Activation steering is a technique for controlling large language models by adding a direction to hidden activations during inference. The idea is that a direction representing a concept can nudge the model's behavior without retraining.
For steering to work well, the direction must be a pure signal for the target concept. If the direction is entangled with other features, such as style or topic, the steering will introduce unwanted changes.
The paper proposes contrastive covariance subspaces as a way to isolate the target concept. By contrasting covariance structures, the method aims to identify a subspace that captures the concept while leaving confounds behind.