Large Language Models and Dynamic Difficulty Adjustment: An Integration Perspective

Dynamic Difficulty Adjustment (DDA) aims to enhance player retention by adjusting the difficulty level of a game according to the player's skills. However, rule-based DDA systems often struggle with scalability, adaptability, and the cognitive burden of defining exhaustive rules. In this paper, we propose the integration of Large Language Models (LLMs) into DDA mechanisms to overcome these limitations, based on prompt creation techniques. As a proof of concept, we apply this approach to the DDA-MAPEKit framework, replacing its static rule-based logic with LLM actions using ChatGPT, and an experiment was conducted in a space shooter game. The results showed promising outcomes, with pertinent adjustments to the game variables, no hallucinations, and values coupled to the current context. Overall, our findings suggest that LLM-based DDA mechanisms hold significant potential for improving adaptivity in digital games.