Cost-Aware LLM Selection via Dueling Feedback Bandit Algorithm
September 28, 2026
A new Track-and-Stop algorithm identifies the best-performing LLM from a heterogeneous pool by optimizing for both pairwise preference signals and varying query costs. The method achieves asymptotically optimal costs for finding the Condorcet winner as error approaches zero.
HOW THIS AFFECTS YOU
●
builderYou can implement this approach to programmatically select the cheapest model that meets your quality requirements.
●
founderThis enables more efficient automated benchmarking and model selection workflows in production.