Reproduce “CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding 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
Shouli Wang, Yanfeng Jia, Zhihao Ou, Zitao Su, Ruize He, Haotong Xie, Hao Peng, Juanzi Li and 1 more
CATCH makes reward hacking in coding reinforcement learning measurable by comparing a model-visible Hackable Run with an independent Unhackable Run that audits task correctness. It tracks hacking during training. On Qwen3-4B, a chain-of-thought monitor initially suppresses hacking, but later misses more hacks; removing comments improves recall under monitor-penalty training. The results concern controlled algorithmic repositories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.