Retrieval on NIAH seq=4096
100AccuracySpectrumKV Greedy 3-tier
Evaluation Results
| Method | Links | |
|---|---|---|
| SpectrumKV Greedy 3-tierModel=Qwen2.5-7B, Budget b=0.52026.06 | 100 | |
| SpectrumKV Greedy 3-tierModel=Qwen2.5-7B, Budget b=0.72026.06 | 100 | |
| SpectrumKV Adaptive 2-tierModel=Qwen2.5-7B, Budget b=0.52026.06 | 100 | |
| SpectrumKV Adaptive 2-tierModel=Qwen2.5-7B, Budget b=0.72026.06 | 100 | |
| SpectrumKV Adaptive verificationModel=Mistral-7B, Budget b=0.32026.06 | 100 | |
| SpectrumKV Adaptive verificationModel=Mistral-7B, Budget b=0.42026.06 | 100 | |
| SpectrumKV Adaptive verificationModel=Mistral-7B, Budget b=0.52026.06 | 100 | |
| SpectrumKV Adaptive verificationModel=Mistral-7B, Budget b=0.72026.06 | 100 | |
| SpectrumKV Greedy/SinkProtectModel=Gemma-2-9B, Budget b=0.32026.06 | 100 | |
| SpectrumKV Greedy/SinkProtectModel=Gemma-2-9B, Budget b=0.42026.06 | 100 | |
| SpectrumKV Greedy/SinkProtectModel=Gemma-2-9B, Budget b=0.52026.06 | 100 | |
| SpectrumKV Greedy/SinkProtectModel=Gemma-2-9B, Budget b=0.72026.06 | 100 | |
| SpectrumKV Greedy rawModel=Mistral-7B, Budget b=0.32026.06 | 89.5 | |
| SpectrumKV Greedy rawModel=Mistral-7B, Budget b=0.42026.06 | 89.5 | |
| SpectrumKV Greedy rawModel=Mistral-7B, Budget b=0.52026.06 | 89.5 | |
| SpectrumKV Greedy rawModel=Mistral-7B, Budget b=0.72026.06 | 89.5 | |
| SpectrumKV Adaptive 2-tierModel=Qwen2.5-7B, Budget b=0.42026.06 | 74.7 | |
| PDTrimModel=Gemma-2-9B, Budget b=0.72026.06 | 74.7 | |
| SWSModel=Gemma-2-9B, Budget b=0.72026.06 | 72.6 | |
| PDTrimModel=Qwen2.5-7B, Budget b=0.72026.06 | 68.4 | |
| SWSModel=Qwen2.5-7B, Budget b=0.72026.06 | 68.4 | |
| PDTrimModel=Mistral-7B, Budget b=0.72026.06 | 62.1 | |
| SWSModel=Mistral-7B, Budget b=0.72026.06 | 61.1 | |
| PDTrimModel=Gemma-2-9B, Budget b=0.52026.06 | 57.9 | |
| SpectrumKV Adaptive 2-tierModel=Qwen2.5-7B, Budget b=0.32026.06 | 52.6 | |
| SWSModel=Gemma-2-9B, Budget b=0.52026.06 | 52.6 | |
| PDTrimModel=Gemma-2-9B, Budget b=0.42026.06 | 49.5 | |
| PDTrimModel=Qwen2.5-7B, Budget b=0.52026.06 | 47.4 | |
| PDTrimModel=Mistral-7B, Budget b=0.52026.06 | 44.2 | |
| SWSModel=Gemma-2-9B, Budget b=0.42026.06 | 44.2 | |
| SWSModel=Qwen2.5-7B, Budget b=0.52026.06 | 43.2 | |
| PDTrimModel=Gemma-2-9B, Budget b=0.32026.06 | 41.1 | |
| SWSModel=Mistral-7B, Budget b=0.52026.06 | 37.9 | |
| PDTrimModel=Qwen2.5-7B, Budget b=0.42026.06 | 36.8 | |
| SWSModel=Gemma-2-9B, Budget b=0.32026.06 | 35.8 | |
| PDTrimModel=Mistral-7B, Budget b=0.42026.06 | 34.7 | |
| SWSModel=Qwen2.5-7B, Budget b=0.42026.06 | 31.6 | |
| SWSModel=Mistral-7B, Budget b=0.42026.06 | 29.5 | |
| PDTrimModel=Qwen2.5-7B, Budget b=0.32026.06 | 26.3 | |
| PDTrimModel=Mistral-7B, Budget b=0.32026.06 | 26.3 | |
| SWSModel=Qwen2.5-7B, Budget b=0.32026.06 | 21.1 | |
| SWSModel=Mistral-7B, Budget b=0.32026.06 | 20 | |
| SpectrumKV Greedy 3-tierModel=Qwen2.5-7B, Budget b=0.32026.06 | 4.2 | |
| SpectrumKV Greedy 3-tierModel=Qwen2.5-7B, Budget b=0.42026.06 | 0 |