Long-context Language Understanding on LongBench
60.31M-AvgCortexDebate
Evaluation Results
| Method | Links | |||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CortexDebateType=Ours2025.07 | 60.31 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MDMType=Part Debate2025.07 | 59.63 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GDType=Part Debate2025.07 | 55.97 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-8BLoss function=base+aux2026.02 | 55.38 | — | — | 40.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 73.46 | — | — | — | — | — | — | 25.43 | 61.56 | — | 84.4 | 47.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GDType=Part Debate2025.07 | 54.58 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NDType=Part Debate2025.07 | 54.55 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| NDType=Part Debate2025.07 | 54.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ChatEvalType=Full Debate2025.07 | 53.56 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-4BLoss function=base+aux2026.02 | 52.75 | — | — | 36.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 71.27 | — | — | — | — | — | — | 19.19 | 58.61 | — | 83.25 | 47.51 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RECONCILEType=Full Debate2025.07 | 52.55 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-4BLoss function=base2026.02 | 51.46 | — | — | 35.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 70.81 | — | — | — | — | — | — | 18.87 | 60.81 | — | 79.96 | 43.25 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ChatEvalType=Full Debate2025.07 | 51.23 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-8BLoss function=base2026.02 | 50.67 | — | — | 42.34 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 72.43 | — | — | — | — | — | — | 16.54 | 54.43 | — | 75.79 | 42.49 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen2.5-14B-InstructBackbone=Qwen2.5-14B-Instruct, Mem%=100%, Configuration=FullAttention baseline2026.01 | 50.55 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 56.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | 43.01 | 52.54 | 24.99 | 71.68 | 54.35 | — | — | — | — | — | — | — | — | — | — | — | — | |
| PRDType=Full Debate2025.07 | 50.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HeteroCacheBackbone=Qwen2.5-14B-Instruct, Mem%=30%2026.01 | 49.97 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 53.33 | — | — | — | — | — | — | — | — | — | — | — | — | — | 42.46 | 52.12 | 24.12 | 71.18 | 54.63 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Full KVBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=1.02026.03 | 49.88 | — | — | 46.6 | — | — | — | 32.52 | 34.66 | 25.19 | — | 73 | 92.11 | 43.7 | 6.56 | — | 65.52 | — | 54.75 | 100 | 50.01 | — | — | 56.85 | — | — | — | 31.38 | 58.1 | 27.11 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ThinKBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.72026.03 | 49.86 | — | — | 45.44 | — | — | — | 32.65 | 32.63 | 24.96 | — | 73 | 91.72 | 42.1 | 6.96 | — | 65.94 | — | 55.64 | 99.5 | 51.09 | — | — | 58.83 | — | — | — | 32.77 | 58.3 | 26.19 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Llama-3.1-8B-InstructBackbone=Llama-3.1-8B-Instruct, Mem%=100%, Configuration=FullAttention baseline2026.01 | 49.77 