Training AI Scientists to Replicate Research
Summary
Inherent Laboratories introduces Faraday, a 27B-parameter AI Scientist that surpasses Claude Opus 4.8 and GPT-5.5 in replicating research tasks. The system trains via long-horizon reinforcement learning on a large Replica task suite (310 tasks across ML/AI-for-science domains), uses an LLM judge to capture 'research taste,' and leverages GPT-5.5 Codex as a tool. The work positions replication as a stepping stone toward autonomous, innovation-capable AI scientists with scalable oversight.