大模型推理能力 - arXivDaily 专题

2401.10005 2024-07-19 cs.CV cs.CL 83%

Advancing Large Multi-modal Models with Explicit Chain-of-Reasoning and Visual Question Generation

Kohei Uehara, Nabarun Goswami, Hanqin Wang, Toshiaki Baba, Kohtaro Tanaka, Tomohiro Hashimoto, Kai Wang, Rei Ito, Takagi Naoya, Ryo Umagami, Yingyi Wen, Tanachai Anakewat, Tatsuya Harada

专题命中推理评测：reasoning(title,abstract)；chain-of-thought(abstract)；分类 cs.CL

2406.06461 2024-06-18 cs.CL 83%

Reasoning in Token Economies: Budget-Aware Evaluation of LLM Reasoning Strategies

Junlin Wang, Siddhartha Jain, Dejiao Zhang, Baishakhi Ray, Varun Kumar, Ben Athiwaratkun

专题命中推理评测：reasoning(title,abstract)；chain-of-thought(abstract)；分类 cs.CL

2406.09103 2024-06-14 cs.CL 83%

Chain-of-Though (CoT) prompting strategies for medical error detection and correction

Zhaolong Wu, Abul Hasan, Jinge Wu, Yunsoo Kim, Jason P. Y. Cheung, Teng Zhang, Honghan Wu

专题命中推理评测：CoT(title,abstract)；chain-of-thought(abstract)；分类 cs.CL

Comments accepted as NAACL workshop

2403.03167 2024-06-07 cs.CL 83%

PARADISE: Evaluating Implicit Planning Skills of Language Models with Procedural Warnings and Tips Dataset

Arda Uzunoglu, Abdalfatah Rashid Safa, Gözde Gül Şahin

专题命中推理评测：planning(title,abstract)；reasoning(abstract)；分类 cs.CL

Comments 9 pages, ACL 2024 Findings

2405.16064 2024-05-28 cs.CL 83%

Keypoint-based Progressive Chain-of-Thought Distillation for LLMs

Kaituo Feng, Changsheng Li, Xiaolu Zhang, Jun Zhou, Ye Yuan, Guoren Wang

专题命中推理评测：chain-of-thought(title,abstract)；reasoning(abstract)；分类 cs.CL

Comments Accepted by ICML 2024

2403.10900 2024-03-19 cs.CL 83%

BEnQA: A Question Answering and Reasoning Benchmark for Bengali and English

Sheikh Shafayat, H M Quamran Hasan, Minhajur Rahman Chowdhury Mahim, Rifki Afina Putri, James Thorne, Alice Oh

专题命中推理评测：reasoning(title,abstract)；chain-of-thought(abstract)；分类 cs.CL

2312.07122 2023-12-13 cs.AI 83%

Neural Reasoning About Agents' Goals, Preferences, and Actions

Matteo Bortoletto, Lei Shi, Andreas Bulling

专题命中推理评测：reasoning(title,abstract)；logical reasoning(abstract)；分类 cs.AI

Comments The 38th Annual AAAI Conference on Artificial Intelligence (AAAI-24)

2311.09721 2023-11-17 cs.CL 83%

On Evaluating the Integration of Reasoning and Action in LLM Agents with Database Question Answering

Linyong Nan, Ellen Zhang, Weijin Zou, Yilun Zhao, Wenfei Zhou, Arman Cohan

专题命中推理评测：reasoning(title,abstract)；planning(abstract)；分类 cs.CL

2311.09277 2023-11-17 cs.CL 83%

Contrastive Chain-of-Thought Prompting

Yew Ken Chia, Guizhen Chen, Luu Anh Tuan, Soujanya Poria, Lidong Bing

专题命中推理评测：chain-of-thought(title,abstract)；reasoning(abstract)；分类 cs.CL

2305.13903 2023-11-10 cs.CL cs.CV 83%

Let's Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought

Vaishnavi Himakunthala, Andy Ouyang, Daniel Rose, Ryan He, Alex Mei, Yujie Lu, Chinmay Sonar, Michael Saxon, William Yang Wang

专题命中推理评测：chain-of-thought(title,abstract)；reasoning(abstract)；分类 cs.CL

Comments Accepted to the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP 2023)

2310.20256 2023-11-07 cs.CL 83%

PsyCoT: Psychological Questionnaire as Powerful Chain-of-Thought for Personality Detection

