Adaptive Neuro-Symbolic Learning Companions for Explainable Cognitive Scaffolding in Educational AI

Abstract The increasing adoption of artificial intelligence (AI) in education has accelerated the development of intelligent tutoring systems capable of providing adaptive instructional support. However, many existing systems remain constrained by limited personalization, weak reasoning transparency, and insufficient integration of cognitive-developmental learning principles. Most approaches rely on either neural conversational intelligence or symbolic pedagogical reasoning independently, restricting their ability to provide explainable and developmentally aligned educational support. This study proposes and empirically evaluates an adaptive neuro-symbolic learning companion framework designed to support explainable cognitive scaffolding through integrated learner modeling and pedagogical reasoning. Using the Khan Academy Tutoring Accuracy Dataset, a neuro-symbolic analytical framework was developed that combines transformer-based conversational analysis, symbolic tutoring inference, emotional state detection, scaffolding evaluation, and explainability assessment. The framework operationalizes scaffolding strategies grounded in Vygotsky’s Zone of Proximal Development (ZPD), including adaptive hint escalation, conceptual reinforcement, reflective prompting, and meta-cognitive support. Statistical analyses, including independent samples t-tests and ANOVA, were conducted to examine relationships between scaffolding strategies and learning effectiveness. Results indicate that existing tutoring interactions are predominantly procedural and direct-answer oriented, with only six explicit meta-cognitive prompts identified across 2,022 tutoring interactions. Although conversational coherence remained high (BERTScore F1 = 0.815), reflective scaffolding was largely absent. The proposed framework successfully identified interpretable pedagogical behaviors and demonstrated measurable explainability through symbolic reasoning structures. The study contributes a novel neuro-symbolic approach to adaptive learner modeling, explainable educational AI, and cognitively aligned tutoring systems, advancing the development of transparent and personalized AI learning companions.