The PIAC framework improves the generalization of LLM-generated algorithm portfolios for combinatorial optimization. It introduces a potential gain metric that estimates generalization benefits by perturbing algorithms, removing the need for high-quality reference solutions during instance generation.
HOW THIS AFFECTS YOU
●
researcherThis enables more efficient co-evolution of algorithms and instances without requiring ground-truth labels.