Multi-Hop Search-augmented Question Answering on 2Wiki (success rate %)
45.2Success RateRole-Agent
Evaluation Results
| Method | Links | |
|---|---|---|
| Role-AgentType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 45.2 | |
| GiGPOType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 37 | |
| + MASABackbone=Qwen3-32B2026.05 | 35.6 | |
| + DS-AdapterBackbone=Qwen3-32B2026.05 | 32 | |
| StepSearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 32 | |
| Search_R1Type=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 31.9 | |
| No SkillBackbone=Qwen3-8B2026.05 | 30.6 | |
| + Base SkillBackbone=Qwen3-14B2026.05 | 30.3 | |
| + MASABackbone=Qwen3-14B2026.05 | 30 | |
| Zero-SearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 30 | |
| No SkillBackbone=Qwen3-32B2026.05 | 29.3 | |
| R1-InstructType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 29.2 | |
| + DS-AdapterBackbone=Qwen3-14B2026.05 | 28.5 | |
| + MASABackbone=Qwen3-4B2026.05 | 27 | |
| No SkillBackbone=Qwen3-14B2026.05 | 26.8 | |
| + Base SkillBackbone=Qwen3-32B2026.05 | 26 | |
| + Base SkillBackbone=Qwen3-8B2026.05 | 25.9 | |
| + MASABackbone=Qwen3-8B2026.05 | 25.7 | |
| + Base SkillBackbone=Qwen3-4B2026.05 | 24.4 | |
| + DS-AdapterBackbone=Qwen3-4B2026.05 | 23.9 | |
| + DS-AdapterBackbone=Qwen3-8B2026.05 | 22.9 | |
| No SkillBackbone=Qwen3-4B2026.05 | 22.8 |