PriVE-Bench - Counting: leaderboard

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: Around 2037. 8 models tracked.

Top models

#ModelScore
1Gemma 3 12B (IT)36.8
2Gemma 3 27B (IT)27.6
3Claude Sonnet 419.5
4GPT-5 (Minimal)15.1
5Qwen 3 VL 8B Instruct14.8
6Qwen 3 VL 32B Instruct13
7InternVL3.5-8B5.1

Interactive version: theaggregate.ai/benchmark?slug=prive-bench-counting · How It Works · Data refreshed daily, snapshot 2026-09-29.