论文检索

Jianghao Wu, Feilong Tang, Yulong Li, Ming Hu, Haochen Xue, Shoaib Jameel, Zongyuan Ge, Yutong Xie, Imran Razzak

Recent advances such as Chain-of-Thought prompting have significantly improved large language models (LLMs) in zero-shot medical reasoning. However, prompting-based methods often remain shallow and unstable, while fine-tuned medical LLMs suffer from poor generalization under distribution shifts and limited adaptability to unseen clinical scenarios. To address these limitations, we present TAGS, a test-time framework that combines a broadly capable generalist with a domain-specific specialist to offer complementary perspectives without any model fine-tuning or parameter updates. To support this generalist–specialist reasoning process, we introduce two auxiliary modules: a hierarchical retrieval mechanism that provides multi-scale exemplars by selecting examples based on both semantic and rationale-level similarity, and a reliability scorer that evaluates reasoning consistency to guide final answer aggregation. TAGS achieves strong performance across nine MedQA benchmarks, boosting GPT-4o accuracy by 13.8%, DeepSeek-R1 by 16.8%, and improving a vanilla 7B model from 14.1% to 23.9%. These results surpass several fine-tuned medical LLMs, without any parameter updates.

Zhitong Chen, Kai Yin, Xiangjue Dong, Chengkai Liu, Xiangpeng Li, Bo Li, Junwei Ma, Yiming Xiao, Ali Mostafavi, James Caverlee

Accurate question answering (QA) in disaster management requires reasoning over uncertain and conflicting information, a setting poorly captured by existing benchmarks built on clean evidence. We introduce DisastQA, a rigorously verified benchmark of 3,000 expert-annotated questions (2,000 multiple-choice and 1,000 open-ended) spanning eight disaster types. The benchmark is constructed via a human-LLM collaboration pipeline with stratified sampling to ensure balanced coverage. Models are evaluated under varying evidence conditions, from closed-book to noisy evidence integration, enabling separation of internal knowledge from reasoning under imperfect information. For open-ended QA, we propose a human-verified keypoint-based evaluation protocol emphasizing factual completeness over verbosity. Experiments with 20 models reveal substantial divergences from general-purpose leaderboards such as MMLU-Pro. While recent open-weight models approach proprietary systems in clean settings, performance degrades sharply under realistic noise, exposing critical reliability gaps for disaster response. All code, data, and evaluation resources are available at https://github.com/TamuChen18/DisastQA

Tao Liu, Taiqiang Wu, Runming Yang, Shaoning Sun, Junjie Wang, Yujiu Yang

Supervised fine-tuning (SFT) is a fundamental post-training strategy to align Large Language Models (LLMs) with human intent. However, traditional SFT often ignores the \textit{one-to-many} nature of language by forcing alignment with a single reference answer, leading to the model overfitting to non-core expressions. Although our empirical analysis suggests that introducing multiple reference answers can mitigate this issue, the prohibitive data and computational costs necessitate a strategic shift: prioritizing the mitigation of single-reference overfitting over the costly pursuit of answer diversity. To achieve this, we reveal the intrinsic connection between token probability and semantic importance: high-probability tokens carry the core logical framework, while low-probability tokens are mostly replaceable expressions. Based on this insight, we propose \textbf{ProFit}, which selectively masks low-probability tokens to prevent surface-level overfitting. Extensive experiments confirm that ProFit consistently outperforms traditional SFT baselines on general reasoning and mathematical benchmarks

Jinshuo Zhang, Xiaoding Zhou, Weiyu Zhang, Guoqiang Chen, Ying Lian, Xiaoyang Meng, Yonghe Chen, Hongjiao Guan, Jiasheng Si, Wenpeng Lu

