Building Resilient Educational AI Through Multimodal Context Fusion

Abstract

Educational systems face escalating challenges: diverse learner needs, resource limits, and rapid technological change. Resilient AI is needed to remain useful as learner behavior, interface conditions, and infrastructure vary. We present a position on resilient educational AI through multimodal context fusion, demonstrated with a virtual reality (VR) algebra learning system. We define resilience operationally through three measurable indicators: (1) action-grounding accuracy, (2) tolerance to latency variation and fallback conditions, and (3) stability under changing learner behaviors across mission phases. In a controlled ablation using 30 scripted traces, the context-grounded system achieved 92% action reference accuracy and reduced hallucination events by 86% relative to a mission-only baseline. Complementary pilot logs from live VR use show sustained learner-initiated interaction and stable response timing, supporting the human-AI collaboration claim while remaining bounded by the absence of direct trust/usability surveys. We position multimodal context fusion as a practical design pattern for robust, adaptable, and scalable educational AI, with implications for broader resilience domains.