27B Faraday model directs GPT-5.5 to reproduce papers, beats Claude Opus 4.8
aigclink · x · 2026-08-15
Inherent Labs released Faraday, a 27B-parameter AI scientist model that outperforms Claude Opus 4.8 and GPT-5.5 on paper reproduction tasks. Key highlights: 1) Task design: given a paper, the agent must reproduce results within limited compute, leveraging the fact that papers only describe successful paths, enabling infinitely scalable task generation. 2) Small model directing large models: Faraday doesn't write code itself but uses GPT-5.5 Codex as a tool, trained with GPT-5.4-mini and generalizing to stronger agents at test time. 3) Addressing the lack of verifiable rewards in research: it uses an auto-generated per-task rubric judge, combined with multiple sampling aggregation and turn-level credit assignment to reduce noise. This approach, similar to Tsinghua's openrsi, shows that small models plus tool use can rival frontier models. The author sees Faraday as addressing the evaluator bottleneck identified by Lilian Weng, marking significant progress in AI for Science.
More from Models
- Why aren't labs training on hundreds of millions of videos for native audio/video? — imjustnewatai · 2026-10-16
- ChatGPT UI screenshot reveals Wallet coming soon plus Email, Calendar, Health and Finances integrations — jxnlco · 2026-10-16
- Why the next OpenAI and Anthropic models may still disappoint: no native audio or video input — imjustnewatai · 2026-10-16
- Thousands Hit by ChatGPT Standard Voice Mode Regression: Wrong Voice, Interruptions — shaunralston · 2026-10-16
- Gemini Argon 4 Quota Exhausted, Reset Countdown Shows Nearly 7 Days — ssh4net · 2026-10-16
- AINews: Meta ships 816B Muse Spark Open; Goodfire argues neural nets are already clean — Latent Space · 2026-10-16