paper / ian-goodfellow / May 9 / failedMONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking