I'm not an expert on transformer networks, but it doesn't logically follow that more computation = a better answer. It may just mean a longer answer. Do you have any evidence to back this up?