LLM-based Stackelberg Program Optimization for Adaptive LNS Operators

Stackelberg Program Optimization (SPO) uses LLMs to discover adaptive destroy-repair programs for Large Neighborhood Search. The framework treats operator discovery as a Stackelberg game, evaluating destroy programs as leaders and repair programs as followers to improve performance on TSP and CVRP benchmarks.