MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking

Open in new window