This paper proposes InfoCausalQA, a new benchmark for evaluating the causal inference capabilities of visual language models (VLMs). InfoCausalQA consists of two tasks: quantitative causal inference and semantic causal inference. InfoCausalQA evaluates causal inference based on infographics, which combine structured visual data with textual information. Using GPT-4, we generated 1,482 multiple-choice question-answer pairs based on 494 infographic-text pairs collected from four publicly available sources. These pairs were manually reviewed to ensure that answers could not be derived solely from superficial clues. Experimental results show that existing VLMs exhibit limited capabilities in both computational and semantic causal inference, significantly outperforming humans. This highlights the need to improve causal inference capabilities using infographic-based information.
