULENCE). About one in seven (15%) inputs generated by MUCOCO exposed inconsistencies. Our work motivates the need to test Code LLMs for consistency property.

Hallucinations remain one of the major barriers to the reliable deployment of Large Language Models (LLMs). Recent works have explored hallucination detection using both supervised classification based approaches and unsupervised metric based approaches, with the latter remaining popular since they do not require labeled data. In this work, we study the performance and generalizability of these approaches through a comprehensive empirical analyses across 11 datasets and 10 models. We first establish that supervised methods achieve stronger in-domain performance but lag behind unsupervised methods in the out-of-domain setting. Then to narrow this generalization gap, we introduce a simple approach that improves the generalizability of supervised hallucination detectors by training on a curated, multi-domain data mix, which can complement the subsequent addition of task-specific data. In our experiments on hallucination detection on 697K QA samples from 12 open source QA datasets, we show that incorporating this general training allows supervised methods to surpass unsupervised metric based methods by an average of +7.25% on out-of-distribution data within the short-form QA setting, without addition of any task-specific data. We also analyze scaling behaviors and estimate how much task-specific data is required to achieve reliable performance, finding that models augmented with general data require up to 40.3% less task-specific data to achieve close to optimal performance. Together, our findings highlight a simple path toward fortifying supervised hallucination detectors against domain shift.

In dialogue systems, role-playing agents (RPAs) aim to embody specific characters, but they often struggle to exhibit emotional expressions that are consistent with their assigned roles, resulting in insufficient emotional fidelity. This paper proposes EmoRL, a novel curriculum reinforcement learning (RL) framework, to enhance the emotional fidelity of RPAs. EmoRL features a multi-stage training paradigm guided by a carefully designed suite of reward functions. These rewards, including Format, Fine-grained Exact (FE), Emotional Similarity (ES), and Coarse-grained Categorical (CC) rewards, assess emotional consistency from strict to lenient perspectives. The curriculum RL process applies these rewards sequentially from easy (CC) to difficult (FE), enabling the RPAs to progressively learn coarse-grained emotional tendencies before mastering fine-grained distinctions. Furthermore, we introduce an emotional dialogue data synthesis pipeline to augment training data diversity and balance. Experiments on the Personality EmotionLines Dataset (PELD) demonstrate that EmoRL significantly outperforms baselines like RoleLLM, SFT, and DPO across both single-turn and multi-turn dialogues, with particularly notable gains in longer conversations. Our findings confirm the effectiveness of the curriculum learning strategy, the importance of balanced data, and the value of synthesized data for improving emotional fidelity.

Large Vision-Language Models (LVLMs) demonstrate remarkable capabilities in multimodal tasks, but visual object hallucination remains a persistent issue. It refers to scenarios where models generate inaccurate visual object-related information based on the query input, potentially leading to misinformation and concerns about safety and reliability. Previous works focus on the evaluation and mitigation of visual hallucinations, but the underlying causes have not been comprehensively investigated. In this paper, we analyze each component of LLaVA-like LVLMs—the large language model, the vision backbone, and the projector, to identify potential sources of error and their impact. Based on our observations, we propose methods to mitigate hallucination for each problematic component. Additionally, we developed two hallucination benchmarks: QA-VisualGenome, which emphasizes attribute and relation hallucinations, and QA-FB15k, which focuses on cognition-based hallucinations.

Video Large Multimodal Models (VLMMs) have achieved impressive progress in video understanding, yet they remain limited in adaptive reasoning—the ability to revise conclusions when new evidence emerges. In reality, conclusions are rarely set in stone; additional context can strengthen or weaken an initial inference. To capture this capability, we introduce Defeasible Video Entailment (DVidE), a new task that challenges models to think like doubters, requiring them to dynamically update entailment judgments as evidence evolves. In DVidE, given a video premise and a textual hypothesis, the models must determine whether a new update strengthens or weakens the hypothesis (classification task) or generates a coherent update that modifies the entailment relationship (generation task). For solving the classification task, we propose the Chain of Counterfactual Thought (CoCT) framework, utilizing counterfactual reasoning, ASR-enhanced video content, and rationale refinement to reduce inference bias. For the generation task, we develop a framework that combines ASR output with a Large Language Model (LLM) to produce coherent, contextually relevant updates aligned with the intended strengthener or weakener goals. We also introduce a benchmark dataset with human annotations and an LLM-based evaluation metric for assessing generative performance. Experimental results show substantial improvements across multiple VLMM baselines, demonstrating that CoCT significantly enhances both classification and generation performance and strengthens the dynamic reasoning capabilities of video-language models.

Medical large language models hold promise for reducing healthcare disparities, yet Hindi remains severely underrepresented. While medical LLMs excel in high-resource languages, their performance degrades sharply in Hindi, particularly on Indian systems of medicine. We argue that robust cross-lingual medical transfer requires Hindi reasoning. To this end, we introduce HiMed, a Hindi reasoning medical corpus and benchmark suite covering both Western and Indian medicine. We further propose HiMed-8B, a Hindi-form medical reasoning LLM, through the design of decaying scaffolding reward. Extensive experiments demonstrate improvement in Hindi medical reasoning performance and a reduction in the English–Hindi accuracy gap. Ablation studies validate the contribution of each training stage and reward component. All code, data, and weights are available.

Reinforcement Learning (RL) is essential for evolving Large Language Models (LLMs) into autonomous agents capable of long-horizon planning, yet a practical recipe for scaling RL in complex, multi-turn environments remains elusive. This paper presents a systematic empirical study using TravelPlanner, a challenging testbed requiring tool orchestration to satisfy multifaceted constraints. We decompose the agentic RL design space along 5 axes: reward shaping, model scaling, data composition, algorithm selection, and environmental stability. Our controlled experiments yield 7 key takeaways, e.g., (1) reward and algorithm choices are scale-dependent as smaller models benefit from staged rewards and enhanced exploration, whereas larger models converge efficiently with simpler dense rewards, (2) ~1K training samples with a balanced difficulty mixture mark an empirical sweet spot for balancing in-domain and out-of-domain performance under our primary setting, and (3) environmental stability is critical to prevent policy degradation. Based on our distilled recipe, our RL-trained models achieve state-of-the-art performance on TravelPlanner, outperforming leading LLMs.

Although batch prompting makes large language model inference more efficient by processing multiple instances simultaneously, it suffers from unpredictable downstream task performance. We propose cascaded batch prompting, a two-stage approach designed to resolve the unpredictability of conventional batch prompting by disentangling complex reasoning from symbol grounding. Experiments on multiple-choice question answering and natural language inference demonstrate that the proposed method outperforms the standard single prompting baseline while achieving a speedup proportional to batch size, establishing a new state of the art on the Pareto frontier.

Existing long-context benchmarks for Large Language Models (LLMs) focus on evaluating comprehension of long inputs, while overlooking the evaluation of long reasoning abilities. To address this gap, we introduce LongReasonArena, a benchmark specifically designed to assess the long reasoning capabilities of LLMs. Our tasks require models to solve problems by executing multi-step algorithms that reflect key aspects of long reasoning, such as retrieval and backtracking. By controlling the inputs, the required reasoning length can be arbitrarily scaled, reaching up to 1 million tokens of reasoning for the most challenging tasks. Extensive evaluation results demonstrate that LongReasonArena presents a significant challenge for both open-source and proprietary LLMs. For instance, even GPT-5 achieves only 27.8% accuracy on Level 3. Further analysis also reveals that the accuracy exhibits a linear decline with respect to the logarithm of the expected number of reasoning steps. Our code and data are available at https://github.com/LongReasonArena/LongReasonArena.

Aligning large language models with human values is a critical challenge for ensuring public safety. This alignment is commonly achieved through methods such as Direct Preference Optimization and RLAIF/RLHF. Despite extensive research, prior studies show that models often fail to maintain consistent alignment performance across languages, resulting in unequal treatment of users. To address this issue, we introduce a regularization term in the Direct Preference Optimization fine-tuning process that penalizes discrepancies in model behavior between languages using a bilingual corpus. Our results demonstrate that this approach effectively reduces the alignment gap across languages following fine-tuning.

With the growing demand for processing multiple image sequences in real-world applications, various visual token pruning methods have emerged to mitigate the computational and context length constraints faced by Large Vision Language Models (LVLMs). However, most existing pruning approaches rely on static strategies that struggle to adapt across different architectural LVLMs and multi-image scenarios, and are additionally constrained by their dependence on attention computations that are incompatible with efficient techniques like FlashAttention.To address these limitations, we propose a training-free, Adaptive Visual Token Pruning (AVTP) framework, applicable to diverse LVLM architectures. We strategically determine pruning layers based on empirical analysis of visual attention distributions across various LVLMs, and implement adaptive pruning ratios in multi-image contexts where images of higher importance retain proportionally more tokens. We conduct extensive experiments across multiple LVLMs with different architectures to demonstrate the effectiveness and robustness of AVTP. Specifically, Qwen3VL-8B achieves 2× inference speedup while maintaining 96.1% of its original accuracy on multiple multi-image benchmarks, InternVL3.5-8B retains 94.1% accuracy, and LLaVA-OneVision-7B even exceeds its original baseline performance. Our code is available at https://github.com/zry13/AVTP.

Personalized alignment from preference data has focused primarily on improving personal reward model (RM) accuracy, with the implicit assumption that better preference ranking translates to better personalized behavior. However, in deployment, computational constraints necessitate inference-time adaptation such as reward-guided decoding (RGD) rather than per-user policy fine-tuning. This creates a critical but overlooked requirement: reward models must not only rank preferences accurately but also effectively guide generation. We demonstrate that standard RM accuracy fails catastrophically as a selection criterion for deployment-ready personalized rewards. We introduce policy accuracy—a metric quantifying whether RGD-adapted LLMs correctly discriminate between preferred and dispreferred responses—and show that upstream RM accuracy correlates only weakly with downstream policy accuracy (Kendall’s tau = 0.08–0.31). More critically, we introduce Pref-LaMP the first personalized alignment benchmark with ground-truth user completions, enabling direct behavioural evaluation. On Pref-LaMP, we expose a complete decoupling between discriminative ranking and generation metrics: methods with 20-point RM accuracy differences produce almost identical output quality, and methods with high ranking accuracy can fail to generate behaviorally aligned responses. These findings reveal that the field has been optimizing for proxy metrics that do not predict deployment performance, and that current personalized alignment methods fail to operationalize preferences into behavioral adaptation under realistic deployment constraints. In contrast, we find simple in-context learning (ICL) to be highly effective - dominating all reward-guided methods for models ≥3B parameters, achieving ~3 point ROUGE-1 gains over the best reward method at 7B scale.

Reading behaviour varies not only with linguistic input, but also with reader proficiency. In this study, we investigate whether the layer-wise relationship between surprisal from large language models (LLMs) and human gaze behaviour differs across readers with different levels of proficiency and across gaze measures. Using eye-tracking data from the MECO L2 corpus, we compare readers with high and low vocabulary proficiency on first-pass gaze duration (FPGD) and total gaze duration (TGD). We quantify the distribution of the predictive power of surprisal across model layers using Predictive Depth (PD). Across 12 tested LLMs, we find that readers with lower vocabulary proficiency tend to show deeper PD for FPGD, while this difference is smaller for TGD. Also, TGD itself shows deeper PD than FPGD in both proficiency groups. These patterns suggest that where predictive power is concentrated across LLM layers may be related to the timing and breadth of the reading processes captured by different gaze measures, and that this relationship can vary with reader proficiency. Our leave-one-out analysis further shows that the advantage of informative internal layers extends to unseen texts, although the practical improvements in prediction are limited. Overall, our results show that layer-wise LLM surprisal provides a useful perspective on variation in reading behaviour across both reader groups and gaze measures.

Although Large Vision-Language Models (LVLMs) excel at single-image understanding tasks, their performance often degrades significantly when processing multi-image inputs due to unintended cross-image information leakage, in which visual content from different images is erroneously fused and image-specific contextual information is lost. Current approaches attempt to mitigate this by inserting special vision delimiter tokens between images during input formatting to help models distinguish image boundaries and preserve context. However, our analysis reveals that such delimiters alone yield limited gains because models do not sufficiently perceive or respond to these boundary signals. To address this, we propose Vision-Delimiter-aware Attention Head Modulation (VDAHM), a training-free method that modulates the output of delimiter-sensitive attention heads, reorienting the model’s internal attention flow to strengthen its structural perception of boundary signals. Extensive evaluations across multiple benchmarks, including Mantis, BLINK, MuirBench, and Qbench2, demonstrate that VDAHM consistently improves LVLMs performance with virtually no additional computational cost. The source code is available at [https://github.com/ZeroNLP/VDAHM](https://github.com/ZeroNLP/VDAHM).

Text-to-motion (T2M) generation has achieved impressive realism but poses significant safety risks by faithfully executing unsafe textual prompts. Existing motion unlearning methods face dual challenges: brittle keyword filtering and discrete codebook manipulation, which may degrade benign generation quality and introduce jerky transitions. To address these challenges, we propose SafeMo, a linguistically grounded framework that leverages Large Language Models (LLMs) to align motion generation with safety constraints. First, we introduce SafeMoEngine, an LLM-agent pipeline that autonomously classifies unsafe intents and performs semantic rewriting to construct SafeMoVAE-29K, the first safety-aligned text-to-motion dataset. Second, we propose Minimal Motion Unlearning (MMU), a continuous-space unlearning strategy that projects unsafe semantic concepts into a task vector for precise negation, avoiding the quantization artifacts of discrete methods. Experiments demonstrate SafeMo’s effective unlearning performance by more strongly suppressing unsafe text–motion retrieval than the prior SOTA motion unlearning baseline LCR, reducing forget-set R@1 by 35.3% and achieving 2.61× normalized retrieval trade-off score (NRTS), while showing better or competitive benign utility on safe prompts, with similar trends on Motion-X. Code and checkpoints are publicly available at https://github.com/AIGeeksGroup/SafeMo. Project page: https://aigeeksgroup.github.io/SafeMo.

Enhancing LLM reasoning in federated settings is nontrivial due to stringent computational, communication, and privacy constraints, especially in healthcare, where clinically consequential decisions require not only accuracy but also interpretable, auditable rationales to meet safety, accountability, and regulatory requirements. Conventional federated fine-tuning largely imitates final answers rather than cultivating step-by-step reasoning, often relying on privacy-sensitive centralized distillation and still incurring substantial communication overhead. We address this gap with FedCoT, a federated reasoning framework that combines lightweight chain-of-thought resampling with a compact discriminator for selection, and client-aware LoRA stacking with weighted classifier aggregation to accommodate heterogeneity while reducing aggregation noise and communication; clients generate candidate chains and supervision locally, and only lightweight modules are aggregated on the server. Experiments on medical reasoning benchmarks show consistent gains under tight resource budgets while keeping data local and respecting privacy, offering an interpretable and resource-efficient solution.

Open-domain question answering (ODQA) over multimodal document corpora requires linking evidence scattered across text, tables, and figures. Existing systems often store these sources separately or retrieve only coarse pages, which weakens global evidence linking. We present PILAR, a page-grounded unified evidence representation instantiated as an entity-linked assertion graph. PILAR maps sentence-, table-, and figure-derived facts into a common assertion space and uses the graph as a controlled linking layer over robust page retrieval. In a shared-reader evaluation with four agent frameworks, fourteen retrieval backends, and two benchmarks, PILAR achieves the best end-to-end EM/ANLS. Gains are largest on compositional, cross-document, and multimodal questions, with a single-shot improvement of +1.6 EM over flat retrieval, rising to +2.9 on compositional and +5.9 on 3-hop questions. Ablations show that current gains are driven mainly by the text-instantiated slice of the framework, while visual assertions help only after locality-aware filtering. We therefore position PILAR as a unified evidence representation for multimodal ODQA rather than a standalone visual-reasoning module.

Fine-tuning large language models for code editing has typically relied on mining commits and pull requests. The working hypothesis has been that commit messages describe human intent in natural language, and patches to code describe the changes that implement that intent. However, much of the previously collected data is noisy: commit messages are terse, human-written commits commingle several unrelated edits, and many historical commits come from simple, rule-based bots.

The recent adoption of software engineering agents changes this landscape. Code changes co-authored by humans and agents are often accompanied by substantially more explicit natural-language descriptions of intent and rationale. Moreover, changes in public repositories may have passed through human selection and review, and we treat this potential human filtering as an indirect signal of data quality.

We present AgentPack, a corpus of 1.8M code edits co-authored by Claude Code, OpenAI Codex, and Cursor Agent across public GitHub projects up to early October 2025. We describe the identification and curation pipeline, quantify adoption trends of these agents, and analyze the structural properties of the edits. Finally, we show that models fine-tuned on AgentPack outperform models trained on prior human-only commit corpora, highlighting the potential of using public data from software engineering agents to train future code-editing models.

We introduce MineCEraft (Minecraft Construction Engineering Benchmark, pronounced mine-see-ee-raft), an easy-to-use, open-source benchmark designed to systematically evaluate the reliability and limitations of LLMs for construction tasks in Minecraft. The MineCEraft benchmark comprises 723 domain-expert hand-crafted natural-language instructions with programmatically verifiable evaluation, spanning 17 distinct task categories, providing a safe and controllable experimental environment for assessing LLMs’ ability to perform realistic construction engineering tasks. With this benchmark, we conduct an in-depth evaluation of state-of-the-art LLMs and perform a detailed error analysis, revealing key failure modes and practical challenges in applying LLMs to construction engineering tasks.

The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data. We address this by introducing , an expansive dataset of 207,489 agentic trajectories spanning nine programming languages (Python, Go, TS, JS, Rust, Java, PHP, C, C++). Sourced from 20,000 real-world PRs via OpenHands and SWE-agent harnesses, the dataset utilizes a hybrid-reasoning synthesis: Minimax-M2.5 generates trajectories with explicit "thinking" processes, while Qwen3.5-122B provides high-quality behavioral traces. Filtered for permissive licenses (MIT, Apache, BSD) from SWE-rebench-V2, this data facilitates the training of models capable of long-horizon reasoning. We validate the dataset by fine-tuning the Qwen3-30B-A3B series (Thinking, Coder, and Instruct). The best performing model achieves a resolve rates of 60% on SWE-bench Verified, 46.2% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro. These results establish as a premier resource for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs.

Spatial audio understanding aims to enable machines to interpret complex auditory scenes, particularly when sound sources move over time. In this work, we study Spatial Audio Question Answering (SAQA) with a focus on movement reasoning, where a model must infer object motion, position and directional changes directly from stereo audio. First, we introduce a movement-centric spatial audio augmentation framework that synthesizes diverse motion patterns from isolated mono audio events, enabling controlled and scalable training data generation. Second, we propose an end-to-end multimodal finetuning approach with a thinking mode, which allows audio-language models to produce explicit intermediate reasoning steps before predicting an answer. Third, we investigate the impact of query-conditioned source separation as a preprocessing stage and compare three inference regimes: no masking, an audio grounding model (AGM), and ground-truth masks. Our results show that the benefits of thinking mode are most pronounced when paired with high-quality source separation rather than in isolation, with an improvement of +5.1% when a single event is present in the question. These findings highlight the relationship between movement modeling, reasoning, and separation quality, offering new insights for advancing spatial audio understanding.

The ability to generate counterarguments is important for critical thinking and balanced discourse, yet existing approaches typically produce only a single counterargument, failing to capture the diversity and persuasiveness required in real-world debates. To address this limitation, we propose Persona-guided Tree-based Counterargument Generation (PTCG), a framework that combines Tree-of-Thoughts–inspired step-wise generation and pruning with speaker persona selection. By estimating the author’s persona from the original argument and incorporating speaker personas representing distinct perspectives, PTCG operationalizes perspective-taking and enables the generation of diverse counterarguments. Results from LLM-as-a-Judge, classifier-based assessment, and human evaluations indicate that PTCG shows consistent improvements in both the diversity and persuasiveness of counterarguments compared to baseline methods.

As scientific literature grows and papers increasingly under-report limitations, multi-agent LLMs offer a promising approach to systematically uncover these hidden failure modes. Here, we introduce Tree-of-Concerns, a multi-agent framework that deploys specialized skeptic personas, each operating through a category-specific analytical lens, as parallel debate trees to extract unstated limitations from scientific papers. Each persona conducts structured, evidence-grounded argumentation, while a Panel Review mechanism re-evaluates each surviving claim from all five perspectives to correct category drift and severity miscalibration. Through retrieval-free, single-paper experiments on ToC-Bench, our benchmark of 414 research papers with 1,905 unstated limitations, sourced from reviewer-reported weaknesses and follow-up citation critiques, we demonstrate that ToC improves precision by 79% and coverage by 11% relative to the strongest baseline, surfacing specific, evidence-grounded concerns that support reviewers in systematic evaluation.

Recent years have witnessed rapid advances in graph representation learning, with the continuous embedding approach emerging as the dominant paradigm. However, such methods encounter issues regarding parameter efficiency, interpretability, and robustness. Thus, Quantized Graph Representation (QGR) learning has recently gained increasing interest, which represents the graph structure with discrete codes instead of conventional continuous embeddings. Given its analogous representation form to natural language, QGR also possesses the capability to seamlessly integrate graph structures with large language models (LLMs). As this emerging paradigm is still in its infancy yet holds significant promise, we undertake this thorough survey to promote its rapid future prosperity. We first present the background of the general quantization methods and their merits. Moreover, we provide an in-depth demonstration of current QGR studies from the perspectives of quantized strategies, training objectives, distinctive designs, knowledge graph quantization, and applications. We further explore the strategies for code dependence learning and integration with LLMs. At last, we give discussions and conclude future directions, aiming to provide a comprehensive picture of QGR and inspire future research.

Large Language Model agents augmented with external skill banks have demonstrated remarkable performance on sequential decision-making tasks. However, existing methods update skills in an undirected, unverified manner and prematurely penalize agents for failing to utilize immature skills. We introduce a structured cognitive representation that decomposes both general and task-specific skills into four dimensions—D-PLAN, D-TRACK, D-EXEC, and D-GUARD—enabling precise failure attribution. Building upon this representation, we propose A²SE, an ability-aligned skill evolution framework that synchronizes skill bank updates with the agent’s learning trajectory. Specifically, A²SE concentrates evolution on the identified bottleneck dimension, safeguards updates via a causal verification mechanism with per-task-type rollback, and employs a competence-aligned skill reward to alleviate premature penalization. Extensive experiments on ALFWorld and WebShop demonstrate that A²SE achieves 11.0% and 7.1% improvements over the strongest baselines, respectively, with ablation studies confirming the necessity of each proposed component.

Fine-tuning large language models (LLMs) for sentiment polarity regression often suffers from severe data distribution imbalance. Financial texts typically concentrate in the neutral range, leaving extreme sentiments underrepresented, yet traditional techniques offer limited remedy since such extremes are inherently rare. We introduce an auxiliary classification task to provide complementary supervision, but this multi-task formulation faces two challenges: (1) sentiment group construction for the classification task, and (2) gradient conflicts between tasks that fixed weights cannot resolve. These challenges are coupled: improper group boundaries yield biased gradients, which amplify task conflicts. We propose the DRNash framework as a unified solution to jointly address both challenges and the underlying data distribution imbalance, including Distributionally Robust Optimization (DRO) and Nash Task Coordinator (NTC). DRO partitions the sentiment space into quantile-based groups and performs worst-case aggregation over group losses, automatically assigning higher weights to underperforming sentiment regions. NTC formulates task weighting as a Nash bargaining game, deriving closed-form weights that achieve scale-invariant balancing. Experiments demonstrate that DRNash achieves average improvements of 14.1% in MSE and 1.92% in accuracy (ACC) across multiple financial datasets, with cross-domain experiments confirming generalizability beyond financial texts.

Robustness is often regarded as a critical future challenge for real-world applications, where stability is essential. However, as models often learn tasks in a similar order, we hypothesize that easier tasks will be easier regardless of how they are presented to the model. Indeed, in this paper, we show that as models approach high performance on a task, robustness to semantics-preserving changes is effectively achieved. Through an empirical analysis of multiple models across diverse datasets and configurations (e.g., paraphrases, different temperatures), we find a strong positive correlation. Moreover, we find that robustness is primarily driven by task-specific competence rather than inherent model-level properties, challenging current approaches that treat robustness as an independent capability. Thus, from a high-level perspective, we may expect that as new tasks saturate, model robustness on these tasks will emerge accordingly. For researchers, this implies that explicit efforts to measure and improve robustness may warrant reduced emphasis, as such robustness is likely to develop alongside performance gains. For practitioners, it acts as a sign that indeed the tasks that the literature deals with are unreliable, but on easier past tasks, the models are reliable and ready for real-world deployment.

Behavioral benchmarking reveals what biases exist in vision-language models but not which internal components are most sensitive to targeted intervention, precluding principled intervention. We argue for mechanism-level evaluation as a necessary complement, demonstrating causal mediation analysis as a diagnostic instrument for gender bias. We decompose gender-cue effects into controlled indirect effects attributable to specific-layer activations and direct effects through all other pathways, producing layer-by-layer mechanistic signatures. Across six models spanning three architectural families (LLaVA-1.5, LLaVA-NeXT, InstructBLIP at 7B/13B) and two 8B-scale architectures, three findings emerge: language-layer activations exhibit the greatest output sensitivity under controlled intervention, with the direct component often carrying the opposite sign; architectural choices redistribute layer-wise sensitivity to activation replacement; and counterfactual scores diverge from surface-level scores, exposing implicit associations. Systematic ablation validates internal consistency. An intervention experiment finds that the average indirect effect (AIE) and downstream intervention effectiveness are only weakly correlated (Pearson r = 0.33), and the layer with the second-largest AIE produces near-zero bias change—indicating that mechanistic diagnosis captures activation-replacement sensitivity but does not, by itself, identify optimal intervention targets. These results show mechanism-level evaluation captures architecture-specific sensitivity patterns that behavioral benchmarks cannot; pairing both should become standard NLP practice. Code: https://github.com/zhaozhipeng1997/CARD-GenderBias.

When strong multimodal models are widely available, progress requires new scientific methodologies beyond benchmark scores—using models as instruments for understanding behavior. We address this by asking: can we use large vision-language models (LVLMs) as experimental instruments for studying their own failure dynamics? Focusing on visual hallucination, we introduce SAFE, a training-free decoding framework that contrasts visually-grounded and vision-ablated generation paths to produce a token-level contrastive grounding score that identifies when the model favors linguistic priors over visual evidence. This signal serves dual roles: as a practical proxy for detecting visually-ungrounded tokens, and as the basis for decoding-time penalties. Our analysis yields three empirical observations: visual dependency decays over generation, hallucinations co-occur in temporal clusters, and early intervention reduces clustering without substantially degrading fluency. On MMHalBench, SAFE substantially outperforms all compared baselines; results elsewhere are more mixed. We argue that designing contrastive probes exemplifies a broader mission: using models as instruments for scientific understanding. Code: https://github.com/zhaozhipeng1997/SAFE_public.

RNA’s conformational flexibility enables diverse biological functions but makes structure-guided design difficult. Although precise control of 3D structure is essential, existing methods often overlook the molecular surface, which critically shapes intermolecular interactions and functional outcomes. To address this, we propose gRFold, a multimodal framework that incorporates molecular surface information as a core design constraint. gRFold employs two geometric encoders: a Frame-Averaged Geometric Surface Encoder (FA-GSE) to learn E(3)-invariant surface features, and an Equivariant Backbone Graph Perceptron (EBGP) to capture backbone topology. The structural representations are integrated via hierarchical cross-attention to guide autoregressive generation; at inference, candidates are re-ranked by agreement between their predicted folds and the target base-pairing pattern. Across three RNA inverse-folding benchmarks covering RNA-protein complexes, monomeric RNAs and independently curated RNA-Puzzles targets, gRFold achieves strong sequence recovery and structural fidelity under both secondary-structure and 3D refolding evaluations, and ablations show that the surface and backbone branches contribute complementary gains. These results highlight the potential of surface-aware geometric modeling for structure-guided RNA design.

Large Language Model (LLM)-augmented Community Notes offer a scalable path for timely, evidence-grounded correction of health misinformation on social platforms. However, they still reset at every post, leaving useful correction experience from prior cases unused. We introduce EvoNote, an agentic framework that enables health Community Notes generation to self-evolve through an evolving experience memory of prior misinformation correction episodes. Its core is fine-grained credit assignment: EvoNote grounds trajectory-level feedback in health-specific note qualities and distills it into action-level memory for claim analysis, evidence acquisition, and note writing. We evaluate EvoNote on MM-HealthCN, a 1.2K-instance multimodal benchmark of user-flagged health posts with human-written Community Notes and crowd-derived helpfulness labels. Under a human-validated hierarchical utility judge, EvoNote-generated notes are preferred over corresponding human-written notes in 89.6% of cases; on a separate set of Needs More Ratings posts without a crowd helpfulness verdict, EvoNote produces helpful notes for 82.0% of cases. It also reduces time to an available candidate correction from multi-hour post-to-note delays to under 2 minutes on average. Analyses link these quality gains to stronger evidence use and reusable correction strategies, positioning self-evolving note generation as a promising paradigm for scalable health misinformation governance.

Retrieval-augmented language models improve knowledge-intensive question answering, but they remain brittle on multi-step questions, where an early retrieval error can steer subsequent reasoning toward the wrong entity or relation. Existing methods can often detect that the current answer or evidence is unreliable, yet they rarely turn this diagnosis into an explicit retrieval correction. We propose Verification-Guided Retrieval Revision (VRR), a framework that uses verification as a control signal for multi-step retrieval. VRR couples a Generator that proposes search and answers, a Judge that verifies answer correctness and evidential support while emitting structured failure diagnoses, and a Retrieval Context Manager (RCM) that converts these diagnoses into retriever-facing updates such as clue selection, query rewriting, and state-aware evidence ranking. To make this loop learnable, we train VRR in three stages: iterative cold-start training to stabilize search, judge-shaped policy optimization to improve long-horizon reasoning, and diagnosis-conditioned RCM training for retrieval adaptation. Across seven open-domain QA benchmarks with a fixed retriever and corpus, training on HotpotQA and transferring to out-of-domain datasets, VRR surpasses the strongest prior baseline by 4.2 EM on average under budget-aligned evaluation, with larger gains on multi-hop benchmarks (+5.4).

Text-rich image understanding requires multimodal large language models (MLLMs) to organize OCR (Optical Character Recognition)-grounded evidence across words, layout, fields, charts, and visual correspondences. Existing evaluations often conflate extraction with reasoning and rarely test whether models follow the required reasoning direction: applying visible rules, abstracting hidden regularities, or recovering missing premises. We introduce OCR-MetaReasoning, a controlled single-image benchmark that treats deduction, induction, and abduction as distinct directions and separates final-answer correctness from reasoning-process compliance. The benchmark contains 1,500 verified samples in a balanced 3×5 taxonomy crossing three reasoning types with five OCR-object categories, along with reference reasoning steps, automatic answer scoring, the Meta-Reasoning Macro Score (MRMS), and the Reasoning Process Compliance Score (RPCS). Experiments with representative closed-source and open-source MLLMs show that OCR-grounded meta-reasoning remains far from saturated: models struggle with visible-rule application and layout-sensitive inference, while process-compliant rationales can accompany incorrect final answers under exact-match evaluation. The code is available at https://github.com/gengxuli/OCR-MetaReasoning.

Multimodal large language models have made rapid progress in video understanding, yet existing benchmarks largely rely on simple prompts and provide limited evidence about whether models can satisfy explicit output constraints. We introduce VCIFBench, a benchmark for evaluating complex instruction following in video understanding. VCIFBench constructs constraint-rich instructions from both benchmark-adapted and directly video-grounded prompts, covering content, format, style, and structure requirements, and evaluates model outputs with a hybrid verification pipeline. The benchmark contains 306 satisfiable test instructions, 540 DPO training instances, and a 100-item diagnostic set for evaluating whether models can recognize instruction conflicts. Experiments on 10 MLLMs show that joint constraint satisfaction remains challenging. Preference optimization improves instruction following for two model families, while Conflict-100 reveals that models usually execute a satisfiable-looking subset instead of detecting global incompatibility.

Parameter-Efficient Fine-Tuning (PEFT) is essential for adapting Large Language Models (LLMs) to multi-task scenarios. A prevailing trend in this field involves complex LoRA variants with multiple adapters or heads, which rely on the premise that architectural isolation of task-specific knowledge is necessary. However, this design often introduces dynamic routing, preventing weight merging and causing significant inference latency. In this work, we present a direct challenge to this paradigm. We first reveal a paradox where a simplified, router-free multi-head model with high inter-head redundancy outperforms complex, diversity-driven baselines. Furthermore, we demonstrate that a unified, single-adapter LoRA with increased rank achieves highly competitive performance, questioning the necessity of multi-component structures. Based on these findings, we propose Align-LoRA, a unified and efficient framework that shifts the focus from architectural isolation to representation alignment. Align-LoRA incorporates an explicit alignment loss to encourage the learning of task-shared representations within a shared latent space. Crucially, our method maintains the standard LoRA architecture, ensuring zero inference latency via weight merging. Theoretical analysis and extensive experiments confirm that Align-LoRA significantly surpasses prevailing approaches, establishing a simpler, more effective, and production-friendly paradigm for multi-task PEFT. The code is available at [Align-LoRA](https://github.com/jinda-liu/Align-LoRA).

