论文检索

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.

Biao Yi, Jiahao Li, Yiming Li, Yu He, Baolei Zhang, Zheli Liu, Dacheng Tao

Membership inference attacks (MIAs) aim to determine whether specific data was used to train a model. While existing MIAs against pre-trained Large Language Models (LLMs) typically require access to complete logits (probabilities), such access is sometimes unavailable in real-world deployments where only the generated text is exposed. Current label-only MIAs relied on surrogate models to estimate the target model’s token probabilities, but we identify fundamental limitations: high sensitivity to surrogate model selection and significant probability estimation errors. To address these challenges, we propose SEAD (Semantic-Aware Density), a novel surrogate-free label-only MIA approach that directly estimates token probabilities through Monte Carlo sampling of the target model itself. This approach eliminates dependency on surrogate models while reducing probability estimation errors by an order of magnitude. Furthermore, we introduce a semantic-aware density approach that enhances attack effectiveness by considering both exact token matches and semantically similar alternatives, inspired by the understanding that LLMs may express memorized information through different but semantically equivalent tokens. Extensive evaluations demonstrate that SEAD consistently outperforms existing label-only attacks and serves as a foundational density estimator in the label-only setting.

Junxian Li, Xinyue Xu, Sai Ma, Di Zhang, Sichao Li

Multimodal Large Language Models (MLLMs) frequently suffer from unfaithfulness, generating reasoning chains that drift from visual evidence or contradict final predictions. We propose Faithful-First Reasoning, Planning, and Acting (RPA) framework in which FaithEvi provides step-wise and chain-level supervision by evaluating the faithfulness of intermediate reasoning, and FaithAct uses these signals to plan and execute faithfulness-aware actions during inference. Experiments across multiple multimodal reasoning benchmarks show that faithful-first RPA improves perceptual faithfulness by up to 24% over prompt-based and tool-augmented reasoning frameworks, without degrading task accuracy. Our analysis shows that treating faithfulness as a guiding principle perceptually faithful reasoning trajectories and mitigates hallucination behavior. This work thereby establishes a unified framework for both evaluating and enforcing faithfulness in multimodal reasoning. Code is at https://github.com/lijunxian111/Faithful-First-RPA.

Mengxiang Zhang, Lingyuan Liu

Knowledge distillation (KD) is a key technique for compressing large language models (LLMs), yet it faces challenges stemming from the teacher–student capacity gap. While existing KD methods address these challenges either by mixing teacher and student distributions in the distillation target or by using curriculum learning to sequence training from easy to hard examples, they typically design these two strategies independently, missing the opportunity for synergistic co-design. To bridge this gap, we propose Calibrated Progressive Distillation (CPD), a white-box KD framework that co-designs curriculum scheduling and target mixing through a unified difficulty-aware principle. CPD uses a difficulty profile to select epoch-specific subsets that ensure a uniform increase in average difficulty, adapting to the dataset’s intrinsic hardness structure. Simultaneously, the mixing coefficient in the distillation target and the distillation temperature are synchronized with this progression, gradually shifting supervision from teacher-dominated to student-informed signals as training advances. Theoretically, CPD ensures bounded gradients and induces an implicit attention shift from easy to hard samples. Empirically, CPD consistently outperforms advanced KD methods across diverse tasks, while reducing training runtime by over 10%. Our work demonstrates that aligning data scheduling with distillation signal design is crucial for effective and efficient LLM distillation.

Chao Wen, Jacqueline Staub, Adish Singla

Vision-language models (VLMs) have been explored for visual programming, where they generate code to solve visual tasks. However, most prior work focuses on visual programming for productivity; it remains unclear how well current VLMs perform on education-oriented visual programming and what factors limit their performance. To bridge this gap, we introduce TURTLEAI, a benchmark containing 823 tasks curated based on real-world visual programming tasks in the Turtle Graphics domain. Solving these tasks requires models to perceive geometric patterns, reason about spatial relationships, and synthesize Python code that faithfully reproduces geometric patterns. We evaluate 20+ VLMs, including GPT-5, GPT-4o, and Qwen2-VL-72B, and find that they struggle significantly, with most achieving success rates below 30%. To address these limitations, we propose a data generation technique that requires only a small set of seed samples. Fine-tuning Qwen2-VL-72B on the resulting synthetic data yields an improvement of about 20% on real-world tasks. Our failure analysis reveals that GPT-4o struggles with spatial reasoning and precise visual replication, whereas fine-tuning primarily improves the alignment between visual reasoning and code implementation.

