Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems
Learned routers beat the best fixed candidate by 0.3 points across 16 benchmarks; an ideal per-input selector gained 5.4 points.
The study tests how systems can select among specialized models and policies while balancing quality against computational cost. Its evaluation matrix contains 374,401 scored runs from 452 candidate configurations across 25 benchmarks. Learned routers delivered gains of 2.1 points on video generation and 1.1 points on robotic manipulation, with the latter using 24% less policy time. The authors also report an average 8.82-fold efficiency improvement over the reference model.
HF Daily Papers' note
The study tests how systems can select among specialized models and policies while balancing quality against computational cost. Its evaluation matrix contains 374,401 scored runs from 452 candidate configurations across 25 benchmarks. Learned routers delivered gains of 2.1 points on video generation and 1.1 points on robotic manipulation, with the latter using 24% less policy time. The authors also report an average 8.82-fold efficiency improvement over the reference model.
HF Daily Papers' note
score 4