UnrealMentor GPT: A System for Teaching Programming Based on a Large Language Model
Hongli Zhu, Jian Xin Xiang, Zhihui Yang · Computer Applications in Engineering Education · 2025
ABSTRACT This paper introduces UnrealMentor GPT, a multiagent debugging framework that combines advanced large language model (LLM) capabilities with a dynamically updated knowledge base. Systems incorporating this framework are used in programming courses for university computer‐related majors. This teaching system based on Generative Pre‐training (GPT) technology guides students through a hierarchical learning process using multiple specialized agents (syntax checking, algorithm analysis, optimization) and retrieval‐augmented generation (RAG). Experimental results based on the effectiveness of undergraduate courses show that students spend less time debugging code in the course, the accuracy of solutions is improved, and the overall learning efficiency is significantly enhanced. Subsequent surveys on teaching effectiveness also showed that students were satisfied with the learning process. Feedback from surveys of relevant teaching staff indicated that the system can simplify the error correction process and deepen students' understanding of concepts. However, there are some limitations to the current research, including the small sample size and short intervention time, which limits the application scenarios of the system. Future research will focus on expanding the participating groups, exploring cross‐language applicability, and conducting longitudinal experiments to verify the effectiveness of UnrealMentor GPT in various educational environments.