MuEvo: LLM-driven evolution of multi-heuristic ensembles outperforms single-heuristic design

MuEvo, proposed in arXiv:2608.03636, advances LLM-based automated heuristic design (LLM-AHD) by moving from single-heuristic optimization to coordinated ensembles. The core challenge: practical solvers rely on multiple interacting components, but naively extending single-heuristic methods fails because early component selection may discard late-blooming options, while independent evolution misses inter-component dependencies. MuEvo addresses this with Dynamic Component Management, which uses short-budget probing and a reversible lifecycle to revise component priorities during search. It also employs LLM-Driven Co-Evolution that coordinates component populations through Multi-Ensemble Evaluation, Cross-Component Information Sharing, Relation-Guided Pair Evolution, and Adaptive Budget Allocation.

In experiments across selection hyper-heuristics and componentized ant colony optimization in four combinatorial optimization domains, MuEvo significantly improved human-designed frameworks and outperformed representative multi-component extensions of state-of-the-art LLM-AHD methods. The results suggest that ensemble-level feedback is critical for discovering effective heuristics, and that LLMs can guide not just individual heuristics but entire interacting systems. The 30-page paper includes 4 figures and 16 tables, with code and data links available via arXiv.