Towards Game Level Generation Through LLM and GAN
Filip Martinović, Danijel Mlinarić, Juraj Dončević, Agneza Krajna, Ivica Botički · Annals of Computer Science and Information Systems · 2025
This paper tackles the challenge of adaptive level generation in video games, focusing on generating content that aligns with player skill.A key limitation of procedural content generation (PCG) is achieving semantic control.Specifically, generating levels of varying difficulty with limited training data.To address this problem, we propose a hybrid approach combining Large Language Models (LLMs) and Generative Adversarial Networks (GANs).An LLM is used to generate a diverse, difficulty-labeled dataset of Snake game levels, which are validated with A* pathfinding to ensure playability.These levels serve as training data for GANs that are able to efficiently generate new levels.The system is evaluated through user study and playability metrics.Results show that the LLMassigned difficulty labels correlate strongly with human perception.The achieved playability is 87% for easy levels and 36% for hard levels.Our findings demonstrate that the hybrid LLM-GAN approach enables scalable and semantically controlled content generation, balancing quality, adaptability, and computational efficiency.