Dynamic difficulty adjustment using a large language model: A case study in magic: The Gathering

Abstract

This paper presents a framework called LLM-MTG-DDA, which uses a large language model (LLM) in the real-world card game Magic: The Gathering (MTG) to act as a player and implement a dynamic difficulty adjustment (DDA) mechanism. LLMs, as a highly useful technology, have been explored across various fields. However, research on using LLMs for DDA in games, particularly in complex turn-based games, is very limited. In this paper, GPT-4o acts as two players in a simplified version of MTG. One GPT-4o player plays as a regular player (LLM-Player), while the other GPT-4o player adjusts its strategy based on the current game state to balance the difficulty (LLM-DDA). Our LLM-MTG-DDA framework, with a suitable objectives for different players, demonstrates reasonable DDA, with the LLM-Player’s win rate and the win rate per round (excluding draws) both approaching 50%. This framework provides insights for applying LLMs as DDA mechanisms in other similar games.