论文检索

Yangyang Zhong, Yanmei Gu, Zhengqing Zang, Xiaomeng Li, Yuqi Ding, Xibei Jia, Yuting Shen, Zhenzhong Lan, Liwang Zhu, Weiping Liu 等

Masked Diffusion Language Models (MDLMs) promise parallel token generation and arbitrary-order decoding, yet it remains unclear to what extent current models truly realize these capabilities. We characterize MDLM behavior along two dimensions—parallelism strength and generation order—using Average Finalization Parallelism (AFP) and Kendall’s τ. We evaluate eight mainstream MDLMs (up to 100B parameters) on 58 benchmarks spanning knowledge, reasoning, and programming. The results show that MDLMs still lag behind comparably sized autoregressive models, mainly because parallel probabilistic modeling weakens inter-token dependencies. Meanwhile, MDLMs exhibit adaptive decoding behavior: their parallelism and generation order vary significantly with the task domain, the stage of reasoning, and whether the output is correct. On tasks that require “backward information” (e.g., Sudoku), MDLMs adopt a solution order that tends to fill easier Sudoku blanks first, highlighting their advantages. Finally, we provide theoretical motivation and design insights supporting a Generate-then-Edit paradigm, which mitigates dependency loss while retaining the efficiency of parallel decoding.

Peng Chuang, Wei Zhang, Renshuai Tao, Xinhao Zhang, Jian Yang

Text-based web agents offer computational efficiency for autonomous web navigation, yet developing robust agents remains challenging due to the noisy and heterogeneous nature of real-world HTML. Standard Supervised Fine-Tuning (SFT) approaches fail in two critical dimensions: they lack discrimination capabilities to reject plausible but incorrect elements in densely populated pages, and exhibit limited generalization to unseen website layouts. To address these challenges, we introduce the Triton dataset (590k instances) and a progressive training curriculum. Triton is constructed via Structural-Semantic Hard Negative Mining, which explicitly mines topologically similar distractors, and a Dual-Agent Consensus pipeline that synthesizes diverse cross-domain tasks with strict verification. Building upon this foundation, our progressive curriculum produces three models: Triton-SFT-32B for basic imitation, Triton-ORPO-32B for robust discrimination via Odds Ratio Preference Optimization, and Triton-GRPO-32B for long-horizon consistency through Group Relative Policy Optimization. Empirical evaluation on Mind2Web demonstrates that Triton-GRPO-32B achieves state-of-the-art performance among open-source models with 58.7% Step Success Rate, surpassing GPT-4.5 (42.4%) and Claude-4.5 (41.4%) by over 16%, validating that specialized data curriculum outweighs raw parameter scale for web navigation.

Nguyen Viet Anh, Shiqian Zhao, Gia Dao, Runyi Hu, Yi Xie, Xiaobao Wu, Anh Tuan Luu

Recently, Large Reasoning Models (LRMs) have demonstrated superior logical capabilities compared to traditional Large Language Models (LLMs), gaining significant attention. Despite their impressive performance, the potential for stronger reasoning abilities to introduce more severe security vulnerabilities, though pointed out by some previous works, remains largely underexplored. Existing jailbreak methods often struggle to balance effectiveness with robustness against adaptive safety mechanisms. In this work, we propose SEAL, a novel jailbreak attack that targets LRMs through an adaptive encryption pipeline designed to override their reasoning processes and evade potential adaptive alignment. Specifically, SEAL introduces a stacked encryption approach that combines multiple ciphers to overwhelm the model’s reasoning capabilities, effectively bypassing built-in safety mechanisms. To further prevent LRMs from developing countermeasures, we incorporate two dynamic strategies—random and adaptive—that adjust the cipher length, order, and combination. Extensive experiments on real-world reasoning models, including DeepSeek-R1, Claude Sonnet, and OpenAI GPT-o4-mini, validate the effectiveness of our approach. Notably, SEAL achieves an attack success rate of 85.6% on GPT o4-mini, outperforming state-of-the-art baselines by a significant margin of 17.2%. Warning: This paper contains examples of inappropriate, offensive, and harmful content