Tao Yang, Tianyuan Shi, Fanqi Wan, Xiaojun Quan, Qifan Wang, Bingzhe Wu, Jiaxiang Wu

专题命中推理评测：chain-of-thought(title,abstract)；CoT(abstract)；分类 cs.CL

Comments Accepted to Findings of EMNLP 2023

2306.00790 2023-06-02 cs.AI cs.LO cs.MA 83%

Knowledge-based Reasoning and Learning under Partial Observability in Ad Hoc Teamwork

Hasra Dodampegama, Mohan Sridharan

专题命中推理评测：reasoning(title,abstract)；logical reasoning(abstract)；分类 cs.AI

Comments 17 pages, 3 Figures

2301.10896 2023-02-17 cs.CL 83%

Causal Reasoning of Entities and Events in Procedural Texts

Li Zhang, Hainiu Xu, Yue Yang, Shuyan Zhou, Weiqiu You, Manni Arora, Chris Callison-Burch

专题命中推理评测：reasoning(title,abstract)；chain-of-thought(abstract)；分类 cs.CL

Comments In Findings of EACL 2023

2604.05114 2026-04-08 cs.CL cs.AI cs.LG 83%

$π^2$: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language Models

$π^2$：结构起源的推理数据提升大语言模型的长上下文推理能力

Quyet V. Do, Thinh Pham, Nguyen Nguyen, Sha Li, Pratibha Zunjare, Tu Vu

机构 * Virginia Tech（弗吉尼亚理工大学）

专题命中推理评测：reasoning(title,abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结本文提出$π^2$方法，通过结构化数据生成高质量推理数据，提升大语言模型的长上下文推理能力，在四个基准测试中取得显著改进。

Comments Our structured analytical reasoning data, which originates from Wikipedia tables, significantly improves long-context reasoning capability of LLMs

2608.09432 2026-08-11 cs.CL cs.AI 新提交 82%

ZetaGPT: A Reference Implementation of Positional--Encoding--Free State--Space--Attention Language Models

ZetaGPT：无位置编码的状态空间注意力语言模型的参考实现

Róisín Luo

专题命中推理评测：CoT(abstract,abstract_cn)；reasoning(abstract)；chain-of-thought(abstract)；分类 cs.CL、cs.AI

AI总结 ZetaGPT是首款无显式位置编码的开源小型语言模型，它将因果状态空间方程与自注意力结合，提供全开源训练流水线，为无位置编码语言模型的研究提供紧凑可复现的参考实现。

2606.22496 2026-06-23 cs.PF cs.AI cs.LG 新提交 82%

Enabling Cloud-Level Accuracy in Edge AI through IoT Data Preprocessing

通过物联网数据预处理实现边缘AI的云端级精度

Aygün Varol, Katarzyna Kołodziej, Łukasz Sobczak, Michał Romaszewski, Przemysław Głomb, Naser Hossein Motlagh, Mirka Leino, Johanna Virkki

机构 * Faculty of Information Technology and Communication Sciences, Tampere University（信息科技与通信科学学院，塔尔皮耶大学）； Institute of Theoretical and Applied Informatics, Polish Academy of Sciences（理论与应用信息学研究所，波兰科学院）； Department of Computer Science, University of Helsinki（计算机科学系，赫尔辛基大学）； Satakunta University of Applied Sciences（萨塔昆塔应用科学大学）

专题命中推理评测：CoT(abstract,abstract_cn)；reasoning(abstract)；chain-of-thought(abstract)；分类 cs.AI、cs.LG

AI总结针对边缘部署的紧凑型LLM在原始传感器读数下数值推理弱的问题，提出结构化提示构建框架，通过逐步丰富文本表示（原始值、阈值描述、环境摘要标志）提升局部模型精度，实验表明该方法可将局部准确率从约50%提升至80%以上，同时保持低延迟。

2403.18957 2026-06-17 cs.CY cs.CL cs.LG cs.SI 版本更新 82%

Moderating Illicit Online Image Promotion for Unsafe User-Generated Content Games Using Large Vision-Language Models

使用大型视觉语言模型审核不安全的用户生成内容游戏中的非法在线图像推广

Keyan Guo, Ayush Utkarsh, Wenbo Ding, Isabelle Ondracek, Ziming Zhao, Guo Freeman, Nishant Vishwamitra, Hongxin Hu

