How is that different from machine learning 101 "regression"? And why don't they just put a regression or softmax head on top of a trained transformer? (or do they?)