缩放定律论文趋势 2023–2026

主要结论

• 缩放定律论文从 2023 年的 60 篇增长到 2026 年的 248 篇，三年间扩大到约 4.1 倍。
• NeurIPS 是发缩放定律论文最多的会议，2023–2026 年共 221 篇，占四个会议总量的 37%。
• 交叉方向里涨得最快的是时间序列：同时属于两个主题的论文从 2023 年的 1 篇增至 2026 年的 24 篇。
• 在「优化与缩放」的 5 个主题里，它的 2026 年论文数排第 5；同组增长最快的低秩适配是 3.0 倍，缩放定律是 4.1 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 Unified Merge Calculus for Learning-Rate Scaling on Neural Computation Graphs
• ICML 2026A Capacity-Based Rationale for Multi-Head Attention
• ICLR 2026$\mathbf{Li_2}$: A Framework on Dynamics of Feature Emergence and Delayed Generalization
• CVPR 2026A Unified Framework for Knowledge Transfer in Bidirectional Model Scaling
• NeurIPS 2025Audits Under Resource, Data, and Access Constraints: Scaling Laws For Less Discriminatory Alternatives
• ICML 2025A Simple Model of Inference Scaling Laws
• ICLR 2025(Mis)Fitting Scaling Laws: A Survey of Scaling Law Fitting Techniques in Deep Learning
• CVPR 2025D^3: Scaling Up Deepfake Detection by Learning from Discrepancy
• NeurIPS 2024An exactly solvable model for emergence and scaling laws in the multitask sparse parity problem
• ICML 2024A Dynamical Model of Neural Scaling Laws
• ICLR 2024Principled Architecture-aware Scaling of Hyperparameters
• CVPR 2024From Activation to Initialization: Scaling Insights for Optimizing Neural Fields
• NeurIPS 2023Abide by the law and follow the flow: conservation laws for gradient flows
• ICML 2023Dissecting the Effects of SGD Noise in Distinct Regimes of Deep Learning
• ICLR 2023Distributionally Robust Post-hoc Classifiers under Prior Shifts
• CVPR 2023LargeKernel3D: Scaling Up Kernels in 3D Sparse CNNs
• NeurIPS 2026ADIS-Law: Unified Scaling Laws for Annealing-Phase Domain Injection in Large Language Models
• ICML 2026Balancing Learning Rates Across Layers: Exact Two-Step Dynamics and Optimal Scaling in Linear Neural Networks
• ICLR 2026A universal compression theory: Lottery ticket hypothesis and superpolynomial scaling laws
• CVPR 2026MatAnyone 2 Scaling Video Matting via a Learned Quality Evaluator

全部缩放定律论文

共 78 篇，第 2 / 2 页

51. ICML 2026Reuse your FLOPs: Scaling RL on Hard Problems by Conditioning on Very Off-Policy Prefixes
52. ICML 2026Revisiting Padded Transformer Expressivity: Which Architectural Choices Matter and Which Don't
53. ICML 2026Sample Margin-Aware Recalibration of Temperature Scaling
54. ICML 2026Scalable Reinforcement Learning via Adaptive Batch Scaling
55. ICML 2026Scaling Behavior in Model Fine-tuning for Audio DeepFake Detection
56. ICML 2026Scaling Beyond Masked Diffusion Language Models
57. ICML 2026Scaling Continual Learning to 300+ Tasks with Bi-Level Routing Mixture-of-Experts
58. ICML 2026Scaling depth capacity via zero/one-layer model expansion
59. ICML 2026Scaling Law for Quantization-Aware Training
60. ICML 2026Scaling Laws and Architectural Frontiers in Metagenomic Foundation Models
61. ICML 2026Scaling Laws for Precision in High-Dimensional Linear Regression
62. ICML 2026Scaling Laws of Global Weather Models
63. ICML 2026Selling Data as a Digital Good with Scaling Valuations
64. ICML 2026Single-Head Attention in High Dimensions: A Theory of Generalization, Weights Spectra, and Scaling Laws
65. ICML 2026Spectral Reach: Understanding Neural Scaling as Progress into the Spectral Tail
66. ICML 2026Statistically Calibrated Scaling for Token Merging in Transformers
67. ICML 2026Temperature Scaling in Discrete Sequence (Language) Models
68. ICML 2026Token Sample Complexity of Attention
69. ICML 2026TQL: Scaling Q-Functions with Transformers by Preventing Attention Collapse
70. ICML 2026Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries
71. ICML 2026Understanding MARS: When Scaling Momentum Provably Helps
72. ICML 2026Universal One-third Time Scaling in Learning Peaked Distributions
73. ICML 2026Utility Boundary of Dataset Distillation: Scaling and Coverage Laws
74. ICML 2026ViTok-v2: Scaling Native Resolution Autoencoders to 5 Billion Parameters
75. ICML 2026WF-Bench: A Benchmark for Neural-Network WaveFunction Expressivity and Scaling Laws
76. ICML 2026When Data Is Scarce: Scaling Sparse Language Models with Repeated Training
77. ICML 2026When Does Adaptation Win? Scaling Laws for Meta-Learning in Quantum Control
78. ICML 2026Who Said Neural Networks Aren't Linear?

常见问题

2023–2026 年四大会议一共有多少篇缩放定律论文？

共 605 篇，按年份分别是 2023 年 60 篇、2024 年 96 篇、2025 年 201 篇、2026 年 248 篇。

NeurIPS 2026 有多少篇缩放定律论文？

NeurIPS 2026 有 82 篇缩放定律论文，上一届是 80 篇。

ICLR 2026 有多少篇缩放定律论文？

ICLR 2026 有 73 篇缩放定律论文，上一届是 48 篇。

ICML 2026 有多少篇缩放定律论文？

ICML 2026 有 78 篇缩放定律论文，上一届是 56 篇。

CVPR 2026 有多少篇缩放定律论文？

CVPR 2026 有 15 篇缩放定律论文，上一届是 17 篇。

哪个会议的缩放定律论文最多？

NeurIPS，2023–2026 年共 221 篇。