“A limitation of training LLMs to reason via Chain of Thought is the requirement for human-labeled reasoning traces, which do not exist for unsolved problems like the Millennium Prize Problems.”