Personalized Game Difficulty with Large Language Models: a Preliminary Study

This paper presents a preliminary study on whether state-of-the-art (SOTA) large language models (LLMs), widely applied in various fields, can serve as dynamic difficulty adjustment (DDA) mechanisms in games. Personalized game mechanisms achieved through software and reflected via consumer electronics are critical for improving player experience. However, research on LLMs' usage in gaming has not addressed personalized game mechanisms, specifically in DDA. This paper picks GPT-4o as the SOTA LLM. Using random values to simulate the human player's skill strength, we investigate whether the LLM can provide a reasonable skill strength of the opponent in a turn-based game. Different instructions are used in the prompts to observe GPT-4o's responses, indicating significant potential to act as a DDA mechanism in games. Some prompts led GPT-40 to provide reasonable difficulty responses and follow the flow theory in games. In addition, pivotal limitations of GPT-40 in acting as DDA were exposed, suggesting prompt optimization and LLMs' fine-tuning as a future direction.