Fetching the paper…

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