AI RESEARCH
Reasoning Models Struggle to Control their Chains of Thought
arXiv CS.AI
•
ArXi:2603.05706v1 Announce Type: new Chain-of-thought (CoT) monitoring is a promising tool for detecting misbehaviors and understanding the motivations of modern reasoning models. However, if models can control what they verbalize in their CoT, it could undermine CoT monitorability. To measure this undesirable capability -- CoT controllability -- we