Adaptive Model Routing

Adaptive model routing studies how to choose, combine, or escalate among models and expert modules for each input instead of relying on one fixed predictor. Common approaches include learned gates and mixture-of-experts, confidence- and uncertainty-based cascades, contextual bandits, and routers using query features or model activations; they are used in settings from LLM serving to edge-cloud inference and time-series forecasting. The goal is to retain strong performance while reducing calls to expensive models, compute, or latency, often under changing budgets and workloads. A central challenge is making reliable, calibrated decisions on unfamiliar inputs and under distribution shift, since routers can learn broad model rankings without predicting which model will succeed on a particular case.