Multi-hop Question Answering on HotpotQA (EM, F1, LLM)
61.13EMSeleCom
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SeleComBackbone=Mistral-7B-Instruct, Top-k=5, Comp. Rate=82x2026.01 | 61.13 | 76.16 | 76.57 | |
| SeleComBackbone=Qwen2.5-7B-Instruct, Top-k=5, Comp. Rate=82x2026.01 | 59.14 | 74.38 | 75.43 | |
| LLM (with RAG*)Backbone=Mistral-7B-Instruct, Top-k=5, Comp. Rate=-2026.01 | 57.22 | 74.36 | 69.96 | |
| LLM (with RAG*)Backbone=Qwen2.5-7B-Instruct, Top-k=5, Comp. Rate=-2026.01 | 56.26 | 72.77 | 74.94 | |
| SeleComBackbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=82x2026.01 | 44.46 | 57.88 | 57.49 | |
| LLM (with RAG*)Backbone=Qwen2.5-7B-Instruct, Top-k=1, Comp. Rate=-2026.01 | 43.19 | 52.61 | 51.71 | |
| LLM (with RAG*)Backbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=-2026.01 | 42.92 | 56.43 | 55.98 | |
| SeleComBackbone=Qwen2.5-7B-Instruct, Top-k=1, Comp. Rate=82x2026.01 | 42.35 | 55.34 | 55.91 | |
| LLM (with RAG)Backbone=Qwen2.5-7B-Instruct, Top-k=1, Comp. Rate=-2026.01 | 36.96 | 49.95 | 57.17 | |
| COCOMBackbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=128x2026.01 | 25.9 | 35.36 | 34 | |
| ICAEBackbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=4x2026.01 | 24.77 | 40.06 | 53.82 | |
| LLM (with RAG)Backbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=-2026.01 | 18.69 | 35.87 | 58.57 | |
| RECOMPBackbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=N/A2026.01 | 18.51 | 35.76 | 56.4 | |
| LLM (w/o RAG)Backbone=Qwen2.5-7B-Instruct, Top-k=1, Comp. Rate=-2026.01 | 15.73 | 23.79 | 24.63 | |
| LLMLingua-2Backbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=4x2026.01 | 12.46 | 26.86 | 45.91 | |
| xRAGBackbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=164x2026.01 | 8.45 | 22.52 | 45.65 | |
| LLM (w/o RAG)Backbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=-2026.01 | 5.97 | 16.28 | 29.97 | |
| PISCOBackbone=Mistral-7B-Instruct, Top-k=1, Comp. Rate=64x2026.01 | 0.95 | 13.08 | 54.18 |