Chain-of-Thought (CoT) monitoring provides actionable evidence for AI oversight, enabling early detection and intervention. Existing approaches are typically task‑specific, which limits their applicability to the broader regulation of reasoning models. We introduce the Factor mechanism, allowing monitoring requirements to be specified in the prompt so that a single monitor can handle multiple tasks by switching factors. Two factor types are defined: fine‑grained (e.g., checking whether a particular email address was leaked) and coarse‑grained (e.g., assessing overall privacy leakage), catering to diverse user needs. Building on this, we develop the CoT monitor OwlSight, which relies solely on factors to track multiple dimensions such as capability, safety, and trustworthiness. Experiments on 15,971 reasoning trajectories show OwlSight outperforms general‑purpose LLMs by 1.32‑30.57 points on fine‑grained and 11.71‑27.81 points on coarse‑grained tasks, and achieves state‑of‑the‑art results on OpenAI’s CoT monitorability benchmark. It also generalizes to unseen models’ CoTs and remains robust to factor rephrasings, CoT compression, and obfuscation, offering a promising paradigm for frontier AI oversight.
Review