论文检索

Songcheng Cai, Zhiheng Lyu, Yuansheng Ni, Xiangchao Chen, Baichuan Zhou, Shenzhe Zhu, Yi Lu, Haozhe Wang, Chi Ruan, Benjamin Schneider 等

Agentic repository-level code understanding is essential for automating complex software engineering tasks, yet the field lacks reliable benchmarks. Existing evaluations often overlook the long tail topics and rely on popular repositories where Large Language Models (LLMs) can cheat via memorized knowledge. To address this, we introduce SWE-QA-Pro, a benchmark constructed from diverse, long-tail repositories with executable environments. We enforce topical balance via issue-driven clustering to cover under-represented task types and apply a rigorous difficulty calibration process: questions solvable by direct-answer baselines are filtered out. This results in a dataset where agentic workflows significantly outperform direct answering (e.g., a ~13-point gap for Claude Sonnet 4.5), confirming the necessity of agentic codebase exploration. Furthermore, to tackle the scarcity of training data for such complex behaviors, we propose a scalable synthetic data pipeline that powers a two-stage training recipe: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning from AI Feedback (RLAIF). This approach allows small open models to learn efficient tool usage and reasoning. Empirically, a Qwen3-8B model trained with our recipe surpasses GPT-4o by 2.3 points on SWE-QA-Pro and substantially narrows the gap to state-of-the-art proprietary models, demonstrating both the validity of our evaluation and the effectiveness of our agentic training workflow.

Bolei Ma, Yina Yao, Anna-Carolina Haensch

Large Language Models (LLMs) are increasingly applied to creative domains, yet their performance in classical Chinese poetry generation and evaluation remains poorly understood. We propose a three-step evaluation framework that combines computational metrics, LLM-as-a-judge assessment, and human expert validation. Using this framework, we evaluate six state-of-the-art LLMs across multiple dimensions of poetic quality, including themes, emotions, imagery, form, and style, in the context of Tang poetry (唐诗) generation. Our analysis reveals a critical "echo chamber" effect: LLMs systematically overrate machine-generated poems that mimic statistical patterns yet fail strict prosodic rules, diverging significantly from human expert judgments. These findings underscore the limitations of using LLMs as standalone evaluators for culturally complex tasks, highlighting the necessity of hybrid human-model validation frameworks.

Loris Bergeron, Ioana Buhnila, Jerome Francois, Radu State

Large Language Models excel at NLP tasks but remain prone to hallucinations, limiting trust in real-world applications. We present HalluGuard, a 4B-parameter Small Reasoning Model (SRM) designed as a guardrail for Retrieval-Augmented Generation (RAG) pipelines, which classify document-claim pairs as grounded or hallucinated in closed-book, document-grounded settings and produces evidence-grounded justifications. Our approach combines (i) a domain-agnostic synthetic dataset derived from FineWeb and refined through multi-stage curation and data reformation, (ii) synthetic grounded and hallucinated claims, and (iii) preference-based fine-tuning with Odds Ratio Preference Optimization (ORPO) to distill large-model reasoning into a smaller backbone. On the RAGTruth subset of the LLM-AggreFact benchmark, HalluGuard achieves 84.4% balanced accuracy (BAcc), surpassing specialized models, MiniCheck (7B; 84.0%) and Granite Guardian 3.3 (8B; 82.2%) while using roughly half their parameters. Across the benchmark, it reaches 77.1% BAcc, surpassing larger general-purpose LLMs such as GPT-4o (75.9%). HalluGuard and datasets will be released upon acceptance.

