Visual context
A useful visual response should connect the question to the relevant parts of an image. The research challenge is to preserve those relationships without introducing details that are not present.
Documents and diagrams
Layout, labels, and spatial relationships each carry information. Evaluation should separate reading visible text from interpreting a chart or understanding the structure of a diagram.
Grounded answers
A model should distinguish direct observations from interpretations. Ambiguous or unreadable regions call for uncertainty, a closer view, or a better source.
A question to explore
Describe the visible components of this diagram. Separate what the image shows from what you infer.
This example illustrates a task direction. It is not a recorded demonstration or a claim of verified performance.
What to evaluate
Fine text, dense diagrams, unfamiliar objects, misleading visual context, and low-resolution inputs.
Explore the evaluation approach