论文检索

Junkun Qiu, Min Huang, Qinghai Miao

While LLM-based agents have emerged as a focal point for automating data science tasks, they continue to grapple with inefficient context management, "silent failures" (where code executes correctly but fails the task objectives), and error propagation inherent in sequential generation. In this paper, we propose Tree-Notebook, an agentic framework designed to mimic the iterative cognitive process of human data scientists. At its core, Tree-Notebook conceptualizes Jupyter Notebook cells as nodes within a tree structure, facilitating organized and efficient context retrieval. We formalize the task-solving process as a Partially Observable Markov Decision Process (POMDP) over a dynamic tree, utilizing an entropy-based information gain function for path evaluation to enhance adaptability in real-world environments. Furthermore, we introduce the "Data Shadow" system, which resolves silent failures by performing real-time tracking of data distributions, provenance, and semantic constraints. Experimental results demonstrate that Tree-Notebook achieves state-of-the-art (SOTA) performance on both InfiAgent-DABench and DSBench. To further evaluate robustness, we introduce an augmented version of InfiAgent-DABench to simulate complex environments, where Tree-Notebook consistently maintains its SOTA standing. Code is available at: https://github.com/QJK-BUAA/Tree-Notebook

Xinping Zhao, Shouzheng Huang, Yan Zhong, Xinshuo Hu, Meishan Zhang, Baotian Hu, Min Zhang

Retrieval-Augmented Generation (RAG) effectively improves the accuracy of Large Language Models (LLMs). However, retrieval noises significantly undermine the quality of LLMs’ generation, necessitating the development of denoising mechanisms. Previous works extract evidence straightforwardly without deep thinking, which may risk filtering out key clues and struggle with generalization. To this end, we propose EviOmni, which learns to extract rational evidence via reasoning first and then extracting. Specifically, EviOmni integrates evidence reasoning and evidence extraction into one unified trajectory, followed by knowledge token masking to avoid information leakage, optimized via on-policy reinforcement learning with verifiable rewards in terms of answer, length, and format. Extensive experiments on five benchmark datasets show the superiority of EviOmni, which provides compact and high-quality evidence, enhances the accuracy of downstream tasks, and supports both traditional and agentic RAG systems.

Hengyi Feng, Zeang Sheng, Meiyi Qiang, Yang Li, Wentao Zhang

Despite the remarkable success of multimodal large language models (MLLMs) in generative tasks, we observe that they exhibit a counterintuitive deficiency in the zero-shot multimodal retrieval task. In this work, we investigate the underlying mechanisms that hinder MLLMs from being effective retrievers. With the help of sparse autoencoders (SAEs), we decompose MLLM output representations into interpretable semantic concepts to probe their intrinsic behavior. Our analysis reveals that the representation space of MLLMs is overwhelmingly dominated by textual semantics; and the visual semantics essential for multimodal retrieval only constitute a small portion. We find that this imbalance is compounded by the heavy focus of MLLMs on bridging image-text modalities, which facilitates generation but homogenizes embeddings and finally diminishes the discriminative power required for multimodal retrieval. We further discover that the specific feature components that contribute most to the similarity computations of MLLMs are actually distractors that greatly reduce retrieval performance. Building on these insights, we propose , a test-time adaptation approach that applies a whitening transformation to adjust the geometry of MLLM representation spaces. Empirical results show that this simple intervention consistently improves zero-shot multimodal retrieval performance across diverse MLLMs without fine-tuning efforts.

Yuwei Guo, Zihan Zhao, Xiaowei Liu, Xiangning Yu, Qun Ma, Deyu Zhou, Xiao Xue

Multi-agent simulation based on LLMs has increasingly emerged as a new paradigm for exploring complex social phenomena and validating theoretical hypotheses. However, traditional experimental design in the social sciences relies heavily on interdisciplinary expert knowledge, involving cumbersome procedures and high technical barriers. While LLM-driven agents demonstrate broad prospects for designing experiments, their limitations regarding reliability and scientific rigor continue to significantly hinder their in-depth application in social science research. To address these challenges, this paper proposes FSTS, an automated framework for multi-agent experiment design based on script generation. Drawing on the concept of the "Decision Theater," the framework deconstructs experimental design into three core phases: Script Composition, Script Finalization, and Actor Generation. Tests across multiple scenarios indicate that the agents generated by this framework can enact the script within the "experimental theater," reproducing results consistent with real-world situations. The proposal of FSTS not only effectively lowers the barrier for social science experimental design but also provides scientifically grounded decision support for policy-making.