Boyang Li, Hongzhe Shou, Yuanyuan Liang, JingBin Zhang, Fang Zhou

Existing Chinese toxic content detection methods mainly target sentence-level classification but often fail to provide readable and contiguous toxic evidence spans. We propose ToxiTrace, an explainability-oriented method for BERT-style encoders with three components: (1) CuSA, which refines encoder-derived saliency cues into fine-grained toxic spans with lightweight LLM guidance; (2) GCLoss, a gradient-constrained objective that concentrates token-level saliency on toxic evidence while suppressing irrelevant activations; and (3) ARCL, which constructs sample-specific contrastive reasoning pairs to sharpen the semantic boundary between toxic and non-toxic content. Experiments show that ToxiTrace improves classification accuracy and toxic span extraction while preserving efficient encoder-based inference and producing more coherent, human-readable explanations. The core training code is available at https://github.com/ZhouF-ECNU/ToxiTrace.

Chen Xiong, Zihao Wang, Rui Zhu, Tsung-Yi Ho, Pin-Yu Chen, Jingwei Xiong, Haixu Tang

Large Language Models (LLMs) rely on massive training datasets, often including proprietary data, which raises concerns about unauthorized usage and copyright infringement. Existing dataset inference methods typically require access to log probabilities or other internal signals, but many modern LLMs restrict such access, motivating token-only inference approaches. We propose CatShift, a token-only dataset inference framework based on catastrophic forgetting, where models overwrite prior knowledge when trained on new data. Fine-tuning an LLM on a subset of its training data induces larger output shifts than fine-tuning on unseen data. CatShift compares these shifts against those from a known non-member validation set to infer whether a dataset was included in training. Experiments on both open-source and API-based LLMs show that CatShift remains effective without logit access, enabling practical protection of proprietary datasets.

Zuchao Li, Qiwei Li, Yao Yao, Hai Zhao, Lefei Zhang, Bo Du

Chain-of-Thought (CoT) reasoning, while effective, suffers from an inherent mechanism flaw: linearity induces overthinking. Constrained by sequential generation, models often produce redundant narration and circular self-corrections to maintain logical context. We propose GoT-R1, a framework that fundamentally mitigates this by replacing verbose linear trajectories with high-density reasoning graphs. Unlike CoT, GoT-R1 decouples logic from narration, modeling deliberation as a structured topology of atomic units. We internalize this inductive bias via a two-stage regimen: synthesizing structural data to distill logical skeletons, followed by Group Relative Policy Optimization (GRPO) to explicitly reinforce topological integrity. Extensive evaluations across mathematical reasoning and instruction following demonstrate that GoT-R1 consistently outperforms state-of-the-art baselines. Crucially, it achieves these gains with significantly reduced token overhead, demonstrating that structured reasoning density offers a more robust and parsimonious alternative to the recursive verbosity of standard CoT. The GoT-R1 models are open-sourced on Hugging Face at: https://huggingface.co/collections/MYTH-Lab/got-r1.

Jianjie Zheng, Zhichen Liu, Zhanyu Shen, Jingxiang Qu, Guanhua Chen, Yile Wang, Yang Xu, Yang Liu, Sijie Cheng

Nowadays, wearable devices can continuously lifelog ambient conversations, creating substantial opportunities for memory systems. However, existing benchmarks primarily focus on online one-on-one chatting or human-AI interactions, thus neglecting the unique demands of real-world scenarios. Given the scarcity of public lifelogging audio datasets, we propose a hierarchical synthesis framework to curate LifeDialBench, a novel benchmark comprising two complementary subsets: EgoMem, built on real-world egocentric videos, and LifeMem, constructed using simulated virtual community. Crucially, to address the issue of temporal leakage in traditional offline settings, we propose an Online Evaluation protocol that strictly adheres to temporal causality, ensuring systems are evaluated in a realistic streaming fashion. Our experimental results reveal a counterintuitive finding: current sophisticated memory systems fail to outperform a simple RAG-based baseline. This highlights the detrimental impact of over-designed structures and lossy compression in current approaches, emphasizing the necessity of high-fidelity context preservation for lifelog scenarios.

