Researchers introduced Mechanist, an agentic system designed to autonomously discover the mechanisms underlying AI intelligence by combining an interpretability-focused knowledge graph of roughly 13,000 papers, a broader database of 43 million papers, and a curated library of 32 foundational analytical methods. The system generates mechanistic hypotheses about AI models, performs causal interventions to test them, and translates findings into practical improvements. In application, Mechanist surfaced safety vulnerabilities, produced theories of how models represent knowledge and beliefs, and demonstrated techniques for steering model behavior toward desired outcomes.
