AdvSpot: 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 sample-weighted grounded VQA accuracy: an answer counts when its text set covers the ground-truth transcription for the target region. ArmorOCR is the authors' own Qwen3-VL-8B model trained for adversarial OCR.
Metric: Grounded VQA Accuracy (%). Source: arxiv.org. Saturation forecast: Not forecast. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 45.6 | #105 |
| 2 | Gemini 3.5 Flash | 42.6 | #68 |
| 3 | Gemini 2.5 Pro | 37.3 | #161 |
| 4 | Qwen 3 VL 32B Instruct | 35.3 | #292 |
| 5 | Qwen 3 VL 235B A22B Instruct | 34.8 | #282 |
| 6 | Qwen 3 VL 30B A3B Instruct | 34.3 | #391 |
| 7 | Gemini 2.5 Flash | 32.3 | #255 |
| 8 | Qwen 3 VL 8B Instruct | 31.2 | #420 |
| 9 | GPT-5 | 30 | #108 |
| 10 | Kimi K2.6 | 29.7 | #115 |
| 11 | Kimi K2.5 | 28.2 | #154 |
| 12 | GLM-4.6V | 27.7 | #331 |
| 13 | Qwen 3.5 397B A17B | 26.7 | #155 |
| 14 | GPT-5.1 | 25.9 | #146 |
| 15 | GLM-4.5V | 24.7 | #357 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=advspot · How It Works · Data refreshed daily, snapshot 2026-11-07.