List your company

Long-context language understanding on LongBench (Comprehensive 16-Task Subset)

30.46NtrvQAFull Cache
15.712819.541423.3727.1986May 29, 2026

Evaluation results

MethodLinks
Full CacheBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%2026.0530.4647.2955.0358.6551.6133.0335.0324.9727.042984.8139.2910.1510051.4344.1945.12
CaMBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0529.2930.0539.2653.793927.628.5122.7623.422282.5539.458.7996.546.9444.1939.63
GRKVBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0529.2432.8739.8653.4742.6826.1329.2723.0423.565081.8341.6810.059753.5546.1542.52
CriticalKVBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%2026.0528.9431.3539.2253.5741.7925.9429.2723.0323.44980.5841.019.5596.552.8945.9942
Full CacheBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%2026.0528.3840.7251.684936.2127.8134.2325.4126.9355.7584.9320.945.059846.9253.8642.86
SnapKVBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%2026.0527.8427.1934.1854.9837.924.4827.7621.8622.723883.0140.947.5589.551.7344.3139.62
GRKVBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=SnapKV2026.0527.7328.334.3654.4536.8525.9927.9822.0423.1141.584.0641.178.559251.4744.8640.28
D2OBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0527.1826.5531.2849.3631.4122.6125.9721.1821.953682.8139.931074.550.8545.4637.32
CaMBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=SnapKV2026.0526.1626.6135.5454.4237.462627.0321.8722.5920.0784.7839.568.6886.546.5342.7237.91
AsymKVBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=SnapKV2026.0525.630.0636.4849.0833.5219.4627.9222.2924.417.589.2426.995.6163.549.144.9535.36
D2OBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=SnapKV2026.0525.1119.7129.1648.4125.5219.8324.9119.9320.883382.4138.917.0466.550.0845.0534.78
AsymKVBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=SnapKV2026.0524.5724.135.8342.8428.9816.7728.7221.6423.7845.8285.6124.711.6160.1745.8148.6834.98
GRKVBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0523.225.8443.0244.0732.9219.6429.0723.0823.694788.0324.34.8893.550.4952.0239.05
AsymKVBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0523.0624.2737.2545.6326.4318.9428.821.5824.075185.3623.252.4354.545.0649.2335.05
CaMBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0522.9324.6441.245.4833.722.2628.2722.8623.4235.0785.4411.634.259048.9550.1736.89
CriticalKVBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%2026.0522.8825.6842.5542.53—18.6828.7623.0223.2247.2586.9923.144.659349.4951.6538.47
AsymKVBackbone=Llama-3.1-8B-Instruct, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0522.7629.7634.1743.0834.2716.627.2122.1224.24783.8721.863.8345.551.9247.1732.21
CaMBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=SnapKV2026.0519.518.335.1340.3521.9420.4127.5721.8822.8136.584.6814.042.4787.1747.6950.8534.46
D2OBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=CriticalKV2026.0519.2815.9635.1336.6425.8416.4826.2821.1321.7641.587.0826.984.4357.550.6151.2633.62
GRKVBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=SnapKV2026.0518.719.4537.8139.527.8221.4428.0322.15234487.1723.134.659052.5152.5436.99
SnapKVBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%2026.0518.0618.4636.6839.725.6720.9527.8321.6822.6243.7586.4223.424.5589.551.4752.636.46
D2OBackbone=Mistral-7B-Instruct-v0.3, Cache Budget=20%, Base Eviction Method=SnapKV2026.0516.2810.7734.0935.2722.6114.7425.5919.9320.9235.7586.4728.095.265550.6552.1632.1