VLM-as-a-Judge

C$^2$: Scalable Auto-Feedback for LLM-based Chart Generation