The adoption of artificial intelligence in education has shown substantial potential for enhancing learning efficiency. In particular, large language models (LLMs) such as ChatGPT and Gemini are expected to significantly improve educational tasks form...
The adoption of artificial intelligence in education has shown substantial potential for enhancing learning efficiency. In particular, large language models (LLMs) such as ChatGPT and Gemini are expected to significantly improve educational tasks formulated in natural language, including the provision of personalized feedback to students. However, these models still face limitations such as hallucinations, inaccuracies in training data, and a lack of transparency in their reasoning processes. To address these issues, this study proposes TutorAgent, a personalized feedback generation system that integrates an enhanced BKTransformer model with an Agentic GraphRAG framework. TutorAgent consists of two major components: analysis of students’ learning histories using a Knowledge Tracing (KT) model, and LLM’s geneartion of feedbacks based on the analysis of the KT model. For the KT component, we employ BKTransformer with RoPE, which incorporates Rotary Position Embeddings(RoPE) into the BKTransformer proposed in previous research. Unlike conventional KT models that merely predict student response correctness, BKTransformer with RoPE explicitly estimates the four interpretable parameters of Bayesian Knowledge Tracing (BKT), thereby enabling a more transparent interpretation of learners’ knowledge states. The natural language feedback generation is handled by two LLM agents. The diagnostic agent interprets the BKT parameters produced by BKTransformer with RoPE and generates a structured diagnosis of a student’s changing knowledge states. The feedback agent then performs GraphRAG on a given knowledge graph using this diagnosis and produces personalized feedbacks. We evaluate both key components—the KT model and LLM agents. Using publicly available datasets, including a large-scale learning history dataset from Icecream Edu on AI-Hub, BKTransformer with RoPE demonstrates an approximate 1% AUC improvement over the original BKTransformer. To assess the quality of generated feedback, we compare the Agentic GraphRAG approach—which distributes tasks across two LLMs—with a single-LLM approach that simultaneously performs the retrieval and feedback generation from the knowledge graph. The results show that Agentic GraphRAG achieves substantially higher performance in both answer faithfulness and context relevance, indicating more accurate retrieval and feedback generation. The knowledge graph used in the evaluation also originates from AI-Hub’s Korean curriculum–based knowledge graph dataset. This study represents the first attempt to integrate a BKT-based deep learning model with GraphRAG using Korean educational data.