Edge of Stability Selectively Shapes Learning Across the Data Distribution

In three linesThe study shows that edge of stability (EoS) selectively redistributes learning across data subgroups. Two conditions enable a group to benefit: alignment of its aggregate gradient with the top Hessian eigenvector, and sustained non-vanishing gradient magnitude. Under cross-entropy loss, gradient saturation favors output-outliers while suppressing progress on confidently classified groups.

Summary generated by Claude — human-verified