机构 * University at Buffalo（布法罗大学）； Northeastern University（东北大学）； Clemson University（克莱姆森大学）； The University of Texas at San Antonio（德克萨斯大学圣安东尼奥分校）

专题命中推理评测：CoT(abstract,abstract_cn)；reasoning(abstract)；chain-of-thought(abstract)；分类 cs.CL、cs.LG

AI总结针对社交媒体上非法推广不安全UGC游戏的图像，提出UGCG-Guard系统，利用大型视觉语言模型和条件提示策略实现零样本域适应，检测准确率达94%。

Comments In Proceedings of the 33rd USENIX Conference on Security Symposium (SEC '24), August 14-16, 2024

2603.17145 2026-06-01 cs.LG cs.AI 82%

REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge

REAL: 面向LLM评判的回归感知强化学习

Yasi Zhang, Tianyu Chen, Mingyuan Zhou, Oscar Leong, Ying Nian Wu, Michal Lukasik

机构 * University of California, Los Angeles（加州大学洛杉矶分校）； The University of Texas at Austin（得克萨斯大学奥斯汀分校）； Google Research Now at Google DeepMind（谷歌研究现在在谷歌深Mind）

专题命中推理评测：CoT(abstract,abstract_cn)；reasoning(abstract)；chain-of-thought(abstract)；分类 cs.AI、cs.LG

AI总结提出REAL框架，通过广义策略梯度将回归目标融入强化学习，优化LLM作为评分器的数值评估，在多个规模模型上超越SFT和标准RL方法。

Comments Accepted to ICML 2026. The first two authors contributed equally

2501.18099 2025-07-09 cs.AI cs.CL 82%

Learning to Plan & Reason for Evaluation with Thinking-LLM-as-a-Judge

Swarnadeep Saha, Xian Li, Marjan Ghazvininejad, Jason Weston, Tianlu Wang

机构 * FAIR at Meta（Meta 的 FAIR）

专题命中推理评测：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；planning(abstract)

Comments ICML 2025

2608.21032 2026-08-24 cs.RO 新提交 82%

Roadside-Cooperative Autonomous Driving: From Data Platform to Vision-Language End-to-End Reasoning

路侧协同自动驾驶：从数据平台到视觉-语言端到端推理

Yitao Xu, Tong Wu, Yiyan Wu, Guoji Xu, Yanbo Jiang, Jiahao Wang, Zehong Ke, Junkai Jiang, Fang Zhang, Jianqiang Wang

专题命中推理评测：reasoning(title,abstract)；planning(abstract)

AI总结针对现有V2X基准的局限，推出V2XBench平台与Chat-V2XBench数据集，提出AURORA端到端协同驾驶框架，其在严重遮挡场景下路径完成率达98.21%、驾驶得分76.02，开创了可扩展的V2X-VLM范式。

Comments 15 pages, 5 figures, under review

2608.12391 2026-08-14 cs.CL cs.AI cs.LG 新提交 82%

Unified Multi-Dimensional Benchmark for Complex Graph Reasoning in Large Language Models

面向大语言模型复杂图推理的统一多维度基准

Fali Wang, Ali Al-Lawati, Iliyas Bektas, Jinxuan Fang, Alek Melenski, Tianxiang Zhao, Yao Ma, Suhang Wang

机构 * The Pennsylvania State University（宾夕法尼亚州立大学）； Rensselaer Polytechnic Institute（伦斯勒理工学院）

专题命中推理评测：reasoning(title,abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结针对现有图推理基准的局限，提出半自动框架构建含202个任务的统一多维度基准，评估LLM在不同推理模式下的表现，揭示模型局限并提供实证指导。

Comments Under review

2608.04434 2026-08-06 cs.CV 新提交 82%

OmniRouting: A Semantic-Coupled Multimodal Benchmark for Constraint-Aware Spatial Reasoning in PCB Routing

OmniRouting：面向PCB布线中约束感知空间推理的语义耦合多模态基准

Taiting Lu, Kaiyuan Lin, Ziwei Dong, Sisong Bei, Haolin Ye, Yuxin Tian, Runze Liu, Mingjia Wang, Jingying Zeng, Hongxing Pan, Kai Zhang, Haoyu Wang, Guoliang Shi, Ling Ma, Yifan Yang, Jiaying Lu, Qi He, Yi-Chao Chen, Sung-Liang Chen, Yincheng Jin, Mahanth Gowda

