An artificial intelligence (AI) teacher has been developed that can grade even messy, handwritten math answers like a human and provide feedback.

Ulsan National Institute of Science and Technology (UNIST) announced on the 17th that Professor Kim Tae-hwan of the Graduate School of Artificial Intelligence and Professor Ko Seong-an’s team from the Department of Computer Science and Engineering at Pohang University of Science and Technology (POSTECH) have developed an AI model called “Bemi” that grades complex, handwritten math answers.

Descriptive math answers are unstructured data in which handwriting styles and answer layouts differ from person to person and formulas, graphs, and figures are mixed together, making it difficult for AI to recognize them accurately and grade them. By contrast, Bemi reads the positions and context of formulas precisely, as if following the flow of a human problem-solving process. When the research team used Bemi to grade solutions to a wide range of math problems, from elementary-level arithmetic to calculus, it demonstrated accuracy comparable to OpenAI’s “GPT-4o” and Google’s “Gemini 2 Flash.”

Behind Bemi’s performance is the “Expression Visual Prompting for Math (EVPM)” technology developed by the team. EVPM is a training method in which Bemi draws virtual boxes around complexly arranged formulas to ensure it does not miss the order of the solution steps. The researchers made Bemi open source so that educational institutions such as schools and private academies can use it free of charge. Professor Kim explained, “Grading handwritten math is one of the toughest challenges in edtech AI,” adding, “It is meaningful that Bemi has secured a level of stability and efficiency that makes it usable in real educational settings.”

Popular News