AdvSpot - Localization IoU: leaderboard
AdvSpot (inclusionAI, from the ArmorOCR paper) tests grounded OCR on adversarial text: rotated, mirrored, tiny, stylized, handwritten, degraded, low-contrast, pattern-overlaid, dot/line/symbol-encoded and AI-generated fusion text, 390 images with 397 region-grounded question-answer pairs. This board reports average intersection-over-union between the predicted and ground-truth box for the target text region. ArmorOCR is the authors' own Qwen3-VL-8B model trained for adversarial OCR.
Metric: Average Localization IoU (%). Source: arxiv.org. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 63 | #105 |
| 2 | Qwen 3 VL 32B Instruct | 51.9 | #292 |
| 3 | GLM-4.6V | 51 | #331 |
| 4 | Qwen 3 VL 30B A3B Instruct | 50.6 | #391 |
| 5 | Qwen 3 VL 8B Instruct | 49.1 | #420 |
| 6 | GLM-4.5V | 48.8 | #357 |
| 7 | Qwen 3 VL 235B A22B Instruct | 47.5 | #282 |
| 8 | Qwen 3.5 397B A17B | 47 | #155 |
| 9 | Kimi K2.5 | 41.6 | #154 |
| 10 | Kimi K2.6 | 41.1 | #115 |
| 11 | Gemini 3.5 Flash | 40.5 | #68 |
| 12 | Qwen 3.5 122B A10B | 38.5 | #186 |
| 13 | Gemini 2.5 Pro | 36.2 | #161 |
| 14 | GPT-5 | 35 | #108 |
| 15 | GPT-5.4 | 34.5 | #85 |
Interactive version: theaggregate.ai/benchmark?slug=advspot-localization-iou · How It Works · Data refreshed daily, snapshot 2026-11-07.