Reproduce “Stress Testing Chain-of-Thought Monitoring Against Covert Misalignment” 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

Yao Huang, Yifan Wang, Yitong Sun, Yichi Zhang, Yinpeng Dong

TraceMRT tests whether task-conditioned red-teaming can evade chain-of-thought monitors and evaluates Bayesian Trajectory Monitoring as a defense. TraceMRT reframes fixed hidden side goals as task-specific sub-goals and searches for less suspicious reasoning traces; the monitor combines whole-trace intent with selected local actions. On SHADE-Arena, evasion rises from 27.56% for MRT to 59.94%, in a controlled setting, not spontaneous misalignment.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.