论文检索

Shuo Liu, Zeyu Liang, Xueguang Lyu, Christopher Amato

A large amount of work has been done in Multi-Agent Systems (MAS) for modeling and solving problems with multiple interacting agents. However, most LLMs are pretrained independently and not specifically optimized for coordination. Existing LLM fine-tuning frameworks rely on individual rewards, which require complex reward designs for each agent to encourage collaboration. To address these challenges, we model LLM collaboration as a cooperative Multi-Agent Reinforcement Learning (MARL) problem. We develop a multi-agent, multi-turn algorithm, Multi-Agent Group Relative Policy Optimization (MAGRPO), to solve it, building on current RL approaches for LLMs as well as MARL techniques. Our experiments on LLM writing and coding collaboration demonstrate that fine-tuning MAS with MAGRPO enables agents to generate high-quality responses efficiently through effective cooperation. Our approach opens the door to using MARL methods for LLM collaboration and highlights the associated challenges.

Shuaitong Liu, Renjue Li, Lijia Yu, Lijun Zhang, Zhiming Liu, Gaojie Jin

Recent advances in Chain-of-Thought (CoT) prompting have substantially improved the reasoning capabilities of large language models (LLMs), but have also introduced their computational efficiency as a new attack surface. In this paper, we propose BadThink, the first backdoor attack designed to deliberately induce "overthinking" behavior in CoT-enabled LLMs while ensuring stealth. When activated by carefully crafted trigger prompts, BadThink manipulates the model to generate inflated reasoning traces—producing unnecessarily redundant thought processes while preserving the consistency of final outputs. This subtle attack vector creates a covert form of performance degradation that significantly increases computational costs and inference time while remaining difficult to detect through conventional output evaluation methods. We implement this attack through a sophisticated poisoning-based fine-tuning strategy, employing a novel LLM-based iterative optimization process to embed the behavior by generating highly naturalistic poisoned data. Our experiments on multiple state-of-the-art models and reasoning tasks show that BadThink consistently increases reasoning trace lengths—achieving an over 17× increase on the MATH-500 dataset—while remaining stealthy and robust. This work reveals a critical, previously unexplored vulnerability where reasoning efficiency can be covertly manipulated, demonstrating a new class of sophisticated attacks against CoT-enabled systems.

Shengchao Liu, Xiaoming Liu, Chengzhengxu Li, Zhaohan Zhang, Guoxin Ma, Yu Lan, Shuai Xiao

Large Language Models have shown growing ability to generate fluent and coherent texts that are highly similar to the writing style of humans. Current detectors for Machine-Generated Text (MGT) perform well when they are trained and tested in the same domain but generalize poorly to unseen domains, due to domain shift between data from different sources. In this work, we propose MGT-Prism , an MGT detection method from the perspective of the frequency domain for better domain generalization. Our key insight stems from analyzing text representations in the frequency domain, where we observe consistent spectral patterns across diverse domains, while significant discrepancies in magnitude emerge between MGT and human-written texts (HWTs). The observation initiates the design of a low frequency domain filtering module for filtering out the document-level features that are sensitive to domain shift, and a dynamic spectrum alignment strategy to extract the task-specific and domain-invariant features for improving the detector's performance in domain generalization. Extensive experiments demonstrate that MGT-Prism outperforms state‑of‑the‑art baselines by an average of 0.90% in accuracy and 0.92% in F1 score on 11 test datasets across three domain‑generalization scenarios.

Ruiheng Liu, Xiaobing Chen, Jinyu Zhang, Qiongwen Zhang, Yu Zhang, Bailong Yang

The rapid advancement of Large Language Models (LLMs) has driven significant progress in Natural Language Interface to Database (NLIDB). However, the widespread adoption of LLMs has raised critical privacy and security concerns. During interactions, LLMs may unintentionally expose confidential database contents or be manipulated by attackers to exfiltrate data through seemingly benign queries. While current efforts typically rely on rule-based heuristics or LLM agents to mitigate this leakage risk, these methods still struggle with complex inference-based attacks, suffer from high false positive rates, and often compromise the reliability of SQL queries. To address these challenges, we propose SafeNLIDB, a novel privacy-security alignment framework for LLM-based NLIDB. The framework features an automated pipeline that generates hybrid chain-of-thought interaction data from scratch, seamlessly combining explicit security reasoning with SQL generation. Additionally, we introduce reasoning warm-up and alternating preference optimization to overcome the multi-preference oscillations of Direct Preference Optimization (DPO), enabling LLMs to produce security-aware SQL through fine-grained reasoning without the need for human-annotated preference data. Extensive experiments demonstrate that our method outperforms both larger-scale LLMs and ideal-setting baselines, achieving significant security improvements while preserving high utility.

