SAE steering on Neutral prefixes dataset
0.35Concept Improvement (C)Dream-Mask
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Dream-MaskLayer=L23, Layer Category=Deep, L0=802026.02 | 0.35 | -0.35 | 0.24 | |
| Dream-UnmaskLayer=L27, Layer Category=Deep, L0=802026.02 | 0.33 | 0.18 | 0.38 | |
| Dream-MaskLayer=L27, Layer Category=Deep, L0=802026.02 | 0.29 | 0.01 | 0.29 | |
| Qwen-2.5-7BLayer=L23, Layer Category=Deep, L0=802026.02 | 0.28 | -0.43 | 0.15 | |
| Dream-UnmaskLayer=L23, Layer Category=Deep, L0=802026.02 | 0.26 | -0.33 | 0.16 | |
| LLaMA-3-8BLayer=L14, Layer Category=Middle, L0=802026.02 | 0.25 | -0.67 | 0.05 | |
| LLaDA-UnmaskLayer=L23, Layer Category=Deep, L0=802026.02 | 0.25 | -0.16 | 0.2 | |
| Qwen-2.5-7BLayer=L14, Layer Category=Middle, L0=802026.02 | 0.22 | -0.26 | 0.14 | |
| Dream-MaskLayer=L14, Layer Category=Middle, L0=802026.02 | 0.22 | -0.37 | 0.11 | |
| LLaDA-MaskLayer=L23, Layer Category=Deep, L0=802026.02 | 0.21 | -0.12 | 0.17 | |
| Dream-UnmaskLayer=L14, Layer Category=Middle, L0=802026.02 | 0.2 | -0.64 | 0.01 | |
| Dream-UnmaskLayer=L10, Layer Category=Middle, L0=802026.02 | 0.19 | -0.31 | 0.1 | |
| Qwen-2.5-7BLayer=L5, Layer Category=Shallow, L0=802026.02 | 0.18 | -0.34 | 0.08 | |
| Qwen-2.5-7BLayer=L10, Layer Category=Middle, L0=802026.02 | 0.18 | -0.35 | 0.08 | |
| LLaMA-3-8BLayer=L5, Layer Category=Shallow, L0=802026.02 | 0.18 | -2.14 | -0.46 | |
| Dream-UnmaskLayer=L5, Layer Category=Shallow, L0=802026.02 | 0.17 | -0.17 | 0.12 | |
| Dream-MaskLayer=L10, Layer Category=Middle, L0=802026.02 | 0.17 | -0.12 | 0.13 | |
| LLaMA-3-8BLayer=L23, Layer Category=Deep, L0=802026.02 | 0.17 | -0.44 | 0.04 | |
| Qwen-2.5-7BLayer=L1, Layer Category=Shallow, L0=802026.02 | 0.16 | -0.39 | 0.04 | |
| LLaDA-UnmaskLayer=L5, Layer Category=Shallow, L0=802026.02 | 0.16 | -0.13 | 0.13 | |
| LLaDA-UnmaskLayer=L27, Layer Category=Deep, L0=802026.02 | 0.16 | -0.26 | 0.08 | |
| LLaDA-MaskLayer=L14, Layer Category=Middle, L0=802026.02 | 0.16 | -0.06 | 0.14 | |
| LLaMA-3-8BLayer=L10, Layer Category=Middle, L0=802026.02 | 0.15 | -1.23 | -0.22 | |
| LLaDA-MaskLayer=L27, Layer Category=Deep, L0=802026.02 | 0.15 | -0.02 | 0.15 | |
| Dream-MaskLayer=L1, Layer Category=Shallow, L0=802026.02 | 0.14 | -0.11 | 0.11 | |
| LLaDA-UnmaskLayer=L1, Layer Category=Shallow, L0=802026.02 | 0.14 | -0.05 | 0.13 | |
| LLaDA-MaskLayer=L10, Layer Category=Middle, L0=802026.02 | 0.14 | 0.04 | 0.15 | |
| Dream-UnmaskLayer=L1, Layer Category=Shallow, L0=802026.02 | 0.13 | -0.08 | 0.1 | |
| Dream-MaskLayer=L5, Layer Category=Shallow, L0=802026.02 | 0.13 | -0.09 | 0.11 | |
| LLaMA-3-8BLayer=L27, Layer Category=Deep, L0=802026.02 | 0.13 | -0.09 | 0.1 | |
| LLaDA-UnmaskLayer=L10, Layer Category=Middle, L0=802026.02 | 0.13 | 0.02 | 0.13 | |
| LLaDA-UnmaskLayer=L14, Layer Category=Middle, L0=802026.02 | 0.13 | -0.04 | 0.11 | |
| LLaDA-MaskLayer=L1, Layer Category=Shallow, L0=802026.02 | 0.13 | 0 | 0.13 | |
| LLaDA-MaskLayer=L5, Layer Category=Shallow, L0=802026.02 | 0.11 | 0 | 0.11 | |
| LLaMA-3-8BLayer=L1, Layer Category=Shallow, L0=802026.02 | 0.08 | -7.89 | -2.29 | |
| Qwen-2.5-7BLayer=L27, Layer Category=Deep, L0=802026.02 | 0.03 | -0.19 | -0.02 |