Insofar as you mean to imply that “negative vectors” are obviously comparable to our technique, I disagree. Those are not activation additions, and I would guess it’s not particularly similar to our approach. These “task vectors” involve subtracting weight vectors, not activation vectors. See also footnote 39 (EDIT: and the related work appendix now talks about this directly).
Redwood Research used to have a project about trying to prevent a model from outputting text where a human got hurt, which IIRC, they did primarily by trying to fine tunes and adversarial training. (Followup). It would be interesting to see if one could achieve better results then they did at the time through subtracting some sort of hurt/violence vector.
Page 4 of this paper compares negative vectors with fine-tuning for reducing toxic text: https://arxiv.org/pdf/2212.04089.pdf#page=4
In Table 3, they show in some cases task vectors can improve fine-tuned models.
Insofar as you mean to imply that “negative vectors” are obviously comparable to our technique, I disagree. Those are not activation additions, and I would guess it’s not particularly similar to our approach. These “task vectors” involve subtracting weight vectors, not activation vectors. See also footnote 39 (EDIT: and the related work appendix now talks about this directly).