GeoMMBench - Spatial: leaderboard

The spatial relation questions of GeoMMBench (2026), an expert-written multimodal geoscience and remote sensing question benchmark answered zero-shot (test set). Micro-averaged accuracy in percent; higher is better.

Metric: Micro-averaged accuracy (%) on the spatial relation questions of the GeoMMBench test set (expert-written multimodal geoscience and remote sensing multiple-choice and numeric questions, zero-shot, LMMs-Eval, rule-based answer extraction, invalid outputs counted wrong); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 35 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2 VL 7B62.2#2275
2Gemini 2.0 Flash59.5#723
3Qwen 2.5 VL 3B59.5#2639
4InternLM-XComposer2-4KHD59.5#5427
5LLaVA-NeXT-34B59.5#3443
6Gemini 1.5 Pro56.8#1063
7Qwen 3 VL 4B56.8#1151
8Qwen 2.5 VL 7B54.1#1862
9O154.1#403
10LLaVA-OneVision-7B54.1#2981
11InternVL2.5-8B54.1#2370
12Phi-3.5-Vision Instruct54.1#6268
13Qwen 2 VL 2B51.4#6157
14Qwen 3 VL 8B51.4#917
15LLaVA-NeXT-7B51.4#5032

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

Interactive version: theaggregate.ai/benchmark?slug=geommbench-spatial · How It Works · Data refreshed daily, snapshot 2026-11-21.