Repository of DUTH

An interactive Python Tutor based on Large Language Models (Master thesis)

Χρυσόπουλος, Μάριος/ Chrysopoulos, Marios/ Προηγμένες Τεχνολογίες Πληροφορικής και Υπολογιστών

The integration of Large Language Models (LLMs) into educational practice creates new opportunities for supporting the teaching of programming, particularly at the high school level where students first encounter fundamental algorithmic concepts. This thesis presents and evaluates PythonTutor, a fully integrated intelligent tutoring system that provides theory, examples, exercises, and personalized AI-guided feedback for the introductory teaching of Python. The system leverages GPT 4.0 Mini to deliver step-by-step hints tailored to each student submission, aiming to promote independent problem-solving, conceptual understanding, and reflective learning without fostering overreliance on the AI. To assess the effectiveness of PythonTutor, a pilot study was conducted with 18 Vocational High School students and three teachers. Students were divided into two groups: a treatment group using the platform and a control group receiving traditional instruction. Results revealed significantly higher normalized learning gains for students who used the application. Moreover, evaluation of the underlying AI classification model demonstrated high accuracy, sensitivity, and consistency in identifying errors and generating appropriate hints—an essential factor for building pedagogically reliable AI-driven learning environments. Student and teacher feedback highlighted the platform’s usability, educational value, and practical relevance in real classroom settings. Students reported that AI-generated hints substantially supported their problem-solving process, while teachers observed improvements in student engagement and performance. Additionally, PythonTutor offers oversight tools enabling teachers to review detailed student–AI interaction logs, providing insight into learners’ difficulties, strategies, and progression. The methodology of this work includes platform development, integration of the LLM, prompt engineering for instructional content generation, analysis of student–AI interaction data, and quantitative assessment of learning outcomes. Findings indicate that even a lightweight model such as GPT 4.0 Mini can serve as a dependable and cost-effective component within an intelligent tutoring system. Overall, PythonTutor appears to be a scalable, pedagogically aligned, and resource-efficient solution for personalized programming instruction in secondary education. The results support the potential of LLM-based tutoring systems to enhance learning, democratize access to high-quality feedback, and inform future developments such as support for additional programming languages, adoption of open-source models, and advanced learning analytics.