Towards Human-Centered Large Language Models for Personalized Programming Education

Abstract

Large Language Models (LLMs) are increasingly used in computing education to support programming practice and generate explanations [5, 9]. However, current LLM-based systems are not pedagogically aligned and fail to adapt to learners’ prior knowledge, often producing feedback that is too advanced or oversimplified [8]. Recent evidence also suggests that LLMs do not consistently outperform traditional machine learning approaches for fine-grained educational tasks, highlighting the need to better understand when and how LLMs should be used. This research aims to develop human-centered LLMs for programming education by leveraging student modeling data, including knowledge states and learning processes. We explore prompt engineering, fine-tuning, and reinforcement learning to improve pedagogical effectiveness and enable scalable personalized support.