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 138 篇，第 2 / 3 页

51. ICML 2026A Fourier perspective on the learning dynamics of neural networks: from sample complexities to mechanistic insights
52. ICML 2026A Narrowing Geometry in Contaminated Reasoning
53. ICML 2026Adversarial Vulnerability from Interference Between Features in Superposition
54. ICML 2026AI Engram: In Search of Memory Traces in Artificial Intelligence
55. ICML 2026All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs
56. ICML 2026Bayesian Gated Non-Negative Contrastive Learning
57. ICML 2026Beyond Fixed Biases: Decoding the Role of Reasoning Uncertainty in MLLM Modality Conflicts
58. ICML 2026BLOCK-EM: Preventing Emergent Misalignment via Latent Blocking
59. ICML 2026Breaking the Block: Preserving Data Continuity to Train Superior SAEs for Instruct Models
60. ICML 2026Certified Circuits: Stability Guarantees for Mechanistic Circuits
61. ICML 2026CircuitPrint: Mechanistic Circuit Fingerprints for Large Language Models
62. ICML 2026Contribution Weights: A Geometrical Analysis of Self-Attention Transformers
63. ICML 2026CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features
64. ICML 2026DB-KSVD: Scalable Alternating Optimization for Disentangling High-Dimensional Embedding Spaces
65. ICML 2026Deep neural networks divide and conquer dihedral multiplication
66. ICML 2026Discovering Interpretable Algorithms by Decompiling Transformers to RASP
67. ICML 2026DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders
68. ICML 2026Do Language Models Track Entities Across State Changes?
69. ICML 2026Do Sparse Autoencoders Identify Reasoning Features in Language Models?
70. ICML 2026Dual Mechanisms of Value Expression: Intrinsic vs. Prompted Values in Large Language Models
71. ICML 2026Endogenous Resistance to Activation Steering in Language Models
72. ICML 2026Faster Query-Key Learning Sharpens Attention in Self-Attention Models
73. ICML 2026From Directions to Regions: Decomposing Activations in Language Models via Local Geometry
74. ICML 2026Hide and Seek in Embedding Space: Geometry-based Steganography and Detection in Large Language Models
75. ICML 2026How Do Language Models Speak Languages? A Case Study on Unintended Code-Switching
76. ICML 2026How do LLMs Compute Verbal Confidence?
77. ICML 2026How Few-Shot Examples Add Up: A Causal Decomposition of Function Vectors in In-Context Learning
78. ICML 2026How Language Models Process Negation
79. ICML 2026Induction Meets Biology: Mechanisms of Repeat Detection in Protein Language Models
80. ICML 2026Interpretability Transfer from Language to Vision via Sparse Autoencoders
81. ICML 2026Interpreting Physics in Video World Models
82. ICML 2026Language as a Wave Phenomenon: Semantic Phase Locking and Interference in Neural Networks
83. ICML 2026Language Model Circuits Are Sparse in the Neuron Basis
84. ICML 2026Learning Adaptive Perturbation-Conditioned Contexts for Robust Transcriptional Response Prediction
85. ICML 2026Learning Protein Structure-Function Relationships through Knowledge-guided Representation Decomposition
86. ICML 2026Learning Self-Interpretation from Interpretability Artifacts: Training Lightweight Adapters on Vector-Label Pairs
87. ICML 2026Left–Right Symmetry Breaking in CLIP-Style Vision-Language Models Trained on Synthetic Spatial-Relation Data
88. ICML 2026Mechanisms of Introspective Awareness
89. ICML 2026Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units
90. ICML 2026Mechanistic Interpretability as Statistical Estimation: A Variance Analysis
91. ICML 2026Mitigating Hallucinations in Large Vision-Language Models via Causal Route Gating
92. ICML 2026Patterning: The Dual of Interpretability
93. ICML 2026PolySAE: Modeling Feature Interactions in Sparse Autoencoders via Polynomial Decoding
94. ICML 2026Predicting the Emergence of Induction Heads in Language Model Pretraining
95. ICML 2026Protein Circuit Tracing via Cross-layer Transcoders
96. ICML 2026Quantifying LLM Attention-Head Stability: Implications for Circuit Universality
97. ICML 2026Query Circuits: Explaining How Language Models Answer User Prompts
98. ICML 2026Reading Between the Tokens: Improving Preference Predictions through Mechanistic Forecasting
99. ICML 2026Representational Curvature Modulates Behavioral Uncertainty in Large Language Models
100. ICML 2026Revisiting Anisotropy in Language Transformers: The Geometry of Learning Dynamics

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。