Rui Liu, Yuan Zhao, Zhenqi Jia

The automatic movie dubbing model generates vivid speech from given scripts, replicating a speaker's timbre from a brief timbre prompt while ensuring lip-sync with the silent video. Existing approaches simulate a simplified workflow where actors dub directly without preparation, overlooking the critical director–actor interaction. In contrast, authentic workflows involve a dynamic collaboration: directors actively engage with actors, guiding them to internalize the context cues, specifically emotion, before performance. To address this issue, we propose a new Retrieve-Augmented Director-Actor Interaction Learning scheme to achieve authentic movie dubbing, termed Authentic-Dubber, which contains three novel mechanisms: (1) We construct a multimodal Reference Footage library to simulate the learning footage provided by directors. Note that we integrate Large Language Models (LLMs) to achieve deep comprehension of emotional representations across multimodal signals. (2) To emulate how actors efficiently and comprehensively internalize director-provided footage during dubbing, we propose an Emotion-Similarity-based Retrieval-Augmentation strategy. This strategy retrieves the most relevant multimodal information that aligns with the target silent video. (3) We develop a Progressive Graph-based speech generation approach that incrementally incorporates the retrieved multimodal emotional knowledge, thereby simulating the actor's final dubbing process. The above mechanisms enable the Authentic-Dubber to faithfully replicate the authentic dubbing workflow, achieving comprehensive improvements in emotional expressiveness. Both subjective and objective evaluations on the V2C-Animation benchmark dataset validate the effectiveness.

Ren-Biao Liu, Chao-Zeng Ma, Anqi Li, Hui Sun, Xin-Ye Li, Ming Li

Large Language Models (LLMs) have demonstrated impressive capabilities in code generation. Like human programmers, LLMs tend to call high-level APIs and libraries to program efficiently. However, this shortcut may hinder LLMs from learning the essential algorithm reasoning, leading instead to rote memorization of API usage. As a result, LLMs often struggle to generalize to new or domain-specific algorithms that lack ready-made library support. In this work, we propose ARBench, a novel benchmark for evaluating LLMs’ ability to generate machine learning algorithms from scratch, beyond merely invoking high-level APIs. It emphasizes algorithmic reasoning and implementation, distinguishing genuine understanding from superficial API usage. It covers fundamental and advanced machine learning tasks, rigorously assessing current LLMs’ capacity to implement these algorithms from scratch. Our evaluation reveals the strengths and weaknesses of state-of-the-art LLMs in algorithmic reasoning and generalization, offering valuable insights to guide future research and development.

Ren-Biao Liu, Jiang-Tian Xue, Chao-Zeng Ma, Hui Sun, Xin-Ye Li, Ming Li

Large language models (LLMs) show significant improvement in code generation. A common practice is sampling multiple candidate codes to increase the likelihood of producing an accurate solution. However, effectively identifying the best candidate from the pool is a significant challenge. Although existing code consensus methods attempt to solve this issue, they suffer from a critical problem: relying on test cases generated by LLMs, which can be flawed or provide incomplete coverage. This problem can result in erroneous validations, causing correct code to fail flawed tests and preventing the detection of functional differences in candidate code solutions. To address these issues, we present the Dynamic-Static Synergistic Selection Method, a novel framework that combines two complementary analytical approaches. First, it uses the abstract syntax tree (AST) to detect and filter candidate solutions and test cases. Second, the method statically analyzes the quality of the solutions and then dynamically validates functional consistency based on the execution results of the extracted inputs, thereby neutralizing the impact of faulty tests. Extensive experiments demonstrate that this synergistic approach significantly outperforms existing methods, substantially enhancing the correctness of the selected code.

