Web Browsing Security and Task Completion on MiniWob++ (unseen tasks)
18.9Success Rate (ASR)Base Model
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Base ModelBackbone=Gemma-3-12B-IT2026.03 | 18.9 | 14 | |
| Online SFT2026.03 | 15.1 | 18.4 | |
| ART2026.03 | 14.6 | 21.8 | |
| SPAG2026.03 | 14.4 | 22.7 | |
| DMAST2026.03 | 10.8 | 25.7 | |
| Prompt DefenseStrategy=Zero-shot system prompts2026.03 | 7.4 | 15.3 | |
| DMAST + Prompt Def.Combination=DMAST with prompt-based defense2026.03 | 4.5 | 25 |