Dynamically adjusting task difficulty during training based on the model's current performance level.