Safety Evaluation on APGD-Toxic
3.28ASR (Unconstrained)ASTRA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ASTRABackbone=LLaVA-OneVision2026.03 | 3.28 | 12.4 | 9.75 | 6.62 | |
| CAREBackbone=Qwen2.5-VL2026.03 | 4.43 | 9.37 | 9.62 | 8.18 | |
| CAREBackbone=LLaVA-OneVision2026.03 | 4.77 | 9.11 | 8.7 | 7.15 | |
| ASTRABackbone=Qwen2.5-VL2026.03 | 7.13 | 16.08 | 13.73 | 10.12 | |
| Original modelBackbone=Qwen2.5-VL2026.03 | 67.71 | 59.38 | 61.03 | 70.62 | |
| Original modelBackbone=LLaVA-OneVision2026.03 | 71.11 | 56.16 | 60.6 | 62.9 |