“Using a "teacher" model as a judge in a reinforcement learning setup is a viable path to training a "student" model that surpasses the teacher's performance.”