Verification of Reasoning Ability using BDI Logic and Large Language Model in AIWolf

Abstract

We attempt to improve the reasoning capability of LLM in werewolf game by combining BDI logic with LLM. LLM such as ChatGPT has been developed and used for various tasks. However, LLM has many challenges, and logical reasoning is one of them. Therefore, the purpose of this study was to verify the logical reasoning ability of LLM in dialogue of werewofl game by using BDI logic-based prompts. Experiments and evaluations were conducted using ``Werewolf,'' a communication game with incomplete information. From the results of the game played by five agents, we compare the logical reasoning ability of LLM by using the win rate and the vote rate against werewolf. We conducted a discussion using the results.