Unifying AI Tutor Evaluation: An Evaluation Taxonomy for Pedagogical Ability Assessment of LLM-Powered AI Tutors

TL;DR

This work addresses the lack of a unified, pedagogy-grounded framework for evaluating AI tutors and introduces MRBench, a benchmark of 192 dialogues (1,596 tutor responses) grounded in middle-school mathematics. It proposes an eight-dimension taxonomy derived from learning sciences to assess student-mistake remediation, and validates it with human annotations and LLM-based critiques. The study finds that current state-of-the-art LLM tutors are capable QA systems but often fall short as pedagogical tutors, and that LLM-based evaluation correlates poorly with human judgments in most dimensions. By releasing the taxonomy and MRBench, the authors provide a standardized resource to guide future development, RLHF fine-tuning, and cross-model comparisons in AI tutoring. The results highlight the need for stronger pedagogical alignment and more robust evaluator tools to reliably track progress in AI-tutor development.

Abstract

In this paper, we investigate whether current state-of-the-art large language models (LLMs) are effective as AI tutors and whether they demonstrate pedagogical abilities necessary for good AI tutoring in educational dialogues. Previous efforts towards evaluation have been limited to subjective protocols and benchmarks. To bridge this gap, we propose a unified evaluation taxonomy with eight pedagogical dimensions based on key learning sciences principles, which is designed to assess the pedagogical value of LLM-powered AI tutor responses grounded in student mistakes or confusions in the mathematical domain. We release MRBench - a new evaluation benchmark containing 192 conversations and 1,596 responses from seven state-of-the-art LLM-based and human tutors, providing gold annotations for eight pedagogical dimensions. We assess reliability of the popular Prometheus2 and Llama-3.1-8B LLMs as evaluators and analyze each tutor's pedagogical abilities, highlighting which LLMs are good tutors and which ones are more suitable as question-answering systems. We believe that the presented taxonomy, benchmark, and human-annotated labels will streamline the evaluation process and help track the progress in AI tutors' development.

Figures (6)

• Figure 1: An example of mistake identification from the validation pilot study.
• Figure 2: The prompt template used to generate responses from the seven considered LLMs for both Bridge and MathDial datasets. The template is adapted from wang2024bridging. The notable differences are: (1) The problems covered in the Bridge dataset are at the elementary school level, whereas those in MathDial are at the middle school level; and (2) The conversation topic is not provided in MathDial.
• Figure 3: An example from the guidelines document provided to annotators, showing a page that details definitions, annotation labels, and associated examples for the Providing Guidance dimension of the taxonomy.
• Figure 4: An example from the annotator training phase for the Mistake Identification dimension.
• Figure 5: An example from the annotator testing phase for the Revealing of the Answer dimension.