Large language models solve complex tasks by generating long reasoning chains, achieving higher accuracy at the cost of increased computational cost and reduced ability to isolate functionally relevant reasoning. Prior work on compact reasoning shortens such chains through probabilistic sampling, heuristics, or supervision from frontier models, but offers limited insight into whether models internally encode token-level functional importance for answer generation. We address this gap diagnostically and propose greedy pruning, a likelihood-preserving deletion procedure that iteratively removes reasoning tokens whose removal minimally degrades model likelihood under a specified objective, yielding length-controlled reasoning chains. We evaluate pruned reasoning in a distillation framework and show that students trained on pruned chains outperform a frontier-model–supervised compression baseline at matched reasoning lengths. Finally, our analysis reveals systematic pruning patterns and shows that attention scores can predict greedy pruning ranks, further suggesting that models encode a nontrivial functional importance structure over reasoning tokens.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
HEALing Entropy Collapse: Enhancing Exploration in Few-Shot RLVR via Hybrid-Domain Entropy Dynamics Alignment
PDF ↗Reinforcement Learning with Verifiable Reward (RLVR) has proven effective for training reasoning-oriented large language models, but existing methods largely assume high-resource settings with abundant training data. In low-resource scenarios, RLVR is prone to more severe entropy collapse, which substantially limits exploration and degrades reasoning performance. To address this issue, we propose **H**ybrid-domain **E**ntropy dynamics **AL**ignment (HEAL), a framework tailored for few-shot RLVR. HEAL first selectively incorporates high-value general-domain data to promote more diverse exploration. Then, we introduce Entropy Dynamics Alignment (EDA), a reward mechanism that aligns trajectory-level entropy dynamics between the target and general domains, capturing both entropy magnitude and fine-grained variation. Through this alignment, EDA not only further mitigates entropy collapse but also encourages the policy to acquire more diverse exploration behaviors from the general domain. Experiments across multiple domains show that HEAL consistently improves few-shot RLVR performance. Notably, using only 32 target-domain samples, HEAL matches or even surpasses full-shot RLVR trained with 1K target-domain samples.
The widespread integration of Large Language Models (LLMs) necessitates rigorous and systematic safety evaluation. Existing paradigms either rely on constructed benchmarks to assess safety from predefined perspectives, or employ dynamic red-teaming to probe potential vulnerabilities. While effective, these approaches face challenges, as they depend heavily on expert domain knowledge, offer limited systematic guarantees, and are vulnerable to rapid obsolescence. To address these limitations, we introduce a novel framework POLARIS that brings the rigor of specification-based software testing to AI safety. POLARIS first compiles unstructured natural-language policies into First-Order Logic (FOL) representations, establishing a traceable link between high-level rules and concrete test cases. This formalization enables the construction of a Semantic Policy Graph, where complex policy violation scenarios are encoded as traversable paths. By systematically exploring this graph, POLARIS uncovers compositional violation patterns, which are then instantiated into executable natural-language test queries, enabling coverage-driven and reproducible safety testing. Experiments demonstrate that POLARIS achieves higher policy coverage and attack success counts compared to established baselines. Crucially, by bridging formal methods and AI safety, POLARIS provides a principled, automated approach to ensuring LLMs adhere to safety-critical policies with verifiable traceability.
While large language models (LLMs) have shown to perform well on monolingual mathematical and commonsense reasoning, they remain unreliable for multilingual medical reasoning applications, hindering their deployment in multilingual healthcare settings. We address this by first introducing CURE-Med-Bench, a high-quality multilingual medical reasoning dataset with open-ended reasoning queries with a single verifiable answer, spanning thirteen languages, including underrepresented languages such as Amharic, Yoruba, and Swahili. Building on this dataset, we propose CURE-Med, a curriculum-informed reinforcement learning framework that integrates code-switching-aware supervised fine-tuning and Group Relative Policy Optimization to jointly improve logical correctness and language stability. Across thirteen languages, our approach consistently outperforms strong baselines and scales effectively, achieving 85.21% language consistency and 54.35% logical correctness at 7B parameters and 94.96% language consistency and 70.04% logical correctness at 32B parameters. These results support reliable and equitable multilingual medical reasoning in LLMs. The code and dataset will be made publicly available upon acceptance.
