Combining generative and verification components could be a productive way to use LLMs for reasoning and planning tasks. What are your thoughts on this?