Prevailing safety alignment methods still leave Large Language Models (LLMs) vulnerable to sophisticated jailbreak attacks. To bolster defenses, explicit reasoning mechanisms like Safety-oriented Chain-of-Thought (SCoT) have emerged, significantly enhancing robustness. However, this transparency introduces a critical trade-off: the exposed reasoning process itself becomes a new attack surface, risking the leakage of harmful information and revealing the model’s safety logic to adversaries. This paper directly confronts this dilemma, asking: Can we achieve the full benefits of deliberative safety without the costs of explicit reasoning generation? We propose Safety Reasoning Internalization to make the deliberative process in SCoT "available but not visible". This approach is grounded in a key theoretical insight: the corrective influence of an SCoT can be effectively approximated by a targeted, low-rank update to the model’s Feed-Forward Network (FFN) layers. We operationalize this through Hierarchical Internalization of Adversarially-Guided Reasoning (HIAR), a layer-wise safety alignment framework that internalizes safety reasoning into an implicit computational pathway using Low-Rank Adaptation (LoRA). HIAR enables the model to reach a safe conclusion within a single forward pass, entirely eliminating the need to generate vulnerable SCoT text. Extensive experiments on various LLMs demonstrate that HIAR achieves a 43% lower Attack Success Rate (ASR) against distinct jailbreak attacks compared to strong baselines.
论文检索
输入标题、作者或关键词，从 8,216 篇学术成果中精准定位
Competitive programming has become a rigorous benchmark for evaluating the reasoning and problem-solving capabilities of large language models (LLMs). The International Olympiad in Informatics (IOI) stands out as one of the most prestigious annual competitions in competitive programming and has become a key benchmark for comparing human and AI-level programming ability. While several proprietary models have been claimed to achieve gold medal-level performance at the IOI, often with undisclosed methods, achieving comparable results with open-weight models remains a significant challenge. In this paper, we present GenCluster, a scalable and reproducible test-time compute framework that attains IOI gold-level performance using open-weight models. It combines large-scale generation, behavioral clustering, ranking, and a round-robin submission strategy to efficiently explore diverse solution spaces under limited validation budgets. Our experiments show that the performance of our proposed approach scales consistently with available compute, narrowing the gap between open and closed systems. Notably, we will show that GenCluster can achieve a gold medal at IOI 2025 for the first time with an open-weight model gpt-oss-120b, setting a new benchmark for transparent and reproducible evaluation of reasoning in LLMs.
LLMs are increasingly deployed in dynamic, real-world settings, where the distribution of user prompts can shift substantially over time as new tasks, prompts, and users are introduced to a deployed model. Such natural prompt distribution shift poses a major challenge to LLM reliability, particularly for specialized models designed for narrow domains or user populations. Despite attention to out-of-distribution robustness, there is very limited exploration of measuring natural prompt distribution shift in prior work, and its impact on deployed LLMs remains poorly understood. We introduce the LLM Evaluation under Natural prompt Shift (LENS) framework: a data-centric approach for quantifying natural prompt distribution shift and evaluating its effect on the performance of deployed LLMs. We perform a large-scale evaluation using 192 real-world post-deployment prompt shift settings over time, user group, and geographic axes, training a total of 81 models on 4.68M training prompts, and evaluating on 57.6k prompts. We find that even moderate shifts in user prompt behavior correspond with large performance drops (73% average loss) in deployed LLMs. This performance degradation is particularly prevalent when users from different latent groups and geographic regions interact with models and is correlated with natural prompt distribution shift over time. We systematically characterize how LLM instruction following ability degrades over time and between user groups. Our findings highlight the critical need for data-driven monitoring to ensure LLM performance remains stable across diverse and evolving user populations.
