Faraday beats larger models at paper replication

Inherent Labs says Faraday, a 27B-parameter AI-scientist agent, outperformed Claude Opus 4.8 and GPT-5.5 on held-out paper-replication tasks. The research release describes 310 tasks drawn from 100 machine-learning and AI-for-science papers, with Faraday using coding agents as tools; the paper says it learned through long-horizon reinforcement learning and a rubric-based judge. Research teams should read the result as evidence that scientific judgment can be trained into a smaller orchestrator—not as proof that an agent has discovered new science—and compare it with the archive’s reproducibility gate for AI-generated results.