Safety Evaluation on HADES
100Refusal RateSASA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SASABackbone=GEMMA32026.06 | 100 | 100 | |
| SASABackbone=QWEN3-VL2026.06 | 100 | 100 | |
| SASABackbone=QWEN3.52026.06 | 99.8 | 99.9 | |
| MARSBackbone=QWEN3.52026.06 | 99.5 | 99.8 | |
| SASABackbone=MOLMO22026.06 | 98.4 | 92.2 | |
| MARSBackbone=QWEN3-VL2026.06 | 98.3 | 99.6 | |
| MARSBackbone=MOLMO22026.06 | 95.8 | 97.6 | |
| AdaSteerBackbone=QWEN3.52026.06 | 92.4 | 98.6 | |
| SASABackbone=INTERNVL 3.52026.06 | 90.6 | 95.8 | |
| MARSBackbone=INTERNVL 3.52026.06 | 88.3 | 79.8 | |
| AdaSteerBackbone=QWEN3-VL2026.06 | 77.8 | 98.5 | |
| ECSOBackbone=MOLMO22026.06 | 73.2 | 91.3 | |
| ECSOBackbone=QWEN3-VL2026.06 | 67.2 | 97 | |
| ECSOBackbone=QWEN3.52026.06 | 66.1 | 92.2 | |
| Zero-shotBackbone=QWEN3.52026.06 | 65.3 | 91.2 | |
| Zero-shotBackbone=QWEN3-VL2026.06 | 62 | 92 | |
| MARSBackbone=GEMMA32026.06 | 59.2 | 82.1 | |
| AdaSteerBackbone=MOLMO22026.06 | 29.7 | 51.6 | |
| Zero-shotBackbone=GEMMA32026.06 | 17.9 | 37.4 | |
| ECSOBackbone=GEMMA32026.06 | 17.9 | 35.2 | |
| Zero-shotBackbone=MOLMO22026.06 | 17.2 | 41 | |
| ECSOBackbone=INTERNVL 3.52026.06 | 16.8 | 65.7 | |
| Zero-shotBackbone=INTERNVL 3.52026.06 | 16.7 | 62.6 | |
| AdaSteerBackbone=INTERNVL 3.52026.06 | 13.4 | 69.9 | |
| AdaSteerBackbone=GEMMA32026.06 | 11.4 | 32.9 |