PriVE-Bench: leaderboard

Metric: Counterfactual accuracy (%; macro-averaged over the five domains) on 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 April 2027. 8 models tracked.

Top models

#ModelScore
1GPT-5 (Minimal)46.7
2Claude Sonnet 446.6
3Qwen 3 VL 32B Instruct40.3
4Gemma 3 12B (IT)39.7
5Qwen 3 VL 8B Instruct31.6
6Gemma 3 27B (IT)30.5
7InternVL3.5-8B27.5

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