The rapid growth of generative artificial intelligence has redefined education, while college-level physics remains a persistent challenge for learners, intensified by their growing reliance on large language models (LLMs). This dependency often weakens independent problem-solving, while the complex definitions of physics require substantial instructional resources. To solve these issues, the authors present a smart course framework, physics intelligent teaching assistant (PITA), for personalized and adaptive learning in college physics education. PITA, an LLM-based platform designed to advance physics education by curbing excessive LLM reliance through adaptive learning, student-tailored dialogue and personalized resource recommendations.
PITA integrates retrieval-augmented generation (RAG) to enhance response quality by retrieving in verified knowledge bases and uses graph neural networks (GNN) for tracking learners’ conceptual mastery. Previous knowledge tracking work showed that students can be divided into two groups based on learning style: memorization or comprehension. Therefore, PITA adopts a Socratic dialogue tailored to individual learning styles, decomposing complex physics problems into manageable subproblems. To assess PITA, the authors contrast its performance against DeepSeek using responses to 30 diverse physics homework tasks, with accuracy as the core metric, supplemented by evaluations from experienced physics professors.
In an evaluation of 30 diverse physics concept comprehension question answering tasks, PITA achieved 96.67% accuracy, exceeding DeepSeek’s 90% (6.67% higher), with professional college physics instructors claiming that its responses outperformed manual explanations in depth and clarity. The improved graph-based knowledge tracing model went beyond traditional methods in accuracy and interpretability, promoting critical thinking and independent inquiry.
By synergizing RAG, GNN and the Socratic method, PITA provides a framework for learning physics and fosters profound conceptual understanding and self-directed inquiry among learners. The research findings have the potential to impact the design, implementation and evaluation of artificial intelligence-enabled teaching assistants in higher education, informing the development of innovative educational tools that can promote student learning outcomes, engagement and satisfaction.
