Researchers from UIUC, the University of Maryland, Nanyang Technological University, Purdue, and the University of Illinois Chicago introduce LLMRouter, a unified framework that casts LLM routing as a sequential decision process built from context encoders, model encoders, scoring functions, decision rules, and learning signals. The framework includes an automated pipeline for building routing training data and a new benchmark, xRouteBench, spanning generic, memory-augmented, vision, time-series, and personalized routing scenarios. In experiments, learned routers built with the framework outperformed fixed-model baselines by roughly 14.6% on average, with router rankings shifting considerably depending on the cost-performance trade-off being optimized.
