Mathematical Reasoning on GSM8K (Speedup & Tau)
6.02SpeedupDFLARE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DFLAREModel=Q3-8B, Temperature=02026.06 | 6.02 | 7.88 | — | |
| DFLAREModel=Q3-4B, Temperature=02026.06 | 6 | 7.93 | — | |
| DFLAREModel=Q3-4B, Temperature=12026.06 | 5.38 | 7.14 | — | |
| DFLAREModel=Q3-8B, Temperature=12026.06 | 5.34 | 7.07 | — | |
| DFlashModel=Q3-4B, Temperature=02026.06 | 5.31 | 6.51 | — | |
| DFlashModel=Q3-8B, Temperature=02026.06 | 5.28 | 6.53 | — | |
| DFlashModel=Q3-8B, Temperature=12026.06 | 4.83 | 5.96 | — | |
| DFlashModel=Q3-4B, Temperature=12026.06 | 4.8 | 6 | — | |
| DFLAREModel=GPT-20B, Temperature=02026.06 | 4.06 | 4.98 | — | |
| DFlashModel=GPT-20B, Temperature=02026.06 | 3.75 | 4.54 | — | |
| DFLAREModel=GPT-20B, Temperature=12026.06 | 3.07 | 3.83 | — | |
| DFlashModel=GPT-20B, Temperature=12026.06 | 2.95 | 3.64 | — | |
| EAGLE3Model=Q3-4B, Temperature=02026.06 | 2.56 | 3.76 | — | |
| EAGLE3Model=Q3-8B, Temperature=02026.06 | 2.43 | 3.53 | — | |
| EAGLE3Model=Q3-4B, Temperature=12026.06 | 2.37 | 3.71 | — | |
| EAGLE3Model=Q3-8B, Temperature=12026.06 | 2.23 | 3.44 | — | |
| Llama 3.2-3Bγ (loss-decay constant)=None, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 2.875 | |
| Llama 3.2-3Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 3.04 | |
| Llama 3.2-3Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=02026.06 | — | — | 3.11 | |
| Llama 3.2-3Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=402026.06 | — | — | 3.507 | |
| Llama-3-8Bγ (loss-decay constant)=None, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 2.279 | |
| Llama-3-8Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 2.382 | |
| Llama-3-8Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=02026.06 | — | — | 2.753 | |
| Llama-3-8Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=402026.06 | — | — | 3.074 | |
| Qwen3-4Bγ (loss-decay constant)=None, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 2.397 | |
| Qwen3-4Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 2.502 | |
| Qwen3-4Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=02026.06 | — | — | 2.598 | |
| Qwen3-4Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=402026.06 | — | — | 2.833 | |
| Qwen3-8Bγ (loss-decay constant)=None, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 2.248 | |
| Qwen3-8Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0, αc (chain-loss coefficient)=02026.06 | — | — | 2.517 | |
| Qwen3-8Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=02026.06 | — | — | 2.608 | |
| Qwen3-8Bγ (loss-decay constant)=10, αf (first-error focal coefficient)=0.3, αc (chain-loss coefficient)=402026.06 | — | — | 2.795 |