Inference Efficiency on HotpotQA
496Time to Last Token (ms)SeleCom
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SeleComTop-k=1, Backbone=Mistral-7B-Instruct2026.01 | 496 | 1,271 | 49 | |
| SeleComTop-k=5, Backbone=Mistral-7B-Instruct2026.01 | 505 | 1,416 | 53 | |
| COCOMTop-k=1, Backbone=Mistral-7B-Instruct2026.01 | 541 | 1,905 | 138 | |
| PISCOTop-k=1, Backbone=Mistral-7B-Instruct2026.01 | 553 | 1,927 | 141 | |
| RECOMPTop-k=1, Backbone=Mistral-7B-Instruct2026.01 | 563 | 1,799 | 89 | |
| LLMLingua-2Top-k=1, Backbone=Mistral-7B-Instruct2026.01 | 716 | 2,082 | 76 | |
| ICAETop-k=1, Backbone=Mistral-7B-Instruct2026.01 | 738 | 3,293 | 164 | |
| LLM (with RAG)Top-k=1, Backbone=Mistral-7B-Instruct2026.01 | 1,006 | 2,694 | 57 | |
| LLM (with RAG*)Top-k=1, Backbone=Mistral-7B-Instruct2026.01 | 1,024 | 2,570 | 63 | |
| LLM (w/o RAG)Top-k=1, Backbone=Mistral-7B-Instruct2026.01 | 1,073 | 1,270 | 47 | |
| xRAGTop-k=1, Backbone=Mistral-7B-Instruct2026.01 | 1,188 | 1,158 | 71 | |
| LLM (with RAG*)Top-k=5, Backbone=Mistral-7B-Instruct2026.01 | 1,459 | 4,038 | 80 |