Physical Commonsense Reasoning on PIQA (Accuracy and Supernodes Pruned)
84AccuracyUnpruned
Evaluation results
| Method | Links | ||
|---|---|---|---|
| UnprunedBackbone=Llama-3.1-8B2026.04 | 84 | 0 | |
| SCAR-LPBackbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 64 | 0 | |
| SCAR-ConnBackbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 64 | 0 | |
| SCAR-ProtBackbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 63 | 0 | |
| Act. L2Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 61 | 19.6 | |
| Wanda (channel)Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 56 | 24.3 | |
| SparseGPT (channel)Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 56 | 34.6 | |
| Magnitude (channel)Backbone=Llama-3.1-8B, Sparsity=50%, Sparsity Target=FFN channel2026.04 | 51 | 41.7 |