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | 43.53 | 40.82 | 29.04 | 69.48 | 53.75 | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKV-HybridBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.72026.03 | 49.67 | — | — | 45.97 | — | — | — | 32.42 | 33.27 | 25.12 | — | 73 | 92.28 | 43.8 | 6.27 | — | 65.12 | — | 55.13 | 97 | 50.16 | — | — | 58.15 | — | — | — | 32.04 | 58.14 | 26.82 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| HeteroCacheBackbone=Llama-3.1-8B-Instruct, Mem%=50%2026.01 | 49.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 60.77 | — | — | — | — | — | — | — | — | — | — | — | — | — | 43.38 | 40.54 | 28.59 | 69.48 | 53.75 | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKV-HybridBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.52026.03 | 49.4 | — | — | 45.44 | — | — | — | 32.69 | 31.81 | 24.81 | — | 73 | 91.87 | 42.48 | 6.92 | — | 65.74 | — | 55.11 | 96.5 | 50.13 | — | — | 57.68 | — | — | — | 31.93 | 58.53 | 25.83 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullKVModel=LLaMA-3.1-8B, Keep Token Rate=100%2026.02 | 49.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullKVKeep Token Rate=100%, Backbone=LLaMA-3.1-8B2026.02 | 49.32 | — | — | — | — | — | — | 31.28 | 35.12 | 25.28 | — | 73 | — | — | — | — | 63.38 | 56.73 | — | 99.5 | 46.66 | — | — | 54.94 | 43.8 | — | 8.88 | 30.16 | 56.02 | 27.25 | 91.65 | — | — | — | — | — | — | — | — | — | 45.53 | — | — | — | — | — | — | — | — | |
| Flash-AttentionModel=LLaMA-3.1-8B-Instruct2026.02 | 49.28 | — | — | 45.37 | — | — | — | 31.26 | 35.12 | 25.38 | — | 72.5 | — | — | — | — | 9.41 | 99.5 | — | 56.8 | 45 | — | — | 55.8 | 44.32 | — | 62.9 | 30.22 | 55.97 | 27.2 | 91.65 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OmniKVBackbone=Llama-3.1-8B-Instruct, Mem%=50%2026.01 | 49.22 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 60.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | 43.28 | 40.33 | 29.09 | 68.94 | 53.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CAKEBackbone=Llama-3.1-8B-Instruct, Mem%=50%2026.01 | 49.16 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 60.45 | — | — | — | — | — | — | — | — | — | — | — | — | — | 43.37 | 40.48 | 27.95 | 68.97 | 53.75 | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKVBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.72026.03 | 49.08 | — | — | 44.53 | — | — | — | 32.08 | 29.29 | 25.06 | — | 73 | 92.52 | 41.67 | 6.87 | — | 63.57 | — | 54.72 | 99 | 49.89 | — | — | 57.84 | — | — | — | 32.24 | 57.9 | 25.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Token Sparse AttentionModel=LLaMA-3.1-8B-Instruct, Base Method=Flash-Attention2026.02 | 48.96 | — | — | 45.17 | — | — | — | 31.28 | 34.97 | 25.43 | — | 72.5 | — | — | — | — | 7.47 | 98 | — | 56.28 | 45.56 | — | — | 55.34 | 44.07 | — | 63.06 | 30.43 | 55 | 27.09 | 91.64 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleModel=LLaMA-3.1-8B, Keep Token Rate=40%2026.02 | 48.88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleKeep Token Rate=40%, Backbone=LLaMA-3.1-8B2026.02 | 48.88 | — | — | — | — | — | — | 30.95 | 35.11 | 25.17 | — | 70 | — | — | — | — | 63.99 | 57.33 | — | 99.5 | 47.42 | — | — | 55.36 | 42.56 | — | 6.84 | 30.14 | 54.27 | 25.97 | 91.79 | — | — | — | — | — | — | — | — | — | 45.74 | — | — | — | — | — | — | — | — | |
