HeyHomees Lab Notes · No. 16 · Published November 4, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Per-input oracle routing scores 5.4 points higher than the best fixed candidate (87.98 vs 82.60) on 19,814 inputs across 24 benchmarks.
- Learned routers beat the best fixed candidate by 0.3 points overall, adding 2.1 points on video generation and 1.1 points on robot manipulation while using 24% less policy time.
- Routing raises average efficiency by 8.82× versus a reference model, with up to 25.5× on code and 7.1× on text QA.
What is it?
Route Anything is a research framework that studies model and policy routing for general-purpose AI systems, providing theory, a large omni-modal trajectory matrix, and code for routers that select among language, image, video, speech, and robot candidates.
How does it work?
The router chooses the candidate with the highest expected value minus a priced cost based on the information it observes; the paper defines a reference policy and quality tolerance to fix the price, and evaluates both oracle and learned routers on the matrix.
How well does it do?
Per-input oracle routing outperforms the best fixed candidate by 5.4 points; learned routers improve by 0.3 points overall, 2.1 points on video generation, and 1.1 points on robot manipulation while cutting policy time by 24%; routing improves efficiency by 8.82× on average.
What does it need to run?
The released repository contains router implementations and evaluation utilities, but the proprietary omni-modal trajectory matrix and trained checkpoints are not included; reproducing the full scores requires separate authorized data.
By the numbers
- Oracle routing score advantage: 5.4 points
- Learned router overall gain: 0.3 points
- Video generation gain: 2.1 points
- Robot manipulation gain and policy time reduction: 1.1 points, 24% less policy time
The honest read
The reported gains come from the authors' own experiments on a proprietary omni-modal trajectory matrix; the oracle and learned router scores are self-reported and cannot be reproduced without the private data. Efficiency numbers compare to a reference policy defined by the authors and may not reflect real-world deployment costs. The code release does not include the matrix or trained model checkpoints, limiting direct replication of the headline results. The code repository provides router implementations and utilities, but reproducing the full set of benchmark scores requires separate authorized data, as noted by the authors. Thus the figures illustrate potential improvements under the study’s conditions, not guaranteed performance for all users.
What we don't know yet
- Exact hardware or compute requirements for running the routers
- Public availability of the omni-modal trajectory matrix or trained checkpoints
- Details of the reference policy used for efficiency comparisons
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.