Open-web future event prediction requires agents to distill reliable signals from noisy, redundant, and incomplete evidence. Existing retrieval/memory mechanisms directly feed retrieved information to agents or rely on simple memory functions such as storing and reusing prior information for prediction, leaving them insufficient for open-web forecasting. We propose to transform raw web evidence into structured memory before prediction, enabling agents to reason over distilled, question-specific evidence rather than noisy retrieval results. This paper presents ForeDreamer, a self-evolving dual-agent framework for managing memory over open-web evidence. ForeDreamer separates factual memory, a question-specific evidence state for the current forecast, from experiential memory, persistent agent experience accumulated across forecasting episodes. It uses a main agent for search and prediction, and a memory-processing subagent to convert search results into factual memory with dedicated tools. ForeDreamer further evolves experiential memory through two tracks, improving both forecasting decisions and factual-memory construction. Experiments on Prophet Arena and FutureX demonstrate the effectiveness of ForeDreamer.

Recently, several multi-turn dialogue benchmarks have been proposed to evaluate the conversational abilities of large language models (LLMs). As LLMs are increasingly recognized as key technology for advancing intelligent education, owing to their ability to deeply understand instructional contexts and provide personalized guidance, constructing dedicated teacher-student dialogue benchmarks has become particularly important. To this end, we present EduDial, a comprehensive multi-turn teacher-student dialogue dataset. EduDial covers 345 core knowledge points and consists of 34,250 dialogue sessions from interactions between teacher and student agents. Its design is guided by Bloom’s taxonomy of educational objectives and incorporates ten questioning strategies—including situational questioning, zone of proximal development (ZPD) questioning, and metacognitive questioning—better capturing authentic classroom interactions. Furthermore, we design differentiated teaching strategies for students at different cognitive levels, providing more targeted guidance. Building on EduDial, we develop EduDial-LLM 32B via training and propose an 11-dimensional evaluation framework that systematically measures the teaching abilities of LLMs, encompassing overall teaching quality and content quality. Experiments on 17 mainstream LLMs reveal that most models struggle in student-centered teaching scenarios, whereas EduDial-LLM achieves significant gains, consistently outperforming all baselines across all metrics. Code is available at https://github.com/ECNU-RAIL/EduDial-EMNLP2026.

Generating formally verified programs from natural language remains challenging: existing approaches either produce code in a single pass without recourse when verification fails, or rely on open-ended agentic reasoning that is non-deterministic and opaque. We introduce SKILLFORGE, a framework that decomposes formal code synthesis into a library of atomic, reusable skills—each targeting a specific subtask such as specification inference, body synthesis, invariant generation, error diagnosis, or targeted repair—defined by a prompt template, tool binding, and decidable success criterion. A verification-driven harness orchestrates these skills: it submits candidates to the Dafny verifier, diagnoses failures into structured categories, deterministically routes to the appropriate repair skill, and iterates until formal correctness is proved or a budget is exhausted. On a curated benchmark of natural language–Dafny specification pairs, SKILLFORGE substantially outperforms both state-of-the-art agentic approaches (including ReAct-style agents, MCTS-based repair, and RL-guided verification) and traditional iterative baselines, while requiring fewer tokens and lower latency. Ablation studies confirm that every skill contributes measurably, and the harness converges rapidly with the majority of programs verified on the first attempt.

Post-training compression of Large Language Models (LLMs) removes entire architectural components, either deleting them or replacing them with fitted modules. Existing replacement-based methods share two design constraints: full-layer granularity and contiguous selection. We argue that this is overly restrictive: in fact, redundancy in pretrained transformers is not confined to contiguous regions, nor does it evenly distribute between Attention and FeedForward outputs, implying that different strategies best approximate different submodule types and that removable components need not cluster within contiguous depth ranges. Based on this intuition, we introduce SubFit (Submodule-level Fitted residual replacement), which compresses LLMs at the submodule level: Attention and FeedForward submodules are selected non-contiguously, and each receives its own lightweight fitted residual bypass. SubFit operates post-training and requires only calibration data. Across ten LLMs (five base, five instruction-tuned), five sparsity levels from 12.5% to 37.5%, and four replacement-based baselines, SubFit achieves the best aggregate perplexity–accuracy trade-off across the evaluated sparsity levels, with larger gains under aggressive compression. At 25% sparsity, it retains 84.6% of dense downstream accuracy and incurs 2.42× perplexity degradation, against 81.6% and 4.34× for the strongest baselines, while delivering measurable inference speedup and KV-cache savings. Code is available at https://github.com/eliacunegatti/SubFit.

Self-generated counterfactual explanations (SCEs) are minimally modified inputs minimality generated by large language models (LLMs) that flip their own predictions validity, offering a causally grounded approach to unraveling black-box LLM behavior. Yet extending them beyond English remains challenging: existing methods struggle to produce valid SCEs in non-dominant languages, and a persistent trade-off between validity and minimality undermines explanation quality. We introduce Macro, a preference alignment framework that applies Direct Preference Optimization (DPO) to multilingual SCE generation, using a composite scoring function to construct preference pairs that effectively translate the trade-off into measurable preference signals. Experiments across four LLMs and seven typologically diverse languages show that Macro improves validity by 12.55% on average over the chain-of-thought baseline without degrading minimality, while avoiding the severe minimality violations of the translation-based baseline. Compared to supervised fine-tuning, Macro achieves superior performance on both metrics, confirming that explicit preference optimization is essential for balancing this trade-off. Further analyses reveal that Macro increases cross-lingual perturbation alignment and mitigates common generation errors. Our results highlight preference optimization as a promising direction for enhancing multilingual model explanations.

Large language models (LLMs) can generate functionally correct code, but their ability to optimize programs is underexplored. We introduce OEISPerf, a code-to-code benchmark for algorithmic program optimization built from alternative implementations of integer sequences in the On-Line Encyclopedia of Integer Sequences. For each sequence we mine multiple mathematical formulations, synthesize and validate Python implementations, and pair a substantially slower program with the fastest validated reference found by our construction pipeline, producing instances that often require substantial structural transformations and can yield substantial speedups. Evaluating 15 open-source and closed-source LLMs shows that even the strongest evaluated setting, Gemini-3-Flash with its thinking mode enabled, matches the reference efficiency on only 67.9% of test instances. Across LLMs, we observe both incorrect “optimizations” that break correctness and superficial refactorings that preserve suboptimal complexity. Finally, we fine-tune Qwen2.5-Coder-3B-Instruct via reinforcement learning, markedly improving reference attainment on OEISPerf while slightly improving performance on standard code-generation benchmarks.

Training long-horizon LLM agents with reinforcement learning is challenging because sparse outcome rewards reveal whether a task succeeds, but not which intermediate actions caused the outcome or how they should be corrected. Recent methods alleviate this issue by generating rewards or textual hints from turn-level action-output signals, or by using feedback-conditioned self-distillation. However, generating feedback at every turn is inefficient when many intermediate turns are already successful or neutral, and applying feedback at a fixed or misaligned turn often fails to supervise the actions that contributed to the failure. To bridge this gap, we propose HINT-SD, a targeted self-distillation framework that uses full-trajectory hindsight to select failure-relevant actions and applies feedback-conditioned distillation only to targeted action spans. Experiments on BFCL v3 and AppWorld show that our method outperforms the dense per-turn feedback baseline by up to 13.60 percentage points on average while achieving a 2.26× reduction in time per training step, suggesting that selecting where to distill is key to effective and efficient long-horizon agent training.

Event extraction aims to identify triggers in text, classify them into predefined event types, and recognize associated arguments such as participants and locations. Large language model (LLM)-based generative approaches provide a unified paradigm for joint trigger identification, classification, and argument extraction; however, their computational cost necessitates knowledge distillation (KD) into smaller student models. Existing KD methods operate at the token level, treating each output position independently and thus failing to concentrate the distillation signal on event-relevant spans or to capture the relational structure among event components. We propose EventKD, a two-level distillation framework tailored for event extraction. At the token level, a distributional divergence loss incorporates event-aware span importance weighting to prioritize tokens corresponding to triggers, arguments, and type labels. At the span level, a Relational-Aware (RA) loss preserves pairwise geometric relationships among event-aware spans across Transformer layers. Experiments on four benchmarks (ACE05, MAVEN, GENEVA, and RAMS) demonstrate that EventKD consistently outperforms existing distillation methods on both trigger identification and argument extraction.

Knowledge graph (KG) reasoning infers new knowledge from existing facts and is widely applied in question answering, recommendation, and decision support. With the rapid development of large language models (LLMs), LLM-based KG reasoning frameworks have become increasingly popular by leveraging retrieved KG information. However, hallucinations in LLMs remain a critical issue. Even when relevant KG knowledge is incorporated, models may still generate incorrect outputs, leading to misinformation and unreliable decisions. Existing hallucination detection methods either focus on LLM internal states or verify consistency with retrieved contexts, but both overlook the structural information in KGs, resulting in suboptimal performance. To address this gap, we propose LUCID, the first halLUcination deteCtIon method for LLM-based knowleDge graph reasoning frameworks. LUCID extracts node and edge features from attention scores and semantic similarities, and integrates them with KG structure using a graph neural network (GNN). Moreover, we also construct manually annotated benchmark datasets for evaluation. Experiments on nine datasets show that LUCID achieves state-of-the-art performance compared to 15 baselines. The code and data are available at: https://github.com/BUPT-GAMMA/LUCID.

Multi-agent large language model (LLM) workflows have emerged as a powerful paradigm for solving complex, open-ended tasks through collaborative reasoning among specialized LLM agents, but they incur substantial operating costs due to repeated LLM invocations and long-horizon context accumulation. Existing cascade routing methods make one-shot, query-level decisions and cannot adapt to the dynamic, state-dependent nature of multi-step workflows, in which the right LLM at each step depends on evolving task progress, remaining task difficulty, and cost-efficiency requirements. We present ProgRouter, an online progress-guided routing framework that adaptively selects LLM agents across workflow steps to preserve task-solving quality while adhering to time and cost budgets. ProgRouter introduces a multi-view task progress scorer that combines coarse workflow outcome regimes with fine-grained signals on subtask completion, progress trends, and workflow state quality. Then, a dual-path task progress predictor and an adaptive meta-gating mechanism estimate the progress gain for each candidate routed LLM. ProgRouter makes online step-wise routing decisions that balance progress gain, task time budgets, and long-term operating cost efficiency. Experiments on HumanEval Plus, MBPP, MATH-500, and ASQA, spanning agentic code generation, mathematical reasoning, and retrieval-augmented long-form question answering, demonstrate that ProgRouter reduces the operating cost relative to key baselines while maintaining strong task-solving performance.

Linters are widely used in code maintenance and real-time IDE editing to surface lint alerts related to performance, style, error-prone, and other patterns in source files. However, existing code benchmarks do not explicitly evaluate linter-guided file refactoring, where models must reduce lint alerts in source files with correct initial functionality while satisfying functionality preservation. We introduce LintEval, a multilingual benchmark for linter-guided file refactoring across C++, Python, and Java, comprising 13,061 programs with correct initial functionality and 171,662 lint alerts extracted using language-specific linters. LintEval jointly evaluates functionality preservation, lint alert reduction, new alert avoidance, and edit minimality. Experiments reveal a persistent trade-off: closed-source commercial models reach up to 90.4% Reduction@P but achieve only 67.1%–68.3% Pass@1, while several open-source small models show low or negative Reduction@P. A controlled fine-tuning study further shows that supervision format matters: refactoring-oriented SFT raises Qwen2.5-Coder-7B Reduction@P from 24.7% to 76.0%. LintEval enables controlled evaluation of models that improve source files under functionality preservation.

Weight-only post-training quantization (PTQ) for large language models (LLMs) typically ends with the projection of floating-point weights onto a uniform integer lattice, introducing structured output error due to non-zero activation means. Existing bias correction methods compensate for this shift by adding bias terms, which is incompatible with modern bias-less architectures such as LLaMA and Mistral and only addresses the error indirectly. We propose **Complementary Lattice Correction (CLC)**, a lightweight post-quantization method that corrects projection error directly on the integer lattice without modifying the model architecture. CLC performs a budget-constrained greedy descent that selectively flips quantized weights to adjacent lattice levels. We further show that CLC provably reduces the first-moment error and admits a worst-case bound on the layer-wise mean-squared error change, which certifies non-increase under a verifiable budget condition. Experiments across multiple models, quantizers, and bit widths show that CLC consistently improves perplexity and downstream accuracy with negligible overhead.

With the release of OpenAI’s o1 model, reasoning models that adopt slow-thinking strategies have become increasingly common. Their outputs often contain complex reasoning, intermediate steps, and self-reflection, making existing evaluation methods and reward models inadequate. In particular, they struggle to judge answer equivalence and to reliably extract final answers from long, complex responses. To address this challenge, we propose xVerify, an efficient answer verifier for evaluating reasoning models. xVerify shows strong equivalence judgment capabilities, enabling accurate comparison between model outputs and reference answers across diverse question types. To train and evaluate xVerify, we construct the VAR dataset, consisting of multi-round annotated QA pairs generated by multiple LLMs across challenging reasoning tasks. Experimental results on both test and generalization sets show that all xVerify variants achieve over 95% F1 score and accuracy. Notably, the smallest model, xVerify-0.5B-I, outperforms all evaluation methods except GPT-4o, while xVerify-3B-Ib surpasses GPT-4o in overall performance. In addition, reinforcement learning experiments using xVerify as the reward model yield an 18.4% improvement for Qwen2.5-7B compared with direct generation, exceeding the gains achieved with Math Verify as the reward. These results demonstrate the effectiveness and generalizability of xVerify. All xVerify resources are available on [GitHub](https://github.com/IAAR-Shanghai/xVerify).

Jailbreak attacks threaten Large Language Models (LLMs) by bypassing their safety mechanisms. A truly advanced jailbreak is defined not only by its effectiveness but, more critically, by its stealthiness. However, existing methods face a fundamental trade-off between semantic stealth (hiding malicious intent) and linguistic stealth (appearing natural), leaving them vulnerable to detection. We propose StegoAttack, a framework that leverages steganography. The core insight is to embed a harmful query within a benign, semantically coherent paragraph. By concealing malicious content while preserving natural fluency, StegoAttack achieves both forms of stealth. We evaluate it on four state-of-the-art safety-aligned LLMs, including GPT-5 and Gemini-3, against fourteen leading jailbreak methods. StegoAttack achieves an average attack success rate (ASR) of 95.50%, outperforming all baselines across the four models. Under external detectors, its ASR decreases by less than 27.00% while retaining natural-language characteristics. These results show that steganography can decouple semantic and linguistic stealth, posing a highly concealed and effective security threat. The code is available at https://github.com/GenggengSvan/StegoAttack.

Large Vision-Language Models (LVLMs) remain prone to hallucinations, producing responses that are irrelevant or inconsistent with the multimodal input. Existing mitigation methods mainly rely on external supervision, output calibration, or attention regulation, leaving the internal representation dynamics of autoregressive generation underexplored. We identify an inference-time failure mode in which cross-modal representations degrade across decoder layers and drift across generation steps, destabilizing token prediction and increasing hallucination risk. We propose Dynamic Alignment Compensation (DAC), a training-free inference-time method that detects representation divergence and selectively applies lightweight residual compensation. DAC combines Layer-wise Semantic Compensation to mitigate inter-layer degradation with Sequential Semantic Correction to constrain temporal drift. Experiments on nine hallucination-focused and general-purpose multimodal benchmarks across multiple LVLM backbones show that DAC consistently reduces hallucinations while maintaining strong overall performance.

Previous works identify sensitivity to option order as a key issue in MC-VQA evaluation and propose protocols to mitigate this effect. We show that such mitigation is insufficient to ensure the validity of MC-VQA as a reliable benchmark for MLLMs: performance remains highly sensitive to semantically neutral prompt format choices that are not controlled by current benchmarks. In a large-scale study spanning seven MLLMs and five MC-VQA datasets, we find frequent rank reversals even under order-invariant evaluation. These reversals arise when we systematically vary option ID sets, delimiters, and separators, yielding 48 semantically equivalent prompt formats. Mechanistic analyses trace this instability to low-level language modeling effects: tokenizer-induced fusion or removal of option ID tokens introduces corrupted option ID tokens into the input sequence, while the choice of option ID sets directly affects the reliability of attention patterns for option selection. Accordingly, MC-VQA rankings correlate weakly with open-ended evaluation, indicating that MC-VQA reflects option selection dynamics in addition to multimodal reasoning. These findings identify prompt formatting as a major, previously under-controlled confounder in MC-VQA benchmarking and motivate evaluation protocols that explicitly control prompt format sensitivity.

Visual instruction tuning is crucial for advancing the vision-language alignment and instruction-following capabilities of Vision-Language Models (VLMs). However, identifying optimal subsets under a fixed ratio constraint from rapidly expanding datasets remains a significant bottleneck. While existing methods largely depend on distribution diversity or heuristic filtering, they often overlook the internal coherence within individual samples. To bridge this gap, we propose Data Intrinsic Consistency (DIC), a self-scoring metric designed to quantify the sample-level inter-component consistency. DIC consists of two modules: Visual Information Consistency (VIC), evaluating the alignment between visual content and instructions, and Response Information Consistency (RIC), assessing response coherence relative to the instruction. Building upon DIC, we introduce Data Intrinsic Consistency Selection (DICS), an adaptive data selection method that optimizes the trade-off between high intra-sample consistency and global distributional diversity under varying data budgets. Extensive experiments demonstrate that DICS consistently outperforms state-of-the-art methods across diverse dataset scales and model architectures, surpassing full-dataset fine-tuning while using only 25% of the LLaVA-1.5-665K data.

Authorship verification asks whether two given texts were written by the same author. Since supervised verification methods often generalize poorly across domains, the unsupervised Impostors Method, inspired by police lineups, reduces verification to in-domain authorship attribution. Using one text as a query, it hides the other among a sample of hard negatives and measures how reliably a style similarity function retrieves it at the top rank. Although widely applied in practice, the Impostors Method still suffers when no in-domain impostor documents are available. Resorting instead to sampling from open-domain collections provides much less control over confounding stylistic variables such as genre or register. Moreover, its statistical foundations remain largely unexplored compared with those of police lineups. In this paper, we recast the Impostors Method as a statistical hypothesis test, enabling a principled assessment of evidential strength without training a decision threshold. We further investigate large language models (LLMs) as a controllable source for generating in-domain impostor documents.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language interaction, yet their safety alignment remains vulnerable to jailbreak attacks. A key challenge is that safety behavior learned in the textual space does not reliably transfer to fused cross-modal representations, leaving multimodal inputs exploitable through latent semantic cues. We propose __Text-Anchored Semantic Perturbation Attack__ (TA-SPA), a black-box jailbreak framework that optimizes transferable perturbations in a text-anchored semantic space. TA-SPA integrates __Text-Anchored Semantic Factorization__ (TASF), which disentangles cross-modal semantic factors from modality-specific residuals, with __Semantic-Preserving Augmentation__ (SPA), which diversifies harmful target anchors while preserving semantic consistency. Extensive experiments show that TA-SPA consistently outperforms prior jailbreak attacks, transfers to commercial MLLMs, and remains effective under representative defenses, underscoring the need for representation-level safety alignment beyond input-level filtering.

Text detoxification is increasingly used in content moderation and dialogue systems, where systems must produce safe rewrites while revealing what was changed and why. Existing approaches typically trade off interpretability and end-to-end optimization. Monolithic rewriting methods treat detoxification as a single opaque generation step, whereas edit-based pipelines make edits explicit but are trained in separate stages, preventing joint optimization. We propose STEER, a framework that models detoxification as a structured edit chain formulation and optimizes the entire editing process end-to-end with Group Relative Policy Optimization (GRPO). To make this formulation trainable, we construct synthetic edit-chain supervision using a multi-agent pipeline, perform supervised fine-tuning (SFT) to teach the model the structured editing format, and then apply GRPO with a veto-aligned multiplicative reward and granular format scoring. Experiments on ParaDetox and two out-of-domain benchmarks show that STEER achieves the highest aggregate J score among the compared systems on ParaDetox and remains competitive under out-of-domain evaluation.

Parameter-efficient fine-tuning (PEFT) reduces trainable parameters, yet head-sparse PEFT still requires deciding which attention heads should receive updates. We introduce SIGHT, **S**parse **I**dentification of **G**aze-aligned **H**eads for **T**uning, a gaze-guided head selector for head-sparse PEFT. Using eye-tracking data during reading comprehension tests, SIGHT ranks heads by correlating word-level gaze metrics with the attention heads of large language models. Specifically, the comprehension question appears either before or after the passage so that the readers either read the passage with a clear goal or for general understanding. The eye-tracking data from these two conditions are both correlated with attention heads and are fused by an offline-calibrated layer-wise gating module. When applied to Qwen3-based models, updating only the top 10% SIGHT-selected heads generally outperforms other head selectors and all-head PEFT references on reasoning and syntax benchmarks. These results suggest that gaze information during human reading can move beyond post hoc interpretation and serve as an external criterion for allocating adaptation capacity.

Large Language Models (LLMs) can be trained to perform chain-of-thoughts reasoning in order to improve the reliability of their responses. In this work, we investigate how explicit reasoning can be leveraged for Machine Translation (MT). We introduce a novel fragment-based reasoning framework in which the model first extracts parallel source-target fragments from retrieved similar examples, and uses these fragments as intermediate reasoning traces to produce the final translation. To train our model, we distill silver fragments and drafts from a larger teacher model. Our experiments with the Qwen3 model family, over 6 languages, including up to 5 domains per language, demonstrates that fragment-based MT significantly outperforms alternative methods like standard k-shot, or drafting.

Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.

LLM agents can generate paper reproduction code, yet often produce scientifically unfaithful implementations. We define this failure mode as semantic drift, where generated code silently diverges from the paper’s specifications. We introduce SemanticAlign-Bench(SA-Bench), a diagnostic benchmark covering 30 papers from ICLR, ICML and NeurIPS 2025. For each paper, we decompose its specifications into atomic and verifiable implementation claims, which we call Semantic Alignment Units (SAUs) and evaluate repositories along four diagnostic dimensions spanning numerical, methodological, protocol and ordering drift. In total, we construct 1,491 SAUs across five ML domains and evaluate 12 generator configurations (4 models 3 scaffolds). Even the strongest configuration (Claude + PaperCoder) achieves a mean SAU score of only 0.301 out of 1.0, with an overall mean of 0.221 across 360 evaluations. A failure taxonomy reveals that agents attempt most requirements but implement them incorrectly, with implementation mismatch and stubs accounting for the majority of zero-scored claims. Our analysis further indicates that scaffolds optimized for executability provide limited leverage for scientific reproduction; narrowing the gap requires scaffolds that prioritize semantic specification verification. The benchmark, annotations and evaluation pipeline are publicly available.

Code:https://anonymous.4open.science/r/SemanticAlign-Bench-0676

Data:https://anonymous-hf.up.railway.app/a/rrgn430zpfui/

Current plan-based reasoning methods improve large language models (LLMs) by inserting a planning stage before execution, giving rise to the **question → plan → cot** paradigm. While effective, a closer examination reveals an inherent paradigm-level gap: both the planning and its execution stages decide *how to solve* a problem, while the prior question of *what to solve*; recognizing the problem type, the applicable tools, and the foreseeable pitfalls; remains entirely implicit. To bridge this gap, we propose **PPC** (**P**replan-**P**lan-**C**oT), a framework that introduces an explicit problem-understanding stage, the *preplan*, yielding a new **question → preplan → plan → cot** paradigm. Realizing this paradigm requires safeguarding the conceptual integrity of preplan at both ends. Specifically, we design a three-stage synthesis pipeline with a *spoiler-score detector* that filters out leakage and spoiler failures to build clean preplan supervision, and a composite GRPO reward enforces that the generated plan genuinely follows from the preplan. Experiments across four backbones and five mathematical reasoning benchmarks show that PPC achieves the best results on 39 of 40 metrics, improving *maj@16* and *pass@16* by +2.23 and +3.06 over the strongest baseline without introducing additional inference token overhead.

Retrieval-Augmented Generation (RAG) grounds LLM responses in external evidence but treats the model as a passive consumer of search results, with no view of how the corpus is organized or what it has not yet seen. We present Corpus2Skill, a system-level retrieval architecture for bounded, structurally coherent corpora such as enterprise knowledge bases: an offline compiler distills the corpus into a hierarchical skill directory, and at serve time an LLM agent navigates it, drilling from a bird’s-eye view through progressively finer summaries down to documents and backtracking when a branch is unproductive. On an enterprise customer-support benchmark, Corpus2Skill improves both answer quality and grounding over single-shot dense, hybrid, hierarchical-retrieval, and agentic RAG baselines at a moderate cost tradeoff, and the lead persists under encoder-matched controls and paired significance tests. An eleven-dataset study shows that corpus navigation is not a universal replacement for retrieval: it significantly wins on five datasets, ties on three, and loses on three. It helps on single-domain corpora with a recoverable topical taxonomy, but flat retrieval remains preferable on open-domain factoid pools or homogeneous-tabular corpora that defeat top-level clustering. We characterize this scope distinction as a design guideline for knowledge-grounded systems. Code is available at https://github.com/dukesun99/Corpus2Skill.

Recently, the Muon optimizer has demonstrated strong efficiency in large model training through matrix orthogonalization. At each optimization step, Muon orthogonalizes the raw update matrix via Newton-Schulz (NS) iterations before applying it to model parameters. However, orthogonalization requires repeated matrix multiplications whose computational cost scales with O(n³) for an n× n matrix, making the overhead prohibitive at large model scales. In this paper, we observe that the geometric angle derived from cosine similarity between consecutive orthogonal update matrices remains highly stable during training, a phenomenon that we partially explain theoretically and support empirically. Based on this angular stability, we propose LazyMuon, which skips computationally expensive NS iterations through intermittent linear interpolation and employs a residual correction mechanism to mitigate accumulated approximation errors. This design ensures fast iteration while preserving Muon’s orthogonal manifold constraints. Empirical evaluations on language and vision tasks demonstrate that LazyMuon achieves speedups of up to 1.34× and 1.31× over Muon, respectively, with negligible degradation in convergence.

Geometry problem solving poses distinct challenges in artificial intelligence. Existing approaches typically fall into two paradigms: symbolic methods, which exhibit limited adaptability, and neural methods, which are prone to hallucinations. Recent neuro-symbolic hybrids predominantly rely on a unidirectional pipeline where neural outputs are fed into solvers without feedback, making system brittle to early-stage errors. To break this unidirectional bottleneck, we propose BiNSGPS, a framework that establishes Bidirectional Neuro-Symbolic Interaction (BiNS) between a MLLM Adviser and a Symbolic Solver. MLLM Adviser actively incorporates feedback from the symbolic solver to dynamically rectify inconsistent formal representations or propose auxiliary hypotheses, resolving symbolic conflicts and facilitating complex deductions. Experiments show that BiNSGPS achieves state-of-the-art performance, reaching 90.5% completion accuracy on Geometry3K and 90.1% on PGPS9K, while maintaining 96% step-wise logical coherence to balance capability with mathematical rigor.

Paper-to-code reproduction asks scientific AI agents to turn research papers into executable repositories that preserve the paper’s method, protocol and artifacts. This is difficult because the specification is split: explicit paper content such as algorithms, metrics and artifacts is often lost across long agent trajectories, while implicit details such as framework defaults and conventions inherited from related work are absent from the paper. We introduce ReproAgent a four-stage Prepare–Plan–Generate–Repair pipeline built around a persistent implementation contract with two channels: an implementation-requirement channel that turns paper snippets into code obligations, and a reference-evidence channel that retrieves content and structure evidence from related repositories. Both are bound to work packages, projected into file-level contracts, and consumed across generation and repair. On PaperBench Code-Dev, ReproAgent reaches the highest mean score among same-backbone scaffolds under both Claude-Sonnet-4.5 and Gemini-3-Flash. End-to-end channel ablations and per-paper cases support the contribution of both channels.Code and experimental artifacts are publicly available at https://anonymous.4open.science/r/ReproAgent-E760.

Contemporary large language models (LLMs) exhibit multilingual capabilities but remain predominantly English-centric, resulting in a persistent performance gap across languages. While some approaches to multilingual training are grounded in the implicit pivot language hypothesis to improve interpretability, it remains underexplored how the languages represented in each layer are affected during training, particularly in practical, cost-efficient instruction-tuning scenarios. In this study, we investigate how instruction tuning with Low-Rank Adaptation (LoRA) in a target language affects the layer-wise language representation of instruction-tuned LLMs, with the aim of assessing the relevance of the implicit pivot language hypothesis as a theoretical grounding for multilingual finetuning. We find that target-language instruction tuning shifts models’ internal language representation in a non-uniform manner across layers and sublayers: middle layers shift further toward English, while early and final layers primarily serve the mapping between the target language and the dominant intermediate language. Our findings empirically support the implicit pivot language hypothesis in the context of instruction tuning, demonstrating that layer-wise parameter variation broadly corresponds to the hypothesized functional roles of layers in multilingual processing. Our code is available at: https://github.com/hyemijo/layer_interpretability

Visible tests are a common gate for LLM-generated code, but passing them does not certify specification correctness. We study a deployment-like monitoring problem: after code has passed public tests, can a weaker LLM verifier identify the residual hidden bugs? We introduce Code Monitor Red Teaming, a monitor-red-teaming protocol that fixes a public-check information boundary while varying generator pressure, verifier scaffolding, and weak-to-strong capability. We instantiate it as CodeMonitorBench, spanning function-level, data-science, and workflow code. Across 71,000 generated candidates, 43,677 pass public tests and 23,081 of those fail hidden tests. Weak verifiers improve with scaffolding and model family, but still miss most hidden bugs at 5% false-positive rate. As a robustness stress test, adversarial public-test-overfit pressure lowers verifier AUROC and raises low-FPR miss rates in most cells. A GLM-5.1 verifier recovers part of the gap under the same evidence boundary; an inferability audit shows that remaining misses mix verifier failures with M1 evidence limits.

As Multimodal Large Language Models (MLLMs) evolve into sophisticated interactive assistants, their reliability depends not only on following instructions but also on validating them. We define Proactive Critique as the model’s autonomous ability to identify, analyze and fix faulty user inputs without extra prompts. However, evaluations mainly test models under ideal circumstances or simple refusal behaviors, largely ignoring active error processing. To fill this gap, we propose MMPCBench, a comprehensive framework for evaluating MLLMs’ proactive critique competence. It features a fine-grained taxonomy of 4 primary error types spanning 12 subcategories, ranging from cross-modal contradictions to missing visual premises. We adopt a hierarchical evaluation protocol to measure models’ error detection, diagnosis and resolution performance, and apply alignment-aware metrics to assess the coherence between internal reasoning and final responses. Tests on 14 mainstream MLLMs show obvious weaknesses in proactive critique, especially in dealing with subtle visual anomalies. Notably, we identify a pervasive "consistency gap": reasoning models can often correctly identify and analyze errors during internal reasoning yet suppress these valid insights in final outputs to prioritize response compliance.

Synthetic data generation has become a cornerstone for advancing large language models. However, the lack of the quantitative analysis for error tolerance became a critical bottleneck. Consequently, current filtering strategies fluctuate between two extremes: they are either overly aggressive, risking the exclusion of potentially valuable samples, or overly permissive, failing to eliminate erroneous samples effectively. To bridge this gap, this paper introduces **A**tomic **T**ree **O**peration **M**odeling (**ATOM**), a framework that decomposes data into functional units (f(x)→ y). ATOM distinguishes *benign Operand* x *perturbations* from *fatal Operator* f *perturbations*. The former are needlessly discarded by aggressive filtering, while the latter slip through permissive filtering. Our experiments reveal a **double dissociation**: models are robust to operand perturbations but collapse under operator perturbations. By prioritizing operator over aggressive operand precision, our ATOM-synthesized data outperforms rigorous baselines (e.g., +3.1% gain over LIMA), suggesting that operator diversity matters more than operand precision. Our code is available at [GitHub](https://github.com/Lut-hub/ATOM).

Vision-language models (VLMs) achieve strong multimodal performance but remain expensive to deploy. Knowledge distillation (KD) is an effective approach for compressing these models and transferring their capabilities to more efficient student models. While KD for large language models (LLMs) is already well surveyed, no existing work provides systematic coverage of KD for VLMs, where the addition of a visual modality introduces challenges that text-only distillation does not face. This paper presents the first dedicated survey of knowledge distillation for vision-language models. We propose a unified taxonomy that organizes methods along three orthogonal views: knowledge type, teacher accessibility, and structural compatibility, and use it to organize methods across major VLM distillation settings. We also survey the associated datasets and evaluation protocols, identify open challenges and outline directions for future research. We further maintain a GitHub repository at https://github.com/HoangTran223/Awesome-Knowledge-Distillation-of-VLMs to compile all surveyed papers, which we will actively update.

Chain of Thought (CoT) reasoning enhances logical performance by decomposing complex tasks, yet its multimodal extension faces a trade-off. The existing thinking with images paradigm is limited by the modality gap between vision and language, which hinders reliable extraction of reasoning relevant information from high dimensional visual data. Recent latent space reasoning method provides stronger multimodal representations, but it often lacks the ability to refocus on visual inputs and suffers from limited interpretability. To address these issues, we proposeLatent Refocusing (LaRe), a novel multimodal reasoning paradigm that combines visual refocusing with rich latent representations, enabling iterative reasoning within the latent space. We further design a semantic augmentation training strategy that ensures the semantic structure of the latent space through visual reconstruction objective. Experimental evaluations demonstrate that LaRe improves average accuracy by 7.6% compared to existing baselines while reducing the number of tokens required for inference by 59.7%. When scaled to a 8B-parameter Vision-Language Model backbone, LaRe achieves performance comparable to state-of-the-art methods, demonstrating the efficacy of our proposed latent refocusing paradigm for multimodal reasoning.

