AI Tutoring Benchmarks
What each one measures, who it’s for, and how to use it, updated as the field evolves
As AI tutors get better at explaining, the harder question is whether they actually help students learn. A growing number of researchers, companies and nonprofits are building benchmarks to answer that, from tests of whether a tutor gives hints instead of answers to studies that measure learning gains with real students.
This page tracks the open benchmarks we find most useful for founders building AI tutors and learning tools: what each one measures, the grade levels and subjects it covers, and how you can access it. We’ll keep it updated as new ones appear.
For founders building AI tutors: check out our take on what these benchmarks capture, what they miss, and the five questions you should be ready to answer in this article: AI Tutors Can Teach. Who Will Learn?
Know of a benchmark we missed, or building one yourself? Reach out to tony[at]reachcapital[dot]com.
Last updated: November 2026
| Benchmark | Creator | What it measures | Level | Published as |
|---|---|---|---|---|
| TutorBench | Scale AI | Teaching moves: explaining, feedback, hints without giving away answers | HS / AP STEM | Leaderboard |
| MathTutorBench | ETH Zurich and partners | Correctness and teaching moves | K-12 math | Code, data |
| MRBench + BEA 2025 | MBZUAI and partners | Spotting and fixing student mistakes | K-12 math | Data |
| DrawEduMath | Allen Institute for AI, Worcester Polytechnic Institute and partners | Understanding students’ handwritten and hand-drawn math work | K-12 math | Code, data |
| GradeBench | Pensive | Grading handwritten student work against instructors’ rubrics | University STEM | Results chart |
| MMTutorBench | Academic | Teaching moves with images and diagrams | K-12 math | Paper |
| TutorMoments | Allen Institute for AI and partners | When to scaffold, push for rigor or step back, using real tutoring sessions | Grades 2–7 math | Code, data (preview) |
| CSTutorBench | UIUC | Teaching quality of small models in block-based programming | Grades 6–8 coding | Rubric, code |
| Pedagogy Benchmark | Fab Inc | Knowledge of teaching and learning science | Teacher level | Gated data |
| EduBench | Beijing Institute of Technology | Teaching knowledge and moves | Mixed | Data |
| ELBench | Academic | Teaching knowledge and soundness, across nine models | Mixed | Paper |
| SafeTutors | Academic | Pedagogical soundness: over-disclosure, reinforcing misconceptions, dropping scaffolding over long conversations | K-12 STEM | Paper, data |
| EduZone | Academic | Resistance to misuse, including disguised cheating requests | K-12 | Paper |
| Safe-Child-LLM | Academic | Age-appropriate responses | Ages 7–17 | Code, data |
| TutorCaseBench | Academic | Supporting students with learning and behavioral challenges | Special populations | Paper (under review) |
| EduClaw-Bench | Academic | Sustained tutoring over 30 days with simulated learners | Simulated learners | Paper |
| StudentBench | Handshake | Learning gains with real learners, compared with human tutors | Adults (GRE prep) | Code, data |
| Learning Outcomes Measurement Suite | OpenAI, University of Tartu, Stanford SCALE | Long-term learning outcomes from real-world AI use | K-12 and beyond | Forthcoming |