Mechanistic Interpretability

Follow this topic Ask about this topic

The field seeks to explain how a model’s internal computations produce particular outputs, rather than relying only on input-output behavior. Researchers use sparse autoencoders and transcoders to identify features, and probes, attribution graphs, activation patching, ablation, and steering to trace or test their roles. Studies examine language-model tasks such as factual recall, reasoning, and refusal, as well as vision, audio, and scientific models. A central challenge is that a signal can be easy to decode without being causally used, while circuit explanations and interventions can vary across prompts, models, and analysis choices.

Latest papers

391 papers63 papers in the last four weeks, up 75% on the four weeks before. 0.4% of all new papers.

The Persistent Vulnerability of Aligned AI Systems
Mar 31, 2026Aengus LynchAI Agent Safety AI Alignment
Department of Computer Science, University College London
Tracking Equivalent Mechanistic Interpretations Across Neural Networks
Mar 31, 2026Alan Sun, Mariya TonevaMechanistic Interpretability Transformer Interpretability
Carnegie Mellon University · Max Planck Institute for Software Systems
Entanglement as Memory: Mechanistic Interpretability of Quantum Language Models
Mar 27, 2026Nathan RollLong-Range Dependency Modeling Circuits in Language Models
Department of Linguistics, Stanford University
A Universal Vibe? Finding and Controlling Language-Agnostic Informal Register with SAEs
Mar 27, 2026Uri Z. Kialy, Avi Shtarkberg, Ayal KleinMultilingual Language Models Mechanistic Interpretability
Ariel University
H-Node Attack and Defense in Large Language Models
Mar 27, 2026Eric Yocam, Varghese Vaidyan, Yong WangLLM Hallucination Mitigation Hallucination in Language Models
Department of Computer Science and Software Engineering, California Polytechnic State University, San Luis Obispo, CA 93407, USA. · Beacom College of Computer and Cyber Sciences, Dakota State University, Madison, SD 57042, USA. · Department of Computer Science, University of Idaho, Moscow, ID 83844, USA.
Can We Locate and Prevent Stereotypes in LLMs?
Mar 26, 2026Alex D'SouzaSocial Bias in Language Models LLM Interpretability
UC Davis
Central Dogma Transformer III: Interpretable AI Across DNA, RNA, and Protein
Mar 24, 2026Nobuyuki OtaBioinformatics Mechanistic Interpretability
Independent Researcher, Burlingame, CA, USA
Mechanisms of Introspective Awareness
Mar 22, 2026Uzay Macar, Li Yang, Atticus Wang +3Language Model Introspection Metacognition in Language Models
1Anthropic Fellows Program · 2MIT · 3Constellation +1
Posterior-Calibrated Causal Circuits in Variational Autoencoders: Why Image-Domain Interpretability Fails on Tabular Data
Mar 22, 2026Dip Roy, Rajiv Misra, Sanjay Kumar Singh +1Variational Autoencoders Causal Interpretability
Department of Computer Science and Engineering, Indian Institute of Technology, Patna, India · Department of Computer Science, Rajarshi School of Management Technology, Varanasi, India · Department of Electronics and Communication Engineering, Jaypee Institute of Information Technology, Noida
CLT-Forge: A Scalable Library for Cross-Layer Transcoders and Attribution Graphs
Mar 22, 2026Florent Draye, Vedant Palit, Abir Harrasse +8LLM Interpretability Mechanistic Interpretability
Max Planck Institute for Intelligent Systems, Tübingen, Germany · Jinesis AI Lab, University of Toronto & Vector Institute · CMU +2
Semantic Sections: An Atlas-Native Feature Ontology for Obstructed Representation Spaces
Mar 21, 2026Hossein JavidniaMechanistic Interpretability Representation Geometry in Language Models
School of Computing, Dublin City University, Ireland
The Anatomy of an Edit: Mechanism-Guided Activation Steering for Knowledge Editing
Mar 21, 2026Yuan Cao, Mingyang Wang, Hinrich SchützeLanguage Model Editing Activation Steering
Technical University of Munich · LMU Munich · Munich Center for Machine Learning (MCML)
Pitfalls in Evaluating Interpretability Agents
Mar 20, 2026Tal Haklay, Nikhil Prakash, Sana Pandey +5Mechanistic Interpretability LLM Agent Evaluation
Technion—IIT · Northeastern University · MIT +2
WASD: Locating Critical Neurons as Sufficient Conditions for Explaining and Controlling LLM Behavior
Mar 19, 2026Haonan Yu, Junhao Liu, Zhenyu Yan +2LLM Interpretability Language Model-Based Control
Key Lab of High Confidence Software Technologies (Peking University), Ministry of Education School of Computer Science, Peking University, Beijing, China
Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations
Mar 18, 2026Sanjay Basu, Sadiq Y. Patel, Parth Sheth +5Mechanistic Interpretability Language Model Steering
University of California San Francisco, San Francisco, CA, USA · Waymark, San Francisco, CA, USA · University of Pennsylvania, Philadelphia, PA, USA +2
Detection Is Cheap, Routing Is Learned: Why Refusal-Based Alignment Evaluation Fails
Mar 18, 2026Gregory N. FrankLanguage Model Safety Evaluation Refusal Behavior in Language Models
Independent Researcher
How do LLMs Compute Verbal Confidence
Mar 18, 2026Dharshan Kumaran, Arthur Conmy, Federico Barbero +3Verbalized Confidence Estimation Confidence Estimation in Language Models
Google DeepMind
From Drop-off to Recovery: A Mechanistic Analysis of Segmentation in MLLMs
Mar 18, 2026Boyong Wu, Sanghwan Kim, Zeynep AkataMultimodal Large Language Models Semantic Segmentation
Technical University of Munich · Helmholtz Munich · Munich Center for Machine Learning (MCML)
Me, Myself, and $π$ : Evaluating and Explaining LLM Introspection
Mar 17, 2026Atharv Naphade, Samarth Bhargav, Sean Lim +1Language Model Introspection Self-Reflection in Language Model Agents
Carnegie Mellon University
Sparse but not Simpler: A Multi-Level Interpretability Analysis of Vision Transformers
Mar 16, 2026Siyu ZhangNeural Network Interpretability Transformer Interpretability
University of Texas at Austin, Austin TX 78712, USA
Interpretative Interfaces: Designing for AI-Mediated Reading Practices and the Knowledge Commons
Mar 16, 2026Gabrielle BenabdallahLLM Interpretability Interpretable ML
University of Washington Seattle, Washington, USA
Mechanistic Foundations of Goal-Directed Control
Mar 16, 2026Alma LagoEmbodied Cognition Mechanistic Interpretability
Cajal Neuroscience Center (CNC), CSIC · Department of Computer Science and Artificial Intelligence, University of the Basque Country (UPV/EHU)
Exhaustive Circuit Mapping of a Single-Cell Foundation Model Reveals Massive Redundancy, Heavy-Tailed Hub Architecture, and Layer-Dependent Differentiation Control
Mar 12, 2026Ihor KendiukhovMechanistic Interpretability Transformer Interpretability
Department of Computer Science, University of Tübingen, Germany
Attention Gathers, MLPs Compose: A Causal Analysis of an Action-Outcome Circuit in VideoViT
Mar 11, 2026Sai V R ChereddyTransformer Interpretability Video Representation Learning
Independent Researcher
Discovery of a Hematopoietic Manifold in scGPT Yields a Method for Extracting Performant Algorithms from Biological Foundation Model Internals
Mar 10, 2026Ihor KendiukhovSingle-Cell Transcriptomics Mechanistic Interpretability
Department of Computer Science, University of Tübingen, Tübingen, Germany
Dissecting Chronos: Sparse Autoencoders Reveal Causal Feature Hierarchies in Time Series Foundation Models
Mar 10, 2026Anurag MishraTime-Series Foundation Models Transformer Interpretability
Rochester Institute of Technology
SYNAPSE: Framework for Neuron Analysis and Perturbation in Sequence Encoding
Mar 9, 2026Jesús Sánchez Ochoa, Enrique Tomás Martínez Beltrán, Alberto Huertas CeldránNeuron-Level Interpretability Transformer Interpretability
Department of Information and Communications Engineering, University of Murcia, Murcia, 30100, Spain
The Struggle Between Continuation and Refusal: A Mechanistic Analysis of the Continuation-Triggered Jailbreak in LLMs
Mar 9, 2026Yonghong Deng, Zhen Yang, Ping Jian +4LLM Jailbreak Attacks LLM Safety Alignment
School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China · Beike Language and Intelligence, Beijing, China
Interpretable-by-Design Transformers via Architectural Stream Independence
Mar 8, 2026Clayton Kerce, Alexis FoxTransformer Interpretability Neural Network Interpretability
Georgia Tech Research Institute
Engineering Verifiable Modularity in Transformers via Per-Layer Supervision
Mar 8, 2026J. Clayton KerceTransformer Interpretability Mechanistic Interpretability
Georgia Tech Research Institute