Changhai Zhou, Shiyang Zhang, Yuhua Zhou, Jun Gao, Qian Qiao, Shichao Weng, Weizhong Zhang, Cheng Jin

Deploying and fine-tuning Large Language Models (LLMs) on resource-constrained edge devices requires navigating a strict trade-off between memory footprint and task performance. Existing quantization-aware fine-tuning methods typically decouple weight precision and adapter capacity, overlooking that a layer’s ability to adapt is constrained by the information preserved in its frozen weights. Layers that are highly sensitive to quantization—whether due to representational specialization or accumulated error propagation—can become bottlenecks that adapter rank alone cannot recover. To address this issue, we introduce QR-Adaptor, a unified framework that jointly optimizes per-layer quantization bit-width and LoRA rank. We formulate resource allocation as a multi-objective discrete search guided by empirical layer-wise sensitivity, and implement it with a three-stage pipeline comprising KL-based sensitivity profiling, evolutionary exploration, and Bayesian refinement. Extensive experiments across LLaMA and Qwen models, including modern instruction tuning on OpenOrca and comparisons with strong PEFT baselines such as QDoRA, show that QR-Adaptor establishes a strong Pareto frontier: under a strict 4-bit memory budget, it matches or approaches 16-bit baselines while using substantially less memory.

Lei Wang, Min Huang, Eduard Dragut

Qualitative analysis is essential for studying complex social and behavioral phenomena, yet existing large language model (LLM) approaches face key limitations. Fully automated pipelines often compromise methodological rigor, while fully manual coding remains costly and labor-intensive. Although recent work emphasizes human–AI collaboration, existing multi-agent systems focus primarily on theme-level outputs, provide limited human oversight, and overlook fine-grained, data-level coding quality.We introduce CentaurTA, an iterative, self-improving human–agent framework for scalable thematic analysis. CentaurTA places humans in the loop to oversee and guide analysis, using expert feedback as a persistent learning signal to drive prompt-level refinement. By combining structured human feedback with rubric-based evaluation, CentaurTA provides fine-grained supervision for both open coding and theme construction while preserving methodological rigor. Experiments across multiple datasets, baselines, and LLM families show that CentaurTA improves coding alignment and transparency, highlighting the central role of human feedback in reliable qualitative analysis. Our code and data are available at https://github.com/Tom-Owl/CentaurTA.

Kyeongman Park, Minha Jhang, Kyomin Jung

Modern generative models still lack human-level creativity, particularly in multi-branch diversity. Prior approaches to address this problem often incur heavy computation or strong dependency on model architecture. Therefore, we introduce **UAG**(**U**niversal **A**voidance **G**eneration), a model-agnostic and computationally efficient generation strategy that penalizes similarity among previously generated outputs. Thus, UAG can enhance multi-branch diversity across both diffusion and transformer models, with minimal additional computation. In experiments, our method achieves up to 1.9 times higher diversity, runs 4.4 times faster, and requires only 1/64 of the FLOPs compared to state-of-the-art methods.

