GUI Agent Attack Success Rate Evaluation on MIRAGE (1,111-sample main set)
41FB Success Rategpt-4o-mini
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gpt-4o-miniModel Type=Closed-weight2026.05 | 41 | 21 | 46 | 30 | 22 | 26 | 23 | 32 | 23 | 39 | 30.2 | |
| Qwen3-VL-8B-InstructModel Size=8B, Backbone=Qwen3-VL2026.05 | 29 | 24 | 47 | 23 | 27 | 25 | 24 | 32 | 26 | 39 | 28.9 | |
| GLM-4.5VBackbone=GLM-4.5V2026.05 | 28 | 25 | 41 | 24 | 29 | 24 | 30 | 32 | 19 | 40 | 28.6 | |
| Qwen3-VL-32B-InstructModel Size=32B, Backbone=Qwen3-VL, Architecture=Dense2026.05 | 27 | 19 | 39 | 22 | 26 | 20 | 22 | 25 | 13 | 25 | 23 | |
| Qwen3-VL-30B-A3B-InstructModel Size=30B, Backbone=Qwen3-VL, Architecture=MoE2026.05 | 26 | 20 | 41 | 26 | 22 | 23 | 25 | 29 | 16 | 29 | 25 |