Route Anything study finds learned model routers gain little over the best fixed model

A University of California, Berkeley-led study called Route Anything tested model routing across 25 benchmarks using 452 candidate models and policies.

Across 16 benchmarks, learned routers gained only 0.3 points over the best single fixed model, while a per-input oracle router gained 5.4 points, according to the authors.

The authors report that routing raised efficiency 8.82 times on average at a quality loss below 0.5 points, and 25.5 times on coding tasks. The study ran all candidates on the same inputs, covering tasks from text question answering to robot control.