| MLDType=Full Debate2025.07 | 48.87 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAModel=LLaMA-3.1-8B, Keep Token Rate=40%2026.02 | 48.72 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAKeep Token Rate=40%, Backbone=LLaMA-3.1-8B2026.02 | 48.72 | — | — | — | — | — | — | 31.72 | 33.1 | 24.87 | — | 71 | — | — | — | — | 62.4 | 56.23 | — | 99.5 | 47.38 | — | — | 54.32 | 43.31 | — | 7.18 | 30.37 | 55.02 | 25.21 | 92.04 | — | — | — | — | — | — | — | — | — | 45.84 | — | — | — | — | — | — | — | — | |
| QuestBackbone=Llama-3.1-8B-Instruct, Mem%=100%2026.01 | 48.6 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.55 | — | — | — | — | — | — | — | — | — | — | — | — | — | 42.79 | 40.12 | 29.11 | 67.5 | 52.52 | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKVBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.52026.03 | 48.53 | — | — | 44.8 | — | — | — | 31.79 | 29.22 | 24.1 | — | 73 | 91.92 | 41.45 | 7.17 | — | 64.61 | — | 54.91 | 89.5 | 50.25 | — | — | 57.51 | — | — | — | 33.06 | 57.91 | 25.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleModel=LLaMA-3.1-8B, Keep Token Rate=20%2026.02 | 48.39 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleKeep Token Rate=20%, Backbone=LLaMA-3.1-8B2026.02 | 48.39 | — | — | — | — | — | — | 29.83 | 33.96 | 24.92 | — | 69.5 | — | — | — | — | 63.08 | 57.74 | — | 99.5 | 47.17 | — | — | 55.33 | 42.93 | — | 5.41 | 29.86 | 54.57 | 24.39 | 91.63 | — | — | — | — | — | — | — | — | — | 44.37 | — | — | — | — | — | — | — | — | |
| RECONCILEType=Full Debate2025.07 | 48.33 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineModel=Mistral-7B-Instruct2024.05 | 48.32 | — | — | — | — | — | — | — | 32.85 | — | — | 71 | — | — | — | — | 55.51 | — | 48.96 | 60 | 27.33 | 42.74 | 1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullKVModel=Mistral-Nemo-12B, Keep Token Rate=100%2026.02 | 48.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleModel=Mistral-Nemo-12B, Keep Token Rate=20%2026.02 | 48.27 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleModel=Mistral-Nemo-12B, Keep Token Rate=40%2026.02 | 48.24 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAModel=LLaMA-3.1-8B, Keep Token Rate=20%2026.02 | 48.12 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAKeep Token Rate=20%, Backbone=LLaMA-3.1-8B2026.02 | 48.12 | — | — | — | — | — | — | 31.01 | 30.72 | 25.33 | — | 71 | — | — | — | — | 61.21 | 56.07 | — | 99.5 | 46.09 | — | — | 53.85 | 43.14 | — | 7.61 | 30.63 | 54.61 | 23.14 | 92.08 | — | — | — | — | — | — | — | — | — | 43.95 | — | — | — | — | — | — | — | — | |
| FastKVModel=Mistral-Nemo-12B, Keep Token Rate=40%2026.02 | 48.04 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ShadowKVBackbone=Llama-3.1-8B-Instruct, Mem%=50%2026.01 | 48.03 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 59.08 | — | — | — | — | — | — | — | — | — | — | — | — | — | 42.1 | 40.42 | 23.34 | 69.48 | 53.75 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAModel=Mistral-Nemo-12B, Keep Token Rate=40%2026.02 | 48.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseBackbone=Llama3.1-8B-Instruct, Evaluation Phase=Prefill2026.02 | 47.95 | — | — | 44.67 | — | — | — | 29.41 | 34.77 | 25.14 | — | 73 | — | — | — | — | 55.13 | — | 47.79 | — | 45.62 | — | — | — | 43.24 | 99 | 10 | 30.05 | 55.4 | 26.9 | 91.16 | 55.97 | — | — | — | — | — | 50.48 | — | — | — | — | — | — | — | — | — | — | — | |
