Abstract:

Large language models (LLMs) can generate natural explanations, but generic LLMs do not consistently retain learner-specific accessibility constraints across user requests. This paper describes a behavior-conditioned Retrieval-Augmented Generation (RAG) system that retrieves learner-profile constraints and conditions educational generation for neurodiverse learners, primarily Autism Spectrum Disorder (ASD) and dyslexia, and secondarily dysgraphia-related motor-writing issues. The proposed system retrieves how to structure an explanation - literalness, line length, spacing, scaffolding, keyword highlighting, and response burden - rather than just what factual information to include, as is the case with content-oriented RAG. The prototype system features a Flask server, web-based learner interface, JSON knowledge base of learner profiles, optional Chroma vector store, OpenAI/Gemini/DeepSeek model orchestration, auto_eval_v4 scoring, dashboard logging, human feedback, and dashboards. In the locked 578-row official analysis snapshot, adaptive RAG responses scored a mean CompositeScore of 87.01, versus 57.84 for baseline responses. A paired analysis of 262 matched model runs yielded a mean RAG gain of 27.83 points. Human feedback included 23 rows with a mean rating of 4.78/5 and 100.0% helpful flags. Four expert review forms were completed and are regarded as formative feedback for educational appropriateness, not clinical validation. The reviewers broadly endorsed the educational suitability of the Adaptive RAG outputs versus the baseline outputs for clarity, reduced ambiguity, suitability for the target audience, and reduced cognitive load. After expert feedback, the prototype was improved with additional ASD and dyslexia strategy families and a Dyslexia Guided Reading pilot interface. Results show prototype-level gains in automated accessibility-oriented measures and positive user feedback. The system is not a diagnostic, therapeutic or clinical decision-making tool. Keywords: Retrieval-Augmented Generation; neurodiversity; dyslexia; autism; adaptive learning; large language models; mixed-method evaluation; educational accessibility; information systems; Arabic educational technology 1. Introduction Generative AI is gaining traction in education as it can generate explanations, examples, summaries, quizzes and scaffolds. But generating natural language is not enough to make a system accessible. Neurodiverse learners may need accommodation in the form of the answer: shorter lines, steps, literal language, less ambiguity, clear separation, or less writing. A typical answer can be semantically correct but inaccessible due to the way it is presented. Learners with ASD may need social or figurative language translated into literal meaning. They may need step structures, cause-effect relationships and no assumptions. Students with dyslexia may need text to be visually stable, short-lined, spaced, and anchored with bolded keywords. These requirements are not the same as lowering standards. The same information can be technically correct, but presented in a more accessible format. Generic LLM interaction does not address this issue. The learner or teacher can prompt the model to be simpler, use bullet points, or avoid idioms, but this requires manual prompting. Manual prompting is not consistent, is user-burdening, and can be lost between sessions. Fine-tuning is also not the best approach for individual accommodations because accommodations need to be transparent, editable, and adaptable to changing learner needs. This paper therefore considers a profile-conditioned RAG system where learner constraints are stored separately and accessed at runtime. Existing research on inclusive educational AI demonstrates the potential of AI for accessibility and adaptive learning, but also stresses the importance of transparency, privacy and human oversight (Pagliara et al., 2024; UNESCO, 2023). Universal Design for Learning focuses on accessibility and multiple representation (CAST, 2024). Thus, the current work proposes an educational assistive prototype rather than a diagnostic, therapeutic or clinical system. The main innovation is the retrieval of behavior. Traditional RAG retrieves evidence or passages to support an answer (Lewis et al., 2020). In the new approach, the system retrieves a learner profile with pedagogical constraints. The model retrieves the answer: line length, literalness, type of scaffolding, number of steps, emphasis on keywords, and language. This positions RAG as an information systems approach to accessibility. The research is also driven by the practical differences of current LLM providers. Models may vary in reasoning, verbosity, multimodal capabilities, speed, and adherence to formatting guidelines. As such, the prototype does not presume a single provider is superior. It runs OpenAI, Gemini and DeepSeek in baseline and RAG modes, evaluates the responses, logs the results and returns the best-scoring personalized response.