ToW: Thoughts of Words Improve Reasoning in Large Language Models