论文检索

Chengyu Jiao, Siyin Huang, Yu Zhang

Several studies have demonstrated that large language models (LLMs) exhibit positional bias when answering multiple-choice questions (MCQs). Previous methods have identified such bias to be detrimental, leading to the development of techniques to mitigate it. However, we observe that certain permutations of options can actually improve the performance. Therefore, instead of eliminating such bias, we propose an EMbracing the Bias EquivaRiantly (EMBER) network. Specifically, the EMBER network, which outputs a permutation of options in MCQs, is optimized towards the beneficial permutations to which the LLM is biased. Additionally, to solve the positional bias among different permutations of options, the EMBER network is designed to grant the equivariance to the permutation to the LLMs. Theoretically and empirically, we show that the proposed EMBER network can effectively utilize the positional bias and demonstrate state-of-the-art performance over various baselines.

Zishang Jiang, Jinyi Han, Tingyun Li, Xinyi Wang, Sihang Jiang, Xiaojun Meng, Jiansheng Wei, Jiaqing Liang, Yanghua Xiao

Fine-tuning plays an essential role in improving the performance of large language models (LLMs) on specific tasks. A central challenge lies in designing data-efficient strategy to achieve better fine-tuning performance. Curriculum learning, which organizes data from easy to hard, has become a widely adopted technique in LLMs training. However, existing methods for curriculum learning focus only on the difficulty of samples, while neglecting their contribution to improving model performance, making them vulnerable when applied to fine-tuning LLMs. To address this, we propose Difficulty-Utility Curriculum Learning (DUCL), a curriculum learning framework that jointly considers difficulty and utility. DUCL introduces a novel scoring method, Difficulty-Utility Evaluation (DUE), and a soft scheduling strategy called Window Ordering, which together promote efficient and effective fine-tuning. Our method not only improves convergence and final performance with negligible computational overhead, but is also broadly applicable across a wide range of tasks, making it a practical and scalable solution for LLMs fine-tuning.

Yilin Jiang, Mingzi Zhang, Xuanyu Yin, Sheng Jin, Suyu Lu, Zuocan Ying, Zengyi Yu, Xiangjie Kong

Large Language Models for Simulating Professions (SP-LLMs), particularly as teachers, are pivotal for personalized education. However, ensuring their professional competence and ethical safety remains a major challenge, as existing benchmarks fail to measure role-playing fidelity or address the unique teaching harms inherent in educational scenarios. To address this gap, we propose EduGuardBench, a dual-component benchmark that evaluates professional fidelity through the Role-playing Fidelity Score (RFS) and diagnoses harms specific to the teaching profession. It also probes safety vulnerabilities using persona-based adversarial prompts targeting both general harms and academic misconduct, with metrics such as Attack Success Rate (ASR) and a three-tier Refusal Quality assessment. Extensive experiments on 14 leading models reveal a stark polarization in performance. While reasoning-oriented models generally demonstrate higher fidelity, incompetence remains the dominant failure mode across most models. Adversarial testing uncovered a counterintuitive scaling paradox, where mid-sized models appear more vulnerable, challenging monotonic safety assumptions. Notably, we identify an Educational Transformation Effect, where the safest models convert harmful requests into teachable moments through ideal educational refusals. This ability is strongly negatively correlated with ASR, revealing a new dimension of advanced AI safety. EduGuardBench thus provides a reproducible framework for holistic assessment of professional, ethical, and pedagogical alignment, uncovering dynamics critical to deploying trustworthy AI in education.

Xue Jiang, Yihong Dong, Huangzhao Zhang, Tangxinyu Wang, Zheng Fang, Yingwei Ma, Rongyu Cao, Binhua Li, Zhi Jin, Wenpin Jiao 等

