AI Generates Game Content by Evolving Program Code

The advent of large language models (LLMs) has opened new avenues for content generation, particularly in the realm of executable programs. Instead of generating individual game levels or assets, this research investigates an approach where LLMs are used to search directly over procedural content generators themselves. This means the AI evolves entire Python programs that can then create game content. The study applied this method to classic games such as Sokoban, Zelda, Dangerous Dave, and Lode Runner. During each experimental run, complete Python generators were evolved through a process involving language-model mutation and crossover. A key innovation introduced is Continual Abstraction Discovery (CAD). CAD works by identifying and extracting reusable primitives—or fundamental building blocks—from programs that demonstrate high fitness (i.e., generate good content). These primitives are then incorporated into a run-specific helper module, making them available for subsequent program evolution. A comprehensive experiment demonstrated that CAD significantly improved the mean final best fitness across all tested domains and API comparisons. The learned libraries, containing these discovered abstractions, were widely adopted by later programs and consistently rediscovered essential utilities like validation, reachability, and structural components. These findings strongly suggest that the ability to discover and reuse primitives is crucial for enhancing evolutionary program search in the context of content generation.