Open-sourced circuit tracing tools
Anthropic open-sourced its attribution-graph tooling, with a Neuronpedia frontend, so anyone can trace circuits in open-weights models.
Releases the method behind 'Tracing the thoughts of an LLM' as a library that generates attribution graphs on open-weights models (demonstrated on Gemma-2-2b and Llama-3.2-1b), plus an interactive explorer at Neuronpedia. Built by Anthropic Fellows with Decode Research.
- Date
- Thursday, 29 May 2025
- Lab
- Anthropic
- Kind
- infra
- Access
- paper only
Open-source code release (library plus Neuronpedia frontend), not model weights; access tag is 'paper only' because the schema has no open-source-tooling value.
Sources
This record was checked against its sources on 6 October 2026. How we check