27B Model Faraday Uses Long-Horizon RL to Replicate AI Papers

teortaxesTex · x · 2026-08-17

Inherent Labs introduced Faraday, a 27B-parameter AI Scientist designed to replicate results from AI research papers. Trained via long-horizon reinforcement learning, it outperforms models like Claude Opus and GPT-4 on replication tasks. Analysis suggests Faraday's improvement stems less from coding ability and more from developing "scientific taste"—knowing which experiments to run, what evidence supports claims, and when the code takes a shortcut. Its RL setup transforms abstract research quality into judgeable tasks, using task-specific rubrics and turn-level credit assignment to identify useful decisions within long trajectories.

Original post →

More from coding & agent

coding & agent channel →