Prior work has shown that large language models (LLMs) often converge to accurate input embedding for numbers, based on sinusoidal representations.In this work, we demonstrate that these representations are in fact strikingly systematic, to the point of being almost perfectly universal: different LLM families develop equivalent sinusoidal structures, and number representations are broadly interchangeable in a large swathe of experimental setups.We show that properly factoring in this characteristic is crucial when it comes to assessing how accurately LLMs encode numeric and other ordinal information, and that mechanistically enhancing this sinusoidality can also lead to reductions of LLMs’ arithmetic errors.
ArgGenBench: Benchmarking the Complex Controlled Argument Generation Capability of Large Language Models
PDF ↗Argument generation is a fundamental NLP task that aims to automatically produce persuasive arguments.Effective human argumentation is inherently complex and multifaceted, integrating argumentative strategies, appropriate styles, and adaptation to target audiences, etc.However, existing studies focus on limited control signals such as topic, stance, or key aspects, failing to capture this complexity.As LLMs advance, the lack of benchmarks evaluating multifaceted argumentative control becomes a critical bottleneck.To address this, we introduce ArgGenBench, a novel benchmark containing complex instructions that integrate multi-dimensional control, including topic, stance, length, style, strategy, audience, and key points.Extensive evaluation across 15 LLMs reveals significant limitations: even the best-performing model achieves only 42.7% win rate against human-verified references.These results highlight the challenge of controlled argument generation and establish ArgGenBench as a rigorous testbed for developing more capable systems.
Previous work examining the Uniform Information Density (UID) hypothesis has shown that while information as measured by surprisal metrics is distributed more or less evenly across documents overall, local discrepancies can arise due to functional pressures corresponding to syntactic and discourse structural constraints. However, work thus far has largely disregarded the relative salience of discourse participants. We fill this gap by studying how overall salience of entities in discourse relates to surprisal using 70K manually annotated mentions across 16 genres of English and a novel minimal-pair prompting method. Our results show that globally salient entities exhibit significantly higher surprisal than non-salient ones, even controlling for position, length, and nesting confounds. Moreover, salient entities systematically reduce surprisal for surrounding content when used as prompts, enhancing document-level predictability. This effect varies by genre, appearing strongest in topic-coherent texts and weakest in conversational contexts. Our findings refine the UID competing pressures framework by identifying global entity salience as a mechanism shaping information distribution in discourse.
Large Language Models (LLMs) are increasingly used in education, yet their default helpfulness often conflicts with pedagogical principles. Prior work evaluates pedagogical quality via answer leakage–the disclosure of complete solutions instead of scaffolding–but typically assumes well-intentioned learners, leaving tutor robustness under student misuse largely unexplored. In this paper, we study scenarios where students behave adversarially and aim to obtain the correct answer from the tutor. We evaluate a broad set of LLM-based tutor models, including different model families, pedagogically aligned models, and a multi-agent design, under a range of adversarial student attacks. We adapt six groups of adversarial and persuasive techniques to the educational setting and use them to probe how likely a tutor is to reveal the final answer. We evaluate answer leakage robustness using different types of in-context adversarial student agents, finding that they often fail to carry out effective attacks. We therefore introduce an adversarial student agent that we fine-tune to jailbreak LLM-based tutors, which we propose as the core of a standardized benchmark for evaluating tutor robustness. Finally, we present simple but effective defense strategies that reduce answer leakage and strengthen the robustness of LLM-based tutors in adversarial scenarios.
