PriVE-Bench - Counting: leaderboard
PriVE-Bench (2026) shows vision-language models counterfactual images whose visible content contradicts a familiar prior, such as an unusual part count or logo layout, in five domains, and a text-only judge classifies each answer as correct, prior-consistent or other. This board covers the counting domain, which asks about visible object-part or instance counts; score is counterfactual accuracy in percent, higher is better.
Metric: Counterfactual accuracy (%) in the counting domain; paired original and counterfactual images in five domains, counterfactual-only input, answers classified by a text-only judge (GPT-4o-mini for closed models, Qwen3-8B for open models); temperature 0, reasoning off or minimal. Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 12B (IT) | 36.8 | #1837 |
| 2 | Gemma 3 27B (IT) | 27.6 | #1289 |
| 3 | Claude Sonnet 4 | 19.5 | #400 |
| 4 | GPT-5 (Minimal) | 15.1 | #167 (GPT-5) |
| 5 | Qwen 3 VL 8B Instruct | 14.8 | #981 |
| 6 | Qwen 3 VL 32B Instruct | 13 | #585 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=prive-bench-counting · How It Works · Data refreshed daily, snapshot 2026-11-13.