Concept-based Steering on AXBENCH (test)
1.102Overall Steering ScorePrOSV
Evaluation Results
| Method | Links | |
|---|---|---|
| PrOSVObjective=SimPO, Model setup=D_Q32B_L322026.05 | 1.102 | |
| PromptModel setup=D_G9B_L202026.05 | 1.075 | |
| PromptModel setup=D_Q32B_L322026.05 | 1.06 | |
| PrOSVObjective=Lang., Model setup=D_Q32B_L322026.05 | 1.049 | |
| FSSV (RePS) + Joint trainingObjective=SimPO, Model setup=D_Q32B_L322026.05 | 0.982 | |
| FSSV (RePS)Objective=SimPO, Model setup=D_Q32B_L322026.05 | 0.947 | |
| FSSV + Joint trainingObjective=Lang., Model setup=D_Q32B_L322026.05 | 0.919 | |
| PrOSVObjective=SimPO, Model setup=D_G9B_L202026.05 | 0.905 | |
| FSSV (RePS)Objective=SimPO, Model setup=D_G9B_L202026.05 | 0.892 | |
| FSSV (RePS) + Joint trainingObjective=SimPO, Model setup=D_G9B_L202026.05 | 0.886 | |
| PrOSVObjective=Lang., Model setup=D_G9B_L202026.05 | 0.859 | |
| FSSV + Joint trainingObjective=Lang., Model setup=D_G9B_L202026.05 | 0.821 | |
| PrOSVObjective=SimPO, Model setup=D_G2B_L102026.05 | 0.803 | |
| FSSVObjective=Lang., Model setup=D_Q32B_L322026.05 | 0.798 | |
| FSSVObjective=Lang., Model setup=D_G9B_L202026.05 | 0.788 | |
| LoReFTModel setup=D_G9B_L202026.05 | 0.777 | |
| FSSV (RePS) + Joint trainingObjective=SimPO, Model setup=D_G2B_L102026.05 | 0.769 | |
| PrOSVObjective=Lang., Model setup=D_G2B_L102026.05 | 0.758 | |
| FSSV (RePS)Objective=SimPO, Model setup=D_G2B_L102026.05 | 0.756 | |
| FSSV + Joint trainingObjective=Lang., Model setup=D_G2B_L102026.05 | 0.736 | |
| LoReFTModel setup=D_G2B_L102026.05 | 0.701 | |
| PromptModel setup=D_G2B_L102026.05 | 0.698 | |
| FSSVObjective=Lang., Model setup=D_G2B_L102026.05 | 0.663 | |
| SAEModel setup=D_G9B_L202026.05 | 0.546 | |
| DiffMeanModel setup=D_G9B_L202026.05 | 0.322 | |
| DiffMeanModel setup=D_G2B_L102026.05 | 0.297 | |
| FSSV + Fixed unit factorObjective=Lang., Model setup=D_G2B_L102026.05 | 0.072 | |
| FSSV + Fixed unit factorObjective=Lang., Model setup=D_G9B_L202026.05 | 0.024 |