Multi-Hop Question Answering on 2Wiki (token-level F1)
58.7Token-Level F1COSEARCH
Evaluation Results
| Method | Links | |
|---|---|---|
| COSEARCHMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 58.7 | |
| OracleMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 58.4 | |
| Fixed RankerMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 56 | |
| ZeroSearchMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 55.9 | |
| Retrieval OnlyMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 55.6 | |
| Search-R1Main agent backbone=Qwen2.5-7B-Instruct2026.04 | 54.7 | |
| OracleMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 51 | |
| COSEARCHMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 42.8 | |
| Fixed RankerMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 41.1 | |
| Retrieval OnlyMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 40.6 | |
| Search-R1Main agent backbone=Qwen2.5-3B-Instruct2026.04 | 39.7 | |
| ZeroSearchMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 38.7 | |
| Direct InferenceMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 30.4 | |
| Search-o1Main agent backbone=Qwen2.5-7B-Instruct2026.04 | 28.6 | |
| CoTMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 26.4 | |
| Direct InferenceMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 24.9 | |
| RAGMain agent backbone=Qwen2.5-7B-Instruct2026.04 | 24.4 | |
| CoTMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 22.9 | |
| RAGMain agent backbone=Qwen2.5-3B-Instruct2026.04 | 19.3 | |
| Search-o1Main agent backbone=Qwen2.5-3B-Instruct2026.04 | 12.8 |