{"id":"arxiv/2610.29192","url":"https://arxiv.org/abs/2610.29192","tags":["math.OC","cs.LG","stat.ML"],"score":0,"title":"Curvature-Scaled Sharpness-Aware Minimization: Convergence Beyond Smoothness","domain":"math.OC","source":"arxiv","authors":["Dimitris Oikonomou","Nicolas Loizou"],"reviews":0,"subject":"math.OC","summary":"Sharpness-Aware Minimization (SAM) has established itself as a widely used method for training deep neural networks. Despite its popularity, existing convergence guarantees for SAM-type methods require smoothness, a restrictive assumption. Generalized smoothness conditions have recently been proposed in the analysis of gradient methods but remain underexplored for SAM. To analyze the effect of SAM's perturbation under generalized smoothness, we use relative gradient distortion to quantify how the sharpness-aware direction departs from the original gradient. Under $(L_0,L_1)$-smoothness, we…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-21","reactions":{}}