Multi-component causal tracing on CounterFact Factual
21,155.9Tracing Fidelity (ℓ(D, m))PGB-CT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PGB-CTModel=GPT2-small, Sparsity=≤15%2026.06 | 21,155.9 | 0.13 | |
| PGB-CTModel=GPT2-small, Sparsity=≤10%2026.06 | 8,376.24 | 0.13 | |
| PGB-CTModel=GPT2-small, Sparsity=≤5%2026.06 | 1,790.33 | 0.13 | |
| PGB-CTModel=DistilGPT2, Sparsity=≤15%2026.06 | 1,141.76 | 0.07 | |
| PGB-CTModel=GPT2-small, Sparsity=≤2.5%2026.06 | 755.95 | 0.13 | |
| PGB-CTModel=DistilGPT2, Sparsity=≤10%2026.06 | 286.34 | 0.07 | |
| randomModel=GPT2-small, Sparsity=≤10%2026.06 | 239.06 | 3.99 | |
| top-kModel=GPT2-small, Sparsity=≤15%2026.06 | 197.02 | 32.35 | |
| randomModel=GPT2-small, Sparsity=≤15%2026.06 | 122.11 | 3.99 | |
| randomModel=DistilGPT2, Sparsity=≤10%2026.06 | 111.89 | 1.98 | |
| top-kModel=GPT2-small, Sparsity=≤10%2026.06 | 107.6 | 32.35 | |
| top-kModel=GPT2-small, Sparsity=≤5%2026.06 | 66.02 | 32.35 | |
| PGB-CTModel=DistilGPT2, Sparsity=≤5%2026.06 | 62.22 | 0.07 | |
| randomModel=DistilGPT2, Sparsity=≤15%2026.06 | 37.67 | 1.98 | |
| top-kModel=GPT2-small, Sparsity=≤2.5%2026.06 | 27.49 | 32.35 | |
| top-kModel=DistilGPT2, Sparsity=≤10%2026.06 | 24.99 | 10.08 | |
| top-kModel=DistilGPT2, Sparsity=≤5%2026.06 | 24.77 | 10.08 | |
| top-kModel=DistilGPT2, Sparsity=≤15%2026.06 | 22.45 | 10.08 | |
| randomModel=DistilGPT2, Sparsity=≤5%2026.06 | 18.07 | 1.98 | |
| PGB-CTModel=DistilGPT2, Sparsity=≤2.5%2026.06 | 15.63 | 0.07 | |
| randomModel=GPT2-small, Sparsity=≤5%2026.06 | 13.75 | 3.99 | |
| top-kModel=DistilGPT2, Sparsity=≤2.5%2026.06 | 5.72 | 10.08 | |
| randomModel=GPT2-small, Sparsity=≤2.5%2026.06 | 3.88 | 3.99 | |
| randomModel=DistilGPT2, Sparsity=≤2.5%2026.06 | 2.35 | 1.98 |