“Dima from Fireworks states that using an LLM as a judge in an RL loop is effective because it is easier to judge an output against a rubric than to generate the output from scratch.”