Question Answering on NQ (EM, ROUGE, F1)
21.2EMGRIP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GRIPBackbone=Qwen2.5-7B-Instruct2026.04 | 21.2 | 27.7 | 32.5 | |
| GRIPBackbone=Qwen-3-4B-Instruct2026.04 | 21.1 | 28.6 | 33.2 | |
| Robust-RAGBackbone=Qwen-3-4B-Instruct2026.04 | 20.9 | 23.8 | 28.1 | |
| Robust-RAGBackbone=Qwen2.5-7B-Instruct2026.04 | 20.5 | 23.8 | 27.8 | |
| Single-RAGBackbone=Qwen2.5-7B-Instruct2026.04 | 18.9 | 24.6 | 28.3 | |
| R1-SearcherBackbone=Qwen2.5-7B-Instruct2026.04 | 14.3 | 19.2 | 23 | |
| Single-RAGBackbone=Qwen-3-4B-Instruct2026.04 | 12.9 | 18.9 | 21.9 | |
| R1-SearcherBackbone=Qwen-3-4B-Instruct2026.04 | 11.8 | 16.8 | 20.1 | |
| InstructBackbone=Qwen2.5-7B-Instruct2026.04 | 11.8 | 16.7 | 20 | |
| InstructBackbone=Qwen-3-4B-Instruct2026.04 | 6.2 | 11.3 | 13.7 |