HELM VHELM - Bingo - I2I: leaderboard
VHELM (Stanford CRFM HELM for vision-language models, run v2.1.2) Bingo I2I split: open-ended questions designed to expose hallucination from image-to-image interference, where similar images are shown together, each response rated 1-5 by a Prometheus-Vision judge; 124 questions. Mean rating on the 1-5 scale, higher is better.
Metric: Prometheus Vision Rating (1-5). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 03-25) | 4.45 | #126 |
| 2 | O3 (2025-04-16) | 4.45 | #125 |
| 3 | O4 Mini (2025-04-16) | 4.23 | #183 |
| 4 | O1 (2024-12-17) | 4.16 | #157 |
| 5 | Llama 4 Maverick Instruct FP8 | 4.02 | #424 |
| 6 | GPT-4o (2024-11-20) | 3.97 | #392 |
| 7 | GPT-4o Mini (2024-07-18) | 3.84 | #568 |
| 8 | GPT-4o (2024-05-13) | 3.79 | #337 |
| 9 | Gemini 2.0 Pro (Preview 02-05) | 3.79 | #318 |
| 10 | GPT-4.1 (2025-04-14) | 3.76 | #257 |
| 11 | Gemini 1.5 Flash (002) | 3.74 | #587 |
| 12 | Llama 4 Scout Instruct | 3.74 | #547 |
| 13 | Claude 3.5 Sonnet (20241022) | 3.73 | #302 |
| 14 | Gemini 1.5 Pro (002) | 3.68 | #368 |
| 15 | GPT-4o (2024-08-06) | 3.68 | #343 |
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-bingo-i2i · How It Works · Data refreshed daily, snapshot 2026-11-05.