HELM VHELM - Blink Perception - Relative Reflectance: leaderboard
VHELM (Stanford CRFM HELM for vision-language models, run v2.1.2) BLINK Relative Reflectance task: comparing the surface reflectance (albedo) of marked points, one of BLINK's perception tasks that humans solve at a glance, asked as multiple choice; 134 questions, prefix quasi-exact match. Percent correct, 0-100, higher is better.
Metric: Prefix Quasi-Exact Match (%). Source: crfm.stanford.edu. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3 Sonnet (20240229) | 68.66 | #694 |
| 2 | Claude 3 Haiku (20240307) | 67.16 | #780 |
| 3 | Claude 3.5 Sonnet (20241022) | 65.67 | #302 |
| 4 | Llama 4 Scout Instruct | 65.67 | #547 |
| 5 | GPT-4o Mini (2024-07-18) | 64.93 | #568 |
| 6 | Llama 4 Maverick Instruct FP8 | 64.18 | #424 |
| 7 | GPT-4 Turbo | 63.43 | #417 |
| 8 | GPT-4o (2024-11-20) | 59.7 | #392 |
| 9 | Claude 3 Opus (20240229) | 59.7 | #428 |
| 10 | Gemini 2.0 Flash Lite | 58.96 | #455 |
| 11 | GPT-4o (2024-05-13) | 58.21 | #337 |
| 12 | O4 Mini (2025-04-16) | 58.21 | #183 |
| 13 | Claude 3.5 Sonnet (20240620) | 57.46 | #308 |
| 14 | Gemini 2.5 Pro (Preview 03-25) | 55.22 | #126 |
| 15 | GPT-4.5 | 55.22 | #278 |
Interactive version: theaggregate.ai/benchmark?slug=helm-vhelm-blink-perception-relative-reflectance · How It Works · Data refreshed daily, snapshot 2026-11-05.