PuMVR: leaderboard
Script consistency in Punjabi visual reasoning from PuMVR (Not Truly Multilingual, Singh et al., 2026): each image question with four options is posed separately in Gurmukhi, Shahmukhi and Roman script, zero-shot, over a fixed 375-instance evaluation split. The Script Consistency Rate is the percent of instances answered correctly in all three scripts; higher is better.
Metric: Script Consistency Rate (%; instances answered correctly in all three scripts, zero-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 85.6 | #452 |
| 2 | Grok 4.1 Fast (Reasoning) | 79.73 | #426 (Grok 4.1 Fast) |
| 3 | GPT-4o | 78.13 | #741 |
| 4 | Claude Sonnet 4 | 77.09 | #397 |
| 5 | Qwen 2 VL 72B Instruct | 54.67 | #1403 |
| 6 | Qwen 2 VL 7B Instruct | 41.33 | #2569 |
| 7 | Llama 3.2 11B Instruct | 27.47 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.