NumerosityVLM: leaderboard
NumerosityVLM (2026) tests numerosity perception in vision-language models with 10,800 synthetic images of apples, butterflies or human figures, varying count, object size and spatial arrangement independently and removing texture, shape and color cues in turn. The model picks the count from 12 options between 1 and 100; score is exact-match accuracy in percent, with chance near 8.3; higher is better.
Metric: Exact-match accuracy (%; the model picks the object count from 12 options, 1 to 100 on a logarithmic grid, across 10,800 synthetic images in six controlled conditions; zero-shot, closed-ended prompt; chance is 8.3). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 83.97 | #981 |
| 2 | Qwen 3 VL 4B Instruct | 77.73 | #1279 |
| 3 | InternVL3-8B | 77.21 | #1688 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=numerosityvlm · How It Works · Data refreshed daily, snapshot 2026-11-13.