Each example’s weight changes with its predicted probability during training, decreasing as the model becomes more confident.