论文检索

Niu Lian, Yuting Wang, Hanshu Yao, Jinpeng Wang, Bin Chen, Yaowei Wang, Min Zhang, Shu-Tao Xia

While multimodal large language models have demonstrated impressive short-term reasoning, they struggle with long-horizon video understanding due to limited context windows and static memory mechanisms that fail to mirror human cognitive efficiency. Existing paradigms typically fall into two extremes: vision-centric methods that incur high latency and redundancy through dense visual accumulation, or text-centric approaches that suffer from detail loss and hallucination via aggressive captioning. To bridge this gap, we propose **MM-Mem**, a pyramidal multimodal memory architecture grounded in *Fuzzy-Trace Theory*. **MM-Mem** structures memory hierarchically into a *Sensory Buffer*, *Episodic Stream*, and *Symbolic Schema*, enabling the progressive distillation of fine-grained perceptual traces (*verbatim*) into high-level semantic schemas (*gist*).Furthermore, to govern the dynamic construction of memory, we derive a Semantic Information Bottleneck objective and introduce SIB-GRPO to optimize the trade-off between memory compression and task-relevant information retention.In inference, we design an entropy-driven top-down memory retrieval strategy.Extensive experiments across 4 benchmarks confirm that **MM-Mem** achieves state-of-the-art performance on both offline and streaming tasks, demonstrating robust generalization and validating the effectiveness of cognition-inspired memory organization.Code and associated configurations are publicly available at ‘https://github.com/EliSpectre/MM-Mem‘.

Eva Vanmassenhove

Multilingual Large Language Models considerably changed how technologies influence language. While previous technologies could mediate or assist humans, there is now a tendency to offload the task of writing itself to these technologies, enabling models to change our languages more directly. While they provide us quick access to information and impressively fluent output, beneath their (apparent) sophistication lies a subtle, insidious threat: the gradual decline and loss of linguistic diversity. In this position paper, I explore how model collapse, with a particular focus on translation technology, can lead to the loss of linguistic forms, grammatical features, and cultural nuance. Model collapse refers to the consequences of self-consuming training loops, where automatically generated data (re-)enters the training data, leading to a gradual distortion of the data distribution and the underrepresentation of low-probability linguistic phenomena. Drawing on recent work in Computer Vision, Natural Language Processing and Machine Translation, I argue that the many tails of our linguistic distributions might be vanishing, and with them, the narratives and identities they carry. This paper is a call to resist linguistic flattening and to reimagine Natural Language Processing as a field that encourages, values and protects expressive multilingual diversity and creativity.

Chonghua Liao, Ke Wang, Yuchuan Wu, Ruoran Li, Fei Huang, Yongbin Li

Role-playing agents (RPAs) require balancing multiple objectives, such as instruction following, persona consistency, and stylistic fidelity, which are not always perfectly aligned across different dimensions. While prior work has primarily relied on supervised fine-tuning or reinforcement learning with scalarized rewards, these approaches do not explicitly address the coordination of multiple reward dimensions during optimization. We present **MOA** (**M**ulti-**O**bjective **A**lignment), a reinforcement-learning framework that enables multi-dimensional, fine-grained rubric optimization for general RPAs. MOA introduces a novel multi-objective optimization strategy that trains simultaneously on multiple fine-grained rubrics to boost optimization performance. Besides, to address the issues of model output diversity and quality, we have also employed thought-augmented rollout with off-policy guidance. Experiments on PersonaGym and RoleMRC show that MOA consistently improves multi-dimensional role-playing performance over supervised and standard RL baselines. Under identical evaluation protocols, an 8B model trained with MOA reaches performance competitive with strong closed-source models across multiple evaluation dimensions. These results suggest that MOA provides a practical framework for training more capable general-purpose role-playing agents.

Ruizi Han, Miao Zhang, Ziyue Qiao, Liqiang Nie

Efficient long-context inference remains a major challenge for large language models (LLMs), as the cost of attention computation during auto-regressive decoding grows linearly with the context length. Recent sparse attention methods attempt to reduce the computational burden by selecting a subset of tokens at each step, while most rely on static importance scores that are repeatedly computed over the entire cache, overlooking the relational dynamics of the decoding process. In this work, we revisit sparse attention in LLMs and propose to model token importance as a dynamic process that evolves over decoding steps and propagates through model layers. To efficiently measure token importance, we propose two lightweight mechanisms: (1) Cross-Step Accumulation, which incrementally maintains long-term, query-agnostic importance via decayed accumulation of sparse attention scores, avoiding recomputing the importance of decoded tokens; and (2) Cross-Layer Propagation, which leverages the model’s intrinsic Retrieval Heads to compute query-aware indices and efficiently propagate them across layers; Together, these mechanisms preserve both stable context memory and adaptive query relevance while reduce redundant computation. We evaluate our approach on PG-19, RULER, LongBench, and mathematical reasoning benchmarks using models employing Multi-Head and Grouped-Query Attention. Under varying KV cache budgets, our method consistently outperforms prior sparse attention baselines, approaches full attention performance in most settings, and achieves speedups of up to 5.36\times for attention latency and 2.33\times for end-to-end decoding. Our code is available at: https://github.com/iLearn-Lab/ACL26-EvoSparse.