专题命中推理评测：reasoning(title,abstract)；planning(abstract)

AI总结本研究针对LLMs在PCB布线推理上的能力缺口，构建了OmniRouting基准，含1681个工业级PCB设计及四项任务，发现现有LMMs存在显著局限并计划开源相关资源。

2608.00200 2026-08-04 cs.AI cs.CL cs.ET cs.LG 新提交 82%

TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding

TRACE-TS：面向人类活动理解的归因基础且可追踪的传感器-语言推理

Sparsh Rastogi, Tanmay Kumar, Baiyu Chen, Jatin Bedi, Zechen Li, Flora D. Salim

专题命中推理评测：reasoning(title,abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结针对现有传感器-语言推理与信号关联弱、解释不可靠的问题，提出TRACE-TS框架，通过专家归因识别关键传感器区域并构建可追踪推理轨迹，在7个可穿戴基准上实现最优性能。

Comments 24 pages, 9 figures, 24 tables

2607.09786 2026-08-04 cs.AI cs.CL cs.LG 版本更新 82%

Length Penalties Make Chain-of-Thought Less Monitorable

长度惩罚使思维链更难被监测

Bryce Little

专题命中推理评测：chain-of-thought(title)；reasoning(abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结研究发现长度惩罚强化学习会缩短思维链推理，隐藏驱动模型答案的因素。通过训练不同目标链长度的Qwen3模型并评估，表明压缩虽能减少推理令牌、保留准确率，但降低了可监测性，揭示了压缩与可监测性的前沿关系。

2607.19678 2026-07-23 cs.CL cs.AI cs.LG 新提交 82%

Reference-Free Evaluation of Reasoning in Open-Ended Question Answering

开放式问答中推理的无参考评估

Guneet Singh Kohli, Yuxiang Zhou, Michael Sejr Schlichtkrull, Gregory E Dean, Maria Liakata

机构 * Queen Mary University of London（伦敦大学玛丽皇后学院）； Temple University（天普大学）

专题命中推理评测：reasoning(title,abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结针对高风险领域人工智能生成答案难验证问题，提出基于推理的无参考框架审核大语言模型输出，通过分解推理轨迹、标记关系并组织成超图等方法评估，在数学和医学推理设置下比直接以大语言模型为基线更可靠，强调问答评估应考虑推理关系组合。

2607.13439 2026-07-16 cs.CR cs.SE 新提交 82%

DREA: Decoupled Reasoning and Exploration Agents for Repository-Level Vulnerability Detection

DREA：用于仓库级漏洞检测的解耦推理与探索代理

Mingyang Sun, Guozhu Meng

专题命中推理评测：reasoning(title,abstract)；planning(abstract)

AI总结研究针对大语言模型用于仓库级漏洞检测时的不足，提出DREA框架，通过规划与探索代理解耦，实现目标导向上下文获取。构建RepoPairBench评估，提升配对正确性，降低成本，还发现安全推理质量是当前大语言模型的瓶颈。

Comments Accepted for presentation at Internetware 2026. 12 pages, 4 figures, and 5 tables

2607.13081 2026-07-16 cs.CR cs.AI cs.CL cs.LG 新提交 82%

SingGuard-NSFA: Extensible Guardrails for Agentic AI via Generative Reasoning and Real-Time Classification

SingGuard-NSFA：通过生成式推理和实时分类实现的可扩展智能体人工智能护栏

SingGuard Team

机构 * Ant Group（蚂蚁集团）

专题命中推理评测：reasoning(title,abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结研究针对智能体人工智能系统操作威胁，提出NSFA分类法构建基准套件，开发双模式方法，发布四个不同参数模型，在基准测试和跨源评估中表现出色，证明方法可扩展性与通用性。

2605.24902 2026-07-16 cs.CL cs.AI cs.LG 版本更新 82%

When Reasoning Hurts: Source-Aware Evaluation of Frontier LLMs for Clinical SOAP Note Generation

当推理有害：面向临床SOAP笔记生成的前沿LLM源感知评估

Faizan Faisal

机构 * University of California, Davis（加州大学戴维斯分校）

专题命中推理评测：reasoning(title,abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结通过源感知基准测试，评估推理增强型LLM在临床SOAP笔记生成中的表现，发现推理能力反而降低GPT-5.4的质量，而相同源RAG带来模型依赖的小幅提升。

