AnnoBench: A Benchmark for Visualization Annotation Generation
The paper introduces a testbed for judging whether automated chart annotations are readable, accurate, and visually compatible.
AnnoBench uses visualizations from professional data journalism and visualization galleries, paired with annotation tasks across representation formats, chart-description conditions, and prompt-specificity levels. The authors run the benchmark with VLM-as-a-judge, using models aligned with manual human assessment. Their experiments test how input representation, semantic context, prompt detail, and model choice affect annotation quality. Source: ArXiv · AI/CL/LG's note
AnnoBench uses visualizations from professional data journalism and visualization galleries, paired with annotation tasks across representation formats, chart-description conditions, and prompt-specificity levels. The authors run the benchmark with VLM-as-a-judge, using models aligned with manual human assessment. Their experiments test how input representation, semantic context, prompt detail, and model choice affect annotation quality. Source: ArXiv · AI/CL/LG's note
score 4