Reproduce “Designing Effective Monitor-Based Interventions for Mitigating Reward Hacking During RL” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aria Wong, Joshua Engels, Neel Nanda
The authors study monitor-based interventions against reward hacking in Qwen3-4B across three task settings. They compare probes and judges with synthetic monitors whose errors are randomized, and test penalty designs. Synthetic monitors can outperform realistic monitors despite lower initial accuracy. A medical judge reduced test-set hacking by 22.4 percentage points with a 0.3-point performance increase, but no intervention eliminated hacking without performance costs across the study.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.