Meta's IdeaScientist: 27B open model beats Claude Code & Codex setups by up to 5.9% at research proposals
dair_ai · x · 2026-10-10
Meta published a paper on IdeaScientist, an AI research agent that splits ideation into three roles trained separately with RL: a gap finder reads related work for limitations, an innovator retrieves mechanisms that solved similar problems in other fields, and a writer produces full proposals.
- Retrieval runs over 2.77M decomposed research ideas
- A 27B open model beats the strongest open autoresearch baseline by 14.0% (mostly on novelty), and beats Claude Code SDK and Codex SDK setups by up to 5.9%
- Evaluation only allows literature before a cutoff date, scoring proposals against directions later explored in 15K human-written papers
Highly relevant for anyone building AI-scientist systems.
Related event: Meta's IdeaScientist: RL-Trained Agents That Generate Novel Research Ideas(2 posts)→
More from coding & agent
- NVIDIA 推出 Jetson Agent Skills，让编码 Agent 懂边缘硬件 — blaizedsouza · 2026-10-11
- Agent auto-fixing JS errors uncovers WebKit WebSocket payload truncation bug — Vjeux · 2026-10-11
- SkyPilot meetup recap: GPU scheduling pain, Holo4 training on K8s, and SGLang's roadmap — skypilot_org · 2026-10-11
- ODS V3 pre-release turns any PC, Mac or Linux box into a private AI server — tom_doerr · 2026-10-10
- Dev praises Hugging Face docs as the best resource for learning LLMs — ayushthakur0 · 2026-10-10
- Building RL envs is an endless loop, says dev echoing PewDiePie's post-training video — mervenoyann · 2026-10-10