Date
Publisher
arXiv
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 confusion 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 LLM as an evaluator 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.
Study at a glance
- Question: How well do leading large language models demonstrate pedagogical skills when remediating student math mistakes?
- Setting and data: 192 elementary and middle school math dialogues comprising 1,596 responses from human tutors and seven large language models.
- Method: Model and benchmark evaluation comparing human expert annotations against LLM evaluator judgments across eight pedagogical dimensions.
Why it matters
Evaluating AI tutors requires measuring complex pedagogical abilities rather than generic text quality, but existing benchmarks use fragmented, subjective, or incomplete criteria that prevent standardized comparisons.
Key findings
- All evaluated models lack key pedagogical abilities needed for tutoring, with GPT-4 revealing final answers roughly 47 percent of the time instead of guiding students.
- Llama-3.1-405B achieved the strongest overall balance across pedagogical dimensions, while Phi3 performed worst due to low coherence and weak contextual understanding.
- Using models such as Prometheus2 and Llama-3.1-8B as automated evaluators proved unreliable, yielding predominantly negative correlations with human judgments across pedagogical dimensions.
Summary drafted with AI from the full paper and may contain errors. See the full paper for details.
What is the application?
Who is the user?
Who age?
Why use AI?
Study design
