SocioSeg - Socio-Name: leaderboard
Segmentation of socially defined urban entities from a satellite image and a rendered digital map by a vision-language model without task training, which proposes box and point prompts for SAM (SocioSeg, Wang et al., 2026), for targets named as a specific place. Mean per-image IoU (gIoU) in percent; higher is better.
Metric: gIoU (%; mean per-image IoU of the predicted mask on the SocioSeg test set, socio-name targets (a named place); a general VLM without fine-tuning reads the rendered digital map and the satellite image and returns box and point prompts for SAM). Source: arxiv.org. Saturation forecast: Around 2033. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 72B Instruct | 29.5 | #845 |
| 2 | O3 | 22.9 | #224 |
| 3 | GPT-5 | 16.1 | #163 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=socioseg-socio-name · How It Works · Data refreshed daily, snapshot 2026-11-12.