Generating radiology reports from 3D volumetric data remains challenging due to the difficulty of grounding fine-grained pathologies within high-dimensional scans. While retrieval-augmented generation (RAG) offers a potential solution, standard approaches struggle with visual-semantic ambiguity and often introduce irrelevant "normal" context that dilutes pathological signals. To address this limitation, we introduce CPR-RAG, a model-agnostic RAG framework that enhances organ-level grounding by integrating clinical priors into the retrieval process. Specifically, we propose a clinical prior-regularized re-ranking module that leverages corpus-derived co-occurrence statistics to align retrieved candidates with latent disease distributions, ensuring clinical consistency beyond mere visual similarity. Furthermore, we employ clinical relevance context refinement to selectively filter out boilerplate normal descriptions, thereby maximizing the information density of the evidence provided to the generator. Extensive experiments on the RadGenome-ChestCT benchmark demonstrate that CPR-RAG significantly improves clinical efficacy across state-of-the-art radiology report generation models. Human evaluation further confirms that our approach achieves superior factual correctness, completeness, and utility compared to the existing models.
To Judge or Not to Judge: Can Large Language Models Leverage the Dispute Focus in Legal Judgment?
PDF ↗Civil judicial cases are highly complicated, posing significant challenges for Large Language Models (LLMs) for Legal Judgment Prediction (LJP). While judges manage this complexity through the dispute focus—a mechanism distilling cases into core issues—existing research largely overlooks this tool in favor of generic reasoning frameworks that lack authentic judicial logic. To bridge this gap, we first introduce \textbf{FocalLaw}, the first dataset aligning full-process Chinese civil judicial data through the dispute focus, comprising 1,000 high-quality cases across six causes of action. Building on this dataset, we examine LLMs’ capability to utilize the dispute focus and uncover a counter-intuitive phenomenon: LLMs fail to leverage the dispute focus even with CoT and SFT, which we identify as the "Clerk Trap".To solve the problem, we propose \textbf{FocalJudge}, a novel framework that leverages the dispute focus to guide LLMs through a structured, judge-like cognitive workflow. Experimental results demonstrate the effectiveness of FocalJudge and offer valuable insights into the interpretability and reliability of LLMs in the legal domain.
Beyond the Panorama: Training-Free Hierarchical Perception-Reasoning for Fine-Grained Vision in MLLMs
PDF ↗Multimodal large language models (MLLMs) enable cross-modal semantic understanding and generation by learning semantic alignment and fusion across modalities. However, existing MLLMs still face challenges in fine-grained visual tasks. Their uniform encoding for global understanding tends to blur or lose local details, while the lack of explicit modeling of intermediate visual evidence leads them to rely on semantic priors or the statistical patterns of language models rather than grounded visual information, resulting in potential hallucinations. To address these issues, we propose HiPerson, a training-free hierarchical perception-reasoning framework that enhances fine-grained visual understanding by simulating human perception mechanisms. Specifically, HiPerson fuses internal relative attention and gradient activation signals to generate a task-aware semantic heatmap, providing explicit perceptual anchors for precise localization. Then, it employs a dual-scale adaptive cropping strategy to extract visual cues for interactive reasoning, simulating the process of human visual focus shifting and detail attention. Finally, by combining local-global dual-image cooperative input with a multi-step reasoning prompting mechanism, HiPerson guides the model to complete a full perception loop from detail observation to contextual verification. Experiments show that HiPerson achieves competitive results on multiple datasets, demonstrating its generalizability and scalability.
Understanding and Mitigating Bias Inheritance in LLM-based Data Augmentation on Downstream Tasks
PDF ↗Generating synthetic datasets via large language models (LLMs) has emerged as a promising approach to improve LLM performance.However, LLMs inherently reflect biases in their training data, leading to a critical challenge: when models are trained on synthetic data, they may propagate and amplify the inherent biases that can significantly impact fairness and robustness on downstream tasks—a phenomenon we term bias inheritance. This work presents the first systematic investigation in understanding, analyzing, and mitigating bias inheritance. We fine-tune LLMs with a combined dataset of real and LLM-augmented data with varied bias ratio as the proportion of augmented data. Through systematic experiments across 10 classification and generation tasks, we analyze how 6 different types of biases manifest. Our results indicate that bias inheritance harms downstream task performance in bias directly-related classification and generation tasks. Then, our analysis identifies three key misalignment factors: misalignment of values, group data, and data distributions. Based on these insights, we propose three mitigation strategies: token-based, mask-based, and loss-based approaches, which can work differently on various tasks and bias, indicating the substantial challenges to mitigate bias inheritance. We hope this work can provide insights to the research of LLM data augmentation.
