大模型推理能力 - arXivDaily 专题

2508.03501 2025-10-14 cs.LG cs.CL cs.SE 62%

Training Long-Context, Multi-Turn Software Engineering Agents with Reinforcement Learning

Alexander Golubev, Maria Trofimova, Sergei Polezhaev, Ibragim Badertdinov, Maksim Nekrashevich, Anton Shevtsov, Simon Karasik, Sergey Abramov, Andrei Andriushchenko, Filipp Fisin, Sergei Skvortsov, Boris Yangel

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

2510.04503 2025-10-13 cs.CR cs.AI cs.CL 62%

P2P: A Poison-to-Poison Remedy for Reliable Backdoor Defense in LLMs

Shuai Zhao, Xinyi Wu, Shiqian Zhao, Xiaobao Wu, Zhongliang Guo, Yanhao Jia, Anh Tuan Luu

机构 * Nanyang Technological University, Singapore（南洋理工大学）； Shanghai Jiao Tong University, Shanghai, China（上海交通大学）

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

2510.05529 2025-10-08 cs.CL cs.LG 62%

H1B-KV: Hybrid One-Bit Caches for Memory-Efficient Large Language Model Inference

Harshil Vejendla

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

Comments MIT URTC 2025 Technical Paper (Oral), 5 pages, 1 figure

2508.13141 2025-10-07 cs.CL cs.LG 62%

OptimalThinkingBench: Evaluating Over and Underthinking in LLMs

Pranjal Aggarwal, Seungone Kim, Jack Lanchantin, Sean Welleck, Jason Weston, Ilia Kulikov, Swarnadeep Saha

机构 * FAIR at Meta（Meta 的 FAIR 研究院）； Carnegie Mellon University（卡内基梅隆大学）

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

Comments 30 pages, 10 tables, 11 figures

2502.06916 2025-10-07 cs.LG cs.AI eess.SP quant-ph 62%

QuIC: Quantum-Inspired Compound Adapters for Parameter Efficient Fine-Tuning

Snehal Raj, Brian Coyle

机构 * QC Ware（QC公司）； LIP6, CNRS, Sorbonne Université（LIP6、CNRS、索邦大学）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

Comments 30 pages, 12 figures, 7 tables, ~8000 words

2510.01253 2025-10-03 cs.AI cs.LG 62%

OR-Toolformer: Modeling and Solving Operations Research Problems with Tool Augmented Large Language Models

Jianzhang Zhang, Jialong Zhou, Chuang Liu

机构 * Alibaba Business School（阿里巴巴商业学院）； Hangzhou Normal University（杭州师范大学）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2510.01135 2025-10-02 cs.LG cs.CL 62%

Prompt Curriculum Learning for Efficient LLM Post-Training

Zhaolin Gao, Joongwon Kim, Wen Sun, Thorsten Joachims, Sid Wang, Richard Yuanzhe Pang, Liang Tan

机构 * Meta Superintelligence Labs（Meta超智能实验室）； Cornell University（康奈尔大学）； University of Washington（华盛顿大学）

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

2510.00911 2025-10-02 cs.LG cs.AI 62%

RiskPO: Risk-based Policy Optimization via Verifiable Reward for LLM Post-Training

Tao Ren, Jinyang Jiang, Hui Yang, Wan Tian, Minhao Zou, Guanghao Li, Zishi Zhang, Qinghao Wang, Shentao Qin, Yanjun Zhao, Rui Tao, Hui Shao, Yijie Peng

机构 * Peking University（北京大学）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2507.16577 2025-10-02 cs.LG cs.CL 62%

Scaling Linear Attention with Sparse State Expansion

Yuqi Pan, Yongqi An, Zheng Li, Yuhong Chou, Ruijie Zhu, Xiaohui Wang, Mingxuan Wang, Jinqiao Wang, Guoqi Li

机构 * Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）； The Hong Kong Polytechnic University（香港理工大学）

