A visual question-answering model produces fluent answers that contradict the image. What must evaluation measure?

1. Only text generation speed
2. Only grammatical correctness
3. Visual grounding as well as language quality
4. Only the number of output tokens