Code Generation on MBPP+ (Pass@1, Avg, ΔScore)
76.42Pass@1Individual
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| IndividualModel Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.06 | 76.42 | — | — | |
| Weight Averaging w/ HARCMerging Strategy=Weight Averaging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.06 | 76.36 | — | — | |
| WUDI-Merging w/ HARCMerging Strategy=WUDI-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.06 | 76.12 | — | — | |
| TIES-Merging w/ HARCMerging Strategy=TIES-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.06 | 76.05 | — | — | |
| Weight AveragingMerging Strategy=Weight Averaging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.06 | 76.03 | — | — | |
| Fisher MergingMerging Strategy=Fisher Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.06 | 75.94 | — | — | |
| WUDI-MergingMerging Strategy=WUDI-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.06 | 75.61 | — | — | |
| Fisher Merging w/ HARCMerging Strategy=Fisher Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.06 | 75.38 | — | — | |
| TIES-MergingMerging Strategy=TIES-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.06 | 75.23 | — | — | |
| DARE w/ HARCMerging Strategy=DARE, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.06 | 75.18 | — | — | |
| DAREMerging Strategy=DARE, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.06 | 74.19 | — | — | |
| R1-OracleBackbone=Qwen2.5-3B-Instruct, Supervision Paradigm=100% GT2026.03 | 73.75 | 79.12 | 0 | |
| Uni-OPDDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.05 | 72.6 | — | — | |
| ExPODistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.05 | 72 | — | — | |
| Uni-OPDDistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.05 | 71.6 | — | — | |
| OPDDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.05 | 70.9 | — | — | |
| ExOPDDistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.05 | 70.7 | — | — | |
| ExOPDDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.05 | 70.6 | — | — | |
| ExPODistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.05 | 70.2 | — | — | |
| TeacherTraining Strategy=RL2026.05 | 69.8 | — | — | |
| SFTDistillation Scenario=Multi-Teacher Distillation, Model Scale=4B2026.05 | 69.6 | — | — | |
| OPDDistillation Scenario=Single-Teacher Distillation, Model Scale=4B2026.05 | 68.8 | — | — | |
| R-ZeroBackbone=Qwen3-4B2026.05 | 68.3 | — | — | |
| BaseBackbone=Qwen3-4B2026.05 | 67.72 | — | — | |
| AZRBackbone=Qwen3-4B2026.05 | 67.5 | — | — | |
| COSEBackbone=Qwen3-4B2026.05 | 67.5 | — | — | |
| Teacher2026.05 | 66.75 | — | — | |
| StudentModel Scale=4B2026.05 | 65.3 | — | — | |
| R1-pBackbone=Qwen2.5-3B-Instruct, Supervision Paradigm=20% GT2026.03 | 65 | 75.67 | -3.45 | |
| MemRewardBackbone=Qwen2.5-3B-Instruct, Supervision Paradigm=20% GT + GNN2026.03 | 63.75 | 77.02 | -2.1 | |
| R1-OracleBackbone=Qwen2.5-1.5B-Instruct, Supervision Paradigm=100% GT2026.03 | 56.25 | 70.47 | 0 | |
| R1-pBackbone=Qwen2.5-1.5B-Instruct, Supervision Paradigm=20% GT2026.03 | 55 | 62.72 | -7.75 | |
| MemRewardBackbone=Qwen2.5-1.5B-Instruct, Supervision Paradigm=20% GT + GNN2026.03 | 55 | 68.1 | -2.37 | |
| AZRBackbone=Llama-3.2-3B-Instruct2026.05 | 54.5 | — | — | |
| Reverse KL + Ours2026.05 | 51.95 | — | — | |
| α-β divergence + Ours2026.05 | 51.95 | — | — | |
| COSEBackbone=Llama-3.2-3B-Instruct2026.05 | 51.9 | — | — | |
| Forward KL2026.05 | 51.59 | — | — | |
| Skew RKL + Ours2026.05 | 51.17 | — | — | |
| R-ZeroBackbone=Llama-3.2-3B-Instruct2026.05 | 50.5 | — | — | |
| Forward KL + Ours2026.05 | 50.39 | — | — | |
| Skew FKL + Ours2026.05 | 50.39 | — | — | |
| Skew RKL2026.05 | 50.39 | — | — | |
| Skew FKL2026.05 | 50.26 | — | — | |
| α-β divergence2026.05 | 50.26 | — | — | |
| Hard KD2026.05 | 50 | — | — | |
| Reverse KL2026.05 | 50 | — | — | |
| Total Variation + Ours2026.05 | 49.97 | — | — | |
| Student (No Distill)2026.05 | 49.61 | — | — | |
| Total Variation2026.05 | 49.61 | — | — | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.05 | 49.47 | — | — |