Scientific Question Answering on GPQA D (Accuracy (mean@4, %) & Speedup)
38.5Accuracy (mean@4)GRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=Qwen3-4B-base2026.05 | 38.5 | 1 | |
| DUETBackbone=Qwen3-4B-base, Budget=Full2026.05 | 38.3 | 1.43 | |
| DUETBackbone=Qwen3-4B-base, Budget=50%2026.05 | 38.1 | 2.38 | |
| DAPOBackbone=Qwen3-4B-base2026.05 | 37.4 | 0.63 | |
| VIPBackbone=Qwen3-4B-base2026.05 | 36.8 | 0.75 | |
| ARRoLBackbone=Qwen3-4B-base2026.05 | 34.5 | 1.19 | |
| DAPOBackbone=Qwen3-1.7B-base2026.05 | 29.1 | 0.81 | |
| DUETBackbone=Qwen3-1.7B-base, Budget=50%2026.05 | 29.1 | 2.51 | |
| DUETBackbone=Qwen3-1.7B-base, Budget=Full2026.05 | 29 | 1.62 | |
| BaselineBackbone=Qwen3-4B-base2026.05 | 28.8 | — | |
| VIPBackbone=Qwen3-1.7B-base2026.05 | 27.8 | 0.95 | |
| GRPOBackbone=Qwen3-1.7B-base2026.05 | 27.4 | 1 | |
| ARRoLBackbone=Qwen3-1.7B-base2026.05 | 26.7 | 1.3 | |
| GRPOBackbone=Llama-3.2-3B-Instruct2026.05 | 25.4 | 1 | |
| ARRoLBackbone=Llama-3.2-3B-Instruct2026.05 | 25.3 | 1.13 | |
| DUETBackbone=Llama-3.2-3B-Instruct, Budget=Full2026.05 | 25.3 | 1.26 | |
| VIPBackbone=Llama-3.2-3B-Instruct2026.05 | 24.9 | 0.95 | |
| DAPOBackbone=Llama-3.2-3B-Instruct2026.05 | 24.5 | 0.82 | |
| DUETBackbone=Llama-3.2-3B-Instruct, Budget=50%2026.05 | 23.2 | 2.04 | |
| BaselineBackbone=Qwen3-1.7B-base2026.05 | 22.2 | — | |
| BaselineBackbone=Llama-3.2-3B-Instruct2026.05 | 20.5 | — |