I would guess that reasoning models would generalize better (i.e. have a smaller discrepency between stuff in the training set and stuff out of it) but it would

I would guess that reasoning models would generalize better (i.e. have a smaller discrepency between stuff in the training set and stuff out of it) but it would be very interesting to check.