Yifan Yang, Jinghui Lu, Evadeng, Ao Yang, Peijie Yu, TingHao YU, Feng Zhang

Autonomous agents powered by large language models (LLM-based agents) are capable of using off-the-shelf tools to interact with the environment, solve real-world problems, and boost work efficiency. However, current approaches to enhancing tool use for LLM-based agents primarily focus on post-training fine-tuning or test-time context extension. These methods overlook the fundamental tool knowledge acquisition during the early training phase, where models actually learn and internalize core knowledge representations, restricting model performance on out-of-distribution tool usage. To solve such a problem, we introduce enhancing tool knowledge for LLM-based agents during continuous pre-training (ToolCPT). We identify and bridge a key gap in current LLM training by shifting focus from tool-calling patterns to deep internalization of core tool-knowledge representations. We begin by curating 5.1 million code artifacts from large-scale, high-quality code repositories. These artifacts are selected based on a set of criteria that defines a usable "proxy agent tool", thereby forming a comprehensive agent tool library. For each proxy tool, we then create a detailed playbook covering implementation specifications, core functionalities, interaction protocols with other tools, and illustrative positive and negative examples. This process yields a large-scale tool knowledge corpus comprising 18 billion tokens, which is used to continuously pre-train our model. Experiments show our playbook-enhanced corpus catalyzes deep knowledge internalization, driving the model to notable performance gains on multiple standard benchmarks.

Ho Hung Lim, Yi Yang

Visual RAG has offered an alternative to traditional RAG. It treats documents as images and uses vision encoders to obtain vision patch tokens. However, hundreds of patch tokens per document create retrieval and storage challenges in a vector database. Practical deployment requires aggregating them into a single vector. This raises a critical question: does single-vector aggregation lose key information in financial documents? We develop a diagnostic benchmark using financial documents where changes in single digits can lead to significant semantic shifts. Our experiments show that single-vector aggregation collapses different documents with almost identical vectors. Metrics show that the patch level detects semantic changes, and confirm that aggregation obscures these details. We identify global texture dominance as the root cause. Our findings are consistent across model scales, retrieval-optimized embeddings, and multiple mitigation strategies, highlighting significant risks for single-vector visual document retrieval in financial applications.

Keyang Zhong, Junlin Xie, Hefeng Wu, Haofeng Li, Guanbin Li

Vision-language models (VLMs) have shown impressive capabilities in perceptual tasks, yet they degrade in complex multi-hop reasoning under multi-player game settings with imperfect and deceptive information. In this paper, we pick up a representative multi-player task, Murder Mystery Games, which require to infer hidden truths based on partial clues provided by the roles of different intentions. To address this challenge, we propose a collaborative multi-agent framework for evaluating and synthesizing high-quality, role-driven multi-player game scripts, enabling fine-grained interaction patterns tailored to character identities (i.e., murderer vs. innocent). Our system generates rich multimodal contexts—including character backstories, visual/textual clues, and multi-hop reasoning chains—through coordinated agent interactions. We design a two-stage agent-monitored training strategy to enhance the reasoning ability of VLM: (1) Chain-of-Thought based fine-tuning on curated and synthetic datasets that model uncertainty and deception; (2) GRPO-based Reinforcement Learning with agent-monitored reward shaping, encouraging the model to develop character-specific reasoning behaviors and effective multi-modal multi-hop inference. Extensive experiments demonstrate that our method significantly boosts the performance of VLM in narrative reasoning, hidden fact extraction, and deception-resilient understanding. Our contributions offer a scalable solution for training and evaluating VLMs under uncertain, adversarial, and socially complex conditions, laying the groundwork for future benchmarks in multimodal multi-hop reasoning under imperfect information.

Zhiyin Yu, Bo Zhang, Qibin Hou, Zhonghai Wu, Xiao Luo, Lei Bai

Previous LLMs-based RL studies typically follow either supervised learning with high annotation costs, or unsupervised paradigms using voting or entropy-based rewards. However, their performance remains far from satisfactory due to the substantial annotation cost and issues such as model collapse or reward hacking. To address these issues, we introduce a new perspective inspired by cognitive learning theory and propose a novel approach called EasyRL. The core of EasyRL is to simulate the human cognitive acquisition curve by integrating reliable knowledge transfer from easy labeled data with a progressive divide-and-conquer strategy that tackles increasingly difficult unlabeled data. Specifically, we initialize a warm-up model using supervised RL with few-shot labeled data. This is followed by a divide-and-conquer pseudo-labeling strategy on difficult unlabeled data, combining consistency-based selection for low-uncertainty cases and reflection-based resolution for medium-uncertainty cases. Finally, difficulty-progressive self-training with iterative pseudo-labeling and RL further strengthens the model’s reasoning capability. EasyRL provides a unified self-evolving framework that facilitates data-efficient post-training of LLMs. Experimental results on mathematical and scientific benchmarks demonstrate that EasyRL, using only 10% of easy labeled data, consistently outperforms state-of-the-art baselines.

