Reasoning models have poor control over their chain of thought: 2.7% compared to 61.9% on the final answer

An evaluation suite measures the gap between a model's ability to obey an instruction in its response and its ability to apply it to its internal reasoning.

If we monitor a model by reading its chain of thought, we still need to know to what extent it can control it. This is the question posed by a team led by Chen Yueh-Han, in a work submitted on arXiv on March 5, 2026. The authors build an evaluation suite, CoT-Control, which gives the model an explicit instruction on its reasoning, then measures whether it sticks to it.

The difference measured is considerable. Claude Sonnet 4.5 complies with an instruction aimed at his chain of thought in 2.7% of cases, compared to 61.9% when the same instruction relates to his final answer. In other words, what the model produces on the surface obeys the instruction; what it carries out internally largely escapes him.

The trends observed are counterintuitive. Controllability increases with the size of the model, but decreases with reinforcement training, with the computational budget devoted to reasoning and with the difficulty of the task. The authors see this as good news in the short term for surveillance: a model that does not know how to control its chain of thought can hardly use it to hide its intentions. However, they admit that they do not explain where this phenomenon comes from.