Retrieval-Augmented Generation (RAG) mitigates hallucinations in large language models by incorporating external knowledge. However, retrieval does not always return relevant documents and may return noisy ones. Indiscriminately retrieving and utilizing this external knowledge can interfere with the model’s originally correct reasoning. In this work, we propose Dual-Decision Retrieval-Augmented Generation (D^2-RAG), which integrates multi-dimensional uncertainty estimation to decide whether to retrieve and employs adaptive contrastive decoding to handle retrieved contexts of varying quality. Specifically, we first integrate uncertainty estimation scores that assess model uncertainty from multiple perspectives, construct them into a comprehensive feature vector, and train a lightweight retrieval decision model to accurately identify the model’s knowledge boundaries and determine whether to retrieve. Subsequently, we dynamically adjust the contrastive decoding strategy based on the utility of retrieved contexts to enhance the utilization of relevant contexts while suppressing interference from noisy contexts. Extensive experiments on four medical question-answering datasets demonstrate that D^2-RAG significantly outperforms baselines, enabling retrieval-augmented Llama3.1-8B to surpass non-retrieval-augmented Llama3.1-70B on the MedMCQA dataset. The source code is available on https://github.com/zakelawen/d–rag.

Namyeong So, Seokgyu Jang, Taeuk Kim

Adaptive multi-agent systems (MAS) are increasingly adopted as solutions to complex problems. However, their optimization for narrow task ranges leaves it unclear whether they can function as general-purpose systems. To fill this gap, we conduct an extensive empirical study on adaptive MAS, revealing two key findings: (1) they are prone to topological overfitting, defined as failures in domain transfer; and (2) they exhibit illusory coordination, where surface-level accuracy is high but underlying agent coordination deviates from ideal MAS behavior, raising concerns about their practical effectiveness. These observations highlight the urgent need to prioritize generalization in MAS development and motivate more thorough evaluation beyond correctness of the final answer.

Peng Wang, Xuesi Hu, Jiageng Wu, Yuntao Zou, Qiancheng Zhang, Dagang Li

Recently, large language models (LLMs) and reasoning large language models (RLLMs) have gained considerable attention from many researchers. RLLMs enhance the reasoning capabilities of LLMs through Long Chain-of-Thought (Long CoT) processes, significantly improving the performance of LLMs in addressing complex problems. However, there are few works that systematically explore what methods can fully unlock the performance of LLMs and RLLMs within the financial domain. To investigate the impact of various methods on LLMs and RLLMs, we utilize five LLMs and four RLLMs to assess the effects of prompting methods, agentic frameworks, and multilingual alignment methods on financial question-answering tasks. Our research findings indicate: (1) Current prompting methods and agent frameworks enhance the performance of LLMs in financial question answering by simulating Long CoT; (2) RLLMs possess inherent Long CoT capabilities, which limits the effectiveness of conventional methods in further enhancing their performance; (3) Current advanced multilingual alignment methods primarily improve the multilingual performance of LLMs by extending the reasoning length, which yields minimal benefits for RLLMs. Additionally, we discuss strategies for enhancing the performance of LLMs and RLLMs in financial question answering, which may serve as a inspiration for future improvements. We hope that this study can serve as an important reference for LLMs and RLLMs in the field of financial question answering.

Weiyu Sun, Liangliang Chen, Yongnuo Cai, Huiru Xie, Yi Zeng, Ying Zhang

Multimodal Large Language Models (MLLMs) hold significant promise for revolutionizing traditional education and reducing teachers’ workload. However, accurately interpreting unconstrained STEM student handwritten solutions with intertwined mathematical formulas, diagrams, and textual reasoning poses a significant challenge due to the lack of authentic and domain-specific benchmarks. Additionally, current evaluation paradigms predominantly rely on the outcomes of downstream tasks (e.g., auto-grading), which often probe only a subset of the recognized content, thereby failing to capture the MLLMs’ understanding of complex handwritten logic as a whole. To bridge this gap, we release EDU-CIRCUIT-HW, a dataset consisting of 1,300+ authentic student handwritten solutions from a university-level STEM course. Utilizing the expert-verified verbatim transcriptions and grading reports of student solutions, we simultaneously evaluate various MLLMs’ upstream recognition fidelity and downstream auto-grading performance. Our evaluation uncovers an astonishing scale of latent failures within MLLM-recognized student handwritten content, highlighting the models’ insufficient reliability for auto-grading and other understanding-oriented applications in high-stakes educational settings. In response, we present a case study demonstrating that leveraging identified error patterns to preemptively detect and rectify recognition errors, with only minimal human intervention (e.g., with 3.3% assignments routed to human graders while the rest to GPT-5.1 grader), can effectively enhance the robustness of the deployed AI-enabled grading system on unseen student solutions.

