Long-context understanding on LongBench V1
31NQAOracle
Evaluation results
| Method | Links | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OracleBackbone Model=Mistral-7B-v0.3, Token Budget=2562026.02 | 31 | 40.2 | 52.4 | 50.3 | 39.4 | 28.8 | 34 | 25.74 | 27.2 | 76 | 89.4 | 5 | 59.3 | 61 | 47.9 | |
| OracleBackbone Model=Llama3.1-8B, Token Budget=2562026.02 | 30.3 | 44.5 | 55 | 54.9 | 44.6 | 32 | 34.8 | 25.1 | 26.9 | 72.5 | 91.5 | 7 | 63.3 | 57.4 | 48.7 | |
| FKVBackbone Model=Llama3.1-8B, Token Budget=2562026.02 | 30 | 45.3 | 55.6 | 55.8 | 43.7 | 30.2 | 35.1 | 25.4 | 27 | 72.5 | 91.7 | 7.1 | 63 | 56.3 | 48.7 | |
| FASABackbone Model=Mistral-7B-v0.3, FCs ratio=25%2026.02 | 29.9 | 42.3 | 53.7 | 51.1 | 39.1 | 28.7 | 34 | 24.8 | 28.2 | 76 | 89.4 | 5 | 57.8 | 58 | 47.8 | |
| FASABackbone Model=Llama3.1-8B, FCs ratio=25%2026.02 | 29.3 | 43.7 | 54.1 | 54.8 | 43.9 | 30.8 | 33.5 | 24.7 | 27 | 72 | 91.1 | 7.5 | 61.8 | 52.7 | 48.2 | |
| FKVBackbone Model=Mistral-7B-v0.3, Token Budget=2562026.02 | 29.1 | 41.6 | 52.9 | 49.4 | 39.5 | 29.1 | 34.8 | 25.7 | 27.8 | 76 | 88.6 | 5.5 | 58.4 | 59.7 | 47.4 | |
| FKVBackbone Model=Qwen2.5-14B-1M, Token Budget=2562026.02 | 28.7 | 46.2 | 53.8 | 65.2 | 64.5 | 43.6 | 43.5 | 23.3 | 22.7 | 80.5 | 89.5 | 11 | 32.3 | 37.5 | 50.3 | |
| OracleBackbone Model=Qwen2.5-14B-1M, Token Budget=2562026.02 | 28.5 | 46.3 | 54.3 | 64.3 | 63.6 | 44.7 | 31.5 | 22.9 | 22.7 | 81 | 88.4 | 10 | 33.6 | 39.7 | 49.4 | |
| FASABackbone Model=Qwen2.5-7B, FCs ratio=25%2026.02 | 28.3 | 43.8 | 51.9 | 57.4 | 46 | 30.1 | 31.2 | 22.8 | 24.3 | 72 | 89.4 | 8 | 60.3 | 64 | 47.9 | |
| SnapKVBackbone Model=Llama3.1-8B, Token Budget=2562026.02 | 27.5 | 34.5 | 51.6 | 52.3 | 44.3 | 28.3 | 23.9 | 24 | 22.7 | 62.5 | 90.9 | 7.5 | 60.1 | 52.6 | 45 | |
| FASABackbone Model=Qwen2.5-14B-1M, FCs ratio=25%2026.02 | 27.2 | 45.5 | 54.5 | 64.4 | 63.9 | 44.5 | 30.4 | 22.8 | 21.9 | 80 | 87.5 | 15.5 | 30.5 | 36.1 | 49.2 | |
| OracleBackbone Model=Llama3.2-3B, Token Budget=2562026.02 | 26.6 | 41.2 | 49.8 | 31.9 | 29.9 | 16.2 | 32.6 | 22.2 | 25 | 71.5 | 89.3 | 3.5 | 53.7 | 54.4 | 42.4 | |
| SnapKVBackbone Model=Qwen2.5-7B, Token Budget=2562026.02 | 26.6 | 36 | 50.8 | 55.6 | 43.8 | 26.5 | 21.9 | 21.9 | 19.3 | 58 | 86.2 | 8 | 55.6 | 60.6 | 42.6 | |
| SnapKVBackbone Model=Qwen2.5-14B-1M, Token Budget=2562026.02 | 26.3 | 40.5 | 51.2 | 63.2 | 62.2 | 43.3 | 22.5 | 22 | 18.3 | 63.5 | 87.5 | 11.5 | 30.4 | 36 | 45.9 | |
| FKVBackbone Model=Llama3.2-3B, Token Budget=2562026.02 | 26 | 40.7 | 50.4 | 32.2 | 29.6 | 15.1 | 33.5 | 22.9 | 25.3 | 71.5 | 88.9 | 3.5 | 52 | 54.2 | 42.2 | |
| FASABackbone Model=Llama3.2-3B, FCs ratio=25%2026.02 | 25.6 | 38.9 | 49.9 | 29.7 | 31.2 | 14.8 | 28 | 24.2 | 26.1 | 71.5 | 89.2 | 3.6 | 53.2 | 50.5 | 41.5 | |
| SnapKVBackbone Model=Mistral-7B-v0.3, Token Budget=2562026.02 | 25.5 | 32.6 | 53.7 | 48.4 | 37.3 | 25.9 | 22.7 | 23.6 | 23.1 | 62.5 | 89.4 | 6.5 | 57.3 | 57 | 44 | |
| OracleBackbone Model=Qwen2.5-7B, Token Budget=2562026.02 | 24.4 | 43 | 52.3 | 57.8 | 46.9 | 30.1 | 31.6 | 23.9 | 24.1 | 72.5 | 89.7 | 8 | 60.5 | 65.3 | 48.7 | |
