General-domain capability retention on AIME-2024, MATH-500, GSM8K, HumanEval, MMLU
62.9AIME-2024 AccuracyBase
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BaseModel=Qwen3-4B-It2026.05 | 62.9 | 94.2 | 92.6 | 86 | 77.6 | 82.6 | |
| MIXSDModel=Qwen3-4B-It, lambda=0.32026.05 | 53.1 | 92.2 | 91.9 | 87.2 | 71.6 | 79.2 | |
| MIXSDModel=Qwen3-4B-It, lambda=0.52026.05 | 52.1 | 93.4 | 92.4 | 84.8 | 73 | 79.1 | |
| MIXSDModel=Qwen3-4B-It, lambda=02026.05 | 43.8 | 91 | 92.8 | 87.2 | 72.9 | 77.5 | |
| OPSDModel=Qwen3-4B-It2026.05 | 43.3 | 91.2 | 92.3 | 88.4 | 73 | 77.7 | |
| OPSD-NLLModel=Qwen3-4B-It, n=1, T=02026.05 | 38.8 | 90 | 91.8 | 86.6 | 65.3 | 74.5 | |
| BaseModel=Qwen3-8B2026.05 | 26 | 83.4 | 91.8 | 86.6 | 73.1 | 72.2 | |
| MIXSDModel=Qwen3-8B, lambda=0.32026.05 | 21.2 | 78.2 | 90.3 | 82.3 | 60.7 | 66.6 | |
| MIXSDModel=Qwen3-8B, lambda=0.52026.05 | 20.4 | 79.2 | 91.3 | 84.1 | 57.8 | 66.6 | |
| OPSDModel=Qwen3-8B2026.05 | 15.6 | 76.6 | 90.8 | 86.6 | 56.8 | 65.3 | |
| MIXSDModel=Qwen3-8B, lambda=02026.05 | 14.4 | 74.2 | 88.6 | 84.1 | 43.3 | 60.9 | |
| OPSD-NLLModel=Qwen3-8B, n=1, T=02026.05 | 12.5 | 74.8 | 88.9 | 81.1 | 53.8 | 62.2 | |
| BaseModel=Qwen3-1.7B2026.05 | 11 | 72.4 | 80.4 | 60.4 | 58.5 | 56.5 | |
| MIXSDModel=Qwen3-1.7B, lambda=0.52026.05 | 5.4 | 57.6 | 75 | 53.7 | 36.3 | 45.6 | |
| SFTModel=Qwen3-8B2026.05 | 5.4 | 41.6 | 63.2 | 84.1 | 73.1 | 53.5 | |
| MIXSDModel=Qwen3-1.7B, lambda=0.32026.05 | 4.2 | 58.6 | 77.7 | 41.5 | 29.1 | 42.2 | |
| OPSDModel=Qwen3-1.7B2026.05 | 3.3 | 54.6 | 75.9 | 43.3 | 36.7 | 42.8 | |
| MIXSDModel=Qwen3-1.7B, lambda=02026.05 | 3.1 | 49 | 70.1 | 39.6 | 28.4 | 38 | |
| OPSD-NLLModel=Qwen3-1.7B, n=1, T=02026.05 | 1.9 | 41.6 | 64.8 | 28 | 33.9 | 34 | |
| SFTModel=Qwen3-1.7B2026.05 | 0 | 2.2 | 3.4 | 8.5 | 1.8 | 3.2 | |
| SFTModel=Qwen3-4B-It2026.05 | 0 | 2.2 | 3 | 50 | 28 | 16.6 |