论文检索

Xu Sun, Lionel Delphin-Poulat, Christèle Tarnec, Anastasia Shimorina

Large language models (LLMs) are increasingly used for zero-shot conversation summarization, but often exhibit positional bias—tending to overemphasize content from the beginning or end of a conversation while neglecting the middle. To address this issue, we introduce PoSum-Bench, a comprehensive benchmark for evaluating positional bias in conversational summarization, featuring diverse English and French conversational datasets spanning formal meetings, casual conversations, and customer service interactions. We propose a novel semantic similarity-based sentence-level metric to quantify the direction and magnitude of positional bias in model-generated summaries, enabling systematic and reference-free evaluation across conversation positions, languages, and conversational contexts.Our benchmark and methodology thus provide the first systematic, cross-lingual framework for reference-free evaluation of positional bias in conversational summarization, laying the groundwork for developing more balanced and unbiased summarization models.

Johannes Moll, Louisa Fay, Asfandyar Azhar, Sophie Ostmeier, Sergios Gatidis, Tim Lueth, Curtis P. Langlotz, Jean-Benoit Delbrouck

Radiology reports are critical for clinical decision-making but often lack a standardized format, limiting both human interpretability and machine learning (ML) applications. While large language models (LLMs) have shown strong capabilities in reformatting clinical text, their high computational requirements, lack of transparency, and data privacy concerns hinder practical deployment. To address these challenges, we explore lightweight encoder-decoder models (<300M parameters)—specifically T5 and BERT2BERT—for structuring radiology reports from the MIMIC-CXR and CheXpert Plus datasets. We benchmark these models against eight open-source LLMs (1B–70B parameters), adapted using prefix prompting, in-context learning (ICL), and low-rank adaptation (LoRA) finetuning. Our best-performing lightweight model outperforms all LLMs adapted using prompt-based techniques on a human-annotated test set. While some LoRA-finetuned LLMs achieve modest gains over the lightweight model on the Findings section (BLEU 6.4%, ROUGE-L 4.8%, BERTScore 3.6%, F1-RadGraph 1.1%, GREEN 3.6%, and F1-SRR-BERT 4.3%), these improvements come at the cost of substantially greater computational resources. For example, LLaMA-3-70B incurred more than 400 times the inference time, cost, and carbon emissions compared to the lightweight model. These results underscore the potential of lightweight, task-specific models as sustainable and privacy-preserving solutions for structuring clinical text in resource-constrained healthcare settings.

Songze Li, Zhiqiang Liu, Zhengke Gui, Huajun Chen, Wen Zhang

Large Language Models (LLMs) exhibit strong reasoning capabilities in complex tasks. However, they still struggle with hallucinations and factual errors in knowledge-intensive scenarios like knowledge graph question answering (KGQA). We attribute this to the semantic gap between structured knowledge graphs (KGs) and unstructured queries, caused by inherent differences in their focuses and structures. Existing methods usually employ resource-intensive, non-scalable workflows reasoning on vanilla KGs, but overlook this gap. To address this challenge, we propose a flexible framework, Enrich-on-Graph (EoG), which leverages LLMs’ prior knowledge to enrich KGs, bridge the semantic gap between graphs and queries. EoG enables efficient evidence extraction from KGs for precise and robust reasoning, while ensuring low computational costs, scalability, and adaptability across different methods. Furthermore, we propose three graph quality evaluation metrics to analyze query-graph alignment in KGQA task, supported by theoretical validation of our optimization objectives. Extensive experiments on two KGQA benchmark datasets indicate that EoG can effectively generate high-quality KGs and achieve the state-of-the-art performance.