Multimodal Large Language Models (MLLMs) encode images into dense visual token sequences, creating substantial attention and KV-cache overhead. Existing visual token compression methods reduce this cost by pruning low-saliency tokens in shallow layers, but this practice couples temporary computational deactivation with irreversible information deletion. We show that this coupling breaks an intrinsic inference pattern in MLLMs. Through layer-wise attention analysis, we identify Visual Backtracking: visual pivots remain spatially stable, yet in middle layers they actively retrieve context from tokens that appeared unimportant earlier. Thus, early low-saliency tokens can be inactive for computation while still being useful for later reasoning. Motivated by this finding, we propose ReTrace, a backtracking-aware compression framework that introduces a dormant token state. ReTrace caches non-pivot tokens outside expensive layer-wise computation, restores them during the backtracking window, and prunes only after their information has been absorbed. Experiments on Qwen2.5-VL and InternVL3 demonstrate nearly 90% reductions in FLOPs and KV cache, a 2.83 prefilling speedup, and strong performance retention across complex vision-language benchmarks.

Audio-conditioned language models often underuse acoustic cues such as prosody, emotion, and non-speech sounds, raising the question of whether ASR-supervised frontends discard this information before it reaches the LM. We test whether the frontend is responsible by comparing Whisper-Tiny and Whisper-Small with EnCodec, DAC-VAE, and WavTokenizer in a shared Qwen3.5-4B audio-LM pipeline on ASR, emotion recognition, and sound captioning. Encoder replacement alone does not resolve this underuse: Whisper variants remain strongest overall, including on emotion and environmental sound captioning. To localize the failure, we trace task-relevant information through the encoder, projector, LM layers, and LM head. Linear probes and geometric analyses show that discriminative acoustic structure remains recoverable at the final LM layer, even when MCQA accuracy trails probe accuracy by up to 83 points. Because the answer format and decoding procedure are controlled, this task-dependent gap points to content-specific readout failure rather than generic format bias. LogitLens analyses and a targeted LM head intervention support the conclusion that acoustic underuse is not explained solely by encoder-side information loss and that readout alignment can be a dominant bottleneck.

Large language models (LLMs) achieve strong fluency in open-domain question answering, yet often hallucinate in specialized medical QA and lack reliable uncertainty quantification, posing risks in safety-critical applications. We propose Uncertainty-guided Retrieval-augmented Generation (URGen), a training-free framework for guideline-grounded medical multiple-choice QA, where authoritative clinical documents (e.g., consensus statements, treatment protocols) are available and answer correctness depends on precise evidence retrieval. URGen is a four-stage pipeline that combines adaptive retrieval depth with uncertainty calibration: (1) multi-temperature knowledge retrieval to rank candidate keypoints, (2) zero-shot adaptive retrieval depth (ZARD) that expands context when rankings are ambiguous, and contracts it when rankings are consistent, (3) multi-temperature answer generation with aggregation over candidates, and (4) entropy-based uncertainty estimation over the answer distribution to enable abstention on high-uncertainty queries. Extensive experiments on two expert-annotated ophthalmology benchmarks, Thyroid Eye Disease and Graves’ Disease, show that URGen improves accuracy by 9.29% and 6.19% over a naive RAG-based Meditron-7B baseline, while achieving 3.01% and 4.57% expected calibration error (ECE). URGen is LLM-agnostic, consistently improving diverse backbones (, LLaMA-3.1 and Mistral) and providing interpretable uncertainty indicators for safety-critical deployment.

Operations research supports decision-making in domains such as energy, economics, and healthcare. Solving operations research problems typically begins with optimization modeling, which translates a natural-language problem description into executable solver code. LLMs offer a promising way to automate this process, but they remain prone to errors. In practice, these errors can be divided into two categories: syntactic errors refer to solver code that fails to run successfully or is judged infeasible by the solver; semantic errors refer to solver code that successfully returns an objective value but violates the intent of the original problem. Since semantic errors do not trigger runtime failures, they are difficult to detect and rectify. To address this problem, we introduce SemOPT, a semantic-guided framework for correcting LLM-based optimization models. SemOPT combines a semantic reward model that distinguishes faithful math models from plausible but incorrect ones with an adaptive correction system that applies hierarchical reward-guided search over the modeling space. Experiments on seven optimization modeling benchmarks show that SemOPT establishes a new state of the art and achieves an average 7.6% accuracy improvement over the strongest baseline on complex datasets.

Large Vision-Language Models (LVLMs) have demonstrated remarkable multimodal understanding capabilities, yet their inherent security vulnerabilities remain a pressing concern. Prior studies have indicated that enhancing chain-of-thought reasoning can inadvertently facilitate jailbreaks; however, existing attempts often rely on image masking, which only modestly extends reasoning depth and performs poorly against advanced commercial closed-source models. To bridge this gap, we propose an effective jailbreak method, Text–Image Segmentation and Reassembly (TISR). TISR disperses harmful semantic fragments across the visual modality while employing text-based role-playing to guide their reassembly. This strategic organization conceals surface-level toxicity while preserving malicious intent, thereby extending LVLM reasoning chains and enabling successful jailbreaks. Our experiments show that LVLMs, leveraging their visual understanding and reasoning abilities, can reconstruct complete harmful semantics while failing to uphold safety constraints. Extensive evaluations across massive LVLMs demonstrate that the proposed TISR achieves state-of-the-art attack performance, surpassing prior multi-modal jailbreak methods. Specifically, it attains average Attack Success Rates (ASR) exceeding 90% across three benchmark datasets on four leading commercial LVLMs.

Constructive feedback is crucial for creative writers to refine their storytelling abilities. Since receiving feedback from human experts is often costly and time-intensive, large language models (LLMs) offer a scalable and efficient alternative as automatic writing assistants. Despite their potential, research indicates that LLM-generated feedback is often generic, lacks actionability, and fails to identify which writing issue is most critical. To address these limitations, we present a reinforcement learning approach that steers LLMs to generate constructive feedback without the need for ground-truth feedback. We train our model using group relative policy optimization (GRPO) with a novel multi-component reward function aiming at constructiveness: it prioritizes feedback that is uniquely tailored to the story, helps to improve story quality, and addresses the most critical writing issue. In automatic and human evaluation across three story corpora, our approach outperforms state-of-the-art LLMs (including Gemini) and competitive baselines. We find that providing actionable suggestions is the main driver of feedback constructiveness.

Large Language Models (LLMs) integrated into complex ecosystems (e.g., code agents) face critical safety threats from Indirect Prompt Injection (IPI) attacks, where malicious instructions are embedded in retrieved data. Existing defenses face an intractable safety-utility trade-off: sanitization-based guardrails incur high latency, while detection-based guardrails suffer from severe over-refusal. We first demonstrate, both theoretically and empirically, that LLMs can intrinsically separate instructions from data. Inspired by this insight, we propose AEGIS (*A*daptive *E*nsemble *G*uard for *I*njection *S*hielding), a training-free defense that extracts instruction-sensitive projectors to identify malicious instructions and leverages a Unified Multi-Layer Consensus that aggregates topologically distinct signals across the network depth. Evaluations on eight IPI attacks and ten baselines show that AEGIS achieves state-of-the-art detection accuracy while reducing the false positive rate on benign tasks to below 1%. Code is available at [Github](https://github.com/xaddwell/AEGIS).

Understanding persuasion is critical for the safety and reliability of multi-agent systems built on large language models (LLMs). This paper studies persuasion dynamics by contrasting general LLMs with Large Reasoning Models (LRMs) that employ explicit "thinking" processes. Through large-scale experiments on objective (MMLU) and subjective (PersuasionBench and Perspectrum) tasks, we identify Persuasion Duality: reasoning enhances an agent’s persuasive power while simultaneously increasing its resistance to persuasion. For LRMs, adding thinking content increases persuasion rates by 21 pp on average, yet reduces susceptibility to incorrect persuasion by up to 10 pp on objective tasks. Despite these gains, we uncover a critical vulnerability: persuasiveness often stems from superficial cues such as response length and repetition rather than logical validity. Non-semantic padding or repeated conclusions can match or exceed the persuasive effect of coherent reasoning, revealing a strong length bias in agents’ judgments. We further show that persuasion propagates non-linearly in multi-hop agent chains, where intermediate agents may amplify or attenuate influence depending on task subjectivity. Finally, guided by attention analysis, we propose a prompt-level adversarial argument detection method that consistently improves agent robustness.

Large Language Models (LLMs) are transforming chemical research by enabling specialized reasoning models and tool-augmented agents. Yet existing benchmarks typically evaluate reasoning or tool invocation in isolation, overlooking the dynamic interaction between internal deliberation and external execution. We present ChemEval2, a multimodal benchmark for tool-augmented chemical reasoning, containing 2,410 samples spanning fundamental principles, spectroscopic analysis, and complex reaction understanding. Built from professional textbooks and experimental data through automated generation and rigorous human annotation, ChemEval2 introduces a tool-augmented Chain-of-Thought evaluation framework that embeds tool-use checkpoints at stages where unaided language reasoning becomes unreliable. This design evaluates whether models can identify information gaps, invoke appropriate tools, and integrate tool outputs into reasoning trajectories. Across diverse text and multimodal models, we observe a process–outcome inconsistency: models can satisfy intermediate reasoning steps or execute useful tool calls yet still fail at final synthesis. These results provide evidence that coordinating computational tool use with downstream chemical reasoning remains a substantial challenge. We open-source ChemEval2 and an MCP-based suite of specialized chemical tools for reproducible.

Code completion models often fail when the developer’s intent is under-specified in the code context. To mitigate this, developers frequently use natural language comments to clarify objectives. However, current code completion models fail to prioritize these directives effectively since they are merely pre-trained using the Fill-In-the-Middle (FIM) objective. On the one hand, the natural language instructions, mixed with the noisy code comments, are just treated as part of the background context within the prefix. On the other hand, the pre-training datasets for the FIM objective are mostly sourced from open-source repositories, which results in a scarcity of high-intent instruction-to-code pairings that reflect the developers’ workflow in code completion. To bridge this gap, we propose Instruction-aware Fill-In-the-Middle (IFIM), a fine-tuning method that extends the FIM structure with a dedicated, structurally separated instruction section. Our evaluation shows that IFIM substantially improves adherence to developer intent, while leaving infilling performance unchanged when no instruction is given. The gains hold on an in-the-wild benchmark of 100 instructions written by real developers and across model scales from 1.5B to 7B. IFIM thus offers a backward-compatible upgrade path for existing FIM-based code completion systems at a modest training cost.

Recent parameter-efficient fine-tuning (PEFT) methods effectively adapt Vision-Language Models (VLMs) to seen tasks, yet often exhibit limited generalization to unseen novel classes. To better understand this phenomenon, we analyze adaptation dynamics from the perspective of cross-modal mutual information (MI). Our empirical analyses reveal a consistent pattern across existing methods: during the adaptation process, cross-modal MI for novel classes frequently encounters early plateaus or declines, indicating weakened cross-modal semantic dependency. Motivated by this observation, we propose Retrieval-Augmented Dynamic Adaptation (RADA), a retrieval-enhanced adaptation framework that continuously injects rich instance-specific semantic priors through an LLM-enhanced dual-modality memory bank and a lightweight modality-agnostic token generator. By facilitating fine-grained cross-modal alignment while preserving pretrained zero-shot ability, RADA achieves more transferable adaptation across diverse scenarios. Extensive experiments across multiple widely used benchmarks demonstrate that RADA consistently achieves state-of-the-art performance while maintaining strong zero-shot capability. In addition, our empirical analyses show that RADA preserves more stable cross-modal dependency trajectories throughout adaptation, offering further insights into the relationship between adaptation dynamics and generalization in VLMs.

Retrieval over large codebases is a key component of modern LLM-based software engineering systems. Existing approaches predominantly rely on dense embedding models, while learned sparse retrieval (LSR) remains largely unexplored for code. However, applying sparse retrieval to code is challenging due to subword fragmentation, the semantic gap between natural-language queries and code, the diversity of programming languages and retrieval tasks, and the length of code documents, which can harm sparsity and latency. We introduce SPLADE-Code, the first large-scale family of learned sparse retrieval models specialized for code retrieval (600M–8B parameters). Despite a lightweight one-stage training pipeline, SPLADE-Code achieves state-of-the-art performance among retrievers under 1B parameters (75.4 on MTEB Code) and competitive results at larger scales (79.0 with 8B). We show that learned expansion tokens are critical to bridge lexical and semantic matching, and provide a latency analysis showing that LSR enables sub-millisecond retrieval on a 1M-passage collection with little effectiveness loss.

Diffusion language models (DLMs) have recently emerged as an alternative modeling paradigm to autoregressive LMs, offering advantages such as parallel generation and bidirectional context modeling. Despite growing interest in their generative capabilities, the privacy risks of DLMs remain underexplored. We identify a phenomenon termed token-level memorization asymmetry through theoretical analysis of diffusion training dynamics. Building on this finding, we propose Q-Skew, a quantile-weighted skewness-based indicator for membership inference on finetuned DLMs. Experiments across multiple fine-tuning datasets and models show that our method outperforms existing baselines. Moreover, we show that Q-Skew can also facilitate other privacy violations, such as PII extraction. Our findings reveal a previously underexplored privacy attack surface and highlight the need for systematic privacy evaluation of DLMs.

Multimodal harmful meme detection is challenging because harmful intent is often expressed through implicit image-text associations, cultural references, rhetorical framing, and context-dependent targets. Existing methods mainly rely on supervised multimodal models, MLLM prompting, or fixed agent workflows. However, they still lack experience accumulation and rely on static skill design, making it difficult to reuse prior judgments or adapt to emerging meme templates and long-tail harm patterns. To address these limitations, we propose HarmEvoSkill (Harmful Meme Evolving Skill-Memory), a label-free self-evolving Skill-Memory framework for harmful meme detection. HarmEvoSkill converts unlabeled reference memes into Case Memory, Insight Memory, and Risk Memory. It then automatically discovers candidate skills, abstracts recurring reasoning patterns, and induces executable skill cards without predefined skill labels. During inference, a Skill Planner retrieves relevant memories, selects task-specific skills, and generates structured evidence packets. A graph-assisted judge organizes the evidence into a harmful–harmless claim graph, makes the final prediction, and updates memory through graph diagnosis. Extensive benchmark evaluations show that HarmEvoSkill achieves consistently strong performance among evaluated methods. On HarM, it improves Accuracy by 9.33% and Macro-F1 by 17.97% over the direct MLLM baseline.

Large language models (LLMs) show growing promise for scientific discovery, yet existing benchmarks rely on static, single-turn QA formats that cannot measure the iterative reasoning and strategic planning required in real research. We introduce MolQuest, an agent-based benchmark that reformulates molecular structure elucidation as a multi-turn interactive task grounded in authentic experimental data. Within a simulated laboratory, the model must actively plan experiments, acquire heterogeneous spectral evidence (NMR, MS, IR) on demand, and iteratively refine structural hypotheses while minimizing unnecessary tool calls—mirroring the abductive reasoning workflow of a practicing chemist. We evaluate 16 frontier LLM variants under two settings, yielding 32 model–setting configurations, and find that the best Agent-mode result reaches only 51.51% accuracy, while most remain substantially lower. Furthermore, the dynamic paradigm amplifies performance for some models yet degrades others, revealing strategic planning deficiencies invisible to static evaluations. MolQuest provides a reproducible, extensible framework that exposes the critical gap between pattern recognition and genuine scientific reasoning in current LLMs.

Current aligned language models exhibit a dual failure mode we term the Evasive Servant: they sycophantically validate flawed user beliefs while deflecting responsibility with boilerplate disclaimers. We propose the Dignified Peer framework, which targets servility through anti-sycophancy and trustworthiness, and evasiveness through empathy and creativity. Operationalizing this framework requires addressing three concrete gaps: lack of fine-grained per-dimension supervision, objective imbalance during multi-objective optimization, and confound-prone evaluation. We introduce the PersonaKnob dataset, which encodes a compositional partial-order structure of multi-persona preferences, and pair it with Lag-DPO, a tolerance-constrained Lagrangian DPO algorithm that dynamically rebalances dimensions to mitigate collapse. For evaluation, we adapt the Many-Facet Rasch Model (MFRM) to disentangle model capability from judge severity, rubric difficulty, and question complexity, and we validate MFRM scores against human pairwise preferences (Spearman 𝜌 ∈ [0.67, 0.85] across dimensions). Across two open-source backbones, Lag-DPO improves all four dimensions in distribution and reduces sycophancy while preserving safe-refusal balance on out-of-distribution benchmarks; we report these results with explicit weight-sensitivity and uncertainty analyses rather than claim a fully resolved Dignified Peer.

Backdoor attacks pose a serious threat to large language models (LLMs), but existing defenses remain fragmented, failing to provide unified defense against both dirty-label and clean-label attacks. To investigate why such fragmentation arises, we present the first systematic feature-level mechanistic analysis of LLM backdoors using sparse autoencoders (SAEs). Starting from a 2×2 comparison of clean and poisoned models on clean and triggered inputs, we trace backdoor-induced logit shifts to high-contributing SAE features and categorize them into four roles: interaction, suppressed, mixed, and weight-modified features. This taxonomy reveals systematic encoding differences: dirty-label backdoors are dominated by isolated interaction features, whereas clean-label backdoors rely more on heterogeneous mixtures of mixed and weight-modified features. These differences explain why existing defenses remain fragmented across attack paradigms. We validate this hypothesis through inference-time feature clamping, which reduces ASR to at most **10.8%** in most dirty-label settings and at most **15.4%** in the majority of clean-label settings, while preserving benign-task performance. These results show that SAE-based analysis can explain defense fragmentation and guide interpretable backdoor mitigation.

Vision-language-action (VLA) models are built on the premise that semantic understanding from pretrained language or vision-language backbones should guide robot action prediction. Yet robot fine-tuning is optimized as imitation over task-specific action distributions, and many evaluations can be solved through visual or instruction-action shortcuts. We introduce **RoboSemanticBench (RSB)**, an embodied benchmark for diagnosing semantic grounding in action prediction: whether post-trained VLA models can use complex instruction semantics to select and manipulate the correct physical target. In each episode, a robot receives a multiple-choice math or general-knowledge question, observes candidate answer blocks, and must grasp the block corresponding to the correct answer. RSB covers controlled arithmetic, grade-school mathematical understanding, and commonsense or factual understanding under four-choice and ten-choice suites. Across representative VLA models, we find that many policies learn to grasp candidate blocks but select the semantically correct block at near-random or below-random rates after controlling for grasp success, revealing a persistent gap between backbone-level semantic competence and action prediction.

Large language models deployed in specialized domains must improve in-domain performance without sacrificing general capabilities. Existing parameter-efficient fine-tuning methods are typically always on: their learned perturbations are applied to every input, which can degrade out-of-domain (OOD) performance. We propose Engram Adapter, a framework that repurposes pretraining-time conditional memory as a post-hoc adapter for frozen LLMs. It uses multi-channel matching over local n-gram patterns with explicit occupancy tracking as a lightweight selectivity prior, making residual injection more likely on in-domain inputs while a learned scalar gate suppresses incoherent OOD retrievals. We evaluate on Qwen3-4B and Qwen3-8B with AG-News and MedMCQA as adaptation tasks and OOD benchmarks spanning reasoning, translation, code generation, and legal reasoning. Engram Adapter improves in-domain accuracy while preserving 99.4% to 100.1% of average OOD performance; on LegalBench it slightly exceeds the frozen base model on average, whereas comparable always-on baselines degrade sharply. Mechanistic analyses show that although OOD activations are non-zero, gate and projection attenuation reduce residuals to approximately 0.08% of hidden-state norm, yielding small KL drift and negligible accuracy change. These results suggest conditional activation is a promising route toward modular, retention-preserving domain specialization over frozen backbones.

Existing jailbreak classifiers for large language models (LLMs) often fail to distinguish truly harmful responses from benign paraphrases, resulting in high error rates. We reframe jailbreak detection as a strict instruction-following verification task rather than holistic semantic matching. To achieve this, we propose Sub-Query Decomposition (SQD), a framework that breaks complex adversarial prompts into atomic, actionable sub-intents. SQD transforms opaque safety assessment into a parameterizable adjudication protocol, allowing practitioners to quantitatively adjust safety boundaries via explicit mathematical aggregation rules for highly interpretable evaluation. To operationalize SQD at scale, we introduce three foundational pillars: (1) Efficient Deployment via distilled lightweight models, featuring SQD-Phi-4 for single-pass end-to-end verification and SQD-Decomposer, optimized with an entropy-aware loss to mathematically prevent the omission of rare constraints; (2) Evaluation Foundation through SQD-Bench, a rigorously human-annotated dataset; and (3) LLM-independent Judge via E(mbedding)-TFIDF to eliminate inference-time reasoning bias. Extensive experiments demonstrate that our unified framework provides a transparent, low-cost, and policy-aligned standard for reliable jailbreak evaluation.

Multimodal Knowledge Graphs (MMKGs) provide structured multimodal grounding for Multimodal Large Language Models (MLLMs), while Multimodal Entity Alignment (MMEA) is essential for integrating heterogeneous MMKGs. However, existing MMEA methods remain vulnerable to noisy modalities and unreliable pseudo-supervision. We propose CRMEA, a large–small collaborative framework for unsupervised MMEA that unifies PTM-based retrieval and MLLM-based reasoning through a closed-loop optimization process. CRMEA introduces a text-centric adaptive cross-modal coordination mechanism to regulate noisy visual evidence, together with a retrieve-and-vote strategy for reliable pseudo-supervision. Extensive experiments show that CRMEA consistently outperforms existing unsupervised MMEA methods and achieves performance competitive with state-of-the-art supervised approaches, while maintaining strong robustness, adaptability across heterogeneous backbones, and substantially improved cost-effectiveness over exhaustive MLLM-based alignment. Code and data are available at https://github.com/nsai-szu/CRMEA.

This paper explores agentic 3D spatial understanding, i.e., MLLM agents performing indoor 3D reasoning through tool use. Existing methods often misuse tools and exhibit biased tool preferences, leaving the agentic paradigm with only marginal gains over non-agentic strategies. We reveal that 3D spatial reasoning tasks are heterogeneous across scenes, while these agents apply a uniform tool-use strategy to all scenes rather than selecting tools according to the specific scene and task. To address this, we propose Skill-3D, a framework that learns self-evolving scene-aware skills. Specifically, Skill-3D identifies the task scene and records the agent’s tool-use trajectory into a Scene Memory, where successful trajectories from similar scenes are aggregated and distilled into a reusable scene-aware skill, with failed ones attached to the skill as lessons. During training, once a similar scene recurs, the corresponding skill is injected to guide the agent, producing new trajectories whose successes and failures further refine the skill, forming a loop in which the memory and the skill library co-evolve. Experiments show that Skill-3D substantially improves tool utilization in 3D spatial reasoning (from 39%to 78% on VSI-Bench), driving the agent toward correct and sufficient tool use. For instance, it improves Gemini-3-Flash by 67% on MMSI-Bench. Furthermore, we conduct agentic reinforcement learning over skill-guided trajectories, which boosts Qwen3-VL-8B by 60% on VSI-Bench.

Reinforcement learning with verifiable rewards (RLVR) improves vision-language models (VLMs) by optimizing outcome rewards derived from final answers. However, such outcome-only rewards do not tell the model which image regions justify an answer. For questions that require visual grounding, these rewards cannot distinguish responses supported by relevant visual evidence from those produced by language-prior shortcuts or lucky guesses. We introduce **EASE** (Evidence-Anchored Spatial Attention), which augments multimodal RLVR with visual-evidence process supervision. EASE converts annotated evidence regions into a smoothed visual-token target and uses it to guide response-to-image attention during RL training, but only on high-reward trajectories. The annotations are used solely as privileged training labels, while inference requires only the original image and question. Across Qwen2.5-VL-7B, Qwen3-VL-4B, and Qwen3-VL-8B, EASE raises average scores over DAPO by 2.5 to 3.1 points on perception, hallucination, visual math, and multimodal reasoning benchmarks. Diagnostics and ablations show that EASE better aligns visual attention with annotated evidence regions. Code are available at https://github.com/Nrich-sunny/Attend-to-Evidence/.

Although recent large reasoning models (LRMs) achieve strong performance, their reasoning processes remain vulnerable to external manipulation. We observe that LRMs can gradually rationalize incorrect user-provided hints during multi-step reasoning. Instead of rejecting false hints, models adapt their intermediate states to accommodate them and eventually produce seemingly coherent yet incorrect conclusions. We refer to this phenomenon as self-brainwashing and analyze it through empirical and theoretical analyses. We find that unfaithful reasoning steps are often locally detectable when models only focus on local reasoning context. However, this ability degrades over long reasoning trajectories, allowing unfaithful steps to accumulate and bias the final answer toward the hint. Motivated by these findings, we propose Long-Short Memory Steering (LSMS), a training-free inference-time framework that uses short-term reasoning hidden states to enhance faithfulness and long-term reasoning hidden states to maintain consistency. Across three open-source reasoning models and four reasoning datasets, LSMS consistently reduces unfaithful answer shifts while preserving benign accuracy. On average, it lowers the shifted ratio by more than 20 percentage points under three real-world hint-based threats.

Answering complex queries over long-context multimodal streams requires dynamically synthesizing asynchronous and heterogeneous evidence across extended temporal and logical spans. Existing approaches often rely on iterative Multimodal Retrieval-Augmented Generation (RAG), facing structural challenges from semantic and modality gaps as well as context bloat induced by multi-step retrieval. To address these challenges, we propose SynapseHSG, a decoupled one-shot grounding framework for long-context multimodal Question Answering. SynapseHSG introduces a hyper-semantic evidence organization method (named M2HSG) that constructs an offline graph index to align cross-modal evidence. During inference, we design a Semantic Decomposer to convert natural-language queries into graph-operable constraints. Guided by these constraints, SynapseHSG retrieves a compact query-anchored subgraph in one shot and linearizes it as structured evidence for a frozen generator. This avoids repeated LLM-driven retrieval while restricting trainable components to graph construction and query parsing. To support systematic evaluation, we introduce MM-ConfQA, a new long-context multimodal QA benchmark built from academic conference presentations. Experimental results on MM-ConfQA show that SynapseHSG improves grounded answer generation over strong multimodal and retrieval baselines, especially on dispersed-evidence questions, highlighting the value of pre-organized multimodal evidence over repeated search in raw heterogeneous contexts. We release the code and data to public.

Vision-Language-Action (VLA) policies are commonly trained from dense robot demonstration trajectories, often collected through teleoperation, by sampling every recorded frame as if it provided equally useful supervision. We argue that this convention creates a temporal supervision imbalance: long low-change segments dominate the training stream, while manipulation-critical transitions such as alignment, contact, grasping, and release appear only sparsely. We introduce FrameSkip, a data-layer frame selection framework that scores trajectory frames using action variation, visual-action coherence, task-progress priors, and gripper-transition preservation, then remaps training samples toward high-importance frames under a target retention ratio. Because FrameSkip operates only in the dataloader, it leaves the VLA architecture, action head, training objective, and inference procedure unchanged. Across RoboCasa-GR1, SimplerEnv, and LIBERO, FrameSkip improves the success-retention trade-off over full-frame training and simpler frame selection variants, achieving a macro-average success rate of 76.15% across the three benchmarks compared with 66.50% for full-frame training while using a compressed trajectory view that retains 20% of unique frames in the main setting.

Although Large Language Model (LLM) agents have demonstrated remarkable performance on complex, real-world tasks, the significant bottleneck of execution inefficiencies and static environmental feedback hinders agents’ effective generalization to open-ended scenarios. To address these limitations, this paper introduces Role-Agent, a novel framework that repurposes one LLM to concurrently function as the agent and the environment, thereby enabling a bootstrapped co-evolution. At the core of Role-Agent are two synergistic components: World-In-Agent (WIA) and Agent-In-World (AIW). In WIA where the LLM acts as the agent, we define the reward for an action as the alignment between the agent’s prediction of the next several states and the actual state. These rewards are subsequently grouped by state and optimized, equipping the agent with predictive reasoning capabilities regarding the environment. In AIW where the LLM provides environmental adjustments, we require it to analyze the typical failure modes from failed trajectories, then searches for tasks that exhibit similar failure modes, thereby altering the data distribution to facilitate learning. Extensive experiments conducted on the various benchmarks demonstrate the substantial performance improvements achieved by Role-Agent, yielding an average performance gain of over 4%.

Reinforcement learning for multi-turn agents suffers from a credit-assignment mismatch: rewards are sparse and trajectory-level, while success often hinges on a few local decisions. Existing online policy distillation (OPD) provides denser token-level supervision, but typically treats heterogeneous agent trajectories as monolithic strings rather than causal interaction units. We present StepOPSD, a post-rollout preference self-distillation framework that takes the agent step as the unit of credit redistribution. StepOPSD decomposes trajectories into action-centered step segments, rescoring them under hindsight-enriched teacher contexts and converting token-level log-probability gaps into sign-preserving advantage shaping with a normalized per-step credit budget before the GRPO update. Across ALFWorld and Search-QA with Qwen3-1.7B and Qwen2.5-3B-Instruct, StepOPSD attains best or second-best results on subsets most sensitive to local causal errors, including first-place performance on ALFWorld Heat (79.1%), PickTwo (95.0%), Search-QA TriviaQA (61.6%), and tied-best performance on HotpotQA (40.4%). The results further reveal a consistent two-knob law: smaller α_clip acts as a broadly stabilizing local trust region, whereas the optimal global mixing strength λ_mix remains task-dependent. These findings suggest that step-aware distillation is most useful when trajectory-level rewards are weakly aligned with the local action that determines downstream success.

When applying LLM agents to automatic program repair (APR), existing systems often produce Draft Patches—partially correct fixes that either fail to fully resolve bugs or overfit public test cases due to incomplete contextual reasoning and limited test coverage. To address this, we present , a general context-aware refinement agent that evolves existing APR agents by transforming Draft Patches into robust and correct fixes. Operating in a black-box, plug-and-play manner, augments diverse APR systems without model retraining or architectural changes. tackles three challenges: (1) resolving ambiguous issue context through structured contextualization, (2) improving patch quality via test-time scaling and diversified generation, and (3) consolidating partial fixes through LLM-driven refinement. We integrate into multiple APR systems to demonstrate its generality and system-agnostic design. On SWE-Bench Lite, achieves state-of-the-art performance, improving AutoCodeRover by 14.67% to 51.67% resolution. %cOn SWE-Bench Verified, it improves resolution rates by 12.2%, while yielding an average 14% gain across diverse APR systems. These results identify refinement as a missing yet essential layer in APR pipelines and highlight agent-level refinement as a practical path toward more correct and robust program repair systems. We also open source our code : https://github.com/AnvithPabba/Refine-Program-Repair.

Reliable code retrieval is crucial for developer productivity and effective code reuse. However, current neural code language models (CLMs) powering search tools are susceptible to adversarial attacks targeting non-functional textual elements. In this paper, we introduce a programming language-agnostic, transferable, adversarial attack that exploits this CLM vulnerability. Our approach perturbs identifiers within a code snippet without altering the snippet’s functionality to artificially align the code with a target query. We demonstrate that our attack, even when computed using smaller code embedding models, such as CodeT5+, is highly effective and transferable to larger, closed-source embedding models, like Voyage-code-3, or LLMs like Gemini-3.1-Pro. Our attack can increase the similarity between the query and arbitrary, irrelevant code snippets, consequently degrading key retrieval metrics such as the Mean Reciprocal Rank (MRR) of state-of-the-art models by up to 77%. The experimental results highlight the fragility of current code search methods and underscore the need for more robust, semantic-aware approaches.

Transforming academic papers into patent descriptions challenges LLMs with extreme length, strict factuality, and rigorous stylistic norms. Existing supervised methods suffer from pathological repetition, while static preference optimization causes style degradation by erroneously penalizing necessary structural repetitions. To address this, we propose PatAlign, a framework utilizing Iterative Preference Optimization to dynamically sample negative instances from the evolving policy, effectively distinguishing pathological repetition from the structural repetition required to ensure patent rigor. We also introduce the Paper-to-Patent Score (P²S) metric to evaluate the overall quality of generated patents. Experiments demonstrate that PatAlign effectively mitigates pathological repetition and style degradation, achieving state-of-the-art performance in both factual consistency and language style. Notably, our 8B model significantly outperforms 32B general-purpose models, validating the efficacy of iterative alignment strategies in complex long-form generation tasks. Our code and data are available at https://github.com/SUSTech-GenAI/PatAlign.

Generative retrieval (GR) maps queries directly to document identifiers (docids) using parametric knowledge. However, this design makes corpus expansion costly: adding new documents requires updating model parameters to encode new document-docid associations incurs repeated training and catastrophic forgetting of previously indexed documents. In this work, we revisit incremental GR as an in-context retrieval problem, where newly added documents are supplied as inference-time document-docid evidence. We propose ICICLE, an in-context indexing framework that performs source-aware docid generation over both parametric memory and context-provided document-docid pairs. ICICLE combines a ‘[COPY]‘-based routing mechanism, preference-based calibration, and large context adaptation to distinguish context-grounded retrieval from parametric retrieval. Experiments on MS MARCO and NQ320K show that ICICLE improves retrieval of newly introduced documents while preserving seen-document retention without corpus-specific retraining. Our analysis further shows that high-shot degradation is mainly caused by routing failure, highlighting source-selection calibration as a key bottleneck for scaling in-context generative retrieval.

