Multiagent Brainstorming Support using IBIS and LLM role-play

Abstract

This research proposes a brainstorming support method using a multi-agent system that incorporates Issue-Based Information System (IBIS) and LLM role-playing. In human discussions, the IBIS is effective for allowing all participants to comprehensively address problems. While maintaining IBIS as the foundation, we aim to achieve more effective discussion support by introducing LLM-based agents with role-playing capabilities. While group brainstorming generates various ideas through diverse perspectives, there are two challenges when using LLMs for such support. The first challenge is that discussions tend to converge prematurely due to insufficient training in multifaceted conversations, which we address through IBIS-based discussion structures. The second challenge is the homogeneity of perspectives, which we tackle by incorporating role-playing functionality to promote diverse viewpoints. Experimental results confirmed that the proposed system enhanced the creativity of human contributions, and particularly for topics not requiring domain-specific knowledge, the agents were shown to produce more creative outcomes.