| FKVBackbone Model=Qwen2.5-7B, Token Budget=2562026.02 | 24.2 | 43.5 | 52.1 | 55.9 | 46.9 | 28.6 | 31.8 | 23.1 | 23.9 | 71.5 | 89.3 | 7.5 | 60.2 | 66.5 | 47.8 | |
| SnapKVBackbone Model=Llama3.2-3B, Token Budget=2562026.02 | 23.5 | 28.9 | 45.6 | 17.7 | 22.9 | 11.8 | 21.7 | 20.9 | 21.1 | 61 | 88.5 | 3.5 | 50.7 | 48.6 | 37 | |
| StreamBackbone Model=Llama3.1-8B, Token Budget=2562026.02 | 21.9 | 23.4 | 31.8 | 45.1 | 36.7 | 24.3 | 20 | 21 | 19.3 | 45.5 | 87.9 | 6.9 | 59.4 | 49.1 | 38.8 | |
| StreamBackbone Model=Qwen2.5-14B-1M, Token Budget=2562026.02 | 19.6 | 26.9 | 29.4 | 46.5 | 48.3 | 29.6 | 17.8 | 18.4 | 15 | 46.5 | 82.5 | 12.5 | 28.7 | 31.2 | 35.3 | |
| StreamBackbone Model=Qwen2.5-7B, Token Budget=2562026.02 | 18.1 | 24.2 | 26.5 | 41.2 | 36.4 | 17.3 | 18.4 | 18.3 | 15.4 | 45 | 82.9 | 8.5 | 49.6 | 52.2 | 31.9 | |
| FlashAttention-2Category=Dense, Model=Llama-3.1-8B-Instruct, KV retention=10%2026.05 | 16.4 | 46.4 | 55 | 54.5 | 47.1 | 30.8 | — | — | — | — | — | — | — | — | 41.7 | |
| LOUVERCategory=Ours, Model=Llama-3.1-8B-Instruct, KV retention=10%2026.05 | 16.2 | 47.1 | 55.2 | 54.5 | 47.1 | 30.9 | — | — | — | — | — | — | — | — | 41.8 | |
| H2OCategory=Eviction, Model=Llama-3.1-8B-Instruct, KV retention=10%2026.05 | 15.9 | 43.4 | 53.7 | 54.3 | 47.2 | 29.1 | — | — | — | — | — | — | — | — | 40.6 | |
| QuestBackbone Model=Mistral-7B-v0.3, Token Budget=2562026.02 | 15.7 | 30.7 | 41 | 37.4 | 27.1 | 11.9 | 29.3 | 21.3 | 26.6 | 57 | 80.7 | 5 | 56.9 | 53 | 38.6 | |
| TwilightCategory=Adaptive, Model=Llama-3.1-8B-Instruct, KV retention=10%2026.05 | 15.7 | 46.7 | 55 | 54.5 | 48.2 | 29.7 | — | — | — | — | — | — | — | — | 41.6 | |
| QuestCategory=Retrieval, Model=Llama-3.1-8B-Instruct, KV retention=10%2026.05 | 15.4 | 47.1 | 54.3 | 53.5 | 47.5 | 28.7 | — | — | — | — | — | — | — | — | 41.1 | |
| StreamingLLMCategory=Eviction, Model=Llama-3.1-8B-Instruct, KV retention=10%2026.05 | 15.1 | 26.6 | 30.7 | 45.6 | 41.1 | 18.6 | — | — | — | — | — | — | — | — | 29.6 | |
| QuestBackbone Model=Qwen2.5-14B-1M, Token Budget=2562026.02 | 14.5 | 31.9 | 39.1 | 38.8 | 36.6 | 16.2 | 16.2 | 20.1 | 25.2 | 43.5 | 72.7 | 10 | 35 | 34 | 34.9 | |
| QuestBackbone Model=Llama3.1-8B, Token Budget=2562026.02 | 13.7 | 33.1 | 38.4 | 35.8 | 32.2 | 12.8 | 26.5 | 20.9 | 26.7 | 38 | 65.6 | 3.8 | 52.5 | 45.7 | 35.4 | |
| StreamBackbone Model=Llama3.2-3B, Token Budget=2562026.02 | 13.2 | 19.7 | 23.6 | 18.1 | 22.7 | 7.8 | 18.2 | 17.9 | 17.9 | 49 | 83.7 | 3.5 | 49.3 | 45.9 | 31.8 | |
| StreamBackbone Model=Mistral-7B-v0.3, Token Budget=2562026.02 | 11.8 | 15.3 | 20.9 | 32.1 | 27.1 | 10.6 | 20.2 | 17.3 | 20.1 | 44.5 | 69 | 1.6 | 56.5 | 49.8 | 26.7 | |
| QuestBackbone Model=Qwen2.5-7B, Token Budget=2562026.02 | 9.1 | 24.5 | 30.4 | 24.7 | 24.1 | 8.8 | 26.8 | 19.9 | 24.4 | 41.8 | 66.7 | 4.4 | 46.5 | 42 | 31.4 | |
| QuestBackbone Model=Llama3.2-3B, Token Budget=2562026.02 | 8.7 | 19.5 | 23.6 | 12.9 | 15.9 | 6.5 | 23.3 | 18.1 | 25.1 | 34.5 | 52.9 | 6.5 | 53.7 | 43.6 | 25.5 |