MOSAIC: Adversarial co-evolution creates specialist LLM heuristics that beat SOTA

Researchers Oguzhan Gungordu, Siheng Xiong, and Faramarz Fekri introduce MOSAIC, a grid-based framework that adversarially co-evolves problem instances and specialist heuristics inside a Quality-Diversity archive indexed by structural instance features. Instances evolve to expose weaknesses in current heuristics, while heuristics evolve to specialize to newly exposed regions. Each archive cell stores a specialist heuristic, representative instances, and insights explaining what works in that region, forming persistent memory. For heuristic pairs from distant grid regions, an LLM-guided evolutionary loop generates discriminative instances, a decision tree identifies where each heuristic wins, and a reflection LLM produces multi-directional insights to guide crossover and mutation. The archive serves as both a co-evolved benchmark of discriminative instances and a pool of region specialists, from which greedy selection extracts a compact complementary portfolio. Across COPs, test sizes, and LLM backbones, the portfolio consistently outperforms state-of-the-art LLM-based AHD methods, while the co-evolved instances achieve higher feature-space coverage and stronger heuristic discrimination than evolutionary instance-generation baselines.