Korean researchers have developed artificial intelligence (AI) that reads free-response math answers written in messy handwriting like a person would, grades them, and even marks them up by identifying the reasons for wrong answers.
A team led by Kim Tae-hwan, a professor at the UNIST Graduate School of Artificial Intelligence, said on the 17th it co-developed with a team led by professor Go Sung-an at Pohang University of Science and Technology POSTECH an AI model, "VEHME," that grades handwritten math solutions. The results were accepted as a full paper at the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP).
In education settings, grading free-response math problems is considered a representative time-consuming task. Because solutions include not only formulas but also graphs and figures, and because each student's handwriting and answer layout differ, existing AI grading methods had clear limits.
To solve these issues, the researchers applied their self-developed "Equation-aware Visual Prompting Module (EVPM)" and a "dual learning technique" to VEHME to boost grading performance. EVPM is designed to keep the model from missing the expansion order of complex, connected formulas by following the flow of the solution and jointly capturing context and formula positions. On top of that, the dual learning technique enables the model not only to judge correctness but also to explain which point in the solution is wrong and why.
For training materials such as handwriting and markup data, the team fed in synthetic data generated using Alibaba's language model. The researchers said this taught VEHME to read the context of a solution like a human grader and pinpoint error locations specifically.
In performance validation across a range of handwritten answers from elementary arithmetic to calculus, the model, despite using a relatively lightweight 7-billion-parameter configuration, showed accuracy comparable to large models such as "GPT-4o" and "Gemini 2.0 Flash." In particular, for difficult answers with misaligned lines or messy handwriting, there were cases where it more precisely identified error locations.
VEHME is open-sourced and can be used free of charge by educational institutions such as schools and private academies. Kim Tae-hwan said, "We have secured enough stability and efficiency for use in real education settings," adding, "The EVPM module excels at automatically structuring complex visual information and can be extended to various industries, including document recognition, design drawing analysis, and digitization of handwritten records."