Uncertainty-Aware Reward Discounting for Mitigating Reward Hacking

Disha Singha

Reinforcement learning from human feedback (RLHF) systems face a compounding alignment challenge: not only are learned reward models uncertain about unseen state-action pairs, but…