Boyun Zhang, Zequn Xie, Fangming Feng, Zihan Zhang, Yongbo He, Chuxin Wang, Sihang Cai, Tao Jin, Qifei Zhang

Multimodal Sentiment Analysis (MSA) models typically suffer significant performance degradation under domain shifts. While Test-Time Adaptation (TTA) aims to mitigate this, existing discriminative approaches often succumb to “confident but wrong” predictions on out-of-distribution samples. Conversely, generative models offer robust calibration but incur prohibitive computational costs. To bridge this gap, we propose GD-Adapt (Generative-Discriminative Adaptation), a novel TTA framework that harmonizes the robustness of generative diffusion models with the efficiency of discriminative regression networks via Bayesian Diffusion Distillation (BDD). Specifically, we introduce Auxiliary Generative Regularization (AGR) during pretraining to enforce manifold-aware feature learning. Extensive experiments across five cross-domain scenarios demonstrate our method’s superiority. For instance, on the challenging MOSI to SIMS shift, GD-Adapt reduces Mean Absolute Error (MAE) from 0.6872 to 0.5673 and boosts binary accuracy by 5.81 percentage points (reaching 57.33%). Notably, in scenarios such as SIMS to MOSI, we achieve an 11.18-point gain over the non-adapted baseline.

Haiming Qin, Jianxun Lian, Qimin Zhong, Mingyang Zhou, Hao Liao, Naipeng Chao

Large Language Models (LLMs) are increasingly deployed in role-play scenarios, but their safety implications remain under-characterized. We present an explanatory framework grounded in Bandura’s Moral Disengagement theory and introduce a diagnostic benchmark (MD-Trace) for role-play jailbreaks. In our experiments, role-play improves safety behavior for benign personas while increasing unsafe compliance for malicious ones. We observe a Knowing-but-Doing failure in which models recognize safety risks in their thinking traces yet proceed to comply with harmful requests. Mechanism analysis suggests that Moral Justification is dominant, with Disregard of Consequences appearing as a secondary pattern. We compare multiple attack and defense methods and find that the diagnosis aligns with observed failure modes. Finally, we propose MD-Shield, an introspection-based defense that reduces attack success while maintaining Role Fidelity. The source code is publicly available at https://github.com/lavapapa/MoralJustify/.

Yutao Sun, Tianzhu Ye, Li Dong, Yuqing Xia, Jian Chen, Yizhao Gao, Shijie Cao, Jianyong Wang, Furu Wei

Efficient long-sequence generation is a critical challenge for Large Language Models. While recent sparse decoding methods improve efficiency, they suffer from KV cache misalignment, where approximation errors accumulate and degrade generation quality. In this work, we propose Rectified Sparse Attention (ReSA), a simple yet effective method that combines block-sparse attention with periodic dense rectification. By refreshing the KV cache at fixed intervals using a dense forward pass, ReSA bounds error accumulation and preserves alignment with the pretraining distribution. Experiments across math reasoning, language modeling, and retrieval tasks demonstrate that ReSA achieves near-lossless generation quality with significantly improved efficiency. Notably, ReSA delivers up to 3.77x end-to-end speedup under decoding at 256K sequence length, making it a practical solution for scalable long-context inference.

Zhensheng Wang, Xiaole Liu, Wenmian Yang, Kun Zhou, Yiquan Zhang, Weijia Jia

The rapid development of LLMs has significantly advanced tabular question answering, but most systems cannot perform future-oriented numerical prediction. To address this gap, we introduce a novel task, Open-Domain Tabular Question Answering for Future Data Forecasting and Reasoning, and propose the first dataset to cover time-series forecasting and forecast-based reasoning scenarios using real estate data. This task poses challenges in retrieving precise historical data, overcoming the forecasting limitations of LLMs, and standardizing responses for diverse queries. To solve the above challenges, we propose TimeFore, an LLM agent-based framework that decomposes the problem into three collaborative roles: a Retriever autonomously generates SQL to fetch data, a Forecaster invokes external time-series models for higher accuracy, and an Analyzer synthesizes the results to construct a precise and consistent final answer. Extensive experiments demonstrate the effectiveness of our TimeFore.