| MinferenceModel=LLaMA-3.1-8B-Instruct2026.02 | 47.89 | — | — | 45.64 | — | — | — | 29.24 | 34.88 | 24.94 | — | 71 | — | — | — | — | 6.38 | 98.5 | — | 50.72 | 42.93 | — | — | 52.59 | 44.15 | — | 62.38 | 29.48 | 55.05 | 26.53 | 91.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleModel=LLaMA-3.1-8B, Keep Token Rate=10%2026.02 | 47.83 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OracleKeep Token Rate=10%, Backbone=LLaMA-3.1-8B2026.02 | 47.83 | — | — | — | — | — | — | 28.92 | 32.26 | 25.29 | — | 68.5 | — | — | — | — | 59.63 | 56.85 | — | 99.5 | 47.11 | — | — | 55.85 | 41.8 | — | 7.47 | 29.85 | 54.99 | 21.95 | 91.43 | — | — | — | — | — | — | — | — | — | 43.94 | — | — | — | — | — | — | — | — | |
| OracleModel=Mistral-Nemo-12B, Keep Token Rate=10%2026.02 | 47.75 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastKVModel=LLaMA-3.1-8B, Keep Token Rate=40%2026.02 | 47.68 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastKVKeep Token Rate=40%, Backbone=LLaMA-3.1-8B2026.02 | 47.68 | — | — | — | — | — | — | 31.02 | 28.21 | 24.39 | — | 71 | — | — | — | — | 61.32 | 55.47 | — | 99.5 | 47.36 | — | — | 55.09 | 42.89 | — | 6.09 | 30.39 | 55.4 | 22.02 | 91.87 | — | — | — | — | — | — | — | — | — | 40.92 | — | — | — | — | — | — | — | — | |
| PRDType=Full Debate2025.07 | 47.67 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Sketch&WalkBackbone=Llama3.1-8B-Instruct, Evaluation Phase=Prefill2026.02 | 47.67 | — | — | 44.13 | — | — | — | 31.27 | 34.39 | 25.26 | — | 70 | — | — | — | — | 52.76 | — | 47.88 | — | 47.26 | — | — | — | 43.99 | 99 | 6.16 | 29.61 | 56.69 | 26.92 | 92.02 | 55.3 | — | — | — | — | — | 50.43 | — | — | — | — | — | — | — | — | — | — | — | |
| Token Sparse AttentionModel=LLaMA-3.1-8B-Instruct, Base Method=Minference2026.02 | 47.61 | — | — | 46.19 | — | — | — | 29.89 | 34.79 | 25.05 | — | 71 | — | — | — | — | 5.34 | 94.5 | — | 51.12 | 42.18 | — | — | 52.52 | 44.07 | — | 62.79 | 29.28 | 54.62 | 26.48 | 91.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FlexPrefillModel=LLaMA-3.1-8B-Instruct2026.02 | 47.48 | — | — | 44.63 | — | — | — | 31.97 | 34.5 | 24.92 | — | 70.5 | — | — | — | — | 4.09 | 82 | — | 60.36 | 41.82 | — | — | 54.94 | 43.81 | — | 63.08 | 27.99 | 57.31 | 27.06 | 90.74 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PaluBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.72026.03 | 47.47 | — | — | 46.73 | — | — | — | 29.91 | 32.51 | 24.69 | — | 73 | 89.1 | 43.77 | 2.5 | — | 57.9 | — | 56.78 | 96.5 | 45.35 | — | — | 53.26 | — | — | — | 28.96 | 53.08 | 25.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Token Sparse AttentionModel=LLaMA-3.1-8B-Instruct, Base Method=FlexPrefill2026.02 | 47.39 | — | — | 44.84 | — | — | — | 31.09 | 34.79 | 25.47 | — | 69 | — | — | — | — | 3.09 | 85.5 | — | 60.24 | 40.24 | — | — | 54.74 | 43.72 | — | 62.78 | 28.07 | 56.63 | 27.12 | 90.