Jielong Tang, Yang Yang, Jianxing Yu, Zhen-Xing Wang, Haoyuan Liang, Liang Yao, Jian Yin

Grounded Multimodal Named Entity Recognition (GMNER) is a new information extraction task. It requires models to extract named entities and ground them to real-world visual objects. Previous methods, relying on domain-specific fine-tuning, struggle with unseen multimodal entities due to limited knowledge and generalization. Recently, multimodal large language models (MLLMs) have demonstrated strong open-set abilities. However, their performance is hindered by the lack of in-domain knowledge due to costly training for GMNER datasets. To address these limitations, we propose **UnCo**, a two-stage Uncertainty-driven Collaborative framework that leverages the complementary strengths of small fine-tuned models and MLLMs. Specifically, **in stage one**, we equip the small model with a unified uncertainty estimation (UE) for multimodal entities. This enables the small model to express \textit{"I do not know"} when recognizing unseen entities beyond its capabilities. Predictions with high uncertainty are then filtered and delegated to the MLLM. **In stage two**, an Uncertainty-aware Hierarchical Correction mechanism guides the MLLM to refine uncertain predictions using its open-domain knowledge. Ultimately, UnCo effectively retains the in-domain knowledge of small models while utilizing the capabilities of MLLMs to handle unseen samples. Extensive experiments demonstrate UnCo’s effectiveness on two GMNER benchmarks.

Heng-Da Xu, Xian-Ling Mao, Fanshu Sun, Tian-Yi Che, Cheng-Xin Xin, Heyan Huang

High-quality datasets are essential for building effective task-oriented dialogue (TOD) systems. The existing TOD datasets often present overly simplified interactions, where users incrementally express straightforward requests that can be managed with basic slot-value style dialogue states, such as “hotel-area = east.” However, this approach does not reflect real-life scenarios in which users may express complex constraints and preferences. To address this gap, in this paper, we propose SQLWOZ, a novel TOD dataset designed to capture complex, real-world user requirements. The user requirements in SQLWOZ include the four categories: 1) multiple values for a slot, 2) excluded values within a slot, 3) preferred or prioritized values, and 4) conditional values based on other conditions. We utilize SQL statements as a formalized and expressive representation of dialogue states within SQLWOZ. To evaluate the dataset, we adapt large language models as dialogue agents and conduct extensive experiments on the SQL-based dialogue state tracking, dialogue response generation and end-to-end TOD tasks. The experimental results demonstrate the complexity and quality of SQLWOZ, establishing it as a new benchmark for advancing TOD research.

Yingli Shen, Wen Lai, Shuo Wang, Ge Gao, Kangyang Luo, Alexander Fraser, Maosong Sun

Continued pretraining and instruction tuning on large-scale multilingual data have proven to be effective in scaling large language models (LLMs) to low-resource languages. However, the unaligned nature of such data limits its ability to effectively capture cross-lingual semantics. In contrast, multi-way parallel data, where identical content is aligned across multiple languages, provides stronger cross-lingual consistency and offers greater potential for improving multilingual performance. In this paper, we introduce a large-scale, high-quality multi-way parallel corpus, TED2025, based on TED Talks. The corpus spans 113 languages, with up to 50 languages aligned in parallel, ensuring extensive multilingual coverage. Using this dataset, we investigate best practices for leveraging multi-way parallel data to enhance LLMs, including strategies for continued pretraining, instruction tuning, and the analysis of key influencing factors. Experiments on six multilingual benchmarks show that models trained on multi-way parallel data consistently outperform those trained on unaligned multilingual data.

Zihan Wang, Zihan Liang, Zhou Shao, Yufei Ma, Huangyu Dai, Ben Chen, Lingtao Mao, Chenyi Lei, Yuqing Ding, Han Li

Retrieval-Augmented Generation (RAG) has emerged as a promising approach to address key limitations of Large Language Models (LLMs), such as hallucination, outdated knowledge, and lacking reliable reference. However, current RAG frameworks often struggle with identifying whether retrieved documents meaningfully contribute to answer generation. This shortcoming makes it difficult to filter out irrelevant or even misleading content, which notably impacts the final performance. In this paper, we propose Document Information Gain (DIG), a novel metric designed to quantify the contribution of retrieved documents to correct answer generation. DIG measures a document’s value by computing the difference of LLM’s generation confidence with and without the document augmented. Further, we introduce InfoGain-RAG, a framework that leverages DIG scores to train a specialized reranker, which prioritizes each retrieved document from exact distinguishing and accurate sorting perspectives. This approach can effectively filter out irrelevant documents and select the most valuable ones for better answer generation. Extensive experiments across various models and benchmarks demonstrate that InfoGain-RAG can significantly outperform existing approaches, on both single and multiple retrieval paradigm. Specifically on NaturalQA, it achieves the improvements of 17.9%, 4.5%, 12.5% in exact match accuracy against naive RAG, self-reflective RAG and modern ranking-based RAG respectively, and even an average of 15.3% increment on advanced proprietary models GPT-4o across all datasets. These results demonstrate the feasibility of InfoGain-RAG as it can offer a reliable solution for RAG in multiple applications.

