Multi-Hop Search-Augmented QA on Bamboogle out-of-domain
72AccuracyGEPO
Evaluation Results
| Method | Links | |
|---|---|---|
| GEPOType=RL Training, Model=Qwen2.5-7B-Instruct2025.10 | 72 | |
| GiGPOType=RL Training, Model=Qwen2.5-7B-Instruct2025.10 | 68.9 | |
| GEPOType=RL Training, Model=Qwen2.5-3B-Instruct2025.10 | 68.2 | |
| GiGPOType=RL Training, Model=Qwen2.5-3B-Instruct2025.10 | 64.1 | |
| StepSearchType=RL Training, Model=Qwen2.5-7B-Instruct2025.10 | 40 | |
| Search-R1Type=RL Training, Model=Qwen2.5-7B-Instruct2025.10 | 36.8 | |
| StepSearchType=RL Training, Model=Qwen2.5-3B-Instruct2025.10 | 34.4 | |
| R1-InstructType=RL Training, Model=Qwen2.5-3B-Instruct2025.10 | 29.3 | |
| ZeroSearchType=RL Training, Model=Qwen2.5-7B-Instruct2025.10 | 27.8 | |
| Search-R1Type=RL Training, Model=Qwen2.5-3B-Instruct2025.10 | 26.4 | |
| R1-InstructType=RL Training, Model=Qwen2.5-7B-Instruct2025.10 | 19.2 | |
| ZeroSearchType=RL Training, Model=Qwen2.5-3B-Instruct2025.10 | 11.1 |