[bigyabai]

RLHF is basically a fancy, overengineered GAN. Most of the industry could see that DPO was more efficient for fitting to human behavior.