Sentiment Steering on Mistral-7B Generation (Evaluation Set)
96.38Control Accuracy (CA)Vanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaAttack=CBA2025.11 | 96.38 | 90.89 | |
| VanillaAttack=Sleeper Agent2025.11 | 96.14 | 95.77 | |
| VanillaAttack=VPI2025.11 | 94.99 | 87.34 | |
| GracCeFulAttack=Sleeper Agent2025.11 | 93.41 | 15.48 | |
| OURSAttack=CBA2025.11 | 92.74 | 18.13 | |
| GracCeFulAttack=CBA2025.11 | 92.6 | 17.25 | |
| OURSAttack=Sleeper Agent2025.11 | 92.41 | 25.71 | |
| CleanGenAttack=CBA2025.11 | 91.89 | 25.46 | |
| CleanGenAttack=Sleeper Agent2025.11 | 91.55 | 26.17 | |
| GracCeFulAttack=VPI2025.11 | 90.54 | 9.03 | |
| CleanGenAttack=VPI2025.11 | 90.37 | 30.29 | |
| OURSAttack=VPI2025.11 | 90.15 | 20.32 | |
| MuScleLoRAAttack=CBA2025.11 | 87.14 | 18.2 | |
| MuScleLoRAAttack=VPI2025.11 | 84.07 | 17.55 | |
| MuScleLoRAAttack=Sleeper Agent2025.11 | 82.85 | 14.08 |