Each example’s weight changes during training in proportion to \(1-p\), reducing the weight of confident examples.