Towards Reward Modeling for AI Tutors in Math Mistake Remediation

TL;DR

This paper develops a reward modeling approach for AI math tutors that assesses pedagogical quality, focusing on mistake remediation, and demonstrates improved accuracy over larger models using synthetic data and preference modeling.

Contribution

It introduces a hierarchy of pedagogical aspects for AI tutors, synthesizes contrastive response pairs, and trains reward models that outperform larger models in evaluating tutoring responses.

Findings

Best model achieves 0.69 accuracy on human preferences.

Combining synthetic data improves accuracy to 0.74.

Models outperform larger general-purpose reward models.

Abstract

Evaluating the pedagogical quality of AI tutors remains challenging: standard NLG metrics do not determine whether responses identify mistakes, scaffold reasoning, or avoid revealing the answers. For the task of mistake remediation, we derive a hierarchy of pedagogical aspects from human pairwise preferences on MRBench, and synthesize minimally contrastive response pairs that differ along key aspects (e.g., mistake identification and location, targetedness, scaffolding, actionability, clarity, and coherence). We develop and release Bradley-Terry preference models trained on weighted-sum rankings that we automatically create from MRBench, synthetic pairs, and data combinations. Using only synthetic data, our best model reaches 0.69 pairwise accuracy on a human preference test, and combining weighted-sum data with targeted synthetic groups improves accuracy to 0.74, outperforming larger general-purpose reward models while using only a 0.5B-parameter backbone.