AI-Powered Grading for Higher Education

Abstract

The rapid rise of Artificial Intelligence (AI) is reshaping many areas of society, and education is no exception. In particular, the growing use of Large Language Models (LLMs) is beginning to transform how teaching and assessment are approached. One area where this shift is especially promising is in the automated grading of student writing, particularly open-ended or essay-style responses [1][2][3]. Traditional grading of such responses is not only time-consuming but also prone to subjectivity and inconsistency among different graders. These challenges are especially pronounced in higher education, where instructors often face large volumes of student work. By leveraging the natural language understanding and generation capabilities of LLMs, educators can achieve faster, more consistent, and scalable assessments, while also improving the quality and timeliness of feedback provided to students. In addition to supporting instructors, this approach can also serve as a powerful learning aid for students. It offers an effective means of exam preparation through the creation of practice tests and self-assessments evaluated by AI, allowing learners to receive immediate, personalized feedback and better understand their progress. In this work, we present a platform that automates the grading of textual student responses using OpenAI’s LLMs. The system is implemented as a web application with a React frontend and a Django backend. Student submissions are processed in the backend, where they are combined with instructor-provided model answers or lecture materials and sent as structured prompts to the LLM. The output consists of both a grade and formative feedback, which are stored in the database and made accessible through the instructor-facing interface for review and adjustment. The modular architecture also supports extensions such as additional LLM providers, multilingual grading, and explainability features. The platform supports two complementary grading approaches: referencebased grading, in which student answers are evaluated against instructor-provided model responses; and generative grading, where the AI generates reference answers based on teaching materials such as lecture notes or textbooks, and compares student responses to these automatically created references. To evaluate the system, we will use authentic student work collected from five different exam sessions across three undergraduate courses at the Faculty of Mathematics (two first-year courses and one fourth-year course) and two master’s-level courses at the University of Arts. Student responses will be graded by both instructors and the platform, and the results compared to assess grading consistency and the quality of feedback. While a large-scale comparative study is planned as a future step, this evaluation aims to demonstrate the potential of the platform to provide instructors with a tool for creating, distributing, and evaluating assessments in an entirely online environment, streamlining the grading process while enabling high-quality AI-generated feedback. To encourage wider use and collaboration, the platform will be made available as open-source software. Looking ahead, we plan to improve the tool by supporting more languages and adding features that explain how the AI comes to its decisions, so everyone can better understand and trust the results. Overall, this project shows how LLMs can play a practical role in modernizing educational assessment, providing scalable and efficient tools that meet the changing needs of both teachers and students.