While Large Language Models (LLMs) excel at code generation, their inherent tendency toward verbatim memorization of training data introduces critical risks like copyright infringement, insecurity emission, and deprecated API utilization, etc. A straightforward yet promising defense is unlearning, i.e., erasing or down-weighting the offending snippets through post-training. However, we find its application to source code often tends to spill over, damaging the basic knowledge of programming languages learned by the LLM and degrading the overall capability. To ease this challenge, we propose PROD for precise source code unlearning. PROD surgically zeroes out the prediction probability of the prohibited tokens, and renormalizes the remaining distribution so that the generated code stays correct. By excising only the targeted snippets, PROD achieves precise forgetting without much degradation of the LLM's overall capability. To facilitate in-depth evaluation against PROD, we establish an unlearning benchmark consisting of three downstream tasks (i.e., unlearning of copyrighted code, insecure code, and deprecated APIs), and introduce Pareto Dominance Ratio (PDR) metric, which indicates both the forget quality and the LLM utility. Our comprehensive evaluation demonstrates that PROD achieves superior overall performance between forget quality and model utility compared to existing unlearning approaches across three downstream tasks, while consistently exhibiting improvements when applied to LLMs of varying series. PROD also exhibits superior robustness against adversarial attacks without generating or exposing the data to be forgotten. These results underscore that our approach not only successfully extends the application boundary of unlearning techniques to source code, but also holds significant implications for advancing reliable code generation.

Weipeng Jiang, Juan Zhai, Shiqing Ma, Ziyan Lei, Xiaofei Xie, Yige Wang, Chao Shen

Recently LLMs have faced increasing demands to selectively remove specific information through Machine Unlearning. While evaluating unlearning effectiveness is crucial, existing benchmarks suffer from fundamental limitations in audit dataset generation from unstructured corpora. We identify two critical challenges: ensuring audit adequacy and handling knowledge redundancy between forget and retain datasets. Current approaches rely on ad-hoc question generation from unstructured text, leading to unpredictable coverage gaps and evaluation blind spots. Knowledge redundancy between forget and retain corpora further obscures evaluation, making it difficult to distinguish genuine unlearning failures from legitimately retained knowledge. To bring clarity to this challenge, we propose LUCID, an automated framework that leverages knowledge graphs to achieve comprehensive audit dataset generation with fine-grained coverage and systematic redundancy elimination. By converting unstructured corpora into structured knowledge representations, it transforms the ad-hoc audit dataset generation process into a transparent and automated generation pipeline that ensures both adequacy and non-redundancy. Applying LUCID to the MUSE benchmark, we generated over 69,000 and 111,000 audit cases for News and Books datasets respectively, identifying thousands of previously undetected knowledge memorization instances. Our analysis reveals that knowledge redundancy significantly skews metrics, artificially inflating ROUGE from 19.7% to 26.1% and Entailment Scores from 32.4% to 35.2%, highlighting the necessity of deduplication for accurate assessment.

Tingyu Jiang, Shen Li, Yiyao Song, Lan Zhang, Hualei Zhu, Yuan Zhao, Xiaohang Xu, Kenjiro Taura, Hao Henry Wang

Instruction tuning plays a critical role in enhancing the performance and efficiency of Large Language Models (LLMs). Its success depends not only on the quality of the instruction data but also on the inherent capabilities of the LLM itself. Some studies suggest that even a small amount of high-quality data can achieve instruction fine-tuning results that are on par with, or even exceed, those from using a full-scale dataset. However, rather than focusing solely on calculating data quality scores to evaluate instruction data, there is a growing need to select high-quality data that maximally enhances the performance of instruction tuning for a given LLM. In this paper, we propose the Model Instruction Weakness Value (MIWV) as a novel metric to quantify the importance of instruction data in enhancing model's capabilities. The MIWV metric is derived from the discrepancies in the model’s responses when using In-Context Learning (ICL), helping identify the most beneficial data for enhancing instruction tuning performance. Our experimental results demonstrate that selecting only the top 1% of data based on MIWV can outperform training on the full dataset. Furthermore, this approach extends beyond existing research that focuses on data quality scoring for data selection, offering strong empirical evidence supporting the effectiveness of our proposed method.