Peiding Wang, Li Zhang, Fang Liu, Chongyang Tao, Yinghao Zhu

Large language models (LLMs) substantially enhance developer productivity in repository-level code generation through interactive collaboration. However, as interactions progress, repository context must be continuously preserved and updated to integrate newly validated information. Meanwhile, the expanding session history increases cognitive burden, often leading to forgetting and the reintroduction of previously resolved errors. Existing memory management approaches show promise but remain limited by natural language-centric representations. To overcome these limitations, we propose CodeMEM, an AST-guided dynamic memory management system tailored for repository-level iterative code generation. Specifically, CodeMEM introduces the Code Context Memory component that dynamically maintains and updates repository context through AST-guided LLM operations, along with the Code Session Memory that constructs a code-centric representation of interaction history and explicitly detects and mitigates forgetting through AST-based analysis. Experimental results on the instruction-following benchmark CodeIF-Bench and the code generation benchmark CoderEval demonstrate that CodeMEM achieves state-of-the-art performance, improving instruction following by 12.2% for the current turn and 11.5% for the session level, and reducing interaction rounds by 2–3, while maintaining competitive inference latency and token efficiency.

Meng Xi, Sihan Lv, Yechen Jin, Guanjie Cheng, Naibo Wang, Ying Li, Jianwei Yin

Retrieval-Augmented Generation (RAG) systems based on Large Language Models (LLMs) have become a core technology for tasks such as question-answering (QA) and content generation. RAG poisoning is an attack method to induce LLMs to generate the attacker’s expected text by injecting poisoned documents into the database of RAG systems. Existing research can be broadly divided into two classes: white-box methods and black-box methods. White-box methods utilize gradient information to optimize poisoned documents, and black-box methods use a pre-trained LLM to generate them. However, existing white-box methods require knowledge of the RAG system’s internal composition and implementation details, whereas black-box methods are unable to utilize interactive information. In this work, we propose the RIPRAG attack framework, an end-to-end attack pipeline that treats the target RAG system as a black box and leverages our proposed Reinforcement Learning from Black-box Feedback (RLBF) method to optimize the generation model for poisoned documents. We designed two kinds of rewards: similarity reward and attack reward. Experimental results demonstrate that this method can effectively execute poisoning attacks against most complex RAG systems, achieving an attack success rate (ASR) improvement of up to 0.72 compared to baseline methods. This highlights prevalent deficiencies in current defensive methods and provides critical insights for LLM security research.

Haeun Jang, Hwan Chang, Hwanhee Lee

The deployment of Large Vision-Language Models (LVLMs) for real-world document question answering is often constrained by dynamic, user-defined policies that dictate information disclosure based on context. While ensuring adherence to these explicit constraints is critical, existing safety research primarily focuses on implicit social norms or text-only settings, overlooking the complexities of multimodal documents. In this paper, we introduce Doc-PP (Document Policy Preservation Benchmark), a novel benchmark constructed from real-world reports requiring reasoning across heterogeneous visual and textual elements under strict non-disclosure policies. Our evaluation highlights a systemic Reasoning-Induced Safety Gap: models frequently leak sensitive information when answers must be inferred through complex synthesis or aggregated across modalities, effectively circumventing existing safety constraints. Furthermore, we identify that providing extracted text improves perception but inadvertently facilitates leakage. To address these vulnerabilities, we propose DVA (Decompose–Verify–Aggregation), a structural inference framework that decouples reasoning from policy verification. Experimental results demonstrate that DVA significantly outperforms standard prompting defenses, offering a robust baseline for policy-compliant document understanding.

Feng Luo, Yu-Neng Chuang, Guanchu Wang, Hoang Anh Duy Le, Shaochen Zhong, Hongyi Liu, Jiayi Yuan, Yang Sui, Vladimir Braverman, Vipin Chaudhary 等

Reasoning-capable large language models (LLMs) achieve strong performance on complex tasks but often exhibit overthinking after distillation, generating unnecessarily long chain-of-thought (CoT) reasoning even for simple inputs and incurring high inference cost. However, naively shortening reasoning length can degrade reasoning accuracy, as concise reasoning may be insufficient for certain inputs and lacks explicit supervision. We propose Auto Long-Short Reasoning (AutoL2S), a distillation framework that empowers non-reasoning LLMs to think thoroughly but only when necessary. AutoL2S first learns a lightweight switching token with verified long-short CoTs to enable instance-wise long-short reasoning selection. Then it leverages long-short reasoning rollouts induced by switching tokens within a GRPO-style loss to improve reasoning efficiency while maintaining accuracy. Experiments demonstrate that AutoL2S effectively reduces reasoning length up to 71% with minimal accuracy loss, yielding markedly better trade-off in token length and inference time while preserving accuracy. The code is available at https://github.com/amandaluof/AutoL2S.

