NormViz-Bench: leaderboard
Grounding multimodal reasoning in global cultural norms: paired images where a norm is followed or violated; group accuracy requires both images of a pair to be judged correctly. Higher is better. Source: arxiv.org/abs/2609.06831.
Metric: Group Accuracy (%). Source: arxiv.org. Saturation forecast: Around 2034. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 25.3 | #103 |
| 2 | Qwen 2.5 7B | 23 | #895 |
| 3 | GPT-5.2 | 22.7 | #115 |
| 4 | Qwen 3 8B | 19.6 | #678 |
| 5 | Qwen 2.5 3B | 18.3 | #1177 |
| 6 | GPT-4o | 16.3 | #357 |
| 7 | Qwen 3 4B | 14.9 | #834 |
| 8 | Gemma 3 4B | 11.7 | #1088 |
| 9 | Molmo2-8B | 5 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=normviz-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.