Route Anything Studies AI Model Routing
- •Route Anything studies routing specialized models and policies across general-purpose AI tasks
- •Dataset spans 25 benchmarks, 374,401 scored runs, and 452 candidate configurations
- •Learned routers gain 0.3 points across 16 benchmarks and use 24% less policy time
Researchers introduce Route Anything, a framework for dynamically choosing and coordinating specialized AI models and policies across tasks. The paper studies how a system can select candidates while balancing task quality against computational cost, instead of relying on one model for every capability.
The researchers build an omni-modal trajectory matrix (scored results across multiple kinds of tasks) covering 25 benchmarks in language understanding, image, video and speech generation, coding, workplace agents, and robotic manipulation. It contains 374,401 scored runs from 452 candidate configurations. They use it to study model routing, learned routers, and when routing improves efficiency.
Across 16 benchmarks, learned routers outperform the best fixed candidate by 0.3 points; gains are 2.1 points on video generation and 1.1 points on robot manipulation, while using 24% less policy time. Per-input oracle routing, which selects the best candidate for each input, exceeds the best fixed candidate by 5.4 points. The authors also report that routing improves efficiency by an average of 8.82x compared with the reference model.