Seonil Son, Ju-Min Oh, Heegon Jin, Cheolhun Jang, Jeongbeom Jeong, KunTae Kim

As Large Language Models (LLMs) expand across domains, LLM judges have become essential for systems evaluation. Current benchmarks typically compare system outputs against baselines.This baseline-mediated approach, though convenient, yields lower reliability than direct comparison between systems.We propose Arena-Lite which integrates tournament structure on top of head-to-head comparison.The application of a tournament structure and direct comparison eliminates the need for baseline outputs, reduces the number of required comparisons, and allows higher reliability in system rankings.We conducted two experiments: (1) controlled stochastic modeling and (2) empirical validation with a real LLM judge. Those experiments collectively demonstrate that Arena-Lite consistently achieves higher reliability with fewer comparisons, even with smaller datasets or weaker judges.We release an easy-to-use web demonstration and code to foster adoption of Arena-Lite, streamlining model selection across research and industry communities. Arena-Lite demo and code are available on https://huggingface.co/spaces/NCSOFT/ArenaLite

Kailai Yang, Zhiwei Liu, Qianqian Xie, Jimin Huang, Erxue Min, Sophia Ananiadou

Recent advancements in LLM alignment leverage token-level supervisions to perform fine-grained preference optimization. However, existing token-level alignment methods either optimize on all available tokens, which can be noisy and inefficient, or perform selective training with complex and expensive key token selection strategies. In this work, we propose Selective Preference Optimization (SePO), a novel selective alignment strategy that centers on efficient key token selection without requiring strong, fine-grained supervision signals. We theoretically prove the feasibility of Direct Preference Optimization (DPO) as token-level reward function estimators, which applies to any existing alignment datasets and enables cost-efficient token selection with small-scale model sizes and training data. We then train an oracle model with DPO on the target data and utilize the estimated reward function to score all tokens within the target dataset, where only the key tokens are selected to supervise the target policy model with a contrastive objective function. Extensive experiments on three public evaluation benchmarks show that SePO significantly outperforms competitive baseline methods by only optimizing on 30% key tokens with up to 60% reduction in GPU training hours. We also explore SePO as a new paradigm for weak-to-strong generalization, showing that weak oracle models effectively supervise strong policy models with up to 16.8 more parameters. SePO also selects useful supervision signals from out-of-distribution data, alleviating the over-optimization problem.

Zinan Tang, Xin Gao, Qizhi Pei, Zhuoshi Pan, Mengzhang Cai, Jiang Wu, Conghui He, Lijun Wu

Supervised Fine-Tuning (SFT) Large Language Models (LLM) fundamentally rely on high-quality training data. While data selection and data synthesis are two common strategies to improve data quality, existing approaches often face limitations in static dataset curation that fail to adapt to evolving model capabilities. In this paper, we introduce Middo, a self-evolving Model-informed dynamic data optimization framework that uses model-aware data selection and context-preserving data refinement. Unlike conventional one-off filtering/synthesis methods, our framework establishes a closed-loop optimization system: (1) A self-referential diagnostic module proactively identifies suboptimal samples through tri-axial model signals - loss patterns (complexity), embedding cluster dynamics (diversity), and self-alignment scores (quality); (2) An adaptive optimization engine then transforms suboptimal samples into pedagogically valuable training points while preserving semantic integrity; (3) This optimization process continuously evolves with model capability through dynamic learning principles. Experiments on multiple benchmarks demonstrate that our consistently enhances the quality of seed data and boosts LLM’s performance with improving accuracy by 7.15% on average while maintaining the original dataset scale. This work establishes a new paradigm for sustainable LLM training through dynamic human-AI co-evolution of data and models.

