Route Anything tests AI routing across 25 benchmarks

A new paper, Route Anything, asks whether model and policy routing works beyond language models. Routing means sending each input to one candidate from a pool: a chat model, an image or video generator, a speech model or a robot policy. The answer from the authors' data is a qualified yes. Routing has real headroom in some places and almost none in others, and a learned router captures only a small part of it.

The paper was posted to arXiv on November 3, 2026 by researchers at the National University of Singapore, InfRec / Cardinal AI Lab, UC Berkeley and HKUST. It had 84 upvotes on Hugging Face daily papers when we checked.

What routing is and why it matters

Most routing work so far has been about language models: a cheap model answers easy prompts and an expensive one gets the hard ones. The bet is that no single candidate is best on every input, so picking per input beats picking once.

The paper tests that bet across seven groups of tasks: text question answering, image-text question answering, image generation, video generation, speech generation, code and workplace agents, and robot manipulation. In total it covers 25 benchmarks and 452 candidate configurations, with 374,401 scored runs. Every candidate was run on the same inputs, which gives the authors a table of what each one would have scored.

How much a perfect router could win

The first result is a ceiling. The authors compare the best fixed candidate per benchmark (chosen on training data) with an oracle that picks the best candidate for each input after the fact. The fixed candidate averages 82.60 and the oracle 87.98, a gap of 5.4 points.

The gap varies a lot by task:

Speech has almost nothing to gain. Video generation and robot manipulation have the most. The oracle is not a real router, since it sees the scores. It only shows how much a perfect one could win.

What learned routers actually get

Real routers do much less. The paper's capability-tree router, which sorts an input into a task type and then picks a candidate, scores 79.98 on 16 benchmarks. The fixed candidate scores 79.68 and the oracle 84.80. That is a gain of 0.30 points. Video generation improves by 2.12 points and code and agents by 0.77, while text QA drops by 0.17. The router's classifier is 86 to 89 percent accurate, yet 83 to 86 percent of inputs still end up at the fixed candidate.

In a separate 11-benchmark study, neither learned router beat the fixed candidate. The outcome router, which predicts how well each candidate will do, scored 67.90 against 68.30. The authors also report that how you handle missing scores matters: dropping them turned a 0.40-point router deficit into a 2.38-point gain, so the comparison has to be matched.

Robots are the clearest win. A router built on the 4-billion-parameter Qwen3.5-4B model, with camera frames and robot state as extra input, reached 71.55 percent success on RoboTwin 2.0. The best single policy, ME-Dex, reached 70.45 percent. The gain is 1.10 points, but the 95 percent confidence interval runs from -0.40 to +2.55, so it could be zero. Time per scene fell from 14.51 to 11.05 seconds, 24 percent less. A decision takes 0.13 seconds on one H20 GPU.

The training target matters too. On robots, regressing each candidate's outcome reached 71.55 percent, against 42.80 percent for a classifier trained to name the winner.

Where the savings are: cost, not quality

The stronger results are about cost. Compared with a reference model, the paper reports efficiency gains of 25.5 times on CodeRouterBench at a quality change of -0.44 points, and 7.1 times on RouterBench at -0.40. For images and robots the gains are 1.35 times and 1.33 times. The paper reports an average of 8.82 times. The reference is Claude Opus 4.6 for CodeRouterBench and GPT-4 for RouterBench.

The authors also build a "certified floor": a statistical test that checks a route will not lose quality on new inputs. On CodeRouterBench the certified zero-loss route held its floor in 200 of 200 test splits, at a median 19.4 times efficiency. A route fitted without certification held in only 12.5 percent of splits. The lesson is that a cheap route that looks safe on training data often is not.

Verifiers are a cost too. A free compiler check gave 4.09 times dollar efficiency at a 2.4-point loss for Qwen3-Max. An LLM judge at list price gave 0.30 times, which is worse than not routing.

Limits

• The omni-modal data is proprietary and not released, according to the code repository. Trained checkpoints are not released either, so the 25-benchmark results cannot be reproduced from the public release. Public code and text data (CodeRouterBench) can be used.
• The theory assumes the inputs do not depend on the router's earlier choices. Multi-step agent tasks, where a model's earlier output becomes part of the next input, break that assumption. The authors say step-level routing for long agentic tasks needs more work.
• Certification needs the reference's outcome on every certification input.
• The robot gain is within the noise of the confidence interval.

Why it matters

If you run a product on several models, the useful part is the map of where routing pays. Code and cost-sensitive text work can save a lot. Video generation and robotics have the most quality headroom, but learned routers capture little of it yet. For speech, the data says do not bother.

The code is on GitHub under the MIT license.