Xinyan Guan, Jiali Zeng, Chunlei Xin, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, Fandong Meng

Large language models generate computationally expensive yet semantically void reasoning on beyond-capability tasks, creating safety risks where plausible-sounding but incorrect derivations mislead users. We characterize this futile reasoning phenomenon through systematic analysis, revealing universal capability overreach and systematic miscalibration towards over-confidence. The dominant failure mode is specious reasoning, superficially valid outputs with subtle hallucinations, which escalates with task difficulty. We demonstrate that prompt engineering proves insufficient to calibrate refusal behavior. To address this, we introduce CaRL (Capability-aligned Reinforcement Learning), which aligns model behavior with capability boundaries through reward shaping that incentivizes refusal over hallucination and hindsight augmentation that converts failures into refusal supervision. Experiments demonstrate a substantial reduction in futile reasoning while preserving performance across task difficulties, effectively achieving capability-aligned behavior without sacrificing utility.

Zouying Cao, Jiaji Deng, Li Yu, Weikang Zhou, Zhaoyang Liu, Bolin Ding, Hai Zhao

Procedural memory enables large language model (LLM) agents to internalize ”how-to” knowledge and thus reduce redundant trial-and-error. However, existing frameworks predominantly suffer from a ”passive accumulation” paradigm, treating memory as a static append-only archive. To bridge the gap between static storage and dynamic reasoning, we propose ReMe (Remember Me, Refine Me), a comprehensive framework for experience-driven agent evolution. ReMe manages the memory lifecycle via three mechanisms: 1) multi-faceted distillation, which extracts fine-grained experiences by recognizing success patterns, analyzing failure triggers and generating comparative insights; 2) context-adaptive reuse, which tailors historical insights to new contexts through scenario-aware indexing; and 3) utility-based refinement, which automatically adds validated memories and prunes outdated ones to maintain a compact, high-quality experience pool. Experiments on BFCL-V3 and AppWorld demonstrate that ReMe establishes a new state-of-the-art in agent memory system. Crucially, we observe a significant memory-scaling effect: Qwen3-8B equipped with ReMe outperforms larger, memoryless Qwen3-14B, indicating that self-evolving memory provides a computation-efficient path for lifelong learning.

Chen Wang, Lai Wei, Yanzhi Zhang, Chenyang Shao, Zedong Dan, Weiran Huang, Ge Lan, Yue Wang

Recent advances in reinforcement learning (RL) have improved the reasoning capabilities of large language models (LLMs) and vision-language models (VLMs). However, the widely used Group Relative Policy Optimization (GRPO) consistently suffers from entropy collapse, causing the policy to converge prematurely and lose diversity. Existing exploration methods introduce additional bias or variance during exploration, making it difficult to maintain optimization stability. We propose Unified Entropy Control for Reinforcement Learning (UEC-RL), a framework that provides targeted mechanisms for exploration and stabilization. UEC-RL activates more exploration on difficult prompts to search for potential and valuable reasoning trajectories. In parallel, a stabilizer prevents entropy from growing uncontrollably, thereby keeping training stable as the model consolidates reliable behaviors. Together, these components expand the search space when needed while maintaining robust optimization throughout training. Experiments on both LLM and VLM reasoning tasks show consistent gains over RL baselines on both Pass@1 and Pass@k. On Geometry3K, UEC-RL achieves a 37.9% relative improvement over GRPO, indicating that it sustains effective exploration without compromising convergence and underscoring UEC-RL as a key for scaling RL-based reasoning in large models. Our code is available at https://github.com/597358816/UEC-RL.