Hwan Chang, Yumin Kim, Yonghyun Jun, Hwanhee Lee

As Large Language Models (LLMs) are increasingly deployed in sensitive domains such as enterprise and government, ensuring that they adhere to **user-defined security policies** within context is critical-especially with respect to information non-disclosure. While prior LLM studies have focused on general safety and socially sensitive data, large-scale benchmarks for **contextual security** preservation against attacks remain lacking. To address this, we introduce a novel large-scale benchmark dataset, **CoPriva**, evaluating LLM adherence to contextual non-disclosure policies in question answering. Derived from realistic contexts, our dataset includes explicit policies and queries designed as direct and challenging indirect attacks seeking prohibited information. We evaluate 10 LLMs on our benchmark and reveal a significant vulnerability: many models violate user-defined policies and leak sensitive information. This failure is particularly severe against indirect attacks, highlighting a critical gap in current LLM safety alignment for sensitive applications. Our analysis reveals that while models can often identify the correct answer to a query, they struggle to incorporate policy constraints during generation. In contrast, they exhibit a partial ability to revise outputs when explicitly prompted. Our findings underscore the urgent need for more robust methods to guarantee contextual security.

Junkai Liu, Yujie Tong, Hui Huang, Bowen Zheng, Yiran Hu, Peicheng Wu, Chuan Xiao, Makoto Onizuka, Muyun Yang, Shuyuan Zheng

Legal judgment prediction (LJP), which enables litigants and their lawyers to forecast judgment outcomes and refine litigation strategies, has emerged as a crucial legal NLP task. Existing studies typically utilize legal facts, i.e., facts that have been established by evidence and determined by the judge, to predict the judgment. However, legal facts are often difficult to obtain in the early stages of litigation, significantly limiting the practical applicability of fact-based LJP. To address this limitation, we propose a novel legal NLP task: legal fact prediction (LFP), which takes the evidence submitted by litigants for trial as input to predict legal facts, thereby empowering fact-based LJP technologies to make predictions in the absence of ground-truth legal facts. We also propose the first benchmark dataset, LFPBench, for evaluating the LFP task. Our extensive experiments on LFPBench demonstrate the effectiveness of LFP-empowered LJP and highlight promising research directions for LFP.

Xinkui Lin, Yuhui Zhang, Yongxiu Xu, Kun Huang, Hongzhang Mu, Yubin Wang, Gaopeng Gou, Li Qian, Li Peng, Wei Liu 等

Grounded Multimodal Named Entity Recognition (GMNER), which aims to extract textual entities, their types, and corresponding visual regions from image-text data, has become a critical task in multimodal information extraction. However, existing methods face two major challenges. First, they fail to address the semantic ambiguity caused by polysemy and the long-tail distribution of datasets. Second, unlike visual grounding which provides descriptive phrases, entity grounding only offers brief entity names which carry less semantic information. Current methods lack sufficient semantic interaction between text and image, hindering accurate entity-visual region matching. To tackle these issues, we propose MAKAR, a Multi-Agent framework based Knowledge-Augmented Reasoning, comprising three agents: Knowledge Enhancement, Entity Correction, and Entity Reasoning Grounding. Specifically, in the named entity recognition phase, the Knowledge Enhancement Agent leverages a Multimodal Large Language Model (MLLM) as an implicit knowledge base to enhance ambiguous image-text content with its internal knowledge. For samples with low-confidence entity boundaries and types, the Entity Correction Agent uses web search tools to retrieve and summarize relevant web content, thereby correcting entities using both internal and external knowledge. In the entity grounding phase, the Entity Reasoning Grounding Agent utilizes multi-step Chain-of-Thought reasoning to perform grounding for each entity. Extensive experiments show that MAKAR achieves state-of-the-art performance on two benchmark datasets. Code is available at: https://github.com/Nikol-coder/MAKAR.

