Multiple Choice Question Answering on GPQA, ARC, MMLU (OOD)
42GPQA AccuracyOverSFT+RL
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| OverSFT+RLBackbone=Qwen3-8B2026.06 | 42 | 75.9 | 61.7 | 59.9 | 52.5 | |
| Rejuvenation+RLBackbone=Qwen3-8B2026.06 | 41 | 78.2 | 62.4 | 60.5 | 53.9 | |
| ModSFT+RLBackbone=Qwen3-8B2026.06 | 40.5 | 77.9 | 62.5 | 60.3 | 53.8 | |
| DFT+RLBackbone=Qwen3-8B2026.06 | 38.8 | 76.8 | 62.5 | 59.3 | 51.6 |