Code generation on CodeForces (pass@1, pass@16)
11Pass@1GRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=Qwen3-4B-Base2026.05 | 11 | 30.1 | |
| FlowRLBackbone=Qwen3-4B-Base2026.05 | 10.9 | 28.6 | |
| DISABackbone=Qwen3-4B-Base2026.05 | 10.84 | 31.2 | |
| FlowRLBackbone=Qwen2.5-7B2026.05 | 10.78 | 21.5 | |
| GSPOBackbone=Qwen3-4B-Base2026.05 | 9.9 | 27.5 | |
| GRPOBackbone=Qwen2.5-7B2026.05 | 9.07 | 20.8 | |
| DISABackbone=Qwen2.5-7B2026.05 | 8.82 | 22.8 | |
| GSPOBackbone=Qwen2.5-7B2026.05 | 7.02 | 23 | |
| VanillaBackbone=Qwen3-4B-Base2026.05 | 2.31 | 17.2 | |
| VanillaBackbone=Qwen2.5-7B2026.05 | 1.38 | 11.8 |