Enhanced Explainable Multi-Agent Generative Artificial Intelligence Framework for Personalized Education

Abstract

Generative Artificial Intelligence (GenAI) and Large Language Models (LLMs) have enabled personalized education via intelligent tutoring, adaptive learning, automated assessments, and personalized feedback. Nevertheless, educational AI systems are still confronted with challenges, including high computational complexity, long inference time, unreliable content generation, limited explainability, and ineffective multiagent orchestration. To this end, this study proposes LITE-Trust, a Lightweight Trust-Aware Explainable multiagent Generative Artificial Intelligence Framework for personalized education, which consists of learner profiling, Knowledge State Estimation (KSE), the Educational Library and Readers Optimization (ELRA) algorithm, relevance-based selective agent orchestration, trust evaluation, and an Explainability Engine. The ELRA algorithm is applied to optimize learner-specific educational configurations by taking learning requirements, trust, and computational efficiency into account, and selective orchestration is conducted to reduce computational overhead by only activating relevant agents. Experiments on the EdNet and Open University Learning Analytics Dataset (OULAD) demonstrate that LITE-Trust achieves 10.5/9.5ms latency, 80.5/81.3 MB memory, and 1.92 M parameters, which is computationally more efficient than the compared methods. In addition, LITE-Trust attains 95.0%/94.6% fidelity, 94.1%/93.4% explanation quality, and 92.6%/92.1% interpretability on the two datasets. Besides, the ELRA algorithm significantly outperforms existing algorithms in terms of various metrics. Overall, the experimental results verify that LITE-Trust can effectively balance personalization, learning quality, trust, explainability, and computational efficiency, thereby facilitating the practical application of lightweight and trustworthy GenAI systems in personalized education.