Yi Zhao, Yajuan Peng, Cam-Tu Nguyen, Zuchao Li, Xiaoliang Wang, Xiaoming Fu, Hai Zhao

Large Reasoning Models (LRMs) achieve strong performance on complex tasks through extended chains of thought but suffer from high inference latency due to autoregressive reasoning. Recent work explores using Small Reasoning Models (SRMs) to accelerate LRM inference, yet existing frameworks such as SpecReason adopt a polling-based design that repeatedly invokes the LRM for verification at every step. This approach is inefficient, as frequent LRM calls introduce a high computational overhead, and is unreliable, since the LRM as a judge is prone to errors. In this paper, we systematically characterize the capability boundaries of SRMs and identify three common types of reasoning risks: (1) path divergence, where SRMs lack the strategic ability to construct an initial plan, causing reasoning to deviate from the most probable path; (2) cognitive overload, where SRMs fail to solve particularly difficult steps; and (3) recovery inability, where SRMs lack robust self-reflection and error correction mechanisms. To address these challenges, we propose TrigReason, a trigger-based collaborative reasoning framework that replaces continuous polling with selective intervention. TrigReason delegates most reasoning to the SRM and activates LRM intervention only when necessary—during initial strategic planning (strategic priming trigger), upon detecting extraordinary overconfidence (cognitive offload trigger), or when reasoning falls into unproductive loops (intervention request trigger). The evaluation results on AIME24, AIME25, and GPQA-D indicate that TrigReason matches the accuracy of full LRMs and SpecReason, while offloading 1.70×–4.79× more reasoning steps to SRMs. Under edge–cloud conditions, TrigReason reduces latency by 43.9% and API cost by 73.3% compared to SpecReason.

Qiong Wu, Tan Yue, Jianxin Liang, Zhen Li, Kai He, Shuai Zhao, Dongyan Zhao

The rapid progress of large language models (LLMs) has increased the demand for efficient and reliable evaluation of question answering (QA) systems. Existing evaluation methods either rely on rule-based matching with shallow semantic understanding or adopt LLM-as-a-Judge approaches that incur high cost and latency while offering limited error interpretability. Accordingly, we propose HiEval, a curriculum learning based hierarchical framework for QA task evaluation that supports both quick scoring and fine-grained error analysis. HiEval contains a quick scoring model (HiEval-QS) that predicts three-level correctness labels, and an error analysis model (HiEval-EA) that identifies incorrect responses into five error types. HiEval incorporates a class-balanced focal loss to handle label imbalance, experience replay to prevent forgetting, and contrastive unlikelihood optimization to improve error discrimination. We also construct two large-scale human-annotated evaluation datasets collected from 50 QA-related datasets, covering 8 task types and release two challenging benchmarks. Extensive experiments show that HiEval achieves state-of-the-art performance on both quick scoring and error analysis tasks, outperforming all baseline methods, including GPT-5, while being approximately 25\times faster.

Cehao Yang, Xueyuan Lin, Xiaojun Wu, Chengjin Xu, Xuhui Jiang, Honghao Liu, Hui Xiong, Jian Guo

A practical approach to activate long chain-of-thoughts reasoning ability in large language models is to perform supervised fine-tuning on instruction datasets synthesized by strong large reasoning models, offering a cost-effective alternative to reinforcement learning. However, large-scale instruction sets incur significant training overhead, while effective strategies for automatic data selection still remain unexplored. We propose Select2Reason, a novel and efficient instruction-tuning data selection framework for long-CoT reasoning. From the perspective of emergence of rethinking behaviors like self-correction and backtracking, we investigate metrics that may determine the quality of long-CoT instructions. Select2Reason leverages a difficulty-aware reward model to estimate the learning value of questions and jointly incorporates a reasoning trace length-based heuristic through a weighted scheme for ranking to prioritize high-utility examples. Empirical results on OpenR1-Math-220k demonstrate that fine-tuning LLM on only 10% of the data selected by our method achieves performance competitive with or superior to full-data tuning and open-source baseline across nine competition-level mathematical benchmarks and four broader reasoning tasks. Further experiments highlight the scalability in varying data size, efficiency during inference, and adaptability to other instruction pools of Select2Reason with minimal cost.