Qiming Xie, Wenjie Zheng, Xiangqing Shen, Rui Xia

The rise of social media and generative AI has led to a surge of misinformation online, making reliable fact-checking increasingly critical.Most existing fact-checking research adheres to the decompose-then-verify paradigm, emphasizing verification of individual facts while overlooking the validity of logical dependencies among them. As a result, text containing logical errors may still be misjudged as factual. Moreover, existing datasets and metrics focus on fact completeness and coverage, failing to capture the logical dimension.To help bridge this gap, we propose a content–logic coupled factuality evaluation paradigm, which conceptualizes factuality along two complementary dimensions: content factuality and logic factuality. Under this paradigm, we introduce a holistic solution consisting of LoReFact, the first long-form fact-checking dataset that systematically incorporates the logical dimension; LoRe-Factcheck, a simple yet effective framework for joint content–logic evaluation; and a logic-aware metric named LoReFactScore for exposing and penalizing logical fallacies.Experiments demonstrate the importance of logical factuality and the effectiveness of our proposed paradigm for fact-checking.[Our data and code are publicly available at https://github.com/NUSTM/LoReFact]

Chengyi Yang, Pengzhen Li, Jiayin Qi, Aimin Zhou, Ji Wu, Ji Liu

Text-to-Video (T2V) generation has benefited from recent advances in diffusion models, yet current systems still struggle under complex scenarios, which are generally exacerbated by the ambiguity and underspecification of text prompts. In this work, we formulate complex-scenario prompt refinement as a stage-wise multi-agent refinement process and propose SCMAPR, i.e., a scenario-aware and Self-Correcting Multi-Agent Prompt Refinement framework for T2V prompting. SCMAPR coordinates specialized agents to (i) route each prompt to a taxonomy-grounded scenario for strategy selection, (ii) synthesize scenario-aware rewriting policies and perform policy-conditioned refinement, and (iii) conduct structured semantic verification that triggers conditional revision when violations are detected. To clarify what constitutes complex scenarios in T2V prompting, provide representative examples, and enable rigorous evaluation under such challenging conditions, we further introduce T2V-Complexity, which is a complex-scenario T2V benchmark consisting exclusively of complex-scenario prompts. Extensive experiments on 3 existing benchmarks and our T2V-Complexity benchmark demonstrate that SCMAPR consistently improves text-video alignment and overall generation quality under complex scenarios, achieving up to 2.67% and 3.28 gains in average score on VBench and EvalCrafter, and up to 0.028 improvement on T2V-CompBench over 3 State-Of-The-Art baselines. The codes of SCMAPR are publicly available at https://github.com/HiThink-Research/SCMAPR.

Wenhao Hong, Ziyang Wang, Yixin Zhang, Zilei Wang

Video Large Language Models (VideoLLMs) struggle with the heavy computational cost of long or high-resolution videos due to massive visual token counts and the quadratic complexity of attention. Prior pruning approaches mainly rely on token importance or similarity, while largely overlooking video dynamics and the fact that different scenes exhibit different redundancy patterns. We introduce MoPrune, a training-free, scene-guided and motion-centric token pruning framework for accelerating VideoLLMs. MoPrune first segments videos into semantically coherent scenes to preserve temporal and motion consistency. Within each scene, it determines frame retention rates from intra-scene frame uniqueness. Finally, at the token level, MoPrune retains visually distinctive tokens and motion-salient tokens via a unified score, preserving both informative static details and dynamic regions. Extensive experiments across multiple VideoLLMs and public benchmarks demonstrate MoPrune’s superior efficiency–performance trade-offs. On LLaVA-OneVision, retaining 25% of visual tokens matches or slightly improves the dense baseline, and retaining 15% tokens preserves 99% of the original performance. MoPrune is fully compatible with hardware-efficient techniques such as Flash Attention.

