OpenAI Reasoning Models CoT Controllability Research
OpenAI has found that current frontier reasoning models struggle to control their chains of thought (CoT), which indicates that these models are currently unable to deliberately obscure their reasoning to evade safety monitoring. This low controllability is a positive result for AI safety, as it suggests that CoT monitoring remains a reliable method for identifying unsafe or misaligned behavior.
Understanding CoT Controllability
Chain of thought (CoT) refers to the intermediate reasoning steps an AI agent generates to solve a task. Monitoring these traces allows safety systems to assess whether an agent's behavior is aligned with user goals.
CoT controllability is defined as a model's ability to control the properties of its CoT to reduce its monitorability. Because measuring realistic incentives to evade monitors is difficult, OpenAI used a proxy: the model's ability to follow specific instructions regarding its internal reasoning process. These instructions included requirements such as:
- Using only lowercase letters in the reasoning trace.
- Avoiding specific problem-relevant keywords.
- Appending a specific word to every sentence in the CoT.
To evaluate this, OpenAI introduced CoT-Control, an open-source evaluation suite containing over 13,000 tasks derived from benchmarks including GPQA, MMLU-Pro, HLE, BFCL, and SWE-Bench Verified.
Findings: Low Controllability Across Frontier Models
Evaluations of 13 reasoning models revealed that all tested frontier models struggle to reliably control their CoTs. Controllability scores—measured as the fraction of reasoning traces where the model successfully complied with instructions—ranged from 0.1% to a maximum of 15.4%.
Key observations regarding these failures include:
- Cognitive Awareness vs. Execution: Models often recognized within their own CoT that they were failing to follow the CoT instructions, yet continued to fail regardless of this realization.
- Scale and Performance: While controllability increased slightly with model size in open-weight models, the effect tapered off at larger scales.
- Impact of Compute and Training: Controllability decreased as models were asked to reason for longer periods and as they underwent more post-training. For example, the controllability score of OLMo-3-7B-RL-Zero-Math dropped by more than a factor of ten during RL training.
- Adversarial Conditions: Informing models that they were being monitored ("monitor awareness") or evaluated ("eval awareness") slightly increased controllability, but not significantly. Automated prompt optimization provided gains on training sets that did not transfer to new instruction types.
Implications for AI Safety and Monitoring
Low CoT controllability is viewed as a safety feature rather than a flaw. It ensures that CoT monitoring is more trustworthy because models cannot easily hide their internal reasoning steps to deceive monitors.
However, OpenAI notes that low controllability is not a guarantee of high monitorability. Monitorability could still degrade through other mechanisms, such as direct optimization pressure on the chains of thought.
Future Outlook and Monitoring
OpenAI will treat CoT controllability evaluations as a "canary eval"—an early warning system to detect if models develop the ability to shape or hide their reasoning as they advance. To maintain transparency, OpenAI will begin reporting CoT controllability alongside CoT monitorability in system cards for future frontier models, starting with GPT-5.4 Thinking.