ExChart-Bench: leaderboard
ExChart-Bench (He et al., Zhejiang University, CHI 2026) gives a multimodal model a chart image with its data labels removed and a CSV template with the values masked, and asks it to fill in the numbers: 3,600 bar, line, scatter, pie and radar charts, 744 of them real-world charts cleaned by hand and 2,856 synthetic charts with rendering noise. The score pools all 33,757 values. Adaptive MAPE divides each value's absolute error by the chart's largest absolute value and caps it at 100%, and unparsable output or a table of the wrong shape counts as 100% error; lower is better.
Metric: Adaptive MAPE (%; 33,757 values in 3,600 unlabeled charts, per-value error relative to the chart's largest absolute value, capped at 100; unparsable or misshapen tables count 100). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-4.5V | 5.94 | #739 |
| 2 | Gemini 2.5 Flash | 6.72 | #452 |
| 3 | GPT-4.1 | 7.49 | #492 |
| 4 | Qwen 2.5 VL 72B Instruct | 9.1 | #845 |
| 5 | Qwen 2.5 VL 32B Instruct | 10.7 | #1055 |
| 6 | GPT-4o | 13.38 | #741 |
| 7 | Qwen 2.5 VL 7B Instruct | 16.01 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=exchart-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.