IDEAgent: Agentic Quality-Diversity Search for Research Idea Generation

TL;DR - IDEAgent frames research idea generation as a quality-diversity search, using multiple agents to refine ideas while avoiding overlap. Across 32 computer-science topics, it achieved 3.89× the best baseline’s Yield.

• Tracks idea lineages and applies multi-objective feedback for targeted repair and refinement.
• Promotes diversity by comparing proposals with completed, ancestral, and rejected ideas.
• Introduces Yield, measuring the largest mutually diverse idea set above a quality threshold.
• Achieved non-zero Yield on eight times more topics than the best baseline.