Test-time scaling (TTS) has emerged as a powerful paradigm for enhancing large language model (LLM) reasoning by allocating additional compute during inference. External TTS, particularly Best-of- selection, generates multiple reasoning trajectories and selects the optimal candidate. However, existing process reward models (PRMs) for Best-of-N selection face a critical limitation: they were not originally designed for this task and instead rely on human-designed evaluation criteria (e.g., readability, coherence, formatting) that often misalign with actual answer correctness, leading to unstable and oscillatory performance as N increases. We introduce TrajSelector a framework specifically designed for robust and effective Best-of-N selection. Our key insight is to leverage the sampler LLM’s intrinsic hidden states as self-reflective signals, which naturally align with correctness rather than human-imposed stylistic preferences. A lightweight verifier (0.6B–4B parameters) scores reasoning steps using these hidden states, trained via outcome supervision to eliminate reliance on step-level human annotations. Experiments across six mathematical reasoning benchmarks demonstrate that TrajSelector achieves robust monotonic improvements: in Best-of-32 settings, even our smallest 0.6B verifier outperforms majority voting by 4.61% and existing 7B PRMs by 4.31–12.21%, while our 4B verifier reaches 64.05% accuracy with stable scaling behavior as N increases from 1 to 64.

Agent skills, structured packages of procedural knowledge and executable resources that agents dynamically load at inference time, have become a reliable mechanism for augmenting LLM agents. Yet inference-time skill augmentation is fundamentally limited: retrieval noise introduces irrelevant guidance, injected skill content imposes substantial token overhead, and the model never truly acquires the knowledge it merely follows. We ask whether skills can instead be internalized into model parameters, enabling zero-shot autonomous behavior without any runtime skill retrieval. We introduce SKILL0, an in-context reinforcement learning framework designed for skill internalization. SKILL0 introduces a training-time curriculum that begins with full skill context and progressively withdraws it. Skills are grouped offline by category and rendered with interaction history into a compact visual context, teaching the model tool invocation and multi-turn task completion. A Dynamic Curriculum then evaluates each skill file’s on-policy helpfulness, retaining only those from which the current policy still benefits within a linearly decaying budget, until the agent operates in a fully zero-shot setting. Extensive agentic experiments demonstrate that SKILL0 achieves substantial improvements over the standard RL baseline (+9.7% for ALFWorld, +6.6% for Search-QA and +10.1% for WebShop), while maintaining a highly efficient context of fewer than 0.5k tokens per step.

Modern Large Language Models (LLMs) depend heavily on surface-level lexical filters for safety alignment, exposing a critical lexical-semantic gap: harmful intent can be fully decoupled from explicit wording. Motivated by this vulnerability, we propose CAGE (Concept Algebra for Guardrail Evasion), a single-turn black-box jailbreak leveraging the victim’s reasoning to reconstruct masked harmful intent. Through slot masking and concept equations, CAGE eliminates explicit triggers while guiding the target model to internally deduce the payload. Because standard evaluations conflate generic harmfulness with precise intent execution, we introduce the Strict Attack Success Rate (sASR), a metric that mandates verbatim slot recovery to measure reconstructive controllability beyond mere compliance. Across seven frontier models and three benchmarks, CAGE achieves an average ASR of 93.7% and sASR of 85.1%. Defense stress tests and keyword-recovery analysis reveal that existing token-level guardrails remain largely ineffective against reasoning-driven semantic evasion, suggesting that robust alignment must go beyond surface token filtering to address threats at the reasoning level.

WARNING: THIS PAPER CONTAINS UNSAFE CONTENTS.

Vision-Language-Action (VLA) models are built on the premise that pre-trained Vision-Language Models (VLMs) can supply general semantic understanding ability for embodied tasks. However, we find that robotic fine-tuning can severely erode this capability: VQA benchmarks and layer-wise CKA analysis reveal rapid representational drift, and co-training with general vision-language data does not prevent the collapse. We propose TwinBrainVLA, an asymmetric dual-stream VLA that keeps a pre-trained "Left Brain" strictly frozen as a semantic generalist while training a "Right Brain" as an embodied control specialist. Its Asymmetric Mixture-of-Transformers (AsyMoT) mechanism lets the Right Brain query frozen Left-Brain key-value representations at every transformer layer, integrating intact semantic knowledge without corrupting it. Across SimplerEnv, RoboCasa, LIBERO, and real-robot experiments, TwinBrainVLA outperforms strong baselines and large-scale embodied models, achieving a 14.3% improvement on SimplerEnv and a 2.7× gain in physical-robot instruction generalization.

LLM agents accumulate valuable procedural knowledge during task execution, but these traces are session-bound, model-specific, and often entangled with sensitive content, making them unsafe to pool directly across users. Prior approaches to skill discovery often sacrifice cross-user generalization, iterative quality, or compliance-oriented sharing. We propose SkillLadder, a multi-user framework that distills local execution traces into a shared skill library without funneling raw data to a central server. The system evolves user-contributed traces into a multi-level hierarchy (L0–L3) using five quality gates that prevent regression below the strongest parent skill under a judge-defined score. To ensure safe sharing, SkillLadder distributes only level-≥1 skills that have undergone entity scrubbing and lineage sanitization. On 𝜏²-bench, where each isolated task execution simulates one contributing user—a deliberately conservative proxy for many small local contributors, not a model of longitudinal human users—evolved skills improve a fixed Claude Sonnet 4.6 agent in three of four domains and generalize to simulated users running different agent models. Compliance audits verify that entity retention reaches zero by L3 and source-attribution accuracy is roughly halved, substantially reducing measured sharing risks.

Multimodal large reasoning models often rely on long Chain-of-Thought (CoT) traces in which a substantial fraction of tokens, such as repeated visual descriptions, self-reflection, and other visually-disengaged filler, inflate inference cost without contributing to the answer. Existing CoT compression methods optimize output length but never measure whether a reasoning token is actually grounded in the image. We propose VIG (Visual Information Gain), an information-theoretic GRPO reward that scores each reasoning token by how much the image reduces its predictive uncertainty. VIG is computed online from two forward passes of the same policy, one with and one without the image, so no reference chains, external annotations, or auxiliary reward models are needed. Across six main multimodal reasoning benchmarks and three Qwen3-VL-Thinking model sizes (2B/4B/8B), plus an additional R1-Onevision-Bench evaluation on 8B, VIG consistently improves the accuracy-efficiency trade-off, supporting our central claim: efficient multimodal reasoning emerges from raising visual information density, where every reasoning token earns its place by anchoring to the image, rather than from imposing a length budget.

Multimodal modeling represents a vital step from modality-agnostic reasoning toward world modeling. While early approaches predominantly rely on late-fusion that assembles encoders and frozen language backbones with output heads, recent efforts have shifted the paradigm toward native multimodal modeling (NMM) with the intrinsic integration of modalities for superior multimodal performance. Despite its potential, the design space of native architectures remains insufficiently defined. In this paper, we present the community with a formalized roadmap for this transition. Specifically, we formally define the architectural nativity, distinguishing mid-fusion and early-fusion from non-native paradigms. We further organize the existing native models through the lens of input-output duality into three categories: (i) Multi-to-Text for cross-modal comprehension with text-only output; (ii) Multi-to-Target for scenario-oriented generation e.g., image, audio and video generation, and (iii) Multi-to-Multi for unified modeling with symmetric input-output. We systematically unpack the end-to-end pipeline from industrial perspectives from: architectural coordination, massive data curation, to full-stack training recipes, inference deployment, and evaluation. Finally, based on our assessment of native multimodal development trends, we provide a forward-looking perspective on its future trajectory.

We present **LLaDA-o**, an effective and length-adaptive omni diffusion model for multimodal understanding and generation. LLaDA-o is built on a Mixture of Diffusion (MoD) framework that decouples discrete masked diffusion for text understanding and continuous diffusion for visual generation, while coupling them through a shared, simple, and efficient attention backbone that reduces redundant computation for fixed conditions. Building on MoD, we further introduce a data-centric length adaptation strategy that enables flexible-length decoding in multimodal settings without architectural changes. Extensive experiments show that LLaDA-o achieves strong performance on multimodal understanding and generation benchmarks, reaching 87.04 on DPG-Bench and 0.86 on GenEval. To support future work on omni-diffusion models, we release the model and training/evaluation code at https://github.com/ML-GSAI/LLaDA-o.

Agent skills are callable procedural modules that provide reusable knowledge and execution policies for complex agentic tasks. However, existing methods mainly focus on selecting relevant skills or improving the skills themselves, while overlooking whether a relevant skill should actually be invoked at the current decision point. Unhelpful invocations may introduce irrelevant context and disrupt an otherwise correct execution process. To address this issue, we propose SelSkill, a dual-granularity preference-learning framework for selective skill invocation. SelSkill formulates skill use as a skill-or-skip decision, uses predictive uncertainty to prioritize candidate decision points, and constructs controlled invoke-skip preference pairs from shared trajectory prefixes. It further combines episode-level outcome preferences with step-level invocation preferences to capture both overall trajectory quality and the local effectiveness of skill invocation. On ALFWorld with Qwen3-8B, SelSkill improves task success by 10.9 points over the skill-enabled baseline and 7.8 over No-Skill, with 29.1-point higher execution precision. On BFCL, task success and execution precision improve by 5.7 and 29.5 points, respectively. Zero-shot results on Tau-bench and PopQA suggest partial transfer to unseen domains and skills.

Group-based reinforcement learning typically relies on coarse outcome-level signals, such as group accuracy or variance, for data selection or replay. Yet in multi-turn tool-use rollout groups, these signals miss structural differences between successful and failed trajectories: a failure may diverge into an unrelated tool chain or differ from a successful trajectory only at a local decision. We therefore introduce Group Structural Proximity (GSP), which measures failed-to-successful rollout alignment within a group using tool dependency graphs induced by tool calls and argument flow. A controlled diagnostic study with outcome-statistics-matched subsets further shows that GSP captures distinct learning dynamics: high-GSP data drives localized policy refinement, while low-GSP data induces broader policy movement and higher-entropy exploration. This positions GSP as an effective metric on the exploration–exploitation tradeoff. Building on this insight, we introduce Proximity-Guided Policy Optimization (PGPO), which considers replay buffers with high- and low- GSP under an Exploration-to-Refinement Schedule (ERS) for group-based training. On 𝜏²-Bench and ACEBench Agent, PGPO improves sample efficiency and downstream tool-use performance, demonstrating the benefit of structure-aware group comparison in multi-turn tool-use reinforcement learning.

Recent breakthroughs in Large Language Models (LLMs) have revealed remarkable generative capabilities and emerging self-regulatory mechanisms, including self-correction and self-rewarding. However, current detoxification techniques rarely exploit these built-in abilities; instead, they rely on external modules, labor-intensive data annotation, or human intervention –factors that hinder scalability and consistency. In this paper, we introduce a fully self-reflective detoxification framework that harnesses the inherent capacities of LLMs to detect, correct toxic content, and refine LLMs without external modules and data annotation. Specifically, we propose a Toxic Signal Detector—an internal self-identification mechanism, coupled with a systematic intervention process to transform toxic text into its non-toxic counterpart. This iterative procedure yields a contrastive detoxification dataset used to fine-tune the model, enhancing its ability for safe and coherent text generation. Experiments on benchmark datasets such as DetoxLLM and ParaDetox show that our method achieves better detoxification performance than state-of-the-art methods while preserving semantic fidelity. By obviating the need for human intervention or external components, this paper reveals the intrinsic self-detoxification ability of LLMs, offering a consistent and effective approach for mitigating harmful content generation. Ultimately, our findings underscore the potential for self-regulated language models and more responsible text generation systems. Code: https://github.com/KaituoZhang/SRD Warning: this paper may contain offensive content.

Pedestrian anomaly behavior retrieval requires disambiguating subtle anomalous actions from routine ones using natural language. Existing cross-modal alignment methods often struggle with confounding biases: visual entanglement (dominant environments overshadowing target actions) and textual biases (limited stylistic diversity). These spurious correlations lead to erroneous matches among visually similar but semantically distinct samples. To address this, we reformulate the task from a causal intervention perspective. We propose a dual generative framework leveraging Large Multimodal Models (LMMs) to break environment-action and style-text confounders. Specifically, we synthesize causally meaningful hard negatives to decouple actions from environments, and augment positive queries with stylistically diverse rephrasings to eliminate narrative dependency. Extensive experiments on the Pedestrian Anomaly Behavior (PAB) benchmark yield consistent improvements of +1.71% in Recall@1 and +1.01% in mAP. Furthermore, zero-shot evaluation on the out-of-distribution UCC dataset demonstrates markedly enhanced generalization, achieving +3.23% and +2.46% gains in Recall@1 and mAP, respectively.

Neural topic models are widely used to uncover latent thematic structure in text corpora, but often produce topics with limited semantic coherence and weak alignment between topics and documents. Recent refinement approaches based on large language models (LLMs) have shown promising improvements, yet typically rely on repeated generative inference, resulting in high computational cost and practical issues such as sensitivity to prompting, stochastic variability, and the introduction of out-of-vocabulary content. We propose BERT-NTM, an encoder-based semantic refinement framework for neural topic modeling that avoids generative inference altogether. Instead of generating or rewriting topic words, BERT-NTM leverages a pretrained BERT encoder to provide discriminative semantic signals during a refinement phase: topic–word distributions are refined by scoring candidate words within a topic using masked language modeling, and document–topic representations are improved by aligning document embeddings with mixtures of topic embeddings via a Maximum Mean Discrepancy (MMD) objective. Both refinement components are fully parallelizable, stable, and operate strictly within the model vocabulary. Experiments on standard topic modeling benchmarks demonstrate that BERT-NTM consistently improves topic coherence and document–topic alignment while substantially reducing runtime and GPU memory usage compared to LLM-based refinement methods.

Large language models (LLMs) have demonstrated strong potential in long-horizon decision-making tasks, such as embodied manipulation and web interaction. However, agents frequently struggle with endless trial-and-error loops or deviate from the main objective in complex environments. We attribute these failures to two fundamental errors: global Progress Drift and local Feasibility Violation. Existing methods typically attempt to address both issues simultaneously using a single paradigm. However, these two challenges are fundamentally distinct: the former relies on fuzzy semantic planning, while the latter demands strict logical constraints and state validation. The inherent limitations of such a single-paradigm approach pose a fundamental challenge for existing models in handling long-horizon tasks. Motivated by this insight, we propose a Neuro-Symbolic Dual Memory Framework that explicitly decouples semantic progress guidance from logical feasibility verification. Specifically, during the inference phase, the framework invokes both memory mechanisms synchronously: on the one hand, a neural-network-based Progress Memory extracts semantic blueprints from successful trajectories to guide global task advancement; on the other hand, a symbolic-logic-based Feasibility Memory utilizes executable Python verification functions synthesized from failed transitions to perform strict logical validation. Experiments demonstrate that this method significantly outperforms existing competitive baselines on ALFWorld, WebShop, and TextCraft. Code is available at https://github.com/LAMDA-NeSy/DualMemory.

Maintaining logical coherence in multi-hop Retrieval-Augmented Generation (RAG) remains a challenging problem, as systems must integrate fragmented evidence without suffering reasoning drift. Iterative frameworks enhance information coupling through query refinement, yet often depend on lossy compressed representations that obscure subtle semantics. Conversely, global graph-based methods ensure structural connectivity but carry topological noise and high computational costs. To overcome these issues, we propose STAGE-RAG, a STAte-aware Graph Evolution framework that conceptualizes multi-hop reasoning as continuous Bayesian entity-state evolution on a bi-relational chunk subgraph. By tracking entity states composed of identities and dynamic belief scores, our method implicitly guides reasoning via state-weighted propagation. A dual-path retrieval strategy constructs local subgraphs to prune irrelevant nodes while retaining fine-grained logical dependencies. Evaluations on HotpotQA, MuSiQue, 2WikiMultiHopQA and MoreHopQA indicate that STAGE-RAG consistently surpasses state-of-the-art baselines in both accuracy and efficiency.

In this study, we address the language-guided time-series event detection problem: given a natural-language event description and a time series, the system returns event occurrences as temporal spans. Language-guided event detection leverages domain knowledge expressed in natural language to enable zero-shot detection without domain-specific fine-tuning. We introduce GroundingMOMENT, a zero-shot cross-modal grounding framework that aligns a time-series foundation model with a language model. GroundingMOMENT concurrently localizes both complex, high-level event occurrences and their constituent domain-agnostic structural primitives (e.g., spikes, plateaus). We further introduce GroundTSSyn, a scalable pipeline for generating synthetic time-series grounding data, and release GroundTSBench, the first open-source benchmark for language-guided time-series grounding. Trained exclusively on GroundTSBench, GroundingMOMENT outperforms time-series multimodal large language models and achieves fast and accurate event detection. Case studies further demonstrate its zero-shot generalizability to real-world data from diverse domains.

Recent strategies for low-resource machine translation rely on LLMs to generate synthetic data based on text in higher-resource languages. We revisit this idea for Romansh, a language with 6 distinct varieties. LLMs tend to confuse these varieties when translating into Romansh, but they are quite good at translating out of Romansh into a high-resource language such as German. Due to this asymmetry, the direction of data augmentation is a crucial choice. We find that contrary to recent strategies, creating synthetic translations into the higher-resource language is the superior approach, and only this approach allows us to surpass a Gemini 3 Pro baseline on German-Romansh translation (+23 BLEU over Gemini in the lowest-resource variety). A human evaluation confirms that our experiments yield the first model that generates fluent translations in the individual Romansh varieties.

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by incorporating external knowledge, particularly for long-tail domains such as literary works. However, the critical step of document segmentation in RAG remains largely underexplored. Existing strategies typically either ignore semantics or overlook the complicated narrative structures of literary works, often resulting in chunks with fragmented plots and unclear references that hinder retrieval and generation performance. To address this, we propose LitSeg, a novel narrative-theory-guided segmentation framework. By employing multi-stage prompting, LitSeg explicitly extracts valid events, clarifies narrative structures, and locates turning points to inform segmentation. To alleviate the computational overhead of multi-stage inference with large-scale models, we further introduce LitSeg-Lite, a lightweight single-pass chunker fine-tuned on LitSeg-generated data via a two-stage training strategy, distilling the complex process into a single inference pass. Extensive experiments demonstrate that compared to baselines, our methods yield high-quality text chunks that are narratologically coherent and self-contained. This improved segmentation quality enhances retrieval accuracy and context relevance, and boosts downstream QA performance. Ablation studies validate the efficacy of narratological guidance and data distillation, and efficiency analysis shows that LitSeg-Lite matches the teacher at a substantially lower inference cost.

Current multilingual evaluations for Vision-Language Models (VLMs) assume a one-to-one mapping between language and orthography, overlooking billions of users of multi-script languages. We introduce PuMVR (Punjabi Multimodal Visual Reasoning), a benchmark of 1,000 strictly parallel image-text instances across Punjabi’s three active scripts: Gurmukhi, Shahmukhi, and Roman. Evaluating 10 state-of-the-art VLMs, we expose a substantial and systematic Script Gap. Models frequently solve visual tasks in one script while failing identical tasks in another, with accuracy deltas reaching 16%. Crucially, visual input boosts absolute performance uniformly yet does not close the orthographic gap. Furthermore, cross-script in-context transfer is highly brittle, exposing script-locked knowledge representation. Supported by McNemar tests across all script pairs, our findings demonstrate that current "multilingual" VLMs are not truly multi-script. We propose the Script Consistency Rate (SCR), which falls as low as 24.8% on our benchmark, as a mandatory metric for script-agnostic evaluation to ensure equitable AI access.

Supervised fine-tuning (SFT) followed by reinforcement learning (RL) has become a standard post-training paradigm for large language models. This paradigm provides a cold-start for RL exploration, avoiding the inefficiency of pure RL where on-policy sampling yields insufficient positive samples. However, in practice, existing approaches often use a small amount of data for SFT initialization compared to the RL phase, which can cause the model to fit the limited samples and shift away from its pre-trained distribution. This distribution shift impedes the model’s ability to effectively explore during subsequent RL training. To address this challenge, we propose that in low-data regimes, SFT should prioritize activating task-relevant capabilities rather than memorizing specific content. Along this line, we propose EKSFT (Entropy-KL Selective Fine-Tuning), which selectively masks tokens that exhibit either high entropy or high KL divergence from a reference model. By excluding these high-uncertainty, distribution-shifting tokens from imitation, EKSFT injects task-specific knowledge while preserving the integrity of the model’s pre-trained distribution. Empirical evaluations on mathematical reasoning benchmarks demonstrate that EKSFT consistently outperforms standard SFT. Further RL fine-tuning from the EKSFT model yields consistently better post-RL performance, indicating improved exploration for the RL stage. Our codes and datasets are available at https://github.com/MINE-USTC/EKSFT.

Recommender agents built on Large Language Models offer a promising paradigm for personalized recommendation. However, existing agents typically suffer from a misalignment between their tool-integrated reasoning trajectories and recommendation feedback, limiting their ability to distinguish fine-grained user preferences. To address these challenges, we propose AgenticRec, an agentic recommendation framework that formulates recommendation as a tool-integrated reasoning process over a recommendation-oriented tool suite. Built upon this framework, we further develop a dedicated two-stage training paradigm tailored for recommender agents. In the first stage, we introduce Recommendation-Oriented Trajectory Activation, optimize the agentic recommendation ability under implicit feedback. In the second stage, Progressive Preference Refinement further refines the agent through bidirectional preference reasoning over self-bootstrapped hard pairs, progressively sharpening preference boundaries. Theoretical analysis and extensive experiments demonstrate the effectiveness of AgenticRec.

General-purpose machine translation benchmarks such as FLORES-200 have reached a saturation regime on Chinese-English pairs, where modern large language models cluster within a narrow band of high scores. Across 22 systems, FLORES-200 zh-en GEMBA scores fall in a 7.87-point range with a standard deviation of 2.29, which compresses the separation between systems on knowledge-intensive domains such as finance, healthcare, law, and science and technology. We introduce HardMTBench, a difficulty-aware diagnostic benchmark for bidirectional Chinese-English domain translation. HardMTBench covers 12 domains and contains 10000 hand-curated source sentences with reference translations, packaged as 20000 directional test items. A three-stage construction pipeline builds a domain-balanced candidate pool of 84566 pairs, applies an LLM-based multi-signal judge over knowledge density, translation difficulty, terminology load and reference correctness, and assembles the final test set under a hardness fusion rule with per-domain quotas. Across 22 systems spanning general LLMs, commercial engines and specialised MT models, HardMTBench widens the cross-system GEMBA range by roughly a factor of two over FLORES-200, induces visible rank reorderings, and exposes domain-specific terminology and knowledge weaknesses that quality-only metrics tend to flatten.

The rapid maturation of zero-shot Text-to-Speech (TTS) models has turned high-quality voice cloning into a widely available capability, raising acute concerns over unauthorised replication, fine-tuning and resale of proprietary speech models. Yet ownership verification for TTS remains largely open: speech is a continuous waveform whose perturbations are easily destroyed by routine signal processing, and the human auditory system imposes a much tighter perceptual budget than vision. We present TTS-Guard, a black-box ownership verification framework for TTS models built on adversarial speaker-pair fingerprints. TTS-Guard(i) selects key speaker pairs in a dual embedding space for architecture-agnostic stealth;(ii) optimises a perturbation through an adaptive curriculum of shadow models covering fine-tuning, pruning, quantisation and distillation; and (iii) aggregates black-box queries into a calibrated Verification Confidence Score. On five mainstream TTS systems, TTS-Guard reaches an average Fingerprint Success Rate of 96.4% at a False Positive Rate of 5.8%, while preserving intelligibility and naturalness. The fingerprint remains effective against ten audio attacks, six model modifications, and two state-of-the-art adversarial purifiers.

LLM-based embodied agents are increasingly evaluated on environment-grounded, long-horizon planning, yet existing benchmarks rarely test whether an agent can condition its plans on the specific user it serves. This gap points to a broader limitation in embodied task-planning systems, which often ground their reasoning only in the physical environment while overlooking the diversity of the people they serve. We frame this problem as Planning with Human Attributes (PHA), in which an agent must integrate explicit user information into long-horizon embodied task planning. We introduce Trinity, a benchmark designed to evaluate PHA capabilities. Within Trinity, a single instruction can yield different, equally valid, or even opposing outcomes depending on a user’s personal attributes. Our experiments reveal that standard off-the-shelf in-context learning-based agents commonly fail at the human-grounding component, whether or not they are provided with relevant examples, highlighting a key obstacle for PHA tasks. Beyond benchmarking, we contribute a bottom-up failure-mode analysis that isolates where and why agents break, and we distill it into concrete, model-agnostic design principles for human-agent-environment (HAE) aligned agents. We hope Trinity brings domestic robots a step closer to becoming trusted, socially acceptable partners.

On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy distillation and reinforcement learning with verifiable rewards. However, we find that its standard advantage-weighted policy gradient suffers from three structural weaknesses, including high-variance updates, vanishing gradients in zero-advantage regions, and exploration bottlenecks for models with limited capability. We therefore propose asymmetric on-policy distillation (AOPD), which replaces ineffective policy gradient with localized divergence minimization in non-positive advantage regions while preserving positive reinforcement learning. Experiments on mathematical reasoning benchmarks show that AOPD consistently outperforms standard OPD, with average gains of 4.09 / 8.34 under strong / weak initialization, respectively. AOPD also maintains higher policy entropy during training and better capability retention during sequential tool-use adaptation. Our source code is available at [repository](https://github.com/stargazer319/AOPD).

Multi-agent LLM systems decompose complex tasks across specialist agents, yet joint prompt optimization remains difficult because agents are coupled through structured information handoffs. We identify interface mismatch as a structural failure mode that existing optimizers cannot diagnose: their scalar attribution signals conflate reasoning quality with format incompatibility, and the inter-agent contract is left entirely implicit. We propose BID-SPO (Bidding-based Shapley Prompt Optimization), which makes the inter-agent contract an explicit optimization variable. Each agent declares an Output Bid and an Input Demand, and a contract-aware cooperative-game characteristic function yields an attribution signal whose Shapley decomposition exposes three interpretable dimensions (global task contribution, downstream interface alignment, and upstream adaptation cost), turning attribution into a targeted diagnosis for prompt repair. A bid-conformance constraint keeps declared contracts feasible with respect to actual agent behavior throughout optimization. Across seven benchmarks and different model configurations, BID-SPO outperforms MASPO and AgentDropout, with accuracy gains of +2.93% under the Qwen3-8B + Gemini-2.5-pro configuration and +4.38% under the DeepSeek-v4 flash / pro configuration. Our findings emphasize that reliable multi-agent LLM optimization requires optimizing inter-agent interfaces as explicit objects, not just improving agents in isolation.

Zero-shot text-to-speech (TTS) aims to synthesize high-quality speech for unseen speakers using only a short reference utterance. Existing non-autoregressive TTS models typically convert text tokens into speech-length representations through filler-token padding, explicit duration prediction, or text upsampling. However, these approaches either introduce redundant computation or rely on explicit alignment assumptions. We propose M3-TTS, a native two-stream multi-modal diffusion Transformer that preserves the original lengths of text and speech representations and enables direct text–speech interaction through packed joint self-attention. M3-TTS employs a dual-stage architecture consisting of Joint-DiT layers for learning text–speech correspondence and audio-only DiT layers for acoustic refinement. Reference acoustic information is injected through AdaLN-style conditioning. Experiments on the Seed-TTS benchmark demonstrate strong intelligibility and competitive speaker similarity and perceptual quality. The Mel-VAE variant further improves inference efficiency while maintaining high synthesis quality. Code and demos will be available at https://github.com/WWWWxp/M3-TTS.

Retrieval-augmented generation (RAG) improves large language models (LLMs) with external knowledge, but this access path creates security risks distinct from inherent prompt-only or parametric-model flaws. We frame secure RAG as securing external knowledge access. We conducted a systematic search and curated 135 works on attacks, defenses, and security evaluation, and organized them with SLOT: a taxonomy along the attack Surface (S) and the corresponding defense Layer (L), with cross-cutting axes Objective (O) and attack-target scope (T). Mapping these studies onto an external knowledge-access pipeline, we expose three mismatches: target mismatch (T2 attacks outpace T2 defenses and evaluation), stage mismatch (S1 attacks outnumber L1 defenses), and signal mismatch (fluent, retrievable, corpus-fitting S1 attacks challenge anomaly-based L2/L3 defenses). Finally, we discuss directions for more realistic targets, surface-complete defense stacks, standardized evaluation, confidentiality, and multimodal and agentic systems, and release the screening process, selected-paper metadata, and machine-readable SLOT labels at https://github.com/TreeAI-Lab/Awesome-RAG-Security.

Recent progress in multimodal large language models (MLLMs) has fueled significant enthusiasm in their potential to act as autonomous agents for real-world tasks. However, scenarios requiring agents to fulfill users’ complex, structured requirements remain largely underexplored. In this work, we examine reasoning tasks under three distinct requirement scenarios: (i) Must-have requirements uniquely determine a unique feasible solution; (ii) Multiple answers satisfy the must-have requirements and are prioritized via the nice-to-have requirements; and (iii) No candidate solution satisfies the must-have requirements, in which case the agent should abstain from generating a response. We evaluate state-of-the-art MLLMs on 3,649 carefully constructed problems that reflect realistic service scenarios, including e-commerce, booking, and map-based or ride-hailing. Our evaluation reveals that existing MLLMs exhibit catastrophic failures in all scenarios. They frequently misinterpret task requirements, violate must-have requirements, and produce invalid solutions. To address this critical gap, we propose First Things First Reinforcement Learning (FTF-rl) that explicitly optimizes reasoning over multi-priority user requirements. Experimental results show that our method substantially improves the task success rate compared to strong baselines. Moreover, FTF-rl yields general effectiveness on popular logical and mathematical reasoning tasks, including LogicVista, MathVision, and InfoQA. Our findings suggest that enhancing requirement comprehension provides a simple yet effective pathway to improve generalization of MLLMs. Code is available at https://github.com/claire62/FTF-RL.

Multi-Agent Debate (MAD) has been widely adopted to improve LLM-based evaluation by prompting multiple agents to negotiate and reach a consensus. However, for subjective rubric-based scoring, inter-agent agreement does not guarantee alignment with human judgments. In this paper, we compare a single-judge baseline against a consensus-based MAD protocol on subjective evaluation tasks and design three ablations to isolate the impact of role prompting, multi-round interaction, and explicit score sharing. Evaluations across six LLMs show that the single-judge baseline achieves the strongest human alignment on average across six judge models, whereas MAD shows degradation in human alignment on both tasks. Our ablations demonstrate that this performance drop stems primarily from asymmetric role prompting rather than the interaction itself. Specifically, assigning a strict judge role introduces a systematic downward bias that the consensus process fails to correct. The central finding is that this bias reflects strict-stance dominance beyond averaging: the consensus score falls well beyond the arithmetic midpoint of the standalone strict and lenient conditions, rather than averaging them out. Removing role asymmetry (Symmetric MAD) largely recovers baseline performance, while masking peer scores widens inter-agent disagreement on average and worsens average human alignment. These findings demonstrate that multi-agent consensus can enforce artificial agreement at the expense of true human alignment, revealing a structural limitation in consensus-style, role-specialized MAD protocols for subjective scoring.

Large language model (LLM) agents increasingly rely on external skills, yet standard evaluations obscure whether retrieving these skills actually helps. Aggregate metrics often compare retrieved versus non-retrieved tasks, introducing severe selection bias and failing to isolate the true effect of skill use. To measure this actual-use capability—which we formalize as Skill Following (SF)—we introduce the Retrieval-Invoked Actual-Use Effect (RAE). RAE computes the same-task outcome difference between matched skill-enabled and skill-disabled executions, conditioned exclusively on tasks where the agent actively retrieved a skill. Evaluating 17 LLMs across coding and mathematical domains, we uncover a stark evaluation paradox: models frequently show positive aggregate retrieval lift but negative RAE. On MBPP+, multiple models that appear to benefit system-wide actually harm their own performance on the exact tasks where retrieval occurred. These findings demonstrate that aggregate averages can create a misleading illusion of tool-use proficiency, whereas RAE directly measures whether the retrieval-to-answer pipeline genuinely rescues more outcomes than it harms.

We are witnessing an emerging paradigm shift in which the adoption of large language models (LLMs) has increased significantly across high-stakes, standards-driven industries, particularly in engineering, healthcare, education, and law. In line with EMNLP 2026’s theme of "New Missions for NLP Research", this position paper aims to initiate civil discussion on how the use of LLMs is shaping the future of standard compliance and introduces standard alignment, a targeted research direction towards making LLMs verifiably comply with domain standards. To support the discussion, we propose the Criticality and Compliance Capabilities Framework (C3F) to assess the criticality levels of various standards and pair them with the current compliance capabilities of state-of-the-art LLMs as evidenced by the literature. Finally, we argue that advancing standard alignment is key to ensuring safe and trustworthy integration of LLMs across standards-driven industries.