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

2505.20152 2025-10-02 cs.CV cs.AI cs.CL 62%

MMGeoLM: Hard Negative Contrastive Learning for Fine-Grained Geometric Understanding in Large Multimodal Models

Kai Sun, Yushi Bai, Zhen Yang, Jiajie Zhang, Ji Qi, Lei Hou, Juanzi Li

机构 * Tsinghua University（清华大学）

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

2509.26625 2025-10-01 cs.LG cs.AI cs.CV cs.MM 62%

Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training

Junlin Han, Shengbang Tong, David Fan, Yufan Ren, Koustuv Sinha, Philip Torr, Filippos Kokkinos

机构 * Meta Superintelligence Labs（Meta 超智能实验室）； University of Oxford（牛津大学）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

Comments Project page: https://junlinhan.github.io/projects/lsbs/

2509.26594 2025-10-01 cs.LG cs.CL cs.CV 62%

Clarification as Supervision: Reinforcement Learning for Vision-Language Interfaces

John Gkountouras, Ivan Titov

机构 * ILLC, University of Amsterdam（伊拉斯谟范例学院、阿姆斯特丹大学）； ILCC, University of Edinburgh（爱丁堡大学）

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

2509.06346 2025-10-01 cs.LG cs.AI 62%

Ban&Pick: Ehancing Performance and Efficiency of MoE-LLMs via Smarter Routing

Yuanteng Chen, Peisong Wang, Yuantian Shao, Nanxin Zeng, Chang Xu, Jian Cheng

机构 * Nanjing University of Science and Technology（南京理工大学）； C 2 \text{C}^{2} DL, Institute of Automation, Chinese Academy of Sciences（C2 DL，自动化研究所，中国科学院）； School of Artificial Intelligence, University of Chinese Academy of Sciences（中国科学院大学人工智能学院）； School of Computer Science, University of Sydney（悉尼大学计算机学院）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

Comments 21 pages, 9 figures

2509.24389 2025-09-30 cs.CL cs.AI 62%

LLaDA-MoE: A Sparse MoE Diffusion Language Model

Fengqi Zhu, Zebin You, Yipeng Xing, Zenan Huang, Lin Liu, Yihong Zhuang, Guoshan Lu, Kangyu Wang, Xudong Wang, Lanning Wei, Hongrui Guo, Jiaqi Hu, Wentao Ye, Tieyuan Chen, Chenchen Li, Chengfu Tang, Haibo Feng, Jun Hu, Jun Zhou, Xiaolu Zhang, Zhenzhong Lan, Junbo Zhao, Da Zheng, Chongxuan Li, Jianguo Li, Ji-Rong Wen

机构 * Renmin University of China（中国人民大学）； Ant Group（蚂蚁集团）； Shanghai Jiao Tong University（上海交通大学）； Zhejiang University（浙江大学）

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

2509.24261 2025-09-30 cs.AI cs.LG 62%

Risk-Sensitive RL for Alleviating Exploration Dilemmas in Large Language Models

Yuhua Jiang, Jiawei Huang, Yufeng Yuan, Xin Mao, Yu Yue, Qianchuan Zhao, Lin Yan

机构 * Tsinghua University（清华大学）； ETH Zurich（苏黎世联邦理工学院）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2506.05183 2025-09-30 cs.LG cs.AI 62%

TreeRPO: Tree Relative Policy Optimization

Zhicheng Yang, Zhijiang Guo, Yinya Huang, Xiaodan Liang, Yiwei Wang, Jing Tang

机构 * The Hong Kong University of Science and Technology (Guangzhou)（香港科技大学（广州））； The Hong Kong University of Science and Technology（香港科技大学）； ETH AI Center, ETH Zurich（苏黎世联邦理工学院人工智能中心）； University of California, Merced（加州大学梅德福分校）； Sun Yat-sen University（中山大学）； MBZUAI

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

Comments 13pages, 6 figures

2509.21792 2025-09-29 cs.LG cs.AI 62%

FastGRPO: Accelerating Policy Optimization via Concurrency-aware Speculative Decoding and Online Draft Learning

