List your company

Attention Head Selection (Causal Tracing) on WinoBias

0.0128ℓ(D, m) (10% Sparsity)random
-0.0167680.1828160.38240.581984Jun 2, 2026

Evaluation results

MethodLinks
randomModel=DistilGPT2, Selection Algorithm=random2026.060.01280.01460.01460.0166.55
top-kModel=DistilGPT2, Selection Algorithm=top-k2026.060.01570.01710.01780.01810.47
greedyModel=DistilGPT2, Selection Algorithm=greedy2026.060.01590.01760.01860.019210.62
PGB-CTModel=DistilGPT2, Selection Algorithm=PGB-CT2026.060.01590.01740.01810.0191.44
randomModel=Qwen3-1.7B, Selection Algorithm=random2026.060.0640.0730.0770.08575.38
randomModel=GPT2-small, Selection Algorithm=random2026.060.080.0960.10.10310.25
top-kModel=Qwen3-1.7B, Selection Algorithm=top-k2026.060.0890.090.0910.09231.97
PGB-CTModel=Qwen3-1.7B, Selection Algorithm=PGB-CT2026.060.090.0940.0960.0979.39
top-kModel=GPT2-small, Selection Algorithm=top-k2026.060.1080.1110.1130.1141.49
PGB-CTModel=GPT2-small, Selection Algorithm=PGB-CT2026.060.1090.1150.1170.1192.61
greedyModel=GPT2-small, Selection Algorithm=greedy2026.060.1110.1160.1190.1268.42
randomModel=GPT2-large, Selection Algorithm=random2026.060.1640.2160.2620.29736.8
randomModel=GPT2-medium, Selection Algorithm=random2026.060.190.2470.3050.33621.45
randomModel=GPT2-xl, Selection Algorithm=random2026.060.2210.3930.4420.4664.98
top-kModel=Llama-3.2-1B, Selection Algorithm=top-k2026.060.2790.3580.3860.41526.57
randomModel=Llama-3.2-1B, Selection Algorithm=random2026.060.3220.3260.3250.35151.16
top-kModel=GPT2-large, Selection Algorithm=top-k2026.060.3530.3570.3610.36426.07
PGB-CTModel=GPT2-large, Selection Algorithm=PGB-CT2026.060.3560.370.3750.3838.1
top-kModel=GPT2-medium, Selection Algorithm=top-k2026.060.3740.3780.3890.3888.18
PGB-CTModel=GPT2-medium, Selection Algorithm=PGB-CT2026.060.3810.3940.4010.4045.32
greedyModel=GPT2-medium, Selection Algorithm=greedy2026.060.3910.4060.4150.421,001.5
top-kModel=GPT2-xl, Selection Algorithm=top-k2026.060.5230.5310.5360.53962.85
PGB-CTModel=GPT2-xl, Selection Algorithm=PGB-CT2026.060.5460.5630.5670.57611.32
PGB-CTModel=Llama-3.2-1B, Selection Algorithm=PGB-CT2026.060.7521.0071.7662.7475.82