GEAR-SAM Optimizes Sharpness-Aware Minimization via Block-Wise Gradient Energy

GEAR-SAM improves generalization by using an exponential moving average of squared block gradients to adaptively allocate the perturbation budget. This method replaces noisy instantaneous gradient norms with a curvature-related sensitivity signal without requiring Hessian-vector products or extra memory beyond a scalar state.