大模型推理能力

Schema Key Wording as an Instruction Channel in Structured Generation under Constrained Decoding

模式键作为在受限解码下的指令通道在结构生成中的应用

Yifan Le

机构 * Zhejiang University（浙江大学） 纠错

专题命中 数学推理 ：CoT(abstract,abstract_cn)；reasoning(abstract)；分类 cs.CL、cs.AI

AI总结 本文探讨了在受限解码中模式键作为隐式指令通道的作用，提出结构生成是多通道指令问题，并通过实验展示模式键词汇对准确性的影响，揭示了不同模型对不同通道的依赖性。

Comments 11 pages, 3 figures

DeepPrune: Parallel Scaling without Inter-trace Redundancy

DeepPrune: 无需跨轨迹冗余的并行扩展

Shangqing Tu, Yaxuan Li, Yushi Bai, Lei Hou, Juanzi Li

机构 * Tsinghua University（清华大学） ； ShanghaiTech University（上海科技大学） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

AI总结 本文提出DeepPrune框架，通过动态剪枝解决并行推理中的冗余问题，实现65.73%-88.50%的token减少，保持高精度。

Comments Accepted by ACL 2026 Findings, please check out the project page: https://deepprune.github.io/

A Closer Look into LLMs for Table Understanding

深入探究用于表格理解的大型语言模型

Jia Wang, Chuanyu Qin, Mingyu Zheng, Qingyi Si, Peize Li, Zheng Lin

机构 * Institute of Information Engineering, Chinese Academy of Sciences（中国科学院信息工程研究所） ； School of Cyber Security, University of Chinese Academy of Sciences（中国科学院大学网络安全学院） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；math reasoning(abstract)；分类 cs.CL、cs.AI

AI总结 本文通过研究16种LLM，探讨其理解表格数据和执行下游任务的机制，发现注意力动态、有效层数、专家激活及输入设计影响显著。

Not all tokens are needed(NAT): token efficient reinforcement learning

并非所有标记都必要（NAT）：标记效率强化学习

Hejian Sang, Yuanda Xu, Zhengze Zhou, Ran He, Zhipeng Wang

机构 * LinkedIn Corporation（LinkedIn公司） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.AI、cs.LG

AI总结 NAT通过局部标记采样提升RL效率，使用50%标记实现与完整标记相同性能。

Hallucination-Free Automatic Question & Answer Generation for Intuitive Learning

无幻觉的自动问答生成用于直观学习

Nicholas X. Wang, Aggelos K. Katsaggelos

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

AI总结 本文提出无幻觉的多代理生成框架，通过结构化协作减少教育内容中的幻觉，提升问答生成的准确性与教育价值。

Post-Training Quantization of OpenPangu Models for Efficient Deployment on Atlas A2

在Atlas A2上对OpenPangu模型进行训练后量化以实现高效部署

Yilun Luo, Huaqing Zheng, Haoqian Meng, Wenyuan Liu, Peng Zhang

机构 * School of Computer Science and Technology, Tianjin University（计算机科学与技术学院，天津大学） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.AI、cs.LG

AI总结 本文提出在Atlas A2上通过低比特量化提升OpenPangu模型的推理效率，实现高效CoT推理并保持高精度。

Ladder Up, Memory Down: Low-Cost Fine-Tuning With Side Nets

向上爬，记忆下降：低成本微调与侧边网络

Estelle Zheng, Nathan Cerisara, Sébastien Warichet, Emmanuel Helbert, Christophe Cerisara

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.LG

AI总结 LST通过轻量级侧边网络实现高效微调，比QLoRA更节省内存，同时在多个任务中保持竞争力。

GEM: Generative Entropy-Guided Preference Modeling for Few-shot Alignment of LLMs

Yiyang Zhao, Huiyu Bai, Xuejiao Zhao

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.AI、cs.LG

Comments This paper has been accepted by AAAI 2026-AIA and designated as an oral presentation paper

