大模型推理能力 - arXivDaily 专题

2407.00390 2024-07-02 cs.CL 70%

Advancing Process Verification for Large Language Models via Tree-Based Preference Learning

Mingqian He, Yongliang Shen, Wenqi Zhang, Zeqi Tan, Weiming Lu

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

2404.03820 2024-06-24 cs.CL 70%

CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues

Makesh Narsimhan Sreedhar, Traian Rebedea, Shaona Ghosh, Jiaqi Zeng, Christopher Parisien

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

2312.06585 2024-04-19 cs.LG 70%

Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models

Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron Parisi, Abhishek Kumar, Alex Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Elsayed, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshiteej Mahajan, Laura Culp, Lechao Xiao, Maxwell L. Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Warkentin, Yundi Qian, Yamini Bansal, Ethan Dyer, Behnam Neyshabur, Jascha Sohl-Dickstein, Noah Fiedel

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

Comments Accepted to TMLR. Camera-ready version. First three authors contributed equally

2404.04817 2024-04-09 cs.CL 70%

FRACTAL: Fine-Grained Scoring from Aggregate Text Labels

Yukti Makhija, Priyanka Agrawal, Rishi Saket, Aravindan Raghuveer

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

Comments 22 pages, 1 figure

2402.13064 2024-02-21 cs.CL 70%

Synthetic Data (Almost) from Scratch: Generalized Instruction Tuning for Language Models

Haoran Li, Qingxiu Dong, Zhengyang Tang, Chaojun Wang, Xingxing Zhang, Haoyang Huang, Shaohan Huang, Xiaolong Huang, Zeqiang Huang, Dongdong Zhang, Yuxian Gu, Xin Cheng, Xun Wang, Si-Qing Chen, Li Dong, Wei Lu, Zhifang Sui, Benyou Wang, Wai Lam, Furu Wei

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

Comments Work in progress

2402.12851 2024-02-21 cs.CL 70%

MoELoRA: Contrastive Learning Guided Mixture of Experts on Parameter-Efficient Fine-Tuning for Large Language Models

Tongxu Luo, Jiahe Lei, Fangyu Lei, Weihao Liu, Shizhu He, Jun Zhao, Kang Liu

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

2311.05821 2023-11-13 cs.CL 70%

Let's Reinforce Step by Step

Sarah Pan, Vladislav Lialin, Sherin Muckatira, Anna Rumshisky

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

Comments NeurIPS 2023 Workshop on Instruction Tuning and Instruction Following

2305.07004 2023-10-24 cs.CL 70%

Not All Languages Are Created Equal in LLMs: Improving Multilingual Capability by Cross-Lingual-Thought Prompting

Haoyang Huang, Tianyi Tang, Dongdong Zhang, Wayne Xin Zhao, Ting Song, Yan Xia, Furu Wei

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

Comments Accepted by EMNLP 2023 Findings

2301.09723 2023-01-26 cs.AI 70%

Mathematics, word problems, common sense, and artificial intelligence

Ernest Davis

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

2202.05086 2022-02-11 cs.AI cs.DB 70%

Complexity of Arithmetic in Warded Datalog+-

Lucas Berent, Markus Nissl, Emanuel Sallinger

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

1609.02584 2017-05-31 cs.AI 70%

Towards Better Response Times and Higher-Quality Queries in Interactive Knowledge Base Debugging

Patrick Rodler

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

2608.23566 2026-08-25 cs.LG cs.AI cs.CL 新提交 67%

How to Train a Critic Stably and Efficiently

如何稳定且高效地训练评价模型

Penghui Qi, Xiangxin Zhou, Wee Sun Lee

机构 * National University of Singapore（新加坡国立大学）； Tencent Hunyuan（腾讯混元）

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

AI总结该研究提出BPCO流程稳定高效训练评价模型，在数学推理任务上，其优于基线，且采样1个响应时可匹配或超过基于组的GRPO基线。

2608.10694 2026-08-13 cs.LG cs.AI cs.CL cs.NE 版本更新 67%

Optimize Cheap, Deploy Strong: Cost-Aware Cross-Tier Transfer for Evolutionary Optimization

优化低成本部署强模型：面向进化优化的成本感知跨层迁移

Tal Oved, Roi Pony, Oshri Naparstek, Udi barzelay

机构 * IBM Research（IBM研究院）

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

AI总结该研究提出成本感知跨层迁移方法，解耦LLM角色，在低成本层级完成大部分搜索，跨层部署提示词，在多任务多模型上实现成本大幅降低且性能不劣于同层级优化。

2602.04718 2026-08-13 cs.LG cs.AI cs.CL 版本更新 67%

Superposition Without Interference? Towards Isolated Interventions via Almost Orthogonal Features in Language Models

