Reproduce “Chain-of-Thought Oversight Requires Separate Evidence for Faithfulness and 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
Sichao Li, Sai Ma, Xiyang Hu, Chudi Zhong, Tongliang Liu
The authors propose a standard separating trace faithfulness from monitorability. It names properties, access tiers, artifacts, and evidence links. In security analysis, probes and an intervention support reasoning-step representation and causal relevance, but the tested jury’s exact accuracy is 0.42, below the 0.53 baseline. Authority-hint detection matches answer-only classification; tested methods do not localize a compact mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.