For some tasks, forcing the model to answer with only a truncated version of its chain of thought often causes it to come to a different answer, indicating that the CoT isn’t just a rationalization. The same is true when we introduce mistakes into the CoT.

→ View original post on X — @anthropicai