SArena-Icon - Understanding: leaderboard

Each model reads only the SVG code of an icon and answers 6,012 four-option multiple-choice questions about its colour, geometry, quantity and meaning, replying with the answer letter alone; the score is overall accuracy. Unlike SGP-Bench, the questions were written by GPT-4o from the rendered icons and checked by people, and the board covers only SArena's understanding task, not its editing or generation tasks.

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1InternSVG-8B85.1
2Claude Sonnet 4 (20250514)77.1#405
3Qwen 2.5 VL 32B Instruct65.5#1074
4InternVL3-78B65.3#755
5Qwen 2.5 VL 72B Instruct63.4#849
6InternVL3-8B59.5#1756
7Qwen 2.5 VL 7B Instruct52.8#1893

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=sarena-icon-understanding · How It Works · Data refreshed daily, snapshot 2026-11-17.