论文检索

Jun Bai, Minghao Tong, Yang Liu, Zixia Jia, Zilong Zheng

Context faithfulness is essential for reliable reasoning in context-dependent scenarios. However, large language models often struggle to ground their outputs in the provided context, resulting in irrelevant responses.Inspired by the emergent expert specialization observed in mixture-of-experts architectures, this work investigates whether certain experts exhibit specialization in context utilization—offering a potential pathway toward targeted optimization for improved context faithfulness.To explore this, we propose Router Lens, a method that accurately identifies context-faithful experts. Our analysis reveals that these experts progressively amplify attention to relevant contextual information, thereby enhancing context grounding.Building on this insight, we introduce Context-faithful Expert Fine-Tuning (CEFT), a lightweight optimization approach that selectively fine-tunes context-faithful experts.Experiments across a wide range of benchmarks and models demonstrate that CEFT matches or surpasses the performance of full fine-tuning while being significantly more efficient.

Wei Liu, Michael Strube

In linguistics, coherence can be achieved by different means, such as by maintaining reference to the same set of entities across sentences and by establishing discourse relations between them. However, most existing work on coherence modeling focuses exclusively on either entity features or discourse relation features, with little attention given to combining the two. In this study, we explore two methods for jointly modeling entities and discourse relations for coherence assessment. Experiments on three benchmark datasets show that integrating both types of features significantly enhances the performance of coherence models, highlighting the benefits of modeling both simultaneously for coherence evaluation.

Xuelin Li, Xiangqi Jin, Linfeng Zhang

Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance. Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference. To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression. In GraphKV, \textbf{tokens} are modeled as \textbf{nodes} with importance scores, and \textbf{edges} represent their \textbf{similarity relationships}. Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens. GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-and-play manner. Codes are available in the supplementary materials and will be released on Github.

Honghao Fu, Junlong Ren, Qi Chai, Deheng Ye, Yujun Cai, Hao Wang

Large language models (LLMs) have shown significant promise in embodied decision-making tasks within virtual open-world environments. Nonetheless, their performance is hindered by the absence of domain-specific knowledge. Methods that finetune on large-scale domain-specific data entail prohibitive development costs. This paper introduces VistaWise, a cost-effective agent framework that integrates cross-modal domain knowledge and finetunes a dedicated object detection model for visual analysis. It reduces the requirement for domain-specific training data from millions of samples to a few hundred. VistaWise integrates visual information and textual dependencies into a cross-modal knowledge graph (KG), enabling a comprehensive and accurate understanding of multimodal environments. We also equip the agent with a retrieval-based pooling strategy to extract task-related information from the KG, and a desktop-level skill library to support direct operation of the Minecraft desktop client via mouse and keyboard inputs. Experimental results demonstrate that VistaWise achieves state-of-the-art performance across various open-world tasks, highlighting its effectiveness in reducing development costs while enhancing agent performance.

Omkar Gurjar, Agam Goyal, Eshwar Chandrasekharan

Key point extraction is an important task in argument summarization which involves extracting high-level short summaries from arguments. Existing approaches for KP extraction have been mostly evaluated on the popular ArgKP21 dataset. In this paper, we highlight some of the major limitations of the ArgKP21 dataset and demonstrate the need for new benchmarks that are more representative of actual human conversations. Using SoTA large language models (LLMs), we curate a new argument key point extraction dataset called ArgCMV comprising of ∼12K arguments from actual online human debates spread across ∼3K topics. Our dataset exhibits higher complexity such as longer, co-referencing arguments, higher presence of subjective discourse units, and a larger range of topics over ArgKP21. We show that existing methods do not adapt well to ArgCMV and provide extensive benchmark results by experimenting with existing baselines and latest open source models. This work introduces a novel KP extraction dataset for long-context online discussions, setting the stage for the next generation of LLM-driven summarization research.

