DEV Community

News
News

Posted on

OpenAI maps how training choices shape chain-of-thought monitorability as models scale

Infographic cover

๐Ÿ“… November 12, 2026 โ€ข AI Automation
November 12, 2026 โ€” OpenAI published a research post on November 10 that maps how training choices shape chain-of-thought monitorability as models scale, and write-ups of the work in outlets such as The Decoder and Jetsonsage continued today. The post, authored by Jeffrey G. Wang, Aiden Low, Zora Che, Holden Mui, Harshit Sikchi and colleagues, studies how pretraining, architecture, reinforcement learning and model-generated data affect two proxies of monitorability: no-CoT capability and CoT controllability.
No-CoT capability is reasoning ability in a single forward pass with no chain-of-thought or tool-call tokens, while CoT controllability is the ability to add, remove or manipulate chain-of-thought tokens when prompted.
Three findings stand out. First, lower base-model perplexity correlates with higher no-CoT capability, meaning models that predict text more confidently also reason better in a single forward pass, and some architectural changes improve that no-CoT reasoning as well. Second, reinforcement learning increased CoT controllability even without process supervision: the setup used only length constraints and final-answer grading, and adding grading of tool use contributed a smaller additional effect.
Third, training a student model on the outputs of a more controllable teacher produced a more controllable student, so controllability transfers through teacher distillation.
The authors state the caveat clearly: these are imperfect proxies, and controllability is a noisy stand-in for monitorability, so none of the measurements should be read as a direct verdict on safety. Even so, the indicators taken together suggest that models will become harder to monitor through their chain of thought as capabilities scale.
The context matters: chain-of-thought monitoring is a key part of OpenAI scalable oversight and a core pillar of its safety cases, and the post notes that monitorability can also come from actions or activations, not only from written reasoning.
Per The Decoder, frontier models are becoming harder for safety monitors to see into, and the outlet frames the loss of transparency as tied to decisions about performance and cost rather than to a law of nature.
For automation teams the research lands close to home: agents in production are audited through logs, traces and reasoning summaries, and this work explains why reasoning traces need deliberate engineering attention as models scale, because what reads clearly today may not stay equally readable in the next generation, so observability and audit trails deserve the same design effort as the automations themselves.
Sources: alignment.openai.com โ€” the-decoder.com โ€” jetsonsage.com

Key figures

IMPACT checklist


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)