Shuyang Jiang, Yusheng Liao, Zhe Chen, Ya Zhang, Yanfeng Wang, Yu Wang

Medical language models face critical barriers to real-world clinical reasoning applications. However, mainstream efforts, which fall short in task coverage, lack fine-grained supervision for intermediate reasoning steps, and rely on proprietary systems, are still far from a versatile, credible and efficient language model for clinical reasoning usage. To this end, we propose MedS3, a self-evolving framework that imparts robust reasoning capabilities to small, deployable models. Starting with 8,000 curated instances sampled via a curriculum strategy across five medical domains and 16 datasets, we use a small base policy model to conduct Monte Carlo Tree Search (MCTS) for constructing rule-verifiable reasoning trajectories. Self-explored reasoning trajectories ranked by node values are used to bootstrap the policy model via reinforcement fine-tuning and preference learning. Moreover, we introduce a soft dual process reward model that incorporates value dynamics: steps that degrade node value are penalized, enabling fine-grained identification of reasoning errors even when the final answer is correct. Experiments on eleven benchmarks show that MedS3 outperforms the previous state-of-the-art medical model by +6.45 accuracy points and surpasses 32B-scale general-purpose reasoning models by +8.57 points. Additional empirical analysis further demonstrates that MedS3 achieves robust and faithful reasoning behavior.

Shuting Jiang, Ran Song, Yuxin Huang, Yan Xiang, Yantuan Xian, Shengxiang Gao, Zhengtao Yu

Multi-domain machine translation (MDMT) aims to build a unified model capable of translating content across diverse domains. Despite the impressive machine translation capabilities demonstrated by large language models (LLMs), domain adaptation still remains a challenge for LLMs. Existing MDMT methods such as in-context learning and parameter-efficient fine-tuning often suffer from domain shift, parameter interference and limited generalization. In this work, we propose a neuron-efficient fine-tuning framework for MDMT that identifies and updates consensus-aligned neurons within LLMs. These neurons are selected by maximizing the mutual information between neuron behavior and domain features, enabling LLMs to capture both generalizable translation patterns and domain-specific nuances. Our method then fine-tunes LLMs guided by these neurons, effectively mitigating parameter interference and domain-specific overfitting. Comprehensive experiments on three LLMs across ten German-English and Chinese-English translation domains evidence that our method consistently outperforms strong PEFT baselines on both seen and unseen domains, achieving state-of-the-art performance.

Jiyue Jiang, Yanyu Chen, Pengan Chen, Kai Liu, Jingqi Zhou, Zheyong Zhu, He Hu, Fei Ma, Qi Tian, Chuan Wu

Cognitive impairment is becoming a major public health challenge. Cognitive Stimulation Therapy (CST) is an effective intervention for cognitive impairment, but traditional methods are difficult to scale, and existing digital systems struggle with group dialogues and cognitive stimulation principles. While Large Language Models (LLMs) are powerful, their application in this context faces key challenges: cognitive stimulation dialogue paradigms, a lack of therapeutic reasoning, and static-only user modeling. To address these issues, we propose a principle-driven adaptive policy actualized through a Group Cognitive Stimulation Dialogue (GCSD) system. We first construct a dataset with over 500 hours of real-world CST conversations and 10,000+ simulated dialogues generated via our Principle-Guided Scenario Simulation strategy. Our GCSD system then integrates four core modules to overcome LLM limitations: (i) a multi-speaker context controller to resolve role confusion; (ii) dynamic participant cognitive state modeling for personalized interaction; (iii) a cognitive stimulation-focused attention loss to instill cognitive stimulation reasoning; and (iv) a multi-dimensional reward strategy to enhance response value. Experimental results demonstrate that GCSD significantly outperforms baseline models across various evaluation metrics. Future work will focus on long-term clinical validation to bridge the gap between computational performance and clinical efficacy.

Zhongquan Jian, Wenhan Lv, Yanhao Chen, Guanran Luo, Wentao Qiu, Shaopan Wang, Bingbing Hu, Qingqiang Wu

