Code Generation on MBPP OOD Generalization (trained on CodeContests)
85.45Pass@128Base (No-train)
Evaluation Results
| Method | Links | |
|---|---|---|
| Base (No-train)Backbone=Qwen2.5-32B2026.03 | 85.45 | |
| LEAFEBackbone=Qwen2.5-32B2026.03 | 85.45 | |
| LEAFEBackbone=Qwen2.5-72B2026.03 | 85.13 | |
| Base (No-train)Backbone=Qwen2.5-72B2026.03 | 83.33 | |
| GRPOBackbone=Qwen2.5-32B2026.03 | 81.22 | |
| GRPOBackbone=Qwen2.5-72B2026.03 | 81.22 | |
| LEAFEBackbone=Llama3-70B2026.03 | 79.63 | |
| Base (No-train)Backbone=Llama3-70B2026.03 | 78.31 | |
| GRPOBackbone=Llama3-70B2026.03 | 74.07 |