If they don't actually help then the attention weight for that connection would tend to 0, right? Then it becomes a problem of overfitting which we have a large arsenal to combat.
Not necessarily. That is what this paper is about. From what I understand, they also consider graph attention networks.