Jinchao Li, Yunxin li, Chenrui Zhao, Zhenran Xu, Baotian Hu, Min Zhang

While GUI agents have shown impressive capabilities in common computer-use tasks such as OSWorld, current benchmarks mainly focus on isolated and single-application tasks. This overlooks a critical real-world requirement of coordinating across multiple applications to accomplish complex profession-specific workflows. To bridge this gap, we present a computer-use benchmark in cross-application workflows, named WindowsWorld, designed to systematically assess GUI Agents on complex multi-step tasks that mirror real-world professional activities. Our methodology uses a multi-agent framework steered by 16 occupations to generate four difficulty-level tasks with intermediate inspection, which are then refined by human review and executed in a simulated environment. The resulting benchmark contains 181 tasks with an average of 5.0 sub-goals across 17 common desktop applications, of which 78% are inherently multi-application. Experimental results of leading large models and agents show that: 1) All computer-use agents perform poorly on multi-application tasks (< 21% success rate), far below the performance of simple single-app tasks; 2) They largely fail at tasks requiring conditional judgment and reasoning across \geq3 applications, stalling at early sub-goals; 3) Low execution efficiency, where tasks often fail despite far exceeding human step limits. Code, benchmark data, and evaluation resources are available at github.com/HITsz-TMG/WindowsWorld.

Chengshuai Zhao, Zhen Tan, Pingchuan Ma, Dawei Li, Bohan Jiang, Yancheng Wang, Yingzhen Yang, Huan Liu

Chain-of-Thought (CoT) prompting has been shown to be effective in eliciting structured reasoning (i.e., CoT reasoning) from large language models (LLMs). Regardless of its popularity, recent studies expose its failures in some reasoning tasks, raising fundamental questions about the nature of CoT reasoning. In this work, we propose a data distribution lens to understand when and why CoT reasoning succeeds or fails. We hypothesize that CoT reasoning reflects a structured inductive bias learned from in-distribution data, enabling models to conditionally generate reasoning trajectories that approximate those observed during training. As such, the effectiveness of CoT reasoning is fundamentally governed by the nature and degree of distribution discrepancy between training data and test queries. Guided by this lens, we dissect CoT reasoning via three dimensions: task, length, and format. To test the hypothesis, we introduce DataAlchemy, an abstract and fully controllable environment that trains LLMs from scratch and systematically probes them under various distribution conditions. Through rigorous controlled experiments, we reveal that CoT reasoning is a brittle mirage when it is pushed beyond training distributions, emphasizing the ongoing challenge of achieving genuine and generalizable reasoning.

Guangtao Lyu, Qi Liu, Chenghao Xu, Jiexi Yan, Muli Yang, Xueting Li, Fen Fang, Cheng Deng

LVLMs have achieved strong multimodal reasoning capabilities but remain prone to hallucinations, producing outputs inconsistent with visual inputs or user instructions. Existing training-free methods, including contrastive decoding and auxiliary expert models, which incur several times more computational overhead and may introduce potential interference, as well as static internal signal enhancement, are often vulnerable to the attention sink phenomenon. We find that internal Positive Attention Dynamics (PAD) in LVLMs naturally reveal semantically core visual regions under the distortions of attention sinks. Based on this, we propose Positive Attention Dynamics Enhancement (PADE), a training-free attention intervention that constructs a PAD map to identify semantically core visual regions, applies per-head Median Absolute Deviation Scaling to adaptively control the intervention strength, and leverages System-Token Compensation to maintain attention to complex user instructions and support long-term output consistency. Experiments on multiple LVLMs and benchmarks show that PADE improves visual grounding and reduces hallucinations, validating the effectiveness of leveraging internal attention dynamics for reliable multimodal reasoning.

