SpaceNum: leaderboard
SpaceNum (Zhang et al., 2026) tests whether vision-language models ground numbers in space with 3,800 simulator-built multiple-choice items: choosing the view that a numeric camera move, rotation or map coordinate produces (Num2Space), or reading the number off the views (Space2Num), for AI2-THOR transitions and 1D-3D object layouts at desk and room scale. Accuracy in percent, macro-averaged over 20 task cells, where random guessing scores 30; higher is better. Models answer with the option letter only.
Metric: Accuracy (%; macro-average of 20 task cells over 3,800 multiple-choice items, answer letter only). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 7B | 33 | #1796 |
| 2 | Gemma 3 27B | 31.2 | #1638 |
| 3 | Gemma 3 12B | 30.6 | #2013 |
| 4 | Gemma 3 4B | 28.5 | #5104 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=spacenum · How It Works · Data refreshed daily, snapshot 2026-11-12.