Yizhou Zhang, Ning Lv, Teng Wang, Jisheng Dang

机构 * Lanzhou University（兰州大学）； The University of Hong Kong（香港大学）； National University of Singapore（新加坡国立大学）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

Comments Submitted to ICLR 2026

2505.12716 2025-09-29 cs.CL cs.AI 62%

Shadow-FT: Tuning Instruct Model via Training on Paired Base Model

Taiqiang Wu, Runming Yang, Jiayi Li, Pengfei Hu, Yik-Chung Wu, Ngai Wong, Yujiu Yang

机构 * The University of Hong Kong（香港大学）； Tsinghua University（清华大学）； Tencent（腾讯）

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

Comments 24 pages, 12 tables, 8 figures. Previous name: Shadow-FT: Tuning Instruct via Base

2506.14407 2025-09-26 cs.CL cs.AI 62%

ImpliRet: Benchmarking the Implicit Fact Retrieval Challenge

Zeinab Sadat Taghavi, Ali Modarressi, Yunpu Ma, Hinrich Schütze

机构 * Center for Information and Language Processing, LMU Munich（信息与语言处理中心，慕尼黑大学）； Munich Center for Machine Learning (MCML)（慕尼黑机器学习中心（MCML））

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

2502.11771 2025-09-26 cs.CL cs.AI 62%

The Validation Gap: A Mechanistic Analysis of How Language Models Compute Arithmetic but Fail to Validate It

Leonardo Bertolazzi, Philipp Mondorf, Barbara Plank, Raffaella Bernardi

机构 * DISI, University of Trento（特伦托大学DISI中心）； MaiNLP, Center for Information and Language Processing, LMU Munich（慕尼黑大学信息与语言处理中心）； Munich Center for Machine Learning (MCML), Munich, Germany（慕尼黑机器学习中心）； Free University of Bozen-Bolzano, Italy（博兹纳自由大学）

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

Comments EMNLP 2025 Main, 38 pages, 33 figures

2509.19803 2025-09-25 cs.LG cs.CL 62%

VCRL: Variance-based Curriculum Reinforcement Learning for Large Language Models

Guochao Jiang, Wenfeng Feng, Guofeng Quan, Chuzhan Hao, Yuewei Zhang, Guohua Liu, Hao Wang

机构 * Alibaba Cloud Computing（阿里云计算）

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

2509.19349 2025-09-25 cs.CL cs.LG 62%

ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution

Robert Tjarko Lange, Yuki Imajuku, Edoardo Cetin

机构 * Sakana AI

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

Comments 52 pages, 14 figures

2509.13232 2025-09-24 cs.LG cs.AI stat.ML 62%

Single-stream Policy Optimization

Zhongwen Xu, Zihan Ding

机构 * Tencent（腾讯）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2509.18851 2025-09-24 cs.LG cs.AI 62%

NGRPO: Negative-enhanced Group Relative Policy Optimization

Gongrui Nan, Siye Chen, Jing Huang, Mengyu Lu, Dexun Wang, Chunmei Xie, Weiqi Xiong, Xianzhou Zeng, Qixuan Zhou, Yadong Li, Xingzhong Xu

机构 * Ant Group Shanghai, China（蚂蚁集团上海）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2509.18383 2025-09-24 cs.AI cs.DM cs.LG 62%

Gödel Test: Can Large Language Models Solve Easy Conjectures?

Moran Feldman, Amin Karbasi

机构 * University of Haifa（海法大学）； Cisco Foundation AI（思科基础人工智能）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2509.15157 2025-09-22 cs.LG cs.CL 62%

Mind the Gap: Data Rewriting for Stable Off-Policy Supervised Fine-Tuning

Shiwan Zhao, Xuyang Zhao, Jiaming Zhou, Aobo Kong, Qicheng Li, Yong Qin

机构 * College of Computer Science, Nankai University（南开大学计算机科学学院）

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

2509.15217 2025-09-19 cs.AI cs.CV cs.LG 62%

Generalizable Geometric Image Caption Synthesis

Yue Xin, Wenyuan Wang, Rui Pan, Ruida Wang, Howard Meng, Renjie Pi, Shizhe Diao, Tong Zhang

机构 * University of Illinois Urbana-Champaign (UIUC)（伊利诺伊大学厄巴纳-香槟分校）； Shanghai Jiao Tong University（上海交通大学）； Rutgers University（罗格斯大学）； NVIDIA

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2509.09013 2025-09-12 cs.CL cs.AI cs.CV 62%

Can Vision-Language Models Solve Visual Math Equations?

