Long-Context Retrieval on RULER 4K-32K context
95.73AccuracyManifoldKV
Evaluation results
| Method | Links | |
|---|---|---|
| ManifoldKVFramework=AdaKV, Comp.=0.20, Model=Llama-3.1-8B-Instruct2026.02 | 95.73 | |
| KeyDiffFramework=AdaKV, Comp.=0.20, Model=Llama-3.1-8B-Instruct2026.02 | 95.66 | |
| DuoAttentionFramework=Architectural/Calibrated, Comp.=0.20, Model=Llama-3.1-8B-Instruct, Note=Requires pre-computed calibration patterns2026.02 | 95.36 | |
| KeyDiffFramework=Standalone, Comp.=0.20, Model=Llama-3.1-8B-Instruct2026.02 | 92.93 | |
| SnapKVFramework=AdaKV, Comp.=0.20, Model=Llama-3.1-8B-Instruct2026.02 | 83.97 | |
| SnapKVFramework=Standalone, Comp.=0.20, Model=Llama-3.1-8B-Instruct2026.02 | 83.97 | |
| CriticalKV (Value-aware)Framework=Standalone, Comp.=0.20, Model=Llama-3.1-8B-Instruct2026.02 | 78.9 | |
| StreamingLLMFramework=Standalone, Comp.=0.20, Model=Llama-3.1-8B-Instruct2026.02 | 59.3 |