Personalized Learning via Generative AI: A Profile-Aware Adaptive RAG Framework for Neurodiverse Student Support

While large language models can produce coherent explanations of educational concepts, they are not necessarily consistent in maintaining learner-specific accessibility requirements when asked repeatedly. This can pose a problem for neurodiverse learners who might need literal language, spacing, structured steps, anchors for key words or a lessening of the writing load. Current educational AI systems and Retrieval-Augmented Generation systems typically aim at providing factual support or general tutoring, with less consideration of retrieving profile-based rules to guide the presentation of an explanation.
This paper presents a Retrieval-Augmented Generation framework for personalized learning support based on behavior. The system fetches the learner-profile constraints for Autism Spectrum Disorder, dyslexia, and motor-writing support for dysgraphia and conditions educational responses based on these constraints. The prototype features a Flask backend, a web interface, a JSON learner-profile knowledge base, optional vector retrieval, orchestration with OpenAI/Gemini/DeepSeek, auto_eval_v4 scoring, dashboard logging, and human feedback collection.
The prototype design was used in the evaluation by a combination of mixed methods. Adaptive RAG responses had a mean CompositeScore of 87.01, while baseline responses had a mean CompositeScore of 57.84, in the locked 578-row analysis snapshot. The mean RAG gain for the 262 matched model runs was 27.83 points. Human feedback had 23 rows with a mean rating of 4.78/5 and 100.0% helpful flags. Formative educational-suitability feedback was provided by four specialist reviews. Results show prototype level output alignment for accessibility, not clinical diagnosis or therapeutic validation.