Korean researchers develop AI to grade handwritten math answers

An artificial intelligence teacher capable of grading handwritten math answers and offering detailed feedback has been developed, even for responses written unevenly or with unclear handwriting.

Ulsan National Institute of Science and Technology said on Dec. 17 that a research team led by Kim Tae-hwan, a professor at UNIST’s Graduate School of Artificial Intelligence, and Ko Sung-an, a professor of computer science at Pohang University of Science and Technology, has created an AI model called Bemi that can evaluate complex handwritten math solutions.

Answers to open-ended math questions pose a challenge for AI systems because handwriting styles and answer layouts vary widely, and formulas, graphs and figures are often combined as unstructured data. Bemi, however, accurately recognizes both the placement and contextual meaning of formulas, much like a human who follows the logical progression of a solution. When the research team tested Bemi by grading math solutions ranging from elementary-level arithmetic to calculus, the model showed accuracy comparable to OpenAI’s GPT-4o and Google’s Gemini 2 Flash.

Bemi’s performance is powered by a technology developed by the research team known as the Equation Visual Prompt Method, or EVPM. The training approach allows Bemi to place virtual boxes around densely arranged formulas, enabling it to follow the sequence of problem solving without losing context. The researchers have released Bemi as open-source software so it can be used free of charge by educational institutions, including schools and private academies.

“Grading handwritten math answers is one of the most difficult challenges in edtech AI,” Kim said. “Bemi is significant because it has secured a level of stability and efficiency that makes it suitable for use in real educational environments.”


최지원 기자 jwchoi@donga.com