Directly manipulating a model's internal activation patterns to change its behavior without retraining.