Opening-token controls raise CoT compliance by 34-41 percentage points
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
LLM chain-of-thought reasoning paths are severely constrained by an entropy collapse in the first few tokens following the opening thinking tag, but banning these frequent opening tokens combined with in-context examples increases reasoning compliance by 34 to 41 percentage points. For production agents, you can actively steer and inject specific information into model reasoning traces at inference time by simply prefilling a single token after the thinking tag. However, while this cheap inference-time intervention excels at forcing the model to add or transform its reasoning steps, it remains largely ineffective at suppressing or hiding specific information from the trace.
The safety assumption that models can't strategically control their reasoning traces is weaker than it looks: entropy collapses in the first few tokens after the thinking tag, and prefilling even a single token—or banning common opening tokens plus in-context examples—raises compliance with controllability demands by 34–41 points. This means CoT monitoring is more spoofable than assumed, but only for adding or restyling reasoning content; suppressing information behaves as a separate, harder-to-steer capability, so don't treat "monitorability" as a single number—evaluate addition/style and suppression controllability independently or your oversight will give false assurance.