LinkS2Bench: leaderboard

LinkS2Bench (LinkS²Bench, 2026) tests whether vision-language models can anchor dynamic drone-video observations to static satellite maps, over 17,903 questions in 12 perception, localization, relation and reasoning tasks. Score is the mean of the 12 task accuracies (%); higher is better.

Metric: Average (%) of the 12 task scores (accuracy, MRA for zone counting, ACC@1s for temporal grounding) on the full LinkS2Bench (17,903 questions over 1,500 UAV videos paired with satellite images), temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)51.1#102
2Seed 1.847.8#258
3GPT-5.446.6#138
4Qwen 3.5 397B A17B45.6#270
5GLM-4.6V-Flash44.3#1049
6Qwen 3.5 35B A3B44.1#494
7Gemini-3.1-Flash [LinkS2Bench checkpoint unspecified]42.8
8Kimi K2.538.4#272
9GPT-5.4 Mini38.2#409
10Claude Sonnet 4.637.7#179
11LLaVA-OneVision-1.5-8B33.7#1876
12LLaVA-OneVision-1.5-4B33.2#1874
13Qwen 3.5 9B30.7#838
14Llama 4 Scout26.6#1825
15Qwen 3.5 4B23.7#1270

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

Interactive version: theaggregate.ai/benchmark?slug=links2bench · How It Works · Data refreshed daily, snapshot 2026-11-22.