On Adaptivity in Zeroth-Order Optimization
Hassan Dbouk, Nidham Gazagnadou, Matthias Reisser, Christos Louizos · 2026
Read On Adaptivity in Zeroth-Order Optimization by Hassan Dbouk, Nidham Gazagnadou, Matthias Reisser, Christos Louizos (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-2605-03869.txt.
Description
Hassan Dbouk, Nidham Gazagnadou, Matthias Reisser, and Christos Louizos investigate adaptive zeroth-order optimization for memory-constrained fine-tuning of large language models. Zeroth-order methods estimate gradients from function values alone, saving memory, and adaptive versions such as ZO-Adam were claimed to converge better. The authors show those claims do not hold against well-tuned plain ZO-SGD.
Their analysis reveals that in high dimensions zeroth-order gradients lack coordinate-wise heterogeneity, so per-coordinate adaptation wastes memory. They propose MEAZO, which tracks only a single scalar for global step-size adaptation, supported by convergence guarantees under standard assumptions. Across several model families and tasks, MEAZO matched ZO-Adam's performance with ZO-SGD's memory footprint, and proved more robust to step-size choices in block-structured settings.
Catalog record
| CREATOR | Hassan Dbouk, Nidham Gazagnadou, Matthias Reisser, Christos Louizos |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 137 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.03869