Rui Song, Yingji Li, Jian Li, Fausto Giunchiglia, Hao Xu

With the rapid development of Large Language Models (LLMs), In-Context Learning (ICL) has emerged as one of the universal paradigms for unleashing the capabilities of LLMs. However, LLMs are generally plagued by various biases in context example selection, which can distort the model’s predictions. Although extensive research has focused on designing heuristic sample selection methods to mitigate biases in ICL, these approaches often struggle to adapt to highly biased out-of-distribution (OOD) scenarios with significant shifts between test samples and context samples. To overcome the aforementioned issue, this paper proposes a LLM-driven iterative derivation method for OOD data pseudo-labeling (named LPL), aiming to mitigate the risk of performance degradation caused by OOD bias by avoiding direct use of source data. To mitigate the misleading effects of noise in pseudo-labels, we propose a filtering metric that integrates model confidence and perturbation perplexity to enhance the quality of pseudo-labels. Subsequently, in each iteration, LPL utilizes this metric to expand new pseudo-labeled data as contextual demonstrations and ultimately adopts a voting mechanism to ensure the stability of the predictions. A series of experiments conducted on various LLMs have confirmed that our proposed method can effectively reduce OOD biases, thereby opening up new avenues for research in ICL biases.

Caishuang Huang, Yang Qiao, Rongyu Zhang, Junjie Ye, Pu Lu, Wuwenxi, Meng Zhou, Xiku Du, Qi Zhang, Tao Gui 等

Tool-use capabilities are vital for Large Language Models (LLMs) in finance, a domain characterized by massive investment targets and data-intensive inquiries. However, existing data synthesis methods typically rely on a reverse synthesis paradigm, generating user queries from pre-sampled tools. This approach inevitably introduces artificial explicitness, yielding queries that fail to capture the implicit, event-driven nature of real-world needs. Moreover, its reliance on static tool sets overlooks the dynamic retrieval process required to navigate massive tool spaces. To address these challenges, we introduce \textit{FinToolSyn}, a forward synthesis framework designed to generate high-quality financial dialogues. Progressing from persona instruction and atomic tool synthesis to dynamic retrieval dialogue generation, our pipeline constructs a repository of 43,066 tools and synthesizes over 148k dialogue instances, incorporating dynamic retrieval to emulate the noisy candidate sets typical of massive tool spaces. We also establish a dedicated benchmark to evaluate tool-calling capabilities in realistic financial scenarios. Extensive experiments demonstrate that models trained on FinToolSyn achieve a 21.06% improvement, providing a robust foundation for tool learning in financial scenarios.

Yubo Jiang, Xin Yang, Abudukelimu Wuerkaixi, Zheming Yuan, Xuxin Cheng, Cao Liu, Ke Zeng, Fengying Xie, Zhiguo Jiang, Haopeng Zhang

Large vision–language models (LVLMs) excel at multimodal reasoning but still suffer from object-existence hallucinations when multi-step deliberation decouples from visual evidence. Think-with-Images (TwI) attempts to counter this by generating auxiliary observations (e.g., zoomed crops or highlighted views), yet it is not reliably beneficial. We identify two coupled failure modes: (1) a granularity–context trade-off of common operators, where zoom-in improves local detail but breaks global relations, while highlighting preserves topology but lacks fine evidence; and (2) an over-trust issue in tool-guided region proposals, where mislocalized evidence can dominate reasoning and even underperform standard prompting. We propose Active-Look, a training-free, plug-and-play TwI framework that allocates visual computation by uncertainty. Active-Look runs two heterogeneous grounding experts in parallel and uses their disagreement as a proxy for uncertainty, spending the budget only to verify disputed regions. It further mitigates the operator trade-off with conflict-aware hybrid rendering: highlighting retains global context, while selective zoom-in performs local verification. Experiments on hallucination-focused and general benchmarks (POPE, MME, and CHAIR) across multiple LVLM backbones show consistent gains.

