MOSAIC: Adversarial Co-evolution for LLM-based Heuristic Design

A novel framework named MOSAIC has been introduced for automated heuristic design (AHD) utilizing large language models (LLMs). This framework tackles the shortcomings of traditional AHD techniques, which often focus on optimizing average performance using a limited dataset and depend on scalar feedback for direction. MOSAIC employs a grid-based strategy that adversarially co-evolves problem instances alongside specialized heuristics within a Quality-Diversity (QD) archive, categorized by structural instance features. In this model, instances develop to reveal the vulnerabilities of existing heuristics, while the heuristics adapt to excel in newly uncovered areas. Each archive cell contains a specialized heuristic, representative instances, and explanations of effectiveness in its domain. The research can be found on arXiv under identifier 2608.07544 and was announced as a cross submission, aiming to enhance heuristic performance across varied instance distributions, surpassing single-heuristic dominance.