Long-context language understanding on LongBench v1 (test)
30.7NrtvQA ScoreTop-k
Evaluation results
| Method | Links | |||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Top-kBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 30.7 | 44.7 | 55.4 | 55 | 46.5 | 31.7 | 34.8 | 25.1 | 26.8 | 71.5 | 92.2 | 44.8 | 7.8 | 100 | 67.1 | 63.6 | 49.8 | |
| IceCacheBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 30.6 | 44.7 | 56.3 | 55.2 | 45.9 | 30.6 | 34.6 | 24.4 | 26.7 | 73 | 92 | 43.5 | 6.7 | 100 | 62.5 | 56.4 | 49 | |
| Full KVBackbone=Llama-3.1-8B-Instruct, Budget=N/A2026.04 | 30.2 | 45.5 | 54.9 | 55.5 | 46.7 | 31.3 | 35.2 | 25.2 | 27.2 | 72.5 | 91.7 | 43.8 | 8.4 | 99.5 | 65.1 | 58.8 | 49.5 | |
| IceCacheBackbone=Llama-3.1-8B-Instruct, Budget=1282026.04 | 30 | 44.7 | 56.5 | 55 | 45.4 | 30 | 33.5 | 24.3 | 26.5 | 73 | 91.3 | 42.4 | 6.5 | 100 | 61.5 | 56.7 | 48.6 | |
| PQCacheBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 28.7 | 43.3 | 52.2 | 55.2 | 45.1 | 28.4 | 27.2 | 24 | 22.8 | 69.5 | 91.1 | 41.2 | 6.2 | 99 | 59 | 54.4 | 47.3 | |
| ThinKBackbone=Mistral-7B-Instruct-v0.3, Cache budget=1024L2026.05 | 27.97 | 38.68 | 49.96 | 49.97 | 38.69 | 25.28 | 28.05 | 24.57 | 25.75 | 74 | 88.72 | 45.83 | 6.5 | 95 | 61.67 | 61.85 | 46.41 | |
| SnapKV+ZipCacheBackbone=Mistral-7B-Instruct-v0.3, Cache budget=512L2026.05 | 27.57 | 35.17 | 48.23 | 51.43 | 36.7 | 25.4 | 28.66 | 24.37 | 25.61 | 75.5 | 89.01 | 46.09 | 6 | 95.5 | 60.34 | 60.7 | 46.02 | |
| IceCacheBackbone=Llama-3.1-8B-Instruct, Budget=642026.04 | 27.4 | 43.2 | 55.7 | 55.3 | 44.4 | 31.2 | 33.4 | 23.7 | 26.2 | 72.5 | 90.3 | 41.9 | 6.6 | 99.5 | 61.7 | 51.6 | 47.8 | |
| ThinKBackbone=Mistral-7B-Instruct-v0.3, Cache budget=512L2026.05 | 27.21 | 36.41 | 49.96 | 50.54 | 37.33 | 25.04 | 25.82 | 23.87 | 24.7 | 72.5 | 88.94 | 45.01 | 6 | 96 | 60.58 | 62.08 | 45.75 | |
| OmniKVBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 27.2 | 40.7 | 52.8 | 55.1 | 45.6 | 29.7 | 27.6 | 23 | 25.4 | 72.5 | 88.9 | 40.4 | 5.6 | 94.5 | 60.8 | 51.5 | 46.3 | |
| RDKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=1024L2026.05 | 27 | 37.96 | 49.88 | 50.56 | 37.63 | 28.2 | 32.33 | 25.05 | 26.57 | 76 | 88.48 | 47.88 | 5 | 97.5 | 59.35 | 60.4 | 46.86 | |
| Full KVBackbone=Mistral-7B-Instruct-v0.2, Budget=N/A2026.04 | 26.8 | 33.1 | 49.3 | 42.8 | 27.3 | 18.8 | 33 | 24.2 | 27.1 | 71 | 86.2 | 42.8 | 2.9 | 87 | 56.9 | 54.3 | 42.2 | |
| SnapKV+ZipCacheBackbone=Mistral-7B-Instruct-v0.3, Cache budget=1024L2026.05 | 26.41 | 37.75 | 48.9 | 51.54 | 35.26 | 25.74 | 30.97 | 24.25 | 26.07 | 76 | 88.7 | 46.25 | 6.06 | 95.5 | 59.18 | 60.71 | 46.21 | |
| AdaKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=1024L2026.05 | 26.31 | 36.49 | 49.95 | 51.76 | 37.37 | 27.07 | 27.11 | 24.84 | 25.01 | 73 | 89.03 | 45.83 | 5.5 | 98.5 | 61.03 | 62.35 | 46.32 | |
| Top-kBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 26.2 | 31.3 | 48.9 | 40 | 26.2 | 19.6 | 33.3 | 24.2 | 27.1 | 73 | 86.6 | 43.3 | 2.3 | 82.9 | 59 | 57.3 | 42.6 | |
| ArkValeBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 26.1 | 35.2 | 47.5 | 51.6 | 45.6 | 28.1 | 22.9 | 22.5 | 22.8 | 53.5 | 90.1 | 40 | 6.7 | 85 | 57.7 | 48.9 | 42.8 | |
| RDKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=512L2026.05 | 26.06 | 37.26 | 49.62 | 50.07 | 36.58 | 26.99 | 30.48 | 25.14 | 26.23 | 75.5 | 89.08 | 46.77 | 5.5 | 96.5 | 59.14 | 60.58 | 46.34 | |
| SnapKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=1024L2026.05 | 26.05 | 35.63 | 50.05 | 50.39 | 36.44 | 27.4 | 27.24 | 23.97 | 25.22 | 73 | 89.19 | 45.84 | 6 | 98.5 | 61.42 | 61.55 | 46.12 | |
| AdaKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=512L2026.05 | 26.04 | 33.06 | 50.75 | 50.29 | 35.81 | 26.58 | 25.03 | 23.39 | 23.65 | 71.5 | 88.86 | 46.03 | 6 | 97.5 | 60.74 | 62.25 | 45.47 | |
| RDKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=256L2026.05 | 26 | 35.68 | 49.89 | 49.11 | 35.5 | 25.99 | 27.78 | 24.48 | 25.22 | 73.5 | 88.58 | 46.4 | 6 | 97 | 59.49 | 60.6 | 45.7 | |
| FullKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=Full2026.05 | 25.91 | 38.13 | 49.62 | 52.07 | 39.03 | 28.1 | 34.11 | 25.73 | 26.47 | 76 | 88.59 | 47.46 | 7.01 | 98 | 61.58 | 62.34 | 47.51 | |
| SnapKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=512L2026.05 | 25.89 | 32.99 | 48.59 | 50.39 | 36.72 | 26.24 | 24.86 | 23.81 | 24.17 | 69 | 89.28 | 45.39 | 6 | 97.5 | 60.21 | 61.03 | 45.13 | |
| SnapKV+ZipCacheBackbone=Mistral-7B-Instruct-v0.3, Cache budget=256L2026.05 | 25.85 | 36.05 | 48.22 | 49.31 | 36.59 | 24.75 | 27.34 | 23.59 | 25.26 | 71 | 88.75 | 45.37 | 5.5 | 94.5 | 59.76 | 59.88 | 45.11 | |
| RDKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=128L2026.05 | 25.61 | 33.76 | 48.58 | 48.65 | 36.19 | 25.37 | 25.39 | 24.14 | 24.14 | 70 | 89.18 | 45.52 | 5 | 97.5 | 59.3 | 59.69 | 44.88 | |
| MagicPigBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 25.5 | 39.9 | 51.8 | 51.4 | 39.5 | 25.7 | 33.9 | 23.5 | 25.9 | 65.5 | 84 | 37 | 7.8 | 99.5 | 47.4 | 44.8 | 44.6 | |
| IceCacheBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 25.2 | 31 | 48.4 | 40.4 | 26 | 18.3 | 31.5 | 23.1 | 26.9 | 71 | 86.3 | 42.8 | 3.9 | 85.1 | 54.7 | 53.2 | 41.7 | |
| IceCacheBackbone=Mistral-7B-Instruct-v0.2, Budget=1282026.04 | 25.1 | 30.5 | 48.7 | 40.4 | 26.7 | 18.7 | 30.3 | 22 | 26.1 | 70.5 | 85.7 | 42.4 | 3.4 | 70.2 | 53.6 | 51.5 | 40.4 | |
| ThinKBackbone=Mistral-7B-Instruct-v0.3, Cache budget=256L2026.05 | 24.82 | 31.21 | 50.22 | 50.79 | 35 | 23.7 | 23.75 | 23.55 | 23.17 | 66 | 89.61 | 43.9 | 5.5 | 94.5 | 59.37 | 59.04 | 44.01 | |
| AdaKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=256L2026.05 | 24.71 | 30.19 | 48.31 | 49.08 | 34.69 | 26.03 | 23.13 | 23.46 | 22.19 | 66 | 88.93 | 44.57 | 4.5 | 96.5 | 59.38 | 59.67 | 43.83 | |
| RDKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=64L2026.05 | 24.7 | 29.46 | 47.51 | 46.48 | 34.37 | 22.73 | 23.4 | 23.23 | 22.73 | 60 | 89.07 | 44.99 | 4.5 | 96 | 57.78 | 58.19 | 42.82 | |
| SnapKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=256L2026.05 | 24.37 | 29.08 | 47.68 | 48.14 | 34.66 | 24.45 | 23.17 | 23.24 | 22.37 | 57 | 88.85 | 44.36 | 6 | 96.5 | 59.32 | 57.98 | 42.95 | |
