ScratchMath

Can MLLMs Read Students' Minds? Unpacking Multimodal Error Analysis in Handwritten Math

AIED 2026 — 27th International Conference on Artificial Intelligence in Education

---

Overview

ScratchMath is a multimodal benchmark for evaluating whether MLLMs can analyze handwritten mathematical scratchwork produced by real students. Unlike existing math benchmarks that focus on problem-solving accuracy, ScratchMath targets error diagnosis — identifying what type of mistake a student made and explaining why.

• 1,720 authentic student scratchwork samples from Chinese primary & middle schools
• 7 expert-defined error categories with detailed explanations
• 2 complementary tasks: Error Cause Explanation (ECE) & Error Cause Classification (ECC)
• 16 leading MLLMs benchmarked; best model reaches 57.2% vs. human experts at 83.9%

---

Dataset Structure

Subsets

Subset	Grade Level	Samples
primary	Grades 1–6	1,479
middle	Grades 7–9	241

Error Categories

Category (zh)	Category (en)	Primary	Middle
计算错误	Calculation Error	453	113
题目理解错误	Problem Comprehension Error	499	20
知识点错误	Conceptual Knowledge Error	174	45
答题技巧错误	Procedural Error	118	17
手写誊抄错误	Transcription Error	95	29
逻辑推理错误	Logical Reasoning Error	73	2
注意力与细节错误	Attention & Detail Error	67	15

Fields

Field	Type	Description
question_id	string	Unique identifier
question	string	Math problem text (may contain LaTeX)
answer	string	Correct answer
solution	string	Step-by-step reference solution
student_answer	string	Student's incorrect answer
student_scratchwork	image	Photo of handwritten work
error_category	ClassLabel	One of 7 error types
error_explanation	string	Expert explanation of the error

---

Quick Start

from datasets import load_dataset

# Load primary school subset
ds_primary = load_dataset("songdj/ScratchMath", "primary")

# Load middle school subset
ds_middle = load_dataset("songdj/ScratchMath", "middle")

# Access a sample
sample = ds_primary["train"][0]
print(sample["question"])
print(sample["error_category"])
sample["student_scratchwork"].show()

---

Citation

If you use this dataset, please cite:

@inproceedings{song2026scratchmath,
  title     = {Can MLLMs Read Students' Minds? Unpacking Multimodal Error Analysis in Handwritten Math},
  author    = {Song, Dingjie and Xu, Tianlong and Zhang, Yi-Fan and Li, Hang and Yan, Zhiling and Fan, Xing and Li, Haoyang and Sun, Lichao and Wen, Qingsong},
  booktitle = {Proceedings of the 27th International Conference on Artificial Intelligence in Education (AIED)},
  year      = {2026}
}

---

License

This dataset is released under the CC BY-NC-SA 4.0 license.