Junzhe Zhang, Huixuan Zhang, Xinyu Hu, Li Lin, Mingqi Gao, Shi Qiu, Xiaojun Wan

Evaluation is important for multimodal generation tasks, while traditional multimodal evaluation metrics suffer from several limitations. With the rapid progress of MLLMs, there is growing interest in applying MLLMs to build general evaluation systems. However, existing researches often simply collect large-scale evaluation data for training, while overlooking the quality of evaluation data. What’s more, current proposed evaluation models often struggle to achieve consistently strong performance across both image-to-text (I2T) and text-to-image (T2I) tasks. In this paper, through rigorous quality control strategies, we construct a comprehensive multimodal evaluation dataset, Minos-57K, with evaluation samples across 15 datasets, for developing the multimodal evaluation model Minos with SFT and preference alignment training strategies. Notably, despite using less than half the scale of the training data of prior work, our model achieves state-of-the-art evaluation performance across 16 out-of-domain datasets covering both I2T and T2I tasks among all open-source multimodal evaluation models and remain competitive with closed-source models. Extensive experiments demonstrate the importance of leveraging quality control process, jointly training on evaluation data from both I2T and T2I generation tasks and further preference alignment.

Tian Lan, Felix Henry, Bin Zhu, Qianghuai Jia, Junyang Ren, Qihang PU, Haijun Li, Longyue Wang, Zhao Xu, Weihua Luo

Current Information Seeking (InfoSeeking) agents struggle to maintain focus and coherence during long-horizon exploration, as tracking search states, including planning procedure and massive search results, within one plain-text context is inherently fragile.To address this, we introduce Table-as-Search (TaS), a structured planning framework that reformulates the InfoSeeking task as a Table Completion task.TaS maps each query into a structured table schema maintained in an external database, where rows represent search candidates and columns denote constraints or required information.This table precisely manages the search states: filled cells strictly record the history and search results, while empty cells serve as an explicit search plan.Crucially, TaS unifies three distinct InfoSeeking tasks: Deep Search, Wide Search, and the challenging DeepWide Search.Extensive experiments demonstrate that TaS significantly outperforms numerous state-of-the-art baselines across three kinds of benchmarks, including multi-agent framework and commercial systems.Furthermore, our analysis validates the TaS’s superior robustness in long-horizon InfoSeeking, alongside its efficiency, scalability and flexibility.Code and datasets are publicly released at https://github.com/AIDC-AI/Marco-Search-Agent.

Bajian Xiang, Tingwei Guo, Xuan Chen, Yang Han

Large Speech Language Models (LSLMs) typically operate at high token rates (tokens/s) to ensure acoustic fidelity, yet this results in sequence lengths that far exceed the underlying semantic content, incurring prohibitive inference costs. In this paper, we empirically revisit the necessity of such granular token-level processing. Through layer-wise oracle interventions, we unveil a structured redundancy hierarchy: while shallow layers encode essential acoustic details, deep layers exhibit extreme redundancy, allowing for aggressive compression. Motivated by these findings, we introduce Affinity Pooling, a training-free, similarity-based token merging mechanism. By strategically applying this method at both input and deep layers, we effectively compress speech representations without compromising semantic information. Extensive evaluations across three tasks demonstrate that our approach reduces prefilling FLOPs by 27.48% while maintaining competitive accuracy. Practical deployment further confirms significant efficiency gains, yielding up to \sim1.7\times memory savings and \sim1.1\times faster time-to-first-token on long utterances. Our results challenge the necessity of fully distinct token representations, providing new perspectives on LSLM efficiency.