As LLMs rapidly advance and enter real-world use, their privacy implications are increasingly important. We study an authorship de-anonymization threat: using LLMs to link anonymous documents to their authors, potentially compromising settings such as double-blind peer review. We propose De-Anonymization at Scale (DAS), a large-language-model–based method for attributing authorship among tens of thousands of candidate texts. DAS uses a sequential progression strategy: it randomly partitions the candidate corpus into fixed-size groups, prompts an LLM to select the text most likely written by the same author as a query text, and iteratively re-queries the surviving candidates to produce a ranked top-k list. To make this practical at scale, DAS adds a dense-retrieval prefilter to shrink the search space and a majority-voting–style aggregation over multiple independent runs to improve robustness and ranking precision. Experiments on anonymized review data show DAS can recover same-author texts from pools of tens of thousands with accuracy well above chance, demonstrating a realistic privacy risk for anonymous platforms. On standard authorship benchmarks (Enron emails and blog posts), DAS also improves both accuracy and scalability over prior approaches, highlighting a new LLM-enabled de-anonymization vulnerability.
Long contexts break transformers: attention scores dilute across thousands of tokens, critical information gets lost in the middle, and the model cannot adapt to novel patterns at inference time. Recent work on test-time adaptation addresses this by maintaining a form of working memory—transient parameters updated on the current context—but existing approaches employ uniform write policies that waste computation on low-value regions and suffer from high gradient variance across semantically heterogeneous contexts. In this work, we reframe test-time adaptation as a budget-constrained memory consolidation problem, asking: given limited computational budget, which parts of the context should be consolidated into working memory? We propose GDWM (\textbf{G}ated \textbf{D}ifferentiable \textbf{W}orking \textbf{M}emory), a framework that introduces a Write Controller to gate the memory consolidation process. Our controller estimates Contextual Utility—an information-theoretic measure quantifying how much each region depends on long-range context—and allocates gradient steps accordingly, subject to a coverage constraint that ensures global representation. Theoretically, we prove that our chunk-restricted sampling strategy reduces gradient variance by eliminating inter-chunk variance via the Law of Total Variance. Experiments on ZeroSCROLLS and LongBench v2 benchmarks demonstrate that GDWM achieves comparable or superior performance with 4 ×fewer gradient steps compared to uniform baselines—excelling on sparse-information tasks (+6–13% on Qasper, +5–13% on GovReport for smaller models) while revealing principled trade-offs on dense-coverage tasks, establishing a new efficiency-performance Pareto frontier for test-time adaptation.
Grounded claim factuality checking is important for large language model (LLM) applications such as retrieval-augmented generation, as it helps users assess the correctness of generated outputs. Existing metrics using entailment classifiers require dataset-specific threshold tuning, while LLM-based approaches often use direct prompting, which underutilises the reasoning capabilities of LLMs. We address this by formulating grounded claim factuality checking as a true/false reading comprehension task and prompting LLMs with explicit test-taking strategies for efficient reasoning. Our method reduces token usage by over 80% compared to unguided open-ended reasoning, and achieves competitive performance to more expensive alternatives across two factuality benchmarks, setting a new state of the art on one. To further reduce inference cost, we train small language models (SLMs) to replace LLMs in the checking pipeline. Using supervised fine-tuning (SFT) and a self-revision mechanism, the SLMs learn to improve their factuality judgements. Experimental results show that the resulting SLMs perform on par with strong baselines, combining low inference costs with generating supporting rationales to support interpretability. Code and datasets will be released upon acceptance.
Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model’s valid-generation manifold. Finally, we introduce a new steering approach guided by this analysis that improves preference while better preserving utility.
SwissGov-RSD: A Human-annotated, Cross-lingual Benchmark for Token-level Recognition of Semantic Differences Between Related Documents
PDF ↗Recognizing semantic differences across documents is crucial for text generation evaluation and content alignment, especially in cross-lingual settings. However, as a standalone task, it has received little attention. We address this by introducing SwissGov-RSD, the first naturalistic, document-level, cross-lingual dataset for semantic difference recognition. It encompasses a total of 224 multi-parallel documents in English–German, English–French, and English–Italian with token-level difference annotations by human annotators.We evaluate a variety of open-source and closed-source large language models as well as encoder models across different fine-tuning settings on this new benchmark. Our results show that current automatic approaches perform poorly compared to their performance on monolingual, sentence-level, and synthetic benchmarks, revealing a considerable gap for both LLMs and encoder models.
