The research team at UNIST has developed VEHME, a cutting-edge Vision-Language Model designed for automated assessment of handwritten mathematics expressions. This AI model addresses the challenges posed by various handwriting styles and unstructured answer formats. VEHME employs a two-phase training process, combining supervised fine-tuning with reinforcement learning to align outputs with grading criteria, focusing on correctness and error localization. Its innovative Expression-Aware Visual Prompting Module enhances spatial comprehension, facilitating accurate evaluation. Tested on AIHub and FERMAT datasets, VEHME demonstrates state-of-the-art performance, rivaling proprietary models like GPT-4 while operating on just 7 billion parameters. This efficiency makes VEHME a valuable resource for educational institutions, capable of providing detailed feedback on complex, unstructured answers. The open-source model aims to transform math assessment in classrooms, offering potential applications beyond education, such as document processing. The research has been accepted for presentation at EMNLP 2025, showcasing advancements in AI for education.
Source link
