Leveraging Large Language Models in a Retriever-Reader Framework for Solving STEM Multiple-Choice Questions

This research presents a novel Retriever-Reader framework to efficiently solve multiple-choice questions (MCQs) in STEM (Science, Technology, Engineering, and Mathematics) fields. Unlike previous approaches that primarily rely on either retrieval-based methods or reader-based models, our frame-work integrates both, significantly enhancing problem-solving performance. The Retriever Network selectively extracts relevant context from a customized STEM corpus, while the Reader Network, leveraging a fine-tuned DeBERTa-based language model, interprets this context to accurately solve MCQs. Through advanced model architecture design, data preprocessing, and comprehensive evaluation, our proposed framework outperforms state-of-the-art baseline models in both accuracy and F1 score. This work provides a new perspective on automated problem-solving in STEM domains by integrating multiple modeling techniques.