Adjusting computational cost during inference by varying model behavior (e.g., loop counts) without retraining.