Attention Head Selection (Causal Tracing) on WinoBias
0.0128ℓ(D, m) (10% Sparsity)random
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| randomModel=DistilGPT2, Selection Algorithm=random2026.06 | 0.0128 | 0.0146 | 0.0146 | 0.016 | 6.55 | |
| top-kModel=DistilGPT2, Selection Algorithm=top-k2026.06 | 0.0157 | 0.0171 | 0.0178 | 0.0181 | 0.47 | |
| greedyModel=DistilGPT2, Selection Algorithm=greedy2026.06 | 0.0159 | 0.0176 | 0.0186 | 0.0192 | 10.62 | |
| PGB-CTModel=DistilGPT2, Selection Algorithm=PGB-CT2026.06 | 0.0159 | 0.0174 | 0.0181 | 0.019 | 1.44 | |
| randomModel=Qwen3-1.7B, Selection Algorithm=random2026.06 | 0.064 | 0.073 | 0.077 | 0.085 | 75.38 | |
| randomModel=GPT2-small, Selection Algorithm=random2026.06 | 0.08 | 0.096 | 0.1 | 0.103 | 10.25 | |
| top-kModel=Qwen3-1.7B, Selection Algorithm=top-k2026.06 | 0.089 | 0.09 | 0.091 | 0.092 | 31.97 | |
| PGB-CTModel=Qwen3-1.7B, Selection Algorithm=PGB-CT2026.06 | 0.09 | 0.094 | 0.096 | 0.097 | 9.39 | |
| top-kModel=GPT2-small, Selection Algorithm=top-k2026.06 | 0.108 | 0.111 | 0.113 | 0.114 | 1.49 | |
| PGB-CTModel=GPT2-small, Selection Algorithm=PGB-CT2026.06 | 0.109 | 0.115 | 0.117 | 0.119 | 2.61 | |
| greedyModel=GPT2-small, Selection Algorithm=greedy2026.06 | 0.111 | 0.116 | 0.119 | 0.12 | 68.42 | |
| randomModel=GPT2-large, Selection Algorithm=random2026.06 | 0.164 | 0.216 | 0.262 | 0.297 | 36.8 | |
| randomModel=GPT2-medium, Selection Algorithm=random2026.06 | 0.19 | 0.247 | 0.305 | 0.336 | 21.45 | |
| randomModel=GPT2-xl, Selection Algorithm=random2026.06 | 0.221 | 0.393 | 0.442 | 0.46 | 64.98 | |
| top-kModel=Llama-3.2-1B, Selection Algorithm=top-k2026.06 | 0.279 | 0.358 | 0.386 | 0.415 | 26.57 | |
| randomModel=Llama-3.2-1B, Selection Algorithm=random2026.06 | 0.322 | 0.326 | 0.325 | 0.351 | 51.16 | |
| top-kModel=GPT2-large, Selection Algorithm=top-k2026.06 | 0.353 | 0.357 | 0.361 | 0.364 | 26.07 | |
| PGB-CTModel=GPT2-large, Selection Algorithm=PGB-CT2026.06 | 0.356 | 0.37 | 0.375 | 0.383 | 8.1 | |
| top-kModel=GPT2-medium, Selection Algorithm=top-k2026.06 | 0.374 | 0.378 | 0.389 | 0.388 | 8.18 | |
| PGB-CTModel=GPT2-medium, Selection Algorithm=PGB-CT2026.06 | 0.381 | 0.394 | 0.401 | 0.404 | 5.32 | |
| greedyModel=GPT2-medium, Selection Algorithm=greedy2026.06 | 0.391 | 0.406 | 0.415 | 0.42 | 1,001.5 | |
| top-kModel=GPT2-xl, Selection Algorithm=top-k2026.06 | 0.523 | 0.531 | 0.536 | 0.539 | 62.85 | |
| PGB-CTModel=GPT2-xl, Selection Algorithm=PGB-CT2026.06 | 0.546 | 0.563 | 0.567 | 0.576 | 11.32 | |
| PGB-CTModel=Llama-3.2-1B, Selection Algorithm=PGB-CT2026.06 | 0.752 | 1.007 | 1.766 | 2.747 | 5.82 |