Each example’s weight is adjusted in proportion to \(1-p\) as its predicted probability changes during training.