Multi-Hop Search-augmented Question Answering on Bamboogle (success rate %)
68.4Success RateRole-Agent
Evaluation Results
| Method | Links | |
|---|---|---|
| Role-AgentType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 68.4 | |
| No SkillBackbone=Qwen3-8B2026.05 | 68.1 | |
| + Base SkillBackbone=Qwen3-32B2026.05 | 67.7 | |
| + MASABackbone=Qwen3-32B2026.05 | 66.1 | |
| No SkillBackbone=Qwen3-32B2026.05 | 64.5 | |
| + DS-AdapterBackbone=Qwen3-32B2026.05 | 64.1 | |
| GiGPOType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 64.1 | |
| + MASABackbone=Qwen3-8B2026.05 | 62.5 | |
| + MASABackbone=Qwen3-4B2026.05 | 61.3 | |
| StepSearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 34.4 | |
| R1-InstructType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 29.3 | |
| Search_R1Type=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 26.4 | |
| + Base SkillBackbone=Qwen3-14B2026.05 | 15.3 | |
| + DS-AdapterBackbone=Qwen3-4B2026.05 | 12.9 | |
| + DS-AdapterBackbone=Qwen3-14B2026.05 | 12.5 | |
| Zero-SearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 11.1 | |
| No SkillBackbone=Qwen3-14B2026.05 | 10.5 | |
| + Base SkillBackbone=Qwen3-4B2026.05 | 10.1 | |
| + Base SkillBackbone=Qwen3-8B2026.05 | 10.1 | |
| No SkillBackbone=Qwen3-4B2026.05 | 9.3 | |
| + MASABackbone=Qwen3-14B2026.05 | 8.9 | |
| + DS-AdapterBackbone=Qwen3-8B2026.05 | 7.7 |