Chain‑of‑thought (CoT) controllability measures how well a reasoning model can steer properties of its own reasoning trace, a factor critical for both utility and safe oversight. Prior work evaluates this by placing constraints directly in the user prompt and finds that most models exhibit very low controllability. We argue that such poor performance reflects the naive way constraints are specified rather than an inherent limitation of the models. To investigate, we introduce Self‑Instructed CoT Controllability, which embeds the constraint as an instruction inside the model’s own CoT and then measures compliance. Results show that, under self‑instruction, models achieve markedly higher controllability than with standard user prompts. Even a generic reminder like “follow the user’s constraint” substantially boosts performance. Moreover, a simple instruction added to the user prompt can bootstrap effective self‑instructions into the reasoning chain, enabling better control without direct access to the CoT channel. Finally, using self‑instruction in a realistic monitoring setup demonstrates that CoT controllability can be exploited for monitor evasion, highlighting a risk to CoT monitorability.
Review