accuracy-cost-tradeoff

This paper proposes correlation-aware contextual bandit algorithms that leverage surrogate reward signals from machine learning models for LLM routing, achieving improved accuracy-cost trade-offs and sample efficiency compared to standard baselines.