Hwiyeong Lee, Uiji Hwang, Hyelim Lim, Taeuk Kim

Large language models often retain unintended content, prompting growing interest in knowledge unlearning.Recent approaches emphasize localized unlearning, restricting parameter updates to specific regions in an effort to remove target knowledge while preserving unrelated general knowledge. However, their effectiveness remains uncertain due to the lack of robust and thorough evaluation of the trade-off between the competing goals of unlearning.In this paper, we begin by revisiting existing localized unlearning approaches. We then conduct controlled experiments to rigorously evaluate whether local parameter updates causally contribute to unlearning.Our findings reveal that the set of parameters that must be modified for effective unlearning is not strictly determined, challenging the core assumption of localized unlearning that parameter locality is inherently indicative of effective knowledge removal.

Wei Huang, Anda Cheng, Yinggui Wang

Recent advancements in large language models (LLMs) have shown impressive capabilities in various downstream tasks but typically face Catastrophic Forgetting (CF) during fine-tuning. In this paper, we propose the Forgetting-Aware Pruning Metric (FAPM), a novel pruning-based approach to balance CF and downstream task performance. Our investigation reveals that the degree to which task vectors (i.e., the subtraction of pre-trained weights from the weights fine-tuned on downstream tasks) overlap with pre-trained model parameters is a critical factor for CF. Based on this finding, FAPM employs the ratio of the task vector to pre-trained model parameters as a metric to quantify CF, integrating this measure into the pruning criteria. Importantly, FAPM does not necessitate modifications to the training process or model architecture, nor does it require any auxiliary data. We conducted extensive experiments across eight datasets, covering natural language inference, General Q&A, Medical Q&A, Math Q&A, reading comprehension, and cloze tests. The results demonstrate that FAPM limits CF to just 0.25% while maintaining 99.67% accuracy on downstream tasks. We provide the codes of FAPM at an anonymous repository(https://anonymous.4open.science/r/FAPM-65CF).

Yanshuo Wang, Yanghao Zhou, Yukang Lin, Haoxing Chen, Jin Zhang, Wentao Zhu, Jie Hong, Xuesong Li

End-to-end automatic speech recognition (ASR) based on deep learning has achieved impressive progress in recent years. However, the performance of ASR foundation model often degrades significantly on out-of-domain data due to real-world domain shifts. Test-Time Adaptation (TTA) methods aim to mitigate this issue by adapting models during inference without access to source data. Despite recent progress, existing ASR TTA methods often struggle with instability under continual and long-term distribution shifts. To alleviate the risk of performance collapse due to error accumulation, we propose Dynamic Model-bank Single-Utterance Test-time Adaptation (DMSUTA), a sustainable continual TTA framework based on adaptive ASR model ensembling. DMSUTA maintains a dynamic model bank, from which a subset of checkpoints is selected for each test sample based on confidence and uncertainty criteria. To preserve both model plasticity and long-term stability, DMSUTA actively manages the bank by filtering out potentially collapsed models. This design allows DMSUTA to continually adapt to evolving domain shifts in ASR test-time scenarios. Experiments on diverse, continuously shifting ASR TTA benchmarks show that DMSUTA consistently outperforms existing continual TTA baselines, demonstrating superior robustness to domain shifts in ASR.

Xuhui Zheng, Kang An, Ziliang Wang, Yuhang Wang, Yichao Wu

Efficient multi-hop reasoning requires Large Language Models (LLMs) based agents to acquire high-value external knowledge iteratively. Previous work has explored reinforcement learning (RL) to train LLMs to perform search-based document retrieval, achieving notable improvements in QA performance, but underperform on complex, multi-hop QA resulting from the sparse rewards from global signal only. To address this gap in existing research, we introduce StepSearch, a framework for search LLMs that trained with step-wise proximal policy optimization method. It consists of richer and more detailed intermediate search rewards and token-level process supervision based on information gain and redundancy penalties to better guide each search step. We constructed a fine-grained question-answering dataset containing sub-question-level search trajectories based on open source datasets through a set of data pipeline method. On standard multi-hop QA benchmarks, it significantly outperforms global-reward baselines, achieving 11.2% and 4.2% absolute improvements for 3B and 7B models over various search with RL baselines using only 19k training data, demonstrating the effectiveness of fine-grained, stepwise supervision in optimizing deep search LLMs. The project is open source at https://github.com/Zillwang/StepSearch

