Procedural Content Metageneration via Program Search and Continual Abstraction Discovery

TL;DR - This paper evolves complete Python generators for four games using LLM-driven mutation and crossover. Continual Abstraction Discovery (CAD) improves program-search fitness by extracting reusable primitives from successful generators.

• Evaluated on Sokoban, Zelda, Dangerous Dave, and Lode Runner across 160 complete evolutionary runs.
• A 2×2 experiment tests CAD both with and without a fixed, hand-written domain API.
• CAD increases mean final-best fitness in all eight domain/API comparisons.
• Later programs widely adopt learned libraries, which repeatedly capture validation, reachability, and structural utilities.