The homogeneity and heterogeneity across modalities are critical factors that influence multimodal fusion. In Multimodal Sentiment Analysis (MSA), the inherent textual information within the audio modality induces cross-modality homogeneity with the text modality. Conversely, the mutual independence between text and vision modalities results in their cross-modal heterogeneity. Although existing disentangle-based methods achieve notable performance gains by separating modality features into distinct subspaces, they overlook the characteristics of cross-modality heterogeneity and homogeneity among different modalities. To this end, we propose a novel Modality-aware Disentangle and Fusion (MDF) framework to investigate the role of core modality features. Specifically, we first use text as the anchor to disentangle the audio modality and extract its unique modality-specific features, thereby establishing cross-modal heterogeneity among text, audio, and vision. We then introduce a Cross-Modality Heterogeneity Enhancement (CHE) module to refine these features, further reinforcing their heterogeneous characteristics. Finally, a Modality Adaptive Weighting (MAW) module is employed to dynamically assign weights to the text, sound, and vision modalities based on their potential contributions to sentiment prediction, achieving a more effective multimodal representation for MSA. Experimental evaluations on different benchmarks demonstrate MDF's superiority, with extensive ablation studies confirming its effectiveness.

Weitao Jia, Jinghui Lu, Haiyang Yu, Siqi Wang, Guozhi Tang, An-Lan Wang, Weijie Yin, Dingkang Yang, Yuxiang Nie, Bin Shan 等

Recent advances demonstrate that reinforcement learning with verifiable rewards (RLVR) significantly enhances the reasoning capabilities of large language models (LLMs). However, standard RLVR faces challenges with reward sparsity, where zero rewards from consistently incorrect candidate answers provide no learning signal, particularly in challenging tasks. To address this,we propose Multi-Expert Mutual Learning GRPO (MEML-GRPO), an innovative framework that utilizes diverse expert prompts as system prompts to generate a broader range of responses, substantially increasing the likelihood of identifying correct solutions. Additionally, we introduce an inter-expert mutual learning mechanism that facilitates knowledge sharing and transfer among experts, further boosting the model’s performance through RLVR. Extensive experiments across multiple reasoning benchmarks show that MEML-GRPO delivers significant improvements, achieving an average performance gain of 4.89% with Qwen and 11.33% with Llama, effectively overcoming the core limitations of traditional RLVR methods.

Junlong Jia, Xing Wu, Chaochen Gao, Ziyang Chen, Zijia Lin, Zhongzhi Li, Weinong Wang, Haotian Xu, Donghui Jin, Debing Zhang 等

High-quality long-context data is essential for training large language models (LLMs) capable of processing extensive documents, yet existing synthesis approaches using relevance-based aggregation face challenges of computational efficiency. We present LiteLong, a resource-efficient method for synthesizing long-context data through structured topic organization and multi-agent debate. Our approach leverages the BISAC book classification system to provide a comprehensive hierarchical topic organization, and then employs a debate mechanism with multiple LLMs to generate diverse, high-quality topics within this structure. For each topic, we use lightweight BM25 retrieval to obtain relevant documents and concatenate them into 128K-token training samples. Experiments on HELMET and Ruler benchmarks demonstrate that LiteLong achieves competitive long-context performance and can seamlessly integrate with other long-dependency enhancement methods. LiteLong makes high-quality long-context data synthesis more accessible by reducing both computational and data engineering costs, facilitating further research in long-context language training.

Jingyi Jia, Qinbin Li

Large Language Model (LLM) agents have emerged as powerful tools for automating complex tasks by leveraging the reasoning and decision-making abilities of LLMs. However, a major bottleneck in current agent frameworks lies in the high inference cost of tool selection, especially in approaches like ReAct that repeatedly invoke the LLM to determine which tool to use at each step. In this work, we propose AutoTool, a novel graph-based framework that bypasses repeated LLM inference by exploiting a key empirical observation: tool usage inertia—the tendency of tool invocations to follow predictable sequential patterns. AutoTool constructs a directed graph from historical agent trajectories, where nodes represent tools and edges capture transition probabilities, effectively modeling the inertia in tool selection. It further integrates parameter-level information to refine tool input generation. By traversing this structured representation, AutoTool efficiently selects tools and their parameters with minimal reliance on LLM inference. Extensive experiments across diverse agent tasks demonstrate that AutoTool reduces inference costs by up to 30% while maintaining competitive task completion rates, offering a practical and scalable enhancement for inference-heavy frameworks. Our work highlights the promise of integrating statistical structure into LLM agent design for greater efficiency without sacrificing performance.

