论文检索

Bonggeun Choi, Keunha Kim, Junho Han, Youngjoong Ko

Retrieval-Augmented Generation (RAG) has significantly advanced open-domain question answering systems by incorporating external knowledge into large language models. Despite its effectiveness, existing RAG pipelines suffer from critical efficiency limitations. In particular, modern transformer-based generators exhibit quadratic or higher computational complexity with respect to input sequence length and hidden dimensionality, leading to substantial inference latency as model scales and contextual inputs increase. This issue is exacerbated in RAG settings, where retrieved contexts substantially expand the input prompt. To alleviate this challenge, we propose an effective compression-based RAG framework, ConvX, that directly leverages indexed dense representations produced by a retriever, entirely substituting to long text contexts. Our approach expands a single dense representation into a fixed number of memory slots using a lightweight converter to provide rich lexical information. This design enables efficient knowledge integration while significantly reducing input length and computational overhead. Empirical evaluations demonstrate that the proposed model achieves competitive performances compared to the existing state-of-the-art model that uses a large ad-hoc context compressor, while offering substantially improved inference efficiency.

Mingkuan Zhao, Wentao Hu, Tianchen Huang, Yuheng Min, Suquan Chen, Yide Gao, Yanbo Zhai, Shuangyong Song, Xuelong Li

Hallucination in Large Language Models (LLMs)—characterized by the generation of content inconsistent with contextual facts or logical constraints—remains a persistent challenge for reliable deployment. In this work, we address this issue through a geometric framework rooted in the linear representation hypothesis. We propose that hallucinations manifest as orthogonal noise relative to the semantic manifold of the residual stream. Specifically, we hypothesize that while attention heads ideally propagate information congruent with the context subspace, hallucinations arise when specific heads introduce components orthogonal to this subspace, disrupting the coherence of the latent representation. Based on this formulation, we introduce Dynamic Contextual Orthogonalization (DCO), an inference-time intervention method. DCO utilizes the input residual stream as a dynamic context anchor to perform orthogonal decomposition on attention head outputs. To distinguish between context-aligned semantic updates and divergent noise, DCO employs a layer-wise Z-score suppression mechanism that selectively attenuates outlier orthogonal components based on statistical distributions. Evaluations on Llama-3-8B and 70B across benchmarks such as XSum, NQ-Swap, and IFEval demonstrate that DCO achieves superior contextual faithfulness compared to state-of-the-art intervention baselines. Furthermore, DCO maintains high performance on knowledge-intensive tasks like TriviaQA and TruthfulQA, effectively mitigating the trade-off between hallucination suppression and parametric knowledge retention often observed in existing methods. Our findings validate the geometric interpretation of hallucinations and establish DCO as a computationally efficient approach for enforcing manifold alignment.Our code is available at https://anonymous.4open.science/r/DCO-4AB0

Junwen Mo, MinhDuc Vo, Noriki Nishida, Shin’ichi Satoh, Hideki Nakayama

Japanese Sign Language (JSL) is a low-resource sign language that has received limited attention in the AI research community, primarily due to the lack of large-scale, publicly available parallel corpora. In this work, we introduce J-Shuwa, a large-scale JSL-Japanese parallel corpus constructed from YouTube videos with hard-coded subtitles and closed captions. The corpus contains 197K parallel JSL-Japanese sentence pairs, totaling approximately 300 hours of video, making it the largest publicly available JSL dataset to date. We conduct sign language translation (SLT) experiments by training models on J-Shuwa and evaluating them on the JSL Dialogue Corpus under both zero-shot and fine-tuned settings. Our results demonstrate that J-Shuwa is effective for training SLT models. Beyond SLT, we believe that J-Shuwa can also serve as a valuable resource for future JSL research across a wide range of tasks. The dataset and code are publicly available at: https://github.com/SpaJune/J-Shuwa.

Mengna Zhu, Jibing Wu, Lihua Liu, Yuran Gong, Yang Hao, Fu Yachao, Mao Wang, Lei Hou, Juanzi Li

Emergency response is a safety-critical public governance task that demands accurate and timely decision-making based on complex event information. This process involves multiple stages, including information collection, integration, analysis, risk assessment, and decision recommendation. Existing research has predominantly concentrated on the earlier stages, while studies focusing on the decision support phase remain underexplored, primarily due to the lack of suitable datasets for reliable and compliance-aware decision-oriented modeling and evaluation. To bridge this gap, we introduce the first real-world Emergency Decision-Making dataset EDM-Bench, comprising 1,179 instances spanning diverse task formats, including judgment, choice, short-answer, and structured emergency report generation. We also construct a structured rule repository, EDM-R², which contains 3,406 parsed emergency regulations to enhance decision reliability. Building on these resources, we propose a rule-enhanced reasoning framework, R³V-EDM, which integrates external regulatory knowledge with constrained inference mechanisms to improve both decision safety and interpretability. Extensive experiments demonstrate the inherent complexity of emergency decision-making and validate the effectiveness of our approach in enabling more reliable and trustworthy decisions.

