Thank you for the work. I have a few questions on experiment 1 following the previous inquiry:
I suspect that the activation vector taken at token '.' of the answer are more game-able than if it was taken around the tokens 'chose to', for instance, because it is a position where the decision is yet-to-be-made. In other words, I think that the activation vector at token '.' has more wiggle-room to influence the AV without changing the model's behaviour than at other positions where the model's behaviour can still change, if that makes sense?
Thank you for the work. I have a few questions on experiment 1 following the previous inquiry: