Question Answering on NQ Open (EM, Retrieval, Latency)
39.4Exact Match (EM)Self-Eval Gate
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Self-Eval GateBackbone=Qwen2.5-7B2025.11 | 39.4 | 22.8 | 1.79 | |
| Self-RAG-liteBackbone=Qwen2.5-7B2025.11 | 39 | 19.8 | 2.19 | |
| SKRBackbone=Qwen2.5-7B2025.11 | 39 | 4.8 | 2.46 | |
| TARG-MarginBackbone=Qwen2.5-7B2025.11 | 39 | 6.2 | 0.71 | |
| Never-RAGBackbone=Qwen2.5-7B2025.11 | 38.8 | 0 | 3.29 | |
| FLAREBackbone=Qwen2.5-7B, θ=0.52025.11 | 38.6 | 99.6 | 1.1 | |
| CRAG-liteBackbone=Qwen2.5-7B2025.11 | 38 | 20 | 2.01 | |
| Always-RAGBackbone=Qwen2.5-7B2025.11 | 37.4 | 100 | 2.92 | |
| FLAREBackbone=Qwen2.5-7B, θ=0.32025.11 | 35.8 | 96 | 14.4 |