arxivcs.LGcs.AI2026-07-11
When Data Imbalance Helps: Robust Generalization Through Shortcut Saturation
Cheng-Ting Chou, Duc Binh Hoang
We study robust generalization under spurious correlations: tasks where a shortcut feature is correlated with the true label in training but anti-correlated in an adversarial held-out split. Varying the spurious ratio $r$ (the fraction of training examples where shortcut = true l…