MechRL: Reinforcement Learning Agents Perform Circuit Discovery for Mechanistic Interpretability

In three linesMechRL uses a PPO agent operating over 144 attention heads of GPT-2 small to automatically discover mechanistic circuits. Trained on induction and IOI tasks, the agent identifies causally relevant heads via zero-ablation and contrastive rewards, generalizing to docstring completion (96% of oracle with best-of-five planning).

Summary generated by Claude — human-verified