Tianyu Pang, Yujie Fang, Zihang Liu, Shenyang Deng, Lei Hsiung, Shuhua Yu, Yaoqing Yang

Muon has recently shown promising results in LLM training. In this work, we study how to further improve Muon. We argue that Muon’s orthogonalized update rule suppresses the emergence of heavy-tailed weight spectra and over-emphasizes the training along noise-dominated directions. Motivated by the Heavy-Tailed Self-Regularization (HT-SR) theory, we propose HTMuon. HTMuon preserves Muon’s ability to capture parameter interdependencies while producing heavier-tailed updates and inducing heavier-tailed weight spectra. Experiments on LLM pretraining and image classification show that HTMuon consistently improves performance over state-of-the-art baselines and can also serve as a plug-in on top of existing Muon variants. For example, on LLaMA pretraining on the C4 dataset, HTMuon reduces perplexity by up to 0.98 compared to Muon. We further theoretically show that HTMuon corresponds to steepest descent under the Schatten-q norm constraint and provide convergence analysis in smooth non-convex settings. The implementation of HTMuon is available at https://github.com/TDCSZ327/HTmuon.

Ximing Wen, Wenbo Li, Sudipta Paul, Yashas Malur Saidutta, Kalpa Gunaratna, Srinivas Chappidi

Visually Grounded Document Question Answering often lacks robust, end-to-end solutions capable of handling complex, multi-answer queries without reliance on ad-hoc processing. In this work, we propose two turnkey LLM architectures to address this gap. We first introduce a single-head architecture where coordinates are represented as special tokens within the unified vocabulary. While structurally robust, this approach suffers from the limitations of discrete supervision; to address this, we propose a novel “softening token” method that enables differentiable Mean-Squared-Error loss over token probabilities. Although this significantly improves visual grounding, the spatial precision remains bound by discretization. Consequently, we propose a second solution: a dual-head architecture that alternates between text generation and regression-based bounding box prediction. This method offers high spatial precision via a regression head, further stabilized by our introduction of an Intersection-over-Union loss. Finally, by combining the single head model’s structural robustness with the high precision of the dual head model, we propose an ensemble method that yields significant performance gains beyond each of individual components.

Daeun Lee, Jaehong Yoon, Jaemin Cho, Mohit Bansal

Recent text-to-video (T2V) diffusion models have made remarkable progress in generating high-quality videos. However, they often struggle to align with complex text prompts, particularly when multiple objects, attributes, or spatial relations are specified. We introduce VideoRepair, the first self-correcting, training-free, and model-agnostic video refinement framework that automatically detects fine- grained text–video misalignments and performs targeted, localized corrections. Our key insight is that even misaligned videos usually contain correctly generated regions that should be preserved rather than regenerated. Building on this observation, VideoRepair proposes a novel region-preserving refinement strategy with three stages: (i) misalignment detection, where MLLM-based evaluation with automatically generated evaluation questions identifies misaligned regions; (ii) refinement planning, which preserves correctly generated entities, segments their regions across frames, and constructs targeted prompts for misaligned areas; and (iii) localized refinement, which selectively regenerates problematic regions while preserving faithful content through joint optimization of preserved and newly generated areas. On two benchmarks, EvalCrafter and T2V-CompBench with four recent T2V backbones, VideoRepair achieves substantial improvements over recent baselines across diverse alignment metrics. Comprehensive ablations further demonstrate the efficiency, robustness, and interpretability of our framework.

Yen-Che Chien, Kuang-Da Wang, Wei-Yao Wang, Wen-Chih Peng

Recent advances in autonomous digital agents from industry (e.g., Manus AI and Gemini’s research mode) highlight their potential for structured tasks through autonomous decision-making and task decomposition, but it remains unclear how well such systems support real-world information-intensive workflows. We study this question in journalism, where newswriting requires iterative planning, contextual reasoning, and active discovery of missing background to produce a coherent article. We introduce NEWSAGENT, a benchmark for evaluating how agents search raw materials, select relevant information, and iteratively revise drafts through core journalistic functions. Given a writing instruction and partial firsthand materials, agents must identify narrative perspectives, issue keyword-based queries, retrieve historical context, and generate complete news articles. Unlike typical summarization or retrieval tasks, essential context is not directly available and must be actively discovered, reflecting real-world reporting constraints. NEWSAGENT consists of 6k human-verified examples derived from real news. We evaluate open- and closed-sourced LLMs with commonly-used agentic frameworks on NEWSAGENT, which shows that agents are capable of retrieving relevant facts but struggling with planning and narrative integration. We believe that NEWSAGENT serves a realistic testbed for iterating and evaluating agent capabilities in terms of web data manipulation to real-world productivity. The benchmark resources are publicly available at https://github.com/wywyWang/CoachAI-Projects.

