Search papers, labs, and topics across Lattice.
This study explores the use of large language models to evolve complete procedural content generators through a novel approach called Continual Abstraction Discovery (CAD), which extracts reusable programming primitives from high-fitness programs. By applying CAD in a 2x2 experimental design across four classic games, the researchers demonstrated that the mean final best fitness of generated content significantly improved when CAD was utilized alongside a fixed hand-written domain API. The findings indicate that the integration of learned libraries enhances the efficiency and effectiveness of evolutionary program search in game content generation.
Evolving procedural content generators with reusable programming primitives boosts fitness scores across multiple classic games, revealing a powerful synergy between abstraction and evolutionary search.
Large language models can generate executable programs, which makes it possible to search directly over procedural content generators rather than individual levels. We study this approach in Sokoban, Zelda, Dangerous Dave, and Lode Runner. Each run evolves complete Python generators through language-model mutation and crossover. We introduce Continual Abstraction Discovery, or CAD, which extracts reusable primitives from high-fitness programs into a run-specific helper module. A 2x2 experiment crosses CAD with access to a fixed hand-written domain API. The completed data set contains 160 complete runs, with at least ten 50-generation runs in every cell. CAD raises mean final best fitness in all eight domain and API comparisons. Across all CAD runs, learned libraries are adopted by most later programs and repeatedly rediscover validation, reachability, and structural utilities. These results support that discovering reusable primitives improves evolutionary program search for content generators.
