R2-Router : A New Paradigm for LLM Routing with Reasoning
Jiaqi Xue, Qian Lou, Jiarong Xing, Heng Huang · 2026
Read R2-Router : A New Paradigm for LLM Routing with Reasoning by Jiaqi Xue, Qian Lou, Jiarong Xing, Heng Huang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02823.txt.
Description
Jiaqi Xue and 3 co-authors let a router choose not just which language model answers, but how long its answer may be, in a 2026 paper. Routers usually predict each model's quality and cost per query, then pick the best affordable one. That assumes a fixed quality and cost per model, ignoring that the same model performs differently at different output lengths.
Their R2-Router treats the output length budget as a controllable variable, enforced through length-constrained instructions. A powerful model with a capped output can beat a weaker model at comparable cost, a configuration invisible to earlier routers. The team also built R2-Bench, the first routing dataset capturing model behavior across output length budgets. Experiments show state-of-the-art routing performance at four to five times lower cost than existing routers. Routing becomes a reasoning problem about model and budget jointly.
Catalog record
| CREATOR | Jiaqi Xue, Qian Lou, Jiarong Xing, Heng Huang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-02823