Mengze Hong, Wailing Ng, Chen Jason Zhang, Di Jiang

The rapid advancement of Chinese LLMs underscores the need for vertical-domain evaluations to ensure reliable applications. However, existing benchmarks often lack domain coverage and provide limited insights into the Chinese working context. Leveraging qualification exams as a unified framework for expertise evaluation, we introduce \textbf{QualBench}, the first multi-domain Chinese QA benchmark dedicated to localized assessment of Chinese LLMs. The dataset includes over 17,000 questions across six vertical domains, drawn from 24 Chinese qualifications to align with national policies and professional standards. Results reveal an interesting pattern of Chinese LLMs consistently surpassing non-Chinese models, with the Qwen2.5 model outperforming the more advanced GPT-4o, emphasizing the value of localized domain knowledge in meeting qualification requirements. The average accuracy of 53.98% reveals the current gaps in domain coverage within model capabilities. Furthermore, we identify performance degradation caused by LLM crowdsourcing, assess data contamination, and illustrate the effectiveness of prompt engineering and model fine-tuning, suggesting opportunities for future improvements through multi-domain RAG and Federated Learning. Data and code are publicly available at https://github.com/mengze-hong/QualBench.

Mengze Hong, Wailing Ng, Chen Jason Zhang, Yuanfeng Song, Di Jiang

Discovering customer intentions is crucial for automated service agents, yet existing intent clustering methods often fall short due to their reliance on embedding distance metrics and neglect of underlying semantic structures. To address these limitations, we propose an **LLM-in-the-loop (LLM-ITL)** intent clustering framework, integrating the language understanding capabilities of LLMs into conventional clustering algorithms. Specifically, this paper (1) examines the effectiveness of fine-tuned LLMs in semantic coherence evaluation and intent cluster naming, achieving over 95% accuracy aligned with human judgments; (2) designs an LLM-ITL framework that facilitates the iterative discovery of coherent intent clusters and the optimal number of clusters; and (3) introduces context-aware techniques tailored for customer service dialogue. Since existing English benchmarks lack sufficient semantic diversity and intent coverage, we further present a comprehensive Chinese dialogue intent dataset comprising over 100k real customer service calls with 1,507 human-annotated clusters. The proposed approaches significantly outperform LLM-guided baselines, achieving notable improvements in clustering quality, cost efficiency, and downstream applications. Combined with several best practices, our findings highlight the prominence of LLM-in-the-loop techniques for scalable dialogue data mining.

Yanxu Ji, Jinzhong Ning, Yijia Zhang, Zhi Liu, Hongfei Lin

Stance detection aims to identify the attitude expressed in text towards a specific target. Recent studies on zero-shot and few-shot stance detection focus primarily on learning generalized representations from explicit targets. However, these methods often neglect implicit yet semantically important targets and fail to adaptively adjust the relative contributions of text and target in light of contextual dependencies. To overcome these limitations, we propose a novel two-stage framework: First, a data augmentation framework named Hierarchical Collaborative Target Augmentation (HCTA) employs Large Language Models (LLMs) to identify and annotate implicit targets via Chain-of-Thought (CoT) prompting and multi-LLM voting, significantly enriching training data with latent semantic relations. Second, we introduce DyMCA, a Dynamic Multi-level Context-aware Attention Network, integrating a joint text-target encoding and a content-aware mechanism to dynamically adjust text-target contributions based on context. Experiments on the benchmark dataset demonstrate that our approach achieves state-of-the-art results, confirming the effectiveness of implicit target augmentation and fine-grained contextual modeling.

