Code Generation on MBPP+ (Accuracy, Avg.)
51.59AccuracyForward KL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Forward KLTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 51.59 | 46.76 | |
| HybKDTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 50.39 | 47.89 | |
| Skew F(R)KLTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 50.26 | 47.29 | |
| α-β divergenceTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 50.26 | 47.09 | |
| Hard KDTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 50 | 44.66 | |
| Reverse KLTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 50 | 46.34 |