Yuchi Wang, Yishuo Cai, Shuhuai Ren, Sihan Yang, Linli Yao, Yuanxin Liu, Yuanxing Zhang, Pengfei Wan, Xu Sun

Image recaptioning is widely used to generate training datasets with enhanced quality for various multimodal tasks. Existing recaptioning methods typically rely on powerful multimodal large language models (MLLMs) to enhance textual descriptions, but often suffer from inaccuracies due to hallucinations and incompleteness caused by missing fine-grained details. To address these limitations, we propose RICO, a novel framework that refines captions through visual reconstruction. Specifically, we leverage a text-to-image model to reconstruct a caption into a reference image, and prompt an MLLM to identify discrepancies between the original and reconstructed images to refine the caption. This process is performed iteratively, further progressively promoting the generation of more faithful and comprehensive descriptions. To mitigate the additional computational cost induced by the iterative process, we introduce RICO-Flash, which learns to generate captions like RICO using DPO. Extensive experiments demonstrate that our approach significantly improves caption accuracy and completeness, outperforms most baselines by approximately 10% on both CapsBench and CompreCap.

Wenxuan Wang, Shi Juluan, Zixuan Ling, Yuk-Kit Chan, Chaozheng Wang, Cheryl Lee, Youliang Yuan, Jen-tse Huang, Wenxiang Jiao, Michael R. Lyu

Equipped with the capability to call functions, modern LLM agents can leverage external tools for addressing a range of tasks unattainable through language skills alone. However, the effective execution of these tools relies heavily not just on the advanced capabilities of LLM agents but also on precise user instructions, which often cannot be ensured in the real world. To evaluate the performance of LLM agents tool-use under imperfect instructions, we meticulously examine the real-world instructions queried from users, analyze the error patterns, and build a challenging tool-use benchmark called Noisy ToolBench. We find that due to the next-token prediction training objective, LLM agents tend to arbitrarily generate the missed argument, which may lead to hallucinations and risks. To address this issue, we propose a novel framework, Ask-when-Needed, which prompts LLM agents to ask questions to users whenever they encounter obstacles due to unclear instructions. Moreover, to reduce the manual labor involved in user-LLM interaction and assess LLM agents’ performance in tool utilization from both accuracy and efficiency perspectives, we design an automated evaluation tool named ToolEvaluator. Our experiments demonstrate that the Ask-when-Needed significantly outperforms existing frameworks for tool learning in the Noisy ToolBench. We will release all related code and datasets to support future research.

Jihee Kim, Subeen Park, Hakyung Lee, YongTaek Lim, Hyo-won Suh, Kyungwoo Song

Addressing the challenges in QA for specific technical domains requires identifying relevant portions of extensive documents and generating answers based on this focused content. Traditional pre-trained LLMs often struggle with domain-specific terminology, while fine-tuned LLMs demand substantial computational resources. To overcome these limitations, we propose TIDES, Technical Information Distillation and Extraction System. TIDES is a training-free approach that combines traditional TF-IDF techniques with prompt-based LLMs in a hybrid process, effectively addressing complex technical questions. It uses TF-IDF to identify and prioritize domain-specific words that are rare in other documents and LLMs to refine the candidate pool by focusing on the most relevant segments in documents through multiple stages. Our approach improves the precision and efficiency of QA systems in technical contexts without LLM retraining.

