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 1 / 3 页

1. CVPR 2026CIGMA Causal Information-Gain Mechanistic Attribution of Attention Heads in Vision
2. CVPR 2026From Weights to Concepts Data-Free Interpretability of CLIP via Singular
3. ICLR 2026AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features
4. ICLR 2026Activation Steering with a Feedback Controller
5. ICLR 2026ActivationReasoning: Logical Reasoning in Latent Activation Spaces
6. ICLR 2026ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack
7. ICLR 2026Attributing Response to Context: A Jensen–Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation
8. ICLR 2026Attribution-Guided Decoding
9. ICLR 2026Characterizing Pattern Matching and Its Limits on Compositional Task Structures
10. ICLR 2026Circuit Insights: Towards Interpretability Beyond Activations
11. ICLR 2026CLUE: Conflict-guided Localization for LLM Unlearning Framework
12. ICLR 2026Correlations in the Data Lead to Semantically Rich Feature Geometry Under Superposition
13. ICLR 2026Decomposing Representation Space into Interpretable Subspaces with Unsupervised Learning
14. ICLR 2026Discovering and Steering Interpretable Concepts in Large Generative Music Models
15. ICLR 2026Does Higher Interpretability Imply Better Utility? A Pairwise Analysis on Sparse Autoencoders
16. ICLR 2026Dyslexify: A Mechanistic Defense Against Typographic Attacks in CLIP
17. ICLR 2026Eliciting Numerical Predictive Distributions of LLMs Without Auto-Regression
18. ICLR 2026Evolution of Concepts in Language Model Pre-Training
19. ICLR 2026Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees
20. ICLR 2026Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition
21. ICLR 2026Hedonic Neurons: A Mechanistic Mapping of Latent Coalitions in Transformer MLPs
22. ICLR 2026How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability
23. ICLR 2026Internal Planning in Language Models: Characterizing Horizon and Branch Awareness
24. ICLR 2026Language Models Use Lookbacks to Track Beliefs
25. ICLR 2026Latent Concept Disentanglement in Transformer-based Language Models
26. ICLR 2026Latent Planning Emerges with Scale
27. ICLR 2026Learning Concept Bottleneck Models from Mechanistic Explanations
28. ICLR 2026Localizing Task Recognition and Task Learning in In-Context Learning via Attention Head Analysis
29. ICLR 2026Mechanism of Task-oriented Information Removal in In-context Learning
30. ICLR 2026Mechanistic Independence: A Principle for Identifiable Disentangled Representations
31. ICLR 2026MICLIP: Learning to Interpret Representation in Vision Models
32. ICLR 2026Mixing Mechanisms: How Language Models Retrieve Bound Entities In-Context
33. ICLR 2026Narrow Finetuning Leaves Clearly Readable Traces in the Activation Differences
34. ICLR 2026On the Predictive Power of Representation Dispersion in Language Models
35. ICLR 2026Path Channels and Plan Extension Kernels: a Mechanistic Description of Planning in a Sokoban RNN
36. ICLR 2026Provably Tracking Equivalent Mechanistic Interpretations Across Neural Networks
37. ICLR 2026Reforming the Mechanism: Editing Reasoning Patterns in LLMs with Circuit Reshaping
38. ICLR 2026Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity
39. ICLR 2026Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs
40. ICLR 2026Task Vectors, Learned Not Extracted: Performance Gains and Mechanistic Insights
41. ICLR 2026The Deleuzian Representation Hypothesis
42. ICLR 2026The Geometry and Topology of Circuits: the Manifolds of Modular Addition
43. ICLR 2026The Price of Amortized inference in Sparse Autoencoders
44. ICLR 2026Thought Branches: Interpreting LLM Reasoning Requires Resampling
45. ICLR 2026Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition
46. ICLR 2026Transformers Don’t Need LayerNorm at Inference Time: Scaling LayerNorm Removal to GPT-2 XL and Implications for Mechanistic Interpretability
47. ICLR 2026Verifying Chain-of-Thought Reasoning via Its Computational Graph
48. ICLR 2026When Thinking Backfires: Mechanistic Insights into Reason-induced Misalignment
49. ICML 2026A Capacity-Based Rationale for Multi-Head Attention
50. ICML 2026A Distributional View for Visual Mechanistic Interpretability: KL-Minimal Soft-Constraint Principle

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。