Rethinking the Role of PPO in RLHF

Rethinking the Role of PPO in RLHF TL;DR: In RLHF, there’s tension between the reward learning phase, which uses human preference in the form of comparisons, and the RL fine-tuning…