Research on Intelligent Grading of Physics Problems Based on Large Language Models

Abstract

The automation of educational and instructional assessment plays a crucial role in enhancing the quality of teaching management. In physics education, calculation problems with intricate problem-solving ideas pose challenges to the intelligent grading of tests. This study explores the automatic grading of physics problems through a combination of large language models and prompt engineering. By comparing the performance of four prompt strategies (one-shot, few-shot, chain of thought, tree of thought) within two large model frameworks, namely ERNIEBot-4-turbo and GPT-4o. This study finds that the tree of thought prompt can better assess calculation problems with complex ideas (N = 100, ACC ≥ 0.9, kappa > 0.8) and reduce the performance gap between different models. This research provides valuable insights for the automation of assessments in physics education.