PriVE-Bench - Prior-Following Error Rate: leaderboard

Metric: Prior-following error rate (%; answers that give the canonical prior-consistent value the image contradicts, 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 September 2027. 8 models tracked.

Top models

#ModelScore
1Claude Sonnet 451
2GPT-5 (Minimal)51.3
3Gemma 3 12B (IT)53.9
4Qwen 3 VL 32B Instruct58.8
5Gemma 3 27B (IT)61.1
6Qwen 3 VL 8B Instruct66.2
7InternVL3.5-8B68.6

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