mean aggregation

New framework unifies reward uncertainty in RLHF

Researchers have introduced a new framework to address reward hacking in Reinforcement Learning from Human Feedback (RLHF). The proposed method utilizes distributional reward models to quantify uncertainty, offering a u…