Population Digital Red Queen: A Model of LLM-Driven Coevolution in Program Ecosystems

This paper studies whether LLM-driven adversarial program evolution changes when the evolving object is a population rather than a single champion. It introduces Population Digital Red Queen, motivates it with finite-game examples, and tests three matched-budget variants in Core War using Gemini 2.5 Flash. Across 40 long-profile runs, archive+niche pressure improves the average best champion found, but does not improve the mean quality of the final live population. The contribution is about search dynamics, population structure, and evaluation discipline in LLM-driven program evolution.