Interactive LLM-based Tutoring of Math and Physics Students

Large language models (LLMs) show promise for educational applications but frequently hallucinate incorrect answers, making them unreliable for tutoring. Even on well-known problems, LLMs often produce factually incorrect responses, misinterpret prompts, or fail to follow instructions. We address this issue with a chain-of-thought-based grounding method, requiring the model to generate and internally verify a structured reasoning process before interacting with students. We implement this approach in CᴏTᴜᴛᴏʀ, a tutoring system designed to provide mathematically rigorous, self-verified explanations.