This research introduces a bandit-based framework for routing prompts to LLM experts to maximize quality under constrained feedback budgets. The proposed algorithms achieve regret of O(dT/m) in full-information settings and O(dTK/m) in bandit settings, optimizing routing efficiency with minimal rewards.
HOW THIS AFFECTS YOU
●
builderYou can implement more efficient MoE routing strategies when ground-truth labels are expensive.
●
researcherThis provides a mathematical framework for minimizing regret in online LLM routing tasks.