Reproduce “Monitorability Is Trainable: Post-Training for Chain-of-Thought Monitorability” 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

Usman Anwar, Tim Bakker, Christos Louizos, David Krueger

The authors train a method to make chain-of-thought traces more informative about model outputs without behavioral labels or monitor feedback. After compressing reasoning, post-training on GPQA-main recovered 56.3% of mean monitorability loss on GPQA-Diamond while preserving 80.2% of token savings. The results come from controlled model-organism experiments and do not establish faithful reasoning or reliable oversight.

Summary by alphaXiv.

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