Code Generation on LiveCodeBench v6 Aug 2024-May 2025 (test)
63.1Pass@1FADE h* = 1.0
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FADE h* = 1.0Backbone=CWM 32B, Budget=30k2026.07 | 63.1 | 77.9 | 85.3 | 86.3 | |
| Power α asym (2,2)Backbone=CWM 32B, Budget=30k2026.07 | 62.8 | 76.2 | 83.3 | 84.6 | |
| GRPOBackbone=CWM 32B, Budget=30k2026.07 | 62 | 76.7 | 83.7 | 85 | |
| Power α = 2Backbone=CWM 32B, Budget=30k2026.07 | 61.2 | 76.7 | 82.9 | 83.9 | |
| FADE h* = 1.0Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 40.7 | 58 | 68.2 | — | |
| Power α = 2Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 40.3 | 56.6 | 66 | — | |
| Power α asym (2,2)Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 40 | 56 | 64.9 | — | |
| Implicit p@2Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 39.7 | 56.2 | 65.2 | — | |
| Asym. normBackbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 38.1 | 53.4 | 63 | — | |
| Implicit p@8Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 38 | 54.3 | 64.7 | — | |
| SFT modelBackbone=CWM 32B, Budget=30k2026.07 | 37.5 | 54 | 69 | 72.5 | |
| GRPOBackbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 37.4 | 53.6 | 62.5 | — | |
| LME β = 0.4Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 35.9 | 52.5 | 62 | — | |
| Asym δ = 0.5Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 35.7 | 51.8 | 60.5 | — | |
| Pass@8Backbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 33.8 | 51.7 | 61.1 | — | |
| REINFORCEBackbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 29.4 | 45.3 | 55.8 | — | |
| SFT modelBackbone=Qwen 2.5 7B SFT, Budget=8k2026.07 | 15.2 | 42.3 | 52.2 | — |