Large Language Models (LLMs) have automated parts of scientific discovery, but existing systems often generate ideas that are either too similar or lack quality. A new paper from Hugging Face researchers argues that research ideation should balance both quality and diversity, framing it as a Quality-Diversity (QD) search problem.
The proposed framework, IDEAgent, uses a multi-agent system to manage idea evolution through lineages. It jointly drives quality via multi-objective feedback for repair and refinement, while ensuring diversity through lightweight sequential memory and explicit comparison against completed ideas, ancestors, and rejected proposals.
To evaluate this QD conjunction, the team developed Yield, a metric that computes the largest set of mutually diverse ideas meeting a quality threshold. Across 32 topics in 8 computer science domains, IDEAgent outperformed the best baseline by 3.89x on Yield and achieved non-zero Yield on 8x more topics.
The authors also analyzed quality improvements, showing that repair and refinement are crucial for building logical rigor and clarity while preserving non-obviousness. The code is open-sourced at GitHub.