Long-context language understanding on LongBench-e
53.04Average ScoreDense
Evaluation Results
| Method | Links | ||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DenseModel=Qwen2.5-14B-Instruct-1M, sparse token budget=5122025.06 | 53.04 | 44.32 | 100 | 65.96 | 88.41 | 36.25 | 45.52 | 76.34 | 53.73 | 60.68 | 31.93 | 22.83 | 22.14 | 41.41 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Exact2026.02 | 53 | 69.96 | 99.67 | — | 87.53 | 62.17 | 38.56 | 64 | — | 43.71 | 34.16 | — | 24.32 | 43.76 | 50.08 | 56.43 | 14.67 | — | — | — | — | — | — | — | — | — | |
| HATAModel=Qwen2.5-14B-Instruct-1M, sparse token budget=5122025.06 | 52.9 | 44.86 | 99.67 | 65.87 | 88.49 | 37.41 | 45.41 | 76.67 | 53.45 | 60.7 | 31.25 | 20.5 | 22.02 | 41.46 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Exact (Baseline)Model=Llama-3.1-8B-Instruct2025.02 | 50.17 | 49.62 | 98.13 | — | 91.85 | 42.73 | 42.36 | 71.67 | — | 38.6 | 31.31 | — | 22.07 | 42.87 | 48.54 | 52.05 | 20.37 | — | — | — | — | — | — | — | — | — | |
| 16bitModel=Qwen3-8B, Precision=16bit2026.05 | 49.56 | — | 95.81 | — | — | — | — | 71.33 | — | — | 29.56 | — | 21.96 | 10.27 | — | 13.1 | — | — | 91.95 | 62.47 | — | — | — | — | — | — | |
| ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.05 | 48.92 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Full KVBackbone=Qwen3-4B, KV eviction ratio (r)=02026.05 | 48.77 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OScaRModel=Qwen3-8B, Quantization=INT2, Group Size=322026.05 | 48.74 | — | 92.78 | — | — | — | — | 72 | — | — | 28.33 | — | 21.46 | 9.94 | — | 12.35 | — | — | 91.46 | 61.57 | — | — | — | — | — | — | |
| NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.05 | 48.69 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.05 | 48.45 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OTTModel=Qwen3-8B, Quantization=INT2, Group Size=322026.05 | 48.21 | — | 92.39 | — | — | — | — | 71 | — | — | 28.56 | — | 21.82 | 9.26 | — | 11.86 | — | — | 91.36 | 59.41 | — | — | — | — | — | — | |
| KIVIModel=Qwen3-8B, Quantization=INT2, Group Size=322026.05 | 47.95 | — | 91.92 | — | — | — | — | 71 | — | — | 28.12 | — | 21.58 | 9.03 | — | 11.86 | — | — | 91.09 | 58.96 | — | — | — | — | — | — | |
| TurboQuant+Model=Qwen3-8B, Quantization=2.5-bit2026.05 | 47.56 | — | 88.92 | — | — | — | — | 71.33 | — | — | 28.47 | — | 21.66 | 8.35 | — | 12.53 | — | — | 90.95 | 58.23 | — | — | — | — | — | — | |
| ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.05 | 47.53 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.05 | 47.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.05 | 47.35 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| COMPRESSKVretained context tokens=first and last 322026.02 | 47.25 | 64.85 | 98 | — | 86.15 | 60.43 | 35.38 | 54.33 | — | 33.37 | 30.3 | — | 22.26 | 33.23 | 38.13 | 43.43 | 14.33 | — | — | — | — | — | — | — | — | — | |
| BalanceKVModel=Llama-3.1-8B-Instruct2025.02 | 46.77 | 48.45 | 96.65 | — | 90.88 | 41.4 | 37.88 | 69 | — | 36.24 | 27.09 | — | 20.84 | 35.75 | 37.04 | 46.37 | 20.39 | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.05 | 46.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ClusterGenModel=Llama-3.1-8B-Instruct2025.02 | 46.62 | 47.23 | 96.57 | — | 90.82 | 39.26 | 39.49 | 67.67 | — | 37.24 | 21.16 | — | 19.31 | 33.93 | 42.31 | 50.85 | 20.2 | — | — | — | — | — | — | — | — | — | |
| SnapKV2026.02 | 45.36 | 69.16 | 89.06 | — | 88.32 | 56.33 | 37.15 | 49.33 | — | 31.8 | 29.7 | — | 22.1 | 25.52 | 30.13 | 44.36 | 16.67 | — | — | — | — | — | — | — | — | — | |
| ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.05 | 45.11 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.05 | 44.97 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.05 | 44.26 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Full Cache (16-bit)Backbone=Llama 3.1 8B Instruct, Precision=16-bit2026.05 | 44.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 19.53 | 16.54 | 30.28 | 68.41 | 56.41 | 61.6 | |
| StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.05 | 44.07 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BlockQuantBackbone=Llama 3.1 8B Instruct2026.05 | 44.03 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 19.55 | 16.31 | 29.72 | 68.49 | 56.48 | 61.42 | |
| EDENBackbone=Llama 3.1 8B Instruct2026.05 | 43.87 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 19.02 | 16.3 | 29.42 | 68.46 | 56.29 | 61.41 | |
| NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.05 | 43.68 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.05 | 43.58 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RabitQBackbone=Llama 3.1 8B Instruct2026.05 | 43.52 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 19.12 | 15.88 | 29.4 | 68.23 | 55.75 | 60.35 | |
| TurboQuantBackbone=Llama 3.1 8B Instruct2026.05 | 43.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 18.6 | 15.79 | 29.11 | 68.19 | 55.31 | 59.72 | |
| PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.05 | 43.07 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RotateKVModel=Qwen3-8B, Quantization=INT2, Group Size=322026.05 | 42.95 | — | 91.57 | — | — | — | — | 71 | — | — | 14.11 | — | 10.59 | 6.78 | — | 12.74 | — | — | 88.81 | 48 | — | — | — | — | — | — | |
| Uniformretained context tokens=first and last 322026.02 | 42.85 | 64.84 | 63.44 | — | 81.71 | 61.34 | 35.37 | 54.33 | — | 25.93 | 30.02 | — | 21.86 | 26.91 | 37.51 | 38.46 | 15.33 | — | — | — | — | — | — | — | — | — | |
| BalanceKVretained context tokens=first and last 322026.02 | 42.75 | 65.48 | 71.67 | — | 73.83 | 62.57 | 34.56 | 55 | — | 23.71 | 30.27 | — | 21.98 | 29.5 | 36.57 | 37.89 | 12.67 | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.05 | 42.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.05 | 42.28 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OScaRModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.05 | 41.75 | — | 44.46 | — | — | — | — | 73.33 | — | — | 29.45 | — | 5.68 | 11.61 | — | 13.68 | — | — | 90.41 | 65.39 | — | — | — | — | — | — | |
