Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems

Route Anything. A Study of Model and Policy Routing for General-Purpose AI Systems. By Pengfei Zhou and six others, from the National University of Singapore, InfRec, Cardinal AI Lab, and two other institutions. Published on the third of November.

Artificial general intelligence, in the authors' framing, needs a method that can understand, generate, reason and act across diverse domains and modalities. Rather than one model mastering every capability, they see a promising path in orchestrating a heterogeneous ecosystem of specialized models, tools and policies, which they call candidates, into one unified system. That raises a basic question. How should a system dynamically select and coordinate the most capable candidate for each task, while balancing performance against computational cost? Route Anything is a unified framework for studying exactly that, and its headline claim is that routing holds substantial opportunities for orchestrating capability.

The study stands on two foundations. The first is theoretical. A unified foundation characterizes optimal routing under trade-offs between quality and cost, and establishes how the information available to a router determines the performance it can reach. The second is empirical: an omni-modal trajectory matrix spanning twenty-five benchmarks. They cover language understanding; image, video and speech generation; coding and workplace agents; and robotic manipulation. The matrix holds three hundred seventy-four thousand four hundred and one scored runs, drawn from four hundred fifty-two candidate configurations.

Different routers are trained and tested on that matrix. The ceiling comes first: per-input oracle routing beats the best fixed candidate by five point four points. Learned routers beat it too. Across sixteen benchmarks their margin is zero point three points. On video generation it is two point one points. On robot manipulation it is one point one points, while using twenty-four percent less policy time. And compared with the reference model, routing raises efficiency by eight point eight two times on average.

The claims have clear edges. Every comparison is made against the best fixed candidate or against a reference model, inside the benchmarks of this one matrix. Across the sixteen benchmarks, the learned routers sit at zero point three points above the best fixed candidate, while the oracle sits at five point four, so most of the measured opportunity is not yet captured by a learned router. The study also asks under which conditions routing improves efficiency, which ties the efficiency gains to conditions rather than presenting them as universal.

Route Anything treats the road to general capability as a selection problem as much as a modeling problem. Instead of waiting for one model that is best at everything, it asks how far a system gets by knowing which of its many candidates to call, and at what cost. The oracle figure marks how much there is to gain by choosing well for each input. The learned routers mark how much of that can be reached so far. And the efficiency figure, more than eight times on average, is the reason the question matters even where the quality gains are a fraction of a point.