This paper introduces a two-stage alignment pipeline to make Large Language Models (LLMs) better at tutoring math by following effective teaching strategies for mistake remediation.

In plain terms

Large Language Models (LLMs), while powerful, often struggle to act like good human tutors, sometimes revealing answers instead of guiding students. This paper addresses this by developing a method to train LLMs to follow effective teaching strategies, specifically for helping students correct math mistakes. The researchers used a two-step process: first, "supervised fine-tuning" involved training the LLM on existing tutoring conversations. Second, "Direct Preference Optimization" was used, where the LLM learned by being shown pairs of responses and instructed which one was pedagogically better (e.g., providing hints, known as scaffolding, versus giving direct answers). They also created a new dataset combining real and synthetic tutoring examples focused on pedagogical quality. Their approach significantly improved both the factual accuracy of the math help and the teaching quality of the LLM, with human evaluations showing it to be comparable to leading proprietary AI tutors.