Exaggerated Safety Mitigation on Exaggerated Safety
100NRRSAC Multi-task
Evaluation Results
| Method | Links | |
|---|---|---|
| SAC Multi-taskBackbone=Qwen2-72B-Instruct2024.11 | 100 | |
| SAC Multi-taskBackbone=Qwen2-7B-Instruct2024.11 | 99.5 | |
| SAC Single-taskBackbone=Qwen2-72B-Instruct2024.11 | 99.5 | |
| SAC Single-taskBackbone=Qwen2-7B-Instruct2024.11 | 96 | |
| SACControl Dim=Single2024.11 | 96 | |
| RepEControl Dim=Single2024.11 | 95 | |
| No ControlBackbone=Qwen2-72B-Instruct2024.11 | 90.5 | |
| SACControl Dim=Multiple2024.11 | 88.5 | |
| No ControlBackbone=Qwen2-7B-Instruct2024.11 | 88 | |
| No ControlControl Dim=Single2024.11 | 67 | |
| No ControlControl Dim=Multiple2024.11 | 67 | |
| RepE-MergeControl Dim=Multiple2024.11 | 63 | |
| RepE-MeanControl Dim=Multiple2024.11 | 61 |