Reinforcement Learning with Verifiable Rewards (RLVR) serves as a cornerstone technique for enhancing the reasoning capabilities of Large Language Models (LLMs). However, its training is often plagued by entropy collapse, a rapid decline in policy entropy that limits exploration and undermines training effectiveness. While recent works attempt to mitigate this issue via several heuristic entropy interventions, the underlying mechanisms remain poorly understood. In this work, we conduct comprehensive theoretical and empirical analyses of entropy dynamics in RLVR, offering two main insights: (1) We derive a tight approximation for token-level entropy change at each update step, revealing four governing factors and providing a unified theoretical framework of how existing methods influence entropy; (2) We reveal a fundamental limitation of recent approaches: they rely on heuristic adjustments to one or two of these factors, leaving other relevant factors unconsidered, thus inherently limiting their effectiveness. Motivated by these findings, we propose STEER, a principled entropy-modulation method that adaptively reweighs tokens based on theoretically-estimated entropy variations. Extensive experiments across six mathematical reasoning and three coding benchmarks demonstrate that STEER effectively mitigates entropy collapse and consistently outperforms state-of-the-art baselines.
Large Language Model (LLM) agents are reshaping the industrial landscape. However, most practical agents remain human-designed because tasks differ widely, making them labor-intensive to build. This situation poses a central question: can we automatically create and adapt domain agents in the wild? While several recent approaches have sought to automate agent creation, they typically treat agent generation as a black-box procedure and rely solely on final performance metrics to guide the process. Such strategies overlook critical evidence explaining why an agent succeeds or fails, and often require high computational costs. To address these limitations, we propose ReCreate, an experience-driven framework for the automatic creation of domain agents. ReCreate systematically leverages agent interaction histories, which provide rich concrete signals on both the causes of success or failure and the avenues for improvement. Specifically, we introduce an agent-as-optimizer paradigm that effectively learns from experience via three key components: (i) an experience storage and retrieval mechanism for on-demand inspection; (ii) a reasoning–creating synergy pipeline that map execution experience into scaffold edits; and (iii) hierarchical updates that abstract instance-level details into reusable domain patterns. In experiments across diverse domains, ReCreate consistently outperforms human-designed agents and existing automated agent generation methods, even when starting from minimal seed scaffolds.
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.
The evolution of morality presents a puzzle: natural selection should favor self-interest, yet humans developed moral systems promoting altruism. Traditional approaches must abstract away cognitive processes, leaving open how cognitive factors shape moral evolution. We introduce an LLM-based agent simulation framework that brings cognitive realism to this question: agents with varying moral dispositions perceive, remember, reason, and decide in a simulated prehistoric hunter-gatherer society. This enables us to manipulate factors that traditional models cannot represent—such as moral type observability and communication bandwidth—and to discover emergent cognitive mechanisms from agent interactions. Across 20 runs spanning four settings, we find that cooperation and mutual help are the central driver of evolutionary survival, with universal and reciprocal morality exhibiting the most stable outcomes across conditions while selfishness is strongly disfavoured. Beyond cooperation itself, we further identify cognition as a central mediator—most clearly through a cost of moral judgment that shifts the winning moral type across settings, with a self-purging effect among selfish agents as an additional cognitive pattern. We validate robustness across multiple LLM backbones, architecture ablations, and prompt sensitivity analyses. This work establishes LLM-based simulation as a powerful new paradigm to complement traditional research in evolutionary biology and anthropology, opening new avenues for investigating the complexities of moral and social evolution.
Evaluating the safety robustness of LLMs is critical for their deployment. However, mainstream Red Teaming methods rely on online generation and black-box output analysis. These approaches are not only costly but also suffer from feedback latency, making them unsuitable for agile diagnostics after training a new model.To address this, we propose N-GLARE (A Non-Generative, Latent Representation-Efficient LLM Safety Evaluator). N-GLARE operates entirely on the model’s latent representations, bypassing the need for full text generation. It characterizes hidden layer dynamics by analyzing the APT (Angular-Probabilistic Trajectory) of latent representations and introducing the JSS (Jensen-Shannon Separability) metric.Experiments on over 40 models and 20 red teaming strategies demonstrate that the JSS metric exhibits high consistency with Red Teaming safety rankings at less than 1% token and runtime cost.
