Reproduce “Chain of Thought Monitoring Can Fail Even When CoT Is Necessary” 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
Lennart Bürger, Piotr Komorowski, Moritz Weckbecker, Evgenii Kortukov, Aakriti Jain, Sebastian Lapuschkin, Wojciech Samek
The authors test whether chain-of-thought (CoT) monitoring reveals input attributes affecting answers. They remove an attribute, compare answer distributions, and assess whether reasoning articulates its influence. For Qwen3-8B on Thyroid Disease, sex and query on thyroxine had at least 19% and 24% systematic influence, yet were almost never articulated. These findings concern benign tasks; they do not establish failures in safety-relevant settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.