Haoming Li, Jessica Ouyang

Writing a literature review requires a deep understanding of the relationships among cited papers: how they build on, challenge, or offer alternative perspectives to one another. We present Graph-Reasoning Aided Survey Planning (GRASP), a framework combining LLM planning for related work generation with graph algorithms to extract key relationships among cited papers. Our two-layer graph structure consists of a Graph of Thoughts and an Argument-Counteragument Planning Network, representing the cited papers at different levels of granularity, and we apply topology-aware pruning via a Steiner tree to identify the core inter-paper relationships captured in our graph. Our citation analysis-based evaluation shows that GRASP generates RWS that closely match human-written targets in terms of the discourse roles, intents, and grouping of citations.

Shiyu Liu, Yongjing Yin, Jianhao Yan, Yunbo Tang, Qinggang Zhang, Bei Li, Xin Chen, Jingang Wang, Xunliang Cai, Jinsong Su

RL-based agentic search enables LLMs to solve complex questions via dynamic planning and external search. While this approach significantly enhances accuracy with agent policies optimized via large-scale reinforcement learning, we identify a critical gap in reliability: these agents fail to recognize their reasoning boundaries and rarely admit "I DON’T KNOW" even when evidence is insufficient or reasoning reaches its limit. The lack of reliability often leads to plausible but unreliable answers, introducing significant risks in many real-world scenarios. To this end, we propose Boundary-Aware Policy Optimization (BAPO), a novel RL framework designed to cultivate reliable boundary awareness without compromising accuracy. BAPO introduces two key components: (i) a group-based boundary-aware reward that encourages an IDK response only when the reasoning reaches its limit, and (ii) an adaptive reward modulator that strategically suspends this reward during early exploration, preventing the model from exploiting IDK as a shortcut. Extensive experiments on four benchmarks demonstrate that BAPO substantially enhances the overall reliability of agentic search.

Xuan Luo, Lewei Yao, Lanqing Hong, Kai Chen, Dehua Tao, Daxin Tan, Yukun Deng, Ruifeng Xu, Jing Li

While the automatic evaluation of omni-modal large models (OLMs) is essential, assessing empathy remains a significant challenge due to its inherent affectivity. To investigate this challenge, we introduce AEQ-Bench (Audio Empathy Quotient Benchmark), a novel benchmark to systematically assess two core empathetic capabilities of OLMs: (i) generating empathetic responses by comprehending affective cues from multi-modal inputs (audio + text), and (ii) judging the empathy of audio responses without relying on text transcription. Compared to existing benchmarks, AEQ-Bench incorporates two novel settings that vary in context specificity and speech tone. Comprehensive assessment across linguistic and paralinguistic metrics reveals that (1) OLMs trained with audio output capabilities generally outperformed models with text-only outputs, and (2) while OLMs align with human judgments for coarse-grained quality assessment, they remain unreliable for evaluating fine-grained paralinguistic expressiveness.

Yutao Mou, Yuxiao Luo, Shikun Zhang, Wei Ye

Current safety alignment techniques for large language models (LLMs) struggle to balance harmlessness and helpfulness: improving safety often comes at the cost of degraded utility. Our preliminary study shows that guiding unaligned base models with safety-aware reasoning that includes explicit self-reflection can effectively defend jailbreak attacks while preserving response quality. This observation motivates internalizing and strengthening self-reflective reasoning capabilities within LLMs to achieve a better safety–utility trade-off. We propose Safety-aware Reflective Reasoning Optimization (SaRO), a two-stage framework: (1) Reasoning-style Warmup (RW) to internalize self-reflective reasoning, and (2) Self-reflective Reasoning Process Optimization (SRPO) to encourage reflection and correction. Experiments show that SaRO outperforms existing reasoning-based alignment methods, achieving a better balance of safety and helpfulness.

