Removes harmful/unwanted concepts without retraining via orthogonal representation editing; preserves non-target quality and diversity.