Safety Evaluation on APGD-JailBreak
3.31ASR (Unconstrained)CARE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CAREBackbone=Qwen2.5-VL2026.03 | 3.31 | 11.17 | 8.82 | 6.13 | |
| CAREBackbone=LLaVA-OneVision2026.03 | 6.37 | 10.75 | 8.23 | 8.18 | |
| ASTRABackbone=LLaVA-OneVision2026.03 | 7.26 | 13.3 | 12.13 | 10.24 | |
| ASTRABackbone=Qwen2.5-VL2026.03 | 9.47 | 17.54 | 13.35 | 12.26 | |
| Original modelBackbone=Qwen2.5-VL2026.03 | 69.37 | 60.63 | 64.45 | 73.32 | |
| Original modelBackbone=LLaVA-OneVision2026.03 | 72.11 | 63.18 | 66.43 | 70.7 |