Grading has long been one of the most demanding and challenging tasks for educators. Even with pre-defined rubrics, issues regarding grading validity and consistency among different graders persist. For over six decades, automated essay scoring (AES) systems have been developed to reduce manual workload and increase efficiency. However, most of these existing AES systems focus on relatively simple, closed-ended questions with known preferred answers, which is far less challenging than assessing a scientific thesis.

With the emergence of Large Language Models (LLMs), could these models present a new opportunity for automating thesis evaluation? This study explores that possibility through an experiment: we systematically manipulated existing theses to represent different quality levels, and used three mainstream LLMs to grade them according to a standardized rubric. The results indicate that all three LLMs reliably recognized thesis quality and assigned scores as hypothesized. These results support the findings of previous research and highlight a positive future for integrating AI into grading tasks in higher education.