PriVE-Bench - Object Existence: leaderboard
Metric: Counterfactual accuracy (%) in the object existence 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 December 2026. 8 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 74.4 |
| 2 | Gemma 3 27B (IT) | 65.6 |
| 3 | Claude Sonnet 4 | 62.3 |
| 4 | GPT-5 (Minimal) | 57.7 |
| 5 | Gemma 3 12B (IT) | 46.7 |
| 6 | Qwen 3 VL 8B Instruct | 31.9 |
| 7 | InternVL3.5-8B | 31 |
Interactive version: theaggregate.ai/benchmark?slug=prive-bench-object-existence · How It Works · Data refreshed daily, snapshot 2026-09-29.