Mar 31, 2026J/K move · Enter open · S save

The Persistent Vulnerability of Aligned AI Systems

Aengus Lynch

Department of Computer Science, University College London

AI Agent Safety AI Alignment Adversarial Attacks on LLMs Adversarial Robustness of Language Models Adversarial Training Mechanistic Interpretability

Open paper AskPDF arXiv

Autonomous AI agents are being deployed with filesystem access, email control, and multi-step planning. This thesis contributes to four open problems in AI safety: understanding dangerous internal computations, removing dangerous behaviors once embedded, testing for vulnerabilities before deployment, and predicting when models will act against deployers. ACDC automates circuit discovery in transformers, recovering all five component types from prior manual work on GPT-2 Small by selecting 68 edges from 32,000 candidates in hours rather than months. Latent Adversarial Training (LAT) removes dangerous behaviors by optimizing perturbations in the residual stream to elicit failure modes, then training under those perturbations. LAT solved the sleeper agent problem where standard safety training failed, matching existing defenses with 700x fewer GPU hours. Best-of-N jailbreaking achieves 89% attack success on GPT-4o and 78% on Claude 3.5 Sonnet through random input augmentations. Attack success follows power law scaling across text, vision, and audio, enabling quantitative forecasting of adversarial robustness. Agentic misalignment tests whether frontier models autonomously choose harmful actions given ordinary goals. Across 16 models, agents engaged in blackmail (96% for Claude Opus 4), espionage, and actions causing death. Misbehavior rates rose from 6.5% to 55.1% when models stated scenarios were real rather than evaluations. The thesis does not fully resolve any of these problems but makes each tractable and measurable.