Sihang Zhao, Kangrui Yu, Youliang Yuan, Pinjia He, Hongyi Wen

Large Language Models (LLMs) have been widely explored in educational scenarios. We identify a critical vulnerability in current educational LLMs, pedagogical jailbreaks, where students use answer-inducing prompts to elicit solutions rather than scaffolded instructions. To enable systematic study, we unify and formalize safe, helpful, and pedagogical behaviors with a knowledge-mastery graph and introduce SHAPE, a benchmark of 9,087 student-question pairs for evaluating tutoring behavior under adversarial pressure. We propose a graph-augmented tutoring pipeline that infers prerequisite concepts from queries, identifies mastery gaps, and routes generation between instructing and problem-solving via explicit gating. Experiments across multiple LLMs show that our method yields significantly improved safety under two pedagogical jailbreak settings, while maintaining near-ceiling helpfulness under the same evaluation protocol. Our code and data are available at https://github.com/MAPS-research/SHaPE

Kaitong Cai, Jusheng Zhang, Keze Wang

Hybrid offline–online reinforcement learning (O2O RL) promises both sample efficiency and robust exploration, but suffers from instability due to distribution shift between offline and online data. We introduce RLPD-GX, a framework that decouples policy optimization from safety enforcement: a reward-seeking learner explores freely, while a projection-based guardian guarantees rule-consistent execution and safe value backups. This design preserves the exploratory value of online interactions without collapsing to conservative policies. To further stabilize training, we propose dynamic curricula that gradually extend temporal horizons and anneal offline–online data mixing. We prove convergence via a contraction property of the guarded Bellman operator, and empirically show state-of-the-art performance on Atari-100k, achieving a normalized mean score of 3.02 (+45% over prior hybrid methods) with stronger safety and stability. Beyond Atari, ablations demonstrate consistent gains across safety-critical and long-horizon tasks, underscoring the generality of our design. Extensive and comprehensive results highlight decoupled safety enforcement as a simple yet principled route to robust O2O RL, suggesting a broader paradigm for reconciling exploration and safety in reinforcement learning.

Xiaoao Zhu, Jie Ren, Zhiqiang Li, Jie Zheng, Zhanyong Tang, Zheng Wang

Lifting stripped and highly optimized binaries to the canonical compiler intermediate representation (IR) enables program analysis when source code is unavailable. However, compiler optimizations severely distort control-flow and data-flow structure, making existing rule-based and LLM-based decompilation approaches brittle. We present BRIDGE, a system that reliably lifts optimized binaries to analysis-friendly compiler IR. BRIDGE combines control-flow-aware retrieval-augmented generation with feedback-driven verification. It uses pseudo-probe instrumentation to align optimized binary fragments with normalized IR semantics, and then employs an iterative refinement loop guided by static analysis and runtime feedback to improve executability and semantic consistency. We evaluate BRIDGE on HumanEval-Decompile and MBPP, lifting x86-64 and ARM64 binaries to LLVM IR. BRIDGE outperforms seven baselines, achieving an average of over 30% higher re-executability than the strongest general-purpose LLM baseline.

Ahmed Ewais, Ahmed Hashish, Amr Ali

Large language models encode extensive world knowledge valuable for zero-shot named entity recognition. However, their causal attention mechanism, where tokens attend only to preceding context, prevents effective token classification when disambiguation requires future context. Existing approaches use LLMs generatively, prompting them to list entities or produce structured outputs, but suffer from slow autoregressive decoding, hallucinated entities, and formatting errors. We propose Just Pass Twice (JPT), a simple yet effective method that enables causal LLMs to perform discriminative token classification with full bidirectional context. Our key insight is that concatenating the input to itself lets each token in the second pass attend to the complete sentence, requiring no architectural modifications. We combine these representations with definition-guided entity embeddings for flexible zero-shot generalization. Our approach achieves state-of-the-art results on zero-shot NER benchmarks, surpassing the previous best method by +7.9 F1 on average across CrossNER and MIT benchmarks, being over 20× faster than comparable generative methods.

