Prototyping an LLM Agent for Autonomous Game Rule Acquisition

Abstract

Large Language Models (LLMs) are attracting attention for the construction of autonomous game agents due to their powerful reasoning capabilities, but research in environments where game rules are not provided beforehand is scarce. While rule acquisition through trial-and-error has been achieved in text-based maze games, a challenge has become evident: self-correction is difficult once an incorrect rule is registered in the repository. To solve this problem, this study proposes a system that diversifies the roles of the conventional process—which executed multiple operations with a single prompt— into multiple modules, and introduces a verification process to evaluate the validity of acquired rules. Experimental results confirmed that the proposed system can accurately exclude intentionally introduced incorrect rules. This demonstrates that the combination of the verification process and module decomposition avoids LLM bias and establishes high robustness for the system. This research provides guidelines for structurally improving the reliability and stability of LLM agents in unknown environments.