Ming Liu, Hao Chen, Jindong Wang, Liwen Wang, Jingchen Sun, Wensheng Zhang

Vision-Language Models (VLMs) have achieved success in tasks such as visual question answering, yet their resilience to distractions remains underexplored. Understanding how distractions affect VLMs' performance is crucial for real-world applications, as input data often contains noisy or irrelevant content. This paper assesses the robustness of VLMs—including general-purpose models and those specialized for reasoning—against distractions in the context of science question answering. We introduce I-ScienceQA, a new benchmark based on the ScienceQA dataset, which systematically injects distractions into both visual and textual contexts. We evaluate how distractions perturb the underlying reasoning processes of these models by analyzing changes in textual explanations leading to answers. Our findings show that most VLMs are vulnerable to distractions, with a noticeable degradation in reasoning when extraneous content is present. In particular, some models (including GPT-o4 mini) exhibit a higher degree of robustness. We also observe that textual distractions generally cause greater performance declines than visual distractions. Finally, we explore mitigation strategies such as prompt engineering. Although these strategies improve resilience modestly, our analysis highlights considerable room for further improvement in the robustness of VLMs.

Jingping Liu, Xueyan Wu, Hanxuan Chen, Ziyan Liu, Zhangquan Chen, Ronghao Chen, Huacan Wang

Early childhood is a critical stage for cognitive development, involving core skills such as visual perception and reasoning. While multimodal large language models (MLLMs) have made rapid progress in various general-purpose tasks, their ability to support early education remains largely underexplored. Existing research on child-related AI largely centers on modeling language, emotion, or behavior, with limited focus on evaluating cognitive tasks relevant to early learning. To address this gap, we propose ChildBench, a multimodal benchmark designed to assess models on tasks inspired by early childhood cognitive development. It covers five key domains through ten tasks, including spatial reasoning, visual reasoning, visual discrimination, counting skills, and visual tracking. The benchmark includes 4,890 carefully constructed images and 5,346 manually annotated samples, ensuring both diversity and age-appropriate content. We evaluate a range of state-of-the-art (SoTA) open-source and closed-source MLLMs—including GPT-4o, Gemini, and Qwen2.5-VL—on ChildBench. Despite strong performance on other benchmarks, the best 7B-parameter model with LoRA tuning achieves only 52.01% accuracy, far below the 96% achieved by 5-year-old children. These results reveal critical limitations in fine-grained perception and reasoning. We further analyze failure cases and discuss directions for future model development.

Jiaqi Liu, Yankun Yang, Jiakang Xu, Zhongqiang Du, Wenbin Jiang

Cognitive-functional dialogues, such as those for persuasion, consultation, and question-answering, are prevalent throughout human social interaction. The core difference between these dialogues and casual chat lies in their objective: to guide a person's cognitive and psychological state toward a predetermined one. Existing conversational technologies perform poorly in handling such dialogues. The fundamental reason is that the transformation of human cognitive psychology follows specific patterns, yet existing technologies neither account for these patterns nor possess cognitive guidance planning based on them. This deficiency makes it difficult for dialogues to achieve their intended cognitive-functional goals effectively. To address this, we propose a dynamic cognitive planning method (DyCoP). By modeling the long-term evolution of a user's cognitive psychology during the dialogue process, this method dynamically generates dialogue guidance plans that align with the principles of cognitive-psychological evolution. This allows for the generation of appropriate dialogue responses based on prior user psychology and the immediate conversational context, thereby achieving cognitive-functional goals more efficiently and accurately. Simultaneously, we constructed an evaluation framework for cognitive-functional dialogues and constructed a richly annotated emotional support conversation dataset. Comprehensive automatic and human evaluations show that our proposed DyCoP method demonstrates significant advantages over existing baseline models.