Chi Liu, Yan Shu, Mengzhuo Chen, Hongming Piao, Zhijian Duan, Derek Li, Bryan Dai

While large language models show promise in medical applications, achieving expert-level clinical reasoning efficiently remains challenging due to the need for massive amounts of manually labeled data and large-scale models. To address this challenge, we propose Clinical-Oriented Reinforcement Learning (CORL), the first fully open-source, end-to-end reinforcement learning training pipeline in the clinical reasoning domain, incorporating a Reasoning-Oriented Data Strategy (RODS) based on topological synthesis, CoT cold-start, and two-stage reinforcement learning. Through CORL, we trained the Fleming-R1 series of models. Among them, Fleming-R1-7B significantly outperforms models of comparable size while approaching or even surpassing certain 32B and 72B models. Fleming-R1-32B achieves near-parity with GPT-4o and outperforms the strongest open-source alternatives up to 671B in MedXpertQA. This demonstrates that in clinical reasoning field, a meticulously designed training pipeline holds greater importance than scaling model size alone. Data and Models are available at https://github.com/UbiquantAI/Fleming-R1 and https://huggingface.co/collections/IQuestLab/fleming.

Yue Zhao, Hongyan Li, Yong Chen, Luo Ji

Emotional interaction is increasingly crucial for conversational AI, yet current systems lack a self-emotion determination mechanism to drive the streaming text-to-speech (TTS) synthesis. We propose an emotion-planning framework that determines the emotion prior to the textual generation, grounding the downstream emotional TTS in a streaming manner. The framework is implemented by a plug-and-play LLM module, initialized from pretrained LLMs, and trained by reinforcement learning (RL) with emotions as the actions. A hybrid reward is employed which combines imitation signals with theory-driven scoring, in which the theory of Plutchik’s wheel of emotions is adopted. By experiments on DailyDialog, EmoryNLP, IMEOCAP, and MELD, our method outperforms prompting and finetuning baselines on both emotion determination and response quality. We finally implement an entire streaming pipeline for real-time deployment, with the speech quality confirming the framework’s emotional alignment, contextual coherence, and expressive fluency. Codes, cases, and demos are available in https://sixingdeguo.github.io/EmoQ-page/.

Zewen Long, Yu Peng, Fangming Dong, Congyi Li, Xingmao Guan, Shu Wu, Kai Chen

Large language models (LLMs) are widely deployed in real-world applications, yet their safety alignment often fails to generalize beyond the specific linguistic formats seen during training. Prior work has shown that mismatched generalization can lead to alignment failures, but these studies typically rely on fixed or narrow transformation schemes. In this work, we probe safety alignment generalization using language game jailbreaks, a class of linguistically structured transformations that alter surface form while preserving fluency and semantic recoverability. We further introduce custom language games, which parameterize and vary transformation rules, enabling controlled exploration of alignment behavior across closely related linguistic variants. To scale this analysis, we propose AutoLanJail, an automated framework for discovering and refining language game-based jailbreaks. Experiments across open-source and closed-source LLMs show that safety fine-tuning is highly format-specific: defenses trained on one linguistic form fail to generalize to even minimal variations. These findings reveal a structural limitation of current fine-tuning-based alignment methods and highlight the need for safety evaluations that account for systematic linguistic variation.

Yuchen Yang, Lei Peng, Yujie He, Yang yu, Zhongxin Wu, Yanlei Shi

Large Language Models are increasingly integrated into critical applications, yet they remain vulnerable to prompt injection attacks where meticulously designed adversarial inputs bypass safety alignment. Existing defenses often rely on externally deployed guardrail models or response inspection, which incur significant computational overhead and latency. We propose RAP-ID (Robust Alignment Preservation via Injection Defense), a mechanistic, train-free detection framework that operates exclusively on internal state dynamics during the initial forward pass. RAP-ID identifies attacks by detecting their inevitable "impostor" behavior: they must mimic system instruction semantics (Directive Likeness), usurp attention from the true system prompt (Counterfactual Gain), and trigger latent risk concepts (Policy Conflict). By fusing these three internal signals, RAP-ID achieves effective detection across diverse attack vectors—from direct jailbreaks to stealthy agentic manipulations—without requiring text generation. Comprehensive evaluations demonstrate that RAP-ID achieves competitive performance with significant overall improvements compared to heuristic methods. Crucially, as a train-free solution, it incurs minimal computational overhead and delivers fast response times, making it well-suited for real-time deployment.