Chain‑of‑thought (CoT) monitoring is seen as a promising way to supervise increasingly capable models, partly because current models appear limited in what they can verbalize during reasoning. We identify the source of this limitation as an entropy collapse occurring in the first few tokens after the opening thinking tag. Leveraging this insight, we propose two inference‑time interventions to recover CoT controllability:
- Prefilling – inserting even a single token at the start can substantially boost controllability.
- Banning frequent opening tokens combined with in‑context examples – prohibiting common opening tokens while providing examples raises average compliance by 34‑41 percentage points.
These interventions mainly restore addition and style controllability (adding information or altering expression), whereas suppression controllability (avoiding or removing information) shows more variable improvements, suggesting distinct underlying mechanisms. Across models, performance within each group correlates positively, while cross‑group correlations are largely negative, indicating separate capabilities. Regarding monitorability, addition and style controllability tend to correlate positively, whereas suppression often shows negative associations.
Our findings motivate the development of controllability evaluations that distinguish the difficulty of initiating a specific opening from the general ability to modify reasoning content, and to report control tasks and categories separately.
Review