A system built from specialized candidates can outperform a single fixed model—but only if its router has enough evidence to make the right choice. This paper treats routing as a general systems problem: what information a router can use determines how much quality it can recover, and whether those gains justify the added compute.
Key Findings
- The evaluation spans 25 benchmarks covering language, image, video and speech generation, coding and workplace agents, and robotic manipulation. Its trajectory matrix contains 374,401 scored runs across 452 candidate configurations, offering a broader view than routing studies limited to one task or model family.
- Per-input oracle routing beats the best fixed candidate by 5.4 points. This is an upper bound rather than a deployable method, but it shows how much performance is left on the table when every task uses the same candidate.
- Learned routers exceed the best fixed candidate by 0.3 points across 16 benchmarks, with larger gains on video generation (+2.1) and robot manipulation (+1.1), while using 24% less policy time. The average efficiency improvement over the reference model is 8.82×.
- The unified theoretical framing connects routing quality to the information available at decision time. In practice, this makes router observability—not just the number of candidate models—a central design constraint.
Who It’s For—and the Trade-off
A good fit for researchers and system builders deciding whether heterogeneous model or policy pools can improve a broad AI workload, especially when inference cost matters. The results are encouraging, but benchmark-wide averages do not guarantee gains on every task or deployment: routing depends on the candidates, workload, available signals, and cost metric. The oracle result should be read as a ceiling, not an achievable production baseline.
Method in Context
The study examines routing at the level of selecting models and policies for tasks across modalities and applications. Its breadth helps compare capability orchestration across domains, while also making the headline averages dependent on the particular benchmark mix and candidate pool. It is most useful as evidence that routing can be measured systematically—not as proof that one routing policy will transfer unchanged to every environment.