Multi-hop Question Answering on HotpotQA in-distribution
66.51Exact Match (EM)CRAFT7B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CRAFT7BType=Trained, CRAFT version=v4, Training Mode=RL2026.02 | 66.51 | 81.91 | 96 | |
| Gemini-2.5-FlashType=API, CRAFT version=v42026.02 | 63.75 | 79.65 | 97.2 | |
| Qwen3-30BType=Pre-Trained, CRAFT version=v42026.02 | 63.35 | 78.13 | 97.1 | |
| Deepseek-V3.2Type=API, CRAFT version=v42026.02 | 61.55 | 78.67 | 98.25 | |
| GPT-5-miniType=API, CRAFT version=v42026.02 | 60.85 | 79.83 | 98.05 | |
| Qwen3-4BType=Pre-Trained, CRAFT version=v42026.02 | 60.7 | 76.06 | 95.6 | |
| Qwen2.5-7BType=Pre-Trained, CRAFT version=v42026.02 | 59.3 | 74.4 | 93.75 | |
| Qwen2.5-7BType=Trained, CRAFT version=v4, Training Mode=SFT2026.02 | 57.9 | 72.43 | 89.35 | |
| CRAFT3BType=Trained, CRAFT version=v4, Training Mode=RL2026.02 | 56.11 | 70.63 | 94.3 | |
| Qwen2.5-3BType=Trained, CRAFT version=v4, Training Mode=SFT2026.02 | 50.25 | 64.37 | 83.85 | |
| Qwen2.5-3BType=Pre-Trained, CRAFT version=v42026.02 | 48.3 | 61.44 | 76.05 | |
| CRAFT1.5BType=Trained, CRAFT version=v4, Training Mode=RL2026.02 | 43.68 | 57.79 | 65.66 | |
| Qwen2.5-1.5BType=Trained, CRAFT version=v4, Training Mode=SFT2026.02 | 33.25 | 44.48 | 51.95 | |
| Qwen2.5-1.5BType=Pre-Trained, CRAFT version=v42026.02 | 26.8 | 36.98 | 44.65 | |
| CRAFT0.5BType=Trained, CRAFT version=v4, Training Mode=RL2026.02 | 25.95 | 36.2 | 40.7 | |
| Qwen2.5-0.5BType=Trained, CRAFT version=v4, Training Mode=SFT2026.02 | 11.75 | 21.27 | 37.4 | |
| Qwen2.5-0.5BType=Pre-Trained, CRAFT version=v42026.02 | 3 | 7.51 | 14.1 |