Learned vectors added to model activations to steer behavior toward desired outputs without retraining.