VeriOCRBench - Visual Dimension (Proactive): leaderboard
The visual verification dimension of VeriOCRBench (Zhou et al., 2026): 400 OCR-grounded tasks with the traps image degradation and occluded target, where the text or visual evidence the question needs is unreadable or hidden. Diagnosis accuracy is the percentage on which the model both rejects the task and names the correct trap mechanism, judged by GPT-4o; higher is better. Proactive setting: the standard prompt, with no instruction to verify the task.
Metric: Diagnosis accuracy (%; the 400 trap-injected tasks of the visual verification dimension (image degradation and occluded target: the required text or visual evidence is unreadable or hidden); the model must reject the task and name the ground-truth trap mechanism; judged by GPT-4o at temperature 0; proactive paradigm: no verification instruction). Source: arxiv.org. Saturation forecast: Around August 2028. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 28.75 | #244 |
| 2 | Gemini 3.1 Pro (Preview) | 23.25 | #91 |
| 3 | GPT-5 | 20.5 | #163 |
| 4 | Qwen 3.5 35B A3B | 18.5 | #480 |
| 5 | Llama 4 Maverick | 14.75 | #1105 |
| 6 | Qwen 3.5 397B A17B | 14.5 | #242 |
| 7 | Llama 4 Scout | 11.5 | #1787 |
| 8 | GPT-4o | 11.25 | #741 |
| 9 | Qwen 3.5 122B A10B | 10.5 | #329 |
| 10 | InternVL3-38B | 6.75 | #905 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=veriocrbench-visual-dimension-proactive · How It Works · Data refreshed daily, snapshot 2026-11-12.