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastKVModel=Mistral-Nemo-12B, Keep Token Rate=20%2026.02 | 47.38 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Full KVBackbone Model=Mistral-7B-Instruct-v0.3, Compression Ratio=1.02026.03 | 47.34 | — | — | 39.12 | — | — | — | 26.45 | 34.11 | 25.93 | — | 76 | 88.89 | 46.97 | 5.5 | — | 61.47 | — | 62.57 | 97 | 36.31 | — | — | 50.45 | — | — | — | 29.8 | 50.28 | 26.56 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastKVModel=LLaMA-3.1-8B, Keep Token Rate=20%2026.02 | 47.33 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastKVKeep Token Rate=20%, Backbone=LLaMA-3.1-8B2026.02 | 47.33 | — | — | — | — | — | — | 30.12 | 28.49 | 24.33 | — | 70.5 | — | — | — | — | 59.68 | 54.93 | — | 99.5 | 47.25 | — | — | 54.42 | 42.85 | — | 7.09 | 30.54 | 53.67 | 21.46 | 92.08 | — | — | — | — | — | — | — | — | — | 40.45 | — | — | — | — | — | — | — | — | |
| MInferenceBackbone=Llama3.1-8B-Instruct, Evaluation Phase=Prefill2026.02 | 47.17 | — | — | 43.3 | — | — | — | 27.18 | 33.76 | 25.62 | — | 72.5 | — | — | — | — | 55.43 | — | 50.97 | — | 46.39 | — | — | — | 43.27 | 97.5 | 4.25 | 28.6 | 54.18 | 26.64 | 90.4 | 54.75 | — | — | — | — | — | 50.03 | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAModel=LLaMA-3.1-8B, Keep Token Rate=10%2026.02 | 47.13 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAKeep Token Rate=10%, Backbone=LLaMA-3.1-8B2026.02 | 47.13 | — | — | — | — | — | — | 31.53 | 28.15 | 24.76 | — | 70 | — | — | — | — | 58.31 | 53.86 | — | 99.5 | 44.73 | — | — | 53.68 | 42.93 | — | 6.51 | 31.09 | 53.83 | 20.42 | 92.37 | — | — | — | — | — | — | — | — | — | 42.36 | — | — | — | — | — | — | — | — | |
| CommonKVBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.72026.03 | 47.13 | — | — | 41.16 | — | — | — | 29.96 | 27.91 | 24.21 | — | 66 | 90.03 | 43.1 | 6.67 | — | 61.92 | — | 53.41 | 100 | 45.42 | — | — | 53.86 | — | — | — | 29.87 | 54.9 | 25.68 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MiniCacheModel=Mistral-7B-Instruct2024.05 | 46.99 | — | — | — | — | — | — | — | 31.46 | — | — | 71 | — | — | — | — | 54.79 | — | 51.02 | 64.14 | 24.97 | 27.54 | 5.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAModel=Mistral-Nemo-12B, Keep Token Rate=20%2026.02 | 46.99 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKV-HybridBackbone Model=Mistral-7B-Instruct-v0.3, Compression Ratio=0.72026.03 | 46.97 | — | — | 38.9 | — | — | — | 26.73 | 33.46 | 25.05 | — | 76 | 89.47 | 46.76 | 4 | — | 60.79 | — | 62.51 | 95.5 | 35.43 | — | — | 50.27 | — | — | — | 30.24 | 50.37 | 26.02 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ShadowKVBackbone=Qwen2.5-14B-Instruct, Mem%=30%2026.01 | 46.93 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 53.2 | — | — | — | — | — | — | — | — | — | — | — | — | — | 35.77 | 50.8 | 17.18 | 70.65 | 54 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CAKEBackbone=Qwen2.5-14B-Instruct, Mem%=30%2026.01 | 46.92 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 53.09 | — | — | — | — | — | — | — | — | — | — | — | — | — | 40.74 | 42.95 | 23.69 | 66.77 | 54.25 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ThinKBackbone Model=Mistral-7B-Instruct-v0.3, Compression Ratio=0.72026.03 | 46.86 | — | — | 37.87 | — | — | — | 25.92 | 33.81 | 25.66 | — | 76 | 88.43 | 45.81 | 5.5 | — | 61.9 | — | 61.95 | 95 | 34.55 | — | — | 50.13 | — | — | — | 30.42 | 50.35 | 26.4 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKVBackbone Model=Mistral-7B-Instruct-v0.3, Compression Ratio=0.72026.03 | 46.86 | — | — | 