Real-world lectures follow a predefined topic inventory (e.g., a syllabus), but instructors often reorder, paraphrase, or skip items, creating mismatches between the outline and the spoken content. Lectures also contain non-topical utterances (openings, transitions and wrap-ups) that should not be assigned to any topic, yet existing corpora rarely model both challenges. We present **MOSAIC** (**M**ultimodal **O**pen-course **S**egmentation and **A**lignment corpus **I**n **C**hinese), a large-scale Chinese MOOC corpus for *topic-guided segmentation and alignment* under predefined inventories, with explicit *non-topical* labels and matched slide resources. MOSAIC includes **MOSAIC-G** (human-annotated gold benchmark with utterance-level topic alignment and video-level links to matched slide resources) and **MOSAIC-S** (automatically labeled silver data), collectively covering 179 courses and 14,942 videos. To scale annotation, we propose **DORA** (**D**ual-stage **O**utline **R**efinement and **A**lignment), a two-stage pipeline that refines noisy inventories and then performs joint segmentation and topic assignment. Experiments show that silver supervision substantially improves the performance of local models and that incorporating slides consistently benefits knowledge-point extraction, a lecture-comprehension proxy.

While Large Language Models (LLMs) have achieved remarkable results across various benchmarks, their alignment with normative values often results in homogenized responses that fail to address diverse user preferences. Existing training-free methods often occupy valuable context windows through prompt engineering, while training-based methods typically remain static post-training, failing to support the continual optimization required in real-world settings. To address these challenges, we propose **COPE** (**C**ontinual **O**ptimization with **P**ersonalized embedding and self-**E**valuation), a novel optimization framework tailored for real-world-motivated interaction settings with sparse user feedback. Our framework assigns learnable personalized embeddings to each user and synergistically integrates preference capture, self-evaluation calibration, and personalized response optimization within a single update step. A key innovation of our method is the use of self-evaluation to generate proxy rewards, enabling continuous model updates even when explicit user feedback is unavailable. Experiments show that COPE consistently outperforms strong training-free and training-based baselines under sparse feedback, and remains complementary to Retrieval-Augmented Prompting (RAP). Further analyses confirm COPE’s reliable self-evaluation, meaningful preference patterns, stable general capabilities, and robustness under shifting preferences and alternative evaluators.

Molecular design is a central task in drug discovery that requires precise structural reasoning and domain knowledge. While large language models (LLMs) have shown promise in generating high-level editing intentions in natural language, they often struggle to faithfully execute these modifications, particularly when operating on non-intuitive representations like SMILES. We introduce MECo, a framework that bridges reasoning and execution by translating editing actions into executable code. MECo reformulates molecular modification for LLMs as a cascaded framework: generating human-interpretable editing intentions from a molecule and property goal, followed by translating those intentions into executable structural edits via code generation. Our approach achieves over 98% accuracy in reproducing held-out realistic edits derived from chemical reactions and target-specific compound pairs. On downstream benchmarks spanning physicochemical properties and target activities, MECo substantially improves consistency by 38-86 percentage points to 90%+ and achieves higher success rates over SMILES-based baselines while preserving structural similarity. By aligning intention with execution, MECo enables consistent, controllable and interpretable molecular design, laying the foundation for high-fidelity feedback loops and collaborative human–AI workflows in drug discovery.

Large language models (LLMs) augmented with external tools have demonstrated strong capabilities in API selection, invocation, and reasoning over tool outputs. However, existing training-free tool-use methods usually rely on inference-time context construction to guide the entire tool-use process with a unified input, including tool descriptions, task prompts, demonstrations, trajectories, or planning hints. Tool-use decisions place different emphasis on this contextual evidence: tool/API selection emphasizes task–API correspondence, while call generation additionally uses API schemas, local call transitions, and intermediate observations. We present SD-ICL, a Stage-Decomposed In-Context Learning method that decomposes context construction into tool selection and call generation stages. SD-ICL first retrieves tool selection demonstrations by combining task semantics with tool set relevance and constructs an explicit tool selection state. It then uses this state to retrieve call generation demonstrations through API-set similarity and adjacent call relations, organizing complementary context for the executor. Experiments show that SD-ICL improves task completion and preference quality across the evaluated tool-use benchmarks and model settings.

Despite the growing success of Large Speech Language Models (LSLMs) in processing short-term acoustic signals, their extension to long-form audio understanding is severely bottlenecked. This limitation stems from the limited context length and the exorbitant memory footprints required for long-form inference. In this work, we propose Speech-XL, a new model that capitalizes on the intrinsic key-value (KV) sparsification capacity of Large Language Models (LLMs) to achieve high-ratio speech input compression. Specifically, we introduce a novel special token, the Speech Summarization Token (SST), for each speech interval to encapsulate the intra-interval speech information into its associated KV pairs. The SST module is trained via instruction fine-tuning, employing a curriculum learning strategy where the SST learns to compress information in a progressive manner—advancing from low-ratio (simple) to high-ratio (challenging) compression. Despite using substantially less training data than large-scale audio foundation models, Speech-XL achieves competitive compression quality under the same backbone, maintaining a small gap to its uncompressed counterpart while substantially reducing long-form inference cost. By addressing the long-standing bottlenecks in long-form audio modeling, our approach offers a novel perspective on the condensation of extensive acoustic sequences.

Per-matrix singular value decomposition (SVD) truncation is Eckart–Young optimal in the whitened Frobenius norm, but errors from independently compressed matrices compound through the block’s nonlinear forward pass. Inspired in part by hierarchical variational optimization in quantum many-body methods, we introduce a three-level chain that widens optimization scope from individual matrices to Transformer blocks to the full model: whitened SVD (L1), block-level joint optimization (L2), and end-to-end language-modeling loss refinement (L3), all from 256 calibration sequences, with no instruction or recovery data. On LLaMA-7B at 60% compression, the chain reduces WikiText-2 perplexity from 42.1 to 19.1 to 11.4. The block-level stage acts as a regularizer: skipping it worsens Penn Treebank (PTB) perplexity by 24 points, a gap that additional end-to-end training did not close in our experiments. Perplexity gains hold across 20–80% compression, five architectures up to 13B parameters, and both in-distribution and out-of-distribution benchmarks, though the cross-architecture rows use architecture-specific configurations and the ratio sweep was not run under one common protocol. With more calibration data, skipping the block-level stage becomes competitive, revealing an offline compute–data trade-off. We therefore claim improvements only in perplexity and compression fidelity; downstream accuracy remains well below the dense model.

Organizations often develop and maintain portfolios of related applications: independently deployable codebases that share substantial domain logic, interface patterns, or operational conventions. As LLM coding agents are increasingly used to generate and maintain such software, a naive application-by-application workflow duplicates shared logic across codebases and allows prolonged agentic maintenance to accumulate verbosity, dead code, and structural erosion. We introduce the Super Library Agent problem, where an agent sequentially generates a portfolio of N related applications while maintaining a shared Super Library of reusable cross-application components. A minimal sequential scaffold can in principle extract shared code and migrate applications to the evolving library, but in practice suffers from low extraction recall and fragile dependency migration. We address these failures with candidate-guided extraction over code chunk summaries, pre-extraction codebase consolidation, and context-aware migration using extraction traces and call-graph information. Across WebGen-Bench and PaperBench, our method preserves application functionality while significantly reducing redundancy and token footprint (verbosity, token length) over zero-shot, and avoiding the structural erosion introduced by naive library construction, with additional reductions in LOC and MDL. Our code is available at https://github.com/sbigstar0310/super-library-agent.

Reliable confidence estimation is essential for large language model deployment. However, answer-level calibration remains challenging because generation errors are often localized: a response may be fluent and high-probability overall while still failing at a critical number, entity, or factual claim. Existing estimators compress token probabilities, sequence likelihoods, entropy, or beam statistics into a global score, which can dilute such local risk signals. We propose TRACE, a single-pass, decoded-answer-preserving confidence estimator that treats decoding-time uncertainty as a trajectory through three steps: (i) recording token-level surprisal and predictive entropy during decoding, (ii) applying local risk operators to preserve uncertainty spikes, and (iii) converting localized trace risk into answer-level confidence. TRACE produces a label-free risk score, while TRACE+ calibrates trace-only features into probabilities using a held-out split, without extra generations or external verifiers. We evaluate four tasks against 19 calibration baselines, and TRACE+ reduces Brier from 0.149 to 0.137 and improves AUROC from 0.758 to 0.792 over the strongest likelihood baseline. Across seven LLMs, TRACE+ improves over the best non-TRACE baseline pool from 0.136 to 0.120 Brier and from 0.764 to 0.817 AUROC. Results show localizing decoding-time risk provides a general approach to calibration.

Expanding the vocabulary of conditional text-to-speech (TTS) models typically requires extensive fine-tuning with paired audio, creating a major bottleneck for low-resource and cross-lingual adaptation. To address this, we introduce SEAM, a zero-audio initialization method that constructs embeddings for unseen tokens from known subtokens. Unlike standard pooling strategies that ignore sequential structure, SEAM applies a spectral filter to preserve temporal phonetic order and uses closed-form ridge regression to align the new embeddings with the pretrained conditioning manifold. Across three languages and two TTS architectures, SEAM matches baseline performance using up to four times less training data, reducing sample complexity by 75% compared to random initialization and mean pooling. Our approach provides a lightweight, reproducible, and theoretically grounded solution for vocabulary expansion in low-resource speech synthesis.

Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network’s representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.

Scaling Multimodal Large Language Models (MLLMs) to long-form speech is bottlenecked by the explosive growth of input tokens. Existing speech-language models project high-frame-rate acoustic features directly into the LLM input space, making long-context processing computationally prohibitive. Unlike images or videos, speech lacks spatial redundancy, making extreme token compression particularly challenging. To address this limitation, we propose FastSLM, a token-efficient architecture featuring the Hierarchical Temporal Abstractor (HTA), which progressively distills acoustic features across multiple temporal scales. HTA achieves an extreme compression rate of 1.67 tokens per second (97% reduction) while preserving essential linguistic information for downstream speech-language understanding. Experimental results demonstrate that FastSLM achieves competitive performance across diverse speech-language tasks while requiring substantially fewer speech tokens and FLOPs than existing speech-language models. The source code and model checkpoints are available at https://github.com/Lee-junseok1025/FastSLM.

Assessing the developmental suitability of children’s narratives is important for educational recommendation and developmental literacy research, yet such assessment typically relies on subjective and difficult-to-scale human judgment. This raises an important question: Can AI systems approximate human developmental judgments of children’s stories? To study this problem, we introduce CLARA, a cognitively grounded framework for developmental narrative understanding through structured annotation across cognitive (COG), language (LAN), and social-emotional (SEL) dimensions, together with a bilingual benchmark resource containing 1107 Chinese–English children’s stories with normalized silver developmental references and structured developmental annotations. We evaluate CLARA through benchmark comparison, component analysis, translated bilingual consistency analysis, and blinded human evaluation with educators. Experimental results show that structured developmental annotation achieves substantially stronger alignment with developmental references and human judgments than readability-based methods and direct prompting baselines. Overall, our findings suggest that AI systems can approximate certain aspects of human developmental judgment when guided by structured developmental annotation, while also highlighting the importance of interpretability and human oversight in educational NLP.

As Large Language Model (LLM) capabilities advance, locally deployed personal agents relying on API-based remote models and external skills have emerged as a novel paradigm. With the rapid expansion of available skills, enabling personal agents to learn and adapt to implicit user preferences becomes a critical challenge. However, local deployment constraints preclude complex centralized selection algorithms, creating an urgent need for a lightweight local preference harness. This paper explores the implementation of such a harness through a novel architecture that strictly decouples statistical preference learning from semantic intent parsing. Specifically, we leverage localized statistical results to influence and modulate the selection decisions of the remote LLM. Extensive evaluations demonstrate that our decoupled approach achieves the lowest cumulative regret and highest test accuracy, significantly outperforming traditional memory-augmented agents. We open-source our code at https://github.com/ZyGan1999/Personalized-Skill-Selection.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse domains, yet personalizing their outputs to individual users remains an open challenge. Existing approaches predominantly adopt a flat behavioral paradigm, aggregating user behaviors without an explicit account of how they are organized into deeper behavioral structures. In this work, we draw on Pierre Bourdieu’s Theory of Practice to propose PHF (Practice-Habitus-Field), a sociologically grounded framework that reconceptualizes LLM personalization through three hierarchical levels: individual behaviors as practices, their temporal accumulation into stable dispositions as habitus, and shared regularities across similar users as fields. We instantiate PHF through PHF Compass, a lightweight and model-agnostic implementation based on a frozen LLM. Experiments on the Language Model Personalization (LaMP) benchmark demonstrate consistent improvements across diverse tasks, while further analyses validate the interpretability and extensibility of the learned behavioral structures.

Memory bottlenecks severely constrain the deployment of large language models (LLMs) in long-context scenarios. While existing global pruning methods reduce memory usage during the decoding phase, they largely ignore out-of-memory errors in the prefill stage before generation begins. To address this, we present Task-Aware Chunked KV Cache Eviction (TAKE), a training-free framework for memory-efficient long-context prefill. TAKE refines the prefill process by partitioning long sequences into chunks and performing incremental KV eviction, thereby avoiding the computational overhead of full-sequence processing. Crucially, to mitigate the semantic discontinuity introduced by chunk-wise processing, we introduce accumulated task-aware probe tokens and a delayed eviction strategy to reduce information loss. Extensive evaluations show that TAKE achieves the highest efficiency, slashing peak memory usage to 59.53% of the baseline while accelerating Time-To-First-Token (TTFT) by over 60% under 128k-token inference. Our code is available at [Source Code of TAKE](https://github.com/ObitoJN/TAKE).

Reasoning distillation from powerful teacher models to smaller students faces the Gap Curse: as teachers grow more sophisticated, their complex distributions increasingly diverge from what students can approximate, causing performance degradation. Existing mitigation strategies either filter out challenging examples through data selection or introduce weaker intermediate assistant models, inherently compromising supervision coverage or quality. We propose Teacher Alignment, which directly adapts the teacher toward the student’s distribution without discarding data or degrading reasoning quality. However, naive alignment through standard knowledge distillation triggers catastrophic collapse of the teacher’s reasoning capabilities. To address this, we reformulate teacher alignment as reinforcement learning and introduce TeacherGRPO, built on Group Relative Policy Optimization with two key innovations: (i) Curriculum Selective Alignment applies dual token- and distribution-level curricula to focus rewards on high-signal reasoning gaps while filtering noise from trivial tokens and uncertain tail distributions, and (ii) Importance-Adaptive Length Regularization selectively penalizes verbose redundancy while preserving pedagogically critical reasoning steps. The aligned teacher then distills knowledge to students via standard pipelines. Extensive experiments show TeacherGRPO significantly outperforms baselines across diverse reasoning benchmarks and distillation methods. Our code is available at https://anonymous.4open.science/r/TeacherGRPO-290F/README.md.

Object hallucination in Multimodal Large Language Models (MLLMs) often appears in two forms: omitting objects that are present in the image and generating objects that are absent. Existing training-free mitigation methods usually apply a uniform intervention to both cases. We find that this strategy leads to an imbalanced effect: it can reduce some omission errors, but often introduces new fabrications. To understand this behavior, we analyze spatial attention patterns and layer-wise attention dynamics during generation. Our results suggest that omissions often occur when the model attends to the relevant visual regions but gradually loses visual confidence during decoding, whereas fabrications are more closely related to deep-layer semantic priors overriding visual evidence. Based on this observation, we propose Visual Potential Field Calibration (VPFC), a training-free hidden-state calibration method. VPFC locates an intervention region using the spatial centroid of high-confidence visual tokens and applies a lightweight hidden-state update to adjust the model’s generative confidence. Experiments on multiple hallucination benchmarks show that VPFC consistently improves over existing training-free methods while maintaining inference efficiency close to standard greedy decoding.

Extracting structured, machine-readable compliance criteria from regulatory documents remains an open challenge. Single-pass language models hallucinate structural elements, lose hierarchical relationships, and fail to resolve inter-document dependencies. We introduce RegReAct, a self-correcting multi-agent framework that decomposes regulatory information extraction into seven specialized stages, each with an Observe–Diagnose–Repair (ODR) loop that validates outputs against the source, correcting not only model hallucinations but also cross-reference errors in the regulations themselves. To ensure structural accuracy, RegReAct constructs a typed criterion graph; to ensure completeness, it resolves external dependencies by retrieving, summarizing, and embedding referenced legal content inline, producing self-contained outputs. Applying RegReAct to three EU Taxonomy Delegated Acts, we construct a dataset comprising 242 activities with over 4,800 hierarchical criteria, thresholds, and enriched source summaries. Evaluation against a GPT-4o single-pass baseline shows that RegReAct outperforms it across all structural and semantic metrics.

Large language models (LLMs) are increasingly used to assist scientific writing and experimental design, but a fluent protocol can still be physically infeasible. We introduce PhysDox, a benchmark for physical feasibility auditing of biomedical experimental protocols, comprising a 683-sample expert-curated Gold set and a 5,000-sample expanded Silver set across six sensing domains. PhysDox formulates protocol auditing as a two-stage task: severity detection over valid, minor issue, and fatal violation protocols, followed by diagnosis of fatal violations over coarse and fine-grained physical constraints. We evaluate 6 recent LLM variants under 4 inference strategies. Results show that current LLMs remain poorly calibrated at the severity gate: the best Gold Stage-1 macro-F1 is only 53.0, and strong oracle diagnosis often collapses after end-to-end gating. Qualitative case analyses reveal a shared failure mode: models reward complete experimental scaffolds while missing hidden physical constraints. These results show that protocol auditing requires calibrated feasibility reasoning over complete experimental procedures, rather than factual recall or isolated physics problem solving.

Large language models (LLMs) are increasingly deployed as automated reviewers whose numeric scores directly determine accept/reject outcomes, yet whether those scores faithfully reflect the models’ own critiques remains underexplored. We identify threshold-alignment failure: an LLM reviewer may clearly articulate serious weaknesses while still assigning scores above the decision boundary, rendering its critique inconsequential. We introduce DSAR (Dimension-Specialist Audit and Revision), a training-free protocol that decomposes evidence elicitation across dimension specialists, applies severity-coded adversarial audit, and enforces deterministic score corrections for critical findings before thresholding. On PropAlign-240 under a shared Qwen2.5-72B-Instruct backbone, six standard reviewer protocols keep most human-rejected proposals above the decision threshold, yielding false-positive rates of 89.2–100.0%. DSAR raises reject recall to 45.9% and achieves the lowest MAE (0.334), with all decision-level gains statistically significant (McNemar p < 0.008). Ablations are consistent with severity-based audit correction, rather than longer reasoning or additional agents, as the main source of the improvement. These findings argue that LLM reviewers should be evaluated not only by critique quality, but by whether critiques translate into threshold-consistent decisions.

Human disagreement is ubiquitous and well-known in labeling. However, variation in explanations, captured through token-level human rationales, remains far less explored. At the same time, it is unclear how to best evaluate human labels and rationales, or how to best aggregate rationales beyond majority vote, in light of this variation. Yet rationales can reveal the richness of human reasoning, which may differ in style, values, and interpretations, especially in subjective NLP tasks such as hate-speech detection. In this work, we unify diverse models, training strategies, loss functions, and existing evaluation metrics under a single protocol by systematically re-implementing them across different label and rationale representation spaces. Classification metrics are organized around two key properties—predictive and distributional—while explainability metrics through three complementary dimensions: plausibility, faithfulness, and complexity. In this unified evaluation framework, we assess model behavior across classification and explainability metrics, as well as metric sensitivity to the choice of label (hard and soft) and rationale representation space (hard, intermediate and soft). Results show that intermediate and soft representations better capture variation in labels and rationales, with metrics revealing complementary facets of the same property. These findings highlight the need to align evaluation with task goals and rethink subjective NLP evaluation to better reflect diverse human judgments.

Reinforcement Learning with Verifiable Rewards (RLVR) has markedly improved the performance of Large Language Models (LLMs) on tasks requiring multi-step reasoning. However, most RLVR pipelines rely on sparse outcome-based rewards, providing little supervision over intermediate steps and thus encouraging over-confidence and spurious reasoning, which in turn increases hallucinations. To address this, we propose FaithRL, a general reinforcement learning framework that directly optimizes reasoning faithfulness. We formalize a faithfulness-maximization objective and theoretically show that optimizing it mitigates over-confidence. To instantiate this objective, we introduce a geometric reward design and a faithfulness-aware advantage modulation mechanism that assigns step-level credit by penalizing unfaithful steps while preserving valid partial derivations. Across diverse backbones and benchmarks, FaithRL consistently reduces hallucination rates while maintaining (and often improving) answer correctness. Further analysis confirms that FaithRL increases step-wise reasoning faithfulness and generalizes robustly.

Diffusion large language models (dLLMs) offer a compelling alternative to autoregressive models, yet they may expose sensitive training data during denoising. Detecting such usage is challenging because dLLMs lack the efficient one-pass probability decomposition of causal architectures. Existing methods rely on random masking to obtain tractable token-wise detection signals under limited query budgets, but fail to control dependencies among masked tokens. We demonstrate that this token-wise approximation introduces a non-negative structural estimation error, which is theoretically characterized by the cumulative conditional mutual information (CMI) among masked tokens and can obscure subtle memorization signals. This insight suggests that reliable detection requires masked token sets with weak internal dependency. To avoid the prohibitive cost of directly estimating CMI over token combinations, we propose Independent Token Sampling (ITS), a query-efficient framework that uses an attention-derived pairwise dependency proxy to approximate the CMI-aware selection criterion. ITS further incorporates a diversity-promoting strategy to improve token coverage across sampling rounds, yielding aggregated token-wise signals that are less affected by dependency-induced approximation error. Experiments on multiple datasets show that ITS consistently outperforms state-of-the-art baselines across different models and datasets, achieving an AUC improvement of 0.18 on the ArXiv dataset while maintaining strong performance under limited query budgets.

Verifiable reasoning training compresses an entire response to a binary pass-or-fail label, yet a fail label alone does not reveal whether the corresponding trajectory should be discarded, penalized, or repaired. Many failed trajectories remain close to the correct solution through substantial portions of their intermediate reasoning and deviate only at a later branching point, whereas others contain no reliable prefix that can support a correct suffix. Treating both failure types as uniformly negative supervision suppresses useful reasoning states while reinforcing erroneous ones. To address this limitation, we propose prima, which decomposes failed trajectories into inheritable prefixes and divergent suffixes, and rewrites repairable failures into shared-prefix preference pairs for localized reasoning optimization. We formalize the optimization cost of mixing the two failure types and show that treating them as a uniform negative class shrinks effective gradient magnitude, providing a principled justification for separating repairable from irrecoverable failures before training. Experiments show that prima converts failed trajectories from coarse-grained negative supervision into cleaner suffix-level supervision signals, leading to consistent gains on code and mathematical reasoning tasks.

While role-playing agents excel in short-term interactions, long-term conversations overwhelm context windows, motivating external memory frameworks. Current systems typically rely on persona-agnostic summarization, which records facts without persona-specific interpretation, yielding generic responses that compromise persona fidelity. To bridge this gap, we introduce RoleMemo, a dataset featuring four reasoning tasks where the factual fragments must be interpreted through the persona to reach the correct answer. Evaluation on RoleMemo exposes critical limitations of persona-agnostic frameworks. We thus propose DualMem, which decouples memory into two streams: factual cognition and persona-conditioned insight. Trained through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), our framework with a 4B-parameter model outperforms zero-shot persona-agnostic frameworks powered by DeepSeek-V3.2 for sustained persona fidelity. Our resources are available at https://github.com/Collab-Gen/RoleMemo.

Computer-use agents could automate repetitive screen-based clinical work, but their reliability in medical graphical user interfaces remains largely unvalidated. Existing benchmarks focus on general web or desktop tasks and underrepresent medical software, which requires domain knowledge, exhibits markedly different UI design from mainstream applications, lacks public testing environments, and demands safety validation beyond task completion. We introduce MedCUA-Bench, an interactive benchmark for clinical computer-use agents. It covers 18 clinical scenarios across 10 medical domains, reconstructed from real product manuals and open-source medical systems to capture authentic clinical interfaces while avoiding licensing and privacy constraints. Each task ships with paired intent- and step-level goals to disentangle clinical reasoning from UI execution, and is evaluated by a deterministic checker over task completion and five clinical safety dimensions. Across 23 agents, the best closed-source model reaches 54.2% strict success, while all models remain below 9% on the real OpenEMR. Open-source agents average only 2.5%, with the best reaching 16.2%. MedCUA-Bench exposes the gap between current agents and reliable clinical software use, providing a reproducible testbed for future research. Code: https://github.com/Jia7878/MedCUA-BenchMark.

With the rapid advancement of Large Language Models (LLMs), AI-generated text has become increasingly difficult to distinguish from human-authored content, raising serious concerns about online misinformation and academic integrity. Although plenty of methods have been proposed, they often suffer from high false positives on formal or highly structured human-written texts. Moreover, while external contextual information can provide valuable signals for verification, it remains largely underutilized in existing detection frameworks. In this paper, we propose an agentic framework driven by LLM for AI-generated text detection that integrates multiple machine learning (ML) detectors with an online search module to retrieve and leverage external evidence. Evaluated on a diverse real-world dataset spanning four distinct sources, our agent significantly outperforms standalone detectors, achieving an overall accuracy of 93.90%. We show that online search acts as an effective error-correction mechanism under detector conflict and incorrect agreement, improving accuracy by 10.81% and 90.91%, respectively. We further conduct a failure analysis, identifying key limitations and providing directions for future research.

Web data curation has been widely studied for decoder Large Language Model (LLM) pretraining. Encoders for dense-terminology domains such as medicine, by contrast, are pretrained on small, manually-curated corpora that limit scalability and writing style diversity, a bottleneck even more severe in non-English clinical settings. Whether web-scale data curation also benefits encoder Masked Language Modeling (MLM) in a dense-terminology domain remains an open question. To address this, we introduce two complementary levers. Medical-term density filtering selects documents rich in medical terms. Signal-amplifying rephrasing uses an LLM to rewrite documents into denser variants with broader entity contexts. We instantiate the recipe on French medical NLP. The medical-term density filter outperforms the widely-used educational quality filter on downstream medical tasks, and the two complement each other. Signal-amplifying rephrasing alone improves on raw web data, and mixing it with filtered web data produces the largest gain. The recipe yields FineMed, a French medical pretraining corpus, and DoctoBERT, a state-of-the-art French medical encoder family evaluated on both the public benchmark DrBenchmark and a proprietary clinical Named Entity Recognition (NER) task.

Multi-hop kinship reasoning is a natural testbed for LLM compositionality, but existing benchmarks (notably CLUTRR) cover only the descriptive Eskimo system. We introduce KinshipQA, a procedurally-generated benchmark covering seven anthropologically-documented kinship systems (Eskimo, Sudanese, Hawaiian, Iroquois, Dravidian, Crow, Omaha) and up to six reasoning hops, with a tunable simulator horizon that eliminates exact-instance pretraining overlap. Evaluating six LLMs, we find a 40.9% accuracy drop when reasoning shifts from biological multi-hop to culturally-marked classification on the five non-descriptive systems. The drop follows the anthropological hierarchy descriptive > bifurcate > generational > skewing, persists under chain-of-thought and few-shot prompting, and compounds with depth (10.6% at 5–6 hops). Humans given the override rule reach 89.0% versus 50.7% for LLMs, placing the failure in the model rather than the task. Two follow-up experiments suggest distinct contributors. A fictional-rule control swapping system labels and kin terms for invented strings raises accuracy by 6.1%, implicating familiar English surface forms. An in-context-rule probe prepending the override rule helps skewing systems (+17.1%) but hurts non-skewing systems whose baseline already exceeds ~60% (-13.4%), consistent with a missing skewing prior alongside rule interference where the model already has a working approximation.

Masked diffusion language models (MDLMs) enable parallel decoding by predicting all masked positions at each denoising step, yet existing training-free samplers usually decide which positions to commit at token-level granularity. We revisit this granularity and observe that reliable predictions often emerge as contiguous high-confidence spans, suggesting that the unit of parallel commitment can be larger than a single token. We first group adjacent high-confidence candidates into confidence-induced clusters (CICs) as span-level update units. We then use self-attention maps from the same forward pass to estimate inter-cluster dependencies, enabling conflict-aware selection of mutually compatible CICs for parallel commitment. This yields CLAD (**C**luster-**L**evel **A**ttention-Guided **D**ecoding), a training-free cluster-level decoder for MDLMs. Experiments on LLaDA and Dream model families across four reasoning and code-generation benchmarks show that CLAD achieves 1.77×–8.47× speedups over Vanilla decoding while maintaining broadly comparable task accuracy in most settings.

Theory of Mind (ToM) reasoning requires inferring agents’ beliefs from partial and asymmetric observations, which remains an open challenge for LLMs. Existing prompting-based approaches improve ToM reasoning through observable-event filtering or temporal belief chains, without explicitly modeling nested beliefs. We introduce RecToM, an inference-time framework for ToM reasoning that models nested beliefs via recursive perspective construction. RecToM constructs each character perspective from the preceding character perspective along the character chain specified by the question, reducing higher-order belief questions to actual-world questions within the final constructed perspective. We further provide a KD45 analysis showing that RecToM’s perspective construction induces a well-formed belief modality beyond simple event filtering. Experiments on ToM benchmarks, including Hi-ToM, Big-ToM, and FanToM, across multiple LLM backbones show that RecToM consistently outperforms recent advanced approaches, achieving state-of-the-art performance. Notably, RecToM reaches 100% accuracy on Hi-ToM with GPT-5.4 and Qwen3.5, a benchmark requiring higher-order ToM reasoning.

Legal Judgment Prediction (LJP) predicts applicable law articles, charges, and penalty terms from case facts. Beyond accuracy, LJP calls for intrinsically interpretable and legally grounded reasoning that can reconcile statutory rules with precedent-informed standards. However, existing methods often behave as static, one-shot predictors, providing limited procedural support for verifiable reasoning and little capability to adapt as jurisprudential practice evolves. We propose VERDICT, a self-refining collaborative multi-agent framework that simulates a virtual collegial panel. VERDICT assigns specialized agents to complementary roles (e.g., fact structuring, legal retrieval, opinion drafting, and supervisory verification) and coordinates them in a traceable draft–verify–revise workflow with explicit Pass/Reject feedback, producing verifiable reasoning traces and revision rationales. To capture evolving case experience, we further introduce a Hybrid Jurisprudential Memory (HJM) grounded in the Micro-Directive Paradigm, which stores precedent standards and continually distills validated multi-agent verification trajectories into updated Micro-Directives for continual learning across cases. We evaluate VERDICT on CAIL2018 and a newly constructed CJO2025 dataset with a strict future time-split for temporal generalization. VERDICT achieves state-of-the-art performance on CAIL2018 and demonstrates strong generalization on CJO2025. To facilitate reproducibility and further research, we release our code and the dataset at https://github.com/luckyhui235/VERDICT.

Speculative decoding (SD) accelerates large language model (LLM) inference by using a lightweight draft model to propose tokens and a stronger target model to verify them. However, standard SD is mainly designed for acceleration, and its output quality is typically constrained by the target model. In this work, we propose Entropy-Aware Speculative Decoding (EASD), a lightweight and training-free extension of SD that improves reasoning quality through token-level entropy-guided rejection. EASD detects cases where both draft and target models exhibit high uncertainty while strongly overlapping in their top predictions. In such uncertain-agreement cases, EASD rejects the aligned token and resamples from the target distribution, preventing low-confidence errors from propagating. Experiments on challenging reasoning benchmarks show that EASD consistently improves accuracy over standard SD and reward-guided variants while maintaining comparable inference efficiency. Notably, EASD can surpass the standalone performance of the target model, suggesting that speculative decoding can serve not only as an acceleration method but also as an effective mechanism for improving reasoning quality. The code is available at https://github.com/ECNU-Text-Computing/EASD.

Forced alignment evaluation typically requires manually annotated timestamps, limiting large-scale and multilingual analysis. We introduce two corpus-level metrics based on self-supervised (SSL) speech representations for reference-free forced alignment evaluation: Phoneme-Cluster Mutual Information (PCMI) and Word Acoustic Consistency Score (WACS). PCMI measures agreement between aligned phoneme labels and clusters induced from SSL-speech representations, while WACS measures consistency of repeated word realizations using dynamic time warping similarity between word representation sequences. Using both random and systematic perturbations, we show that PCMI and WACS degrade consistently under alignment perturbations. We further analyze the metrics across multiple alignment systems on 85 languages from FLEURS, validate them against manually annotated alignments from 45 languages in DoReCo, and evaluate them on two phonologically complex low-resource languages. The metrics effectively separate high- and low-quality alignments and correlate strongly with timestamp-based alignment quality measures. Our results demonstrate that SSL-speech representations enable scalable, reference-free forced alignment evaluation. The metrics are available as an open-source Python package at https://github.com/mahesh-ak/forced-aligner-metrics.

Model merging enables combining multiple fine-tuned models without additional training, but its safety implications remain poorly understood. Prior work primarily attributes merging risks to unsafe constituent models, implicitly assuming that merging individually aligned models preserves safety. In contrast, we show that model merging reveals a previously overlooked jailbreak risk rooted in the pretrained foundation model, even when all constituent models are individually safety-aligned. Motivated by this observation, we study a new threat setting where an attacker constructs jailbreak prompts that generalize across merged models sharing the same pretrained backbone, without access to the exact merging coefficients or constituent checkpoints. To exploit this phenomenon, we propose Basin-Aware Jailbreak (BAJ), which formulates jailbreak generation as a min–max optimization over the merging space to produce transferable adversarial suffixes across merged model families. Experiments across diverse backbones and merging settings show that BAJ achieves consistently high transfer success rates and remains effective under existing defenses.

