>>106335825
I don't think LLMs can meta reason about <think> tokens. This is a typical case of schizo prompt.