Explainable and Fine-Grained Safeguarding of LLM Multi-Agent Systems via Bi-Level Graph Anomaly Detection
PDF ↗Large language model (LLM)-based multi-agent systems (MAS) have shown strong capabilities in solving complex tasks. As MAS become increasingly autonomous in various safety-critical tasks, detecting malicious agents has become a critical security concern. Although existing graph anomaly detection (GAD)-based defenses can identify anomalous agents, they mainly rely on coarse sentence-level information and overlook fine-grained lexical cues, leading to suboptimal performance. Moreover, the lack of interpretability in these methods limits their reliability and real-world applicability. To address these limitations, we propose , an explainable and fine-grained safeguarding framework for detecting malicious agents in MAS. To incorporate both coarse and fine-grained textual information for anomalous agent identification, we utilize a bi-level agent encoder to jointly model the sentence- and token-level representations of each agent. A theme-based anomaly detector further captures the evolving discussion focus in MAS dialogues, while a bi-level score fusion mechanism quantifies token-level contributions for explanation. Extensive experiments across diverse MAS topologies and attack scenarios demonstrate robust detection performance and strong interpretability of XG-Guard.
Rerankers play a pivotal role in refining retrieval results for Retrieval-Augmented Generation. However, current reranking models are typically optimized on static human annotated relevance labels in isolation, decoupled from the downstream generation process. This isolation leads to a fundamental misalignment: documents identified as topically relevant by information retrieval metrics often fail to provide the actual utility required by the LLM for precise answer generation. To bridge this gap, we introduce ReRanking Preference Optimization (RRPO), a reinforcement learning framework that directly aligns reranking with the LLM’s generation quality. By formulating reranking as a sequential decision-making process, RRPO optimizes for context utility using LLM feedback, thereby eliminating the need for expensive human annotations. To ensure training stability, we further introduce a reference-anchored deterministic baseline. Extensive experiments on knowledge-intensive benchmarks demonstrate that RRPO significantly outperforms strong baselines, including the powerful list-wise reranker RankZephyr. Further analysis highlights the versatility of our framework: it generalizes seamlessly to diverse readers (e.g., GPT-4o), integrates orthogonally with query expansion modules like Query2Doc, and remains robust even when trained with noisy supervisors.
Existing facial forgery detection methods typically focus on binary classification or pixel-level localization, providing little semantic insight into the nature of the manipulation. To address this, we introduce Forgery Attribution Report Generation, a new multimodal task designed to provide post-hoc forensic evidence for manipulated images. This task jointly localizes forged regions (“Where“) and generates natural language explanations grounded in the editing process (“Why“). This dual-focus approach goes beyond traditional binary forensics, providing a comprehensive, interpretable understanding of the manipulation. To enable research in this domain, we present Multi-Modal Tamper Tracing (MMTT), a large-scale dataset of 152,217 samples. Each sample features a process-derived ground-truth mask and a human-authored textual description, ensuring high annotation precision and linguistic richness. We further propose ForgeryTalker, a unified end-to-end baseline that integrates vision and language via a shared encoder and dual decoders for mask and text generation. Experiments show that ForgeryTalker achieves competitive performance on both subtasks, i.e., 59.3 CIDEr and 73.67 IoU, establishing a strong baseline for explainable multimedia forensics. Our dataset and code are available at: https://github.com/NattyLianJc/Generating-Attribution-Reports.
