[5L1-OS-30] OS-30

9:30 AM - 9:45 AM JST(12:30 AM - 12:45 AM UTC)

[5L1-OS-30-03] Evolving collaborative research ideas with multi-agent grounding in lab-specific contexts and literature

〇Yu Chinen¹, Haruka Ozaki¹ (1. The Institute of Physical and Chemical Research)

Keywords:

AI for Science,Large Language Models,Research Idea Generation,LLM-as-a-Judge

Collaborative research tends to have high scientific and societal impact, but designing collaborative research ideas is often difficult due to the siloing of lab knowledge. In this study, we propose a multi-agent system that continuously generates and improves inter-laboratory collaborative research ideas based on each lab’s knowledge. The proposed system consists of six types of agents, including a Reflection Agent, a Review Agent, and an Evolution Agent: the Reflection Agent grounds ideas through literature search, the Review Agent conducts pairwise evaluation using criteria such as validity and novelty, and the Evolution Agent evolves the selected ideas. This idea evolution by these agents is repeated multiple times. Comparing the initial ideas with the ideas after six update rounds, the updated ideas were superior with a win rate of 77.8%. However, since the top-ranked proposals tended to be biased toward specific lab pairings, it is important for future work to improve idea quality while maintaining diversity, for example by introducing methods that explicitly consider Quality-Diversity.