GenEdXR: A Generative AI-Driven Extended Reality Framework for Adaptive Educational Entertainment Ecosystems

Abstract Generative artificial intelligence and extended reality (XR) offer unprecedented opportunities for adaptive educational entertainment, yet existing systems suffer from static content pipelines and weak integration between immersive interaction and intelligent adaptation. Here we introduce GenEdXR, a framework that unifies a transformer-based deep knowledge tracing (DKT) module, a large-language-model content synthesis pipeline, a proximal-policy-optimisation adaptive orchestration engine, and multi-modal XR rendering within a closed-loop architecture. These components are bound through a scaled dot-product cross-modal attention fusion mechanism that integrates semantic, cognitive and temporal representations. We evaluated GenEdXR through a large-scale simulation protocol involving 33,510 replayed learner trajectories from three complementary public datasets --- ASSISTments 2009--2010, the Open University Learning Analytics Dataset, and KDD Cup 2010 Algebra 2005--2006 --- spanning question-level, session-level and step-level interaction granularities. Under five-fold cross-validation across four framework configurations, GenEdXR-Full attained the highest normalised learning gain (M = 0.523; Cohen's d = 2.71 versus baseline), highest engagement and flow-zone membership, and lowest cognitive-load proxy. Ablation and pairwise-comparison analyses demonstrated that neither the generative content channel nor the immersive rendering channel alone recovered the full performance profile, revealing a synergistic integration effect. Runtime characterisation confirms deployability on a single commodity XR headset paired with an A100-class GPU. GenEdXR offers an empirically grounded architectural blueprint for next-generation adaptive educational entertainment ecosystems.