Code Generation on LiveCodeBench v5 (p@1, p@10, p@100, p@200)
78.9Pass@1FADE h* = 1.0
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FADE h* = 1.0Backbone=CWM 32B, Token Budget=30k, Steps=15k2026.07 | 78.9 | 89.3 | 92.5 | 93.2 | |
| Power α asym (2,2)Backbone=CWM 32B, Token Budget=30k, Steps=15k2026.07 | 78.7 | 88.9 | 92 | 92.6 | |
| Power α = 2Backbone=CWM 32B, Token Budget=30k, Steps=17.5k2026.07 | 77.9 | 89.4 | 92.3 | 92.7 | |
| GRPOBackbone=CWM 32B, Token Budget=30k, Steps=13.4k2026.07 | 77.3 | 89 | 92 | 92.6 | |
| Power α = 2Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=50k2026.07 | 56.3 | 73.8 | 82.1 | — | |
| Power α asym (2,2)Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=20k2026.07 | 56.1 | 73.6 | 81.5 | — | |
| FADE h* = 1.0Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=30k2026.07 | 56.1 | 73.6 | 82 | — | |
| Implicit p@8Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=30k2026.07 | 55.7 | 72.5 | 80.9 | — | |
| Implicit p@2Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=30k2026.07 | 55.3 | 73.3 | 80.9 | — | |
| Asym. normBackbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=20k2026.07 | 54.1 | 70.9 | 79.2 | — | |
| Asym δ = 0.5Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=40k2026.07 | 53.4 | 71.1 | 79.6 | — | |
| GRPOBackbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=20k2026.07 | 52.6 | 71.8 | 80.6 | — | |
| SFT modelBackbone=CWM 32B, Token Budget=30k, Steps=02026.07 | 51.6 | 75.6 | 85.1 | 86.9 | |
| LME β = 0.4Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=20k2026.07 | 51 | 70.8 | 79.9 | — | |
| Pass@8Backbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=30k2026.07 | 48.3 | 70.5 | 80 | — | |
| REINFORCEBackbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=17.5k2026.07 | 40.7 | 60.2 | 69.8 | — | |
| SFT modelBackbone=Qwen 2.5 7B SFT, Token Budget=8k, Steps=02026.07 | 21.9 | 59.7 | 71.2 | — |