A comparative analysis of encoder only and decoder only models for challenging LLM-generated STEM MCQs using a self-evaluation approach

Abstract

Large Language Models (LLMs) have demonstrated impressive capabilities in various tasks, including Multiple-Choice Question Answering (MCQA) evaluated on benchmark datasets with few-shot prompting. Given the absence of benchmark Science, Technology, Engineering, and Mathematics (STEM) datasets on Multiple-Choice Questions (MCQs) created by LLMs, we employed various LLMs (e.g., Vicuna-13B, Bard, and GPT-3.5) to generate MCQs on STEM topics curated from Wikipedia. We evaluated open-source LLM models such as Llama 2-7B and Mistral-7B Instruct, along with an encoder model such as DeBERTa v3 Large, on inference by adding context in addition to fine-tuning with and without context. The results showed that DeBERTa v3 Large and Mistral-7B Instruct outperform Llama 2-7B, highlighting the potential of LLMs with fewer parameters in answering hard MCQs when given the appropriate context through fine-tuning. We also benchmarked the results of these models against closed-source models such as Gemini and GPT-4 on inference with context, showcasing the potential of narrowing the gap between open-source and closed-source models when context is provided. Our work demonstrates the capabilities of LLMs in creating more challenging tasks that can be used as self-evaluation for other models. It also contributes to understanding LLMs’ capabilities in STEM MCQs tasks and emphasizes the importance of context for LLMs with fewer parameters in enhancing their performance.