Eye-Q - Iterative Refinement - English: leaderboard
Eye-Q (Najar et al., 2026), English visual word puzzles (300 synthetic scenes written by the authors and rendered with text-to-image models): each image encodes a hidden word or phrase through visual cues, and the model must output the exact target. The model may revise twice after being told its answer is wrong, and the puzzle counts if any of the three attempts is right. Exact-match accuracy in percent; higher is better.
Metric: Exact-Match Accuracy (%; up to three attempts, told when wrong, English subset). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 46.46 | #262 |
| 2 | Grok 4.1 Fast (Reasoning) | 46.13 | #426 (Grok 4.1 Fast) |
| 3 | GPT-5.2 (Non-reasoning) | 29.29 | #172 (GPT-5.2) |
| 4 | Gemini 2.5 Flash | 28.28 | #452 |
| 5 | Qwen 3 VL 235B A22B Instruct | 14.81 | #521 |
| 6 | Llama 4 Scout | 8.08 | #1787 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=eye-q-iterative-refinement-english · How It Works · Data refreshed daily, snapshot 2026-11-12.