Prompt Engineering for Automated Essay Scoring in Higher Education: A Case Study in Academic Reading

Abstract

This study proposes a prompt engineering framework leveraging GPT-4o to automate the evaluation of academic reading reports. Integrating Role-Setting, FewShot Calibration, and Structured JSON constraints, our method achieved a strong correlation with human grading (r = 0.83, p<0.001) and reduced Mean Absolute Error by 52.2%. The framework successfully quantified a significant “learning gain” (+5.80 points) between pre- and post-revision drafts, offering a scalable solution for process-oriented mentorship in collaborative learning.