JavaTutor: An Integrated Large Language Model Tutoring Platform for Learning Java

Publication Date

1-1-2026

Document Type

Conference Proceeding

Publication Title

International Conference on Artificial Intelligence Computer Data Sciences and Applications Acdsa 2026

DOI

10.1109/ACDSA67686.2026.11468147

Abstract

Large Language Models (LLMs) are rapidly transforming how students and developers approach programming, offering new ways to learn, solve problems, and automate coding tasks. In this study, we present JavaTutor, an LLM-powered tutoring platform designed to help students practice and master programming in Java. Unlike general-purpose coding assistants, JavaTutor emphasizes structured problem-solving, with a threestep methodology that guides learners through understanding the problem, generating solutions, and refining code. To evaluate its effectiveness, we experimented with a diverse set of models: Gemini-2.0 Flash, Llama-3.2-11B-Vision, DeepSeek-R1, DeepSeek-V3, and Mistral 7B and conducted a benchmarking comparison with GitHub Copilot. Our results show that JavaTutor not only provides syntactically correct solutions more consistently but also fosters a deeper learning process for students. Overall, the benchmarking revealed that while GitHub Copilot often delivered quick solutions, it lacked structured pedagogical guidance, leading to shallow engagement with problem-solving. Conventional systems also struggled with exposing students to debugging strategies, causing gaps in conceptual understanding. In contrast, JavaTutor's three-step debug process: problem comprehension, solution generation, and iterative refinement, not only improved correctness rates but also enhanced students' ability to internalize problem-solving strategies, making it a more effective platform for long-term mastery of Java programming.

Keywords

Generative AI, LLMs for Education

Department

Computer Science

Link to Full Text

Share

COinS