(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
Nov 11, 2026 · 6:17 PM UTC
(2/10) First, we study a set of internal and released frontier models.
Across many evals, we see that no-CoT capability (reasoning capability in a forward pass, a proxy for worse CoT monitorability) smoothly improves with better pretrained models (lower base model perplexity).
(3/10) We ask the same about CoT controllability, which measures a model’s ability to control its CoT when prompted to add, remove, or manipulate CoT tokens.
There, we also see a (weaker) correlation between lower base model perplexity and higher CoT controllability.
(4/10) Next, we focus on architecture. We compare three different architectures at shared base model perplexity levels. For the majority of no-CoT evaluations, perplexity is predictive across architectures.
(5/10) For multi-hop planning and serial computation questions, certain architectures perform better than base model perplexity predicts. In these cases, however, we note that no-CoT performance scales smoothly, but at a steeper rate with respect to perplexity.
(6/10) Afterwards, we study RL. In an Astra-class model, we find that CoT controllability increases substantially during RL even with no process supervision other than length constraints and final answer grading.
(7/10) CoT controllability increases slightly when we add tool access and tool supervision during RL on a Terra-class model. Our current results indicate that the effect of full tool process supervision is small compared to decreasing model perplexity.
1
24
2,361
(8/10) Finally, we study how decisions made in one model generation can affect the next via supervised fine-tuning (SFT) on model-generated data. To investigate, we selected two Astra-class teachers with ~matched downstream capabilities but differing CoT controllability.
(9/10) We find that CoT controllability can be distilled: training on rollouts from a more controllable teacher results in a more controllable student!
(10/10) For this pair of similarly capable models, we did not see meaningful association between CoT controllability and monitor evasion.
We report this, and elaborate more in the blog, to underscore our thinking that CoT controllability is a noisy proxy for adversarial monitorability. We are expanding our evaluation suite and interested in understanding precursors to worst-case monitorability better.
really cool work!! one thing that stood out to me was that controllability increases even without directly rewarding the reasoning text? but app B seems important for interpreting that, since the more controllable teacher wasnt better at evasion. did you test evasion in the distilled students too?
449
