Long-context Question Answering on 2WikiMQA (Passage Split)
52.53ScoreBaseline
Evaluation results
| Method | Links | |
|---|---|---|
| BaselineModel=ChatGLM, Setting=Passage Split2026.03 | 52.53 | |
| BaselineModel=Qwen, Setting=Passage Split2026.03 | 51.61 | |
| Our + ReorderModel=Qwen, Setting=Passage Split2026.03 | 50.58 | |
| OurModel=Qwen, Setting=Passage Split2026.03 | 50.19 | |
| OurModel=ChatGLM, Setting=Passage Split2026.03 | 48.9 | |
| Our + ReorderModel=ChatGLM, Setting=Passage Split2026.03 | 46.66 | |
| BaselineModel=LLaMA, Setting=Passage Split2026.03 | 45.88 | |
| EPIC (15%)Model=ChatGLM, Setting=Passage Split2026.03 | 45.21 | |
| Our + ReorderModel=LLaMA, Setting=Passage Split2026.03 | 44.17 | |
| CacheBlendModel=Qwen, Setting=Passage Split2026.03 | 43.3 | |
| OurModel=LLaMA, Setting=Passage Split2026.03 | 42.08 | |
| CacheBlendModel=LLaMA, Setting=Passage Split2026.03 | 39.76 | |
| EPIC (15%)Model=LLaMA, Setting=Passage Split2026.03 | 38.85 | |
| CacheBlendModel=ChatGLM, Setting=Passage Split2026.03 | 37.57 | |
| EPIC (15%)Model=Qwen, Setting=Passage Split2026.03 | 36.97 | |
| No RecomputeModel=ChatGLM, Setting=Passage Split2026.03 | 35.23 | |
| No RecomputeModel=LLaMA, Setting=Passage Split2026.03 | 30.66 | |
| No RecomputeModel=Qwen, Setting=Passage Split2026.03 | 11.62 |