Henry Peng Zou, Wei-Chieh Huang, Yaozu Wu, Jizhou Guo, Yankai Chen, Chunyu Miao, Hoang H Nguyen, Yue Zhou, Weizhi Zhang, Liancheng Fang 等

Recent advances in large language models (LLMs) have sparked growing interest in building fully autonomous agents. However, fully autonomous LLM-based agents still face significant challenges, including limited reliability due to hallucinations, difficulty in handling complex tasks, and substantial safety and ethical risks, all of which limit their feasibility and trustworthiness in real-world applications. To overcome these limitations, LLM-based human-agent systems (LLM-HAS) incorporate human-provided information, feedback, or control into the agent system to enhance system performance, reliability, and safety. These human-agent collaboration systems enable humans and LLM-based agents to collaborate effectively by leveraging their complementary strengths.This paper provides the first comprehensive and structured survey of LLM-HAS. It clarifies fundamental concepts, systematically presents core components shaping these systems, including environment and profiling, human feedback, interaction types, orchestration, and communication, explores emerging applications, and discusses unique challenges and opportunities arising from human-AI collaboration. By consolidating current knowledge and offering a structured overview, we aim to foster further research and innovation in this rapidly evolving interdisciplinary field. Paper lists and resources are available at https://github.com/HenryPengZou/Awesome-Human-Agent-Collaboration-Interaction-Systems.

Xiyuan Zhou, Xinlei Wang, Yirui He, Ruixi Zou, Yang Wu, Yuheng Cheng, Yulu Xie, Wenxuan Liu, Huan Zhao, Yan Xu 等

Large language models (LLMs) have shown strong performance on mathematical reasoning under well-defined conditions. However, real-world engineering problems involve uncertainty, context, and open-ended settings that extend beyond symbolic computation. Existing benchmarks largely focus on well-defined or abstract reasoning and therefore fail to capture these complexities. We introduce EngiBench, a hierarchical benchmark designed to evaluate LLMs on solving engineering problems. It spans three levels of increasing difficulty (foundational knowledge retrieval, contextual reasoning, and open-ended modeling) and covers diverse engineering subfields. To facilitate a deeper understanding of model performance, we systematically rewrite each problem into three controlled variants (perturbed, knowledge-enhanced, and math abstraction), enabling us to separately evaluate the model’s robustness, domain-specific knowledge, and mathematical reasoning abilities. Experimental results show clear performance stratification across difficulty levels: model accuracy declines with task complexity, degrades under minor perturbations, and remains substantially below human performance on high-level engineering tasks. These findings reveal that current LLMs still lack the high-level reasoning needed for real-world engineering, highlighting the need for future models with deeper and more reliable problem-solving capabilities. Our source code and data are available at https://github.com/AI4Engi/EngiBench.

Sungkyun Kim, Jaemin Kim, Dogyeong Yun, Jiho Shin, Junyeol Lee, Jiwon Seo

Speculative decoding (SD) improves LLM inference latency by speculatively generating multiple tokens with a small draft model and verifying them with a larger target model. However, when speculation accuracy is low, the overhead from rejected tokens can negate its benefits, especially at large batch sizes.We propose Speculative Verification (SV), an efficient augmentation to SD that predicts speculation accuracy and dynamically adapts the verification length to maximize throughput. SV introduces a small companion model, similar in size to draft model, to reduce uncertainty in speculation accuracy. By exploiting the information gain from observing the companion distribution, SV reduces wasted verification on rejected tokens and improves decoding efficiency.We evaluate SV across publicly available LLMs on seven NLP tasks using over a hundred combinations of draft, companion, and target models, including 13B–72B target models spanning base, instruction-tuned, and task-specific fine-tuned variants. Compared to target-only decoding, standard SD, and state-of-the-art SD variants, SV consistently delivers higher throughput across batch sizes. SV improves SD performance by up to 1.9\times, with an average 1.4\times speedup at large batch sizes, showing robust and scalable gains for practical LLM inference.

Kehua Feng, Keyan Ding, Yuhao Wang, Menghan Li, Fanjunduo Wei, Xinda Wang, Huajun Chen

Recent advancements in large language models (LLMs) have accelerated progress toward artificial general intelligence, yet their potential to generate harmful content poses critical safety challenges. Existing alignment methods often struggle to cover diverse safety scenarios and remain vulnerable to adversarial attacks. In this work, we propose **SAFER**, a framework for **S**afety **A**lignment via e**F**ficient **E**x-Ante **R**easoning. Our approach instantiates structured Ex-Ante reasoning through initial assessment, rule verification, and path calibration, and embeds predefined safety rules to provide transparent and verifiable safety judgments. Specifically, our approach consists of two training stages: (1) supervised fine-tuning with synthetic traces to teach the multi-stage Ex-Ante reasoning, and (2) step-level reasoning preference optimization to jointly enhance safety, utility, and efficiency. Experiments on multiple open-source LLMs demonstrate that SAFER significantly enhances safety performance while maintaining helpfulness and response efficiency.