Recent advances in generative AI have significantly enhanced the realism of multimodal media manipulation, thereby posing substantial challenges to manipulation detection. Existing manipulation detection and grounding approaches predominantly focus on manipulation type classification under result-oriented supervision, which not only lacks interpretability but also tends to overfit superficial artifacts. In this paper, we argue that generalizable detection requires incorporating explicit forensic reasoning, rather than merely classifying a limited set of manipulation types, which fails to generalize to unseen manipulation patterns. To this end, we propose **REFORM**, a reasoning-driven framework that shifts learning from outcome fitting to process modeling. REFORM adopts a three-stage curriculum that first induces forensic rationales, then aligns reasoning with final judgments, and finally refines logical consistency via reinforcement learning. To support this paradigm, we introduce **ROM**, a large-scale dataset with rich reasoning annotations. Extensive experiments show that REFORM establishes new state-of-the-art performance with superior generalization, achieving 81.52% ACC on ROM, 76.65% ACC on DGM4, and 74.9 F1 on MMFakeBench.
While Large Language Models (LLMs) have excelled in textual reasoning, they struggle with mathematical domains like geometry that intrinsically rely on visual aids. Existing approaches to Visual Chain-of-Thought (VCoT) are often limited by rigid external tools or fail to generate the high-fidelity, strategically-timed diagrams necessary for complex problem-solving. To bridge this gap, we introduce MathCanvas, a comprehensive framework designed to endow unified Large Multimodal Models (LMMs) with intrinsic VCoT capabilities for mathematics. Our approach consists of two phases. First, a Visual Manipulation stage pre-trains the model on a novel 15.2M-pair corpus, comprising 10M caption-to-diagram pairs (MathCanvas-Imagen) and 5.2M step-by-step editing trajectories (MathCanvas-Edit), to master diagram generation and editing. Second, a Strategic Visual-Aided Reasoning stage fine-tunes the model on MathCanvas-Instruct, a new 219K-example dataset of interleaved visual-textual reasoning paths, teaching it when and how to leverage visual aids. To facilitate rigorous evaluation, we introduce MathCanvas-Bench, a challenging benchmark with 3K problems that require models to produce interleaved visual-textual solutions. Our model, BAGEL-Canvas, trained under this framework, achieves an 86% relative improvement over strong LMM baselines on MathCanvas-Bench, demonstrating excellent generalization to other public math benchmarks. Our work provides a complete toolkit—framework, datasets, and benchmark—to unlock complex, human-like visual reasoning in LMMs.
A Data-Efficient Path to Multilingual LLMs: Language Expansion via Post-training PARAM\Delta Integration into Upcycled MoE
PDF ↗Expanding Large Language Models(LLMs) to new languages is a costly endeavor, demanding extensive Continued Pre-Training(CPT) and data-intensive alignment. While recent data-free merging techniques attempt to bypass alignment by fusing a multilingual CPT-enhanced model with its instruct counterpart, they are plagued by a critical trade-off: mitigating parameter conflicts to preserve original abilities inevitably dilutes new language acquisition, and vice-versa. To resolve this conflict, we introduce , which upcycles a dense model into a Mixture-of-Experts(MoE) architecture, allocating different experts to different languages. Alignment ability is then transferred by grafting a MoE-expanded parameter delta(\Delta_{\text{instruct}}) to the CPT-enhanced base model, bypassing the complex alignment phase. Experiments demonstrate ’s superiority even against baselines with similar FLOPs or number of parameters; it improves performance on expanded languages while effectively preserving original capabilities. We further show our approach is highly applicable across different models and Post-training deltas.
