Science Question Answering on GPQA (Mean Accuracy)
24.6Mean AccuracyOPD
Evaluation Results
| Method | Links | |
|---|---|---|
| OPDTraining Budget=sufficient, Rollout=full, Zero-shot=true2026.02 | 24.6 | |
| OPD (prefix scheduling)Training Budget=sufficient, Zero-shot=true2026.02 | 23.6 | |
| SeqKDSteps=300, Training Budget=limited, Zero-shot=true2026.02 | 20.2 | |
| SeqKDTraining Budget=sufficient, Zero-shot=true2026.02 | 19.8 | |
| OPDSteps=10, Training Budget=limited, Zero-shot=true2026.02 | 18.3 | |
| Qwen3-1.7B-BaseZero-shot=true2026.02 | 11.2 | |
| OPDPrefix Length=2048, Training Budget=limited, Zero-shot=true2026.02 | 4.5 | |
| OPDPrefix Length=512, Training Budget=limited, Zero-shot=true2026.02 | 1.1 | |
| OPDPrefix Length=1024, Training Budget=limited, Zero-shot=true2026.02 | 1.1 | |
| OPDPrefix Length=256, Training Budget=limited, Zero-shot=true2026.02 | 1 |