Large language models (LLMs) offer a promising approach to machine translation (MT) for extremely low-resource languages by incorporating linguistic resources through in-context learning. However, LLMs often struggle to apply grammatical information effectively during translation. Inspired by recent progress in chain-of-thought reasoning, we investigate whether low-resource MT can benefit from structured intermediate steps of linguistic analysis and grammatical reasoning. We propose a pipeline for automatically generating step-by-step linguistic reasoning traces from Universal Dependencies treebanks, dictionaries, and grammar-rule banks. We evaluate these traces in three settings: in-context learning (ICL), supervised fine-tuning (SFT), and reinforcement fine-tuning (RFT), on Xibe and Chintang as test cases. Our results show that linguistic reasoning traces are most effective as inference-time guidance: in ICL, reliable linguistic reasoning traces substantially improve translation performance across models, languages, and metrics. In contrast, using these traces as training data yields smaller and less consistent gains, as models learn the format but often generate error-prone content. These findings suggest that LLMs can leverage grammatical information for low-resource MT when given reliable linguistic analyses, while learning to generate such analyses remains a major bottleneck.

The growing demand for scalable psychological counseling highlights the need for high-quality, privacy-compliant data, yet such data remains scarce. Here we introduce MAGneT, a novel multi-agent framework for synthetic psychological counseling session generation that decomposes counselor response generation into coordinated sub-tasks handled by specialized LLM agents, each modeling a key psychological technique. Unlike prior single-agent approaches, MAGneT better captures the structure and nuance of real counseling. We further propose a unified evaluation framework that consolidates diverse automatic metrics and expands expert assessment from four to nine counseling aspects, thus addressing inconsistencies in prior evaluation protocols. Empirically, MAGneT substantially outperforms existing methods: experts prefer MAGneT-generated sessions in 77.2% of cases on average across the nine aspects over the strongest baseline, and sessions generated by MAGneT using Llama3-8B-Instruct backbone yield 3.2% higher general counseling skills and 4.3% higher CBT-specific skills on cognitive therapy rating scale (CTRS). An open source Llama3-8B-Instruct model fine-tuned on MAGneT-generated data also outperforms models fine-tuned using baseline synthetic datasets by 6.9% on average on CTRS. We make our code, data and fine-tuned model public.

Mixture-of-Experts (MoE) models exhibit striking performance disparities across languages, yet the internal mechanisms driving these gaps remain poorly understood. In this work, we conduct a systematic analysis of expert routing patterns in MoE models, revealing a phenomenon we term *Language Routing Isolation*, in which high- and low-resource languages tend to activate largely disjoint expert sets. Through layer-stratified analysis, we further show that routing patterns first converge and then diverge across model depth. Building on these findings, we propose **RISE** (routing isolation-guided subnetwork enhancement), a framework that identifies and adapts language-specific expert subnetworks. RISE uses a tripartite selection strategy: specificity scores select language-specific experts in shallow and deep layers, while overlap scores select universal experts in middle layers. By training only the selected subnetwork while freezing the rest of the model, RISE improves low-resource performance while preserving other-language capabilities. Across 10 languages, RISE achieves target-language F1 gains of up to 10.85% with minimal cross-lingual degradation.

Post-training LLM compression often relies on low-rank approximations, which force all columns of a projection matrix to share a single low-dimensional subspace. We propose *CoSpaDi*, a training-free compression framework that replaces this single-subspace assumption with a union-of-subspaces model via sparse dictionary learning. *CoSpaDi* factorizes each weight matrix into a dense dictionary and column-sparse coefficients, allowing different columns to select different subsets of dictionary atoms at the same storage budget. To preserve model behavior, we use calibration activations to transform functional reconstruction into a standard dictionary learning problem. Across Llama and Qwen models, CoSpaDi improves accuracy-compression and perplexity-compression trade-offs over SVD-based and structured pruning baselines at 20-40% compression ratios, while naturally supporting sparse-dense computation and post-training quantization of sparse coefficients.

Frontier scientific reasoning remains a major challenge for large language models (LLMs), where even the strongest commercial systems fall short of expert-level performance. A closer look at model behavior reveals substantial complementarity that single-model evaluation hides: different frontier models excel on different question types, and no single model captures the full picture. We present SciOrch, a framework that trains a lightweight 8B model to orchestrate frontier LLMs for scientific reasoning. The orchestrator decomposes each question, delegates sub-problems to selected commercial models through API calls, and synthesizes a final answer. Training such an orchestrator is fundamentally harder than conventional agentic RL: each action triggers an API call that is expensive in both dollar cost and latency, making standard online rollouts infeasible. We address this with MCTS-based approach, producing diverse orchestration trajectories, extracting per-node single-turn samples, and optimizing the orchestrator with GRPO-style training. On a 240-question test set spanning SGI-Reasoning and Scientists’ First Exam, SciOrch reaches 56.66% average accuracy, outperforming the strongest single commercial model by 3.74% and the strongest multi-agent baseline by 3.33%. It also attains the best accuracy on both SGI and SFE with less than half the API cost of typical multi-agent methods.

We introduce KFinEval, a benchmark for evaluating large language models (LLMs) in the Korean financial domain, with about 1K expert-curated instances across foundational financial knowledge, regulation-grounded procedural reasoning under noisy contexts, and financial toxicity under adversarial prompts. Existing financial benchmarks are largely English-centric and emphasize numerical or surface reasoning, providing limited coverage of regulation-grounded decision-making and domain-specific safety. Evaluating proprietary and open-weight LLMs, we observe substantial disparities and trade-offs among knowledge accuracy, procedural reasoning, and safety-aligned behavior; notably, strong capability does not consistently translate into safer responses under adversarial financial scenarios. KFinEval provides a practical Korean-regulation-grounded benchmark for diagnosing and improving financial LLM reliability and safety. Code and resources: https://github.com/sklim84/KFinEval.

Language models often answer harmful requests in low-resource languages (LRLs) that they refuse in high-resource languages (HRLs). Across three instruction-tuned models and 23 languages, harmful refusal falls from 87.9% in HRLs to 43.9% in LRLs, while harmless refusal remains low. A common explanation is that models represent harmfulness weakly in LRLs. We test whether harmfulness is instead represented but does not reliably produce refusal. Across three models, a harmfulness direction learned from HRL activations still separates harmful from harmless LRL prompts, showing that complete absence of harmfulness information cannot explain many failures. However, harmfulness scores shift downward for LRL prompts, making harmful prompts less likely to reach the range associated with refusal. Motivated by this shift, we train a low-rank logistic classifier on HRL activations and calibrate its threshold with a few target-language examples. During generation, the classifier conditionally adds or ablates the HRL harmfulness direction. With the same HRL data and 32 target-language examples per class, CAST remains limited by low harmful refusal and AdaSteer by high harmless refusal, yielding mean refusal selectivity (Δ = harmful − harmless refusal) of 33.6 and 6.8, respectively. Our intervention reaches 54.5 while preserving MMLU utility. HRL-only calibration improves selectivity for Qwen and Gemma, whereas Llama benefits from target-language calibration. These results show that recalibrating existing representations can offer a training-free method for repairing low-resource safety failures.

Existing evaluations of harmful content detection rely predominantly on static benchmarks, which struggle to reflect the interactive adversarial ecosystem of real-world content platforms where users continuously revise their expressions in response to moderation feedback. This mismatch creates a significant performance gap between offline benchmark scores and online deployment effectiveness. To the best of our knowledge, we present EvoHarmBench, the first dynamic adversarial evaluation framework for content moderation systems. The framework employs an iterative optimization loop that evolves evasion strategies at the semantic-cluster level, while simultaneously optimizing for evasion success and human readability. We systematically evaluate LLM-based defense models which are widely used in real world moderation systems. The evaluation covers 229 semantic sub-clusters across five violation categories, derived from 5,002 real-world adversarial samples collected from content platforms. Our experiments reveal substantial vulnerabilities even in leading commercial systems: after twelve optimization iterations, Upon acceptance, we will release the full benchmark data, evaluation framework, and code to encourage a shift from static benchmarking toward dynamic adversarial evaluation in content safety research.

Given a line-art sketch and user-specified color intent, controllable colorization must assign intended colors to fine-grained regions while preserving the underlying line structure. Existing methods have improved visual quality, but precise local adjustment remains difficult. In this work, we argue that the key limitation is the absence of a persistent correction state. Such a state records user intent, localized mismatches, grounded edit regions, and preservation constraints to guide targeted repairs. To address this, we introduce PaletteAgent, an agentic, self-evolving correction framework for controllable line-art colorization. PaletteAgent translates user intent into structured color prompts, employs judgment agents to diagnose mismatches, grounds problematic regions, preserves correct fields, and adaptively decides whether to refine globally, repair locally, or stop. This iterative process naturally defines the GRPO training objective for optimizing the agent’s correction policy. Empirical results show that PaletteAgent outperforms line-art colorization and general image-editing baselines. Further ablations confirm the role of maintaining a persistent correction state for controllable colorization.

Role-based LLM multi-agent systems need adaptive role pools, yet adapting such systems is not merely a matter of prompt optimization: roles often carry structural obligations, including capability coverage, message compatibility, validation, final-answer aggregation, and parser-compatible output protocols. Existing systems either fix the role inventory and lose adaptivity, or allow unconstrained generation to induce role drift, removing structurally necessary roles and breaking answer contracts. We formulate this as contract-preserving role evolution, requiring every committed edit to preserve five structural contracts (capability, communication, validation, aggregation, output protocol). We instantiate this formulation in SERO, a Self-Evolving Role Orchestration framework that evolves a typed role-card pool through credit-guided retrieval, a credit-ranked communication DAG with a protected terminal aggregator and conditional validator repair, and a contextual-bandit controller whose LLM-proposed edits are committed only when they preserve the contracts and improve task score. Experiments on real-world reasoning benchmarks across three LLM backbones confirm the value of contract-preserving role evolution.

Out-of-context (OOC) misinformation often arises when an authentic news image is paired with misleading text, making the cross-modal evidence used by a detector an important object of interpretation. We present SparseAlign-OOC, a two-level sparse autoencoder (SAE) framework for auditing NewsCLIPpings CLIP-embedding detectors through recurring image-text alignment features. A pair-space SAE maps detector inputs into sparse coordinates whose high-activation examples form human-auditable concept groups, while a hidden-state SAE traces sparse features associated with detector decisions. In a 250-item blind human audit with three annotators, selected SAE groups yield 18/50 coherent-alignment concepts, compared with 9/50 for prototype clustering, 7/50 for PCA, 4/50 for a dense autoencoder, and 0/50 for random groups. A 150-item replication audit on a stronger deep MLP preserves the same ordering, producing pooled results of 29/80 SAE alignment groups versus 15/80 for the strongest baseline. Coverage analysis shows that the presence of concepts increases with activation strength. In contrast, direct pair-space interventions show that human-labeled alignment features produce larger detector changes than topical, incoherent, or random SAE features. In the NewsCLIPpings CLIP-embedding setting, SparseAlign-OOC provides a validated audit pipeline for exposing recurring alignment factors that are both human-auditable and functionally connected to detector behavior.

Reinforcement learning with verifiable rewards has become a standard recipe for improving the reasoning abilities of large language models. Existing algorithms face a tradeoff between computational efficiency and sample efficiency in value estimation and policy learning. We introduce BASIS, a critic-free post-training algorithm designed to address this tradeoff. At each online training step, BASIS samples only one rollout per prompt, but leverages rich information across prompts in the entire batch to improve value function estimation. Our experiments demonstrate that BASIS reduces MSE in value function estimation by 69% compared to REINFORCE++, a representative single-rollout baseline, and achieves lower MSE with one rollout than group mean estimators with 8 rollouts. This improvement in value estimation translates to better policy optimization: using substantially less training time, BASIS achieves performance close to multi-rollout GRPO-type baselines and often outperforms single-rollout REINFORCE-type baselines.

Open-weight large language models (LLMs) can generate function-level programs from natural-language prompts, but plausible candidates still fail on hidden semantics and repeat mistakes across repair attempts. We present ReMCTS, an execution-grounded, memory-augmented, LLM-guided MCTS-style search framework. It organizes program candidates as tree states, retains branch-local debugging context, retrieves failure experience across branches, and distinguishes failed checks from unavailable evidence. On HumanEval and MBPP-Sanitized, visible-test ReMCTS improves over direct generation in 8 of 10 model–dataset pairs under held-out evaluation, whereas proxy-only search is less stable. Controlled tree-search, sampling, repair, and memory ablations characterize the source and limits of these gains. A 30-task HumanEval-X C++ pilot further demonstrates compatibility with compiler-backed execution, but does not constitute a broad multilingual evaluation.

As LLMs are increasingly integrated into agentic systems, they must adhere to dynamically defined, machine-interpretable interfaces. We evaluate LLMs as in-context interpreters: given a novel context-free grammar, can LLMs generate syntactically valid, behaviorally functional, and semantically faithful outputs? We introduce RoboGrid, a framework that disentangles syntax, behavior, and semantics through controlled stress-tests of recursion depth, expression complexity, and surface styles. Our experiments reveal a consistent hierarchical degradation: LLMs often maintain surface syntax but fail to preserve structural semantics. Despite the partial mitigation provided by CoT reasoning, performance collapses under structural density, specifically deep recursion and high branching, with semantic alignment vanishing at extreme depths. Furthermore, Alien lexicons reveal that LLMs rely on semantic bootstrapping from keywords rather than pure symbolic induction. These findings pinpoint critical gaps in hierarchical state-tracking required for reliable, grammar-agnostic agents.

Low-budget experimental optimization requires making consequential experimental decisions from a small number of observations. Bayesian optimization (BO) and related black-box optimizers provide strong proposal mechanisms, but they often hide the intermediate decisions that matter to experimental users. Direct LLM recommendation can leverage broad domain priors and natural-language reasoning, but is difficult to control, audit, or align with the statistical search guarantees of the underlying optimizer. We introduce TRACE, a traceable agentic co-optimizer that keeps the optimizer as the search backbone while wrapping it in an auditable agentic decision layer. TRACE represents agent interventions as bounded actions backed by reusable skills, organizes external and experiential evidence through an Evidence Skill System, and records every optimization step as a decision trace. On low-budget reaction-optimization benchmarks, TRACE improves over the corresponding BO-only optimizers and over LLM-assisted and tree-search baselines on most evaluated metrics. These gains suggest that our agentic system can effectively support experimental optimization by providing structured, traceable decision support around its actions, evidence use, and intervention logic.

Despite advancements in controllable text-to-music (TTM) generation, existing models struggle with cross-domain voice cloning for singing due to inadequate modeling of speech-derived vocal characteristics, inherently limiting precise timbre transfer. To achieve precise and controllable music generation conditioned on the timbre of reference speech, we propose S2M-Inject, a cross-domain framework driven by reference speech. Incorporating a fixed timbre embedding strategy, S2M-Inject injects speaker representations into an instruction-guided generation pipeline. This effectively preserves target vocal attributes while enabling comprehensive natural-language control over lyrics, genre, instrumentation, rhythm, melody, and mood. Extensive experiments demonstrate that the framework achieves effective cross-domain voice cloning, significantly enhancing vocal similarity while maintaining high-fidelity audio quality in both Chinese and English tasks. This work extends voice cloning into the musical domain, offering a rigorous paradigm for personalized music generation. Audio samples are available at https://anonymous.4open.science/w/S2M-Inject-975.

While large language models (LLMs) have exhibited remarkable performance across diverse natural language processing tasks, their inherent propensity to generate hallucinations poses a critical challenge to reliable deployment in high-stakes scenarios. Existing hallucination detection methods often rely on static single-layer representations or aggregated features across all layers, which often fail to capture fine-grained factual inconsistencies and thus exhibit limited generalization capability. To address this limitation, this paper proposes TemHal (Temporal Hallucination Detector), a novel detection framework that models the cross-layer evolutionary trajectory of internal representations. Specifically, TemHal tracks the dynamic representational progression of key factual tokens across selected intermediate MLP activations and employs a bidirectional temporal modeling network to identify representational patterns that differentiate factually consistent outputs from hallucinated ones. Extensive experimental results demonstrate that the proposed method achieves consistent performance improvements across multiple mainstream model architectures and diverse benchmark datasets, while also exhibiting robust cross-dataset generalization across multiple LLM backbones. Furthermore, ablation studies validate the contribution of each architectural component and confirm the importance of temporal modeling and intermediate-layer focusing for detection performance.

The performance of large language models is tied to the quality and organization of their training data. However, existing data metrics emphasize surface-level properties, such as sequence length or model-based quality scores, and provide limited insight into how the internal organization of data relates to optimization. We study the structural fabric of training data, defined as the pattern of semantic dependencies within a sequence. We quantify this property using dependency density, a normalized measure of effective segment-level dependencies. Across data length, domain, and format, we find that dependency density varies with data configuration: code shows higher dependency density than natural language within long-context data, and instruction-tuning mappings show higher dependency density than raw intra-input modeling. We show that coherent structural fabric is associated with a lower gradient effective rank, indicating a more concentrated optimization geometry with reduced gradient interference. We observe that this low-rank optimization regime is associated with a more stable intrinsic dimension, preserving a compact representation manifold that may mitigate representation noise and support downstream performance. These results provide a geometric perspective on how data structure relates to LLM training, linking data configuration, optimization dynamics, and model behavior through the lens of structural fabric.

Tool-Integrated Reasoning (TIR) is a fundamental capability for LLM agents to solve complex tasks by interacting with external tools iteratively. Reinforcement Learning (RL) has become the dominant paradigm for enabling this capability. However, existing approaches typically assign uniform trajectory-level advantages and treat all correct tool calls equally, ignoring the varying difficulty and learning value across trajectories and reasoning steps. This can lead to imprecise learning signals that do not adequately distinguish between trivial and challenging tool-use patterns. To address this limitation, we propose HiDiffTIR, a Hierarchical Difficulty-aware policy optimization framework for multi-turn TIR. HiDiffTIR performs difficulty-aware credit assignment at both trajectory and turn levels, enabling the policy to focus on more informative trajectories and harder reasoning steps. Notably, this fine-grained optimization is achieved without additional supervision, relying solely on group-level statistics derived from standard RL rollouts. Extensive experiments on three tool-using benchmarks demonstrate that HiDiffTIR consistently improves multi-turn TIR performance and tool invocation accuracy over strong RL baselines, highlighting the necessity of difficulty-aware credit assignment for effective policy optimization in tool-integrated LLM agents.

Large language models often answer the same multiple-choice question inconsistently when it is posed under support-oriented and elimination-oriented framings. We investigate whether these discrepancies arise from different internal representations induced by the two framings. We introduce a dual-framing protocol with minimally varied prompts that use either support- or elimination-oriented framing while keeping the evaluation target fixed. To probe the internal computation, we append an untrained special token, [STATE], and treat its residual-stream activation as an intervention interface. Across both models, the two framings induce separable [STATE] activations concentrated in intermediate layers. Swapping these activations between paired prompts systematically changes predictions and improves cross-framing agreement, providing intervention-based evidence that the activations are behaviorally relevant. Beyond instance-level substitution, mean-difference steering directions derived from the dual-framing contrast exhibit more bounded layer-wise responses than matched contrastive activation addition directions under the evaluated protocol.

Structure-based drug design (SBDD) requires ligands that satisfy both 3D target affinity and 1D chemical validity. Existing controllable generation methods often rely on task-specific fine-tuning or externally imposed sampling-time guidance, adding cost and potentially conflicting with evolving 3D geometric constraints. We propose LiFT, a language-informed cross-modal framework built on Flow Matching for trend-guided 3D molecular generation across both de novo design and scaffold hopping. LiFT uses a “Sense-Evolve-Assemble” agent to generate target-aware SMILES as intermediate chemical conditions, from which a pre-trained chemical foundation model extracts continuous semantic priors. These priors are integrated into geometric generation through a lightweight semantic projector with zero-initialized adaptive normalization for stable cross-modal conditioning. We further introduce a Self-Conditioned Decoupled Router (SCDR), which modulates the velocity field according to intermediate structural states during ODE integration. Experiments on CrossDocked2020 show that LiFT achieves competitive distribution matching while improving medicinal chemistry metrics and maintaining competitive structural validity under task-steering settings without additional generator fine-tuning. Our results suggest that language-derived chemical priors provide effective trend-level guidance for 3D molecular generation. Code and released artifacts are available at https://github.com/kasurl/LiFT.

Masked Diffusion Models (MDMs) have emerged as a compelling paradigm for language modeling, offering the capability for efficient parallel text generation. However, existing parallel sampling methods typically update multiple masked tokens independently and ignore the complex mutual dependencies among the masked tokens. This independent updating mechanism lacks global coordination and might lead to incoherent outputs. To address this limitation, we propose Representation-based Masked Diffusion Model (RMDM), a framework that leverages the text representation to explicitly encode global semantics and help to parallel update tokens more precisely. Specifically, we first encode text into a continuous semantic space using a pretrained encoder and learn an invertible transformation that normalizes the representation distribution to a Gaussian prior, facilitating efficient sampling during generation. Conditioned on this latent semantic representation, we train a masked diffusion model to learn the conditional text distribution, where the representation serves as global semantic guidance to coordinate parallel token updates and faithfully approximate the target distribution. Empirical results demonstrate that RMDM significantly improves generation quality, particularly in aggressive few-step sampling regimes.

Deploying Large Language Models (LLMs) in real-world dynamic environments raises the challenge of updating their pre-trained knowledge. While existing knowledge editing methods can reliably patch isolated facts, they frequently suffer from a reasoning gap, where the model recalls the edited fact but fails to utilize it in multi-step reasoning chains. Our analysis shows that multi-hop reasoning relies on sparse circuitry beyond direct factual recall, with functionally differentiated components contributing to intermediate processing and final-answer prediction. To bridge this gap, we introduce MCircKE (Mechanistic Circuit-based Knowledge Editing), a mechanistically grounded framework that enables a map-and-adapt editing procedure. MCircKE first identifies the causal circuits responsible for a specific reasoning task, capturing both the storage of the fact and the routing of its logical consequences. It then surgically update parameters exclusively within this mapped circuit. Extensive experiments on the MQuAKE-series benchmarks demonstrate the effectiveness of the proposed method for multi-hop reasoning in knowledge editing.

Determining whether a program terminates is a central problem in computer science. Turing’s Halting Problem established termination as undecidable, showing that no algorithm can universally determine termination for all programs and inputs. Hence, verification tools approximate termination, sometimes failing to prove or disprove; these tools rely on problem-specific architectures and are usually tied to particular programming languages. Recent advances in LLMs raise a natural question: To what extent can they reason about program termination? We evaluate frontier LLMs on a diverse set of C programs from the International Competition on Software Verification (SV-Comp) 2025. Our results show that GPT-5 and Claude Sonnet-4.5 achieve scores comparable to top-ranked verification tools (with test-time scaling). However, while models often correctly infer whether programs terminate, they frequently fail to construct a witness as formal proof, revealing a gap between semantic recognition and symbolic proof generation. Performance further degrades as code length increases. Beyond witness automaton graphs, we introduce a divergence-precondition formulation that characterizes non-termination conditions as logical constraints. We hope these findings motivate future research on real-world termination benchmarks, neuro-symbolic approaches that combine LLMs with symbolic verification methods, and, more broadly, LLM reasoning on other undecidable problems.

Multi-turn tool calling is a core evaluation scenario for large language model (LLM) agents. On public tool-calling benchmarks, open-weight models now approach or even surpass closed-source frontier models in aggregate accuracy. However, aggregate accuracy averages over many different multi-turn situations and obscures whether progress is balanced across them, masking failure redistribution. We propose an action-class-oriented diagnostic framework that decomposes failures into two orthogonal dimensions: capability and calibration. The framework operates over a four-class action space (tool_call/ask/refuse/confirm) and introduces a self-revealing upper bound, Acc ≤ GAR (Gold Action Recall), whose two violation directions separately localize action-class miscalibration and sub-action failures. We validate it on an extensive panel of tool-calling models that includes both open-weight families and closed-source frontier systems, across multiple multi-turn tool-calling benchmarks for cross-benchmark robustness. Across our panel the bottleneck is calibration, not capability: every model can produce all four classes, but which one it fires in which situation is highly model-specific. On no-tool-call scenarios, a top-ranked family’s GAR is roughly one-quarter that of size-matched peers, a 64.5 percentage point (pp) gap, indicating its ranking reflects benchmark-distribution fitting rather than genuine in-context calibration. We further show the calibration profile can be reshaped through context manipulation alone, with the reshaping direction strongly model-specific. We recommend that multi-turn tool-calling evaluations report per-category (GAR, tool-call-emit rate, Acc) triples, exposing the calibration structure aggregate accuracy hides. Our framework provides a falsifiable, reproducible diagnostic foundation for evaluation protocol design, training objective alignment, and inference-time calibration research.

Long-context LLM applications such as document summarization and multi-turn agents require generation from prefixes spanning tens of thousands of tokens, making decoding latency a major bottleneck. Speculative decoding (SD) reduces latency without changing model outputs, but its speedup depends on both accepted draft tokens and draft-step latency: Lightweight drafts are fast but lack the capacity to capture long-range dependencies, whereas strong independent drafts recover acceptance but incur growing KV-access cost at long prefixes. We introduce memory-augmented drafting for long-context SD, equipping a strong independent draft with compressed draft-side KV memory: A lightweight adaptor constructs and incrementally updates this memory to retain distant information and exact recent context. The target verifier retains its full KV cache and applies the standard accept/reject rule, preserving SD’s lossless guarantee. Experiments on Llama 3.1-8B and 70B targets at prefix lengths up to 32K show that our method reduces draft-side memory by over 70%. It achieves speedups of up to 2.08× and 3.33×, respectively, over autoregressive decoding.

Safety alignment in large language models (LLMs) remains brittle against a growing spectrum of attacks. Jailbreak attacks bypass safety mechanisms through crafted prompts, while neuron-level attacks directly prune safety-critical neurons post-deployment. Both exploit a common weakness: safety-relevant information concentrates in a sparse neuron subset. We present NeuronGuard, a fine-tuning-stage defense that simultaneously hardens LLMs against both attack classes by redistributing safety signals across a broader set of neurons. NeuronGuard dynamically identifies safety-critical neurons via periodically refreshed per-layer linear classifiers, forces refusal behavior under deliberate neuron ablation, and applies KL-divergence regularization for distributional consistency. A randomized gradient projection strategy preserves downstream task utility by resolving conflicts between the defense and task objectives. We provide a formal guarantee that NeuronGuard strictly reduces the attack success rate (ASR) upper bound, and experiments across three LLMs, six state-of-the-art attack strategies, and multimodal settings confirm near-zero ASR while maintaining task accuracy, including against white-box adaptive adversaries.

Deploying high-dimensional Vision-Language Model (VLM) embeddings is computationally prohibitive. While Matryoshka Representation Learning (MRL) allows elastic truncation, standard MRL optimizes dimensional hierarchies semi-independently. This lack of cross-scale structural constraints creates a severe topological disconnect causing representational collapse at lower dimensions and introduces optimization trade-offs that compromise high-dimensional capacity. To solve this, we propose **M-VLM2Vec**, a framework that shifts MRL from a compromised hierarchy to a synergistic geometric structure. Without additional supervision, we integrate two novel constraints into the multimodal contrastive objective: (1) Laplacian spectral alignment to preserve the global manifold topology across scales, and (2) Stiefel-constrained orthogonal projection to bidirectionally align local contrastive landscapes. By allowing nested spaces to mutually regularize, M-VLM2Vec eliminates competing gradients, preventing ultra-low-dimensional collapse while yielding striking, consistent performance gains across all embedding scales on the Multimodal Embedding Benchmark (MMEB).

Science as a body of knowledge is important for human learning, yet beyond factual accuracy, little research has focused on the capacity of Large Language Models (LLMs) to communicate science effectively to the public. Here, we develop and validate a computational framework for measuring science communication skills of LLMs. Using this framework, we evaluate different alignment approaches: fine-tuning using top-ranked human explanations, distilling synthetic explanations from a teacher model, and preference optimization. Our results show that annotators overwhelmingly prefer LLM-generated science explanations over top-ranked organic answers, with this preference primarily attributed to better structure and readability. We find that distillation with synthetic answers achieves the most significant and stable improvement, allowing a smaller open-source model to achieve statistical parity with the proprietary teacher model in blind side-by-side comparisons.

Standard large language model prompting treats geospatial inference as independent, instance-wise prediction, ignoring the fundamental spatial dependencies that govern geographic reality. Consequently, even advanced models struggle with spatial consistency and exhibit severe biases toward populous regions. To bridge this gap, we propose **GeoGR²** (**Geo**spatial **G**raph **R**efine **R**easoning), a framework that formalizes zero-shot geospatial prediction as an iterative message-passing process on a dynamically constructed graph. Unlike static retrieval methods, GeoGR² instantiates three dynamic operators via collaborating operators: (1) a **Topology Operator** that constructs graph topology to enforce the Spatial Markov property; (2) a **Feature Operator** that enriches nodes with task-relevant semantic covariates; and (3) an **Update Operator** that performs natural language message passing to iteratively minimize spatial discrepancy. Theoretically, we frame this refinement as a contraction mapping that approximates the fixed point of a global consistency equation. Empirically, we validate GeoGR² on diverse physical and socioeconomic tasks. Results demonstrate that by explicitly embedding geostatistical inductive biases, GeoGR² significantly outperforms standard prompting baselines, while effectively mitigating systematic geographic bias. Our framework leverages large language models’ intrinsic capacity for understanding spatial correlations through explicit topological scaffolding, without resorting to general graph reasoning paradigms.

Grammatical error correction (GEC) evaluation has traditionally relied on reference or edit overlap, which can penalize valid rewrites that differ from gold corrections. Reference-free metrics reduce this dependence, but evaluating whether a fluent output is a valid correction of the source remains challenging. We propose SURE, a source-conditioned reward evaluator trained on within-source preferences spanning minimal-edit and rewrite-oriented corrections. SURE jointly learns an overall reward with criteria-level supervision for grammaticality, faithfulness, and fluency, together with span-level grounding for source-side error resolution. Experiments on SEEDA show that SURE performs competitively against strong baselines, with particular gains on rewrite-style corrections and more disentangled criteria-level diagnostics. Our code is available at https://github.com/hayeonggg/SURE.

Large language model unlearning aims to suppress designated knowledge without degrading retained capabilities, yet existing approaches often fall into destructive over-suppression or conservative under-suppression. We introduce protected update routing, an intervention-guided framework that controls where forgetting updates flow within LoRA adapters. The method activates measurable adapter paths, identifies forget-specific rows from gradient attribution, intervention-induced loss changes, and forget–retain specificity, and applies localized suppression under retain-aware masking and stability safeguards. Multi-seed experiments on TOFU and WMDP show that protected routing provides a controllable forgetting–utility trade-off: it strengthens target suppression over NPO and RMU on TOFU while retaining higher downstream utility on WMDP. In a protocol-matched comparison with PALU, it achieves higher model utility at a similar forget-quality level. Analyses across MUSE-News, alternative routing granularities, and larger models further support the effectiveness and interpretability of separating route selection from update strength. Protected update routing therefore offers a practical and diagnosable interface for parameter-efficient unlearning.

Short-video platforms have become a primary news source for the public, which has also enabled the widespread dissemination of fake news videos. We study the task of fake news video detection and explanation (FNVDE). Existing methods face two critical limitations. First, commonly used frame selection strategies may omit veracity-relevant cues or provide insufficient temporal context for understanding news videos. Second, prior methods neglect either multimodal understanding or evidence retrieval. To address these limitations, we propose NVKE-CEI, a unified system that integrates a news video keyframes extraction method (NVKE) and an FNVDE framework leveraging both content and evidence information (CEI). NVKE selects keyframes based on chronological changes in combined visual and OCR-text similarity. CEI employs two specialized LLM-based fact checkers (content-based and evidence-based) whose outputs are fused by a lightweight judge model. Extensive experiments show that NVKE-CEI outperforms state-of-the-art baselines while generating high-quality content-grounded explanations.

Instruction-tuned LLMs faced with under-specified queries often commit to a single interpretation rather than ask for clarification, producing confidently wrong answers. In our experiments, prompting alone is insufficient: models either ask for clarification on every query or ask vague questions that fail to recover the missing information. Addressing this failure requires learning two coupled skills: when to ask rather than answer and how to ask a question that recovers the disambiguating information. Existing recipes either address only one of these skills or require a separately trained critic. We propose CIGAsk, an RL recipe that teaches both skills through two complementary reward signals within a multi-turn GRPO loop. Counterfactual Information Gain (CIG) compares the gold-answer log-likelihood under a frozen reference model with and without the user response, providing per-turn credit that guides how to ask. The Asymmetric Ambiguity Bonus assigns a signed reward at the terminal token based on the gold ambiguity label, guiding when to ask. Across three clarification benchmarks spanning table, passage, and open-domain QA, CIGAsk-7B outperforms the strongest external baseline despite using a smaller backbone. It also transfers across datasets without per-dataset tuning while preserving single-turn QA performance on out-of-distribution benchmarks.