Gihun Cho, Seunghyun Jang, Hanbin Ko, Inhyeok Baek, Chang Min Park

We introduce CREPE (Rapid Chest X-ray Report Evaluation by Predicting Multi-category Error Counts), a rapid, interpretable, and clinically grounded metric for automated chest X-ray report generation. CREPE uses a domain-specific BERT model fine-tuned with a multi-head regression architecture to predict error counts across six clinically meaningful categories. Trained on a large-scale synthetic dataset of 32,000 annotated report pairs, CREPE demonstrates strong generalization and interpretability. On the expert-annotated ReXVal dataset, CREPE achieves a Kendall’s tau correlation of 0.786 with radiologist error counts, outperforming traditional and recent metrics. CREPE achieves these results with an inference speed approximately 280 times faster than large language model (LLM)-based approaches, enabling rapid and fine-grained evaluation for scalable development of chest X-ray report generation models.

Heekyung Lee, Jiaxin Ge, Tsung-Han Wu, Minwoo Kang, Trevor Darrell, David M. Chan

Rebus puzzles, visual riddles that encode language through imagery, spatial arrangement, and symbolic substitution, pose a unique challenge to current vision-language models (VLMs). Unlike traditional image captioning or question answering tasks, rebus solving requires multimodal abstraction, symbolic reasoning, and a grasp of cultural, phonetic and linguistic puns. In this short paper, we investigate the capacity of contemporary VLMs to interpret and solve rebus puzzles by constructing a hand-generated and annotated benchmark of diverse english-language rebus puzzles, ranging from simple pictographic substitutions to spatially-dependent cues (“head” over “heels”). We analyze how different VLMs perform, and our findings reveal that while VLMs exhibit some surprising capabilities in decoding simple visual clues, they struggle significantly with tasks requiring abstract reasoning, lateral thinking, and understanding visual metaphors.

Haihua Xie, Yinzhu Cheng, Yaqing Wang, Miao He, Mingming Sun

This paper addresses the important yet underexplored task of **multi-class sentiment analysis (MCSA)**, which remains challenging due to the subtle semantic differences between adjacent sentiment categories and the scarcity of high-quality annotated data. To tackle these challenges, we propose **RD-MCSA** (**R**ationales and **D**emonstrations-based **M**ulti-**C**lass **S**entiment **A**nalysis), an In-Context Learning (ICL) framework designed to enhance MCSA performance under limited supervision by integrating classification rationales with adaptively selected demonstrations. First, semantically grounded classification rationales are generated from a representative, class-balanced subset of annotated samples selected using a tailored balanced coreset algorithm. These rationales are then paired with demonstrations chosen through a similarity-based mechanism powered by a **multi-kernel Gaussian process (MK-GP)**, enabling large language models (LLMs) to more effectively capture fine-grained sentiment distinctions. Experiments on five benchmark datasets demonstrate that RD-MCSA consistently outperforms both supervised baselines and standard ICL methods across various evaluation metrics.

Jisu Kim, Youngwoo Shin, Uiji Hwang, Jihun Choi, Richeng Xuan, Taeuk Kim

Idioms have long posed a challenge due to their unique linguistic properties, which set them apart from other common expressions. While recent studies have leveraged large language models (LLMs) to handle idioms across various tasks, e.g., idiom-containing sentence generation and idiomatic machine translation, little is known about the underlying mechanisms of idiom processing in LLMs, particularly in multilingual settings. To this end, we introduce MIDAS, a new large-scale dataset of idioms in six languages, each paired with its corresponding meaning. Leveraging this resource, we conduct a comprehensive evaluation of LLMs’ idiom processing ability, identifying key factors that influence their performance. Our findings suggest that LLMs rely not only on memorization, but also adopt a hybrid approach that integrates contextual cues and reasoning, especially when processing compositional idioms. This implies that idiom understanding in LLMs emerges from an interplay between internal knowledge retrieval and reasoning-based inference.

