But a new paper highlights how many state-of-the-art "vision learning Models" (VLMs) often fail at simple, low-level visual analysis tasks that are trivially easy for a human.