Question Answering Performance and Latency on TriviaQA
62.2Exact Match (EM)SKR
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SKRBackbone=Qwen2.5-7B2025.11 | 62.2 | 3.8 | 2.27 | |
| TARG-MarginBackbone=Qwen2.5-7B2025.11 | 61.8 | 0.1 | 0.21 | |
| Never-RAGBackbone=Qwen2.5-7B2025.11 | 60.8 | 0 | 2.95 | |
| Self-RAG-liteBackbone=Qwen2.5-7B2025.11 | 59.8 | 20 | 2.21 | |
| Self-Eval GateBackbone=Qwen2.5-7B2025.11 | 58.4 | 15 | 2.39 | |
| Always-RAGBackbone=Qwen2.5-7B2025.11 | 57.6 | 100 | 3.46 | |
| CRAG-liteBackbone=Qwen2.5-7B2025.11 | 56.6 | 18 | 2.33 | |
| FLAREBackbone=Qwen2.5-7B, θ=0.52025.11 | 56.6 | 100 | 1.31 | |
| FLAREBackbone=Qwen2.5-7B, θ=0.32025.11 | 55.8 | 97.8 | 1.17 |