Hellina Hailu Nigatu, Bethelhem Yemane Mamo, Bontu Fufa Balcha, Debora Taye Tesfaye, Elbethel Daniel Zewdie, Ikram Behiru Nesiru, Jitu Ewnetu Hailu, Senait Mengesha Yayo

As low-resourced languages are increasingly incorporated into NLP research, there is an emphasis on collecting large-scale datasets. But in prioritizing quantity over quality, we risk 1) building language technologies that perform poorly for these languages and 2) producing harmful content that perpetuates societal biases. In this paper, we investigate the quality of Machine Translation (MT) datasets for three low-resourced languages–Afan Oromo, Amharic, and Tigrinya, with a focus on the gender representation in the datasets. Our findings demonstrate that while training data has a large representation of political and religious domain text, benchmark datasets are focused on news, health, and sports. We also found a large skew towards the male gender–in names of persons, the grammatical gender of verbs, and in stereotypical depictions in the datasets. Further, we found harmful and toxic depictions against women, which were more prominent for the language with the largest amount of data, underscoring that quantity does not guarantee quality. We hope that our work inspires further inquiry into the datasets collected for low-resourced languages and prompts early mitigation of harmful content. WARNING: This paper contains discussion of NSFW content that some may find disturbing.

Adrian de Wynter, Tangming Yuan

Large language models (LLMs) are excellent at maintaining high-level, convincing dialogue, but it remains unclear whether their persuasive success reflects genuine understanding of the discourse. We examine this question through informal debates between humans and LLMs, first by measuring their persuasive skills, and then by relating these to their understanding of what is being talked about: namely, their comprehension of argumentative structures and the pragmatic context on the same debates. We find that LLMs effectively maintain coherent, persuasive debates, and can sway the beliefs of both participants and audiences. We also note that awareness or suspicion of AI involvement encourage people to be more critical of the arguments made. However, we also find that LLMs are unable to show comprehension of deeper dialogical structures, such as argument quality or existence of supporting premises. Our results reveal a disconnect between LLM comprehension and dialogical skills, raising ethical and practical concerns on their deployment on explanation-critical contexts. From an argumentation-theoretical perspective, we experimentally question whether an agent, if it can convincingly maintain a dialogue, is required to show it knows what is talking about.

Donghao Li, Yifan Deng, Jinta Weng, Xingsheng Zhang, Chenxu Niu, Jingyuan Tian, Heyan Huang, Yue Hu

Asynchronous psychological counseling (APC) represents a crucial mental health service modality that transcends temporal and spatial constraints. However, its development faces significant data scarcity challenges: due to stringent privacy protection requirements and professional ethical considerations, direct collection of conversational data from authentic APC scenarios is virtually impossible. To address this challenge, we design a self-optimizing multi-agent framework for counseling dialogue generation, CFlowPsy, which utilizes a small amount of real anonymized counseling cases as seed data to synthesize diverse problem-solving-oriented APC conversations through large language models. Specifically, the framework employs a Persona-Flow module to continuously track and update clients’ basic information, real-time emotions, and counseling progress, providing dynamic personalized analytical support for counselors and enabling self-optimization of generated dialogues. Simultaneously, the framework ensures that generated interventions contain explicit reasoning processes, demonstrating clear psychological analysis and logic, thereby enhancing the accuracy and consistency of responses. Under this framework, we develop the first Chinese APC dataset, CFlowPsyD, comprising 1,700 high-quality extended conversations. Extensive experiments and human evaluations confirm that the proposed CFlowPsyD dataset successfully simulates human-like APC processes.

