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