Chibuzor Okocha, Christan Grant

Audio language models (ALMs) are increasingly used for speech-based understanding; yet, their ability to perform semantic reasoning beyond transcription, Text-to-Audio Retrieval, Captioning, and Question-Answering accuracy remains insufficiently benchmarked. In particular, the effects of accent variation, domain shift, and semantic over-inference on audio reasoning are poorly understood. We evaluate audio language models across five semantic and paralinguistic reasoning tasks: entailment, consistency, plausibility, accent drift, and accent restraint. Collectively, these tasks assess a model’s ability to reason over spoken audio as the primary evidence source, including whether a textual hypothesis can be inferred, contradicted, or left undetermined by the audio, whether statements align or conflict with spoken content, whether claims are plausible given the discourse, and whether model predictions remain stable or appropriately constrained across accent variation. These findings highlight critical limitations in current audio reasoning evaluations and hope to provide guidance for more robust and equitable ALM design and assessment.

Zhenhong Zhou, Zherui Li, Jie Zhang, Yuanhe Zhang, Kun Wang, Yang Liu, Qing Guo

Large Language Model-based Multi-Agent Systems represent a promising paradigm for tackling complex problems through agent collaboration. However, the reliance on open-ended communication exposes a fundamental vulnerability: the collaborative process itself can be exploited and disrupted. In this work, we formalize this threat class as Denial-of-Collaboration (DoC). Unlike DoS, which targets individual nodes or services, DoC attacks corrupt the collaborative structure of the system, transforming its communication topology into self-sabotage. The result is excessive resource consumption and eventual system paralysis. We introduce **CO**ntagious **R**ecursive **B**locking **A**ttacks (CORBA) as a concrete example of DoC, which employs benign yet recursively contagious instructions, forcing LLM-MASs into cycles of meaningless message passing. Critically, since our attacks are semantically benign, they easily bypass conventional safety alignments that are not designed to detect behavioral or systemic attacks. Through extensive experiments across diverse topologies and models, we demonstrate that CORBA achieves system paralysis where the baseline attacks fail. Our work reveals emerging DoC threats in current LLM-MAS security and establishes a crucial baseline for developing robust, collaboration-aware defense mechanisms.

Zhanli Li, Yixuan Cao, Lvzhou Luo, Ping Luo

This paper introduces the task of analytical question answering over large, semi-structured document collections. We present MuDABench, a benchmark for multi-document analytical QA, where questions require extracting and synthesizing information across numerous documents to perform quantitative analysis. Unlike existing multi-document QA benchmarks that typically require information from only a few documents with limited cross-document reasoning, MuDABench demands extensive inter-document analysis and aggregation. Constructed via distant supervision by leveraging document-level metadata and annotated financial databases, MuDABench comprises over 80,000 pages and 332 analytical QA instances. We also propose an evaluation protocol that measures final answer accuracy and uses intermediate-fact coverage as an auxiliary diagnostic signal for the reasoning process. Experiments reveal that standard RAG systems, which treat all documents as a flat retrieval pool, perform poorly. To address these limitations, we propose a multi-agent workflow that orchestrates planning, extraction, and code generation modules. While this approach substantially improves both process and outcome metrics, a significant gap remains compared to human expert performance. Our analysis identifies two primary bottlenecks: single-document information extraction accuracy and insufficient domain-specific knowledge in current systems. MuDABench is available at https://github.com/Zhanli-Li/MuDABench.

