Tutor AI: Phương pháp hướng dẫn từng bước cá nhân hóa trong các hệ thống học tập thích ứng

Abstract

Context-aware adaptive feedback remains a significant challenge in education, particularly when guiding learners through complex problem-solving tasks across diverse disciplines. This study introduces a generalized stepwise tutoring framework that leverages large language models (LLMs) to support learning by decomposing tasks into micro-steps, assessing progress, and generating feedback tailored to the learner’s current state. The framework is implemented in the Tutor AI system and was empirically evaluated in a Programming Techniques course. In this environment, students interact with a real-time coding interface, where the learning path is clearly structured, and feedback is automatically generated via the Gemini LLM. Each student’s submission is dynamically integrated with their learning history to construct context-rich prompts, enabling the system to precisely identify the current learning step and provide adaptive, scaffolded guidance accordingly. Experimental results from a semester-long Programming Techniques (C language) course demonstrated the system’s effectiveness in significantly increasing engagement, fostering self-regulated learning, and reducing reliance on direct solutions. The architecture of Tutor AI is highly generalizable, allowing for expansion beyond programming to various other disciplines such as engineering, data analysis, social sciences, or any subject with clearly defined stepwise exercises. This research contributes to the development of explainable and scalable intelligent tutoring systems, laying a foundation for AI-supported, stepwise learning in modern educational contexts.