Yang Xiao, Eun-Jung Holden, Ting Dang

Recent speech foundation models excel at multilingual automatic speech recognition (ASR) for high-resource languages, but adapting them to low-resource languages remains challenging due to data scarcity and efficiency constraints. Full-model fine-tuning is computationally expensive and prone to overfitting, while parameter-efficient methods like LoRA apply adaptation uniformly across layers, overlooking internal representations thus compromising effectiveness and efficiency. We analyze multilingual ASR models and reveal a U-shaped adaptability pattern: early and late layers are language-specific and require more adaptation, while intermediate layers retain shared semantics and need less. Building on this observation, we propose DAMA, a Depth-Aware Model Adaptation framework that allocates adaptation capacity according to each layer’s role. DAMA also introduces Singular Value Decomposition (SVD)-based initialization to constrain adaptation and preserve the U-shaped pattern, as well as a frozen middle-layer basis for further efficiency. Evaluated on 18 low-resource languages across two benchmark datasets, DAMA matches or surpasses state-of-the-art accuracy with 80% fewer trainable parameters, achieves a 29% error reduction under extreme data scarcity, and significantly improves memory, training time, and computational efficiency over baselines. These results highlight the benefits of structure-aware adaptation for efficient, scalable multilingual ASR.

Yuqi Yang, Weiqi Wang, Baixuan Xu, Wei Fan, Qing Zong, Chunkit Chan, Zheye Deng, Xin Liu, Yifan Gao, Changlong Yu 等

Session history is a common way of recording user interacting behaviors throughout a browsing activity with multiple products. For example, if an user clicks a product webpage and then leaves, it might because there are certain features that don’t satisfy the user, which serve as an important indicator of on-the-spot user preferences. However, all prior works fail to capture and model customer intention effectively because insufficient information exploitation and only apparent information like descriptions and titles are used. There is also a lack of data and corresponding benchmark for explicitly modeling intention in E-commerce product purchase sessions. To address these issues, we introduce the concept of an intention tree and propose a dataset curation pipeline. Together, we construct a sibling multimodal benchmark, SessionIntentBench, that evaluates L(V)LMs’ capability on understanding inter-session intention shift with four subtasks. With 1,952,177 intention entries, 1,132,145 session intention trajectories, and 13,003,664 available tasks mined using 10,905 sessions, we provide a scalable way to exploit the existing session data for customer intention understanding. We conduct human annotations to collect ground-truth label for a subset of collected data to form an evaluation gold set. Extensive experiments on the annotated data further confirm that current L(V)LMs fail to capture and utilize the intention across the complex session setting. Further analysis show injecting intention enhances LLMs’ performances.

Shiwei Lyu, Xidong Wang, Hao Zhu, Lei Liu, Chaohe Zhang, Jian Wang, Jinjie Gu, Benyou Wang, Yue Shen

Although large language models (LLMs) demonstrate expert-level medical knowledge, aligning their open-ended outputs with fine-grained clinician preferences remains challenging. Existing methods often rely on coarse objectives or unreliable automated judges that are weakly grounded in professional guidelines. We propose a two-stage framework to address this gap. First, we introduce HealthRubrics, a dataset of 7,034 physician-verified preference examples in which clinicians refine LLM-drafted rubrics to meet rigorous medical standards. Second, we distill these rubrics into HealthPrinciples: 119 broadly reusable, clinically grounded principles organized by clinical dimensions, enabling scalable supervision beyond manual annotation. We use HealthPrinciples for (1) offline alignment by synthesizing rubrics for unlabeled queries and (2) an inference-time tool for guided self-revision. A 30A3B model with our framework achieves 33.4% on HealthBench-Hard, outperforming much larger models including Deepseek-R1 and o3, establishing a resource-efficient baseline for clinical alignment.

