A model must answer questions about both a chart and its caption. What distinguishes the task from text-only modeling?

1. The answer may require integrating visual and linguistic evidence
2. Text tokens always encode the complete chart automatically
3. Only the image's filename can provide visual information
4. Every correct answer must appear verbatim in the caption