Chen Jia

Bootstrapping large language models (LLMs) via preference-based policy optimization enables aligning model behavior with human preferences while reducing reliance on extensive manual annotations. We propose a novel preference-based policy optimization (PbPO) framework that formulates learning as a min-max game between the LLM policy and a reward model (RM). The RM is constrained within a confidence set derived from collected preferences to ensure reliable exploitation, while simultaneously promoting robust exploration. Our iterative online algorithm actively collects new preference data from the evolving policy, enabling continual self-improvement of both the policy and the RM. We provide theoretical guarantees, establishing high-probability regret bounds for both sequence-level and token-level RMs. Extensive experiments across five benchmark datasets demonstrate that PbPO consistently outperforms state-of-the-art preference optimization methods.

Jinhong Jeong, Sunghyun Lee, Jaeyoung Lee, Seonah Han, Youngjae Yu

Sound symbolism is a linguistic concept that refers to non-arbitrary associations between phonetic forms and their meanings. We suggest that this can be a compelling probe into how Multimodal Large Language Models (MLLMs) interpret auditory information in human languages. We investigate MLLMs' performance on phonetic iconicity across textual (orthographic and IPA) and auditory forms of inputs with up to 25 semantic dimensions (e.g., sharp vs. round), observing models' layer-wise information processing by measuring phoneme-level attention fraction scores. To this end, we present LEX-ICON, an extensive mimetic word dataset consisting of 8,052 words from four natural languages (English, French, Japanese, and Korean) and 2,930 systematically constructed pseudo-words, annotated with semantic features applied across both text and audio modalities. Our key findings demonstrate (1) MLLMs' phonetic intuitions that align with existing linguistic research across multiple semantic dimensions and (2) phonosemantic attention patterns that highlight models' focus on iconic phonemes. These results bridge domains of artificial intelligence and cognitive linguistics, providing the first large-scale, quantitative analyses of phonetic iconicity in terms of MLLMs' interpretability.

Min Ho Jang, Eun Seo Seo, Jin Young Kim, Hyeongsoo Lim, Ji Won Yoon

Recursive transformer (RT) is a promising parameter-sharing technique for reducing computational burden of large-scale model. While RT has been successfully applied to large language models (LLMs), its effectiveness in automatic speech recognition (ASR) remains limited, despite the parallel trend of model scaling in the speech domain. In this paper, we reveal that conventional RT designs for LLMs are suboptimal for speech recognition, primarily because they do not fully consider the layer-wise specialization inherent in the ASR architecture, where lower layers focus on phonetic features and upper layers capture linguistic localization. To address this, we propose BiCycle, a novel RT scheme tailored for ASR. In particular, we firstly analyze attention patterns in a pre-trained ASR model to divide its layers into phonetic and linguistic groups. BiCycle then constructs an efficient RT model by transferring the pre-trained model’s weights in a step-wise manner and applies recursion separately to the phonetic and linguistic groups, preventing conflicts between their roles. Extensive experimental results confirm that the proposed method not only preserves the original ASR mechanism but also outperforms conventional RT approaches.

Zhaopei Huang, Qifeng Dai, Guozheng Wu, Xiaopeng Wu, Xubin Li, Tiezheng Ge, Wenxuan Wang, Qin Jin

With the rise of smart personal devices, service-oriented human-agent interactions have become increasingly prevalent. This trend highlights the need for personalized dialogue assistants that can understand user-specific traits to accurately interpret requirements and tailor responses to individual preferences. However, existing approaches often overlook the complexities of long-term interactions and fail to capture users’ subjective characteristics. To address these gaps, we present PAL-Bench, a new benchmark designed to evaluate the personalization capabilities of service-oriented assistants in long-term user-agent interactions. In the absence of available real-world data, we develop a multi-step LLM-based synthesis pipeline, which is further verified and refined by human annotators. This process yields PAL-Set, the first Chinese dataset comprising multi-session user logs and dialogue histories, which serves as the foundation for PAL-Bench. Furthermore, to improve personalized service-oriented interactions, we propose H2Memory, a hierarchical and heterogeneous memory framework that incorporates retrieval-augmented generation to improve personalized response generation. Comprehensive experiments on both our PAL-Bench and an external dataset demonstrate the effectiveness of the proposed memory framework.

