Research Shows Models Know Answers Before Finishing Chain-of-Thought Reasoning