Code Generation on HumanEval (Accuracy, Avg.)
81.71AccuracyQwen3 8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3 8BGen. Mode=AR, TPF=1.002026.07 | 81.71 | — | |
| Nemotron-Labs-Diffusion-8BGen. Mode=Linear SS, TPF=5.992026.07 | 81.71 | — | |
| Nemotron-Labs-Diffusion-8BGen. Mode=AR, TPF=1.002026.07 | 80.49 | — | |
| SDAR-8B ChatGen. Mode=Diff., TPF=1.752026.07 | 79.27 | — | |
| Nemotron-Labs-Diffusion-8BGen. Mode=Quad. SS, TPF=6.382026.07 | 79.27 | — | |
| SDAR-8B ChatGen. Mode=Diff., TPF=1.002026.07 | 78.7 | — | |
| Nemotron-Labs-Diffusion-8BGen. Mode=Diff., TPF=2.572026.07 | 78.66 | — | |
| Qwen2.5 7BGen. Mode=AR, TPF=1.002026.07 | 77.44 | — | |
| Ministral3-8B Instruct-2512Gen. Mode=AR, TPF=1.002026.07 | 71.04 | — | |
| Dream-7B InstructGen. Mode=Diff., TPF=1.002026.07 | 55.5 | — | |
| LLaDA-8B InstructGen. Mode=Diff., TPF=1.002026.07 | 49.4 | — | |
| Skew F(R)KLTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 42.07 | 47.29 | |
| α-β divergenceTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 41.46 | 47.09 | |
| HybKDTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 41.46 | 47.89 | |
| Reverse KLTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 39.02 | 46.34 | |
| Forward KLTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 38.41 | 46.76 | |
| Hard KDTeacher-Student=DeepSeek-Coder-6.7B -> 1.3B2026.05 | 35.37 | 44.66 |