AI Researchers from NVIDIA and the University of Maryland Propose ODIN: A Reward Disentangling Technique that Mitigates Hacking in Reinforcement Learning from Human Feedback (RLHF)
