理论 · learning theory论文趋势 2023–2026

主要结论

• 理论 · learning theory论文从 2023 年的 86 篇增长到 2026 年的 482 篇，三年间扩大到约 5.6 倍。
• NeurIPS 是发理论 · learning theory论文最多的会议，2023–2026 年共 540 篇，占四个会议总量的 40%。
• 交叉方向里涨得最快的是社会层面 · accountability transparency and interpretability：同时属于两个主题的论文从 2023 年的 0 篇增至 2026 年的 33 篇。
• 在「理论与统计」的 5 个主题里，它的 2026 年论文数排第 2；同组增长最快的理论 · optimization是 4.7 倍，理论 · learning theory是 5.6 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 dynamical systems theory of reward-modulated learning in linear recurrent networks
• ICML 2026A Fine-Grained Understanding of Uniform Convergence for Halfspaces
• ICLR 2026A Law of Data Reconstruction for Random Features (And Beyond)
• NeurIPS 2025A Geometric Analysis of PCA
• ICML 2025A Generalization Theory for Zero-Shot Prediction
• ICLR 2025A Solvable Attention for Neural Scaling Laws
• CVPR 2025Dynamic Stereotype Theory Induced Micro-expression Recognition with Oriented Deformation
• NeurIPS 20244+3 Phases of Compute-Optimal Neural Scaling Laws
• ICML 2024A Statistical Theory of Regularization-Based Continual Learning
• ICLR 2024A Sublinear Adversarial Training Algorithm
• CVPR 2024Accurate Training Data for Occupancy Map Prediction in Automated Driving Using Evidence Theory
• NeurIPS 2023$\varepsilon$-fractional core stability in Hedonic Games.
• ICML 2023Robust Collaborative Learning with Linear Gradient Overhead
• ICLR 2023Bispectral Neural Networks
• NeurIPS 2026A Mechanistic Investigation of Theory of Mind in a Large Language Model
• ICML 2026A Random Matrix Theory of Masked Self-Supervised Learning
• ICLR 2026A New Approach to Controlling Linear Dynamical Systems
• NeurIPS 2025Approximation theory for 1-Lipschitz ResNets
• ICML 2025A Two-Stage Learning-to-Defer Approach for Multi-Task Learning
• ICLR 2025A Theoretical Analysis of Self-Supervised Learning for Vision Transformers

全部理论 · learning theory论文

共 147 篇，第 3 / 3 页

101. ICLR 2025Progress or Regress? Self-Improvement Reversal in Post-training
102. ICLR 2025Provable Uncertainty Decomposition via Higher-Order Calibration
103. ICLR 2025Provable weak-to-strong generalization via benign overfitting
104. ICLR 2025Quality over Quantity in Attention Layers: When Adding More Heads Hurts
105. ICLR 2025Quantitative Approximation for Neural Operators in Nonlinear Parabolic Equations
106. ICLR 2025Regret Bounds for Episodic Risk-Sensitive Linear Quadratic Regulator
107. ICLR 2025Rethinking Self-Distillation: Label Averaging and Enhanced Soft Label Refinement with Partial Labels
108. ICLR 2025RNNs are not Transformers (Yet): The Key Bottleneck on In-Context Retrieval
109. ICLR 2025Robust Feature Learning for Multi-Index Models in High Dimensions
110. ICLR 2025Robustness Auditing for Linear Regression: To Singularity and Beyond
111. ICLR 2025Satisficing Regret Minimization in Bandits
112. ICLR 2025Second Order Bounds for Contextual Bandits with Function Approximation
113. ICLR 2025Selective Task Group Updates for Multi-Task Optimization
114. ICLR 2025Self-Improvement in Language Models: The Sharpening Mechanism
115. ICLR 2025Semialgebraic Neural Networks: From roots to representations
116. ICLR 2025Separation Power of Equivariant Neural Networks
117. ICLR 2025Sharper Guarantees for Learning Neural Network Classifiers with Gradient Methods
118. ICLR 2025Sketching for Convex and Nonconvex Regularized Least Squares with Sharp Guarantees
119. ICLR 2025State Space Models are Provably Comparable to Transformers in Dynamic Token Selection
120. ICLR 2025Statistical Advantages of Perturbing Cosine Router in Mixture of Experts
121. ICLR 2025Statistical Tractability of Off-policy Evaluation of History-dependent Policies in POMDPs
122. ICLR 2025Stochastic Bandits Robust to Adversarial Attacks
123. ICLR 2025Strong Model Collapse
124. ICLR 2025Task Descriptors Help Transformers Learn Linear Models In-Context
125. ICLR 2025The Breakdown of Gaussian Universality in Classification of High-dimensional Linear Factor Mixtures
126. ICLR 2025The Computational Complexity of Positive Non-Clashing Teaching in Graphs
127. ICLR 2025The Foundations of Tokenization: Statistical and Computational Concerns
128. ICLR 2025Theory on Mixture-of-Experts in Continual Learning
129. ICLR 2025Theory on Score-Mismatched Diffusion Models and Zero-Shot Conditional Samplers
130. ICLR 2025Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization
131. ICLR 2025Towards Generalization Bounds of GCNs for Adversarially Robust Node Classification
132. ICLR 2025Towards Understanding the Universality of Transformers for Next-Token Prediction
133. ICLR 2025Tractable Multi-Agent Reinforcement Learning through Behavioral Economics
134. ICLR 2025Trained Transformer Classifiers Generalize and Exhibit Benign Overfitting In-Context
135. ICLR 2025Training Nonlinear Transformers for Chain-of-Thought Inference: A Theoretical Generalization Analysis
136. ICLR 2025Transformer Learns Optimal Variable Selection in Group-Sparse Classification
137. ICLR 2025Transformers are Universal In-context Learners
138. ICLR 2025Transformers Handle Endogeneity in In-Context Linear Regression
139. ICLR 2025Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought
140. ICLR 2025Transformers Provably Learn Two-Mixture of Linear Classification via Gradient Flow
141. ICLR 2025Transformers Provably Solve Parity Efficiently with Chain of Thought
142. ICLR 2025U-Nets as Belief Propagation: Efficient Classification, Denoising, and Diffusion in Generative Hierarchical Models
143. ICLR 2025Understanding Factual Recall in Transformers via Associative Memories
144. ICLR 2025uniINF: Best-of-Both-Worlds Algorithm for Parameter-Free Heavy-Tailed MABs
145. ICLR 2025When does compositional structure yield compositional generalization? A kernel theory.
146. ICLR 2025When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers
147. ICLR 2025When narrower is better: the narrow width limit of Bayesian parallel branching neural networks

常见问题

2023–2026 年四大会议一共有多少篇理论 · learning theory论文？

共 1,350 篇，按年份分别是 2023 年 86 篇、2024 年 393 篇、2025 年 389 篇、2026 年 482 篇。

NeurIPS 2026 有多少篇理论 · learning theory论文？

NeurIPS 2026 有 112 篇理论 · learning theory论文，上一届是 106 篇。

ICLR 2026 有多少篇理论 · learning theory论文？

ICLR 2026 有 200 篇理论 · learning theory论文，上一届是 147 篇。

ICML 2026 有多少篇理论 · learning theory论文？

ICML 2026 有 170 篇理论 · learning theory论文，上一届是 132 篇。

CVPR 2025 有多少篇理论 · learning theory论文？

CVPR 2025 有 4 篇理论 · learning theory论文，上一届是 1 篇。

哪个会议的理论 · learning theory论文最多？

NeurIPS，2023–2026 年共 540 篇。