Scholarlink

<p>Educational AI systems increasingly rely on generative models for personalization, yet most remain blind to learners’ emotional and cognitive states during instruction. This limitation particularly affects neurodivergent learners, including those with autism, whose performance depends heavily on emotional stability and appropriately calibrated instructional support. Current systems optimize for semantic accuracy but fail to adapt content complexity or pacing to individual readiness. To address this gap, we introduce DYNAMIC-RAG, a retrieval-augmented generation framework that integrates real-time learner-state modeling into adaptive tutoring. The system combines behavioral engagement, emotional valence-arousal, and clinical indicators into a unified Pedagogical Readiness Score (PRS) that drives coordinated adaptation across retrieval and generation. The PRS governs three adaptive mechanisms: retrieval depth (3–7 documents), content complexity (conceptual density), and generation parameters (response determinism vs. flexibility). Low-readiness learners receive scaffolded, simplified explanations with increased retrieval support; high-readiness learners access concise, abstract content promoting independent reasoning. Proof-of-concept evaluation with elementary mathematics (grades 3–5, n=40 queries) demonstrated meaningful improvements: exact match accuracy increased 17.5 percentage points (47.5% to 65.0%) compared to baseline LLM, with adaptive scaffolding benefiting low-readiness learners most substantially (+28.6 pp vs. +9.1 pp for high-readiness learners). Hallucination rates decreased 68% while maintaining real-time responsiveness (sub-2&#xa0;s latency). These exploratory findings suggest that jointly adapting retrieval and generation based on learner state provides complementary benefits for factual accuracy and pedagogical appropriateness, though confirmatory validation across broader domains and populations remains necessary before large-scale deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic-rag: a multimodal emotion-aware prompting and retrieval-augmented LLM framework for autism support in adaptive learning systems

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Nisrine El Ayat,
  • Mohammed Boutalline,
  • Adil Tannouche,
  • Chaimae Ouazri,
  • Mohammed Yahyaoui

摘要

Educational AI systems increasingly rely on generative models for personalization, yet most remain blind to learners’ emotional and cognitive states during instruction. This limitation particularly affects neurodivergent learners, including those with autism, whose performance depends heavily on emotional stability and appropriately calibrated instructional support. Current systems optimize for semantic accuracy but fail to adapt content complexity or pacing to individual readiness. To address this gap, we introduce DYNAMIC-RAG, a retrieval-augmented generation framework that integrates real-time learner-state modeling into adaptive tutoring. The system combines behavioral engagement, emotional valence-arousal, and clinical indicators into a unified Pedagogical Readiness Score (PRS) that drives coordinated adaptation across retrieval and generation. The PRS governs three adaptive mechanisms: retrieval depth (3–7 documents), content complexity (conceptual density), and generation parameters (response determinism vs. flexibility). Low-readiness learners receive scaffolded, simplified explanations with increased retrieval support; high-readiness learners access concise, abstract content promoting independent reasoning. Proof-of-concept evaluation with elementary mathematics (grades 3–5, n=40 queries) demonstrated meaningful improvements: exact match accuracy increased 17.5 percentage points (47.5% to 65.0%) compared to baseline LLM, with adaptive scaffolding benefiting low-readiness learners most substantially (+28.6 pp vs. +9.1 pp for high-readiness learners). Hallucination rates decreased 68% while maintaining real-time responsiveness (sub-2 s latency). These exploratory findings suggest that jointly adapting retrieval and generation based on learner state provides complementary benefits for factual accuracy and pedagogical appropriateness, though confirmatory validation across broader domains and populations remains necessary before large-scale deployment.