ChartJudgeBench - CPA Text Occlusion: leaderboard
The text occlusion criterion of ChartJudgeBench Chart Perception Alignment (Wu et al., 2026): 157 chart pairs in which one chart carries overlapping text labels that impair readability, and a multimodal judge must prefer the unflawed chart. Accuracy in percent, counting a pair only when the judge picks the better chart in both presentation orders; higher is better.
Metric: Pairwise judging accuracy (%; the 157 text occlusion pairs; two-pass position-swapped protocol: each of ChartJudgeBench's 1,003 human-filtered chart pairs (an original chart and a copy with one injected flaw) is judged twice with the order reversed and counts as correct only if the judge picks the better chart both times; temperature 0.1). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 96.2 | #135 |
| 2 | GPT-5.2 | 77.1 | #172 |
| 3 | Qwen 3 VL 32B Instruct | 63.7 | #582 |
| 4 | Claude Sonnet 4.5 | 57.3 | #244 |
| 5 | Qwen 2.5 VL 72B Instruct | 52.9 | #845 |
| 6 | Qwen 2 VL 72B Instruct | 43.3 | #1403 |
| 7 | Qwen 3 VL 8B Instruct | 41.4 | #982 |
| 8 | Qwen 3 VL 30B A3B (Thinking) | 29.3 | #710 |
| 9 | Qwen 2 VL 7B Instruct | 13.4 | #2569 |
| 10 | Qwen 2.5 VL 7B Instruct | 3.8 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chartjudgebench-cpa-text-occlusion · How It Works · Data refreshed daily, snapshot 2026-11-12.