Attention Head Selection (Causal Tracing) on WinoGender
0.0082Loss (10% Sparsity)random
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| randomModel=DistilGPT2, Selection Algorithm=random2026.06 | 0.0082 | 0.0093 | 0.0091 | 0.01 | 2.23 | |
| top-kModel=DistilGPT2, Selection Algorithm=top-k2026.06 | 0.01 | 0.0117 | 0.0123 | 0.0128 | 0.18 | |
| greedyModel=DistilGPT2, Selection Algorithm=greedy2026.06 | 0.0103 | 0.0122 | 0.0134 | 0.014 | 3.57 | |
| PGB-CTModel=DistilGPT2, Selection Algorithm=PGB-CT2026.06 | 0.0103 | 0.0118 | 0.013 | 0.0137 | 0.43 | |
| randomModel=GPT2-small, Selection Algorithm=random2026.06 | 0.027 | 0.032 | 0.036 | 0.037 | 3.54 | |
| randomModel=Qwen3-1.7B, Selection Algorithm=random2026.06 | 0.032 | 0.035 | 0.041 | 0.038 | 29.36 | |
| top-kModel=GPT2-small, Selection Algorithm=top-k2026.06 | 0.037 | 0.041 | 0.042 | 0.044 | 0.5 | |
| PGB-CTModel=GPT2-small, Selection Algorithm=PGB-CT2026.06 | 0.039 | 0.044 | 0.048 | 0.049 | 0.79 | |
| greedyModel=GPT2-small, Selection Algorithm=greedy2026.06 | 0.04 | 0.045 | 0.049 | 0.05 | 23.03 | |
| top-kModel=Qwen3-1.7B, Selection Algorithm=top-k2026.06 | 0.041 | 0.042 | 0.045 | 0.044 | 13.01 | |
| PGB-CTModel=Qwen3-1.7B, Selection Algorithm=PGB-CT2026.06 | 0.042 | 0.046 | 0.048 | 0.05 | 3.21 | |
| randomModel=GPT2-xl, Selection Algorithm=random2026.06 | 0.047 | 0.078 | 0.094 | 0.098 | 24.83 | |
| randomModel=GPT2-large, Selection Algorithm=random2026.06 | 0.085 | 0.099 | 0.111 | 0.128 | 13.43 | |
| randomModel=GPT2-medium, Selection Algorithm=random2026.06 | 0.138 | 0.153 | 0.165 | 0.173 | 7.23 | |
| top-kModel=GPT2-large, Selection Algorithm=top-k2026.06 | 0.146 | 0.159 | 0.165 | 0.168 | 9.46 | |
| PGB-CTModel=GPT2-large, Selection Algorithm=PGB-CT2026.06 | 0.152 | 0.171 | 0.184 | 0.192 | 2.43 | |
| top-kModel=GPT2-xl, Selection Algorithm=top-k2026.06 | 0.155 | 0.175 | 0.186 | 0.188 | 29.38 | |
| PGB-CTModel=GPT2-xl, Selection Algorithm=PGB-CT2026.06 | 0.159 | 0.19 | 0.205 | 0.212 | 3.45 | |
| top-kModel=GPT2-medium, Selection Algorithm=top-k2026.06 | 0.191 | 0.201 | 0.203 | 0.205 | 2.76 | |
| PGB-CTModel=GPT2-medium, Selection Algorithm=PGB-CT2026.06 | 0.203 | 0.218 | 0.227 | 0.233 | 1.56 | |
| greedyModel=GPT2-medium, Selection Algorithm=greedy2026.06 | 0.208 | 0.224 | 0.232 | 0.237 | 357.28 | |
| randomModel=Llama-3.2-1B, Selection Algorithm=random2026.06 | 0.326 | 0.383 | 0.387 | 0.418 | 20.71 | |
| PGB-CTModel=Llama-3.2-1B, Selection Algorithm=PGB-CT2026.06 | 0.366 | 0.444 | 0.477 | 2.449 | 1.9 | |
| top-kModel=Llama-3.2-1B, Selection Algorithm=top-k2026.06 | 0.45 | 0.416 | 0.4 | 0.366 | 10.31 |