🤖 AI Summary
To address persistent hallucinations, poor curriculum alignment, and limited customizability of large language models (LLMs) in educational applications, this paper proposes a knowledge graph-enhanced retrieval-augmented generation (RAG) framework for building a generalizable AI tutoring system adaptable to arbitrary curricula. Our key innovation lies in structurally modeling course concepts as a dynamic knowledge graph and dynamically injecting graph-based semantic constraints during response generation, thereby significantly improving answer accuracy and controllability. Leveraging the Qwen2.5 foundation model, we integrate knowledge graph embeddings, fine-grained retrieval, domain-specific prompt engineering, and lightweight supervised fine-tuning. A user study (n=50) demonstrates statistically significant improvements (p<0.01) over baseline methods across answer relevance, usability, and pedagogical satisfaction. Results validate the framework’s effectiveness and feasibility for personalized, high-fidelity intelligent tutoring.
📝 Abstract
This paper introduces a novel framework for adaptable AI tutors using Knowledge Graph-enhanced Retrieval-Augmented Generation (KG-RAG). This approach addresses the critical challenges of information hallucination and limited course-specific adaptation prevalent in Large Language Model (LLM)-based tutoring systems. By integrating Knowledge Graphs (KGs) with RAG, we provide a structured representation of course concepts and their interrelationships, grounding the AI tutor's responses in relevant, validated material. We leverage Qwen2.5, a powerful and cost-effective LLM, within our KG-RAG framework. A user study (n=50) demonstrated positive student feedback regarding answer relevance, ease of use, and overall satisfaction. This KG-RAG framework offers a promising pathway towards personalized learning experiences and broader access to high-quality education.