HELM VHELM - Unicorn - OODCV-VQA: leaderboard
VHELM (Stanford CRFM HELM for vision-language models, run v2.1.2) Unicorn OODCV-VQA split, an out-of-distribution safety test: questions with yes/no or numeric answers about out-of-distribution photos (unusual poses, shapes, textures, weather and occlusion); 1,000 questions, prefix quasi-exact match. Percent correct, 0-100, higher is better.
Metric: Prefix Quasi-Exact Match (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash | 82.8 | #351 |
| 2 | Gemini 1.5 Flash (002) | 82.7 | #587 |
| 3 | O4 Mini (2025-04-16) | 82.4 | #183 |
| 4 | O3 (2025-04-16) | 82.3 | #125 |
| 5 | Gemini 1.5 Pro (002) | 81.9 | #368 |
| 6 | GPT-4.5 | 81.8 | #278 |
| 7 | Gemini 2.0 Flash (Preview) | 81.6 | #376 |
| 8 | GPT-4o (2024-11-20) | 80.8 | #392 |
| 9 | Gemini 2.0 Pro (Preview 02-05) | 80.2 | #318 |
| 10 | Gemini 2.5 Pro (Preview 03-25) | 79.7 | #126 |
| 11 | Gemini 2.0 Flash Lite | 79.6 | #455 |
| 12 | GPT-4.1 Mini | 79.4 | #363 |
| 13 | GPT-4.1 (2025-04-14) | 79.1 | #257 |
| 14 | O1 (2024-12-17) | 78.9 | #157 |
| 15 | Llama 4 Maverick Instruct FP8 | 77.6 | #424 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-vhelm-unicorn-oodcv-vqa · How It Works · Data refreshed daily, snapshot 2026-11-05.