Benchmark · 2026

PolyChartQA: multi-chart figures

A multi-chart scientific-figure benchmark whose name collides with a separate multilingual chart benchmark.

Open the primary source ↗

Multiple charts and documents

Answer 2,694 questions over 534 images containing 2,297 subcharts from computer-science papers.

Human and model-judge accuracy, BERTScore, source and difficulty ablations, and manually verified questions.

Shows that localization and reasoning across subcharts are harder, especially for human-authored questions.

The paper-domain sample and mostly homogeneous subchart types do not represent every multi-chart document or audience.

Where this appears in the research.

Audience routes that point here.