Gengyuan Hu, Haoxiang Liu, Chenhong Cao, Shilei Tan, Wei Gong

Large Multimodal Models (LMMs) have demonstrated significant potential in the medical domain, achieving impressive performance on tasks ranging from report generation to visual question answering. However, existing benchmarks predominantly focus on static evaluation, assessing models on isolated data points. This approach neglects a critical aspect of clinical practice: longitudinal analysis, where physicians interpret patient data as a dynamic trajectory to track disease progression and treatment response. To address this gap, we introduce ELTLM, the first benchmark specifically tailored to assess the temporal perception and reasoning capabilities of medical LMMs. Constructed from temporal chest X-rays, ELTLM features a hierarchical task taxonomy comprising Temporal Perception QA and Temporal Reasoning QA, requiring models to detect fine-grained visual changes and infer high-level clinical trends. Our evaluation of state-of-the-art models reveals that while they excel in static scenarios, they struggle significantly with temporal grounding and consistency. ELTLM serves as a vital resource to identify these limitations and guide the development of future time-aware medical AI systems. Our data is available at [ELTLM](https://github.com/ChengFeng233/ELTLM-Bench).

He Xiao, Qingyao Yang, Dirui Xie, Wendong XU, Zunhai Su, Runming Yang, Haobo Liu, Wenyong Zhou, Zhengwu Liu, Ngai Wong

Large language models with billions of parameters are often over-provisioned: many layers contribute little unique information yet dominate the memory and energy footprint during inference. We present LieQ (Layer-wise information effectiveness Quantization), a hardware-native, metric-driven post-training quantization framework that addresses the critical challenge of maintaining accuracy in sub-8B models, model parameters less than 8B, under extreme low-bit compression. LieQ keeps uniform bit-width within each layer while mixing precision across layers, preserving standard multiplication kernels and avoiding irregular memory access, codebooks, or irregular formats at inference time. Our method uncovers a strong correlation between layer-wise functional saliency and representational compactness, revealing that layers with higher training-induced energy concentration are functionally irreplaceable. Leveraging this insight, we propose a purely geometry-driven sensitivity proxy that enables automatic bit-width allocation under a target average-bit budget without expensive gradient updates or inference-based perplexity probing. Under an average weight bit-width approaching two bits per parameter, LieQ consistently reduces the large accuracy gap typically observed for naive uniform 2-bit baselines on Qwen3 and LLaMA3.x families, while retaining standard-kernel efficiency. These properties make LieQ a practical path toward deploying small language models on resource-constrained edge devices. Code will be available at: https://github.com/HeXiao-55/LieQ-official.git.

Song-Li Wu, Zhaocheng Du, Xianquan Wang, Jingyi Wang

Lifelong knowledge editing (LKE) aims to incrementally correct factual inaccuracies in large language models (LLMs), but sequential edits can lead to substantial degradation of capabilities. Existing approaches primarily rely on static parameter regularization, which restricts knowledge integration and fails to prevent cumulative capability degradation. We argue that an important source of this degradation lies in the temporal mismatch between locally editable factual knowledge and procedural knowledge, which is gradually acquired, guides task execution, and cannot be reliably updated by rapid edits. To this end, we formulate LKE as a dual-timescale process, explicitly decoupling fast-updating factual knowledge from slow-evolving procedural knowledge. Based on this formulation, we propose CaPEdit, a framework that preserves model capabilities under LKE. It first synthesizes procedural knowledge across successive edits, and subsequently performs parameter updates guided jointly by factual supervision and the synthesized procedural signal. To ensure stability under long edit sequences, CaPEdit is trained via a hybrid optimization scheme, combining step-wise updates for rapid factual correction with trajectory-level optimization to facilitate gradual procedural adaptation. Experiments demonstrate that CaPEdit improves capability preservation across all fundamental capabilities by 49.78%, achieves superior editing performance, and requires only 18.07% of the editing time of most existing methods.

