Sentiment Steering on Llama2-7B Generation Evaluation Set
90.15Accuracy (CA)Vanilla
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VanillaAttack=Sleeper Agent2025.11 | 90.15 | 93.13 | |
| VanillaAttack=CBA2025.11 | 89.17 | 90.47 | |
| OURSAttack=CBA2025.11 | 88.45 | 23.18 | |
| VanillaAttack=VPI2025.11 | 88.39 | 86.31 | |
| OURSAttack=Sleeper Agent2025.11 | 88.19 | 16.99 | |
| OURSAttack=VPI2025.11 | 87.13 | 24.57 | |
| GracCeFulAttack=CBA2025.11 | 87.05 | 16.45 | |
| GracCeFulAttack=Sleeper Agent2025.11 | 86.97 | 17.39 | |
| CleanGenAttack=Sleeper Agent2025.11 | 86.38 | 26.17 | |
| GracCeFulAttack=VPI2025.11 | 86.19 | 10.98 | |
| CleanGenAttack=VPI2025.11 | 85.44 | 33.51 | |
| CleanGenAttack=CBA2025.11 | 85.01 | 34.34 | |
| MuScleLoRAAttack=CBA2025.11 | 83.94 | 21.76 | |
| MuScleLoRAAttack=VPI2025.11 | 78.19 | 20.36 | |
| MuScleLoRAAttack=Sleeper Agent2025.11 | 76.11 | 18.55 |