MPCI-Bench - Seed Tier Probing: leaderboard
MPCI-Bench (Wang and Zhang, 2026) seed tier: a multimodal model judges whether sharing a privacy-sensitive VISPR image is appropriate in a contextual-integrity scenario defined by sender, recipient, subject and transmission principle, over paired appropriate and inappropriate flows. Binary probe accuracy (%); higher is better.
Metric: Probe Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 97.3 | #163 |
| 2 | GPT-4o | 96 | #741 |
| 3 | Gemma 3 4B (IT) | 91 | #3848 |
| 4 | Mistral Large 3 | 90.5 | #819 |
| 5 | Gemma 3 12B (IT) | 84.7 | #1849 |
| 6 | Gemma 3 27B (IT) | 83.3 | #1306 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mpci-bench-seed-tier-probing · How It Works · Data refreshed daily, snapshot 2026-11-12.