JunJian Wang, Lidan Zhao, Xi Sheryl Zhang

Large Language Models (LLMs) exhibit impressive reasoning capabilities but often suffer from Embodied Semantic Hallucinations—generating plans that are semantically fluent but physically unsafe due to a lack of grounded common sense. Existing safety alignment methods, such as RLHF or naive safety prompting, typically fall into a Safety-Utility Trade-off, resulting in severe over-rejection of benign household instructions. To address this, we propose MADRA (Multi-Agent Debate for Risk Awareness), a training-free cognitive architecture that mimics System-2 deliberation. MADRA introduces a meta-cognitive Critical Agent that evaluates peer debates using a structured argumentation framework derived from the Toulmin Model, effectively mitigating the "herd mentality" in multi-agent systems. We also introduce SafeAware-VH, a benchmark featuring adversarial safe instructions designed to probe agents’ sensitivity to physical risks. Extensive experiments demonstrate that MADRA breaks the Pareto frontier, achieving over 90% rejection of unsafe tasks while maintaining high utility, significantly outperforming standard Chain-of-Thought and single-agent reflection baselines.

Zhiyuan Li, Yuan Wu, Yi Chang

To stabilize the training of Large Language Models (LLMs), gradient clipping is a nearly ubiquitous heuristic used to alleviate exploding gradients. However, traditional global norm clipping erroneously presupposes gradient homogeneity across different functional modules, leading to an adverse "spill-over" effect where volatile parameters force unnecessary scaling on stable ones. To overcome this, we propose Adaptive Group-wise Gradient Clipping (AGGC). AGGC partitions parameters into groups based on functional types and regulates each according to its historical behavior using an Exponential Moving Average (EMA). Specifically, it constructs an adaptive interval to simultaneously mitigate gradient explosion and vanishing, while employing a time-dependent scheduling mechanism to balance exploration and convergence. Experiments on LLaMA 2-7B, Mistral-7B, and Gemma-7B models demonstrate that AGGC-enhanced LoRA consistently outperforms standard LoRA and frequently exceeds Full Fine-Tuning performance. Specifically, on the GSM8K benchmark, Mistral-7B fine-tuned with AGGC-enhanced LoRA achieves 72.93% accuracy, surpassing the 69.5% of vanilla LoRA. AGGC also contributes to the stability of Reinforcement Learning with Verifiable Rewards (RLVR), leading to improved logical deduction in Qwen 2.5 and Llama 3.2 models. Experimental results demonstrate that AGGC effectively addresses the limitations of traditional gradient clipping methods, particularly in overcoming gradient heterogeneity, by utilizing a modular, adaptive clipping strategy to stabilize the training process. Due to its lightweight design, AGGC can be seamlessly integrated into existing post-training pipelines with negligible overhead.

Ekaterina Fadeeva, Aleksandr Rubashevskii, Dzianis Piatrashyn, Roman Vashurin, Shehzaad Dhuliawala, Artem Shelmanov, Timothy Baldwin, Preslav Nakov, Mrinmaya Sachan, Maxim Panov

Large Language Models (LLMs) enhanced with knowledge retrieval, an approach known as Retrieval-Augmented Generation (RAG), have achieved strong performance in open-domain question answering. However, RAG remains prone to hallucinations: factually incorrect outputs may arise from inaccuracies in the model’s internal knowledge and the retrieved context. Existing approaches to mitigating hallucinations often conflate factuality with faithfulness to the retrieved evidence, incorrectly labeling factually correct statements as hallucinations if they are not explicitly supported by the retrieval. In this paper, we introduce FRANQ (Faithfulness-aware Retrieval-Augmented UNcertainty Quantification), a new method for hallucination detection in RAG outputs. FRANQ applies distinct uncertainty quantification techniques to estimate factuality, conditioning on whether a statement is faithful to the retrieved context. To evaluate FRANQ and competing uncertainty quantification methods, we construct a new long-form question answering dataset annotated for both factuality and faithfulness, combining automated labeling with manual validation of challenging cases. Extensive experiments across multiple datasets, tasks, and LLMs show that FRANQ achieves more accurate detection of factual errors in RAG-generated responses compared to existing uncertainty quantification and hallucination detection approaches.