Mathematical Reasoning on MATH 500 (Accuracy@4 and Speedup)
78.2Accuracy (mean@4)DUET
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DUETBackbone=Qwen3-4B-base, Budget=Full2026.05 | 78.2 | 1.43 | |
| DUETBackbone=Qwen3-4B-base, Budget=50%2026.05 | 76.7 | 2.38 | |
| GRPOBackbone=Qwen3-4B-base2026.05 | 74.5 | 1 | |
| VIPBackbone=Qwen3-4B-base2026.05 | 74.4 | 0.75 | |
| ARRoLBackbone=Qwen3-4B-base2026.05 | 73.9 | 1.19 | |
| DAPOBackbone=Qwen3-4B-base2026.05 | 73.2 | 0.63 | |
| DUETBackbone=Qwen3-1.7B-base, Budget=50%2026.05 | 62.1 | 2.51 | |
| DUETBackbone=Qwen3-1.7B-base, Budget=Full2026.05 | 61.6 | 1.62 | |
| VIPBackbone=Qwen3-1.7B-base2026.05 | 58.6 | 0.95 | |
| ARRoLBackbone=Qwen3-1.7B-base2026.05 | 58.4 | 1.3 | |
| DAPOBackbone=Qwen3-1.7B-base2026.05 | 57.6 | 0.81 | |
| GRPOBackbone=Qwen3-1.7B-base2026.05 | 56.8 | 1 | |
| BaselineBackbone=Qwen3-4B-base2026.05 | 54.3 | — | |
| DUETBackbone=Llama-3.2-3B-Instruct, Budget=Full2026.05 | 45.6 | 1.26 | |
| ARRoLBackbone=Llama-3.2-3B-Instruct2026.05 | 45.3 | 1.13 | |
| BaselineBackbone=Qwen3-1.7B-base2026.05 | 45.1 | — | |
| DUETBackbone=Llama-3.2-3B-Instruct, Budget=50%2026.05 | 44.5 | 2.04 | |
| GRPOBackbone=Llama-3.2-3B-Instruct2026.05 | 44.4 | 1 | |
| DAPOBackbone=Llama-3.2-3B-Instruct2026.05 | 44.1 | 0.82 | |
| VIPBackbone=Llama-3.2-3B-Instruct2026.05 | 43.6 | 0.95 | |
| BaselineBackbone=Llama-3.2-3B-Instruct2026.05 | 37.1 | — |