Multi-Hop Search-augmented Question Answering on HotpotQA (success rate (%))
38.8Success RateRole-Agent
Evaluation Results
| Method | Links | |
|---|---|---|
| Role-AgentType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 38.8 | |
| GiGPOType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 36.9 | |
| StepSearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 34.5 | |
| + MASABackbone=Qwen3-32B2026.05 | 34.2 | |
| + DS-AdapterBackbone=Qwen3-32B2026.05 | 34 | |
| + Base SkillBackbone=Qwen3-32B2026.05 | 33.8 | |
| + MASABackbone=Qwen3-14B2026.05 | 32.8 | |
| + Base SkillBackbone=Qwen3-14B2026.05 | 32.7 | |
| Search_R1Type=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 32.4 | |
| No SkillBackbone=Qwen3-32B2026.05 | 32.2 | |
| No SkillBackbone=Qwen3-14B2026.05 | 31.7 | |
| + DS-AdapterBackbone=Qwen3-14B2026.05 | 31.6 | |
| + DS-AdapterBackbone=Qwen3-4B2026.05 | 28.6 | |
| + Base SkillBackbone=Qwen3-8B2026.05 | 28.6 | |
| + MASABackbone=Qwen3-8B2026.05 | 28.6 | |
| + Base SkillBackbone=Qwen3-4B2026.05 | 28.5 | |
| + DS-AdapterBackbone=Qwen3-8B2026.05 | 27.8 | |
| No SkillBackbone=Qwen3-4B2026.05 | 27.7 | |
| + MASABackbone=Qwen3-4B2026.05 | 27.4 | |
| Zero-SearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 27.4 | |
| No SkillBackbone=Qwen3-8B2026.05 | 24.8 | |
| R1-InstructType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 23.7 |