Code generation on LiveCodeBench (pass@1, pass@16)
19.6pass@1GSPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GSPOBackbone=Qwen3-4B-Base2026.05 | 19.6 | 31.9 | |
| DISABackbone=Qwen3-4B-Base2026.05 | 19.3 | 30.8 | |
| GRPOBackbone=Qwen3-4B-Base2026.05 | 19.1 | 30.4 | |
| FlowRLBackbone=Qwen3-4B-Base2026.05 | 18.9 | 28.9 | |
| GRPOBackbone=Qwen2.5-7B2026.05 | 16.85 | 23.4 | |
| DISABackbone=Qwen2.5-7B2026.05 | 15.41 | 25.4 | |
| FlowRLBackbone=Qwen2.5-7B2026.05 | 14.7 | 24.5 | |
| GSPOBackbone=Qwen2.5-7B2026.05 | 14.4 | 26.2 | |
| VanillaBackbone=Qwen3-4B-Base2026.05 | 7.48 | 24 | |
| VanillaBackbone=Qwen2.5-7B2026.05 | 5.6 | 21.5 |