Mathematical Reasoning on GSM8K (Accuracy by Sample Count)
94.39Task AccuracyDynaSteer
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DynaSteerBackbone=Qwen3-8B2026.06 | 94.39 | — | — | — | — | — | |
| ROMEBackbone=Qwen3-8B2026.06 | 77.94 | — | — | — | — | — | |
| MEMITBackbone=Qwen3-8B2026.06 | 76.95 | — | — | — | — | — | |
| DELBackbone=Qwen2.5-1.5B2026.05 | 57 | 70 | 67 | 67.2 | 64.9 | 61.7 | |
| NTLBackbone=Qwen2.5-1.5B2026.05 | 55.1 | 68.9 | 65.1 | 64.6 | 64.9 | 61.7 | |
| EMOBackbone=Qwen2.5-1.5B2026.05 | 54.5 | 69.3 | 64.1 | 66.3 | 66.4 | 59.6 | |
| DIST2LossBackbone=Qwen2.5-1.5B2026.05 | 54.2 | 67.9 | 63.5 | 64.6 | 61.1 | 53.2 | |
| MLEBackbone=Qwen2.5-1.5B, Loss Function=Maximum Likelihood Estimation2026.05 | 54 | 66.8 | 62.4 | 63.7 | 66.4 | 61.7 | |
| MixCEBackbone=Qwen2.5-1.5B2026.05 | 52.8 | 66.6 | 62.5 | 65.6 | 65.6 | 57.4 |