| 16bitModel=Llama-3.1-8B, Precision=16bit2026.05 | 41.7 | — | 43.57 | — | — | — | — | 72.67 | — | — | 29.88 | — | 5.89 | 11.86 | — | 13.76 | — | — | 90.15 | 65.78 | — | — | — | — | — | — | |
| PyramidKV2026.02 | 41.64 | 67.4 | 61.06 | — | 86.77 | 59.2 | 38.27 | 43.33 | — | 30.24 | 27.12 | — | 21.13 | 21.59 | 29.96 | 39.38 | 15.88 | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.05 | 41.53 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.05 | 41.53 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.05 | 41.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.05 | 41.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.05 | 41.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.05 | 41.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.05 | 40.78 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OTTModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.05 | 40.74 | — | 39.39 | — | — | — | — | 73 | — | — | 29.45 | — | 4.65 | 10.7 | — | 13.27 | — | — | 90.02 | 65.45 | — | — | — | — | — | — | |
| KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.05 | 40.38 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QuaRotModel=Qwen3-8B, Quantization=INT2, Group Size=322026.05 | 40.13 | — | 91.33 | — | — | — | — | 33 | — | — | 20.13 | — | 15.01 | 7.47 | — | 11.7 | — | — | 89.52 | 52.88 | — | — | — | — | — | — | |
| ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.05 | 40.11 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TurboQuant+Model=Llama-3.1-8B, Quantization=2.5-bit2026.05 | 40.03 | — | 41.51 | — | — | — | — | 72.33 | — | — | 26.6 | — | 3.18 | 10.04 | — | 13.85 | — | — | 89.91 | 62.82 | — | — | — | — | — | — | |
| SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.05 | 40.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.05 | 40.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KIVIModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.05 | 39.84 | — | 38.95 | — | — | — | — | 73 | — | — | 24.41 | — | 2.7 | 10.71 | — | 14.66 | — | — | 89.83 | 64.42 | — | — | — | — | — | — | |
| Full KVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=02026.05 | 39.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.05 | 38.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.05 | 38.57 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RotateKVModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.05 | 37.98 | — | 38.83 | — | — | — | — | 72.33 | — | — | 17.82 | — | 1.28 | 10.97 | — | 13.13 | — | — | 90.56 | 58.95 | — | — | — | — | — | — | |
| StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.05 | 37.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QuaRotModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.05 | 37.94 | — | 42.46 | — | — | — | — | 73 | — | — | 18.23 | — | 1.75 | 9.24 | — | 13.73 | — | — | 89.68 | 55.46 | — | — | — | — | — | — | |
| PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.05 | 37.76 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.05 | 37.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.05 | 37.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.05 | 37.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLM2026.02 | 36.4 | 68.48 | 24.83 | — | 75.17 | 56.22 | 35.79 | 52.33 | — | 20.89 | 29.65 | — | 22.25 | 23.17 | 25.49 | 26.45 | 12.5 | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.05 | 36.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.05 | 35.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.05 | 34.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.05 | 34.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.05 | 33.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.05 | 33.57 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.05 | 33.56 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.05 | 33.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.05 | 32.81 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.05 | 31.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.05 | 31.47 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.05 | 31.39 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.05 | 30.76 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.05 | 30.24 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.05 | 29.73 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.05 | 29.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.05 | 28.99 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.05 | 28.33 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.05 | 27.17 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.05 | 27.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.05 | 24.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.05 | 24.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.05 | 24.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.05 | 24.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.05 | 23.47 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.05 | 17.52 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DASHBackbone=Llama-3.1-8B, Token reduction ratio=20%2026.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 46.11 | — | — | — | — | — | — | — | — | |
| DASHBackbone=Qwen-3-8B, Token reduction ratio=20%2026.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 54.25 | — | — | — | — | — | — | — | — | |
| DASHBackbone=Qwen-2.5-7B-Instruct-1M, Token reduction ratio=20%2026.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 46.25 | — | — | — | — | — | — | — | — | |
| LazyLLMBackbone=Llama-3.1-8B, Token reduction ratio=20%2026.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 43.11 | — | — | — | — | — | — | — | — | |
| LazyLLMBackbone=Qwen-3-8B, Token reduction ratio=20%2026.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 52.12 | — | — | — | — | — | — | — | — | |
| LazyLLMBackbone=Qwen-2.5-7B-Instruct-1M, Token reduction ratio=20%2026.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 41.37 | — | — | — | — | — | — | — | — | |
| SlimInferBackbone=Llama-3.1-8B, Token reduction ratio=20%2026.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 47.03 | — | — | — | — | — | — | — | — |