Paper: Stability Analysis of Sharpness-Aware Minimization
Authors: Hoki Kim, Jinseong Park, Yujin Choi, Jaewook Lee
Venue: ICML 2026
Our lab paper “Stability Analysis of Sharpness-Aware Minimization” has been accepted to ICML (International Conference on Machine Learning) 2026, one of the most prestigious venues in machine learning.
Sharpness-Aware Minimization (SAM) has recently drawn attention as a leading optimization technique for improving generalization performance. Contrary to common intuition, however, this work proves theoretically that SAM can become trapped at saddle points, and validates the finding through extensive experiments. The team further identifies momentum and batch size as key factors that mitigate this instability.
By clarifying the limits and stabilization conditions of SAM-based optimization, the paper provides theoretical and practical guidelines that practitioners can rely on when deploying SAM in real training pipelines.
The paper will be presented at ICML 2026.