Web cognitive reasoning on Web-CogBench
91.4Element Attribute AccuracyWeb-CogReasoner
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Web-CogReasoner2025.08 | 91.4 | 93.5 | 87.5 | 69.2 | 79 | 61.4 | 98.3 | 95.2 | 84.4 | |
| Gemini 2.5 Pro2025.08 | 79.8 | 94.6 | 84.4 | 62.6 | 73.5 | 51.9 | 96.6 | 98.4 | 80.2 | |
| Claude Sonnet 42025.08 | 79.7 | 93.5 | 62.5 | 62.8 | 54.3 | 64.7 | 100 | 96.8 | 76.8 | |
| UI-TARsModel size=7B, Training=SFT2025.08 | 63.5 | 88 | 31.3 | 48 | 48 | 32.4 | 25.9 | 33.9 | 46.4 | |
| Qwen2.5-VLModel size=7B2025.08 | 53.2 | 83.9 | 65.6 | 60 | 62 | 51.9 | 91.4 | 90.3 | 69.8 |