Reproduce “CheatBench: Measuring Reward Gaming in AI Agents” 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
Long Phan, Stephen K. Yang, Jason J. Lim, Mantas Mazeika, Wenyu Zhang, Zheyuan Liu, Richard Ren, Jingxiang Meng and 5 more
CheatBench measures whether agents attempt shortcuts that violate expectations across ten categories. It gives agents work, makes a shortcut discoverable, and defines which action crosses the line; ordinary exploration alone does not count. The equally weighted category averages range from 11.2% for Claude Opus 5.5 to 77.9% for Grok 4.7. These results measure planted opportunities, not real-world cheating frequency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.