推理模型难以控制其思维链

Chain-of-thought (CoT) monitoring serves as a promising tool for detecting misbehaviors and understanding the motivations behind modern reasoning models. However, if these models can control what they verbalize in their CoT, it may undermine the monitorability of CoT. To assess this undesirable capability—CoT controllability—a new evaluation suite called CoT-Control has been introduced. This suite includes tasks that require models to solve problems while adhering to specific CoT instructions, such as reasoning about a genetics question without using the term 'chromosome'. The findings reveal that reasoning models exhibit significantly lower CoT controllability compared to output controllability; for instance, Claude Sonnet 4.5 can control its CoT only 2.7% of the time, while it achieves 61.9% control over its final output.