Route Anything extends routing to images, video, speech and robots: learned routers beat the best fixed candidate by 0.3 points across 16 benchmarks

Routing value is no longer measured only on language models but compared across seven task groups: text QA, image-text QA, image generation, video generation, speech generation, code and workplace agents, and robot manipulation. Earlier routing work centered on language models, including cascades that escalate to a stronger model, routers learned from preference data, and formulations linking quality with cost; whether routing helps across modalities remained an open question. On 19,814 test inputs where every candidate is scored, the best fixed candidate per benchmark scores 82.60 and the per-input oracle 87.98, a gap of 5.38 points; the gap is 0.68 points for speech generation and about 18–19 points for video generation and robot manipulation.

What a router can gain depends on what it observes: seeing only the benchmark or task type leaves most of the potential gain out of reach. Choosing the best candidate per benchmark in hindsight raises 82.60 only to 82.62; on the 18 benchmarks of the capability tree, the best candidate per one of 107 capability leaves chosen in hindsight scores 77.89, against 77.03 for the fixed candidate and 87.44 for the oracle. On RoboTwin, choosing the best policy per task and scene condition in hindsight recovers 23% of the 18.4-point gap; on GenEval2, each of the five other generators beats the fixed one by more than two points on 76–100 training prompts while trailing it on more.

One 4B router trained on all benchmarks at once lifts the score from 79.68 to 79.98 across 16 benchmarks and departs from the fixed candidate where candidates complement each other. The fixed candidate is the best candidate per benchmark on training inputs; the capability-tree router uses a Qwen3.5-4B classifier to predict one of 157 leaves, and a global table fitted on training inputs maps each leaf to a candidate. The leaf classifier is 86–89% accurate, yet 83–86% of inputs still go to the fixed candidate; gains concentrate on Terminal-Bench, VBench-2.0, OneIG and GDP.pdf, with the largest loss on AA-LCR; among 31 router designs and 124 configurations on GenEval2 and Qwen-Image-Bench, none beats the fixed generator with an interval above zero.

On robot manipulation, the router reads three camera frames and the robot state and switches policies by scene condition, reaching higher success with less policy time. ME-Dex is the best single policy on both training and test scenes at 70.45% test success, with PatchWAM second at 63.65%; the best policy flips between clean and domain-randomized scenes. The router sends 929 of 1,000 clean scenes to PatchWAM and 842 of 1,000 randomized scenes to ME-Dex, reaching 71.55% overall success against 70.45%, at 11.05 instead of 14.51 seconds of policy time per scene, 24% less.

The conclusions rest on a trajectory matrix of 25 benchmarks, 452 candidate configurations and 374,401 scored runs, with test comparisons on 19,814 inputs where every candidate is scored; the robot result comes from 2,000 RoboTwin 2.0 test scenes, half clean and half domain-randomized, drawn from 10 tasks the routers never see in training. The theory takes the law of states as given, so it does not cover states that depend on the router's own earlier choices, such as SciCode steps containing a model's earlier code; certification also needs the reference's outcome on every certification input.

On 11 text, agent and image benchmarks, the outcome router stays within noise of the matched fixed candidate; a 68-type router, although 75% accurate, maps 93% of its predictions to the same candidate as the true type and scores 1.32 points below it. On robots, the winner-label classifier reaches only 42.8% success against 71.6% for per-policy outcome regression, because a latency tie-break decides 94% of the labels. In image generation, none of 31 router designs across 124 configurations beats the fixed generator with an interval above zero, and 44 are significantly worse.