Fact Chaining & Relational Reasoning on Long-context benchmarks (8k-128k context)
52.8Accuracy (8k Context)DCA
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DCABackbone=Qwen3 8B2026.03 | 52.8 | 47 | 46 | 40.8 | 30 | |
| BaseBackbone=Qwen3 8B2026.03 | 52.8 | 47.2 | 44 | 27.6 | 0.4 | |
| SelfExtendBackbone=Qwen 2.5 7B2026.03 | 52.2 | 47.6 | 41.8 | 38.2 | 34.6 | |
| DCABackbone=Qwen 2.5 7B2026.03 | 51.6 | 47.8 | 44.8 | 39.8 | 35.2 | |
| BaseBackbone=Qwen 2.5 7B2026.03 | 51.6 | 47.8 | 43 | 41.8 | 21.4 | |
| FlexPrefillBackbone=Qwen 2.5 7B2026.03 | 47.6 | 42.6 | 37.8 | 33.4 | 35.8 | |
| CLPBackbone=Qwen 2.5 7B2026.03 | 47.4 | 43.6 | 40 | 31.4 | 38.6 | |
| L2ABackbone=Qwen3 8B2026.03 | 47.2 | 47.6 | 47.6 | 38.6 | 37.8 | |
| S^2-attnBackbone=Qwen 2.5 7B2026.03 | 45.2 | 43.8 | 35.2 | 38.2 | 39.2 | |
| FlexPrefillBackbone=Qwen3 8B2026.03 | 44.8 | 39 | 39 | 34.2 | 31 | |
| L2A (Ours)Backbone=Qwen 2.5 7B2026.03 | 44.6 | 42.2 | 40.6 | 33.4 | 40.4 | |
| CLPBackbone=Qwen 2.5 1.5B2026.03 | 43.6 | 39.8 | 35.8 | 28.2 | 29.4 | |
| S^2-attnBackbone=Qwen 2.5 1.5B2026.03 | 43.4 | 40.2 | 36.2 | 31.6 | 26.4 | |
| CLPBackbone=Qwen3 8B2026.03 | 43.4 | 43.6 | 46.6 | 37.8 | 35.6 | |
| FlexPrefillBackbone=Qwen 2.5 1.5B2026.03 | 43.2 | 37.6 | 36 | 32 | 27.4 | |
| NSABackbone=Qwen 2.5 7B2026.03 | 42.2 | 39.2 | 31.2 | 24 | 17.6 | |
| L2ABackbone=Qwen 2.5 1.5B2026.03 | 41.4 | 36 | — | 29.2 | 30 | |
| NSABackbone=Qwen 2.5 1.5B2026.03 | 40.4 | 30.6 | 23 | 18 | 11 | |
| BaseBackbone=Qwen 2.5 1.5B2026.03 | 38.2 | 35.2 | 31 | 24 | 13 | |
| DCABackbone=Qwen 2.5 1.5B2026.03 | 37 | 35 | 35.6 | 27.8 | 23.4 | |
| SelfExtendBackbone=Qwen 2.5 1.5B2026.03 | 37 | 35.6 | 34 | 26.2 | 23.4 |