Exploring the Integration of Large Language Models and Monte Carlo Tree Search in Team Formation for Turn-Based Games: a case-study with the VGC AI Competition

The integration of Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS) offers a novel approach to tackle decision-making challenges in turn-based strategy games, a domain where traditional methods face scalability and adaptability limitations. The objective is to explore the feasibility of LLMs in improving MCTS strategic reasoning by processing dynamic and probabilistic game states. Although prior research has demonstrated the strengths of MCTS in tactical optimization and the potential of LLMs in natural language reasoning, the synergy between these approaches remains underexplored. To explore this gap, this study faces the team formation problem in the Pokémon VGC AI framework through “PokeHit”, a hybrid AI agent that incorporates LLM-driven reasoning within the MCTS pipeline. By testing open-source models as Llama and Mistral, this work evaluates their ability to guide macro-strategy decisions and compares their performance against baseline-heuristic algorithms. Results reveal that LLMs can provide strategies dynamically but suffer from limitations, including hallucinations and inconsistent improvements over traditional methods. However, this research supports the feasibility and challenges of hybrid AI approaches in fine-tuning and policy design to improve their efficiency. The findings aim to contribute to the field of intelligent game agents through the applicability of LLMs in decision-making domains.