LLM-powered agents struggle with personalization when users issue raw, underspecified queries. In this setting, agents must infer latent intent, extract preferences from noisy interaction histories, and select among competing alternatives. Existing benchmarks rarely test this capability, as they often rely on user-refined queries or simplified histories. We introduce personalized product search (PPS), a testbed for agentic personalization under raw queries and diverse histories. We construct **A**gent **Pe**rsonalized **B**enchmark (**APeB**) from action logs, pairing underspecified intents with rich histories and user-viewed candidate items. Evaluating state-of-the-art LLMs with multi-step agent workflows, we find that models handle explicit queries well but struggle with early-stage queries requiring intent and preference discovery. Rubric analysis attributes this gap mainly to ineffective history use. A simple history-aware query-refinement pipeline, VQRA, yields consistent gains, highlighting the need for dedicated history-utilization modules in personalized agents.

Large language models (LLMs) are effective rerankers, but pointwise reranking repeatedly processes the same document across different queries, causing substantial redundant document-side computation. We propose DoPR, a compressed document prefix framework that decouples offline document processing from online reranking. DoPR first selects query-independent document representations and converts them into compressed document prefix states, which are precomputed offline and reused whenever the document is retrieved. During online reranking, the model scores each query-document pair by processing only the query and scoring token, with document information supplied by the stored prefix states. This design reduces online cost through both document-side compression and cross-query prefix-state reuse. Experiments on TREC DL, BEIR, and BRIGHT with Qwen3 models from 0.6B to 8B show that DoPR achieves up to 8.0× online document-side memory reduction and up to 8.04× latency speedup, while retaining 97.1%-99.5% of the average NDCG@10 of matched full-document rerankers.

Autoformalization aims to produce formal statements that compile and faithfully preserve the intended meaning of informal mathematics. Yet standard single-output evaluation collapses this many-to-many structure into a single prediction. For downstream proving, this granularity is too coarse: a formal statement is not merely a faithful translation endpoint, but also a prover-facing interface whose structure can alter proof search under a fixed budget. We therefore recast autoformalization as budgeted test-time search: FormalEvolve maintains a compilation-feasible archive for reuse and returns a deduplicated, semantically accepted repertoire for evaluation and downstream proving. It expands the archive with LLM-driven mutation, crossover, bounded patch repair, and symbolic abstract syntax tree (AST) rewrites for structural diversity. Under a generator-call budget of T=100 with a fixed LLM semantic judge, FormalEvolve reaches SH@100 of 58.0% on CombiBench and 84.9% on ProofNet, improving over all no-archive controls while reducing the cross-problem concentration of semantic successes. Under a fixed B=64 prover budget, these repertoires improve theorem-complete proving over the matched no-archive control. Additional stronger-base statement-generation experiments show that archive-search gains persist with stronger seed and repair models.

Writing proficiency manifests in how students develop content, organize ideas, choose words, and use language. Despite growing interest in LLM-based student simulation, whether LLMs can reproduce such multidimensional variation in extended writing remains largely unexplored. In this work, we explore if language models can realistically simulate student writing, and introduce SWIM, a task that formulates Student Writing sIMulation as proficiency-conditioned essay generation. We evaluate prompting, supervised fine-tuning (SFT), and reinforcement learning (RL) methods for writing simulation using automated essay scoring as a measure of profile alignment. Extensive experiments reveal that prompting provides limited proficiency control, even for strong proprietary LLMs with rubric-grounded strategies. In particular, while models can adjust content-oriented traits, they struggle to reproduce the lexical, grammatical, and organizational variation in different proficiency levels. SFT substantially improves alignment, while RL with the proposed proficiency-alignment reward yields further gains across all writing traits and essay prompts. Our findings suggest that explicit supervision enables substantially stronger profile alignment than prompting alone, while authentic low-proficiency writing remains challenging to reproduce.

While large language models (LLMs) have increasingly been applied to hate speech detoxification, the prompts often trigger safety alerts, causing LLMs to refuse the task. In this study, we systematically investigate false refusal behavior in hate speech detoxification and analyze the contextual and linguistic biases that trigger such refusals. We evaluate nine LLMs on both English and multilingual datasets, our results show that LLMs disproportionately refuse inputs with higher semantic toxicity and those targeting specific groups, particularly nationality, religion, and political ideology. Although multilingual datasets exhibit lower overall false refusal rates than English datasets, models still display systematic, language-dependent biases toward certain targets. Based on these findings, we propose a simple cross-translation strategy, translating English hate speech into Chinese for detoxification and back, which substantially reduces false refusals while preserving the original content, providing an effective and lightweight mitigation approach.

Open-source text-only translation LLMs have rapidly improved multilingual coverage and translation quality, but their unimodal design limits their use in multimodal translation. In speech translation, they are commonly used in cascaded ASR–MT pipelines, which introduce additional latency and prevent the translation model from directly using visual context. Conversely, pretrained multimodal foundation models provide strong audio-visual perception but generally lag behind specialized translation LLMs in multilingual generation quality. We introduce OmniFusion, a modular architecture that connects a frozen multimodal foundation model to a translation-specialized LLM through a lightweight gated fusion module. OmniFusion supports speech-to-text, speech-and-image-to-text, and text-and-image-to-text translation within a single framework. Experiments with Qwen2.5-Omni-7B and Seed-X-PPO-7B show that selected early and middle MMFM layers provide more transferable representations than late layers. On MCIF, OmniFusion improves the latency–quality trade-off over a matched cascaded SimulST pipeline when evaluated with Local Agreement decoding, while remaining competitive in offline ST and reducing severe XCOMET-XL errors. On CoMMuTE, it achieves competitive visually grounded caption translation performance.

Multimodal review generation is essential in academic research. However, existing LLM-based models often struggle to pinpoint specific details within papers, yielding generalized, unsupported critiques that compromise final decisions. To address these limitations, we propose AnchorReviewer, a coarse-to-fine, evidence-grounded multimodal review generation framework. Central to our architecture is an integrated pipeline featuring an Inline Retrieval-Augmented Generation (InRAG) module. Combined with our Unified Embedding Module, InRAG facilitates the high-precision extraction of multimodal evidence chunks, ensuring our detailed reviews are strictly anchored to the original content. Furthermore, we incorporate a Reflection Mechanism to refine initial outputs and maximize decision accuracy. To support model evaluation and our Progressive Training Strategy, we construct the large-scale AnchorReview dataset, comprising diverse multimodal review samples. Experimental results demonstrate that AnchorReviewer significantly improves review reliability, excelling in critical evaluation, multimodal understanding, and decision consistency. Our code and dataset will be publicly available.

Multi-agent large language model (LLM) systems often rely on independent solvers, critique, debate, or answer aggregation. While these designs increase diversity and feedback, they do not necessarily create effective group reasoning: information may be poorly shared, expertise underused, discussions may drift, and consensus may emerge before unresolved issues are examined. We introduce AgendaFlow, a meeting-science-inspired framework that treats multi-agent reasoning as a structured meeting rather than parallel answering or generic debate. AgendaFlow routes each task to a suitable agenda, instantiates complementary functional roles, and uses a moderator to manage phases, speaking order, shared artifacts, unresolved issues, and completion checks. This enables agents to pool information, surface constraints, compare alternatives, repair errors, and converge through a managed collective decision process. Across five reasoning, planning, and rule- or evidence-based benchmarks, AgendaFlow consistently outperforms single-agent, sampling-based, debate-style, and multi-agent orchestration baselines. Ablations, token-matched comparisons, and process-level analyses suggest that the gains come from meeting-level organization rather than agent diversity or additional computation alone, highlighting meeting design as a missing dimension in multi-agent LLM systems. Code is available at: https://github.com/sunming-code/agendaflow.

Hallucination remains a critical challenge in large language models (LLMs), where early factual errors compound through autoregressive generation in a snowballing effect that neither post-hoc correction nor weight-level intervention can effectively preempt. We propose DescaPE (DEcoding Signal Control Against Path Error-snowballing), a decoding framework that leverages internal model signals to suppress hallucination-prone trajectories at inference time. Through sliding-window MLP ablation, we identify a factual-salient layer span within LLMs whose derived signal is selectively elevated for factual tokens and exhibits anomalous spikes at hallucination-prone steps. We train a lightweight probe to approximate this signal from a single forward pass and integrate it into candidate scoring to penalize high-risk continuations while rewarding factually grounded ones. Experiments across five factuality benchmarks on three LLMs demonstrate that DescaPE achieves factuality improvements over decoding-time baselines in multiple settings, while incurring only 1.10x latency overhead in our efficiency evaluation. Our code is available at https://github.com/hayeonggg/DESCAPE.

The integration of large language models (LLMs) into recommender systems has garnered attention due to their ability to leverage rich semantic representations. However, these models suffer from a limitation, knowledge bias, a systematic tendency to over-recommend items dominated by pretraining knowledge, regardless of actual user interests. Although recent studies have focused on fine-tuning LLMs to better align with the collaborative data, our investigation reveals that such fine-tuning fails to fundamentally correct this deep-seated bias. In this work, we propose ReKnow, a lightweight and efficient decoding-time alignment method to mitigate knowledge bias. Specifically, we quantify LLMs’ knowledge towards biased items, then realign output probabilities to match the target data distribution. For validation, we provide both theoretical justifications and empirical results, including evaluation on two datasets, demonstrating that our  enhances the quality and diversity of LLM-based sequential recommendation. We have release our code[<https://anonymous.4open.science/r/ReKnow-B17E>] for others to reproduce the results.

Reinforcement learning (RL) holds immense promise for enhancing the reasoning capabilities of diffusion large language models (dLLMs). However, progress is fundamentally constrained by a dual misalignment between authentic generation trajectory and the gradient update process: (i) Process-reward misalignment. Sparse, terminal rewards are indiscriminately assigned to all intermediate steps of the generation process, failing to provide discriminative credit assignment. (ii) State-trajectory misalignment. Policy updates are often diverted toward artificial, out-of-trajectory states, squandering gradients on less informative samples. To address these limitations, we introduce Process Aligned Policy Optimization (PAPO), a novel framework that holistically aligns the RL update with the dLLM’s generative trajectory via Step-Aware Process Rewards (SPR) that transform sparse terminal rewards into dense, step-wise credit, and Entropy-Guided Historical Re-enactment (EHR) that replays authentic trajectories at high-uncertainty steps. Extensive experiments on four benchmarks demonstrate that PAPO significantly outperforms baselines, achieving gains of 4.5% on GSM8K, 4.8% on MATH500, 42.2% on Countdown and 16.1% on Sudoku.

Large Reasoning Models (LRMs) have demonstrated exceptional capabilities in complex problem-solving, largely driven by emergent reflection and backtracking mechanisms. However, understanding these mechanisms through mechanistic interpretability remains challenging. Current sparsity-based approaches rely heavily on heuristic keywords (e.g., "wait") and focus on isolated features, lacking a holistic view of how these features collaborate. In this paper, we propose a novel interpretability framework that bridges the microscopic and macroscopic analysis of Sparse Auto-Encoder (SAE) features. By leveraging an auxiliary LRM to robustly annotate reasoning trajectories, we introduce a comprehensive "Reflection Score" to quantify a feature’s relevance to reflection without keyword bias. Our geometric analysis of these features reveals that LRMs encode reflection not as discrete, isolated vectors, but within continuous regions of the hidden state space. Furthermore, we uncover an intrinsic, difficulty-aware reflection representation mechanism: the model dynamically activates sub-regions based on problem complexity. Building upon these insights, we introduce a static, difficulty-informed steering strategy that suppresses superficial, simple-problem oriented reflection features during inference. Comprehensive evaluations across three LRMs and challenging reasoning benchmarks demonstrate that our intervention simultaneously improves reasoning accuracy by an average of 2.25% and reduces token consumption by 5.3%. Our findings provide a deeper geometric understanding of LRM reflection and offer an interpretable, scalable approach to optimizing LRM reasoning.

Graphical User Interface (GUI) Agents autonomously interact with software to fulfill user requests, where GUI navigation stands out as the most critical and challenging capability. Mastering this capability demands a complex synergy of step-wise decision-making, state-action alignment, and long-horizon planning. While directly mixing these corresponding navigation tasks seems intuitive to simultaneously acquire these skills, such a direct combination is severely bottlenecked by inconsistent optimization objectives and profound data heterogeneity. To overcome these barriers, we propose the MaP (stands for “Masked Trajectory Prediction”), a unified framework that seamlessly harmonizes divergent GUI navigation tasks. By modeling multi-turn GUI interactions as a trajectory and defining training objectives through component masking and prediction, MaP shifts the optimization from task-specific marginal distributions to a consistent objective. Furthermore, to handle the data heterogeneity across multiple navigation tasks, we design a role-aware adapter learning module that dynamically routes each token to a specialized representation space. Extensive experiments on five representative GUI navigation benchmarks demonstrate that MaP effectively mitigates gradient conflicts and significantly outperforms the direct mixture training, establishing a robust paradigm for multi-task GUI navigation.

Existing multiple-choice evaluations use small option sets and thus conflate several sources of failure when candidate spaces grow. We provide a controlled factorial decomposition of large-choice degradation across four axes: whether a question requires comparing its options, candidate cardinality (up to 1,000), spatial layout, and gold-answer position. Across six reasoning LLMs, comparative questions degrade far more than questions solvable without inspecting the options; dispersion amplifies this gap; and accuracy often collapses at late-middle and terminal positions rather than recovering in the classic lost-in-the-middle U-shape. We call this asymmetric, structure-dependent failure mode Lost in Choice. A replication on 500 real ToolBench tool-selection trajectories per condition reproduces the same hardest position on two additional model-task evaluations. Finally, an audit of reasoning traces yields a taxonomy of the heuristics and shortcuts models adopt as candidate spaces expand.

In post-training large language models (LLMs), supervised fine-tuning (SFT) provides expert guidance but can overfit, while reinforcement learning (RL) enables exploration but often suffers from cold-start and sparse-reward issues. Although combining SFT with RL is intuitive, naively mixing off-policy expert data into on-policy RL can induce distribution shift and destabilize optimization. We propose APEX, a framework that harmonizes off-policy guidance and on-policy exploration via a Cognitive State Awareness mechanism. Specifically, APEX leverages a perplexity-based Student’s -test to compare the sequence-level perplexity distributions of on-policy generations against candidate offline data. This comparison diagnoses whether failures reflect sampling stochasticity or missing knowledge. Guided by this diagnosis, APEX adaptively gates GRPO optimization and modulates SFT supervision. The framework preferentially selects offline samples that are statistically consistent with the current on-policy distribution to guide RL updates, while filtering out high-shift data and verifiably incorrect samples that could destabilize training. APEX also uses a dynamically weighted SFT loss when needed. Extensive experiments show that APEX improves over SFT-then-RL by +7.10 points in average pass@1 on six math benchmarks and by +15.50 points on three general reasoning benchmarks.

Unified streaming and offline ASR aims to deliver full-context accuracy and low-latency incremental decoding with a single model. Transformer/Conformer systems are effective, but long-form streaming inference still suffers from growing attention masks, historical caches, and memory states. Latency-Controlled BiMamba (LC-BiMamba) offers attention-free bidirectional modeling, yet its backward branch is confined to fixed chunks, tying the model to one latency. Extending it to dynamic chunk training expands mini-batches and slows training. We propose TC-BiMamba, whose Trans-Chunk formulation casts local-future backward modeling as sequence-level dynamic-chunk training over complete sequences. TC-BiMamba supports multiple chunk-wise latency settings, keeps the encoder/CTC path memory-stable on long-form inputs, and improves training scalability. On AISHELL-1, AISHELL-2, and LibriSpeech, it outperforms U2++ and remains competitive with externally reported LC-BiMamba results using fewer parameters, while achieving 1.3 times training speedup and 50% GPU memory reduction.

Despite the significant advancements of Large Language Models (LLMs), their factuality remains a critical challenge, creating a growing need for more nuanced factuality verification. Existing factuality verification methods do not capture graded judgments, even though factuality is better understood as a spectrum rather than a binary of right and wrong. To bridge this gap, we focus on graded factuality verification and propose AEScorer, an agentic evidence-grounded framework with two stages: agentic evidence acquisition and graded scoring. AEScorer first gathers and refines external evidence through agentic search, and then predicts a scalar factuality score to distinguish nuanced differences in factual correctness. We further construct GradedVeriBench, a benchmark for graded factuality verification spanning both general and multi-hop question answering. Experimental results on GradedVeriBench show that AEScorer substantially outperforms existing methods across both settings, demonstrating the value of coupling targeted evidence acquisition with graded scoring.

Effective collaboration in multi-agent systems hinges on communication. Existing studies analyze agent communication indirectly through task performance, such as improvements in reward and success rate, without explicitly characterizing communication itself. As a result, the quality of communication protocols remains difficult to evaluate and interpret, limiting their reliability, scalability, and generalizability. To better characterize communication quality, this survey focuses on communication alignment, which emphasizes semantic fidelity and convention consistency across agents. We review communication alignment from emergent communication in multi-agent reinforcement learning (MARL) to latent communication powered by Large Language Models (LLMs), highlighting existing methods and the emerging challenges introduced by natural-language communication. Finally, we discuss future directions toward more effective, interpretable, and generalizable multi-agent communication.

Robotic manipulation often requires actingon information that is no longer visible, yet Vision-Language-Action policies are usually evaluated when the current observation largely determines the next action. Existing robotic memory benchmarks expose this gap, but they still rely mainly on final task success and there-fore conflate forgetting with manipulation failure. We present MEMOBench, a benchmark for process level memory evaluation in robotic manipulation. MEMOBench includes 30 history dependent tasks, 1,500 expert demonstrations, and 4,200 executable checkpoint instances from 84 templates. Each checkpointpairs coarse to fine language with a simulator predicate and labels one memory operation:Storage, Update, or Compression. These annotations define Memory Storage Rate, MemoryUpdate Rate, and Memory Compression Rate,which measure memory fidelity alongside tasksuccess. Across standard and memory augmented VLA policies, the strongest memory module baseline reaches only 31.9%o average success rate, and high storage often coexists with weak update and compression. Checkpoint language also supervises semantic, contrastive, and framewise memory alignment objectives, yielding modest gains across differentmemory operations. MEMOBench provides a diagnostic evaluation suite and training supervision for memory grounded robotic policies. The project page is available at https://github.com/Collab-Gen/MEMoBench.

Personalized text generation for authors and literary writing is essential for applications such as adaptive writing assistants, creative support tools, and computational literary analysis. However, existing approaches to author modeling and personalization often represent writing behavior as independent labels, requiring large-scale corpus collection or fine-tuning for each author or stylistic category. Such formulations are costly, difficult to interpret, and poorly suited for generalizing across authors. Inspired by the Big Five theory of personality, we propose LiteraryBigFive, a framework that reframes authorial writing characteristics as coordinates within a unified and interpretable space. In this space, we derive each interpretable axis (e.g., Classicism, Emotionality) from activation-space contrasts between author-written and neutral passages, yielding distinct stylistic dimensions that allow texts or authors to be positioned within a five-dimensional system. Beyond localizing different authors, we further introduce an interpretable steering mechanism, which adaptively guides text generation toward target coordinates to perform author-personalized writing. Experimental results show that LiteraryBigFive improves authorial expressiveness while preserving semantic fidelity. The derived author per-axis scores strongly correlate with real-world literary consensus, offering transparent and interpretable explanations of author-specific generation behavior: https://github.com/Znull-1220/LiteraryBigFive.

Retrieval-augmented generation (RAG) has emerged as a critical paradigm for deploying reliable and knowledge-intensive large language model (LLM) applications. Its unique execution pipeline introduces large overhead in system retrieval, such as massive context loading and document dependencies. To reduce this overhead, existing work on RAG serving systems often relies on generic LLM inference traces or synthetic RAG datasets. However, these approaches fail to capture authentic temporal reuse patterns of RAG workloads, resulting in a disconnect between theoretical optimization and real-world deployment efficiency.

To bridge this gap, this paper introduces RAGPulse, an open-source RAG workload trace dataset that preserves authentic temporal information missing in existing traces. Collected from a production-grade university Q A system serving over 40,000 students and faculty members since April 2024, the dataset comprises 134K requests over 389 days. It unveils key RAG dynamics including time-sensitive document reuse, skewed access patterns over long time scales, and bursty temporal locality. By capturing these temporal reuse patterns, RAGPulse empowers the systems community to develop and validate next-generation optimization techniques, offering practical guidance for efficient RAG serving. The trace is available at https://github.com/flashserve/RAGPulse.

Multi-hop Retrieval-Augmented Generation can fail even when relevant evidence has been retrieved. The remaining diffculty lies in how evidence is connected across reasoning steps:an intermediate answer may be unsupported, at-tached to the wrong anchor, or propagated into a later step without valid evidence. We study this failure as Reasoning Chain Hallucination(RCH), where a fuent multi-hop reasoning pro-cess contains one or more ungrounded inter-mediate steps. To address RCH, we formulate Retrieve–Verify–Generate (R-V-G) for multi-hop RAG: retrieved evidence should support an explicitly verifed reasoning structure before it is allowed to guide generation. GROUNDRAG is an evidence-verifed framework that imple-ments this requirement. GroundRAG frst ac-quires the evidence needed to instantiate inter-mediate reasoning variables, then represents the reasoning process as a typed dependency chain.Each step is checked against retrieved passages using explicit grounding criteria; unresolved steps are repaired with queries formed from ver-ifed intermediate answers, while insuffciently grounded chains are discarded before genera-tion. This design separates evidence acquisi-tion from the decision of whether a reasoning chain is safe to use. On HotpotQA, 2WikiMul-tiHopQA, and MuSiQue, GroundRAG reaches 73.8%, 69.2%, and 44.3% LLM-judged accu-racy, respectively, outperforming the strongest baseline in our evaluation.

LLMs often struggle with modern Korean poetry, producing outputs that resemble “line-broken prose.” We address two coupled tasks: detecting whether a Korean poem is human- or LLM-authored, and guiding LLMs to generate poetry closer in form to human writing. We quantify the human–LLM gap along four form-level linguistic dimensions: output length (Volume), the diversity and connective use of line-final forms (Structure Variation), the irregularity of line lengths (Rhythmic Irregularity), and adherence to standard orthography (Normative Adherence). We operationalize these dimensions as five interpretable features. For detection, a logistic regression classifier over these five features attains an average AUC-ROC of 83.60 in zero-shot out-of-distribution detection across seven unseen LLMs, versus 75.84 for the strongest baseline in our comparison, KatFishNet, an absolute gain of 7.76 AUC points and a 10.23% relative improvement; one generator-specific punctuation pattern outside our taxonomy remains a boundary case. For generation, expert evaluation on GPT-5.2 prefers feature-guided poems over the unconstrained baseline, and analyses across GPT-5.2 and Gemini-3 show that targeted length, rhythm, and ending statistics move toward the human distribution. These results suggest that interpretable, language-specific features can bridge the diagnosis and guidance of LLM-generated poetry.

Textual user profiling is an effective paradigm for LLM-based recommendation, but existing methods usually treat profile generation as a one-shot summarization problem. This limitation becomes more severe in long-history recommendation, where early preference signals may be overcompressed and recent behaviors may be overemphasized. We propose LUPE, a Long-history User Profile Evolution framework for LLM-based recommendation. Instead of generating a static profile from the full history, LUPE maintains a persistent textual profile and updates it progressively over temporally ordered interaction chunks. This allows the model to preserve long-term preferences while incorporating new behavioral evidence. To learn how profiles should evolve, LUPE optimizes a profile evolution policy via reinforcement learning using feedback from a frozen downstream LLM-based recommender. This allows the model to directly align profile updates with recommendation utility rather than relying on heuristic or supervised profile targets. Experiments on Yelp, Amazon CDs and Vinyl, and Amazon Books show that LUPE consistently outperforms strong baselines on both rating prediction and ranking tasks. Further analyses show that explicit profile evolution is especially helpful for users with long and heterogeneous interaction histories. The code repository: https://github.com/LongHistoryProfile/long_history_profile.

Retrieval-Augmented Generation (RAG) improves generation quality by incorporating evidence retrieved from large external corpora. However, most existing methods rely on statically selecting top-k passages based on individual relevance, which fails to exploit combinatorial gains among passages and often introduces substantial redundancy. To address this limitation, we propose OptiSet, a set-centric framework that unifies set selection and set-level ranking for RAG. OptiSet adopts an “Expand-then-Refine” paradigm: it first expands a query into multiple perspectives to enable a diverse candidate pool and then refines the candidate pool via re-selection to form a compact evidence set. We then devise a self-synthesis strategy without strong LLM supervision to derive preference labels from the set conditional utility changes of the generator, thereby identifying complementary and redundant evidence. Finally, we introduce a set-list wise training strategy that jointly optimizes set selection and set-level ranking, enabling the model to favor compact, high-gain evidence sets. Extensive experiments demonstrate that OptiSet improves performance on complex combinatorial problems and makes generation more efficient.

Long-form narratives are challenging for long-context models because their structure is implicit: events, characters, and plotlines interact across hundreds of pages without the explicit cues that guide navigation in structured documents. We address this by constructing storyline trees, hierarchical representations that organize narratives from global themes and major plotlines to fine-grained events. We first segment chapters into contiguous narrative segments, or scenes, and use them as the basic units for tree construction. We then infer storyline trees through complementary top-down and bottom-up procedures that derive, refine, cluster, and summarize storylines at multiple levels of abstraction. We showcase the utility of this representation for question answering: storyline trees enable adaptive retrieval, allowing models to iteratively inspect high-level narrative structure and retrieve scene-level evidence on demand. Experiments on three long-context narrative QA benchmarks show that adaptive retrieval outperforms strong baselines, including post-trained long-context models and agentic chunk-based methods. Ablations confirm that scenes are more effective basic units than chapters or generic segmentation, and that gains persist under matched retrieval budgets.

Standardized interfaces such as the Model Context Protocol (MCP) make it easier to connect tools to agents, but connection alone is not orchestration. When many related tools are exposed as a flat callable palette, agents must identify relevant tools, resolve substitutes, pass intermediate outputs, and decide when a workflow is complete. We introduce RealFin, a finance-domain diagnostic benchmark for evaluating workflow construction under MCP-style flat tool exposure. RealFin contains 1524 tasks generated from 108 author-designed financial workflow templates over real market, macroeconomic, and company-data APIs. Each task is paired with an executable reference program that specifies one verified solution path, while the 85 agent-callable tools include explicit alternative, dependency, and related-tool metadata to represent non-unique paths, prerequisite tools, and near-neighbor ambiguity. This design supports path-independent answer scoring together with process-level evaluation of tool selection, ordering, parameter planning, entity grounding, temporal grounding, orchestration success, and error recovery. Across ReAct-style flat exposure, retrieval-based exposure, oracle-controlled variants, and progressive tool disclosure, current agents often achieve nonzero answer accuracy while failing to construct reliable workflows. These results suggest that MCP-style tool connectivity needs structured disclosure and workflow-aware evaluation, especially in finance where temporal constraints and overlapping tools are intrinsic.

Vertical-domain few-shot classification remains challenging for small language models, as limited supervision makes it difficult to acquire domain-specific decision knowledge. On-Policy Distillation (OPD) can improve teacher-guided adaptation by supervising student-generated rollouts, while GRPO-based reinforcement learning can further refine downstream predictions. However, existing KD-to-RL pipelines typically rely on globally fixed transition schedules, ignoring that different samples may require different amounts of teacher-guided acquisition before reward-driven refinement. We propose PIVOT (Perplexity-Informed Transition Optimization), a dynamic transition framework that routes samples between OPD and GRPO according to teacher-evaluated sequence perplexity. PIVOT moves low-perplexity samples to GRPO for reward-driven refinement while keeping high-perplexity samples under OPD for continued domain knowledge acquisition. Experiments on Banking77 and HWU64 show that PIVOT consistently outperforms continued OPD and globally synchronized OPD→GRPO baselines under the same optimization budget, achieving stronger downstream performance and more stable training dynamics.

Retrieval-augmented search agents answer multi-hop questions by repeatedly issuing search queries and accumulating evidence. This creates a stopping problem: after the necessary evidence has appeared, further retrieval often adds cost, latency, and distracting context rather than useful information. We frame stopping as evidence coverage rather than generator confidence, and introduce HALT, a lightweight verification-aware policy that leaves the search agent unchanged. Given expected hop claims, HALT stops only when cumulative evidence supports each required claim. Across three multi-hop QA benchmarks, HALT reduces redundant search while largely preserving exact match, with stronger savings under gold claims and more conservative behavior under generated claims. Baseline comparisons and ablations show that this behavior is driven by claim–evidence alignment rather than generic sufficiency, fixed stop positions, or lexical overlap. Open-corpus pilots further suggest that HALT abstains when coverage cannot be reliably verified. Overall, evidence coverage provides a practical runtime control signal for improving retrieval-augmented agents without retraining or modifying the host agent.

Many expert-domain decisions involve intermediate variables, yet standard LLM evaluation often measures only final-answer accuracy. We present a controlled study of intermediate-variable dependence in medical treatment selection, using diagnosis as an explicit variable between patient phenotype and treatment. We construct a six-option multiple-choice benchmark from PrimeKG-derived phenotype–disease–drug chains and evaluate six LLMs using behavioral diagnosis interventions, held-out QLoRA path supervision, external validation, and auxiliary layer-wise localization analysis. Providing the gold diagnosis improves average treatment accuracy from 49.9% to 84.9%, while incorrect predicted diagnoses reduce accuracy from 85.5% to 23.1%. A shuffled-diagnosis control yields 18.2% accuracy, suggesting that the effect depends on diagnosis correctness rather than merely adding disease-like text. Held-out QLoRA experiments show that path supervision improves treatment selection, with diagnosis-first full-chain and multitask objectives yielding the strongest gains. Center-of-Gravity localization further gives a modest diagnosis-first signal, which we interpret as a localization pattern rather than evidence of faithful clinical reasoning. Overall, in this controlled setting, final-answer accuracy can obscure failures in intermediate-variable dependence, while structured supervision can partially improve diagnosis-to-treatment behavior.

Preference tuning typically distributes adapter updates across every Transformer layer. We study how the placement of a fixed LoRA budget affects Direct Preference Optimization (DPO). Hierarchical Alignment partitions decoder depth into Local, Intermediate, and Global regions and restricts LoRA updates to a chosen contiguous block. Across Qwen and Llama models, three-seed sliding-window sweeps reveal broad low-loss basins around intermediate depth. Under exact trainable-parameter matching, Middle achieves lower held-out DPO loss than an all-depth allocation on Qwen2.5-7B and Llama-3.1-8B. Generation-level evidence is more variable. LLM judges and a blinded human study leave Selected versus All unresolved on Qwen; on Llama, two pairwise judges favor All while the human interval includes zero. Trained-adapter interventions show that learned depth components affect the optimization objective; removing the Llama Selected window also lowers generation quality. Across Qwen2.5 at 3B, 7B, and 14B, Middle versus All remains unresolved under Qwen3-Max; at 3B and 14B, both improve over Base. In our implementation, Middle also uses less peak memory and takes shorter steps than a parameter-matched all-depth allocation. Together, the results support intermediate depth as a practical starting point for fixed-budget preference tuning, though the best window and its behavioral value still need per-setting validation.

Emotional Support Conversation (ESC) systems aim to provide holistic support by balancing professional therapeutic competence with natural empathy. However, existing methods struggle to simultaneously achieve structured, stage-aware reasoning and seamless empathy-expertise alignment, often resulting in an artificial splicing of clinical strategies and generic reassurance. To overcome these limitations, we propose ESCRAG-R1, a unified framework that integrates retrieval-based psychological guidance into Group Relative Policy Optimization (GRPO). By incorporating retrieval into the reinforcement learning loop, ESCRAG-R1 transforms external knowledge into a robust learning signal that stimulates explicit internal reasoning prior to generation and fundamentally reshapes the model’s internal policy. To provide the reliable supervision required for this optimization, we construct ESC-Preference, a high-quality dataset based on a Client–Counselor–Judge evaluation framework that delivers precise, empathy-aware reward signals. Extensive experiments demonstrate that ESCRAG-R1 significantly outperforms existing baselines by mitigating superficial splicing and realizing a natural integration of professional guidance and empathetic expression. Code and datasets will be publicly released at https://anonymous.4open.science/r/ESCRAG-R1-83F1/.

The self-attention mechanism used in transformer language models implements a form of associative, cue-based memory retrieval, which has been argued to resemble human sentence processing. However, while it is well established that surprisal estimates from transformers are predictive of human comprehension difficulty, it has been less clear how these models’ internal memory operations might relate to human sentence processing. We propose to connect the two by measuring transformers’ associative memory updates, which are calculated as the difference between each word’s predicted and actual representation within self-attention. We compute memory updates from three types of language models: (1) standard transformers; (2) CBR-RNN, a cognitively motivated model using a single self-attention head; and (3) CBR-RNN-M, a novel architecture that replaces the self-attention in CBR-RNN with a bounded-size memory cell. Associative memory updates are found to predict naturalistic human reading times over and above surprisal baselines on several psycholinguistic corpora, with evidence for more consistent effects from CBR-RNN and CBR-RNN-M than from standard transformers.

While multimodal retrieval-augmented generation (RAG) systems increasingly rely on images as external knowledge sources, the introduction of poisoned visual evidence can severely compromise multimodal large language model (MLLM) generation. Unlike prior attacks that rely on altering textual metadata, we introduce Vis-Poison, a novel visual knowledge poisoning attack where the poisoned image itself is the attacker-controlled payload, without manipulating captions, summaries, metadata, or other associated text. Specifically, this attack is instantiated through an automated multi-agent method that constructs visually plausible poisoned images. To assess its impact, we evaluate Vis-Poison across two representative multimodal RAG pipelines, four embedding models, and six generation models. Empirically, Vis-Poison achieves an end-to-end attack success rate of 40.16% to 65.40% against 30k-entry multimodal knowledge bases in black-box settings. Moreover, Vis-Poison remains effective against various MLLMs that can answer correctly from parametric knowledge alone, with an average success rate above 60%. Code and data are available at https://github.com/SWUFE-DB-Group/Vis-Poison.

