Long-context language modeling on LongBench-E 1.0 (test)
49.92S-Doc QA Perf.Elastic Attention
Evaluation Results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Elastic AttentionBackbone=Llama-3.1-8B-Instruct, variant=FA-SSA2026.01 | 49.92 | 0.64 | 48.92 | 0.63 | 30.17 | 0.73 | 67.99 | 0.74 | 54 | 0.64 | 60.7 | 0.79 | 53.35 | 0.69 | |
| Elastic AttentionBackbone=Llama-3.1-8B-Instruct, variant=FA-XA2026.01 | 49.4 | 0.71 | 52.94 | 0.69 | 30.3 | 0.8 | 68.55 | 0.79 | 49.66 | 0.72 | 56.49 | 0.87 | 52.71 | 0.77 | |
| XAttentionBackbone=Llama-3.1-8B-Instruct2026.01 | 48.82 | — | 51.73 | — | 30.26 | — | 68.57 | — | 40.83 | — | 56.53 | — | 51 | — | |
| Llama-3.1-8B-InstructBackbone=Llama-3.1-8B-Instruct2026.01 | 48.75 | — | 51.85 | — | 30.26 | — | 68.16 | — | 56 | — | 55.81 | — | 53.28 | — | |
| Elastic AttentionBackbone=Llama-3.1-8B-Instruct, variant=XA-SSA2026.01 | 48.3 | — | 46.4 | — | 30.27 | — | 67.83 | — | 41.37 | — | 60.81 | — | 50.71 | — | |
| MoBABackbone=Llama-3.1-8B-Instruct2026.01 | 46.63 | — | 43.91 | — | 30.72 | — | 66.78 | — | 36 | — | 64.48 | — | 49.69 | — | |
| Elastic AttentionBackbone=Qwen3-8B, variant=FA-SSA2026.01 | 46.15 | 0.64 | 46.54 | 0.65 | 28.19 | 0.72 | 67.52 | 0.71 | 48.07 | 0.65 | 62.95 | 0.78 | 51.51 | 0.69 | |
| Qwen3-8BBackbone=Qwen3-8B2026.01 | 45.57 | — | 51.59 | — | 28.34 | — | 66.64 | — | 50.16 | — | 61.2 | — | 52.16 | — | |
| MoBABackbone=Qwen3-8B2026.01 | 44.73 | — | 43.28 | — | 30.33 | — | 64.22 | — | 48.95 | — | 62.55 | — | 50.47 | — | |
| Elastic AttentionBackbone=Qwen3-4B, variant=FA-XA2026.01 | 44.4 | 0.68 | 39.42 | 0.71 | 28.49 | 0.82 | 65.26 | 0.76 | 44.35 | 0.74 | 54.29 | 0.87 | 47.59 | 0.76 | |
| Elastic AttentionBackbone=Qwen3-8B, variant=FA-XA2026.01 | 44.01 | 0.75 | 49.99 | 0.76 | 28.3 | 0.83 | 66.23 | 0.8 | 50.95 | 0.77 | 60.57 | 0.86 | 51.66 | 0.8 | |
| Qwen3-4BBackbone=Qwen3-4B2026.01 | 43.69 | — | 38.48 | — | 28.46 | — | 66.21 | — | 49.59 | — | 54.38 | — | 48.45 | — | |
| XAttentionBackbone=Qwen3-8B2026.01 | 43.48 | — | 48.99 | — | 28.39 | — | 66.26 | — | 42.86 | — | 60.62 | — | 50.13 | — | |
| NSABackbone=Llama-3.1-8B-Instruct2026.01 | 42.33 | — | 40.37 | — | 29.93 | — | 66.65 | — | 15.17 | — | 57.65 | — | 44.03 | — | |
| Elastic AttentionBackbone=Qwen3-4B, variant=FA-SSA2026.01 | 42.2 | 0.66 | 38.86 | 0.68 | 28.5 | 0.76 | 65.73 | 0.73 | 48.43 | 0.71 | 54.34 | 0.82 | 48.08 | 0.73 | |
| Elastic AttentionBackbone=Qwen3-4B, variant=XA-SSA2026.01 | 41.92 | — | 38.67 | — | 28.45 | — | 65.25 | — | 49.25 | — | 55.07 | — | 48.14 | — | |
| XAttentionBackbone=Qwen3-4B2026.01 | 41.58 | — | 38.85 | — | 28.76 | — | 65.45 | — | 39.01 | — | 54.52 | — | 46.44 | — | |
| NSABackbone=Qwen3-8B2026.01 | 40.63 | — | 37.45 | — | 27.97 | — | 67.14 | — | 29 | — | 62.19 | — | 46.12 | — | |
| Elastic AttentionBackbone=Qwen3-8B, variant=XA-SSA2026.01 | 39.43 | — | 41.38 | — | 28.27 | — | 66.19 | — | 46.73 | — | 61.85 | — | 49.25 | — | |
| NSABackbone=Qwen3-4B2026.01 | 39.13 | — | 34.81 | — | 27.38 | — | 63.44 | — | 23.8 | — | 58.1 | — | 43.02 | — | |
| MoBABackbone=Qwen3-4B2026.01 | 38.16 | — | 34.3 | — | 29.46 | — | 64.57 | — | 37.42 | — | 54.62 | — | 45.09 | — | |
