Evaluating Multimodal Large Language Models on Educational Textbook Question Answering

Abstract: Multimodal LLMs (MLLMs) have recently achieved significant success in vision--language tasks. However, their capacity to reason over complex, long lessons and intricate educational diagrams that cannot be represented as a single natural image remains largely untested. In this work, we present the first evaluation of state-of-the-art MLLMs on the textbook question answering (TQA) task using the CK12-QA dataset. We assess the performance of recent vision-LLMs, including LLaVA and LLaMA 3.2-Vision, across various input configurations. Additionally, we introduce a lightweight multimodal retrieval-augmented generation (RAG) pipeline that integrates both paragraphs and diagrams from the lesson into the prompt. Our results demonstrate the influence of retrieved educational context on model accuracy and reasoning, while also revealing current limitations in handling question-context relationships and the potential for noise, pointing to key directions for future research in multimodal AI-driven learning.