While Reinforcement Learning (RL) has advanced LLM reasoning, applying it to long-context scenarios is hindered by sparsity of outcome rewards. This limitation fails to penalize ungrounded "lucky guesses," leaving the critical process of needle-in-a-haystack evidence retrieval largely unsupervised. To address this, we propose EAPO (Evidence-Augmented Policy Optimization). We first establish the Evidence-Augmented Reasoning paradigm, validating via Tree-Structured Evidence Sampling that precise evidence extraction is the decisive bottleneck for long-context reasoning. Guided by this insight, EAPO introduces a specialized RL algorithm where a reward model computes a Group-Relative Evidence Reward, providing dense process supervision to explicitly improve evidence quality. To sustain accurate supervision throughout training, we further incorporate an Adaptive Reward-Policy Co-Evolution mechanism. This mechanism iteratively refines the reward model using outcome-consistent rollouts, sharpening its discriminative capability to ensure precise process guidance. Comprehensive evaluations across eight benchmarks demonstrate that EAPO significantly enhances long-context reasoning performance compared to SOTA baselines.
MavenCoder: Competitive Code Generation via Model Adaptive Planning Strategies and Multi-Perspective Verification Enhancement
PDF ↗With the rapid advancement of large language models (LLMs), automated code generation has made remarkable progress. Recent studies explore multi-agent collaboration and adopt planning–coding–debugging workflows to enhance performance. However, these approaches are constrained by rigid, predefined workflows that fail to flexibly adjust their plans and lack effective verification of intermediate reasoning steps. In this work, we propose MavenCoder, a model-adaptive and verification–enhanced framework for competition-level code generation. MavenCoder leverages adaptive assessment aligned with the model’s capabilities to select planning strategies, while providing timely feedback and correction via multi-perspective verification. This adaptive problem-solving paradigm mitigates earlier limitations by enabling flexible planning and timely error correction. Compared with existing state-of-the-art approaches, MavenCoder achieves superior pass@1 results across multiple benchmarks, achieving 87.5% on LiveCodeBench, 93.9% on HumanEval+, 81.7% on MBPP+, and 46.1% on CodeContests, outperforming recent agent-based systems with improvement exceeding 3%–40%.
Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a mechanistic account based on the geometry of feature superposition. Because features are encoded in overlapping, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this mechanism and empirically test it across multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, gpt-oss 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach—filtering training samples nearest to toxic features—reduces misalignment by 34.5%, substantially outperforming random removal and achieving stronger mitigation than LLM-as-a-judge–based filtering. Our study explains emergent misalignment through feature superposition, providing a basis for understanding and mitigating this phenomenon.
The goal of this work is to develop a universal approach for aligning subtitles (i.e., spoken language text with corresponding timestamps) to continuous sign language videos. Prior approaches typically rely on end-to-end training tied to a specific language or dataset, which limits their generality. In contrast, our method Segment, Embed, and Align (SEA) provides a single framework that works across multiple languages and domains. SEA leverages two pretrained models: the first to segment a video sequence into individual signs and the second to embed each sign video clip into a shared latent space with text. Alignment is subsequently performed with a lightweight dynamic programming procedure that runs efficiently on CPU within a minute, even for hour-long episodes. SEA is flexible and can adapt to a wide range of scenarios, utilizing resources from small lexicons to large continuous corpora. Experiments on four sign language datasets demonstrate state-of-the-art alignment performance, highlighting the potential of SEA to generate high-quality parallel data for advancing sign language processing.
Translation with Thought: Difficulty-Adaptive Reasoning via Reinforcement Learning for Multi-Domain Machine Translation
PDF ↗Multi-domain machine translation (MDMT) poses a unique challenge due to varying levels of linguistic complexity across domains. Inspired by human translators’ ability to adapt reasoning effort based on difficulty, we propose TwT (Translation with Thought), a resource-rational framework that learns to modulate inference between intuitive and deliberate reasoning. TwT is trained in two stages: (1) supervised fine-tuning on difficulty-aware long chain-of-though traces distilled from DeepSeek-R1 and rewritten by GPT-4o to reflect human-like reasoning economy, and (2) reinforcement learning with a hybrid reward to optimize translation quality and reasoning efficiency. Evaluated on 15 benchmarks spanning in-domain and out-of-domain settings, as well as 3 seen and 59 unseen languages, with ablations across three backbone models, TwT-7B and TwT-14B outperform much larger SOTA reasoning models in translation quality, while reducing token usage by 32–60%. These results confirm that aligning translation behavior with cognitive principles enables robust generalization, high translation quality, and efficient reasoning in MDMT.