Multi-hop Question Answering on MuSiQue (EM, Retrieval Rate, Latency)
15.2Exact Match (EM)TARG-Margin
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TARG-MarginBackbone=Qwen2.5-7B2025.11 | 15.2 | 1.6 | 0.16 | |
| Never-RAGBackbone=Qwen2.5-7B2025.11 | 14.6 | 0 | 0.18 | |
| Self-Eval GateBackbone=Qwen2.5-7B2025.11 | 14.6 | 24.6 | 5.66 | |
| CRAG-liteBackbone=Qwen2.5-7B2025.11 | 14.2 | 23 | 1.76 | |
| SKRBackbone=Qwen2.5-7B2025.11 | 14.2 | 13.2 | 0.92 | |
| Self-RAG-liteBackbone=Qwen2.5-7B2025.11 | 13.4 | 48.2 | 1.8 | |
| Always-RAGBackbone=Qwen2.5-7B2025.11 | 10 | 100 | 12.24 | |
| FLAREBackbone=Qwen2.5-7B, θ=0.52025.11 | 9.8 | 100 | 4.73 | |
| FLAREBackbone=Qwen2.5-7B, θ=0.32025.11 | 9 | 98.8 | 17.86 |