MM-InstructEval: Zero-Shot Evaluation of (Multimodal) Large Language Models on Multimodal Reasoning Tasks

The rising popularity of multimodal large language models (MLLMs) has sparked a significant increase in research dedicated to evaluating these models. However, current evaluation studies predominantly concentrate on the ability of models to comprehend and reason within a unimodal (vision-only) context, overlooking critical performance evaluations in complex multimodal reasoning tasks that integrate both visual and text contexts. Furthermore, tasks that demand reasoning across multiple modalities pose greater challenges and require a deep understanding of multimodal contexts. In this paper, we introduce a comprehensive assessment framework named MM-InstructEval, which integrates a diverse array of metrics to provide an extensive evaluation of the performance of various models and instructions across a broad range of multimodal reasoning tasks with vision-text contexts. MM-InstructEval enhances the research on the performance of MLLMs in complex multimodal reasoning tasks, facilitating a more thorough and holistic zero-shot evaluation of MLLMs. We firstly utilize the "Best Performance" metric to determine the upper performance limit of each model across various datasets. The "Mean Relative Gain" metric provides an analysis of the overall performance across different models and instructions, while the "Stability" metric evaluates their sensitivity to variations. Historically, the research has focused on evaluating models independently or solely assessing instructions, overlooking the interplay between models and instructions. To address this gap, we introduce the "Adaptability" metric, designed to quantify the degree of adaptability between models and instructions. Evaluations are conducted on 31 models (23 MLLMs) across 16 multimodal datasets, covering 6 tasks, with 10 distinct instructions. The extensive analysis enables us to derive novel insights.