Evaluation and Analysis of Chart Reasoning Accuracy in Multimodal Large Language Models: An Empirical Study on Influencing Factors

This study presents a comprehensive empirical evaluation of chart reasoning capabilities in multimodal large language models (MLLMs), examining critical factors that influence performance accuracy across diverse visualization types. Through systematic experimentation with six leading MLLMs including GPT-4V, LLaVA, and BLIP-2, we analyze their proficiency in interpreting statistical charts, graphs, and data visualizations. Our methodology encompasses a curated dataset of 2,400 charts spanning bar graphs, line plots, scatter plots, pie charts, and complex multi-panel visualizations, each annotated with ground-truth reasoning tasks. Performance evaluation reveals significant variations based on chart complexity, data density, textual annotation presence, and visual design elements. Statistical analysis demonstrates that model accuracy decreases substantially with increased data point density (correlation coefficient: -0.73) and increased visual complexity. The study identifies optimal configurations for different chart types and provides actionable insights for improving MLLM deployment in data analysis applications. Our findings contribute to understanding multimodal AI limitations and establishing benchmarks for future chart comprehension research.