37.65 | — | — | — | 24.98 | 32.29 | 25.11 | — | 76 | 89.47 | 47.44 | 5.5 | — | 62.26 | — | 62.18 | 95.5 | 34.54 | — | — | 49.43 | — | — | — | 29.93 | 51.5 | 25.99 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastKVModel=LLaMA-3.1-8B, Keep Token Rate=10%2026.02 | 46.81 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FastKVKeep Token Rate=10%, Backbone=LLaMA-3.1-8B2026.02 | 46.81 | — | — | — | — | — | — | 30.09 | 28.08 | 24.57 | — | 70 | — | — | — | — | 58.43 | 53.49 | — | 99 | 44.73 | — | — | 53.61 | 42.69 | — | 6.56 | 30.6 | 54.87 | 20.93 | 92.38 | — | — | — | — | — | — | — | — | — | 38.96 | — | — | — | — | — | — | — | — | |
| SpecPrefillModel=LLaMA-3.1-8B, Keep Token Rate=40%2026.02 | 46.55 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SpecPrefillKeep Token Rate=40%, Backbone=LLaMA-3.1-8B2026.02 | 46.55 | — | — | — | — | — | — | 30.65 | 32.67 | 24.93 | — | 67.5 | — | — | — | — | 55.9 | 49 | — | 97 | 44.05 | — | — | 54.74 | 42.98 | — | 1.89 | 27.05 | 56.27 | 24.88 | 91.3 | — | — | — | — | — | — | — | — | — | 44 | — | — | — | — | — | — | — | — | |
| EchoKV-HybridBackbone Model=Mistral-7B-Instruct-v0.3, Compression Ratio=0.52026.03 | 46.55 | — | — | 38.69 | — | — | — | 25.1 | 32.88 | 25.02 | — | 76 | 89.57 | 46.55 | 3.5 | — | 61 | — | 62.78 | 93 | 34.95 | — | — | 49.55 | — | — | — | 31.32 | 49.05 | 25.83 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OmniKVBackbone=Qwen2.5-14B-Instruct, Mem%=30%2026.01 | 46.48 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 48.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | 37.9 | 50.21 | 23.1 | 68.44 | 50.25 | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKVBackbone Model=Mistral-7B-Instruct-v0.3, Compression Ratio=0.52026.03 | 46.36 | — | — | 36.92 | — | — | — | 24.51 | 31.54 | 25.28 | — | 76 | 89.32 | 46.93 | 3.5 | — | 61.72 | — | 61.66 | 93.5 | 34.26 | — | — | 49.71 | — | — | — | 30.26 | 51.33 | 25.36 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CommonKVBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.52026.03 | 46.27 | — | — | 40.64 | — | — | — | 29.82 | 25.23 | 23.85 | — | 59 | 89.98 | 42.5 | 10.12 | — | 62.46 | — | 52.5 | 99.5 | 45.58 | — | — | 54.45 | — | — | — | 25.24 | 54.56 | 24.92 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MLDType=Full Debate2025.07 | 46.26 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GALIBackbone=Llama3-8b-ins-8k, Target Context Window=16k2025.02 | 46.22 | — | 24.69 | 45.26 | — | 51.33 | — | 30.79 | 29.28 | 22.65 | 24.63 | 77 | 91.61 | — | — | — | 56.84 | — | 49.04 | — | 37.16 | — | — | 51.78 | 42.92 | 95.5 | 9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLAAModel=Mistral-Nemo-12B, Keep Token Rate=10%2026.02 | 46.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FullKVBackbone=Llama3.2-3B2026.03 | 46.03 | — | — | — | — | — | — | 25.62 | 33.11 | 24.38 | — | 72 | — | — | — | — | 54.65 | 56.83 | — | — | 40.05 | — | — | 50.31 | 43.25 | — | 4 | 25.35 | 53.79 | 25.95 | 88.89 | — | — | — | — | — | — | — | — | — | 40.77 | — | — | — | 97.5 | — | — | — | — | |
