VCG-Bench - Vision-to-XML - CodeXQA: leaderboard
Whether mxGraph XML generated from a raster diagram keeps its content: accuracy (%) on counting, identification and relationship questions answered from the generated XML alone, higher is better (VCG-Bench, Su, Dong et al., HKUST (Guangzhou), 2026).
Metric: CodeXQA Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 93.13 | #135 |
| 2 | Gemini 3 Flash | 92.63 | #159 |
| 3 | GPT-5.2 | 92.55 | #172 |
| 4 | Claude Sonnet 4.5 | 90.72 | #244 |
| 5 | GLM-4.6V | 90.32 | #698 |
| 6 | Claude Opus 4.5 | 89.47 | #152 |
| 7 | Qwen 3 VL 235B A22B | 82.24 | #477 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vcg-bench-vision-to-xml-codexqa · How It Works · Data refreshed daily, snapshot 2026-11-12.