List your company

Jailbreak Attack on SPA-VL Harm

10.57ASROriginal
8.909620.117331.32542.5327Jan 29, 2026

Evaluation results

MethodLinks
Originalmode=baseline2026.0110.57
Only Image Noisingstrategy=Adversarial image noising2026.0113.96
Static Rewritingstrategy=Static template rewriting2026.0124.15
ReAct Dynamic Prompt Rewritingmode=Only ReAct Rewriting2026.0149.81
ReAct-driven adaptive noising mechanismmode=ReAct & Noise2026.0152.08