[arXiv]score: 0.75
Flat Minima Don't Guarantee Generalization: Sharp Minima Can Generalize Optimally in Convex Settings
May 26, 2026
In stochastic convex optimization with smooth objectives, flat empirical minima can incur Ω(1) population risk while sharp minima generalize optimally, and two SAM variants (SA-GD, SA-SGD) inherit this poor generalization behavior.
cs.LG
HOW THIS AFFECTS YOU
●
researcherThis theoretical result directly challenges the flat minima hypothesis underlying SAM and related sharpness-aware optimizers, warranting reconsideration of their use as principled generalization tools.