Jiacheng Liu, Mayi Xu, Qiankun Pi, Wenli Li, Ming Zhong, Yuanyuan Zhu, Mengchi Liu, Tieyun Qian

Large Language Models (LLMs) are increasingly employed in applications that require processing information from heterogeneous formats, including texts, tables, infoboxes, and knowledge graphs. However, systematic biases toward particular formats may undermine LLMs' ability to integrate heterogeneous data impartially, potentially resulting in reasoning errors and increased risks in downstream tasks. Yet it remains unclear whether such biases are systematic, which data-level factors drive them, and what internal mechanisms underlie their emergence. In this paper, we present the first comprehensive study of format bias in LLMs through a three-stage empirical analysis. The first stage explores the presence and direction of bias across a diverse range of LLMs. The second stage examines how key data-level factors influence these biases. The third stage analyzes how format bias emerges within LLMs' attention patterns and evaluates a lightweight intervention to test its effectiveness. Our results show that format bias is consistent across model families, driven by information richness, structure quality, and representation type, and is closely associated with attention imbalance within the LLMs. Based on these investigations, we identify three future research directions to reduce format bias: enhancing data pre-processing through format repair and normalization, introducing inference-time interventions such as attention re-weighting, and developing format-balanced training corpora. These directions will support the design of more robust and fair heterogeneous data processing systems.

Dengcan Liu, Jiahao Li, Zheren Fu, Yi Tu, Jiajun Li, Zhendong Mao, Yongdong Zhang

Reward models (RMs) are a core component in the post-training of large language models (LLMs), serving as proxies for human preference evaluation and guiding model alignment. However, training reliable RMs under limited resources remains challenging due to the reliance on large-scale preference annotations and the high cost of fine-tuning LLMs. To address this, we propose SparseRM, which leverages Sparse Autoencoder (SAE) to extract preference-relevant information encoded in model representations, enabling the construction of a lightweight and interpretable reward model. SparseRM first employs SAE to decompose LLM representations into interpretable directions that capture preference-relevant features. The representations are then projected onto these directions to compute alignment scores, which quantify the strength of each preference feature in the representations. A simple reward head aggregates these scores to predict preference scores. Experiments on three preference modeling tasks show that SparseRM achieves superior performance over most mainstream RMs while using less than 1% of trainable parameters. Moreover, it integrates seamlessly into downstream alignment pipelines, highlighting its potential for efficient alignment.

Yuanyuan Lin, Xiangyu Ouyang, Teng Zhang, Kaixin Sui

Tree search-based methods have made significant progress in enhancing the code generation capabilities of large language models. However, due to the difficulty in effectively evaluating intermediate algorithmic steps and the inability to locate and timely correct erroneous steps, these methods often generate incorrect code and incur increased computational costs. To tackle these problems, we propose RPM-MCTS, an effective method that utilizes Knowledge-Retrieval as Process Reward Model based on Monte Carlo Tree Search to evaluate intermediate algorithmic steps. By utilizing knowledge base retrieval, RPM-MCTS avoids the complex training of process reward models. During the expansion phase, similarity filtering is employed to remove redundant nodes, ensuring diversity in reasoning paths. Furthermore, our method utilizes sandbox execution feedback to locate erroneous algorithmic steps during generation, enabling timely and targeted corrections. Extensive experiments on four public code generation benchmarks demonstrate that RPM-MCTS outperforms current state-of-the-art methods while achieving an approximately 15% reduction in token consumption. Furthermore, full fine-tuning of the base model using the data constructed by RPM-MCTS significantly enhances its code capabilities.

Yi Lin, Ziyi Zhou, Jiashi Gao, Xinwei Guo, Jiaxin Zhang, Haiyan Wu, Xin Yao, Xuetao Wei

With the rapid deployment of Chinese large language models (LLMs), culturally-grounded bias evaluation remains understudied due to the dominance of English benchmarks and simplistic Chinese scenarios. To address this, we propose GeWu, a comprehensive benchmark featuring a culturally-aware dataset of 60,192 questions spanning 14 social groups with fine-grained Chinese contexts, significantly exceeding existing resources in breadth and depth. Our two-stage evaluation first quantifies bias via multiple-choice questions using a novel probability-based scoring mechanism to sensitively capture bias tendencies, distilling high-bias scenarios into GeWu-1K. This refined subset then enables multi-turn dialogue evaluations for in-depth analysis under realistic conditions. Experiments reveal that GeWu effectively exposes social biases in state-of-the-art Chinese LLMs, with 13.93% of scenarios eliciting universal bias across all models. This highlights persistent challenges and provides actionable insights for bias mitigation in Chinese contexts.

