Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?
arXiv:2603.30036v1 Announce Type: new
Abstract: Chain-of-Thought (CoT) monitoring, in which automated systems monitor the CoT of an LLM, is a promising approach for effectively overseeing AI systems. However, the extent to which a model’s CoT helps us…