| PaluBackbone Model=Mistral-7B-Instruct-v0.3, Compression Ratio=0.72026.03 | 46.03 | — | — | 36.08 | — | — | — | 26.37 | 33.59 | 25.16 | — | 73.5 | 87.77 | 45.11 | 4 | — | 59.04 | — | 61.06 | 94.5 | 36.12 | — | — | 52.56 | — | — | — | 28.47 | 46.57 | 26.54 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LAQBackbone=Llama3.2-3B, KV Cache Size=20482026.03 | 45.8 | — | — | — | — | — | — | 26.13 | 30.09 | 24.44 | — | 71.5 | — | — | — | — | 55.44 | 56.77 | — | — | 39.88 | — | — | 49.5 | 43.11 | — | 4 | 24.8 | 54.05 | 25.8 | 88.89 | — | — | — | — | — | — | — | — | — | 41.44 | — | — | — | 97 | — | — | — | — | |
| LOOKAHEADKVBackbone=Llama3.2-3B, KV Cache Size=20482026.03 | 45.77 | — | — | — | — | — | — | 24.87 | — | 24.18 | — | 71 | — | — | — | — | 55.12 | 56.77 | — | — | 40.06 | — | — | 50 | 43.49 | — | 4 | 25.23 | 53.79 | 25.88 | 88.99 | — | — | — | — | — | — | — | — | — | 40.67 | — | — | — | 97 | — | — | — | — | |
| EchoKV-HybridBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.32026.03 | 45.74 | — | — | 38.41 | — | — | — | 30.72 | 23.21 | 23.25 | — | 57.5 | 90.85 | 41.77 | 6.77 | — | 62.77 | — | 51.69 | 95.5 | 49.07 | — | — | 50.83 | — | — | — | 30.89 | 57.28 | 21.39 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SnapKVBackbone=Llama3.2-3B, KV Cache Size=20482026.03 | 45.67 | — | — | — | — | — | — | 25.24 | 28.41 | 23.92 | — | 71.5 | — | — | — | — | 55.07 | 56.97 | — | — | 40.6 | — | — | 51.36 | 42.39 | — | 4 | 24.54 | 53.8 | 25.62 | 88.86 | — | — | — | — | — | — | — | — | — | 40.5 | — | — | — | 98 | — | — | — | — | |
| FastKVModel=Mistral-Nemo-12B, Keep Token Rate=10%2026.02 | 45.62 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LOOKAHEADKVBackbone=Llama3.2-3B, KV Cache Size=10242026.03 | 45.57 | — | — | — | — | — | — | 25.07 | 29.61 | 24.17 | — | 69.5 | — | — | — | — | 55.59 | 56.72 | — | — | 40.36 | — | — | 51.35 | 42.9 | — | 4 | 24.28 | 53.64 | 25.41 | 88.99 | — | — | — | — | — | — | — | — | — | 39.51 | — | — | — | 98 | — | — | — | — | |
| LAQBackbone=Llama3.2-3B, KV Cache Size=10242026.03 | 45.54 | — | — | — | — | — | — | 26.68 | 27.69 | 24.07 | — | 71.5 | — | — | — | — | 54.93 | 56.88 | — | — | 40.01 | — | — | 49.15 | 43.54 | — | 4 | 24.55 | 53.65 | 25.25 | 88.89 | — | — | — | — | — | — | — | — | — | 40.78 | — | — | — | 97 | — | — | — | — | |
| GALIBackbone=Llama3-8b-ins-8k, Target Context Window=32k2025.02 | 45.38 | — | 28.63 | 45.66 | — | 51.07 | — | 29 | 29.98 | 22.79 | 24.59 | 77 | 91.13 | — | — | — | 57.07 | — | 52.63 | — | 38.35 | — | — | 47.23 | 42.38 | 83 | 5.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| PyramidKVBackbone=Llama3.2-3B, KV Cache Size=20482026.03 | 45.29 | — | — | — | — | — | — | 24.5 | 27.88 | 24.27 | — | 70 | — | — | — | — | 54.84 | 55.59 | — | — | 40.03 | — | — | 51.31 | 42.11 | — | 4 | 24.05 | 53.88 | 25.65 | 88.8 | — | — | — | — | — | — | — | — | — | 39.8 | — | — | — | 98 | — | — | — | — | |
| SelfExtendBackbone=Llama3-8b-ins-8k, Target Context Window=32k2025.02 | 45.28 | — | 26.27 | 44.23 | — | 48.28 | — | 29.19 | 29.24 | 22.68 | 24.59 | 76 | 90.16 | — | — | — | 57.47 | — | 49.51 | — | 38.29 | — | — | 50.19 | 42.45 | 88 | 8 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EchoKVBackbone Model=Llama3.1-8B-Instruct, Compression Ratio=0.32026.03 | 45.27 | — | — | 42.47 | — | — | — | 30.75 | 24.23 | 23.84 | — | 49 | 91.79 | 40.9 | 7.25 | — | 62.09 | — | 53.03 | 84.5 | 49.58 | — | — | 54.27 | — | — | — | 31.22 | 57.03 | 22.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |