PuMVR - Roman: leaderboard
Zero-shot accuracy on PuMVR (Not Truly Multilingual, Singh et al., 2026) image questions written in Roman (Latin-alphabet) Punjabi, common in the diaspora and online: the model sees an image with a four-option question and must output the exact text of the correct option. Percent correct over the fixed 375-instance evaluation split, with empty responses left out of the denominator; higher is better.
Metric: Zero-shot Accuracy (%; Roman script). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 92.27 | #452 |
| 2 | Grok 4.1 Fast (Reasoning) | 90.13 | #426 (Grok 4.1 Fast) |
| 3 | GPT-4o | 86.93 | #741 |
| 4 | Claude Sonnet 4 | 85.83 | #397 |
| 5 | Qwen 2 VL 72B Instruct | 80.25 | #1403 |
| 6 | Qwen 2 VL 7B Instruct | 65.07 | #2569 |
| 7 | Llama 3.2 11B Instruct | 60.27 | #5039 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pumvr-roman · How It Works · Data refreshed daily, snapshot 2026-11-12.