Dynamic Multi-view RAG: Mitigating Hallucinations of Large Language Models in Education

Nowadays, LLMs are increasingly used in education. However, hallucination severely undermines the reliability and pedagogical value of their responses. While Retrieval-Augmented Generation (RAG) helps to mitigate hallucinations with external knowledge, existing methods not only rely on limited retrieval sources, but also lack mechanisms to verify whether generated answers are sufficiently supported by retrieved evidence. To address these limitations, we propose a Dynamic Multi-View RAG (DMVR) framework. First, DMVR dynamically rewrites each original query from textbook and example perspectives to perform multi-view retrieval. In addition, a verification mechanism generation process encourages cross-checking across multi-view sources and explicitly grounds answers in supporting evidence. Experiments on question answering tasks show that DMVR outperforms baselines in answer accuracy, showing its effectiveness in mitigating hallucinations.