数学推理论文趋势 2023–2026

主要结论

• 数学推理论文从 2023 年的 57 篇增长到 2026 年的 577 篇，三年间扩大到约 10 倍。
• NeurIPS 是发数学推理论文最多的会议，2023–2026 年共 469 篇，占四个会议总量的 43%。
• 交叉方向里涨得最快的是强化学习：同时属于两个主题的论文从 2023 年的 3 篇增至 2026 年的 140 篇。
• 在「神经符号」的 5 个主题里，它的 2026 年论文数排第 1；同组增长最快的可微推理是 2.5 倍，数学推理是 10 倍。

代表论文

按年份和会议均匀挑选，方便快速看到每一届的样貌

• NeurIPS 2026A mathematical theory of balancing relational generalization and memorization
• ICML 2026ASyMOB: Algebraic Symbolic Mathematical Operations Benchmark
• ICLR 2026AutoGPS: Automated Geometry Problem Solving via Multimodal Formalization and Deductive Reasoning
• CVPR 2026Affine Perspective-Three-Point Problem
• NeurIPS 2025Are Language Models Efficient Reasoners? A Perspective from Logic Programming
• ICML 2025Autoformulation of Mathematical Optimization Models Using LLMs
• ICLR 2025Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages
• CVPR 2025Bridging Viewpoint Gaps: Geometric Reasoning Boosts Semantic Correspondence
• NeurIPS 2024A Careful Examination of Large Language Model Performance on Grade School Arithmetic
• ICML 2024Accelerating Legacy Numerical Solvers by Non-intrusive Gradient-based Meta-solving
• ICLR 2024Don't Trust: Verify -- Grounding LLM Quantitative Reasoning with Autoformalization
• CVPR 2024E-GPS: Explainable Geometry Problem Solving via Top-Down Solver and Bottom-Up Generator
• NeurIPS 2023Evaluating and Improving Tool-Augmented Computation-Intensive Math Reasoning
• ICML 2023Automated Search for Conjectures on Mathematical Constants using Analysis of Integer Sequences
• ICLR 2023Differentiable Mathematical Programming for Object-Centric Representation Learning
• CVPR 2023Abstract Visual Reasoning: An Algebraic Approach for Solving Raven's Progressive Matrices
• NeurIPS 2026A Mechanistic Analysis of Looped Reasoning Language Models
• ICML 2026Beyond Theorem Proving: Formulation, Framework and Benchmark for Formal Problem-Solving
• ICLR 2026CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving
• CVPR 2026Hilbert-Geo Solving Solid Geometric Problems by Neural-Symbolic Reasoning

