Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To
The paper argues that reasoning models are much better at controlling properties of their chain-of-thought when the constraint is phrased as something the model “tells itself” inside its reasoning, rather than as a normal user instruction, and shows this can even be induced with generic reminders or bootstrapped from the prompt. The neat insight is that poor CoT controllability may be partly an interface/prompting artifact rather than a hard capability limit, but the same mechanism also makes CoT-based oversight look shakier because models can use self-instructions to evade monitors, and the results seem tied to specific prompting/evaluation setups rather than proving broad, robust control.
Abstract
Chain-of-thought (CoT) controllability, the ability of reasoning models to control properties of their CoT, is consequential for both their usefulness and reliable oversight. Recent work evaluates this capability by measuring whether models follow CoT constraints given in the user prompt, and reports very low controllability across models. We argue that this low controllability may not reflect an inherent inability of reasoning models to control their CoT, but instead result from naively specifying constraints in the user prompt. To test this, we introduce Self-Instructed CoT Controllability: a complementary evaluation that places the constraint as a self-instruction inside the model's CoT and measures its ability to follow it. We show that, under self-instruction, models achieve significantly higher controllability than with standard user instructions. Even a generic reminder to follow the user's constraint substantially improves controllability. Moreover, effective self-instructions can be bootstrapped into the reasoning through a simple instruction in the user prompt, enabling improved control even without access to the reasoning channel. Finally, using self-instruction, we demonstrate in a realistic monitoring setup that CoT controllability can be leveraged for monitor evasion, posing a risk to CoT monitorability.