Distilling LLM Agent into Small Models with Retrieval and Code Tools

Minki Kang, Jongwon Jeong, Seanie Lee, Jaewoong Cho, Sung Ju Hwang

机构 * KAIST（韩国科学技术院） ； University of Wisconsin-Madison（威斯康星大学麦迪逊分校） ； KRAFTON（KRAFTON公司） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

Comments NeurIPS 2025 Spotlight

WebAgent-R1: Training Web Agents via End-to-End Multi-Turn Reinforcement Learning

Zhepei Wei, Wenlin Yao, Yao Liu, Weizhi Zhang, Qin Lu, Liang Qiu, Changlong Yu, Puyang Xu, Chao Zhang, Bing Yin, Hyokun Yun, Lihong Li

机构 * University of Virginia（弗吉尼亚大学） ； Amazon（亚马逊） ； Georgia Institute of Technology（佐治亚理工学院） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.LG

Comments EMNLP 2025. Code: https://github.com/weizhepei/WebAgent-R1

No Need for Explanations: LLMs can implicitly learn from mistakes in-context

Lisa Alazraki, Maximilian Mozes, Jon Ander Campos, Tan Yi-Chern, Marek Rei, Max Bartolo

机构 * Imperial College London（伦敦帝国学院） ； Cohere 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；math reasoning(abstract)；分类 cs.CL、cs.AI

Comments EMNLP 2025

How Do Humans Write Code? Large Models Do It the Same Way Too

Long Li, Xuzheng He, Haozhe Wang, Linlin Wang, Liang He

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

Forgotten Polygons: Multimodal Large Language Models are Shape-Blind

William Rudman, Michal Golovanevsky, Amir Bar, Vedant Palit, Yann LeCun, Carsten Eickhoff, Ritambhara Singh

机构 * Brown University（布朗大学） ； Tel Aviv University（特拉维夫大学） ； IIT Kharagpur（印度理工学院卡里帕尔分校） ； New York University（纽约大学） ； University of Tübingen（图宾根大学） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

Probing for Arithmetic Errors in Language Models

Yucheng Sun, Alessandro Stolfo, Mrinmaya Sachan

机构 * ETH Zürich（苏黎世联邦理工学院） 纠错

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；self-correction(abstract)；分类 cs.CL、cs.AI

Teaching LLM to Reason: Reinforcement Learning from Algorithmic Problems without Code

Keqin Bao, Nuo Chen, Xiaoyuan Li, Binyuan Hui, Bowen Yu, Fuli Feng, Xiangnan He, Dayiheng Liu

专题命中 数学推理 ：reasoning(abstract)；CoT(abstract)；logical reasoning(abstract)；分类 cs.CL、cs.LG

Achieving >97% on GSM8K: Deeply Understanding the Problems Makes LLMs Better Solvers for Math Word Problems

Qihuang Zhong, Kang Wang, Ziyang Xu, Juhua Liu, Liang Ding, Bo Du

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

Comments The article has been accepted by Frontiers of Computer Science (FCS), with the DOI: { 10.1007/s11704-025-41102-z }

Thinking Slow, Fast: Scaling Inference Compute with Distilled Reasoners

Daniele Paliotta, Junxiong Wang, Matteo Pagliardini, Kevin Y. Li, Aviv Bick, J. Zico Kolter, Albert Gu, François Fleuret, Tri Dao

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

S$^2$R: Teaching LLMs to Self-verify and Self-correct via Reinforcement Learning

Ruotian Ma, Peisong Wang, Cheng Liu, Xingyan Liu, Jiaqi Chen, Bang Zhang, Xin Zhou, Nan Du, Jia Li

专题命中 数学推理 ：reasoning(abstract)；CoT(abstract)；self-correction(abstract)；分类 cs.CL、cs.LG

SketchFill: Sketch-Guided Code Generation for Imputing Derived Missing Values

Yunfan Zhang, Changlun Li, Yuyu Luo, Nan Tang

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.LG

Comments 19 pages, 6 figures

Prompt Baking

Aman Bhargava, Cameron Witkowski, Alexander Detkov, Matt Thomson

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；math reasoning(abstract)；分类 cs.CL、cs.AI

