Fix the Loss, Not the Radius: Rethinking the Adversarial Perturbation of Sharpness-Aware Minimization
Jinping Wang, Qinhan Liu, Zhiwu Xie, Zhiqiang Gao · 2026
Read Fix the Loss, Not the Radius: Rethinking the Adversarial Perturbation of Sharpness-Aware Minimization by Jinping Wang, Qinhan Liu, Zhiwu Xie, Zhiqiang Gao (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-2605-10183.txt.
Description
Wang and three colleagues rethink Sharpness-Aware Minimization, a training method that improves generalisation by shrinking the worst-case loss inside a fixed parameter-space radius. SAM and its variants lean on a first-order linearised surrogate, yet flat minima are by nature a second-order curvature notion. That mismatch motivates Loss-Equated SAM, which flips the mechanism: a fixed loss-space budget replaces the fixed perturbation radius.
Fixing the loss budget removes gradient-norm-dominated learning signals and shifts optimisation toward curvature-dominated terms. Across diverse benchmarks and tasks, LE-SAM consistently beats SAM and its variants, reaching state-of-the-art performance. The paper's message is in its title: fix the loss, not the radius, since the radius is the wrong knob for a curvature concept.
Catalog record
| CREATOR | Jinping Wang, Qinhan Liu, Zhiwu Xie, Zhiqiang Gao |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.10183