OpenAI maps what makes model reasoning harder to monitor
OpenAI Alignment researchers examined how pretraining, architecture, reinforcement learning, and model-generated data affect chain-of-thought monitorability across the model lifecycle. They found that lower base-model perplexity and scaled reinforcement learning were associated with greater no-CoT capability and CoT controllability, while architecture and supervision also influenced those proxies. The researchers warned that these evaluations are noisy and imperfect proxies for monitor evasion, although OpenAI has tracked such capabilities in safety cases since GPT-6 Astra.
The findings show that relying on visible chains of thought as a scalable oversight tool becomes more difficult as training choices improve capability, motivating additional monitoring research for frontier AI systems.
What affects chain-of-thought monitorability across the model lifecycle?
How pretraining, architecture, reinforcement learning, and model-generated data affect chain-of-thought (CoT) monitorability.
November 10, 2026 Jeffrey G. Wang*, Aiden Low*, Zora Che*, Holden Mui*, Harshit Sikchi* Yong Zheng-Xin, Tomek Korbak❦, Bill Stavroulakis, Alex Barron, Shibani Santurkar, Kristian Georgiev Micah Carroll†, Ben Sokolowsky†, Katherine Lee†, Jakub Pachocki† * Equal contribution · ❦ Work done while at OpenAI · † Equal advising
Summary
We studied how pretraining, architecture, reinforcement learning, and model-generated data shape two proxies of CoT monitorability: no-CoT capability and CoT controllability.
Lower base-model perplexity and some architectural changes improve no-CoT reasoning, while RL and model-generated data can increase CoT controllability. These are imperfect proxies for monitorability, but together they indicate that models will become harder to monitor through their CoT as we scale toward more intelligent systems.