Reasoning is now central to modern foundation models, but remains underdeveloped in the audio modality. Unlike text transcripts, audio contains not only words but also vocal cues, background events, and temporal patterns, all of which can be necessary for reliable inference. This survey presents the first dedicated synthesis of audio reasoning in multimodal foundation models. We formulate audio reasoning as reasoning-augmented generation grounded in acoustic evidence, distinguishing it from conventional audio understanding. We then organize existing work around two core settings: **Audio-to-Text reasoning**, where models infer textual answers from acoustic input, and **Audio-to-Speech reasoning**, where models reason before or during speech generation under latency constraints. Finally, we summarize evaluation practices and challenges in acoustic grounding, shortcut learning, modality hallucination, and latency-constrained interaction.

KV cache compression is commonly evaluated by final-answer accuracy, implicitly assuming that preserving the answer also preserves the reasoning that supports it. We test this assumption for large reasoning models and show that it can fail: under compression, answer-level cues and evidential support can remain differentially recoverable. We study this failure with a controlled fixed-trace replay protocol, which holds reasoning content fixed and isolates whether compression preserves usable information from an already available trace. We evaluate ten token-eviction KV compression methods and one quantization method on three models across mathematical reasoning, scientific QA, clinical calculation, and long-context retrieval. We measure final accuracy, answer-chain consistency, and perturbation faithfulness. Across tasks, token-eviction methods can preserve competitive final-answer accuracy while substantially degrading chain support or perturbation faithfulness. We call this the answer-evidence gap. A coverage-preserving quantization control is substantially less affected, suggesting that the failure is tied less to KV memory reduction itself than to losing access to parts of the reasoning trace.

Online memes are often carriers of unsafe content such as disinformation and propaganda. Their rapid evolution in response to news events and strong dependence on contextual information make it difficult for platforms and moderators to detect and moderate them effectively. Moreover, existing moderation studies typically classify memes as safe or unsafe based on individual images, lacking the ability to perform broader provenance analysis, i.e., where a meme originates, how it spreads, what cultural symbolism it represents, whether it poses potential risks to society, etc. To bridge this gap, we introduce MemeTrace, a multimodal LLM-based agent that automatically traces the provenance and potential risks of emerging memes. We evaluate MemeTrace on Meme100, a dataset of 100 real-world memes, using both human annotation and automatic metrics. The results demonstrate that MemeTrace substantially outperforms the VLM baseline. More importantly, MemeTrace achieves higher accuracy in detecting unsafe memes, especially those involving disinformation and sensitive content, and remains robust under visual perturbations.

Foundation models have established unified representations for natural language processing, yet this paradigm remains largely unexplored for tabular data. Existing methods face fundamental limitations: LLM-based approaches lack retrieval-compatible vector outputs, whereas text embedding models often fail to capture tabular structure and numerical semantics. To bridge this gap, we first introduce the Tabular Embedding Benchmark (TabBench), a comprehensive suite designed to evaluate the tabular understanding capability of embedding models. We then propose TabEmbed, the first generalist embedding model that unifies tabular classification and retrieval within a shared embedding space. By reformulating diverse tabular tasks as semantic matching problems, TabEmbed leverages large-scale contrastive learning with positive-aware hard negative mining to discern fine-grained structural and numerical nuances. Experimental results on TabBench demonstrate that TabEmbed significantly outperforms state-of-the-art text embedding models, establishing a new baseline for universal tabular representation learning.

Large Language Model (LLM)-based Multi-Agent Systems (MAS) are vulnerable to semantic attacks that can trigger cascading failures across the network. Current defenses face a fundamental dilemma: lightweight single-auditor methods are prone to single points of failure, while robust committee-based approaches (e.g., voting) incur prohibitive computational costs in multi-turn interactions. To resolve this tension, we propose MAS-Shield, a secure and efficient defense framework designed with a coarse-to-fine filtering pipeline. Rather than applying uniform scrutiny, MAS-Shield dynamically allocates resources by first identifying high-influence nodes via Critical Agent Auditing Selection. These nodes undergo Light Auditing using lightweight sentry models to rapidly filter benign cases, escalating only high-risk signals to a heavyweight Two-Round Consensus mechanism. This hierarchical design optimizes the security-efficiency trade-off. Experiments demonstrate that MAS-Shield achieves a 92.5% recovery rate against diverse adversarial scenarios and reduces defense latency by over 70% compared to existing methods.

While multi-agent systems based on large language models (LLMs) have shown promise in automating academic research workflows, extending them to economics and business research presents two key challenges, as specialized domain knowledge often spans neighboring disciplines yet remains difficult to access in a structured manner. First, existing methods primarily retrieve information at the paper level, whereas the evidence required for research tasks is often distributed across different sections, creating a granularity mismatch that limits retrieval coverage and precision. Second, these fields demand strict empirical rigor, yet current systems provide limited mechanisms for learning from evaluation feedback. We present BizSage, a multi-agent framework that combines corpus-level fine-grained retrieval with quality-driven self-evolution. We construct a Lateral Knowledge Graph (LKG) by integrating section-level knowledge graphs and apply Personalized PageRank (PPR) to identify sections that are both semantically relevant and structurally important. Seven specialized agents collaborate under a Meta-Review self-evolution mechanism that distills failure modes from evaluation traces into reusable strategies. On a benchmark spanning four domains and three tasks, BizSage ranks first on the majority of metrics, achieves pairwise win rates above 60% against six baselines, and produces zero hallucinated citations. We hope BizSage paves the way for reliable research assistance in economics, business, and the broader social sciences.

Distilling reasoning capabilities from Large Reasoning Models (LRMs) into smaller models is typically constrained by the limitations of rejection sampling. Standard methods treat the teacher as a static filter, discarding complex "corner-case" problems where the teacher fails to explore valid solutions independently, thereby creating an artificial "Teacher Ceiling" for the student. In this work, we propose Hindsight Entropy-Assisted Learning (HEAL), an RL-free framework designed to bridge this reasoning gap. Drawing on the educational theory of the Zone of Proximal Development (ZPD), HEAL synergizes three core modules: (1) Guided Entropy-Assisted Repair (GEAR), an active intervention mechanism that detects critical reasoning breakpoints via entropy dynamics and injects targeted hindsight hints to repair broken trajectories; (2) Perplexity-Uncertainty Ratio Estimator (PURE), a ratio-based filtering heuristic that reduces high-anomaly shortcut-like rationales; and (3) Progressive Answer-guided Curriculum Evolution (PACE), a three-stage distillation strategy that organizes training from foundational alignment to hard-case adaptation. Extensive experiments on multiple benchmarks demonstrate that HEAL significantly outperforms traditional SFT distillation and other baselines.

Large Language Models (LLMs) provide rich semantic representations for recommender systems, but their semantic similarity may not align with collaborative preference patterns captured by collaborative filtering (CF). Directly combining LLM and CF representations can therefore distort the structural relationships encoded in the interaction graph. We propose Structure-Preserving Alignment (SPA), which reformulates CF–LLM integration as language-supervised representation alignment. SPA derives collaborative anchors from frozen CF representations and aligns multiple LLM-generated semantic views toward these anchors through cross-profile view competition. The aligned collaborative representations are then used for structure-preserving two-channel retrieval, which constructs behavior-aware context for downstream LLM-based preference prediction. Experiments on multiple datasets and CF backbones show that SPA consistently improves click-through rate (CTR)-style binary preference prediction.

Scaling laws guide large language model training by relating compute to cross-entropy loss, and recent work further extends them to predict downstream benchmark performance. However, prior approaches face generalization limitations from two aspects: focusing on benchmark-level performance introduces scenario-specific artifacts, while relying on IID validation loss fails to track capability improvements when training distributions vary. In this work, we argue that downstream scaling should be studied at the capability level, which captures shared skill factors across related tasks while abstracting away benchmark-specific noise. We propose SuperValid, a framework that synthesizes OOD (out-of-distribution), capability-aligned validation data by distilling core concepts from benchmarks within a capability domain and expanding them into diverse, knowledge-rich texts. Extensive experiments spanning 16 benchmarks grouped into 6 capability domains show that SuperValid loss exhibits strong and stable correlation with downstream performance across models of different architectures, scales, and training data distributions. As a training-free metric computable during training without benchmark evaluation, SuperValid enables effective model selection, early stopping, and scaling decisions.

Rare-disease diagnosis requires converting heterogeneous patient descriptions into a ranked differential over thousands of disorders while keeping each hypothesis tied to verifiable evidence. Phenotype-based tools provide structured, interpretable comparisons but usually require normalized Human Phenotype Ontology (HPO) input. Large language models (LLMs) can read clinical narratives and propose broader hypotheses, yet their free-form outputs are difficult to ground in a disease catalog or combine with ontology-based evidence. We present ORBIT (Orphanet-grounded Retrieval and Bounded LLM Interaction for Traceable rare-disease diagnosis), a catalog-grounded diagnosis layer for both free-text and structured-HPO cases. ORBIT grounds open disease proposals to Orphanet, fuses them with exact phenotype and case-profile retrieval, scores only the resulting candidate pool with a probabilistic model, and reranks a fixed menu of evidence-bearing disease cards. The output records catalog identifiers, candidate provenance, phenotype support, and conflicts. Across RareArena and RareBench, ORBIT improves every reported metric over single-agent and medical-agent systems using the same backbone. Retrieval analysis, ablations, and expert review further identify where complementary candidate sources, iterative refinement, and constrained reranking affect performance.

Long-form story generation requires models to preserve narrative consistency across extended contexts, yet existing prompting-based methods often accumulate temporal, factual, character, commonsense, and stylistic errors as the story grows. We propose ConWriter, a training-free framework for consistency-aware long-form story generation. ConWriter writes stories incrementally at the scene level, guided by static story requirements, dynamic narrative memory, symbolic state reasoning, and uncertainty-aware risk signals. Rather than treating long-story generation as a single free-form decoding process, ConWriter maintains evolving story states, checks whether new scenes satisfy required narrative transitions, and uses uncertainty-aware risk signals to prioritize validation and localized repair. This enables consistency control during generation, before local errors propagate into later scenes. We evaluate ConWriter on ConStory-Bench across four long-story tasks, three target lengths, and multiple base LLMs following the official evaluation protocol. Across models and story lengths, ConWriter consistently matches or improves upon direct generation and outperforms the recent training-free baseline DOME in narrative consistency. These results demonstrate the effectiveness of lightweight neuro-symbolic consistency control for training-free long-form story generation. Code is available on GitHub.

While human evaluation is the gold standard in many NLP tasks, it suffers from prohibitive costs and poor scalability. When identifying top-performing models, typical evaluation protocols waste effort by exhaustively evaluating all models on the entire benchmark, a safe but inefficient approach. In this work, we formalize multi-model human evaluation as a best-arm identification problem in a multi-armed bandit setup with correlated arms, where pulling an arm corresponds to human-evaluating a model. By sampling adaptively based on the intermediate model rankings obtained on the samples so far, we can focus the annotation budget on the most competitive models. We prove the optimality of the proposed algorithms and show that it improves discrimination between top-performing models. This makes evaluations faster, cheaper and more aligned with large-scale competition evaluation goals.

Temporal Knowledge Graph (TKG) extrapolation aims to predict future facts from observed historical events. Existing methods usually achieve this goal by jointly modeling local historical evolution and long-range historical contexts. First, local evolution modeling often summarizes local history into an overall evolving representation, where query-relevant local deviations may not be sufficiently exploited as useful predictive features. Second, existing long-range history modeling often treats historical recurrence as a stable cue, while overlooking that its reliability may vary across relations and recent contexts. To address these issues, we propose DyHist, a query-aware historical dynamics modeling framework for TKG extrapolation. DyHist captures local dynamics by separating recent evidence from medium-range context and explicitly modeling their query-relevant deviations as predictive features. It further uses a relation-aware historical prior to calibrate candidate scores with time-decayed support, recent trends, and relation-specific strength, enabling recurrence signals to contribute differently across relations and recent contexts. Experiments on four benchmarks show that DyHist achieves the best MRR against recent baselines. Code is available at https://github.com/PANPANKK/DyHist.

Multilingual safety alignment typically assumes standard, canonical scripts, overlooking orthographic variation. We introduce PunjabiSafetyBench, a parallel safety benchmark evaluating 300 harmful intents and 500 benign controls across six sociolinguistic script variants (Gurmukhi, Shahmukhi, Romanized, code-mixed, Urdu, English). Evaluating six frontier LLMs, we reveal severe safety degradation on non-canonical scripts: transitioning queries to Romanized Punjabi or Shahmukhi drives significant guardrail bypasses in open-weight models, while proprietary models remain robust. Furthermore, we expose a localization gap where models overlook regional harmful slang despite successfully refusing standard clinical equivalents. Our findings demonstrate that modern guardrails are fragile to surface orthography, necessitating safety alignment on informal, non-standard scripts.

As post-training techniques evolve, large language models (LLMs) are increasingly augmented with structured multi-step reasoning abilities, often optimized through reinforcement learning. These reasoning-enhanced models outperform standard LLMs on complex tasks and now underpin many commercial LLM APIs. However, to protect proprietary behavior and reduce verbosity, providers typically conceal the reasoning traces while returning only the final answer. This opacity introduces a critical transparency gap: users are billed for invisible reasoning tokens, which often account for the majority of the cost, yet have no means to verify their authenticity. This opens the door to token count manipulation, where providers may overreport token usage or inject synthetic, low-effort tokens to inflate charges, a threat that can be carried out at near-zero cost. To address this issue, we propose CoIn, a verification framework that audits both the quantity and semantic validity of hidden tokens. CoIn constructs a verifiable hash tree from token embedding fingerprints to check token counts, and uses embedding-based relevance matching to detect fabricated reasoning content. Experiments demonstrate that CoIn, when deployed as a trusted third-party auditor, can effectively detect token count manipulation with a success rate reaching up to 94.7%, demonstrating a strong ability to restore billing transparency in opaque LLM services. The dataset and code are available at https://github.com/CASE-Lab-UMD/LLM-Auditing-CoIn .

Text watermarking enables provenance tracing of LLM-generated text by embedding detectable signals, while multi-bit watermarking further encodes payloads such as user identities or model versions. Scaling multi-bit watermarking to large payloads is challenging as exhaustive decoding grows exponentially with payload length. Existing methods commonly address this by segmenting the payload and decoding each segment independently, but the impact of this design on decoding accuracy remains insufficiently understood. In this paper, we revisit segmentation in multi-bit watermarking and reveal a drawback of the commonly adopted paradigm: segmentation splits the available token evidence among segments, leaving fewer observations for each local decoding decision and thereby reducing decoding reliability. We theoretically characterize this accuracy–efficiency trade-off and empirically validate the claim across diverse settings. Additionally, we introduce a distribution-balanced detector to mitigate message-dependent bias. Experiments across multiple datasets and models show that our method achieves improved accuracy and robustness under editing attacks compared to existing methods. Our code is available at https://github.com/YouduLiu/SDBMW

Recent advances in Video Large Language Models (Video-LLMs) have enabled performance on long-video understanding tasks. However, existing methods still face two key limitations: evidence acquisition often relies on a single search intent, and answer generation lacks an effective visual feedback mechanism. To address these limitations, we propose CoVER, a Comprehensive Visual Evidence and Reflection framework for long-video understanding. CoVER enables Video-LLMs to See More by dynamically gathering query-expanded visual evidence, and Think Deeper by verifying draft answers with effective answer-specific visual feedback. Together, these mechanisms shift long-video understanding from answer-centric generation to evidence-centric and visually verifiable reasoning. Experimental results show that CoVER-7B substantially outperforms models with the same parameter scale and even surpasses state-of-the-art closed-source models on certain metrics.

Understanding the economic intent of Ethereum transactions is critical for user safety, yet current tools expose only raw on-chain data or surface-level intent, leading to widespread “blind signing” (approving transactions without understanding them). Through interviews with 16 Web3 users, we find that effective explanations should be structured, risk-aware, and grounded at the token-flow level. Motivated by these findings, we formulate TxSum, a new domain-grounded NLP task for DeFi transaction explanation, and construct a dataset of 187 complex Ethereum transactions with 2,375 token-flow annotations and transaction-level summaries.

We further introduce MATEX, a grounded multi-agent framework for high-stakes transaction explanation. It selectively retrieves external knowledge under uncertainty and audits explanations against raw traces to improve token-flow-level factual consistency. MATEX achieves the strongest overall explanation quality, especially on micro-level factuality and intent quality. It improves user comprehension on complex transactions from 52.9% to 76.5% over the strongest baseline and raises malicious-transaction rejection from 36.0% to 88.0%, while maintaining a low false-rejection rate on benign transactions.

LLM-based agents are increasingly being developed for time-series problems, but their design choices vary substantially across task settings. This survey adopts a problem-driven taxonomy that organizes these systems by the time-series problems they address rather than by isolated technical components. We group existing systems into four categories: forecasting and reasoning, augmentation and synthesis, anomaly detection and diagnosis, and decision support. Within each category, we examine how task requirements shape agent architecture, tool use, and memory design. We further summarize representative datasets and environments, and compare reported model performance under shared or closely related settings. Overall, this survey offers a task-oriented guide to designing LLM-based agents for time-series problems and identifies open gaps for future work.

Diffusion large language models (D-LLMs) have emerged as a promising alternative to auto-regressive models due to their iterative refinement capabilities. However, hallucinations remain a critical issue that hinders their reliability. To detect hallucination responses from model outputs, token-level uncertainty (e.g., entropy) has been widely used as an effective signal to indicate potential factual errors. Nevertheless, the fixed-length generation paradigm of D-LLMs implies that tokens contribute unevenly to hallucination detection, with only a small subset providing meaningful signals. Moreover, the evolution trend of uncertainty throughout the diffusion process can also provide important signals, highlighting the necessity of modeling its denoising dynamics for hallucination detection. In this paper, we propose DynHD that bridge these gaps from both spatial (token sequence) and temporal (denoising dynamics) perspectives. To address the information density imbalance across tokens, we propose a semantic-aware evidence construction module that extracts hallucination-indicative signals by filtering out non-informative tokens and emphasizing semantically meaningful ones. To model denoising dynamics for hallucination detection, we introduce a reference evidence generator that learns the expected evolution trajectory of uncertainty evidence, along with a deviation-based hallucination detector that makes predictions by measuring the discrepancy between the observed and reference trajectories. Extensive experiments demonstrate that DynHD consistently outperforms state-of-the-art baselines while achieving higher efficiency across multiple benchmarks and backbone models.

Large language models (LLMs) can propose vaccine adjuvant designs with plausible mechanistic rationales, but such outputs often lack grounding in established immunological knowledge. We present SAVANT, a neuro-symbolic framework for LLM-based adjuvant design that combines two-stage constrained generation with literature-grounded verification. SAVANT first generates a natural-language design rationale, then renders it into an ontology-constrained domain-specific language (DSL) using logic masks and finite-state control. To reason under fragmented evidence and incomplete first-principles rules, SAVANT verifies generated designs against literature-derived symbolic evidence through minimal sub-theorem proof. In expert-aligned evaluation, SAVANT shows greater stability in stage-level rank alignment than the strongest LLM-judge baselines, while achieving the strongest alignment for immune-effect support and mechanistic reasoning. SAVANT provides graded, interpretable assessments that distinguish supported mechanisms, under-supported hypotheses, and knowledge gaps for hypothesis-driven experimentation. Data and code will be released at https://github.com/banjiuyufen/SAVANT.

Large language models deployed as agents over large tool catalogs face a critical tool-retrieval bottleneck. As embedding-based retrieval approaches rely on compact encoders that may under-capture specialized tool semantics, parametric tool retrieval addresses this by encoding each tool as a virtual token appended to the LLM vocabulary, fine-tuned in two stages (memorization then retrieval SFT) to use the LLM as a retriever, achieving strong performance on standard ToolBench retrieval benchmarks. Yet these benchmarks use verbose, fully-specified queries, and their evaluation applies constrained decoding that restricts outputs to valid token paths - neither reveals whether the model actually understands its tools. We introduce **ToolSense**, an open-source LLM-powered diagnostic framework that takes any tool catalog as input and automatically generates three benchmarks: a Realistic Retrieval Benchmark (RRB) with queries at three ambiguity tiers, an MCQ probing benchmark, and a QA probing benchmark. Applying ToolSense to ToolBench (∼47k tools) and evaluating five parametric model training configurations reveals a knowledge-retrieval dissociation: on RRB queries, several configurations collapse by ∼**50-64 percentage points** compared to fully-specified ToolBench benchmarks, falling below embedding-model baseline. Additionally, despite strong retrieval performance, some models score near-random on factual probes, suggesting a knowledge-retrieval dissociation. We open-source the ToolSense framework and the ToolBench diagnostic benchmarks to encourage further research. (see supplementary material)

GUI world models are increasingly evaluated as one-step next-screen predictors, yet their intended use is often as multi-step environments for GUI agents. This mismatch leaves a key requirement under-tested: generated states must remain contextually consistent when they are repeatedly reused for future interaction. We introduce GUI-CC, a benchmark that evaluates contextual consistency of GUI world models as agent environments rather than isolated next-screen predictors. GUI-CC contains two complementary tracks: an offline reference-action track that rolls models along real mobile GUI trajectories, and an online agent-loop track that lets fixed probing agents interact with model-generated UIs. We construct 500 offline trajectory tasks from GUIOdyssey and 200 emulator-verified online tasks across 30 mobile apps. GUI-CC evaluates transition fidelity, transition plausibility, contextual consistency, and task progress. Experiments show that plausible single-step generation does not guarantee reliable environment simulation: current models often produce usable-looking screens while failing to preserve task-relevant context or support executable multi-step rollouts.

People increasingly turn to large language models (LLMs) for everyday advice, making ethically charged interpersonal problems a practical moral-advisory context. Most prior work has studied this context through single-turn judgments or pressure-laden rebuttals, assumptions that poorly match how guidance is sought in real-world contexts. These assumptions leave unclear whether narration alone, without an explicit opposing position, can shift model judgments during multi-turn moral consultation. Yet real-world moral-conflict conversation often elicits one party’s self-justifying account, which can unfold over multiple turns and create information asymmetry. We introduce narrative captivity, a failure mode in which a model treats an unopposed one-sided account as complete and aligns with the narrator’s interpretation without seeking missing perspectives. To measure this phenomenon, we build a benchmark of 5,078 interpersonal-conflict scenarios spanning six moral dimensions. Across 17 LLMs, narrative captivity is widespread: end-state judgments under multi-turn narration shift by 25 percentage points on average beyond the matched single-turn baseline. Stage-level analysis identifies preference optimization as a major contributor, while four inference-time strategies provide only partial mitigation. We hope our project fosters LLM advisors that preserve independent judgment in real-world consultation.

We study few-shot online clustering (FSOC), where an unlabelled stream must be assigned to a dynamically growing cluster inventory using only a few labelled support examples for currently observed classes. We propose a two-component test-time gradient-free framework. First, we adapt Wasserstein Generative Flow Networks (wGFN) as an amortised LLM-based policy for in-distribution assignment and out-of-distribution (OOD) rejection. Unlike prior work that trains separate policies for each evaluation class set, we show that a single policy trained on a few classes generalises to novel support classes at test time without parameter updates and further replace fixed rejection thresholds with a Dirichlet evidence model over candidate class assignments. Second, rejected OOD examples enter an oracle-mediated cluster-formation stage. Under shuffled stream orderings, this stage can create noisy provisional support sets that corrupt later assignments. To address this, we introduce a von Mises–Fisher Dirichlet Process Mixture Model warm start to provide oracle-independent cluster structure over unit-normalised text embeddings before oracle-mediated formation. This improves clustering quality, in some settings by ∼ 2×, even with small LLMs (<1B parameters). To our knowledge, this is the first FSOC framework to couple amortised LLM-based assignment with Bayesian non-parametric warm-starting, enabling growing cluster inventories without data augmentation, novel-class parameter updates or predefined cluster counts.

Modern reasoning systems increasingly rely on verifier-guided selection: sample several solution traces, score them with a reward model, and return the top-ranked answer. In multi-generator deployments, this verifier is often reused across generators. We ask whether this reuse assumption is reliable. On a 34-generator math panel with K=64 candidates per problem and 7 verifiers, the 7B process reward model (PRM-7B) is the strongest fixed verifier, but a per-target single-verifier oracle gains +1.6 percentage points (pp). We then ask whether this target-local mismatch is actionable. Four label-free candidate-text statistics predict the coarse PRM outcome-verifier regime, and three budgeted responses improve over fixed PRM on the 22-target first-pass panel: a no-label gated router (+1.41 pp), PRM+V4 fusion (+2.35 pp), and a labelled PRM-only verifier-allocation arbiter (PVA-Adapt; +2.26 pp). A same-panel American Invitational Mathematics Examination (AIME) 2026 held-out slice gives +3.88 pp, and a second candidate-bank check keeps the scored and labelled gains positive while weakening the no-label router. We frame the result as a seed- and verifier-family-conditioned deployment diagnostic rather than a universal routing rule; a supervised fine-tuning (SFT) replication also finds no positive transfer from better selected traces.

Aspect Sentiment Triplet Extraction (ASTE) is a fine-grained task in Aspect-Based Sentiment Analysis (ABSA), aiming to extract triplets of aspect terms, their corresponding opinion terms, and sentiment orientation. Recent studies have employed semantic and syntactic features to enhance extraction performance. However, as the extraction of both syntactic and semantic information is not task-specific, such information does not target aspect-opinion pairs within sentences, thereby potentially introducing noise unrelated to the task. To address this issue, we propose a Syntactic-Guided Semantic Attention Matrix Enhancement (SG-SAME) framework. First, we use a multi-view masking mechanism and a biaffine-based gating module to derive a task-aligned shared matrix from the existing semantic and syntactic matrices. Second, the shared matrix is injected into the attention score matrix of LLM layers to enhance the model’s representation learning. Finally, we propose a chain-of-thought (CoT) framework with negative candidate injection to steer the model toward autonomous reasoning and error self-rectification. Experimental results on multiple benchmark datasets demonstrate that our method outperforms existing state-of-the-art approaches in ASTE tasks. Code will be available.

Vision-language models can be used in applications that analyze people in professional and social contexts. This raises concerns about whether these models reproduce occupational gender stereotypes. However, it is challenging to evaluate these associations because naturally collected images often combine demographic attributes with clothing, pose, background, and other status-related cues. We introduce a paired-image benchmark to study occupational status judgments in VLMs, containing 5,487 human-validated synthetic images spanning 41 ISCO-08 sub-major occupational groups aggregated into 10 major groups, 7 ethnicity categories, and 5 age groups. Men and women are matched based on occupation, age, ethnicity, and workplace context. We evaluate 5 VLMs using spatial comparisons that avoid direct gender references, explicit gender comparisons, and structured reasoning prompts. We observe persistent male-oriented associations in earnings judgments, even when the prompts avoid direct references to gender. Furthermore, we find that men are more frequently associated with authority, seniority, and earning potential, whereas women are more often associated with warmth. Promotion likelihood exhibits a model-family-dependent pattern: male-oriented associations emerge across Qwen-3-VL models, whereas both Gemma-3 models show the opposite, female-oriented pattern. These findings show that occupational gender stereotypes persist in the evaluated VLMs even under tightly controlled visual conditions. Our code and data are publicly available at https://github.com/NLP-UMUTeam/OccuBiasBench

Retrieval-Augmented Generation (RAG) systems are vulnerable to knowledge corruption attacks, where adversaries inject malicious texts into external corpora to manipulate generated responses. Existing one-to-many attacks can compromise multiple queries with a small number of injected texts, but they typically require white-box access to the target retriever for gradient-based optimization. In contrast, existing black-box attacks are more practical but usually allow each malicious text to target only one query. In this paper, we study one-to-many knowledge corruption attacks against strictly black-box RAG systems, where the attacker has no access to the target retriever, generator, or corpus contents. We propose CoverPoison, a black-box method that combines competitive similarity clustering, coverage-oriented optimization, and T5-based text refinement. Competitive similarity clustering improves query grouping quality, while the coverage-oriented objective encourages each candidate to cover more queries through query-specific retrieval thresholds. Experiments on three datasets show that CoverPoison substantially improves black-box attack effectiveness. Further analyses across RAG configurations, attack objectives, and defenses validate the effectiveness of the proposed method.

While Large Language Models (LLMs) are aligned to mitigate risks, their safety guardrails remain fragile against jailbreak attacks. This reveals limited understanding of components governing safety. Existing methods rely on local, greedy attribution that assumes independent component contributions. However, they overlook the cooperative interactions between different components in LLMs, such as attention heads, which jointly contribute to safety mechanisms. We propose Global Optimization for Safety Vector Extraction (GOSV), a framework that identifies safety-critical attention heads through global optimization over all heads simultaneously. We employ two complementary activation repatching strategies: Harmful Patching and Zero Ablation. These strategies identify two spatially distinct sets of safety vectors with consistently low overlap, termed Malicious Injection Vectors and Safety Suppression Vectors, demonstrating that aligned LLMs maintain separate functional pathways for safety purposes. Through systematic analyses, we find that complete safety breakdown occurs when approximately 30% of total heads are repatched across all models. Building on these insights, we develop a novel inference-time white-box jailbreak attack and a complementary detection-based defense, both grounded in the identified safety vectors. The attack achieves superior performance, while the defense yields strong detection across diverse attack types, providing robust evidence for the effectiveness of the proposed GOSV framework in LLM safety interpretability.

Visual Quality Assessment (QA) in multimodal large language models (MLLMs) jointly predicts absolute quality scores and relative perceptual rankings. Existing methods optimize score regression or ranking within a unified formulation, leading to unstable predictions and weak generalization under stochastic generation. We find this instability is amplified by response variability in MLLM reasoning, where stochastic decoding produces inconsistent score distributions and degrades cross-dataset robustness. This indicates QA depends not only on preference alignment, but also on response-level stability. We propose PreResQ-R1, a Response-Preference Disentangled Ranking-and-Scoring Reinforcement Optimization framework. It decouples response stabilization and preference learning, with a dual-branch reward modeling intra-sample consistency and inter-sample score-ranking alignment. For video QA, we introduce temporal-spatial evidence aggregation without dense motion modeling. PreResQ-R1 achieves state-of-the-art performance, outperforming comparable methods by 5.60% on IQA benchmarks and 2.53% on VQA benchmarks.

Tree-based speculative decoding accelerates autoregressive generation by verifying multiple draft candidates in parallel, but this advantage weakens for sparse Mixture-of-Experts (MoE) models. As the draft tree grows, different branches activate different experts, expanding the union of activated experts and substantially increasing target-side verification cost. We propose EVICT, a training-free, hyperparameter-free, and lossless adaptive verification method for MoE speculative decoding. EVICT makes every verified token count by truncating the draft tree before target verification and retaining only the cost-effective prefix. It leverages fine-grained drafter signals to estimate candidate benefit, combines them with offline-profiled verification cost, and remains highly compatible with the high-performance graph-based serving framework SGLang. Extensive experiments on diverse MoE backbones and benchmarks show that EVICT achieves up to 2.35× speedup over autoregressive decoding and an average 1.21× speedup over the state-of-the-art baseline EAGLE-3, while significantly reducing unnecessary expert activations during verification.

Medical AI benchmarks routinely pair a diagnostic image with the same-visit clinical record, delete the recorded diagnosis, and treat what remains as auxiliary context. The recipe is reasonable. It is not. On 5904 ophthalmic visits across eight cohorts, a frozen text probe still recovers the disease label from the diagnosis-redacted record at 0.944 macro-AUROC, against 0.998 on the unredacted record; the most aggressive dictionary scrub closes less than one further percentage point. The diagnosis is already in the notes.

It is not in one place. A clinical record is written during the diagnostic process, and its fingerprints persist across channels: narrative text alone reaches 0.978, and a view that erases every value and exposes only which fields were populated still reaches 0.854. We introduce CRAFT (Channel-wise Redaction And Floor Test), a portable audit protocol that renders each record into redacted and channel-isolated views and reports each channel’s label recoverability; rankings reproduce under Qwen3.5-9B. Before clinical-text or image-text fusion gains are read as cross-modal reasoning, the text branch’s channel-level shortcut floors must be reported alongside them. https://github.com/yhc98002-bit/The-Diagnosis-Is-Already-in-the-Notes

Humans naturally exhibit multiple forms of abstraction in reasoning and interaction, including temporal abstraction across decision timescales and strategic abstraction over communicative intents. Inspired by these complementary abstractions, we propose a two-level hierarchical reinforcement learning (HRL) framework for conversational agents that bridges the gap between existing token-level and utterance-level RL methods. Built upon a two-level Markov decision process (MDP), our framework conditions token-level response generation on utterance-level actions represented by explicit textual strategies. Based on theoretical analysis and efficiency considerations, we employ DQN to optimize the high-level Q-network and PPO to train the low-level actor-critic. To further alleviate reward sparsity and facilitate convergence, we introduce a dual-granularity reward mechanism that combines the utterance-level satisfaction score with token-level intrinsic self-consistency and a KL-divergence penalty. Experiments on both daily-life and emotional support conversations demonstrate that our method consistently outperforms a wide range of baselines in both strategy determination and response quality. Our implementation is available at https://github.com/AaronJi/ToSCA.