Is there any difference between this and the knowledge distillation of GPT-4o ?

[ningkp]

reasoning by GPT-4o + Rule-based Reward + GRPO = reasoning by GPT-4o + SFT

This is not a real RL; it is supervised learning. Just like the image classification, and the reward is the $1{\hat{y}==y^*}$. It also can employ RL optimization (like PPO or GRPO) to learn the supervised model, but it is not a real RL.

[HaitaoWuTJU]

agree with that, more like SFT