arXiv: MONA Extension — Reward-Hacking Mitigation via Myopic Optimization with Non-Myopic Approval in Agent Environments

arXiv·low signal

Heath extends the MONA (Myopic Optimization with Non-myopic Approval) framework in the Camera Dropbox environment, demonstrating that restricting an agent's planning horizon while supplying far-sighted approval as a training signal can mitigate multi-step reward hacking. The paper reproduces core MONA results and introduces learned approval as a scalable alternative to ground-truth approval signals, with implications for autonomous agent deployment where reward hacking remains a persistent failure mode.