[3L2-GS-5f] Agents

11:15 AM - 11:30 AM JST(2:15 AM - 2:30 AM UTC)

[3L2-GS-5f-02] Diversifying Evacuation Necessity Outputs in Large Language Models Using Observer Prompts and Big Five Personality Trait Prompts

〇Kimihiro Sato¹, Kenta Hakoishi^1,2 (1. Nippon Koei Co., Ltd., 2. Graduate School, University of Tsukuba)

Keywords:

LLM,Evacuation Behavior Simulation,Prompt Engineering,Synthetic Persona,Big Five

Evacuation behavior simulations are useful for examining effective evacuation guidance measures. By assigning personas to a Large Language Model (LLM) and generating corresponding evacuation behaviors, it becomes possible to construct a simulation that enables sensitivity analysis based on linguistic descriptions of the measures. However, such generated behaviors do not necessarily exhibit sufficient diversity, and established methods for such diversification have yet to be developed. Furthermore, evaluations across multiple LLMs remain limited. In this study, we propose Observer Prompts and Big Five Personality Trait Prompts as prompting strategies for diversifying outputs related to the expressions of evacuation necessity. We found that LLMs in the Gemma 3 series and OpenAI's large-scale LLMs tend to produce more diverse outputs when the proposed strategies are applied.