CMU & Meta's IdeaScientist: multi-agent framework automates cross-domain scientific ideation, boosting novelty 57%
量子位 · wechat · 2026-10-18
Researchers from CMU, Meta and others propose IdeaScientist, a multi-agent framework for grounded scientific ideation: it identifies research gaps in a target field, searches other domains for mechanisms that solved structurally similar problems, and assembles a complete research proposal.
The system uses three main roles (GapFinder, Innovator, ReportWriter) plus an orchestrator, supported by paper-reading and review agents. The team also built SvalbardIdeaVault, a knowledge base of 2.77M structured research ideas drawn from 4.7M papers, and trained each role separately with asynchronous GRPO and per-role LoRA adapters. On Qwen-based evaluation it scored 74.6 overall, 14 points above the best baseline, with novelty jumping from 43.8 to 68.8 (+57%). Untuned, it reached 77.7 with Claude-Opus, beating a Claude Code SDK baseline. Still, it only won 43% of head-to-head comparisons against DeepScientist.
The authors frame this as automating the master-skill of borrowing insights across disciplines, while noting proposals still need real experiments to become reliable results.
More from coding & agent
- Claude cut a 2.6-hour video export to 20 minutes — but only after being told to rebuild the server — byhewar · 2026-10-23
- 3 flight-booking rules my AI agent checks that I'd never do by hand — Efistoffeles · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23
- How do you audit AI agent decisions in regulated workflows? A field survey — Ok_Way_1993 · 2026-10-23