New Study Finds CoT Control Starts Early

The result complicates the reassuring idea that models simply cannot control their reasoning: some apparent limits may be decoding-start artifacts that inference-time interventions can unlock.

• –A single prefixed token can substantially increase controllability.
• –Banning frequent openings plus in-context examples produces the largest gains.
• –Addition and style control improve more reliably than suppression, suggesting distinct capabilities.
• –CoT-monitoring evaluations may underestimate models’ ability to reshape reasoning after the opening tokens.
• –The work does not demonstrate deliberate monitor evasion, so the safety implications remain suggestive rather than conclusive.