Automated Interpretability and Feature Discovery in Language Models with Agents

Arnau Marin-Llobet, Javier Ferrando

We introduce an autonomous multiagent framework for mechanistic interpretability that automates both explaining and finding internal features in large language models. The system r…