Towards Reward Modeling for AI Tutors in Math Mistake Remediation
This work addresses the problem of assessing pedagogical quality in AI tutors for math education, representing an incremental improvement in reward modeling for specific educational tasks.
The paper tackles the challenge of evaluating AI tutors in math mistake remediation by developing a reward model based on pedagogical aspects derived from human preferences, achieving 0.74 pairwise accuracy on a human preference test with a 0.5B-parameter backbone.
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.