Scale-VQA: leaderboard
Scale-VQA (Phan et al., 2026) poses multiple-choice pathology questions over regions of TCGA whole-slide images at 10x, 40x and 200x magnification that need evidence from at least two magnifications, screened so that text-only models cannot answer them and validated by pathologists. Accuracy in percent on the 1,005-question test split, averaged over five reasoning dimensions; higher is better.
Metric: Accuracy (%; multiple-choice questions on paired 10x/40x/200x views of TCGA whole-slide images, Scale-VQA test split, 1,005 questions, mean of five dimensions). Source: arxiv.org. Saturation forecast: Around November 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 60.9 | #159 |
| 2 | GPT-5.2 | 58.51 | #172 |
| 3 | Qwen 2.5 VL 7B Instruct | 53.63 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scale-vqa · How It Works · Data refreshed daily, snapshot 2026-11-12.