Mechanistic Interpretability papers, 2023–2026

Key findings

• Mechanistic Interpretability went from 17 papers in 2023 to 138 in 2026, roughly 8.1x in three years.
• ICML publishes the most Mechanistic Interpretability work: 119 papers from 2023 to 2026, 38% of the four-venue total.
• The fastest-growing overlap is with Deep learning · large language models: papers tagged with both rose from 1 in 2023 to 23 in 2026.
• Among the 5 topics under Interpretability, it ranks #2 by 2026 paper count; the fastest grower in the group, Feature Attribution, is up 1.9x versus 8.1x here.

Representative papers

Spread across years and venues so every edition shows up

• 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

All Mechanistic Interpretability papers

311 papers, page 5 of 7

201. ICML 2025Scaling Sparse Feature Circuits For Studying In-Context Learning
202. ICML 2025Strategy Coopetition Explains the Emergence and Transience of In-Context Learning
203. ICML 2025Taming Knowledge Conflicts in Language Models
204. ICML 2025The Logical Implication Steering Method for Conditional Interventions on Transformer Generation
205. ICML 2025The underlying structures of self-attention: symmetry, directionality, and emergent dynamics in Transformer training
206. ICML 2025To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models
207. ICML 2025Towards Global-level Mechanistic Interpretability: A Perspective of Modular Circuits of Large Language Models
208. ICML 2025Towards scientific discovery with dictionary learning: Extracting biological concepts from microscopy foundation models
209. ICML 2025Towards Understanding Fine-Tuning Mechanisms of LLMs via Circuit Analysis
210. ICML 2025Validating Mechanistic Interpretations: An Axiomatic Approach
211. ICML 2025Which Attention Heads Matter for In-Context Learning?
212. NeurIPS 2025A Implies B: Circuit Analysis in LLMs for Propositional Logical Reasoning
213. NeurIPS 2025Angular Steering: Behavior Control via Rotation in Activation Space
214. NeurIPS 2025Attention Sinks: A 'Catch, Tag, Release' Mechanism for Embeddings
215. NeurIPS 2025Better Training Data Attribution via Better Inverse Hessian-Vector Products
216. NeurIPS 2025Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits
217. NeurIPS 2025Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models
218. NeurIPS 2025Capturing Polysemanticity with PRISM: A Multi-Concept Feature Description Framework
219. NeurIPS 2025Causal Head Gating: A Framework for Interpreting Roles of Attention Heads in Transformers
220. NeurIPS 2025Dense SAE Latents Are Features, Not Bugs
221. NeurIPS 2025DePass: Unified Feature Attributing by Simple Decomposed Forward Pass
222. NeurIPS 2025Detecting High-Stakes Interactions with Activation Probes
223. NeurIPS 2025Do different prompting methods yield a common task representation in language models?
224. NeurIPS 2025EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification
225. NeurIPS 2025FaCT: Faithful Concept Traces for Explaining Neural Network Decisions
226. NeurIPS 2025Failure by Interference: Language Models Make Balanced Parentheses Errors When Faulty Mechanisms Overshadow Sound Ones
227. NeurIPS 2025Improved Representation Steering for Language Models
228. NeurIPS 2025Interpreting Emergent Features in Deep Learning-based Side-channel Analysis
229. NeurIPS 2025Interpreting vision transformers via residual replacement model
230. NeurIPS 2025LLM Interpretability with Identifiable Temporal-Instantaneous Representation
231. NeurIPS 2025LLM Layers Immediately Correct Each Other
232. NeurIPS 2025Manipulating Feature Visualizations with Gradient Slingshots
233. NeurIPS 2025Mechanistic Interpretability of RNNs emulating Hidden Markov Models
234. NeurIPS 2025Mitigating Overthinking in Large Reasoning Models via Manifold Steering
235. NeurIPS 2025Pinpointing Attention-Causal Communication in Language Models
236. NeurIPS 2025Prompting as Scientific Inquiry
237. NeurIPS 2025Remarkable Robustness of LLMs: Stages of Inference?
238. NeurIPS 2025Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates
239. NeurIPS 2025Revising and Falsifying Sparse Autoencoder Feature Explanations
240. NeurIPS 2025Same Task, Different Circuits: Disentangling Modality-Specific Mechanisms in VLMs
241. NeurIPS 2025The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability?
242. NeurIPS 2025The Structure of Relation Decoding Linear Operators in Large Language Models
243. NeurIPS 2025Uncovering a Universal Abstract Algorithm for Modular Addition in Neural Networks
244. NeurIPS 2025Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning
245. NeurIPS 2025What One Cannot, Two Can: Two-Layer Transformers Provably Represent Induction Heads on Any-Order Markov Chains
246. CVPR 2024SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology
247. ICLR 2024Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models
248. ICLR 2024Circuit Component Reuse Across Tasks in Transformer Language Models
249. ICLR 2024Feature emergence via margin maximization: case studies in algebraic tasks
250. ICLR 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking

FAQ

How many Mechanistic Interpretability papers were published at NeurIPS, ICLR, ICML and CVPR from 2023 to 2026?

311 in total: 17 in 2023, 49 in 2024, 107 in 2025, 138 in 2026.

How many Mechanistic Interpretability papers are at NeurIPS 2026?

NeurIPS 2026 has 16 Mechanistic Interpretability papers, versus 34 the year before.

How many Mechanistic Interpretability papers are at ICLR 2026?

ICLR 2026 has 46 Mechanistic Interpretability papers, versus 38 the year before.

How many Mechanistic Interpretability papers are at ICML 2026?

ICML 2026 has 74 Mechanistic Interpretability papers, versus 33 the year before.

How many Mechanistic Interpretability papers are at CVPR 2026?

CVPR 2026 has 2 Mechanistic Interpretability papers, versus 2 the year before.

Which conference publishes the most Mechanistic Interpretability papers?

ICML, with 119 papers from 2023 to 2026.