Taeyun Roh, WonJune Jang, Junha Jung, Jaewoo Kang

Large language model agents heavily rely on external memory to support knowledge reuse and complex reasoning tasks. Yet most memory systems store experiences in a single global retrieval pool which can gradually dilute or corrupt stored knowledge. This problem is especially pronounced for small language models (SLMs), which are highly vulnerable to irrelevant context. We introduce CLAG, a CLustering-based AGentic memory framework where an agent actively organizes memory. CLAG employs an SLM-agent driven router to assign each new memory to a semantically coherent cluster. By performing continual evolution within the cluster, it effectively reduces cross-topic interference. During the retrieval phase, CLAG targets a small set of relevant clusters for retrieval, thereby excluding distractors and reducing the search space. Experiments on multiple QA datasets with three SLM backbones show that CLAG consistently improves answer quality and robustness over prior memory systems for agents, remaining lightweight and efficient.

Chenglong Wang, Canjia Li, Xingzhao Zhu, Yifu Huo, Huiyu Wang, Weixiong Lin, Yun Yang, Qiaozhi He, Tian Hua Zhou, Changxiaojia 等

Due to the dynamically evolving nature of real-world query streams, relevance models struggle to generalize to practical search scenarios. A sophisticated solution is self-evolution techniques. However, in large-scale industrial settings with massive query streams, this technique faces two challenges: (1) informative samples are often sparse and difficult to identify, and (2) pseudo-labels generated by the current model could be unreliable. To address these challenges, in this work, we propose a Self-Evolving Relevance Model approach (SERM), which comprises two complementary multi-agent modules: a multi-agent sample miner, designed to detect distributional shifts and identify informative training samples, and a multi-agent relevance annotator, which provides reliable labels through a two-level agreement framework. We evaluated SERM on a large-scale industrial platform, which serves billions of user requests daily. Experimental results demonstrate that SERM can achieve significant performance gains through iterative self-evolution, as validated by extensive offline multilingual evaluations and online testing.

Pengyu Xu, JingRen Hou, Liping Jing, Jian Yu

Large language models (LLMs) enable zero-shot and few-shot multi-label text classification via in-context learning, yet most approaches perform static inference and degrade under streaming test data due to distribution shift and long-tail labels. We study online test-time adaptation for LLM-based multi-label generation without any parameter updates, and identify two bottlenecks: (1) standard generation probabilities provide unreliable confidence because they ignore label competition at key decoding branches; (2) naive confidence-based caching overfits to frequent and easy examples, reducing label coverage and diversity. We propose SCOTTA, a structured confidence-guided online adaptation framework. SCOTTA introduces Label-set Local Likelihood Ratio (L3R), a label-level confidence measure that compares a target label against its valid competitors at critical decision positions. Using L3R as a unified signal, SCOTTA maintains an in-context exemplar cache via streaming submodular maximization, balancing label coverage, semantic diversity, and sample quality under a fixed context budget. Across four benchmarks, SCOTTA consistently improves Micro-F1 and Macro-F1 over strong LLM and non-LLM baselines, with the largest gains on long-tail labels.

Xudong Wang, Chaoning Zhang, Chenghao Li, Shuxu Chen, Qigan Sun, Jiaquan Zhang, Fachrina Dewi Puspitasari, Tae-Ho Kim, Jiwei Wei, Malu Zhang 等

