Using AI to understand AI
We develop scalable, AI-driven techniques for understanding all parts of the machine learning pipeline, from internal representations to model behaviors.
Research Updates
research report
Scalably Extracting Latent Representations of Users
Constructing datasets and training decoders to extract user models from language models
25 November 2025
research report
Language Model Circuits Are Sparse in the Neuron Basis
A new technique for tracing sparse and faithful circuits directly on a model's MLPs
20 November 2025
technical demonstration
Monitoring SWE-bench Agents
Partnering with SWE-bench to enable reliable monitoring of AI coding agents
19 November 2025
research report
Training Language Models to Explain Their Own Computations
We trained explainer models to verbalize the content of AI systems' internal computations
11 November 2025
research report
Automatically Jailbreaking Frontier Language Models with Investigator Agents
Discovering cost-effective attacks with reinforcement learning
3 September 2025
research report
Surfacing Pathological Behaviors in Language Models
Improving our investigator agents with propensity bounds
5 June 2025
research report
Investigating Truthfulness in a Pre-Release o3 Model
o3 frequently fabricates actions it took to fulfill user requests, and elaborately justifies the fabrications when confronted
16 April 2025
technical demonstration
Introducing Docent
A system for analyzing and intervening on agent behavior
24 March 2025
research report
Scaling Automatic Neuron Explanation
Open-source AI systems trained to describe components of other AI systems at the level of a human expert
23 October 2024
research report
Eliciting Language Model Behaviors with Investigator Agents
Language models trained to automatically surface harmful behaviors in language models
23 October 2024
technical demonstration
Monitor: An AI-Driven Observability Interface
An interface designed to help humans observe, understand, and steer computations inside models
23 October 2024