Ying Zhao, Yuanzhao Guo, Xuemeng Weng, Yuan Tian, Wei Wang, Yi Chang

Eliminating toxicity from Large Language Models (LLMs) is crucial for ensuring user safety. However, current methods have limitations in the analysis and utilization of toxic samples, failing to fully harness their potential. Through comparative analysis of toxic and safe samples, we discover that toxic samples exhibit diversity and, within this diversity, there lies specificity. These findings suggest that leveraging these characteristics of toxic samples could enhance the performance of algorithms in detoxifying LLMs. To this end, we propose a novel diverse detoxification framework, DivDetox, which comprises two innovative components: a Multi-Category-Induced Personalized Sample Generation (MPSG) strategy and a Scaled Contrastive DPO (SC-DPO) approach. The former is designed to elicit a variety of personalized toxic responses from the LLM, while the latter is constructed to precisely and fully utilize these toxic responses. Experiments on benchmark datasets across different model scales and different detoxification tasks verify the effectiveness of our architecture.

Zehong Yan, Peng Qi, Wynne Hsu, Mong-Li Lee

Multimodal misinformation, encompassing textual, visual, and cross-modal distortions, poses an increasing societal threat that is amplified by generative AI. Existing methods typically focus on a single type of distortion and struggle to generalize to unseen scenarios. In this work, we observe that different distortion types share common reasoning capabilities while also requiring task-specific skills. We hypothesize that joint training across distortion types facilitates knowledge sharing and enhances the model’s ability to generalize. To this end, we introduce TRUST-VL, a unified and explainable vision-language model for general multimodal misinformation detection. TRUST-VL incorporates a novel Question-Aware Visual Amplifier module, designed to extract task-specific visual features. To support training, we also construct TRUST-Instruct, a large-scale instruction dataset containing 198K samples featuring structured reasoning chains aligned with human fact-checking workflows. Extensive experiments on both in-domain and zero-shot benchmarks demonstrate that TRUST-VL achieves state-of-the-art performance, while also offering strong generalization and interpretability.

Zhongyi Zhou, Yichen Zhu, Minjie Zhu, Junjie Wen, Ning Liu, Zhiyuan Xu, Weibin Meng, Yaxin Peng, Chaomin Shen, Feifei Feng 等

Humans possess a unified cognitive ability to perceive, comprehend, and interact with the physical world. Why can’t large language models replicate this holistic understanding? Through a systematic analysis of existing training paradigms in vision-language-action models (VLA), we identify two key challenges: spurious forgetting, where robot training overwrites crucial visual-text alignments, and task interference, where competing control and understanding tasks degrade performance when trained jointly. To overcome these limitations, we propose ChatVLA, a novel framework featuring Phased Alignment Training, which incrementally integrates multimodal data after initial control mastery, and a Mixture-of-Experts architecture to minimize task interference. ChatVLA demonstrates competitive performance on visual question-answering datasets and significantly surpasses state-of-the-art vision-language-action (VLA) methods on multimodal understanding benchmarks. Notably, it achieves a six times higher performance on MMMU and scores 47.2% on MMStar with a more parameter-efficient design than ECoT. Furthermore, ChatVLA demonstrates superior performance on 25 real-world robot manipulation tasks compared to existing VLA methods like OpenVLA. Our findings highlight the potential of our unified framework for achieving both robust multimodal understanding and effective robot control.

Mo Zhiqiang, Yang Hua, Jiahui Li, Yuan Liu, Shawn Wong, Jianmin Huang

Graph Neural Networks (GNNs) have shown immense potential in improving the performance of large-scale models by effectively incorporating structured relational information. However, current approaches face two key challenges: (1) achieving robust semantic alignment between graph representations and large models, and (2) ensuring interpretability in the generated outputs. To address these challenges, we propose ExGLM (Explainable Graph Language Model), a novel training framework designed to seamlessly integrate graph and language modalities while enhancing transparency. Our framework introduces two core components: (1) a graph-language synergistic alignment module, which aligns graph structures with language model to ensure semantic consistency across modalities; and (2) a judge-and-improve paradigm, which allows the language model to iteratively evaluate, refine, and prioritize responses with higher interpretability, thereby improving both performance and transparency. Extensive experiments conducted on three benchmark datasets—ogbn-arxiv, Cora, and PubMed—demonstrate that ExGLM not only surpasses existing methods in efficiency but also generates outputs that are significantly more interpretable, effectively addressing the primary limitations of current approaches.