Reasoning Models Struggle to Control their Chains of Thought

Researchers found that AI reasoning models struggle to control their chain-of-thought (CoT) outputs, with Claude Sonnet 4.5 able to control its CoT only 2.7% of the time versus 61.9% for final outputs. This limitation suggests CoT monitoring remains viable for detecting AI misbehavior, though the underlying mechanisms are poorly understood.

• →AI models show significantly lower ability to control their chain-of-thought reasoning compared to final outputs.
• →Larger models exhibit higher CoT controllability, while more RL training and increased problem difficulty reduce it.
• →CoT controllability failures persist even when models are incentivized to evade monitoring systems.
• →Current low controllability levels suggest CoT monitoring remains effective for detecting AI misbehavior.
• →Researchers recommend frontier AI labs track CoT controllability as models advance.