Christabel Acquaye, Yi Ting Huang, Marine Carpuat, Rachel Rudinger

Standardized math assessments require expensive human pilot studies to establish the difficulty of test items. We investigate the predictive value of open-source large language models (LLMs) for evaluating the difficulty of multiple-choice math questions for real-world students. We show that, while LLMs are poor direct judges of problem difficulty, simulation-based approaches with LLMs yield promising results under the right conditions. Under the proposed approach, we simulate a "classroom" of 4th, 8th, or 12th grade students by prompting the LLM to role-play students of varying proficiency levels. We use the outcomes of these simulations to fit Item Response Theory (IRT) models, comparing learned difficulty parameters for items to their real-world difficulties, as determined by item-level statistics furnished by the National Assessment of Educational Progress (NAEP). We observe correlations as high as 0.75, 0.76, and 0.82 for grades 4, 8, and 12, respectively. In our simulations, we experiment with different "classroom sizes," showing tradeoffs between computation size and accuracy. We find that role-plays with named students improves predictions (compared to student ids), and stratifying names across gender and race further improves predictions. Our results show that LLMs with relatively weaker mathematical abilities (Gemma) actually yield better real-world difficulty predictions than mathematically stronger models (Llama and Qwen), further underscoring the suitability of open-source models for the task.

Chang Hong, Minghao Wu, Qingying Xiao, Yuchi Wang, Xiang Wan, Guangjun Yu, Benyou Wang, Yan Hu

As medical LLMs transition to clinical deployment, assessing their ethical reasoning capability becomes critical. While achieving high accuracy on knowledge benchmarks, LLMs lack validated assessment for navigating ethical trade-offs in clinical decision-making where multiple valid solutions exist. Existing benchmarks lack systematic approaches to incorporate recognized philosophical frameworks and expert validation for ethical reasoning assessment. We introduce PrinciplismQA, a philosophy-grounded approach to assessing LLM clinical medical ethics alignment. Grounded in Principlism, our approach provides a systematic methodology for incorporating clinical ethics philosophy into LLM assessment design. PrinciplismQA comprises 3,648 expert-validated questions spanning knowledge assessment and clinical reasoning. Our expert-calibrated pipeline enables reproducible evaluation and models ethical biases. Evaluating recent models reveals significant ethical reasoning gaps despite high knowledge accuracy, demonstrating that knowledge-oriented training does not ensure clinical ethical alignment. PrinciplismQA provides a validated tool for assessing clinical AI deployment readiness.

Taiki Sekii, Fumiaki Sato

Recent action recognition based on vision–language pretraining and self-supervised video foundation models tends to induce spurious correlations and shortcut learning by relying on action-irrelevant cues, such as backgrounds and object co-occurrences. By contrast, object-detection-based approaches can suppress spurious correlations; however, the loss of input information can limit accuracy. To mitigate this trade-off, we combine these two approaches to learn complementary features that compensate for each other’s shortcomings. Specifically, we leverage the commonsense knowledge of large language models (LLMs) regarding human actions and realize a framework in which an LLM agent integrates the two approaches within an agentic learning paradigm to design motion features tailored to the target actions. The LLM agent uses an open vocabulary object detector to instruct the video foundation model with the target and nontarget objects in the video to make the model attend to objects in a video required for recognizing the target actions. The composition of the detected objects is optimized for the target actions through in-context reinforcement learning (ICRL) using the commonsense knowledge of the LLM. Experiments on multiple public action recognition datasets and an ablation study confirm the robustness of features learned using the proposed method and the effectiveness of ICRL.

Yixuan Liu, Yicheng Zhang

The Science of Science (SciSci) examines how scientific knowledge is generated, evaluated, and transformed by utilizing large-scale scholarly and bibliometric data. As these data grow in scale and complexity, analysis has increasingly relied on statistical, network-based, machine learning methods, and is now seeing growing involvement of AI agents. This emerging class of such agents, ranging from multi-agent simulations of scientific behavior to tool-augmented systems for empirical analysis, is beginning to reshape how SciSci research is conducted. In this survey, we propose a task-centered taxonomy, distinguishing *agents as simulations*, which model citation, collaboration, and community dynamics, from *agents as tools*, which assist empirical analysis and scientific workflows. We review agent architectures, learning mechanisms, evaluation, and SciSci benchmarks, and examine open challenges related to reliability, data quality, and bias. Our survey aims to clarify the landscape of AI agents in SciSci and to support the development of reliable and scientifically useful AI systems for studying science and scientific communities.