Blind-Spots-Bench - Multimodal-to-Text: leaderboard
Questions that combine an image with text and ask for a text answer, targeting perception and reasoning blind spots that people find easy; from Blind-Spots-Bench (EPFL, 2026). Accuracy in percent averaged over four samples; higher is better.
Metric: Accuracy (%; mean@4 over four samples on the questions with image input and text output, vision-language models only; 235 human-authored questions with reference solutions graded correct or incorrect by gemini-3-flash with code execution; no tools; thinking enabled at medium effort where available; max 32,768 output tokens). Source: arxiv.org. Saturation forecast: Around February 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (Medium) | 66.9 | #94 (Gemini 3.1 Pro (Preview)) |
| 2 | Gemini 3 Flash (Preview) (Medium) | 64.5 | #144 (Gemini 3 Flash (Preview)) |
| 3 | GPT-5.5 | 58.7 | #49 |
| 4 | GPT-5.4 (Medium) | 58.1 | #123 (GPT-5.4) |
| 5 | GPT-5.2 (Medium) | 50 | #176 (GPT-5.2) |
| 6 | Qwen 3.5 122B A10B | 48.8 | #331 |
| 7 | Qwen 3.5 397B A17B | 48.3 | #246 |
| 8 | Qwen 3.5 35B A3B | 47.1 | #482 |
| 9 | GPT-5.4 Mini (Medium) | 45.3 | #389 (GPT-5.4 Mini) |
| 10 | Kimi K2.5 | 42.4 | #256 |
| 11 | Kimi K2.6 | 41.9 | #178 |
| 12 | gemma-4-26B-A4B-it (Thinking) | 40.7 | |
| 13 | GPT-5 | 37.8 | #167 |
| 14 | Gemini 2.5 Pro | 36 | #266 |
| 15 | GPT-5 Mini | 33.1 | #348 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=blind-spots-bench-multimodal-to-text · How It Works · Data refreshed daily, snapshot 2026-11-13.