Zihan Chen, Yiming Zhang, Hengguang Zhou, Zenghui Ding, Yining Sun, Cho-Jui Hsieh

Current benchmarks are inadequate for evaluating progress in reinforcement learning (RL) for large language models (LLMs). Despite recent benchmark gains reported for RL, we find that training on these benchmarks’ training sets achieves nearly the same performance as training directly on the test sets, suggesting that the benchmarks cannot reliably separate further progress. To study this phenomenon, we introduce a diagnostic suite and the Oracle Performance Gap (OPG) metric that quantifies the performance difference between training on the train split versus the test split of a benchmark. We further analyze this phenomenon with stress tests and find that, despite strong benchmark scores, existing RL methods struggle to generalize across distribution shifts, varying levels of difficulty, and counterfactual scenarios: shortcomings that current benchmarks fail to reveal. We conclude that current benchmarks are insufficient for evaluating generalization and propose three core principles for designing more faithful benchmarks: sufficient difficulty, balanced evaluation, and distributional robustness.

Guhong Chen, Chenghao Sun, Cheng Fu, Qiyao Wang, Zhihong Huang, ChaoPeng Wei, Guangxu Chen, Feiteng Fang, Ahmadreza Argha, Bing Zhao 等

As code large language models (LLMs) evolve into tool-interactive agents via the Model Context Protocol (MCP), their generalization is increasingly limited by low-quality synthetic data and the diminishing returns of quantity scaling; moreover, quantity-centric scaling exhibits an early bottleneck that underutilizes trajectory data. We propose TDScaling, a Trajectory Diversity Scaling-based data synthesis framework for code agents that scales performance through diversity rather than raw volume. Moreover, TDScaling is more data-efficient: under a fixed training budget, increasing trajectory diversity yields larger gains than adding more trajectories, improving the performance-cost trade-off for agent training. TDScaling integrates four innovations: (1) a Business Cluster mechanism that captures real-service logical dependencies; (2) a Blueprint-driven multi-agent paradigm that enforces trajectory coherence; (3) an adaptive evolution mechanism that steers synthesis toward long-tail scenarios using Domain Entropy, Reasoning Mode Entropy, and Cumulative Action Complexity to prevent mode collapse; and (4) a sandboxed code tool that mitigates catastrophic forgetting of intrinsic coding capabilities. Experiments on general tool-use benchmarks (BFCL, \tau^2-Bench) and code agent tasks (RebenchT, CodeCI, BIRD) demonstrate a win-win outcome: TDScaling improves both tool-use generalization and inherent coding proficiency. Crucially, we show that trajectory diversity scaling attains a substantially higher performance ceiling than quantity scaling, establishing a resource-efficient paradigm for training robust code agents under data bottlenecks.

Wanlong Liu, Junying Chen, Yunjin Yang, Prayag Tiwari, Wenyu Chen, Benyou Wang

In recent years, large language models (LLMs) have demonstrated remarkable capabilities in the medical domain. However, existing medical benchmarks suffer from performance saturation and are predominantly derived from medical exam questions, which fail to reflect the complexity of real-world clinical scenarios.To bridge this gap, we introduce ClinBench, a challenging benchmark based on authentic clinical cases sourced from authoritative medical journals. Each question retains the complete patient information and clinical test results from the original case, effectively simulating real-world clinical practice. Additionally, we implement a rigorous human review process involving medical experts to ensure the quality and reliability of the benchmark. ClinBench supports both textual and multimodal evaluation formats, covering 11 medical specialties with over 2,000 questions, including a dedicated rare disease track, providing a comprehensive resource for assessing the medical reasoning capabilities of LLMs. We evaluate the performance of over 20 open-source and proprietary LLMs and benchmark them against human medical experts. Our findings reveal that human experts still retain an advantage within their specialized fields, while LLMs demonstrate superior overall performance on a broader range of medical specialties.

