Multi-hop Question Answering on Bamboogle (EM, LJ)
44.56EMM2R
Evaluation results
| Method | Links | ||
|---|---|---|---|
| M2RBackbone=Qwen2.5-7B-Instruct2026.04 | 44.56 | 56.89 | |
| ReSearchBackbone=Qwen2.5-7B-Instruct2026.04 | 42.4 | 54.4 | |
| M2RBackbone=Qwen2.5-3B-Instruct2026.04 | 39.58 | 47.2 | |
| SUREBackbone=Qwen2.5-7B-Instruct2026.04 | 39.58 | 52.81 | |
| ReSearchBackbone=Qwen2.5-3B-Instruct2026.04 | 38.11 | 48.12 | |
| COFTBackbone=Qwen2.5-7B-Instruct2026.04 | 35.71 | 49.23 | |
| SUREBackbone=Qwen2.5-3B-Instruct2026.04 | 31.2 | 39.95 | |
| COFTBackbone=Qwen2.5-3B-Instruct2026.04 | 25.3 | 31.38 | |
| IRCoTBackbone=Qwen2.5-7B-Instruct2026.04 | 24.8 | 36.8 | |
| Iter-RetGenBackbone=Qwen2.5-7B-Instruct2026.04 | 21.6 | 35.2 | |
| Naive RAGBackbone=Qwen2.5-7B-Instruct2026.04 | 20.8 | 32 | |
| IRCoTBackbone=Qwen2.5-3B-Instruct2026.04 | 17.6 | 26.13 | |
| Iter-RetGenBackbone=Qwen2.5-3B-Instruct2026.04 | 16.03 | 22.61 | |
| Naive RAGBackbone=Qwen2.5-3B-Instruct2026.04 | 14.43 | 20 | |
| Naive GenerationBackbone=Qwen2.5-7B-Instruct2026.04 | 10.4 | 22.4 | |
| Naive GenerationBackbone=Qwen2.5-3B-Instruct2026.04 | 5.01 | 8.5 |