Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery
Ballad matched or beat the best individual model in the reported tests while cutting API costs by up to 65.6% versus always using the strongest model.
The researchers tested four models across five NP-hard combinatorial optimization problems and found no consistent advantage for larger, costlier models in discovering heuristics. Ballad learns which model to call as the search progresses, favoring rare, high-scoring results over steady averages. It traces credit through the ancestry of generated heuristics and discounts older observations as each model’s usefulness changes. In several cases, the ensemble surpassed every model tested individually.
ArXiv · AI/CL/LG's note
The researchers tested four models across five NP-hard combinatorial optimization problems and found no consistent advantage for larger, costlier models in discovering heuristics. Ballad learns which model to call as the search progresses, favoring rare, high-scoring results over steady averages. It traces credit through the ancestry of generated heuristics and discounts older observations as each model’s usefulness changes. In several cases, the ensemble surpassed every model tested individually.
ArXiv · AI/CL/LG's note
score 4