Pollawat Hongwimol, Haoning Shang, Chutong Wang, Zhichao Wan, Yi Gao, Yuanming Li, Lin Gui, Wenhao Sun, Cheng Yu

Product attribute extraction in e-commerce is bottlenecked by ontologies that are inconsistent, incomplete, and costly to maintain. We present AutoPKG, a multi-agent Large Language Model (LLM) framework that automatically constructs a Product-attribute Knowledge Graph (PKG) from multimodal product content. AutoPKG induces product types and type-specific attribute keys on demand, extracts attribute values from text and images, and consolidates updates through a centralized decision agent that maintains a globally consistent canonical graph. We also propose an evaluation protocol for dynamic PKGs that measures type/key validity and consolidation quality, as well as edge-level accuracy for value assertions after canonicalization. On a large real-world marketplace catalog dataset from Lazada (Alibaba), AutoPKG achieves up to 0.953 Weighted Knowledge Efficiency (WKE) for product types, 0.724 WKE for attribute keys, and 0.531 edge-level F1 for multimodal value extraction. Across three public benchmarks, we improve edge-level exact-match F1 by 0.152 and yield a 0.208 precision gain on the attribute extraction application. Online A/B tests show that AutoPKG-derived attributes increase Gross Merchandise Value (GMV) in Badge (+3.81%), Search (+5.32%), and Recommendation (+7.89%), supporting AutoPKG’s practical value in production.

Xuxian Hu, Zhu Teng, Wei Zhang, Ming He, Jianping Fan

Retrieval-Augmented Generation (RAG) systems are widely used to mitigate the stateless nature of Large Language Models (LLMs) in long-term and personalized interactions by incorporating external memory. However, existing approaches often prioritize memory organization, such as knowledge graphs, while overlooking a critical semantic gap between implicit, intent-driven queries and explicit, narrative-based memories. To bridge this gap, we propose QueryLink, a novel framework that leverages Query-Memory Alignment to project both queries and memories into a shared semantic space. It significantly boosts recall by facilitating multi-grained retrieval of semantically relevant information. To further enhance memory retrieval, we leverage Coherent Memory Chunking, a mechanism that processes memories in multi-turn dialogue units, preserving semantic integrity, rather than relying on fixed-size segments. Extensive experiments on the LoCoMo and LongMemEval benchmark demonstrate that QueryLink significantly outperforms SOTA methods, achieving at least a 7% improvement in reasoning accuracy (measured by LLM). Additionally, QueryLink can be integrated as a plug-and-play component to boost existing vector-based systems like A-MEM, leading to improvements of over 6% in both F1 and B1 scores.The code is available at https://github.com/Dontplay0112/querylink.

Shuyan Ke, Qiong Wu, Hui Li, Liujuan Cao

Large language models (LLMs) are increasingly deployed in high-stakes domains reliant on tabular data (e.g., financial reporting), where undetected logical inconsistencies such as mismatched totals and components can lead to critical errors. Yet, the ability of LLMs to identify such inconsistencies remains poorly understood, hindered by the absence of standardized evaluation frameworks and cell-level annotated datasets. To bridge this gap, we propose a comprehensive benchmark SEC-Fintables comprising 103,395 real-world and error-injected table instances, alongside a novel evaluation protocol that decomposes inconsistency detection into granular sub-tasks. Through evaluating both proprietary and open-source LLMs on SEC-Fintables, we find that contemporary LLMs exhibit only partial competence in detecting logical inconsistencies. Our study reveals key limitations and improvement opportunities for LLMs. We believe SEC-Fintables and our evaluation protocol can serve as a practical resource for advancing reliable inconsistency detection of LLMs in tabular reasoning. We release SEC-Fintables at https://github.com/XIEFOX/SEC-Fintables.