Towards Fully Automated Exam Grading: Fairness-Aware Recognition of Handwritten Answers with Foundation Models

The study presents a method for fully automated grading of handwritten exams using vision-language foundation models (VLMs), achieving a recognition accuracy of 98.4% on a benchmark of 61 anonymized exams, significantly improving upon previous methods which only reached 88-91%. The approach emphasizes fairness by distinguishing between false negatives and false positives, with a lightweight prompt reducing the false-negative rate to 0.58%. This advancement offers a scalable solution for automated grading while maintaining assessment integrity, making it relevant for practitioners seeking to implement fair and accurate grading systems in educational settings.