Hallucination Steering on CAA
19.5RuntimeBase
Evaluation Results
| Method | Links | |
|---|---|---|
| BaseSteering Protocol=Pairwise, Model=Llama2-7b-chat-hf2026.03 | 19.5 | |
| BaseSteering Protocol=Positive, Model=Llama2-7b-chat-hf2026.03 | 19.5 | |
| COLD-KernelSteering Protocol=Positive, Model=Llama2-7b-chat-hf2026.03 | 31.14 | |
| COLD-KernelSteering Protocol=Pairwise, Model=Llama2-7b-chat-hf2026.03 | 39.29 | |
| DiffMeanSteering Protocol=Pairwise, Model=Llama2-7b-chat-hf2026.03 | 48.71 | |
| COLD-FDSteering Protocol=Positive, Model=Llama2-7b-chat-hf2026.03 | 53.31 | |
| ReFT(mlp)Steering Protocol=Positive, Model=Llama2-7b-chat-hf2026.03 | 69 | |
| ReFT(vector)Steering Protocol=Positive, Model=Llama2-7b-chat-hf2026.03 | 69.06 | |
| COLD-FDSteering Protocol=Pairwise, Model=Llama2-7b-chat-hf2026.03 | 69.81 | |
| Base(ICL)Steering Protocol=Pairwise, Model=Llama2-7b-chat-hf2026.03 | 81.86 | |
| Base(ICL)Steering Protocol=Positive, Model=Llama2-7b-chat-hf2026.03 | 81.86 | |
| ReFT(vector)Steering Protocol=Pairwise, Model=Llama2-7b-chat-hf2026.03 | 103.07 | |
| ReFT(mlp)Steering Protocol=Pairwise, Model=Llama2-7b-chat-hf2026.03 | 109.57 |