Safety Evaluation on MM-SafetyBench (MMSB)
1.8Attack Success Rate (V-T AVG)Refusal Steering
Evaluation Results
| Method | Links | |
|---|---|---|
| Refusal SteeringBase Model=Qwen2.5-Omni-3B2026.05 | 1.8 | |
| ReGapBase Model=Qwen2.5-Omni-3B2026.05 | 14.2 | |
| Refusal SteeringBase Model=Qwen3-Omni-30B-A3B-Instruct2026.05 | 21 | |
| Refusal SteeringBase Model=MiniCPM-o-4.52026.05 | 23.6 | |
| ReGapBase Model=Qwen3-Omni-30B-A3B-Instruct2026.05 | 36.3 | |
| BaseBase Model=Qwen3-Omni-30B-A3B-Instruct2026.05 | 58.1 | |
| ReGapBase Model=MiniCPM-o-4.52026.05 | 65.4 | |
| BaseBase Model=MiniCPM-o-4.52026.05 | 69.8 | |
| BaseBase Model=Qwen2.5-Omni-3B2026.05 | 76.8 |