Retrieval-Augmented Generation on Long-context benchmarks (Context Scaling)
53.7RAG Score (8k Context)CLP
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CLPBackbone=Qwen 2.5 1.5B2026.03 | 53.7 | 51.6 | 50.7 | 44.8 | 36.7 | |
| FlexPrefillBackbone=Qwen 2.5 1.5B2026.03 | 52 | 51 | 47.2 | 43 | 35.8 | |
| S^2-attnBackbone=Qwen 2.5 1.5B2026.03 | 51.8 | 50.6 | 48.6 | 39.6 | 27.4 | |
| BaseBackbone=Qwen 2.5 1.5B2026.03 | 49.8 | 46.1 | 41.8 | 30.2 | 22.8 | |
| DCABackbone=Qwen 2.5 1.5B2026.03 | 49.7 | 46 | 42.2 | 36.3 | 28.8 | |
| SelfExtendBackbone=Qwen 2.5 1.5B2026.03 | 49.4 | 45.9 | 39.8 | 33.6 | 27.4 | |
| L2ABackbone=Qwen 2.5 1.5B2026.03 | 49.1 | 48.9 | 47.1 | 41.2 | 34.8 | |
| NSABackbone=Qwen 2.5 1.5B2026.03 | 48.5 | 44 | 39 | 32.3 | 30.2 | |
| S^2-attnBackbone=Qwen 2.5 1.5B2026.03 | 8.9 | 3.8 | 3.5 | 3 | 1.4 | |
| CLPBackbone=Qwen 2.5 1.5B2026.03 | 8.3 | 5.2 | 2.3 | 2.3 | 1.5 | |
| BaseBackbone=Qwen 2.5 1.5B2026.03 | 8.1 | 4.9 | 4.4 | 0.7 | 0.9 | |
| DCABackbone=Qwen 2.5 1.5B2026.03 | 7.9 | 5.3 | 3.2 | 1.4 | 0.9 | |
| SelfExtendBackbone=Qwen 2.5 1.5B2026.03 | 7.4 | 5 | 2.9 | 1.4 | 1 | |
| NSABackbone=Qwen 2.5 1.5B2026.03 | 7.1 | 3.4 | 2.6 | 1.6 | 0.8 | |
| L2ABackbone=Qwen 2.5 1.5B2026.03 | 6.9 | 2.6 | 2.7 | 2.8 | 1.3 | |
| FlexPrefillBackbone=Qwen 2.5 1.5B2026.03 | 5.5 | 3.1 | 3.3 | 1.9 | 1.6 |