Current Practices for Building LLM-Powered Reasoning Tools Are Ad Hoc -- and We Can Do Better

Open in new window