Towards Pedagogically Aligned LLM Tutors for Math Mistake Remediation

The paper presents a two-stage alignment pipeline for enhancing large language models (LLMs) as intelligent tutors for math mistake remediation. It involves supervised fine-tuning on tutoring dialogues and Direct Preference Optimization using a dataset that combines existing tutoring corpora with synthetic data focusing on pedagogical dimensions. Results indicate improved factual accuracy and pedagogical quality compared to baseline models, demonstrating the potential for more effective and transparent tutoring systems, although challenges in evaluating tutoring quality remain.