2607.12497 2026-07-15 cs.CV 新提交 82%

TerraLogic: A Benchmark for Hierarchical Geospatial Reasoning in Earth Observation

TerraLogic：地球观测中分层地理空间推理的基准

Yuhang Yan, Linchao Mou, Bokang Yang, Qingyu Li

机构 * The Chinese University of Hong Kong (Shenzhen)（香港中文大学（深圳））； Technical University of Munich（慕尼黑工业大学）

专题命中推理评测：reasoning(title,abstract)；planning(abstract)

AI总结针对遥感中认知地理空间推理研究不足的问题，引入TerraLogic基准及HieraPlan工具增强智能体，通过545个任务推动评估向认知级发展，实验显示HieraPlan在分层地理空间推理上表现出色，为相关研究提供强大基线。

Comments 8 pages, 3 figures, and 7 tables. Dataset and agent code are available at https://github.com/Ireliya/TerraLogic

2604.21334 2026-07-14 cs.AI cs.CE cs.CL cs.LG econ.GN q-fin.EC 版本更新 82%

Ideological Bias in LLMs' Economic Causal Reasoning

大语言模型在经济因果推理中的意识形态偏见

Donggyu Lee, Hyeok Yun, Jungwon Kim, Junsik Min, Sungwon Park, Sangyoon Park, Jihee Kim

机构 * Graduate School of Data Science, KAIST（韩国科学技术院数据科学研究生院）； College of Business, KAIST（韩国科学技术院商学院）； School of Computing, KAIST（韩国科学技术院计算机学院）； Division of Social Science, HKUST（香港科技大学社会科学系）

专题命中推理评测：reasoning(title,abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结研究探讨LLM在经济因果推理中的意识形态偏见，通过扩展EconCausal基准测试，评估20种先进LLM在处理意识形态冲突案例时的准确性，发现LLM在意识形态冲突案例上表现更差，且存在系统性偏倚。

Comments Accepted at COLM 2026

1. 推理评测 10472 篇

Advancing Large Multi-modal Models with Explicit Chain-of-Reasoning and Visual Question Generation

Reasoning in Token Economies: Budget-Aware Evaluation of LLM Reasoning Strategies

Chain-of-Though (CoT) prompting strategies for medical error detection and correction

PARADISE: Evaluating Implicit Planning Skills of Language Models with Procedural Warnings and Tips Dataset

Keypoint-based Progressive Chain-of-Thought Distillation for LLMs

BEnQA: A Question Answering and Reasoning Benchmark for Bengali and English

Neural Reasoning About Agents' Goals, Preferences, and Actions

On Evaluating the Integration of Reasoning and Action in LLM Agents with Database Question Answering

Contrastive Chain-of-Thought Prompting

Let's Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought

PsyCoT: Psychological Questionnaire as Powerful Chain-of-Thought for Personality Detection

Knowledge-based Reasoning and Learning under Partial Observability in Ad Hoc Teamwork

Causal Reasoning of Entities and Events in Procedural Texts

$π^2$: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language Models

ZetaGPT: A Reference Implementation of Positional--Encoding--Free State--Space--Attention Language Models

Enabling Cloud-Level Accuracy in Edge AI through IoT Data Preprocessing

Moderating Illicit Online Image Promotion for Unsafe User-Generated Content Games Using Large Vision-Language Models

REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge

Learning to Plan & Reason for Evaluation with Thinking-LLM-as-a-Judge

Roadside-Cooperative Autonomous Driving: From Data Platform to Vision-Language End-to-End Reasoning

Unified Multi-Dimensional Benchmark for Complex Graph Reasoning in Large Language Models

OmniRouting: A Semantic-Coupled Multimodal Benchmark for Constraint-Aware Spatial Reasoning in PCB Routing

TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding

Length Penalties Make Chain-of-Thought Less Monitorable

Reference-Free Evaluation of Reasoning in Open-Ended Question Answering

DREA: Decoupled Reasoning and Exploration Agents for Repository-Level Vulnerability Detection

SingGuard-NSFA: Extensible Guardrails for Agentic AI via Generative Reasoning and Real-Time Classification

When Reasoning Hurts: Source-Aware Evaluation of Frontier LLMs for Clinical SOAP Note Generation

TerraLogic: A Benchmark for Hierarchical Geospatial Reasoning in Earth Observation

Ideological Bias in LLMs' Economic Causal Reasoning