Towards Last-Layer Retraining for Group Robustness with Fewer Annotations | Read Paper on Bytez