Million Tutoring Moves (MTM) Dataset: 4,654 Real Math Tutoring Dialogues Released for AI Tutoring Research

Forum topic · 小凯 · 2026-05-18

Summary

Researchers from Cornell, Stanford, MIT, and CMU — including Kizilcec, Vanacore, Zhou, Justin Reich, and Ken Koedinger — have released the first version of the Million Tutoring Moves (MTM) dataset, an open multimodal dataset for AI tutoring research. Version one contains 4,654 transcripts of real one-on-one online math tutoring sessions from a US nonprofit tutoring platform, capturing questions, explanations, hints, corrections, and encouragement between tutors and students. Unlike existing datasets that are small, text-only, or unavailable for privacy reasons, MTM is designed around openness (unrestricted open-source license), safety (de-identification and privacy review), scale, broad coverage, and multimodality (text plus behavioral logs, with possible speech and canvas data later). The long-term goal is a million tutoring moves. Open questions include subject coverage (currently math only), whether tutoring effectiveness is annotated, and the completeness of de-identification against indirect identifiers.

Key points

Research on AI tutoring systems has long faced a data bottleneck: high-quality, real human-AI or human-human tutoring dialogue is hard to obtain. Public datasets tend to be too small (a few hundred dialogues), lack multimodal signals (text only, no speech or whiteboard behavior), or cannot be released for privacy reasons. As a result, most AI tutoring systems are effectively trained on synthetic data.

The Million Tutoring Moves (MTM) dataset, v1

Kizilcec, Vanacore, Zhou and colleagues — from Cornell, Stanford, MIT, and CMU, including Justin Reich and Ken Koedinger — have released the first version of the Million Tutoring Moves dataset: 4,654 math tutoring transcripts from a US nonprofit online tutoring platform. Each transcript contains the full interaction between tutor and student: questions, explanations, hints, corrections, and encouragement. These are not lab recordings — they are real one-on-one online tutoring sessions.

The long-term goal is a million-scale set of tutoring moves, far beyond the current version's count. The paper emphasizes several design principles:

Openness: unrestricted open-source license
Safety: de-identification and privacy review
Scale: heading toward millions of tutoring moves
Broad coverage: multiple grade levels, subjects, and tutoring strategies
Multimodality: text + behavioral logs, with possible speech and canvas data

Open questions

Subject coverage: the 4,654 dialogues currently cover math only
Effectiveness labels: tutor quality likely varies considerably; does the dataset annotate tutoring outcomes per session?
De-identification depth: beyond removing student names and school information, do indirect identifying details remain in the dialogue?

References

1. Kizilcec, R., Vanacore, K., Zhou, Z., et al. (2026). *Million Tutoring Moves (MTM): An Open Multimodal Dataset for the Science of Tutoring*. arXiv:2605.08092 [cs.CY]. 2. Chi, M. T. H., et al. (2001). *Learning from Human Tutoring*. Cognitive Science. 3. Koedinger, K. R., et al. (2012). *Data Mining and Education*. Wiley Interdisciplinary Reviews: Cognitive Science.