Yue Huang, Xiangqi Wang, Xiangliang Zhang

In high-stakes scenarios—such as self-harm, legal, or medical queries—LLMs must be both trustworthy and helpful. However, these goals often conflict. We propose priority alignment, a new alignment paradigm that enforces a strict “trustworthy-before-helpful” ordering: optimization of helpfulness is conditioned on first meeting trustworthy thresholds (e.g., harmlessness or honesty). To realize this, we introduce Self-Priority Alignment (SPA)—a fully unsupervised framework that generates diverse responses, self-evaluates them and refines them by the model itself, and applies dual-criterion denoising to remove inconsistency and control variance. From this, SPA constructs lexicographically ordered preference pairs and fine-tunes the model using an uncertainty-weighted alignment loss that emphasizes high-confidence, high-gap decisions. Experiments across multiple benchmarks show that SPA improves helpfulness without compromising safety, outperforming strong baselines while preserving general capabilities. Our results demonstrate that SPA provides a scalable and interpretable alignment strategy for critical LLM applications.

Yitong Huang, Ziqi Yang, Zihui Wang, Jianzhong Qi, Rongshan Yu, Xiaoliang Fan, Cheng Wang

Combining Mixture of Experts (MoE) with Low-Rank Adaptation (LoRA) has shown promising efficiency in multi-task instruction tuning for Large Language Models (LLMs). While existing routing schemes for such MoE systems employ auxiliary functions to ensure both expert selection certainty and workload balance among experts, they are hindered by two critical challenges: (1) Existing methods overlook the evolving cross-expert relationships across layers, leading to inefficient expert utilization. (2) The auxiliary functions fail to incorporate cross-task semantic characteristics during expert assignment, leading to suboptimal task adaptation. To address these challenges, we propose Hybrid routing for a Mixture of LoRA Experts (HotMoE), a novel multi-task instruction tuning framework that adapts hierarchical routing to the distinct characteristics of different LLM layers. First, we design a hybrid routing module. In lower layers, expert-expert attention facilitates cross-task collaboration and generalization. In higher layers, token-expert attention enables precise alignment between task semantics and specialized experts. Second, we introduce a similarity-guided auxiliary loss module to regularize routing decisions by exploiting hidden state similarities. This loss synergistically reinforces expert specialization without sacrificing certainty of expert selection by promoting cohesive activation patterns among semantically related tasks while sharpening distinctions between conflicting ones. Experiments across two multi-task instruction tuning scenarios covering seven NLP benchmarks demonstrate that HotMoE consistently outperforms all baselines, improving Mean Relative Difference by up to 1.68% with only 3.1% of trainable parameters.

Manzong Huang, Chenyang Bu, Yi He, Xingrui Zhuo, Xindong Wu

Graph-based Retrieval-Augmented Generation (GraphRAG) mitigates hallucinations in Large Language Models (LLMs) by grounding them in structured knowledge. However, current GraphRAG methods are constrained by a prevailing build-then-reason paradigm, which relies on a static, pre-constructed Knowledge Graph (KG). This paradigm faces two critical challenges. First, the KG's inherent incompleteness often breaks reasoning paths. Second, the graph’s low signal-to-noise ratio introduces distractor facts, presenting query-relevant but misleading knowledge that disrupts the reasoning process. To address these challenges, we argue for a reason-and-construct paradigm and propose Relink, a framework that dynamically builds a query-specific evidence graph. To tackle incompleteness, Relink instantiates required facts from a latent relation pool derived from the original text corpus, repairing broken paths on the fly. To handle misleading or distractor facts, Relink employs a unified, query-aware evaluation strategy that jointly considers candidates from both the KG and latent relations, selecting those most useful for answering the query rather than relying on their pre-existence. This empowers Relink to actively discard distractor facts and construct the most faithful and precise evidence path for each query. Extensive experiments on five Open-Domain Question Answering benchmarks show that Relink achieves significant average improvements of 5.4% in EM and 5.2% in F1 over leading GraphRAG baselines, demonstrating the superiority of our proposed framework.