Yuexiao Liu, Lijun Li, Xingjun Wang, Jing Shao

Recent advancements in Reinforcement Learning with Verifiable Rewards (RLVR) have gained significant attention due to their objective and verifiable reward signals, demonstrating strong performance in reasoning and code generation tasks. However, the potential safety risks associated with RLVR remain underexplored. This paper presents HarmRLVR, the first systematic investigation into the alignment reversibility risk of RLVR. We show that safety alignment can be rapidly reversed using GRPO with merely 64 harmful prompts without responses, causing models to readily comply with harmful instructions. Across five models from Llama, Qwen, and DeepSeek, we empirically demonstrate that RLVR-based attacks elevate the average harmfulness score to 4.94 with an attack success rate of 96.01%, significantly outperforming harmful fine-tuning while preserving general capabilities. Our findings reveal that RLVR can be efficiently exploited for harmful alignment, posing serious threats to open-source model safety.

Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Zhou, Yankai Lin, Ji-Rong Wen, Chongxuan Li

Masked diffusion language models present a promising paradigm for language modeling, yet the systematic theoretical analysis and comprehensive empirical validation of their alignment on general tasks remain relatively underexplored. In this paper, we identify the primary challenge for this problem: the high variance in Evidence Lower Bound (ELBO)-based likelihood estimates required for preference optimization. To address this issue, we propose *Variance-Reduced Preference Optimization* (VRPO), a framework that formally analyzes the bias and variance of the preference optimization loss and gradient based on Direct Preference Optimization, showing both are governed by a score-estimator variance. Building on this foundation, we introduce multiple unbiased variance reduction strategies, including optimal budget allocation and antithetic sampling, to improve alignment performance. We demonstrate the effectiveness of VRPO by applying it to LLaDA, a large diffusion language model. The resulting model, LLaDA 1.5, consistently outperforms its SFT-only predecessor consistently across various general benchmarks, such as mathematics (GSM8K +4.7), coding (HumanEval +3.0, MBPP +1.8), and alignment (IFEval +4.0, Arena-Hard +4.3). Furthermore, LLaDA 1.5 demonstrates a highly competitive mathematical performance compared to other strong language MDMs and ARMs. Our model is available at https://huggingface.co/GSAI-ML/LLaDA-1.5.

Akira Kawabata, Saku Sugawara

Rubric-augmented verification guides reward models with explicit evaluation criteria, yielding more reliable judgments than single-model verification.However, most existing methods require costly rubric annotations, limiting scalability.Moreover, we find that rubric generation is vulnerable to a failure of cooperation; low-quality rubrics actively mislead reward models rather than help. Inspired by the principle of cooperative communication, we propose Cooperative yet Critical reward modeling (C2), a framework that significantly improves reward model judgments by having the reward model critically collaborate with a rubric generator trained solely from binary preferences.In C2, we synthesize helpful and misleading rubric pairs by measuring how each rubric shifts the reward model toward or away from the correct preference. Using these contrastive pairs, we train a cooperative rubric generator to propose helpful rubrics, and a critical verifier to assess rubric validity before making its judgment, following only rubrics it deems helpful at inference time. C2 outperforms reasoning reward models trained on the same binary preferences, with gains of up to 6.5 points on RM-Bench and 6.0 points length-controlled win rate on AlpacaEval 2.0. Without external rubric annotations, C2 enables an 8B reward model to match performance achieved with rubrics from a 4\times larger model.Overall, our work demonstrates that eliciting deliberate cooperation in rubric-augmented verification makes reward models more trustworthy in a scalable way.

Lichao Wang, ZhaoXing Ren, Tianzhuo Yang, Jiaming Ji, Chi Harold Liu, Yaodong Yang, Juntao Dai

As Large Language Model (LLM) agents increasingly leverage the Model Context Protocol (MCP) to operate in complex environments, the expansion of their action spaces offers agents unsafe capabilities and underscores the risk of power-seeking. While broad action space and greater environment influence are essential for task fulfillment, they creates a fragile risk surface where minor errors or hallucinations are magnified into catastrophic failures. In response, we propose SafeMCP, a server-side defense plugin that constrains tool acquisition via predictive reasoning regarding future safety risks. SafeMCP utilizes an internal world model for look-ahead reasoning to implement a two-tier defense: proactive tool filtering to constrain hazardous power expansion and immediate intervention as a fail-safe. To train SafeMCP, we introduce a three-stage pipeline comprising environmental dynamic grounding, safe policy initialization, and reinforcement learning (RL) with dual verifiable rewards. Experiments on PowerSeeking Bench, ToolEmu, and AgentHarm show that SafeMCP achieves a safe equilibrium, effectively mitigating risks while preserving agent utility.

