SpatialQuery-1M - Closest-Instance Distance: leaderboard
SpatialQuery-1M (Nguyen et al., 2026) asks a vision-language model, from a single indoor RGB image, to find the visible instance of an object category closest to a reference object and estimate their floor-plane distance. This board is the percentage of 5,000 fixed queries answered within 0.2 m, credited only when the answer is correct on the image and not on a black image of the same size; higher is better.
Metric: Acc@0.2m (%; share of 5,000 fixed closest-instance distance queries covering all 200 scenes whose estimated gravity-aligned floor-plane distance between the reference object and its nearest same-category candidate is within 0.2 m of the ground truth; black-image sanity filter: a query earns credit only when the model returns valid outputs for both the original RGB image and a same-size uniform black image, is correct on the original and incorrect on the black image; invalid outputs score zero; zero-shot, identical decoding settings). Source: arxiv.org. Saturation forecast: Around 2036. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o Mini | 27.86 | #1623 |
| 2 | Qwen 2.5 VL 7B | 15.5 | #1787 |
| 3 | Gemini 2.5 Pro | 13.68 | #266 |
| 4 | Gemini 2.5 Flash | 11.32 | #453 |
| 5 | GPT-4o | 6.04 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=spatialquery-1m-closest-instance-distance · How It Works · Data refreshed daily, snapshot 2026-11-13.