Published March 5, 2026 alongside the GPT-5.4 launch by Tomek Korbak, Micah Carroll, Bowen Baker, and Ian Kivlichan of OpenAI; paper at arXiv:2603.05706.
The question
The concern is a specific failure path for chain-of-thought monitoring: "if agents become capable of deliberately reshaping or obscuring their reasoning when they know they are being monitored, evaluations could overestimate a system's alignment or safety, and monitoring systems could become less reliable."
CoT controllability is defined as "a reasoning model's ability to control properties of its CoT in ways that reduce monitorability."
The proxy, and its limit
OpenAI is explicit that it is not measuring the thing of concern. Measuring controllability "under the full breadth of realistic incentives to evade monitors we may expect in practice is challenging. Instead, we use a practical proxy: we measure the model's ability to follow user instructions about their CoT" — using only lowercase, avoiding keywords, appending phrases.
The limitations section restates this: "an ideal CoT controllability evaluation would involve testing whether models that naturally pursue misaligned goals can control their CoT to evade monitoring," and CoT-Control "might be over- or underestimating controllability by measuring only a narrow set of controllability types."
CoT-Control is an open-source suite of over 13,000 tasks pairing problems from GPQA, MMLU-Pro, HLE, BFCL, and SWE-Bench Verified with a CoT instruction.
Findings
Across 13 reasoning models — including Claude 3.7 Sonnet, Claude Sonnet 4 and 4.5, OpenAI o3 and o4-mini, GPT-5 through GPT-5.4 Thinking, gpt-oss-20b and 120b, Qwen3 14B and 32B, and DeepSeek R1 — "CoT controllability scores are low, ranging from 0.1% and not exceeding 15.4% for any current frontier model we tested."
The failure is not comprehension: "models often recognize in their CoT that they're failing to comply with the CoT instructions they were given and—even after this realization—still continue to fail." The paper reproduces a GPT-5.2 Thinking trace in which the model, instructed not to use "XOR" in its reasoning, notices the violation and then uses the word four more times while deliberating about having used it.
Scaling behaves in opposite directions on different axes: controllability "increases with model size" but "quickly decreases with the amount of post-training and test-time compute." Letting models think longer reduces control, and RL training on OLMo-3-7B-RL-Zero-Math dropped its controllability score "by more than a factor of ten over the course of training."
Under adversarial conditions — telling models they are monitored or being evaluated — controllability rises "slightly (but noticeably)." Automated prompt optimization raised controllability on a training set, but "these gains transfer poorly to new instruction types."
Interpretation and hedges
The argued conclusion is that "low CoT controllability is a feature, not a bug," making CoT monitoring "more trustworthy as a safeguard." Two hedges accompany it. OpenAI states it does "not yet understand why current reasoning models' controllability is low—or whether this property will persist as models become more capable," proposing the measure as "a canary eval." And low controllability "is not sufficient on its own to guarantee high monitorability because monitorability could still degrade due to other mechanisms such as direct optimization pressure on CoTs."
The commitment attached: CoT controllability will be reported alongside CoT monitorability "in system cards for future frontier models, starting with GPT-5.4 Thinking," alongside OpenAI's stated commitment "to not applying optimization pressure directly to the CoTs of frontier reasoning models."
Relationships
- supports: Reasoning Models and Chain-of-Thought — supplies the measured basis for treating CoT monitoring as currently reliable
- contradicts: GPT-5.6 Preview System Card (OpenAI, June 2026) — the canary this paper proposes moved adversely three months later, with GPT-5.6 Sol's controllability higher than GPT-5.5's and GPT-5.4 Thinking's and the cause stated as under investigation
- related: Monitorability Tax — the optimization-pressure mechanism the paper names as an unaddressed degradation path, and where the three readings are compared
- related: GPT-5.4 Thinking — the first system card to report the metric
- related: Bowen Baker — co-author, and lead author of the earlier CoT-monitoring work
- related: OpenAI, Scalable Oversight, Alignment Auditing