LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers
The paper frames LLM routing as a five-part sequential decision problem and ships a benchmark plus modular open-source router stack.
It defines routers through context encoders, model encoders, scoring functions, decision rules, and learning signals. The authors say xRouteBench evaluates both response quality and inference cost across generic, memory-augmented, vision, time-series, and personalized routing tasks. Their study reports learned routers beating the strongest fixed-model baseline by 14.6% relatively, with lightweight routers gaining ground under tighter budgets. User-conditioned routing is reported to improve personalization consistently. HF Daily Papers' note
It defines routers through context encoders, model encoders, scoring functions, decision rules, and learning signals. The authors say xRouteBench evaluates both response quality and inference cost across generic, memory-augmented, vision, time-series, and personalized routing tasks. Their study reports learned routers beating the strongest fixed-model baseline by 14.6% relatively, with lightweight routers gaining ground under tighter budgets. User-conditioned routing is reported to improve personalization consistently. HF Daily Papers' note
score 4