Examining Vision Language Models through Multi-dimensional Experiments with Vision and Text Features

Recent research on Vision Language Models (VLMs) suggests that they rely on inherent biases learned during training to respond to questions about visual properties of an image. These biases are exacerbated when VLMs are asked highly specific questions that require focusing on specific areas of the image. For example, a VLM tasked with counting stars on a modified American flag (e.g., with more than 50 stars) will often disregard the visual evidence and fail to answer accurately. We build upon this research and develop a multi-dimensional examination framework to systematically determine which characteristics of the input data, including both the image and the accompanying prompt, lead to such differences in performance. Using open-source VLMs, we further examine how attention values fluctuate with varying input parameters (e.g., image size, number of objects in the image, background color, prompt specificity). This research aims to learn how the behavior of vision language models changes and to explore methods for characterizing such changes. Our results suggest, among other things, that even minor modifications in image characteristics and prompt specificity can lead to large changes in how a VLM formulates its answer and, subsequently, its overall performance.