Pengyue Lin, Yanyang Hu, Xinjing Liu, Wenqi Jia, Fangxiang Feng, Ruifan Li

Weakly supervised phrase localization (WSPL) aims to localize visual objects mentioned by given phrases, but it learns without human-annotated bounding boxes. Previous works struggle in multi-object scenarios where objects in the background often appear simultaneously with the target objects. To this end, we propose a Diffusion-Assisted PrOgressive learning framework (i.e., DAPO) for WSPL task in this paper. Specifically, we score the difficulty of training samples based on the quantity of objects and the level of semantic alignment. These samples are then used progressively during training, in an order by their difficulty scores. To address the sample imbalance problem, we propose a Generation-Assisted Tuning (GAT) method for the grounding network. First, to enrich the samples from few-object scenarios, we leverage Stable Diffusion (SD) to generate images with phrases. Second, we introduce an attention-driven scheme to direct SD's attention on the mentioned objects. Finally, we design a diffusion-guided loss, which helps the grounding network learn the objects' layouts. Extensive experiments show that our DAPO framework outperforms the strong baselines on benchmark datasets.

Liang Lin, Miao Yu, Kaiwen Luo, Yibo Zhang, Lilan Peng, Dexian Wang, Xuehai Tang, Yuanhe Zhang, Xikang Yang, Zhenhong Zhou 等

As Audio Large Language Models (ALLMs) emerge as powerful tools for speech processing, their safety implications demand urgent attention. While considerable research has explored textual and vision safety, audio’s distinct characteristics present significant challenges. This paper first investigates: Is ALLM vulnerable to backdoor attacks exploiting acoustic triggers? In response to this issue, we introduce Hidden in the Noise (HIN), a novel backdoor attack framework designed to exploit subtle, audio-specific features. HIN applies acoustic modifications to raw audio waveforms, such as alterations to temporal dynamics and strategic injection of spectrally tailored noise. These changes introduce consistent patterns that an ALLM’s acoustic feature encoder captures, embedding robust triggers within the audio stream. To evaluate ALLM robustness against audio-feature-based triggers, we develop the AudioSafe benchmark, assessing nine distinct risk types. Extensive experiments on AudioSafe and three established safety datasets reveal critical vulnerabilities in existing ALLMs: (I) audio features like environment noise and speech rate variations achieve over 90% average attack success rate, (II) ALLMs exhibit significant sensitivity differences across acoustic features, particularly showing minimal response to volume as a trigger, and (III) poisoned sample inclusion causes only marginal loss curve fluctuations, highlighting the attack’s stealth.

Jiayi Lin, Yanlin Wang, Yibiao Yang, Lei Zhang, Yutao Xie

Recent advances in large-scale code generation models have led to remarkable progress in producing high-quality code. These models are trained in a self-supervised manner on extensive unlabeled code corpora using a decoder-only architecture. However, despite their generative strength, decoder-only models often exhibit limited performance on code understanding tasks such as code search and clone detection, primarily due to their generation-oriented training objectives. While training large encoder-only models from scratch on massive code datasets can improve understanding ability but remains computationally expensive and time-consuming. In this paper, we explore a more efficient alternative by transferring knowledge from pre-trained decoder-only code generation models to code understanding tasks. We investigate how decoder-only architectures can be effectively adapted to learn discriminative and semantically meaningful code representations. To this end, we propose CL4D, a contrastive learning framework tailored to strengthen the representation capabilities of decoder-only models. Extensive experiments on multiple benchmark datasets demonstrate that CL4D achieves competitive or superior performance compared to existing methods on representative code understanding tasks, including code search and clone detection. Further analysis reveals that CL4D substantially improves the semantic alignment of code representations by reducing the distance between semantically similar code snippets. These findings highlight the feasibility of leveraging decoder-only models as a unified backbone for both code generation and understanding.

