Multitype Game Optimisation: A Two-Stage Fine-Tuning Framework for Multi‑Game Optimisation With Large Language Models

Abstract

Large language models (LLMs) have made remarkable advances in natural language processing, demonstrating great potential in modelling structured sequences. However, adapting these capabilities to machine gaming tasks such as Go remains challenging due to limitations in strategy generalisation and optimisation efficiency. This paper presents multitype game optimisation (MyGO), a two-stage fine-tuning framework tailored for two-player perfect information board games, exploring the applicability of LLMs to nonlinguistic decision-making domains. In the supervised fine-tuning stage, we propose a unified structural encoding method, action semantic unit (ASU), which efficiently converts heterogeneous game records into discrete token sequences compatible with LLMs. In the reinforcement learning stage, we design TA-PPO (token-level adaptive proximal policy optimisation), an enhanced PPO-based algorithm to address the issue of sparse feedback commonly encountered in game reinforcement learning. Experimental results demonstrate that the fine-tuned models achieve superior or comparable performance to traditional game-playing algorithms in terms of strategy quality, rule generalisation and inference efficiency. This work provides a scalable paradigm for fine-tuning LLMs in complex decision-making tasks and lays a foundation for future research in game AI and generalisable strategy optimisation.