Single-Hop Search-augmented Question Answering on TriviaQA
61.8Success Rate+ MASA
Evaluation Results
| Method | Links | |
|---|---|---|
| + MASABackbone=Qwen3-14B2026.05 | 61.8 | |
| + MASABackbone=Qwen3-32B2026.05 | 61.6 | |
| + DS-AdapterBackbone=Qwen3-32B2026.05 | 61.5 | |
| + Base SkillBackbone=Qwen3-32B2026.05 | 61.4 | |
| + Base SkillBackbone=Qwen3-14B2026.05 | 60.5 | |
| Role-AgentType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 60.4 | |
| No SkillBackbone=Qwen3-14B2026.05 | 60.2 | |
| + DS-AdapterBackbone=Qwen3-14B2026.05 | 60.2 | |
| No SkillBackbone=Qwen3-32B2026.05 | 59.8 | |
| GiGPOType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 59.5 | |
| + Base SkillBackbone=Qwen3-8B2026.05 | 58.5 | |
| + DS-AdapterBackbone=Qwen3-8B2026.05 | 57.6 | |
| + Base SkillBackbone=Qwen3-4B2026.05 | 57.4 | |
| Zero-SearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 57.4 | |
| + MASABackbone=Qwen3-8B2026.05 | 56.7 | |
| + DS-AdapterBackbone=Qwen3-4B2026.05 | 56.5 | |
| + MASABackbone=Qwen3-4B2026.05 | 55.3 | |
| Search_R1Type=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 54.5 | |
| R1-InstructType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.06 | 53.7 | |
| No SkillBackbone=Qwen3-4B2026.05 | 51 | |
| No SkillBackbone=Qwen3-8B2026.05 | 46.5 |