Abstract:

Personalizing educational content at scale is still a major challenge for e-learning platforms. Deep learning has been shown to automate feature identification and prediction of learner performance in e-learning systems; however, such discriminative models are not designed to generate explanations, hints, or conversational feedback. This paper presents a conceptual architecture, LLM-Smart-E-Learning, that extends this line of work by combining deep knowledge tracing with retrieval-augmented generation (RAG) and a multi-agent Socratic tutoring loop in which prompts are aligned with Bloom's taxonomy. We describe the proposed system architecture, compare it with classical deep-learning approaches, and outline the technical and moral challenges that ought to be addressed before practical evaluation. The work is presented as a research-in-progress proposal intended to solicit feedback before implementation and pilot testing.