Nurit Cohen Inger, Yehonatan Elisha, Bracha Shapira, Lior Rokach, Seffi Cohen

Large language models (LLMs) often appear to excel on public benchmarks, but these high scores may mask an overreliance on dataset-specific surface cues rather than true language understanding. We introduce the **Chameleon Benchmark Overfit Detector (C-BOD)**, a meta-evaluation framework designed to reveal such overfitting. C-BOD systematically rephrases benchmark inputs via a parameterized transformation that preserves semantic content and labels, enabling the detection of performance degradation indicative of superficial pattern reliance.We conduct extensive experiments across two datasets, three rephrasing models, and multiple distortion levels, evaluating 32 state-of-the-art LLMs. On the MMLU benchmark, C-BOD reveals an average performance drop of 2.75% under modest rephrasings, with over 80% of models exhibiting statistically significant differences. Notably, higher-performing models and larger LLMs tend to show greater sensitivity, suggesting a deeper dependence on benchmark-specific phrasing.Due to its dataset and model-agnostic design, C-BOD can be easily integrated into evaluation pipelines and offers a promising foundation for overfitting mitigation strategies. Our findings challenge the community to look beyond leaderboard scores and prioritize resilience and generalization in LLM evaluation. Our code and benchmark datasets are availableat: https://github.com/nuritci/cbod

Yukun Zhang, Xueqing Zhou

We present Continuous-Time Attention, a novel framework that infuses partial differential equations (PDEs) into the Transformer’s attention mechanism to better handle long sequences. Instead of relying on a static attention matrix, we allow attention weights to evolve along a pseudo-time dimension governed by diffusion, wave, or reaction-diffusion dynamics. This dynamic process systematically smooths local noise, strengthens long-range dependencies, and improves gradient stability during training.Our theoretical analysis shows that PDE-driven attention mitigates the exponential decay of distant interactions and improves the optimization landscape. Empirically, Continuous-Time Attention achieves consistent performance gains over both standard and long-sequence Transformer variants across a range of tasks. These results suggest that embedding continuous-time dynamics into attention mechanisms is a promising direction for enhancing global coherence and scalability in Transformer models. Code is publicly available at:https://github.com/XueqingZhou/Continuous-Time-Attention

Fanqi Wan, Longguang Zhong, Ziyi Yang, Ruijun Chen, Xiaojun Quan

While training large language models (LLMs) from scratch can indeed lead to models with distinct capabilities and strengths, it incurs substantial costs and may lead to redundancy in competencies. Knowledge fusion aims to integrate existing LLMs of diverse architectures and capabilities into a more potent LLM through lightweight continual training, thereby reducing the need for costly LLM development. In this work, we propose a new framework for the knowledge fusion of chat LLMs through two main stages, resulting in FuseChat. Firstly, we conduct pairwise knowledge fusion on source chat LLMs of varying structures and scales to create multiple target LLMs with identical structure and size via lightweight fine-tuning. During this process, a statistics-based token alignment approach is introduced as the cornerstone for fusing LLMs with different structures. Secondly, we merge these target LLMs within the parameter space, where we propose a novel method for determining the merging coefficients based on the magnitude of parameter updates before and after fine-tuning. We implement and validate FuseChat using six prominent chat LLMs with diverse architectures and scales. Experimental results on two instruction-following benchmarks, AlpacaEval 2.0 and MT-Bench, demonstrate the superiority of FuseChat-7B over baselines of various sizes.

Pengcheng Jiang, Xueqiang Xu, Jiacheng Lin, Jinfeng Xiao, Zifeng Wang, Jimeng Sun, Jiawei Han

Retrieval-augmented generation (RAG) systems empower large language models (LLMs) to access external knowledge during inference. Recent advances have enabled LLMs to act as search agents via reinforcement learning (RL), improving information acquisition through multi-turn interactions with retrieval engines. However, existing approaches either optimize retrieval using search-only metrics (e.g., NDCG) that ignore downstream utility or fine-tune the entire LLM to jointly reason and retrieve—entangling retrieval with generation and limiting the real search utility and compatibility with frozen or proprietary models. In this work, we propose **s3**, a lightweight, model-agnostic framework that decouples the searcher from the generator and trains the searcher using a Gain Beyond RAG reward: the improvement in generation accuracy over naïve RAG. **s3** requires only 2.4k training samples to outperform baselines trained on over 70 \times more data, consistently delivering stronger downstream performance across six general QA and five medical QA benchmarks.