Monjoy Narayan Choudhury, Junling Wang, Yifan Hou, Mrinmaya Sachan

机构 * IIIT Bangalore（班加罗尔印度理工学院）； ETH Zürich（苏黎世联邦理工学院）

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

Comments Monjoy Narayan Choudhury and Junling Wang contributed equally to this work. Accepted at EMNLP2025 main. Code and datasets are open-sourced with links in the paper

Journal ref Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

2508.03682 2025-09-11 cs.LG cs.AI 62%

Self-Questioning Language Models

Lili Chen, Mihir Prabhudesai, Katerina Fragkiadaki, Hao Liu, Deepak Pathak

机构 * Carnegie Mellon University（卡内基梅隆大学）

专题命中数学推理：reasoning(abstract)；分类 cs.AI、cs.LG

2506.01115 2025-09-05 cs.LG cs.CL 62%

Is Random Attention Sufficient for Sequence Modeling? Disentangling Trainable Components in the Transformer

Yihe Dong, Lorenzo Noci, Mikhail Khodak, Mufan Li

机构 * Princeton University（普林斯顿大学）； ETH Zurich（苏黎世联邦理工学院）

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

1. 数学推理 3246 篇

Training Long-Context, Multi-Turn Software Engineering Agents with Reinforcement Learning

P2P: A Poison-to-Poison Remedy for Reliable Backdoor Defense in LLMs

H1B-KV: Hybrid One-Bit Caches for Memory-Efficient Large Language Model Inference

OptimalThinkingBench: Evaluating Over and Underthinking in LLMs

QuIC: Quantum-Inspired Compound Adapters for Parameter Efficient Fine-Tuning

OR-Toolformer: Modeling and Solving Operations Research Problems with Tool Augmented Large Language Models

Prompt Curriculum Learning for Efficient LLM Post-Training

RiskPO: Risk-based Policy Optimization via Verifiable Reward for LLM Post-Training

Scaling Linear Attention with Sparse State Expansion

MMGeoLM: Hard Negative Contrastive Learning for Fine-Grained Geometric Understanding in Large Multimodal Models

Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training

Clarification as Supervision: Reinforcement Learning for Vision-Language Interfaces

Ban&Pick: Ehancing Performance and Efficiency of MoE-LLMs via Smarter Routing

LLaDA-MoE: A Sparse MoE Diffusion Language Model

Risk-Sensitive RL for Alleviating Exploration Dilemmas in Large Language Models

TreeRPO: Tree Relative Policy Optimization

FastGRPO: Accelerating Policy Optimization via Concurrency-aware Speculative Decoding and Online Draft Learning

Shadow-FT: Tuning Instruct Model via Training on Paired Base Model

ImpliRet: Benchmarking the Implicit Fact Retrieval Challenge

The Validation Gap: A Mechanistic Analysis of How Language Models Compute Arithmetic but Fail to Validate It

VCRL: Variance-based Curriculum Reinforcement Learning for Large Language Models

ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution

Single-stream Policy Optimization

NGRPO: Negative-enhanced Group Relative Policy Optimization

Gödel Test: Can Large Language Models Solve Easy Conjectures?

Mind the Gap: Data Rewriting for Stable Off-Policy Supervised Fine-Tuning

Generalizable Geometric Image Caption Synthesis

Can Vision-Language Models Solve Visual Math Equations?

Self-Questioning Language Models

Is Random Attention Sufficient for Sequence Modeling? Disentangling Trainable Components in the Transformer