Average Performance across Context Lengths (8k-128k) on Average across tasks
57.2Performance @ 8k Context LengthCLP
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CLPBackbone=Qwen3 8B2026.03 | 57.2 | 54.1 | 52.1 | 49.3 | 44 | |
| DCABackbone=Qwen3 8B2026.03 | 56.8 | 53.7 | 49.5 | 38 | 31.7 | |
| BaseBackbone=Qwen3 8B2026.03 | 56.8 | 53.9 | 50 | 34.5 | 4.1 | |
| CLPBackbone=Qwen 2.5 7B2026.03 | 56 | 52.1 | 49 | 45.4 | 40.7 | |
| L2ABackbone=Qwen3 8B2026.03 | 55.7 | 53.4 | 50.6 | 47.2 | 42.5 | |
| SelfExtendBackbone=Qwen 2.5 7B2026.03 | 55.2 | 51.4 | 43.6 | 36 | 27.8 | |
| DCABackbone=Qwen 2.5 7B2026.03 | 54.7 | 51.5 | 47.7 | 39.4 | 31.8 | |
| BaseBackbone=Qwen 2.5 7B2026.03 | 54.7 | 51.6 | 47.4 | 34.1 | 21 | |
| FlexPrefillBackbone=Qwen3 8B2026.03 | 53.4 | 49.2 | 48.3 | 46 | 41 | |
| FlexPrefillBackbone=Qwen 2.5 7B2026.03 | 53.3 | 48.7 | 45.7 | 42.5 | 37.8 | |
| L2A (Ours)Backbone=Qwen 2.5 7B2026.03 | 53 | 49.7 | 47.5 | 44.3 | 41.3 | |
| S^2-attnBackbone=Qwen 2.5 7B2026.03 | 52.8 | 49.9 | 45.8 | 40.9 | 30.7 | |
| NSABackbone=Qwen 2.5 7B2026.03 | 51.5 | 40.3 | 29.6 | 25.3 | 21.7 |