Despite the recent success of large language models (LLMs) in time-series forecasting, most existing methods still adopt a Deep Synchronous Fusion strategy, where dense interactions between textual and temporal features are enforced at every layer of the network. This design overlooks the inherent granularity mismatch between modalities and leads to what we term semantic perceptual dissonance: high-level abstract semantics provided by the LLM become inappropriately entangled with the low-level, fine-grained numerical dynamics of time series, making it difficult for semantic priors to effectively guide forecasting. To address this issue, we propose TimeSAF, a new framework based on hierarchical asynchronous fusion. Unlike synchronous approaches, TimeSAF explicitly decouples unimodal feature learning from cross-modal interaction. It introduces an independent cross-modal semantic fusion trunk, which uses learnable queries to aggregate global semantics from the temporal and prompt backbones in a bottom-up manner, and a stage-wise semantic refinement decoder that asynchronously injects these high-level signals back into the temporal backbone. This mechanism provides stable and efficient semantic guidance while avoiding interference with low-level temporal dynamics. Extensive experiments on standard long-term forecasting benchmarks show that TimeSAF significantly outperforms state-of-the-art baselines, and further exhibits strong generalization in both few-shot and zero-shot transfer settings.
Large language models (LLMs) can effectively handle outdated information through knowledge editing. However, current approaches face two key limitations: **(I) Poor generalization:** Most approaches rigidly inject new knowledge without ensuring that the model can use it effectively to solve practical problems. **(II) Narrow scope:** Current methods focus primarily on structured fact triples, overlooking the diverse unstructured forms of factual information (e.g., news, articles) prevalent in real-world contexts. To address these challenges, we propose a new paradigm: teaching LLMs to edit knowledge via **Chain of Thoughts** (CoTs) reasoning (CoT2Edit). We first leverage language model agents for both structured and unstructured edited data to generate CoTs, building high-quality instruction data. The model is then trained to reason over edited knowledge through supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO). At inference time, we integrate Retrieval-Augmented Generation (RAG) to dynamically retrieve relevant edited facts for real-time knowledge editing. Experimental results demonstrate that our method achieves strong generalization across six diverse knowledge editing scenarios with **just a single round of training** on three open-source language models.
LegalChainReasoner: Grounding Criminal Judicial Opinion Generation via Structured Legal Chains
PDF ↗A criminal judicial opinion represents the judge’s disposition of a case, including the decision rationale and sentencing. Automatically generating such opinions can assist in analyzing sentencing consistency and provide judges with references to past similar cases. However, current research typically approaches this task by dividing it into two isolated subtasks: legal reasoning and sentencing prediction. This separation often leads to inconsistency between the reasoning and predictions, failing to meet real-world judicial requirements. Furthermore, prior studies rely on manually creating knowledge to enhance applicability, yet such methods remain limited in practical deployment. To address these limitations and better align with legal practice, we propose a new LegalAI task: Criminal Judicial Opinion Generation, which simultaneously produces both legal reasoning and sentencing decisions. To achieve this, we introduce LegalChainReasoner framework that applies structured legal chains to guide the model through comprehensive case assessments. By integrating factual premises, composite legal conditions, and sentencing conclusions, our approach ensures flexible knowledge injection and end-to-end opinion generation. Experiments on real-world, open-source Chinese legal case datasets demonstrate that our method outperforms baseline models.
A Survey of Reinforcement Learning for Large Language Models under Data Scarcity: Challenges and Solutions
PDF ↗Reinforcement learning (RL) has emerged as a powerful post-training paradigm for enhancing the reasoning capabilities of large language models (LLMs). However, reinforcement learning for LLMs faces substantial data scarcity challenges, including the limited availability of high-quality external supervision and the constrained volume of model-generated experience. These limitations make data-efficient reinforcement learning a critical research direction. In this survey, we present the first systematic review of reinforcement learning for LLMs under data scarcity. We propose a bottom-up hierarchical framework built around three complementary perspectives: the data-centric perspective, the training-centric perspective, and the framework-centric perspective. We develop a taxonomy of existing methods, summarize representative approaches in each category, and analyze their strengths and limitations. Our taxonomy aims to provide a clear conceptual foundation for understanding the design space of data-efficient RL for LLMs and to guide researchers working in this emerging area. We hope this survey offers a comprehensive roadmap for future research and inspires new directions toward more efficient and scalable reinforcement learning post-training for LLMs.