Code Generation on LCB v6 (fixed 500-problem slice)
53.5Pass@4CPPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CPPOBackbone=Qwen3.5-4B, K=42026.05 | 53.5 | 1,556 | 3.438 | |
| PKPOBackbone=Qwen3.5-4B, K=42026.05 | 48.8 | 2,137 | 2.284 | |
| UpSkillBackbone=Qwen3.5-4B, K=42026.05 | 41.1 | 2,218 | 1.853 | |
| Plan-and-SolveBackbone=Qwen3.5-4B, K=42026.05 | 25.5 | 3,975 | 0.642 | |
| PlanSearchBackbone=Qwen3.5-4B, K=42026.05 | 23.6 | 2,137 | 1.104 | |
| Direct SolveBackbone=Qwen3.5-4B, K=42026.05 | 21.4 | 1,731 | 1.236 |