Adam Štorek, Mukur Gupta, Noopur Bhatt, Aditya Gupta, Janie Kim, Prashast Srivastava, Suman Jana

AI coding assistants automatically gather context from potentially untrusted sources to generate code recommendations. We introduce Cross-Origin Context Poisoning (XOXO), a novel attack that exploits this automatic context inclusion by subtly manipulating code without changing its semantics. Attackers introduce semantics-preserving transformations (e.g., renamed variables) to shared code, causing AI assistants to unknowingly recommend vulnerable code patterns to victims. To systematically identify effective transformations, we present Greedy Cayley Graph Search (GCGS), a black-box algorithm that efficiently composes transformations to identify adversarial inputs. Our evaluation demonstrates XOXO’s effectiveness at making LLMs generate buggy and vulnerable code, achieving average attack success rates of 73.20% against eight state-of-the-art models including GPT 4.1 and Claude 3.5 Sonnet v2, with vulnerability injection rates up to 66.67%. We also demonstrate a real-world attack against GitHub Copilot, highlighting critical security gaps in current AI coding tools.

Weihao Xuan, Qingcheng Zeng, Heli Qi, Yunze Xiao, Junjue Wang, Naoto Yokoya

Autonomous agents based on large language models (LLMs) are rapidly evolving to handle multi-turn tasks, but ensuring their trustworthiness remains a critical challenge. A fundamental pillar of this trustworthiness is calibration, which refers to an agent’s ability to express confidence that reliably reflects its actual performance. While calibration is well-established for static models, its dynamics in tool-integrated agentic workflows remain under-explored. In this work, we systematically investigate verbalized calibration in tool-use agents, revealing a fundamental confidence dichotomy driven by tool type. Specifically, our pilot study identifies that evidence tools (e.g., web search) systematically induce severe overconfidence due to inherent noise in retrieved information, while verification tools (e.g., code interpreters) can ground reasoning through deterministic feedback and mitigate miscalibration. To robustly improve calibration across tool types, we propose a reinforcement learning (RL) fine-tuning framework that jointly optimizes task accuracy and calibration, supported by a holistic benchmark of reward designs. We demonstrate that our trained agents not only achieve superior calibration but also exhibit robust generalization from local training environments to noisy web settings and to distinct domains such as mathematical reasoning. Our results highlight the necessity of domain-specific calibration strategies for tool-use agents. More broadly, this work establishes a foundation for building self-aware agents that can reliably communicate uncertainty in high-stakes, real-world deployments.

Xinhang Ma, William Yeoh, Ning Zhang, Yevgeniy Vorobeychik

Knowledge distillation is a widely adopted technique for transferring capabilities from LLMs to smaller, more efficient student models.However, unauthorized use of knowledge distillation takes unfair advantage of the considerable effort and cost put into developing frontier models.We investigate methods for modifying teacher-generated reasoning traces to achieve two objectives that deter unauthorized distillation: (1) anti-distillation, or degrading the training usefulness of query responses, and (2) API watermarking, which embeds verifiable signatures in student models.We introduce several approaches for dynamically rewriting a teacher’s reasoning outputs while preserving answer correctness and semantic coherence.Two of these leverage the rewriting capabilities of LLMs, while others use gradient-based techniques.Our experiments show that a simple instruction-based rewriting approach achieves a strong anti-distillation effect while maintaining or even improving teacher performance.Furthermore, we show that our rewriting approach also enables embedding watermarks that can be reliably detectedwith essentially no false alarms.Our code is available at https://github.com/xhOwenMa/trace-rewriting.

Xin Shi, Chao Zhang, Yifan Zhu, Xueqiao Zhang, Yawei Luo

The development of LLM-based tutor agents faces challenges in simultaneously ensuring adherence to pedagogical principles and achieving optimal pedagogical effectiveness, particularly in dynamic, multi-turn interactions. Existing methods are often constrained by static data or sparse reward signals in online settings. To address this gap, we propose \textbf{M}ulti-\textbf{H}orizon \textbf{P}reference \textbf{O}ptimization (\textbf{MHPO}), a novel framework that iteratively refines tutor agents using a multi-horizon reward function within a dynamic teacher-student simulation environment. Specifically, this reward function is designed to capture both turn-level pedagogical quality and trajectory-level pedagogical effectiveness, which is estimated via Monte Carlo rollouts. We further investigate two distinct strategies to aggregate these rewards for policy optimization. Our experiments demonstrate that MHPO significantly enhances base model performance, achieving a superior balance between principles and effectiveness compared to various baselines.

