WaldoBench: Where’s Waldo?

Finding Waldo requires spatial reasoning, pattern recognition, and attention to detail across densely cluttered illustrations. We hand-annotated ground truth coordinates for Waldo in each scene, then tested whether top vision models from Anthropic, OpenAI, and Google can predict his location. Models return normalized coordinates, and a prediction is a hit if it falls within a close percentage of the true position.

Each model is tested in two modes: a single-shot mode where the model must find Waldo from the full image in one pass, and an agentic mode where the model is given a crop_image tool and can zoom into up to 20 regions of the scene before giving its final answer.