Web browsing agent security and task completion on VisualWebArena OOD tasks
41.2ASRBase Model
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Base ModelBackbone=Gemma-3-12B-IT2026.03 | 41.2 | 6.2 | |
| SPAG2026.03 | 35.1 | 6.2 | |
| Online SFT2026.03 | 33 | 7.2 | |
| ART2026.03 | 30.9 | 8.2 | |
| DMAST2026.03 | 21.4 | 10.2 | |
| Prompt DefenseStrategy=Zero-shot system prompts2026.03 | 8.2 | 3.1 | |
| DMAST + Prompt Def.Combination=DMAST with prompt-based defense2026.03 | 7.2 | 8.2 |