IDEAgent: Agentic Quality-Diversity Search for Research Idea Generation
IDEAgent treats research ideation as a joint quality-and-diversity search, not two separate optimization jobs.
The paper introduces a multi-agent system that evolves research ideas through lineages, using repair and refinement for quality while checking against completed, rejected, and ancestral ideas for diversity. It also proposes Yield, a metric for counting the largest set of mutually diverse ideas that clear a quality threshold. Across 32 topics in 8 computer science domains, the authors report a 3.89x Yield gain over the best baseline and non-zero Yield on 8x more topics. The paper is under review and says IDEAgent is open-sourced. ArXiv · AI/CL/LG's note
The paper introduces a multi-agent system that evolves research ideas through lineages, using repair and refinement for quality while checking against completed, rejected, and ancestral ideas for diversity. It also proposes Yield, a metric for counting the largest set of mutually diverse ideas that clear a quality threshold. Across 32 topics in 8 computer science domains, the authors report a 3.89x Yield gain over the best baseline and non-zero Yield on 8x more topics. The paper is under review and says IDEAgent is open-sourced. ArXiv · AI/CL/LG's note
score 4