Physical Commonsense Reasoning on PIQA (Accuracy, Drop, TC2b)
93.74AccuracyQwen3-32B (FP16)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-32B (FP16)Model=Qwen3-32B, Precision=FP162026.06 | 93.74 | — | — | |
| Qwen3-32B (+P + L)Model=Qwen3-32B, Precision=2-bit + FP16 plan + loop rescue2026.06 | 89.01 | -5.49 | 99.67 | |
| Qwen3-32B (+P)Model=Qwen3-32B, Precision=2-bit + FP16 plan2026.06 | 88.52 | -5.49 | 99.67 | |
| Qwen3-32B (+L)Model=Qwen3-32B, Precision=2-bit + loop rescue2026.06 | 88.47 | -5.49 | 99.67 | |
| Qwen3-8B (FP16)Model=Qwen3-8B, Precision=FP162026.06 | 88.36 | — | — | |
| Qwen3-32B (2bit)Model=Qwen3-32B, Precision=2-bit2026.06 | 88.25 | -5.49 | 99.67 | |
| DenseModel=LLaMA-2-7B, Sparsity=0%, Evaluation=Zero-shot2026.07 | 79.1 | — | — | |
| PALSModel=LLaMA-2-7B, Sparsity=50%, Evaluation=Zero-shot2026.07 | 73.8 | — | — | |
| Qwen3-8B (+P + L)Model=Qwen3-8B, Precision=2-bit + FP16 plan + loop rescue2026.06 | 72.6 | -20.73 | 53.1 | |
| WandaModel=LLaMA-2-7B, Sparsity=50%, Evaluation=Zero-shot2026.07 | 72.6 | — | — | |
| Qwen3-8B (+P)Model=Qwen3-8B, Precision=2-bit + FP16 plan2026.06 | 71.8 | -20.73 | 53.1 | |
| Qwen3-8B (+L)Model=Qwen3-8B, Precision=2-bit + loop rescue2026.06 | 67.79 | -20.73 | 53.1 | |
| Qwen3-8B (2bit)Model=Qwen3-8B, Precision=2-bit2026.06 | 67.63 | -20.73 | 53.1 | |
| MagnitudeModel=LLaMA-2-7B, Sparsity=50%, Evaluation=Zero-shot2026.07 | 63.7 | — | — |