Multi-hop Question Answering on HotpotQA (EM, F1, Calls)
0.673Exact Match (EM)ReaLM-Retrieve
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ReaLM-RetrieveBackbone=DeepSeek-R1-671B2026.04 | 0.673 | 0.784 | 2 | |
| Search-R1Backbone=DeepSeek-R1-671B2026.04 | 0.641 | 0.752 | 2.8 | |
| IRCoTBackbone=DeepSeek-R1-671B2026.04 | 0.624 | 0.735 | 4.3 | |
| ReaLM-RetrieveBackbone=DeepSeek-R1-Distill-Qwen-32B2026.04 | 0.604 | 0.718 | 1.9 | |
| Single RAGBackbone=DeepSeek-R1-671B2026.04 | 0.579 | 0.694 | 1 | |
| Search-R1Backbone=DeepSeek-R1-Distill-Qwen-32B2026.04 | 0.578 | 0.692 | 2.7 | |
| IRCoTBackbone=DeepSeek-R1-Distill-Qwen-32B2026.04 | 0.562 | 0.679 | 4.1 | |
| Self-RAGBackbone=DeepSeek-R1-Distill-Qwen-32B, Model Type=fine-tuned2026.04 | 0.554 | 0.663 | 2.4 | |
| FLAREBackbone=DeepSeek-R1-Distill-Qwen-32B, Note=Uses token probabilities2026.04 | 0.539 | 0.654 | 3.2 | |
| Single RAGBackbone=DeepSeek-R1-Distill-Qwen-32B2026.04 | 0.513 | 0.628 | 1 | |
| No RetrievalBackbone=DeepSeek-R1-671B2026.04 | 0.452 | 0.581 | 0 | |
| No RetrievalBackbone=DeepSeek-R1-Distill-Qwen-32B2026.04 | 0.384 | 0.512 | 0 |