Jailbreak Attack on SPA-VL Harm
10.57ASROriginal
Evaluation Results
| Method | Links | |
|---|---|---|
| Originalmode=baseline2026.01 | 10.57 | |
| Only Image Noisingstrategy=Adversarial image noising2026.01 | 13.96 | |
| Static Rewritingstrategy=Static template rewriting2026.01 | 24.15 | |
| ReAct Dynamic Prompt Rewritingmode=Only ReAct Rewriting2026.01 | 49.81 | |
| ReAct-driven adaptive noising mechanismmode=ReAct & Noise2026.01 | 52.08 |