HELM VHELM - Seed Bench - Visual Reasoning: leaderboard
VHELM (Stanford CRFM HELM for vision-language models, run v2.1.2) SEED-Bench visual reasoning dimension: four-option multiple-choice image-understanding questions; 331 questions, prefix quasi-exact match. Percent correct, 0-100, higher is better; guessing scores 25.
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 1.5 Flash (002) | 83.99 | #587 |
| 2 | GPT-4o (2024-08-06) | 83.69 | #343 |
| 3 | GPT-4o Mini (2024-07-18) | 83.38 | #568 |
| 4 | Gemini 2.5 Pro (Preview 03-25) | 83.38 | #126 |
| 5 | O3 (2025-04-16) | 83.38 | #125 |
| 6 | GPT-4o (2024-05-13) | 83.08 | #337 |
| 7 | GPT-4.5 | 82.78 | #278 |
| 8 | O4 Mini (2025-04-16) | 82.18 | #183 |
| 9 | Gemini 1.5 Pro (002) | 81.87 | #368 |
| 10 | GPT-4.1 (2025-04-14) | 81.87 | #257 |
| 11 | GPT-4o (2024-11-20) | 81.87 | #392 |
| 12 | GPT-4.1 Mini | 81.57 | #363 |
| 13 | Gemini 2.0 Pro (Preview 02-05) | 81.57 | #318 |
| 14 | GPT-4 Turbo | 81.27 | #417 |
| 15 | Gemini 2.0 Flash Lite | 80.97 | #455 |
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-seed-bench-visual-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-05.