Yong Wu, Weihang Pan, Ke Li, Chen Binhui, Ping Li, Binbin Lin

Large language models (LLMs) have shown remarkable reasoning capabilities, yet aligning such abilities to small language models (SLMs) remains a challenge due to distributional mismatches and limited model capacity. Existing reasoning datasets, typically designed for powerful LLMs, often lead to degraded performance when directly applied to weaker models. In this work, we introduce Dynamic Adaptation of Reasoning Trajectories (DART), a novel data adaptation framework that bridges the capability gap between expert reasoning trajectories and diverse SLMs. Instead of uniformly imitating expert steps, DART employs a selective imitation strategy guided by step-wise adaptability estimation via solution simulation. When expert steps surpass the student’s capacity—signaled by an Imitation Gap—the student autonomously explores alternative reasoning paths, constrained by outcome consistency. We validate DART across multiple reasoning benchmarks and model scales, demonstrating that it significantly improves generalization and data efficiency over static fine-tuning. Our method enhances supervision quality by aligning training signals with the student’s reasoning capabilities, offering a scalable solution for reasoning alignment in resource-constrained models.

Xuan Xu, Zhongliang Yang, Haolun Li, Rui Tian, Beilin Chu, J Song, Yu Li, Shaolin Tan, Linna Zhou

LLMs have become foundational across many NLP applications, driving a shift from an algorithm-centric to a context-centric paradigm. As an important task in text mining, the landscape of topic modeling (TM) is similarly being reshaped by a growing body of LLM-driven research.We review recent TM developments and categorize existing methods into three groups: Classical Algorithm-Centric, LLM-Assisted, and LLM-Centric. For traditional algorithm-centric methods, we refine prior taxonomies and highlight recent advances. For the LLM-Assisted and LLM-Centric settings, we introduce a new taxonomy that emphasizes the role of LLMs and the design of end-to-end workflows, respectively. We examine two key transformations brought by LLM-centric TM: expanded task scope and a shift from model-level improvements to system-level engineering. We also propose a future roadmap for more optimized LLM-Centric TMs and identify ongoing critical challenges. We aim for this survey to spur closer integration between TM and LLMs and to further drive the progress of modern TM.

Huacan Chai, Zijie Cao, Maolin Ran, Yingxuan Yang, Jianghao Lin, Xin Peng, Hairui Wang, Renjie Ding, Ziyu Wan, Muning Wen 等

Large language models (LLMs) have achieved impressive success in single-turn function calling, yet real-world applications such as travel planning or multi-stage data analysis typically unfold across multi-turn conversations. In these settings, LLMs must not only issue accurate function calls at each step but also maintain progress awareness, the ability to summarize past interactions and plan future actions to ensure coherent, long-horizon task execution. Existing approaches, however, either reduce multi-turn training to isolated single-turn samples, which neglects task-level planning, or employ end-to-end reinforcement learning (RL) that struggles with redundancy and lacks explicit integration of progress awareness. To overcome these limitations, we introduce Progra, a framework that explicitly incorporates progress awareness into LLM training for multi-turn function calling. Progra combines (i) a Progress Awareness Generation (PAG) pipeline, which automatically constructs datasets coupling conversation summaries with future task planning, and (ii) a Progress Awareness-Guided Reinforcement Learning (PAG-RL) algorithm, which integrates progress awareness into RL training to reduce contextual redundancy and improve alignment between local actions and global task completion. Empirical results on two public benchmarks demonstrate that Progra significantly outperforms existing methods, highlighting the effectiveness of progress awareness in enabling robust and efficient multi-turn function calling. Our code is available at https://github.com/FatCatCHC/Progra .

Sungmok Jung, Yeonkyoung So, Joonhak Lee, Sangho Kim, Yelim Ahn, Jaejin Lee

Although negation is known to challenge large language models (LLMs), benchmarks for evaluating negation understanding—especially in Korean—are scarce. We conduct a corpus-based analysis of Korean negation and show that LLM performance degrades under negation. We then introduce *Thunder-KoNUBench*, a sentence-level negation understanding benchmark that reflects the empirical distribution of Korean negation phenomena. Evaluating 47 LLMs on Thunder-KoNUBench, we analyze the effects of model size and instruction tuning, and perform error analysis to better understand model behavior. We further show that fine-tuning on Thunder-KoNUBench improves negation understanding and broader contextual comprehension in Korean.