LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss

The growing use of generative models in daily life calls for efficient mechanisms to control their generation, to e.g.