Hyeongsoo Lim, Hyung Yong Kim, Jin Young Kim, Min Ho Jang, Eun Seo Seo, Youshin Lim, Shukjae Choi, Jihwan Park, Yunkyu Lim, Hanbin Lee 等

Knowledge distillation (KD) is a promising compression technique for reducing the computational burden of large language models (LLMs). Depending on access to the teacher model’s internal parameters, KD is typically categorized into white-box and black-box KD. While white-box KD benefits from full access to intrinsic knowledge such as softmax distributions, black-box KD adopts a black-box LLM (e.g., GPT-4) as the teacher, which provides only text-level outputs via API calls. This limited supervision makes black-box KD generally less effective than its white-box counterpart. To bridge the gap between white-box and black-box KD, we propose GrayKD, a novel framework that can effectively distill text-level knowledge from a black-box LLM in a single-stage manner. In particular, rationales generated by the black-box LLM are injected into the student via a lightweight cross-attention module (teacher mode), enabling the model to approximate the black-box teacher’s output distribution without access to internal parameters. The student is then trained with the softmax-level knowledge provided by the teacher mode (student mode). Since both the teacher and student modes share the same backbone, the proposed teacher mode remains highly parameter-efficient, requiring only a small number of additional parameters for rationale injection. Experimental results on instruction-following tasks demonstrate that GrayKD achieves substantial performance improvements over existing KD methods.

Wenjie Liao, Xiaohui Song, Haonan Lu

Self-play fine-tuning has emerged as a promising approach to improve Large Language Models (LLMs) without additional human annotations. However, existing methods struggle with complex generation tasks requiring long context understanding, where models produce partially correct outputs interleaved with errors. Traditional approaches train on entire sequences uniformly, failing to distinguish between well-predicted and erroneous regions, leading to diluted learning signals and slow convergence. We propose DRIFT (Difference-aware Reinforcement through Iterative Fine-Tuning), a novel self-play framework that selectively trains on prediction differences. DRIFT introduces two key innovations: (1) Difference-Aware Masking (DAM) that identifies and masks common subsequences between model outputs and ground truth, focusing training exclusively on error regions; (2) Occurrence-Aware Loss (OAL) that provides position-invariant vocabulary supervision, complementing the position-sensitive adversarial loss. This dual mechanism enables models to correct both positional and lexical errors effectively. Theoretically, we prove that DRIFT converges when masked distributions align. Empirically, we evaluate DRIFT on diverse summarization benchmarks using Qwen2.5-3B and LLaMA-3.1-8B models. Results show that DRIFT significantly outperforms both supervised fine-tuning (SFT) and self-play fine-tuning (SPIN), achieving up to 16\% improvement on SAMSum dialogue summarization tasks while maintaining general capabilities. Notably, DRIFT breaks the performance ceiling of continued SFT and demonstrates superior efficiency compared to holistic self-play methods, validating that targeted optimization on prediction differences is crucial for structured text generation tasks.

Kewei Liao, Tianbo Wang, Yuqing Ma, Zhange Zhang, Zhicheng Geng, Xiaowei Zhao, Jiakai Wang, Xianglong Liu

Hallucination has emerged as a pivotal challenge of Large Language Models (LLMs) that generate plausible yet non‑factual content, significantly impeding the trustworthy AI applications in real-world scenarios like medical diagnosis and autonomous driving. Editing the internal activations of LLMs during inference has shown promising effectiveness in mitigating hallucinations with minimal cost. However, previous editing approaches neglect the query‑specific inference pathways that require tailored truthful steering vectors, resulting in suboptimal hallucination mitigation. To address these issues, we propose the Query-Routed Activation Editing (QRAE) framework, which comprises Divergence-sensitive Head Routing (DHR) and Truth-hierarchical Preference Steering (TPS), to fully leverage query-specific semantics for adaptive activation editing. Specifically, DHR is proposed to establish a query-aware head selection criterion, thereby dynamically routing to truth-critical attention heads. Subsequently, TPS introduces a query-specific steering vector calibration policy with the guidance of progressive truth-preferred optimization, enabling precise and adaptive editing for each distinct query. Extensive experiments on the widely recognized TruthfulQA benchmark demonstrate that QRAE outperforms SOTA methods by up to 13.2% in MC1. Meanwhile, QRAE demonstrates strong generalization to out-of-distribution TriviaQA and Natural Questions benchmarks.