Large Language Models (LLMs) have demonstrated strong capabilities in complex reasoning tasks, while recent prompting strategies such as Chain-of-Thought (CoT) have further elevated their performance in handling complex logical problems. Despite these advances, high-quality reasoning remains heavily reliant on manual static prompts and is sensitive to decoding configurations and task distributions, leading to performance fluctuations and limited transferability. Existing automatic prompt optimization methods typically adopt single-agent local search, failing to simultaneously optimize prompts and decoding hyperparameters within a unified framework to achieve stable global improvements. To address this limitation, we propose Agent-GWO, a dynamic prompt optimization framework for complex reasoning. Specifically, we unify prompt templates and decoding hyperparameters as inheritable agent configurations. By leveraging the leader-follower mechanism of the Grey Wolf Optimizer (GWO), we automatically select three leader agents (\alpha, \beta, and \delta) to guide the collaborative updates of the remaining agents, enabling iterative convergence toward robust optimal reasoning configurations that can be seamlessly integrated for inference. Extensive experiments on multiple mathematical and hybrid reasoning benchmarks across diverse LLM backbones show that Agent-GWO consistently improves accuracy and stability over existing prompt optimization methods.

Xiaoqian Liu, Zhengkun Ge, Jianjin Wang, Haoran Zhang, Yuan Ge, Kaiyan Chang, Chen Xu, Tong Xiao, Zhengtao Yu, Linfeng Zhang 等

Recent advancements in audio diffusion models have significantly improved text-to-audio editing via inversion techniques. However, these models typically rely on dense, fixed-step sampling trajectories to maintain structural integrity during inversion and generation, leading to prohibitive computational costs. We propose AdaTE, a model-agnostic Adaptive Trajectory Extrapolation framework that accelerates the inversion-based editing process by dynamically evaluating only the most critical generative phases. Specifically, we introduce a hierarchical probing mechanism that monitors curvature acceleration and information gain to detect pivotal transitions within the latent flow. This allows the model to selectively skip redundant segments via linear extrapolation while preserving dense neural evaluations for complex semantic changes. Extensive experiments across AudioLDM2, Auffusion, and Tango2 demonstrate that AdaTE achieves up to a 3.9\times speedup with negligible loss in fidelity. AdaTE significantly shifts the Pareto frontier, providing an efficient solution for high-fidelity audio synthesis and editing.

Manjie Xu, Isabella Yin, Xinyi Tu, Chi Zhang, Yixin Zhu

LLMs struggle with Semantic Inertia: the inability to inhibit pre-trained priors (e.g., “Lava is Dangerous”) when dynamic, in-context rules contradict them. We probe this phenomenon using , where physical laws are mutable text rules, enabling precise evaluation of models’ ability to override learned priors when rules change. We quantatively observe that larger models can exhibit inverse scaling: they perform worse than smaller models when natural language reasoning requires suppressing pre-trained associations (e.g., accepting “Lava is Safe”). Our analysis attributes this to natural language encoding, which entangles descriptive semantics and logical rules, leading to persistent hallucinations of familiar physics despite explicit contradictory rules. Here we show that representing dynamics as executable code, rather than descriptive text, reverses this trend and enables effective prior inhibition. We introduce LCV, which fine-tunes models on counterfactual pairs and identifies states with contradictory rules, thereby forcing attention to logical constraints rather than visual semantics. This training-time approach outperforms expensive inference-time search methods in both efficiency and accuracy. Our results demonstrate that representation fundamentally determines whether scaling improves or impairs contextual reasoning. This challenges the assumption that larger models are universally better, with implications for domains that require dynamic overriding of learned priors.

Bichen Wang, Junzhe Wang, Yixin Sun, Xing Fu, Yanyan Zhao, Bing Qin

In recent years, Large Language Models (LLMs) have made significant progress in automated psychological counseling. However, current research focuses on single-session counseling, which doesn’t represent real-world scenarios. In practice, psychological counseling is a process, not a one-time event, requiring sustained, multi-session engagement to progressively address clients’ issues. To overcome this limitation, we introduce a dataset for Multi-Session Psychological Counseling Conversation Dataset (Muspsy). Our Muspsy dataset is constructed using real client profiles from publicly available psychological case reports. It captures the dynamic arc of counseling, encompassing multiple progressive counseling conversations from the same client across different sessions. Leveraging our dataset, we also developed our Muspsy model, which aims to track client progress and adapt its counseling direction over time. Experiments show that our model performs better than baseline models across multiple sessions.