Code Generation on HumanEval (Pass@4 Accuracy and GRPO Speedup)
77Accuracy (Pass@4)DUET
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DUETBackbone=Qwen3-4B-base, Budget=50%2026.05 | 77 | 2.38 | |
| DUETBackbone=Qwen3-4B-base, Budget=Full2026.05 | 77 | 1.43 | |
| DAPOBackbone=Qwen3-4B-base2026.05 | 75.2 | 0.63 | |
| VIPBackbone=Qwen3-4B-base2026.05 | 75.2 | 0.75 | |
| GRPOBackbone=Qwen3-4B-base2026.05 | 73.6 | 1 | |
| ARRoLBackbone=Qwen3-4B-base2026.05 | 73.6 | 1.19 | |
| BaselineBackbone=Qwen3-4B-base2026.05 | 62.8 | — | |
| VIPBackbone=Llama-3.2-3B-Instruct2026.05 | 53 | 0.95 | |
| DUETBackbone=Llama-3.2-3B-Instruct, Budget=50%2026.05 | 52.9 | 2.04 | |
| GRPOBackbone=Llama-3.2-3B-Instruct2026.05 | 52.8 | 1 | |
| ARRoLBackbone=Llama-3.2-3B-Instruct2026.05 | 52.8 | 1.13 | |
| DUETBackbone=Qwen3-1.7B-base, Budget=Full2026.05 | 52.7 | 1.62 | |
| DAPOBackbone=Llama-3.2-3B-Instruct2026.05 | 52.7 | 0.82 | |
| DUETBackbone=Llama-3.2-3B-Instruct, Budget=Full2026.05 | 52.1 | 1.26 | |
| DUETBackbone=Qwen3-1.7B-base, Budget=50%2026.05 | 51.7 | 2.51 | |
| BaselineBackbone=Llama-3.2-3B-Instruct2026.05 | 50.5 | — | |
| VIPBackbone=Qwen3-1.7B-base2026.05 | 48.9 | 0.95 | |
| ARRoLBackbone=Qwen3-1.7B-base2026.05 | 47.7 | 1.3 | |
| GRPOBackbone=Qwen3-1.7B-base2026.05 | 47.6 | 1 | |
| DAPOBackbone=Qwen3-1.7B-base2026.05 | 46.8 | 0.81 | |
| BaselineBackbone=Qwen3-1.7B-base2026.05 | 34.5 | — |