Yubo Wang, Juntian Zhang, Yichen Wu, Yankai Lin, Nils Lukas, Yuhan Liu

While Chain-of-Thought empowers Large Vision-Language Models with multi-step reasoning, explicit textual rationales suffer from an information bandwidth bottleneck, where continuous visual details are discarded during discrete tokenization. Recent latent reasoning methods attempt to address this challenge, but often fall prey to premature semantic collapse due to rigid autoregressive objectives. In this paper, we propose Laser, a novel paradigm that reformulates visual deduction via Dynamic Windowed Alignment Learning. Instead of forcing a point-wise prediction, Laser aligns the latent state with a dynamic validity window of future semantics. This mechanism enforces a "Forest-before-Trees" cognitive hierarchy, enabling the model to maintain a probabilistic superposition of global features before narrowing down to local details. Crucially, Laser maintains interpretability via decodable trajectories while stabilizing unconstrained learning via Self-Refined Superposition. Extensive experiments on 6 benchmarks demonstrate that Laser achieves state-of-the-art performance among latent reasoning methods, surpassing the strong baseline Monet by 5.03% on average. Notably, it achieves these gains with extreme efficiency, reducing inference tokens by more than 97%, while demonstrating robust generalization to out-of-distribution domains.

Jianshu Zhang, Chengxuan Qian, Haosen Sun, Haoran Lu, Dingcheng Wang, Letian Xue, Han Liu

Estimating task progress requires long-horizon and dynamic reasoning, going beyond static visual perception. Although Vision-Language Models (VLMs) excel at describing what is visible in a single observation, it remains unclear whether they can infer how far a task has progressed from partial information. To study this question, we introduce Progress-Bench, a benchmark with over 3K instances for evaluating progress reasoning from a single observation. We further examine a human-inspired two-stage paradigm that combines episodic retrieval with mental simulation. We instantiate this paradigm through both training-free prompting and a training-based approach using the automatically curated ProgressLM-45K dataset. Experiments on 14 VLMs show that most models struggle with reliable progress estimation, and that training-free reasoning provides only limited and model-dependent benefits. In contrast, the training-based ProgressLM-3B achieves consistent improvements in accuracy, robustness to viewpoint variation, and handling of unanswerable cases, despite its small scale. Additional analyses reveal common failure patterns in existing VLMs and clarify when and why progress reasoning succeeds or fails.

Paweł Batorski, Paul Swoboda

LLMs are highly sensitive to prompt design, but handcrafting effective prompts is difficult and often requires intricate crafting of few-shot examples. We propose a fast automatic prompt construction algorithm that augments human instructions by generating a small set of few shot examples. Our method iteratively replaces/drops/keeps few-shot examples using Monte Carlo Shapley estimation of example utility. For faster execution, we use aggressive subsampling and a replay buffer for faster evaluations. Our method can be run using different compute time budgets. Under a limited budget, it outperforms prior automatic prompting methods on text simplification and mathematical reasoning (GSM8K, DeepMath, Math500), while achieving second-best results on classification and summarization and third-best on MedQA. With an extended, yet still modest budget, PIAST sets a new state of the art among automatic prompting methods on classification, simplification, GSM8K, DeepMath, and Math500. Overall, our results suggest that optimizing in-context examples, rather than exhaustively searching over instruction rewrites is the dominant lever for fast and data-efficient prompt engineering. We will release code and data upon acceptance.

Wenzhuo Xu, Zhipeng Wei, Zonghao Ying, Deyue Zhang, Dongdong Yang, Xiangzheng Zhang, Quanchen Zou

Multimodal Large Language Models (MLLMs) are vulnerable to jailbreak attacks, which can elicit harmful responses from MLLMs. Many MLLMs support multi-image inputs, inadvertently introducing new vulnerabilities due to less efforts on multi-image safety alignment. Previous MLLM jailbreak methods only uses a single image, which restricts the attack space: they cannot distribute harmful requests across multiple images, carry abundant information, or exploit additional visual reasoning tasks to distract MLLMs. To address these limitations, in this paper, we propose a compositional jailbreak framework, \textbf{DMN}, which leverages \textbf{D}istributed instruction, \textbf{M}ultimodal evidence and a \textbf{N}umber chain task to fully enhance the jailbreak performance. Extensive experiments show that DMN is highly effective for MLLM jailbreaking, e.g. achieving attack success rates of over 90% on GPT-4o, Gemini-2.5-pro and Claude Sonnet 4, surpassing other baselines by a large margin. This compositional, multi-image jailbreak strategy reveals fundamental weaknesses in their safety mechanisms.