通过语言模型中几乎正交的特征实现孤立干预

Moritz Miller, Florent Draye, Bernhard Schölkopf

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

AI总结该研究基于机械可解释性前提，针对语言模型特征纠缠问题，受“独立因果机制”启发，提出将内部特征约束为几乎正交，通过形式化问题、界定干扰传播并关联正则化，实现对数学推理概念更孤立的干预且保留模型性能。

Comments Published as a conference paper at the Conference on Language Modeling (COLM) 2026

2602.09924 2026-08-12 cs.CL cs.AI cs.LG 版本更新 67%

LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations

LLMs编码其失败：从预生成激活中预测成功

William Lugoloobi, Thomas Foster, William Bankes, Chris Russell

机构 * Oxford Internet Institute, University of Oxford（牛津大学牛津互联网研究所）； FLAIR, University of Oxford（牛津大学FLAIR）； Department of Computer Science, University College London（伦敦大学学院计算机科学系）

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

AI总结研究通过预生成激活预测LLM在数学和编程任务中的成功率，发现模型内部表示能有效指导更高效的推理，且在MATH任务中通过模型池路由可提升性能并降低70%的推理成本。

Comments Accepted at COLM 2026

2608.09805 2026-08-11 cs.LG cs.AI cs.CL 新提交 67%

Parameter Exploration for RLVR via Variational Learning

基于变分学习的RLVR参数探索

Vatsal Venkatkrishna, Nico Daheim, Iryna Gurevych

机构 * INSAIT, Sofia University “St. Kliment Ohridski”（INSAIT，圣克莱门特奥赫里德斯基索非亚大学）； National Research Center for Applied Cybersecurity ATHENE（ATHENE国家应用网络安全研究中心）

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

AI总结本文针对LLM强化学习的探索问题，提出参数空间探索思路，引入3PO方法，在OLMo-3-1025-7B等模型的数学推理等任务上，以相近计算成本相比GRPO提升性能，减少训练中的异常分组与轨迹。

2608.07528 2026-08-11 cs.AI cs.CL cs.LG 新提交 67%

The Knowing-Saying Gap: When Probes See Errors that Confidence Misses

知-言差距：当探测模型发现错误而置信度未察觉时

Jyotin Goel, Ipshita Bandyopadhyay, Justin Shenk

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

AI总结该研究发现线性探测模型能精准检测语言模型的上下文破坏，但无法可靠预测最终答案正确性，推翻了相关假设，表明基于探测模型的监控需结合模型与错误类型的路由策略。

2506.01952 2026-08-11 cs.CL cs.AI cs.LG 版本更新 67%

WebChoreArena: Evaluating Web Browsing Agents on Realistic Tedious Web Tasks

WebChoreArena：在现实繁琐网页任务上评估网页浏览智能体

Atsuyuki Miyai, Zaiying Zhao, Kazuki Egashira, Atsuki Sato, Tatsumi Sunada, Shota Onohara, Hiromasa Yamanishi, Mashiro Toyooka, Kunato Nishina, Ryoma Maeda, Kiyoharu Aizawa, Toshihiko Yamasaki

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

AI总结本文提出WebChoreArena，作为WebArena的扩展，包含532个耗时300多小时开发的繁琐网页任务，从大规模记忆、计算、长期记忆三维度评估网页浏览智能体，实验显示其能清晰衡量LLM进展且难度高于WebArena。

Comments COLM2026. Project Page: https://webchorearena.github.io/

2601.03895 2026-08-07 cs.LG cs.AI cs.CL 版本更新 67%

All-Quadrant Bounded Clipping GRPO: Closing the Unbounded Blind Spot for Stable and Generalizable Training

自适应边界裁剪GRPO：确保有界比率以实现稳定且可推广的训练

Chi Liu, Xin Chen

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

AI总结 ABC-GRPO通过自适应边界裁剪提升GRPO的灵活性和泛化能力，实现更稳定和可推广的训练效果。

Comments 13 pages, 3 figures

2608.02087 2026-08-06 cs.AI cs.CL cs.LG 版本更新 67%

Instruction-Conditioned Exploration for Reinforcement Learning with Self-Distillation to an Unconditioned Policy

基于非对称强化学习与自蒸馏的指令条件探索

Jim Dilkes, Vahid Yazdanpanah, Sebastian Stein

机构 * University of Southampton（南安普顿大学）

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

AI总结该研究针对LLM强化学习中的探索挑战，提出指令条件探索（ICE）方法，结合非对称RL/SD训练目标，使Qwen3-1.7B数学推理性能提升5.0%且长上下文下仍有效。

