机制解释论文趋势 2023–2026

主要结论

• 机制解释论文从 2023 年的 17 篇增长到 2026 年的 138 篇，三年间扩大到约 8.1 倍。
• ICML 是发机制解释论文最多的会议，2023–2026 年共 119 篇，占四个会议总量的 38%。
• 交叉方向里涨得最快的是深度学习 · large language models：同时属于两个主题的论文从 2023 年的 1 篇增至 2026 年的 23 篇。
• 在「可解释性」的 5 个主题里，它的 2026 年论文数排第 2；同组增长最快的特征归因是 1.9 倍，机制解释是 8.1 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 Mechanistic Interpretability Study of an Astronomical Foundation Model
• ICML 2026Contribution Weights: A Geometrical Analysis of Self-Attention Transformers
• ICLR 2026Circuit Insights: Towards Interpretability Beyond Activations
• CVPR 2026CIGMA Causal Information-Gain Mechanistic Attribution of Attention Heads in Vision
• NeurIPS 2025Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits
• ICML 2025Algorithm Development in Neural Networks: Insights from the Streaming Parity Task
• ICLR 2025Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics
• CVPR 2025Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability
• NeurIPS 2024InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques
• ICML 2024Mechanistic Neural Networks for Scientific Machine Learning
• ICLR 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking
• CVPR 2024SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology
• NeurIPS 2023A Logic for Expressing Log-Precision Transformers
• ICML 2023Mechanistic Mode Connectivity
• ICLR 2023Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small
• NeurIPS 2026Discovering Mechanistic Models of Neural Activity: System Identification in an in Silico Zebrafish
• ICML 2026CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features
• ICLR 2026CLUE: Conflict-guided Localization for LLM Unlearning Framework
• CVPR 2026From Weights to Concepts Data-Free Interpretability of CLIP via Singular
• NeurIPS 2025Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models

全部机制解释论文

共 102 篇，第 2 / 3 页

51. ICLR 2025Controllable Context Sensitivity and the Knob Behind It
52. ICLR 2025Density estimation with LLMs: a geometric investigation of in-context learning trajectories
53. ICLR 2025Differential learning kinetics govern the transition from memorization to generalization during in-context learning
54. ICLR 2025Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient
55. ICLR 2025Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models
56. ICLR 2025Eliminating Position Bias of Language Models: A Mechanistic Approach
57. ICLR 2025Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable?
58. ICLR 2025Explaining Modern Gated-Linear RNNs via a Unified Implicit Attention Formulation
59. ICLR 2025HyperDAS: Towards Automating Mechanistic Interpretability with Hypernetworks
60. ICLR 2025Improving Instruction-Following in Language Models through Activation Steering
61. ICLR 2025Interpreting Emergent Planning in Model-Free Reinforcement Learning
62. ICLR 2025Interpreting the Second-Order Effects of Neurons in CLIP
63. ICLR 2025Linear Representations of Political Perspective Emerge in Large Language Models
64. ICLR 2025Mechanistic Permutability: Match Features Across Layers
65. ICLR 2025Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity
66. ICLR 2025Monet: Mixture of Monosemantic Experts for Transformers
67. ICLR 2025Not All Language Model Features Are One-Dimensionally Linear
68. ICLR 2025ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability
69. ICLR 2025Residual Stream Analysis with Multi-Layer SAEs
70. ICLR 2025Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words
71. ICLR 2025Revealing and Mitigating Over-Attention in Knowledge Editing
72. ICLR 2025Revisiting In-context Learning Inference Circuit in Large Language Models
73. ICLR 2025Robust LLM safeguarding via refusal feature adversarial training
74. ICLR 2025Scalable Mechanistic Neural Networks
75. ICLR 2025Sparse Autoencoders Do Not Find Canonical Units of Analysis
76. ICLR 2025Systematic Outliers in Large Language Models
77. ICLR 2025The Computational Complexity of Circuit Discovery for Inner Interpretability
78. ICLR 2025Towards a Unified and Verified Understanding of Group-Operation Networks
79. ICLR 2025Towards Interpreting Visual Information Processing in Vision-Language Models
80. ICLR 2025Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control
81. ICLR 2025Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures
82. ICLR 2025Transformers Struggle to Learn to Search
83. ICLR 2025Uncovering Latent Memories in Large Language Models
84. ICLR 2025Words in Motion: Extracting Interpretable Control Vectors for Motion Transformers
85. ICLR 2024Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models
86. ICLR 2024Circuit Component Reuse Across Tasks in Transformer Language Models
87. ICLR 2024Feature emergence via margin maximization: case studies in algebraic tasks
88. ICLR 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking
89. ICLR 2024Function Vectors in Large Language Models
90. ICLR 2024Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching
91. ICLR 2024Linearity of Relation Decoding in Transformer Language Models
92. ICLR 2024Mechanistically analyzing the effects of fine-tuning on procedurally defined tasks
93. ICLR 2024Overthinking the Truth: Understanding how Language Models Process False Demonstrations
94. ICLR 2024PRIME: Prioritizing Interpretability in Failure Mode Extraction
95. ICLR 2024Scaling Laws for Associative Memories
96. ICLR 2024Successor Heads: Recurring, Interpretable Attention Heads In The Wild
97. ICLR 2024The mechanistic basis of data dependence and abrupt learning in an in-context classification task
98. ICLR 2024Towards Best Practices of Activation Patching in Language Models: Metrics and Methods
99. ICLR 2023Concept Gradient: Concept-based Interpretation Without Linear Assumption
100. ICLR 2023Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small

常见问题

2023–2026 年四大会议一共有多少篇机制解释论文？

共 311 篇，按年份分别是 2023 年 17 篇、2024 年 49 篇、2025 年 107 篇、2026 年 138 篇。

NeurIPS 2026 有多少篇机制解释论文？

NeurIPS 2026 有 16 篇机制解释论文，上一届是 34 篇。

ICLR 2026 有多少篇机制解释论文？

ICLR 2026 有 46 篇机制解释论文，上一届是 38 篇。

ICML 2026 有多少篇机制解释论文？

ICML 2026 有 74 篇机制解释论文，上一届是 33 篇。

CVPR 2026 有多少篇机制解释论文？

CVPR 2026 有 2 篇机制解释论文，上一届是 2 篇。

哪个会议的机制解释论文最多？

ICML，2023–2026 年共 119 篇。