| IceCacheBackbone=Mistral-7B-Instruct-v0.2, Budget=642026.04 | 23.9 | 29 | 47.4 | 40.5 | 23.5 | 18.3 | 30.6 | 21.8 | 26 | 70.5 | 85.9 | 41.5 | 3.5 | 56.5 | 53.2 | 50.4 | 39 | |
| SnapKVBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 23.7 | 27.3 | 46.3 | 52.3 | 40.8 | 24.2 | 19.2 | 22.6 | 19.2 | 29 | 84.1 | 39 | 8.6 | 97.5 | 57.3 | 56.2 | 40.8 | |
| SnapKV+ZipCacheBackbone=Mistral-7B-Instruct-v0.3, Cache budget=128L2026.05 | 23.62 | 32.3 | 48.55 | 50.23 | 35.41 | 26.96 | 24.22 | 22.23 | 23.58 | 67.5 | 88.63 | 45.19 | 5.5 | 92 | 59.01 | 58.3 | 43.95 | |
| SnapKV+ZipCacheBackbone=Mistral-7B-Instruct-v0.3, Cache budget=64L2026.05 | 22.8 | 25.83 | 47.55 | 47.38 | 35.23 | 24.03 | 21.54 | 21.59 | 21.61 | 55 | 89.05 | 44.84 | 4 | 88 | 56.59 | 56.61 | 41.35 | |
| AdaKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=128L2026.05 | 21.97 | 23.31 | 46.28 | 46.67 | 34.52 | 21.47 | 20.92 | 21.85 | 19.71 | 49 | 89.19 | 42.63 | 6.5 | 91.5 | 56.41 | 55.6 | 40.47 | |
| ThinKBackbone=Mistral-7B-Instruct-v0.3, Cache budget=128L2026.05 | 21.9 | 29.83 | 48.32 | 47.25 | 34.76 | 23.27 | 21.73 | 22.37 | 21.18 | 47.5 | 88.36 | 43.1 | 5 | 93 | 57.62 | 56.39 | 41.35 | |
| SnapKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=128L2026.05 | 21.25 | 22.92 | 42.3 | 44.75 | 33.34 | 21.24 | 21 | 21.37 | 19.3 | 44.5 | 89.35 | 42.47 | 4.5 | 91 | 55.98 | 54.12 | 39.34 | |
| MagicPigBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 21 | 28 | 46.5 | 38.4 | 19.5 | 17.5 | 30.8 | 23.3 | 25.8 | 70 | 83.7 | 39.5 | 2.5 | 85 | 49.4 | 48.3 | 39.1 | |
| PQCacheBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 21 | 25.8 | 42.5 | 18.1 | 20.3 | 16 | 29.5 | 21.7 | 27.1 | 70 | 85.8 | 39.7 | 2.5 | 75.5 | 54.2 | 49.2 | 37.4 | |
| ThinKBackbone=Mistral-7B-Instruct-v0.3, Cache budget=64L2026.05 | 20.27 | 24.9 | 40.42 | 45 | 33.54 | 21.44 | 19.83 | 21.26 | 18.45 | 39.5 | 87.94 | 40.66 | 3 | 83 | 53.19 | 51.51 | 37.74 | |
| SnapKVBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 18.3 | 15.1 | 38.3 | 30.4 | 19.2 | 13.8 | 16.5 | 21.4 | 19.9 | 32 | 81.7 | 38.1 | 4 | 59.1 | 49 | 51.8 | 31.8 | |
| ArkValeBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 18 | 15.2 | 41.5 | 30.7 | 17 | 13.2 | 21.6 | 21.6 | 23.1 | 58 | 85.8 | 40.6 | 2.1 | 41.5 | 51.4 | 47.2 | 33 | |
| SnapKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=64L2026.05 | 17.53 | 17.78 | 32.36 | 40.94 | 30.84 | 19.34 | 16.27 | 20.33 | 13.77 | 38 | 87.64 | 38.8 | 5 | 76.5 | 48.46 | 47.72 | 34.45 | |
| AdaKVBackbone=Mistral-7B-Instruct-v0.3, Cache budget=64L2026.05 | 17.33 | 18.43 | 34.33 | 42.13 | 31.16 | 21.08 | 16.79 | 20.64 | 14.23 | 38 | 88.18 | 39.61 | 3.5 | 83 | 51.28 | 48.93 | 35.54 | |
| StreamingLLMBackbone=Llama-3.1-8B-Instruct, Budget=2562026.04 | 17 | 23.2 | 25.7 | 21 | 29.3 | 6.8 | 20.8 | 17.5 | 20.7 | 45.5 | 84.3 | 41.2 | 5 | 71.5 | 59.5 | 47.5 | 33.5 | |
| OmniKVBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 16.5 | 22.8 | 43.8 | 34.7 | 19.2 | 16.6 | 24.4 | 22 | 24.8 | 70.5 | 81.7 | 38.6 | 3.1 | 35.2 | 36.5 | 38.4 | 33 | |
| StreamingLLMBackbone=Mistral-7B-Instruct-v0.2, Budget=2562026.04 | 14.2 | 12.3 | 26.4 | 23.2 | 14.8 | 10.1 | 17.5 | 19.8 | 19.8 | 51 | 80.5 | 39.9 | 4 | 15.6 | 52 | 45.4 | 27.8 |