Comment by Crye

It mentioned not penalizing/rewarding the model for thoughts only rewarding the answer after the thought. I am curious how back propagation works then.