IdeaScientist trains AI agents to turn research gaps into proposals

IdeaScientist, shared as research from a Meta internship, trains specialized AI agents to spot research gaps, connect fields and draft proposals. The researchers report 14% higher overall performance than open autoresearch baselines, a 25% improvement in novelty and 75–92% human preference win rates against six open baselines in blind evaluations. They also introduce Svalbard Idea Vault, a collection of 2.77 million decomposed research ideas for training and evaluating scientific ideation.