Muown: Row-Norm Control for Muon Optimization
Kai Lion 1 Florian Hübler 1,2 Bingcong Li 1 Antonio Orvieto 3 Niao He 1 1 Department of Computer Science, ETH Zurich, Switzerland 2 Department of Mathematics, School of Computation, Information and Technology; Technical University of Munich, Germany 3 ELLIS Institute Tübingen, MPI-IS, Tübingen AI Center, Germany {kai.lion,florian.huebler,bingcong.li,niao.he}@inf.ethz.ch [email protected] · 2026
Read Muown: Row-Norm Control for Muon Optimization by Kai Lion 1 Florian Hübler 1,2 Bingcong Li 1 Antonio Orvieto 3 Niao He 1 1 Department of Computer Science, ETH Zurich, Switzerland 2 Department of Mathematics, School of Computation, Information and Technology; Technical University of Munich, Germany 3 ELLIS Institute Tübingen, MPI-IS, Tübingen AI Center, Germany {kai.lion,florian.huebler,bingcong.li,niao.he}@inf.ethz.ch [email protected] (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-2605-10797.txt.
Description
Kai Lion and four colleagues present Muown, a 2026 tweak to the Muon optimizer, a training method that rivals AdamW for preparing large language models. Muon training can let weight matrices drift upward in size, and the authors trace that drift to the size of matrix rows rather than to how aligned those rows are.
Muown treats the row-size vector as its own variable to control, while leaving Muon in charge of the remaining direction part. The authors prove it reaches the best known rates for rough, non-convex problems in both exact and noisy settings, with noise levels below Muon's in tests. Practitioners get a drop-in swap that steadies large-scale training without changing the rest of the recipe.
Catalog record
| CREATOR | Kai Lion 1 Florian Hübler 1,2 Bingcong Li 1 Antonio Orvieto 3 Niao He 1 1 Department of Computer Science, ETH Zurich, Switzerland 2 Department of Mathematics, School of Computation, Information and Technology; Technical University of Munich, Germany 3 ELLIS Institute Tübingen, MPI-IS, Tübingen AI Center, Germany {kai.lion,florian.huebler,bingcong.li,niao.he}@inf.ethz.ch [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 82136 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.10797