ArabGeoText - OSM Test-I - YOLO26 Zero-Shot WER: leaderboard
arXiv2611.03339v1 TableVI; pretrained reader plus identical YOLO26 detector and rectified crops; no LoRA. Test-I Riyadh,Test-II Jeddah. Recognition conditioned on matched labels, same detections for both readers; E2E H includes localization errors. Direct/adapted checkpoint results are separate and withheld from base models.
Metric: Word error rate (%). Source: arxiv.org. Saturation forecast: Around June 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 4B | 19 | #514 |
| 2 | Qwen 3.5 0.8B | 40.7 | #1410 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=arabgeotext-osm-test-i-yolo26-zero-shot-wer · How It Works · Data refreshed daily, snapshot 2026-11-05.