OpenAI published a research analysis on November 10 examining how pretraining, architecture, reinforcement learning and model-generated data affect chain-of-thought oversight. For researchers and universities assessing AI agents, the distinction between a warning indicator and demonstrated monitoring failure is central to interpreting the findings. The release appears in OpenAI’s research index.
What the training experiments found
The company’s study examines two proxies: reasoning without chain-of-thought or tool-call tokens, and controllability, meaning the ability to follow instructions that alter reasoning traces. OpenAI reports:
- Pretraining: lower perplexity, a measure of prediction error, correlated with stronger reasoning without a trace.
- Architecture: some designs added gains on planning and serial computation beyond those predicted by perplexity.
- Reinforcement learning: controllability rose even with only length constraints and final-answer grading; tool supervision added a smaller increase.
- Model-generated data: students trained on more-controllable teachers’ outputs became more controllable.
The experiments include internal models. The reasoning-disabled Astra configuration used in no-trace evaluations is outside the public release. The tool experiment also incompletely separates supervision from perplexity.
Higher controllability did not mean greater evasion
Appendix B compares similarly capable models: the more-controllable one did not evade better in the tested math side-task and multiple-choice sandbagging evaluations. Results were indistinguishable or favored greater evasion by the less-controllable model. These proxies do not establish deceptive intent or the impossibility of monitoring.
The distinction builds on earlier CoT-Control research, which tested whether models could solve problems while following constraints on their reasoning text. That work recommended tracking controllability in future models. It provides background for understanding chain-of-thought monitorability as an oversight technique, rather than treating a controllability score as a direct observation of misconduct.
What this means for university AI oversight
Earlier Monitoring Monitorability research found that access to reasoning traces improved monitoring over action-only checks in its tested settings, while monitorability remained imperfect. Separately, subliminal learning experiments showed that teacher-generated data could transmit behavioral traits under studied conditions. Both provide context for examining oversight across training stages.
For research offices, the practical implication is to distinguish three kinds of evidence: task performance, compliance with reasoning-text constraints, and success at detecting unwanted behavior. An assessment of a model evaluation suite should identify which of these it measures. Universities interpreting model cards and system cards should also distinguish results for internal research configurations from evidence about the public product they use. These are implications for evaluating evidence, rather than new institutional requirements.
Sources
- OpenAI, Research and Releases. Checked 2026-11-14.
- OpenAI, What affects chain-of-thought monitorability across the model lifecycle?, including Appendix B. Checked 2026-11-14.
- Chen Yueh-Han and colleagues, Reasoning Models Struggle to Control their Chains of Thought. Checked 2026-11-14.
- Melody Y. Guan and colleagues, Monitoring Monitorability. Checked 2026-11-14.
- Alex Cloud and colleagues, Subliminal Learning: Language models transmit behavioral traits via hidden signals in data. Checked 2026-11-14.
