Route Anything studies when routing between models beats one fixed model
A new paper builds a 374,401-run trajectory matrix over 25 benchmarks and 452 candidate configurations to test how far learned routers can beat a single fixed candidate.
The question
Rather than relying on a single model to master every capability, the Route Anything paper (arXiv 2611.05275, published 2026-11-03) looks at systems that orchestrate a heterogeneous ecosystem of specialized models, tools and policies, which the authors call candidates. The central question is how such a system can dynamically select the most capable candidate for each task while balancing performance against computational cost.
The authors present Route Anything as a unified framework for studying model and policy routing across general-purpose AI systems. On the theory side, the abstract says it characterizes optimal routing under quality-cost trade-offs and establishes how the information available to a router determines its attainable performance. Put simply, what a router gets to see about an input limits how well it can choose.
The trajectory matrix
To study routing across very different capabilities, the authors first built what they call an omni-modal trajectory matrix. It spans 25 benchmarks covering language understanding, image, video and speech generation, coding and workplace agents, and robotic manipulation. The matrix holds 374,401 scored runs from 452 candidate configurations.
On top of this matrix the authors train and test different routers. They look at three things: the potential of omni-modal model routing, the capabilities of learned routers, and the conditions under which routing improves efficiency.
Results as reported
All figures below are the authors’ own, taken from the abstract.
- Per-input oracle routing, which picks the best candidate for every input, outperforms the best fixed candidate by 5.4 points.
- Learned routers outperform the best fixed candidate by 0.3 points across 16 benchmarks.
- On video generation, learned routers are ahead by 2.1 points.
- On robot manipulation they are ahead by 1.1 points while using 24% less policy time.
- Compared with the reference model, routing raises efficiency by 8.82x on average.
The distance between the oracle’s 5.4 points and the learned routers’ 0.3 points across 16 benchmarks shows how much of the possible gain current learned routers leave unused, even though they already beat the best single candidate.
What the abstract leaves open
The abstract does not name the 452 candidate configurations or describe the router architectures. It does not identify the reference model behind the 8.82x efficiency figure, and it does not explain how efficiency is measured beyond the policy-time figure for robot manipulation. It also does not say which 16 of the 25 benchmarks the 0.3-point result covers. The paper lists a code repository: route-anything on GitHub.