Cross-lingual Retrieval-Augmented Generation on Hotpot-ENKB5 (all)
86.54Language Consistency (ALL-AVG)Llama-3.2-3B-Instruct
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Llama-3.2-3B-InstructBackbone=Llama-3.2-3B-Instruct2026.07 | 86.54 | 35.62 | 53.75 | 58.7 | |
| Knowledge distillationBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 86.51 | 49.34 | 68.92 | 68.25 | |
| Prompt-ControlBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 86.42 | 39.29 | 53.9 | 59.87 | |
| DITBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 86.39 | 37.06 | 48.8 | 57.42 | |
| TR-RAGBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 86.37 | 57.8 | 77.5 | 73.88 | |
| SDFTBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 86.36 | 52.74 | 73.47 | 70.86 | |
| On-Policy DistillationBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 86.35 | 53.69 | 73.42 | 71.15 | |
| Naive-RLBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 86.34 | 55.03 | 76.61 | 72.66 | |
| Llama-3.3-70B-InstructBackbone=Llama-3.3-70B-Instruct2026.07 | 85.99 | 63.22 | 83.8 | 77.67 | |
| TranslateBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 85.88 | 29.86 | 36.12 | 50.62 | |
| SFTBackbone=Llama-3.2-3B-Instruct, Teacher=Llama-3.3-70B-Instruct2026.07 | 83.52 | 50.66 | 64.84 | 66.34 |