Comments Submitted to ACL Rolling Review (ARR) May 2026 cycle. OpenReview submission record at https://openreview.net/forum?id=PV945lekMa

2608.02139 2026-08-05 cs.CL cs.AI cs.LG 版本更新 67%

Self-Improving Large Language Models via Progressive Experience Evolution

基于渐进式经验演化的自我改进大语言模型

Shijie Ren, Xiting Wang, Meng Li, Yujie Guo, Yunhang Yao, Ziheng Peng, Xunlong Wang, Yuetan Chen, Haoyang Zhou, Yunlong Liang, Fandong Meng

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

AI总结本研究提出统一后训练框架SPEE，通过显式经验演化与隐式策略优化结合，在五种数学推理基准上提升了大语言模型的自我改进能力，优于现有基线方法。

Comments 10 pages, 5 figures

2608.01522 2026-08-04 cs.LG cs.AI cs.CL 新提交 67%

Question Begets Question: Self-Evolving Curriculum for Reinforcement Fine-Tuning on Competition Mathematics

问题催生问题：面向竞赛数学强化微调的自演进课程

Longtian Bao, Jianyou Wang, Yang Zhang, Youze Zheng, Ramamohan Paturi

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

AI总结本文针对竞赛数学强化微调的三大困难，提出Question-begets-Question（QbQ）方法及自演进课程，使Qwen2.5-Math-7B的AIME任务pass@1从5.6%提升至16.5%，且能解决更难问题。

2608.00004 2026-08-04 cs.CL cs.AI cs.LG 新提交 67%

Cost-Effective Automated Judging of Natural-Language Mathematical Proofs

自然语言数学证明的高性价比自动评判

Benjamin Grayzel

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

AI总结该研究以 IMO-GradingBench 为基准，验证 GPT-OSS 120B 等三个低成本模型作为数学证明评判器的效果，发现全票通过规则的低成本方案与前沿模型效果相当，成本低达 100 倍。

Comments 7 pages, 5 figures, 4 tables

2511.20639 2026-08-04 cs.CL cs.AI cs.LG 版本更新 67%

Latent Collaboration in Multi-Agent Systems

多智能体系统中的潜在协作

Jiaru Zou, Ruizhong Qiu, Gaotang Li, Xiyuan Yang, Katherine Tieu, Pan Lu, Ke Shen, Hanghang Tong, Yejin Choi, Jingrui He, James Zou, Mengdi Wang, Ling Yang

机构 * University of Washington（华盛顿大学）

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

AI总结提出LatentMAS框架，使LLM智能体在连续潜在空间直接协作，无需文本中介，实现更高精度、更低开销和更快推理。

Comments ICML2026 Spotlight, Project: https://github.com/Gen-Verse/LatentMAS

2506.06382 2026-08-03 stat.ML cs.AI cs.CL cs.GT cs.LG 67%

On the Fundamental Impossibility of Hallucination Control in Large Language Models

Michał P. Karpowicz

机构 * Samsung AI Center（三星人工智能中心）

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

Comments Mathematics debugged, added examples and illustrations, corrected claims, and re-edited, typos removed

2607.22837 2026-07-28 cs.LG cs.AI cs.CL 新提交 67%

Frustratingly Simple Black-Box Adaptation of Language Models via Logit Bias

通过对数偏差对语言模型进行极其简单的黑箱适应

Ofek I. Cohen, Lior Shani, Aviv Rosenberg, Ankur Samanta, Tal Wagner, Yonathan Efroni

机构 * Tel Aviv University（特拉维夫大学）； Google Research（谷歌研究院）； Columbia University（哥伦比亚大学）

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

AI总结研究如何通过简单API级控制调整语言模型，开发黑箱方法学习对数偏差向量，无需修改模型权重或梯度，该方法在数学和推理基准上优于基础模型，是一种轻量级的语言模型适应机制。

Comments 36 pages, 4 figures, 6 tables. Appendix included. Code available at https://github.com/Ofek-Israeli/logit_bias_lm_adaptation

2607.22554 2026-07-28 cs.AI cs.CL cs.LG 新提交 67%

Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy

相同问题，不同答案：超越准确率评估大语言模型的可靠性

Kazem Faghih, Yize Cheng, Shoumik Saha, Mobina Pournemat, Armin Gerami, Soheil Feizi

