OrcaRouter: A Production-Oriented LLM Router with Hybrid Offline-Online Learning
Researchers have developed OrcaRouter, a system designed to intelligently route incoming requests to the most suitable large language model. This router employs a hybrid offline-online learning approach, utilizing contextual bandits with lexical and sentence-embedding features. In offline testing, OrcaRouter achieved a high accuracy rate and ranked second on the RouterArena leaderboard, demonstrating its efficiency and cost-effectiveness for production environments. AI
IMPACT Optimizes LLM deployment by dynamically selecting the best model for each query, potentially reducing costs and improving performance.