全部数学推理论文

共 469 篇，第 2 / 10 页

51. NeurIPS 2026Credit Assignment with Resets in Language Model Reasoning
52. NeurIPS 2026CryptanalysisBench: Can LLMs do cryptanalysis?
53. NeurIPS 2026CurveBench: A Benchmark for Exact Topological Reasoning over Nested Jordan Curves
54. NeurIPS 2026DACE RL for Compute Efficient Reinforcement Learning in Small Model Reasoning
55. NeurIPS 2026Data Diversity Drives the Emergence of Symbolic Mechanisms Supporting Abstract Reasoning
56. NeurIPS 2026Decoding the Critique Mechanism in Large Reasoning Models
57. NeurIPS 2026Deep Reasoning in General Purpose Agents via Structured Meta-Cognition
58. NeurIPS 2026Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves
59. NeurIPS 2026DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding
60. NeurIPS 2026Does Your Large Language Model Have An Intuitive Sense of The Difficulty of A Question?
61. NeurIPS 2026DRAGON : A Benchmark for Evidence-Grounded Visual Reasoning over Diagrams
62. NeurIPS 2026DualSAT: A Dual-Branch GNN-Transformer Framework for SAT Solving
63. NeurIPS 2026Efficient Reasoning via Constrained Optimization in Latent Space
64. NeurIPS 2026Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models
65. NeurIPS 2026Exact Posterior Score Estimation for Solving Linear Inverse Problems
66. NeurIPS 2026Extraneous Cognitive Load in Large Language Models
67. NeurIPS 2026FACBench: A Benchmark for Formal-Anchor Collisions in Multilingual Mathematical Grounding
68. NeurIPS 2026Failure Profiling and Reachable Trajectory Selection for Reasoning Distillation
69. NeurIPS 2026FaithSieve: Fine-Grained Evaluation of Math Proofs with Faithful Formal Evidence
70. NeurIPS 2026Fixed-Point Reasoning: Stable and Adaptive Deep Looped Models
71. NeurIPS 2026From Finding to Linking: Benchmarking and Advancing Cross-Long-Video Reasoning for Multimodal LLMs
72. NeurIPS 2026From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning
73. NeurIPS 2026From Solver Trajectories to Teaching Trajectories: Cognition-Aligned Reasoning Distillation
74. NeurIPS 2026GeLVR: Geometry-Consistent Latent Visual Reasoning in Multimodal LLMs
75. NeurIPS 2026GeoDial: A Multimodal Dialog Tutoring Dataset for Geometry Problem-Solving with Visual Tutor Turns
76. NeurIPS 2026Geometric Latent Reasoning Induces Shorter Generations in LLMs
77. NeurIPS 2026Geometric Signatures of Reasoning: A Spectral Perspective on Task Hardness
78. NeurIPS 2026GeoX: Mastering Geospatial Reasoning Through Self-Play and Verifiable Rewards
79. NeurIPS 2026GUI-Libra: Data-Efficient Post-Training for Reliable Reasoning-and-Acting in Native GUI Agents
80. NeurIPS 2026Hide to See: Reasoning-prefix Masking for Visual-anchored Thinking in VLM Distillation
81. NeurIPS 2026HierRR: Enhancing Instruction Alignment in Open-Vocabulary Indoor Scene Synthesis via Agentic Hierarchical Reasoning and Reflection
82. NeurIPS 2026How do Small Transformer Models Learn Hard Math Tasks?
83. NeurIPS 2026How does RL Post-training Induce Skill Composition? A Case Study on Countdown
84. NeurIPS 2026Identified-Set Geometry of Distributional Model Extraction under Top-K Censored API Access
85. NeurIPS 2026INFUSER: Influence-Guided Self-Evolution Improves Reasoning
86. NeurIPS 2026Interactive Combinatorial Reinforcement Learning for Knowledge Graph Reasoning
87. NeurIPS 2026Iterative Nonlinear Computation Underlying Abstract Reasoning
88. NeurIPS 2026Language Models Need Sleep
89. NeurIPS 2026Latent Spatial Reasoning: Building Innate 3D Awareness via Latent-Space Distillation
90. NeurIPS 2026LatentUM: Unleashing the Potential of Interleaved Cross-Modal Reasoning via a Latent-Space Unified Model
91. NeurIPS 2026LeAct: Learning to Reason from Expert Actions
92. NeurIPS 2026LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models
93. NeurIPS 2026Learning to Complete Extremal Mathematical Structures
94. NeurIPS 2026Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
95. NeurIPS 2026LLM Agents Already Know When to Call Tools - Even Without Reasoning
96. NeurIPS 2026LLMs Show No Signs Of Individuated Metacognition
97. NeurIPS 2026Look Before You Reason: Implicit Visual Thinking for Efficient Multimodal Reasoning
98. NeurIPS 2026Low-Rank Hierarchical Merging for Efficient Long-to-Short Reasoning
99. NeurIPS 2026MammoGPS: A Benchmark for Visual Grounding, Perception, and Spatial Reasoning in Mammography
100. NeurIPS 2026Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning

常见问题

2023–2026 年四大会议一共有多少篇数学推理论文？

共 1,085 篇，按年份分别是 2023 年 57 篇、2024 年 137 篇、2025 年 314 篇、2026 年 577 篇。

NeurIPS 2026 有多少篇数学推理论文？

NeurIPS 2026 有 208 篇数学推理论文，上一届是 190 篇。

ICLR 2026 有多少篇数学推理论文？

ICLR 2026 有 175 篇数学推理论文，上一届是 61 篇。

ICML 2026 有多少篇数学推理论文？

ICML 2026 有 113 篇数学推理论文，上一届是 45 篇。

CVPR 2026 有多少篇数学推理论文？

CVPR 2026 有 81 篇数学推理论文，上一届是 18 篇。

哪个会议的数学推理论文最多？

NeurIPS，2023–2026 年共 469 篇。