Comments 25 pages, 8 figures

SarcasmBench: Towards Evaluating Large Language Models on Sarcasm Understanding

Yazhou Zhang, Chunwang Zou, Zheng Lian, Prayag Tiwari, Jing Qin

专题命中 数学推理 ：reasoning(abstract)；CoT(abstract)；logical reasoning(abstract)；分类 cs.CL、cs.AI

Auto-Regressive Next-Token Predictors are Universal Learners

Eran Malach

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.LG

Are Human-generated Demonstrations Necessary for In-context Learning?

Rui Li, Guoyin Wang, Jiwei Li

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.AI、cs.LG

Towards Revealing the Mystery behind Chain of Thought: A Theoretical Perspective

Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, Liwei Wang

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.LG

Comments 42 pages; Camera-ready version for NeurIPS 2023 (Oral Presentation)

Large Language Models are Better Reasoners with Self-Verification

Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Shengping Liu, Bin Sun, Kang Liu, Jun Zhao

专题命中 数学推理 ：reasoning(abstract)；CoT(abstract)；logical reasoning(abstract)；分类 cs.CL、cs.AI

Comments Accept in EMNLP 2023 Findings

SatLM: Satisfiability-Aided Language Models Using Declarative Prompting

Xi Ye, Qiaochu Chen, Isil Dillig, Greg Durrett

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；planning(abstract)；分类 cs.CL、cs.AI

Comments NeurIPS 2023

Legal Prompting: Teaching a Language Model to Think Like a Lawyer

Fangyi Yu, Lee Quartey, Frank Schilder

专题命中 数学推理 ：reasoning(abstract)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.CL、cs.AI

Comments 12 pages, 6 figures, 4 tables. Accepted by NLLP 2022 (EMNLP workshop)

Building Math Agents with Multi-Turn Iterative Preference Learning

Wei Xiong, Chengshuai Shi, Jiaming Shen, Aviv Rosenberg, Zhen Qin, Daniele Calandriello, Misha Khalman, Rishabh Joshi, Bilal Piot, Mohammad Saleh, Chi Jin, Tong Zhang, Tianqi Liu

专题命中 数学推理 ：reasoning(abstract,comments)；chain-of-thought(abstract)；CoT(abstract)；分类 cs.LG

Comments A multi-turn direct preference learning framework for tool-integrated reasoning tasks

BRUCE: Benchmarking Robustness Under Corruption Escalation for Scientific Vision-Language Reasoning

BRUCE：面向科学视觉-语言推理的污染升级下鲁棒性基准测试

Saim Rehman, Muhammad Shafique

机构 * New York University Abu Dhabi（纽约大学阿布扎比分校） 纠错

专题命中 数学推理 ：reasoning(title,abstract)

AI总结 该研究提出BRUCE基准测试框架，通过RCI与T-RCI指标分析VLMs在科学视觉-语言推理任务中随污染升级的鲁棒性退化，实现可解释的失败分析。

Diagnosing Long-Video Quantitative Reasoning in Multimodal LLMs via Enumeration and Counting

EC-Bench：超长视频的枚举与计数基准

Fumihiko Tsuchiya, Taiki Miyanishi, Shunsuke Yasuki, Mahiro Ukai, Nakamasa Inoue, Shuhei Kurita, Yusuke Iwasawa, Yutaka Matsuo

机构 * The University of Tokyo, Japan（东京大学） ； Institute of Science Tokyo, Japan（东京科学大学） ； National Institute of Informatics, Japan（国立信息学研究所） 纠错

专题命中 数学推理 ：reasoning(title,abstract)

AI总结 EC-Bench针对超长视频的枚举与计数问题，通过152部超过30分钟的视频和1699个查询，评估多模态大语言模型的性能，揭示模型在时间推理和计数任务中的局限性。

Comments The first two authors are equally contributed. The data and code are publicly available at: https://github.com/matsuolab/EC-Bench