Appear2Meaning: leaderboard
Appear2Meaning (2026) tests whether vision-language models infer structured cultural metadata from an image of a museum object, over Getty and Metropolitan Museum objects from East Asia, the ancient Mediterranean, Europe and the Americas, judged against the museum records. Share of objects with every predicted field judged correct, in percent; higher is better.
Metric: Exact-match accuracy (%, times 100; every predicted metadata field judged correct) on the Appear2Meaning benchmark (750 Getty and Met museum objects, 50 per culture-type combination over four cultural regions and four object types; image-only input, structured JSON prediction; GPT-4.1-mini judge against the museum metadata); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 9 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 2.9 | #607 |
| 2 | Qwen 3 VL 8B Instruct | 2.4 | #1043 |
| 3 | Qwen-VL-Max | 1.4 | #1402 |
| 4 | Qwen 3 VL Plus | 1.4 | #514 |
| 5 | Qwen3-VL-Flash | 1.4 | #734 |
| 6 | GPT-4.1 Mini | 1.3 | #799 |
| 7 | Claude Haiku 4.5 | 1.2 | #587 |
| 8 | Pixtral-12B | 0.9 | #2611 |
| 9 | GPT-5.4 Mini | 0.5 | #409 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=appear2meaning · How It Works · Data refreshed daily, snapshot 2026-11-22.