| QuestBase Model=Llama-3-8B-262K, Retrieval/Cache Method=Block-level sparse attention, Sparsity Budget (p')=10242025.12 | 30.76 | — | 19.97 | — | 27.04 | — | 78.84 | — | — | — | 50.26 | — | -0.39 | — | |
| ShadowKVBase Model=Llama-3-8B-262K, Retrieval/Cache Method=Block-level sparse attention (Mean pooling), Sparsity Budget (p')=10242025.12 | 30.73 | — | 17.88 | — | 30.71 | — | 78.95 | — | — | — | 49.15 | — | -0.27 | — | |
| FlatBase Model=Llama-3-8B-262K, Retrieval/Cache Method=Exact KNN, Sparsity Budget (p')=10242025.12 | 29.94 | — | 24.01 | — | 28.92 | — | 78.35 | — | — | — | 50.26 | — | 0.54 | — | |
| ShadowKVBase Model=Yi-9B-200K, Retrieval/Cache Method=Block-level sparse attention (Mean pooling), Sparsity Budget (p')=10242025.12 | 29.9 | — | 30.32 | — | 27.71 | — | 83.61 | — | — | — | 69.31 | — | -0.98 | — | |
| CTkvrBase Model=Llama-3-8B-262K, Retrieval/Cache Method=Centroid-then-Token Indexing, Sparsity Budget (p')=10242025.12 | 29.68 | — | 21.2 | — | 28.98 | — | 78.12 | — | — | — | 50.29 | — | -0.11 | — | |
| CTkvrBase Model=Llama-3-8B-262K, Retrieval/Cache Method=Centroid-then-Token Indexing, Sparsity Budget (p')=5122025.12 | 29.55 | — | 21.25 | — | 29.32 | — | 78.12 | — | — | — | 50.41 | — | -0.03 | — | |
| FullKVBase Model=Llama-3-8B-262K, Retrieval/Cache Method=None (FullKV)2025.12 | 29.13 | — | 21.82 | — | 29.33 | — | 79.04 | — | — | — | 49.5 | — | 41.76 | — | |
| FlatBase Model=Yi-9B-200K, Retrieval/Cache Method=Exact KNN, Sparsity Budget (p')=10242025.12 | 29.02 | — | 37.99 | — | 26.29 | — | 83.61 | — | — | — | 69.12 | — | 0.05 | — | |
| CTkvrBase Model=Yi-9B-200K, Retrieval/Cache Method=Centroid-then-Token Indexing, Sparsity Budget (p')=10242025.12 | 28.77 | — | 38.08 | — | 26.24 | — | 83.61 | — | — | — | 69.35 | — | 0.06 | — | |
| CTkvrBase Model=Yi-9B-200K, Retrieval/Cache Method=Centroid-then-Token Indexing, Sparsity Budget (p')=5122025.12 | 28.72 | — | 38.16 | — | 26.15 | — | 83.61 | — | — | — | 69.72 | — | 0.12 | — | |
| QuestBase Model=Yi-9B-200K, Retrieval/Cache Method=Block-level sparse attention, Sparsity Budget (p')=10242025.12 | 28.19 | — | 37.39 | — | 25.97 | — | 83.34 | — | — | — | 68.46 | — | -0.48 | — | |
| FullKVBase Model=Yi-9B-200K, Retrieval/Cache Method=None (FullKV)2025.12 | 28.04 | — | 38.54 | — | 25.86 | — | 83.5 | — | — | — | 69.81 | — | 49.15 | — | |
| MagicPIGBase Model=Yi-9B-200K, Retrieval/Cache Method=Token-level sparse attention (LSH), Sparsity Budget (p')=10242025.12 | 26.67 | — | 38.21 | — | 24.74 | — | 83.66 | — | — | — | 69.14 | — | -0.66 | — | |
| MagicPIGBase Model=Llama-3-8B-262K, Retrieval/Cache Method=Token-level sparse attention (LSH), Sparsity Budget (p')=10242025.12 | 18.17 | — | 9.59 | — | 26.79 | — | 78.08 | — | — | — | 48.82 | — | -5.47 | — | |
| SnapKVBase Model=Yi-9B-200K, Retrieval/Cache Method=Eviction-based, Sparsity Budget (p')=10242025.12 | 8.83 | — | 10.13 | — | 11.29 | — | 52.38 | — | — | — | 25.16 | — | -27.59 | — | |
| SnapKVBase Model=Llama-3-8B-262K, Retrieval/Cache Method=Eviction-based, Sparsity Budget (p')=10242025.12 | 5.42 | — | 6.25 | — | 12.35 | — | 40.12 | — | — | — | 27.06 | — | -23.52 | — |