Stopping a model's computation before completion when sufficient confidence is reached, reducing computational cost.