Generation w/ citations
34.8Citation Quality (8k Context)Base
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| BaseBackbone=Qwen3 8B2026.03 | 34.8 | 27.3 | 15.2 | 6.4 | 0.7 | |
| DCABackbone=Qwen3 8B2026.03 | 33.3 | 26.8 | 16.4 | 3.9 | 1.6 | |
| CLPBackbone=Qwen3 8B2026.03 | 32.6 | 26.6 | 17 | 12.3 | 5.9 | |
| L2ABackbone=Qwen3 8B2026.03 | 30.4 | 27.1 | 15.2 | 5.2 | 2.1 | |
| CLPBackbone=Qwen 2.5 7B2026.03 | 24.2 | 18.3 | 8.6 | 4.8 | 2.8 | |
| SelfExtendBackbone=Qwen 2.5 7B2026.03 | 22.5 | 16.3 | 5.9 | 2.5 | 2.3 | |
| DCABackbone=Qwen 2.5 7B2026.03 | 21.9 | 15.5 | 9.5 | 2.1 | 1.2 | |
| BaseBackbone=Qwen 2.5 7B2026.03 | 21.7 | 16.3 | 8.2 | 2.2 | 1.3 | |
| S^2-attnBackbone=Qwen 2.5 7B2026.03 | 20.9 | 14.9 | 8.4 | 2.9 | 2.5 | |
| FlexPrefillBackbone=Qwen3 8B2026.03 | 19.3 | 7.7 | 5.7 | 5 | 3.6 | |
| L2A (Ours)Backbone=Qwen 2.5 7B2026.03 | 17.2 | 10.5 | 8.5 | 4.1 | 1.9 | |
| NSABackbone=Qwen 2.5 7B2026.03 | 15.3 | 6.4 | 2.8 | 2.1 | 1.5 | |
| FlexPrefillBackbone=Qwen 2.5 7B2026.03 | 13.3 | 6.8 | 5.1 | 3.6 | 2.4 |