机构 * University of Maryland, College Park（马里兰大学帕克分校）

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

AI总结研究大语言模型在不同等效表述问题下的答案变化，发现其输出依赖措辞，准确率指标掩盖不稳定性，实例级行为不稳定，提出自释义策略可恢复潜在知识提升性能，强调评估一致性对展现模型可靠性的重要性。

2607.04763 2026-07-28 cs.LG cs.AI cs.CL stat.ML 版本更新 67%

Multi-Turn On-Policy Distillation with Prefix Replay

基于前缀重放的多轮在线策略蒸馏

Baohao Liao, Hanze Dong, Christof Monz, Xinxing Xu, Li Dong, Furu Wei

机构 * Microsoft Research（微软研究院）； University of Amsterdam（阿姆斯特丹大学）

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

AI总结研究用于智能体任务的在线策略蒸馏，提出重放前缀在线策略蒸馏方法，复用预收集教师轨迹作重放前缀，解决多轮在线策略蒸馏的前缀陷阱问题，提高效率且保持或提升精度。

2607.16872 2026-07-21 cs.CL cs.AI cs.LG 新提交 67%

Trace-Based On-Policy Distillation for Masked Diffusion Language Models

基于轨迹的策略蒸馏用于掩码扩散语言模型

Haolin Ren, Ziyang Huang, Chenhao Yuan, Jun Zhao, Kang Liu

机构 * Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）； University of Chinese Academy of Sciences（中国科学院大学）

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

AI总结研究针对扩散大语言模型推理训练后处理难的问题，提出基于轨迹的策略蒸馏（TOPD）框架，通过在目标模型去噪轨迹上监督，利用教师模型获取令牌分布并以反向KL目标更新，在数学推理基准上提升了模型准确率且减少计算量。

2607.16255 2026-07-21 cs.LG cs.AI cs.CL 新提交 67%

Feature Generation Using LLMs: An Evolutionary Algorithm Approach

使用大语言模型进行特征生成：一种进化算法方法

Aria Nourbakhsh, Benoît Alcaraz, Christoph Schommer

机构 * University of Luxembourg（卢森堡大学）

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

AI总结该研究利用大语言模型解决表格数据特征生成问题，创建管道结合属性与提示生成新特征，经选择算法筛选，应用于八个不同数据集，结果显示语言模型生成的新特征有助于给定任务并提升分类效果。

Journal ref Commun. Comput. Inf. Sci. 2471, 48-64 (2025)

1. 数学推理 3240 篇

Advancing Process Verification for Large Language Models via Tree-Based Preference Learning

CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues

Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models

FRACTAL: Fine-Grained Scoring from Aggregate Text Labels

Synthetic Data (Almost) from Scratch: Generalized Instruction Tuning for Language Models

MoELoRA: Contrastive Learning Guided Mixture of Experts on Parameter-Efficient Fine-Tuning for Large Language Models

Let's Reinforce Step by Step

Not All Languages Are Created Equal in LLMs: Improving Multilingual Capability by Cross-Lingual-Thought Prompting

Mathematics, word problems, common sense, and artificial intelligence

Complexity of Arithmetic in Warded Datalog+-

Towards Better Response Times and Higher-Quality Queries in Interactive Knowledge Base Debugging

How to Train a Critic Stably and Efficiently

Optimize Cheap, Deploy Strong: Cost-Aware Cross-Tier Transfer for Evolutionary Optimization

Superposition Without Interference? Towards Isolated Interventions via Almost Orthogonal Features in Language Models

LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations

Parameter Exploration for RLVR via Variational Learning

The Knowing-Saying Gap: When Probes See Errors that Confidence Misses

WebChoreArena: Evaluating Web Browsing Agents on Realistic Tedious Web Tasks

All-Quadrant Bounded Clipping GRPO: Closing the Unbounded Blind Spot for Stable and Generalizable Training

Instruction-Conditioned Exploration for Reinforcement Learning with Self-Distillation to an Unconditioned Policy

Self-Improving Large Language Models via Progressive Experience Evolution

Question Begets Question: Self-Evolving Curriculum for Reinforcement Fine-Tuning on Competition Mathematics

Cost-Effective Automated Judging of Natural-Language Mathematical Proofs

Latent Collaboration in Multi-Agent Systems

On the Fundamental Impossibility of Hallucination Control in Large Language Models

Frustratingly Simple Black-Box Adaptation of Language Models via Logit Bias

Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy

Multi-Turn On-Policy Distillation with Prefix Replay

Trace-Based On-Policy Distillation for Masked Diffusion Language Models

Feature Generation Using LLMs: An Evolutionary Algorithm Approach