Physical Commonsense Reasoning on PIQA (val test)
79.42AccuracyLLaDA (Base)
Evaluation results
| Method | Links | |
|---|---|---|
| LLaDA (Base)Pruning Ratio=0.02026.02 | 79.42 | |
| Sink-Aware (Ours)Pruning Ratio=0.32026.02 | 69.55 | |
| LLaDA-structurePruning Ratio=0.32026.02 | 68.34 | |
| Sink-Aware (Ours)Pruning Ratio=0.52026.02 | 60.37 | |
| LLaDA-structurePruning Ratio=0.52026.02 | 58.98 | |
| GPT2-201Mmode=Zero-shot, parameters=201M2026.05 | 54.9 | |
| SNN-194Mmode=Zero-shot, parameters=194M2026.05 | 53.8 | |
| Randommode=Zero-shot2026.05 | 50 |