Multi-objective Retrieval on NeedleBench M-RT (32k context stress slice)
100Slot-set ExactnessFullKV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FullKVBackbone=Qwen3-8B [39], Context=32k2026.05 | 100 | 37.5 | |
| FullKVBackbone=Llama-3.1-8B [43], Context=32k2026.05 | 100 | 75 | |
| FullKVContract Type=reference2026.05 | 95.8 | 37.5 | |
| FullKVBackbone=Average, Context=32k2026.05 | 95.8 | 37.5 | |
| FullKVBackbone=Mistral-7B [44], Context=32k2026.05 | 87.5 | 0 | |
| MII-QDBackbone=Qwen3-8B [39], Context=32k2026.05 | 75 | 50 | |
| MII-QDBackbone=Llama-3.1-8B [43], Context=32k2026.05 | 75 | 62.5 | |
| MII-QDContract Type=question-routed2026.05 | 72.2 | 54.2 | |
| MII-QDBackbone=Average, Context=32k2026.05 | 72.2 | 54.2 | |
| MII-pooledBackbone=Qwen3-8B [39], Context=32k2026.05 | 66.7 | 37.5 | |
| SnapKVBackbone=Llama-3.1-8B [43], Context=32k2026.05 | 66.7 | 50 | |
| MII-QDBackbone=Mistral-7B [44], Context=32k2026.05 | 66.7 | 50 | |
| SnapKVBackbone=Qwen3-8B [39], Context=32k2026.05 | 62.5 | 25 | |
| SnapKVContract Type=fixed contract2026.05 | 61.1 | 25 | |
| SnapKVBackbone=Average, Context=32k2026.05 | 61.1 | 25 | |
| SnapKVBackbone=Mistral-7B [44], Context=32k2026.05 | 54.2 | 0 | |
| MII-pooledContract Type=pooled proxy2026.05 | 52.8 | 25 | |
| MII-pooledBackbone=Average, Context=32k2026.05 | 52.8 | 25 | |
| MII-pooledBackbone=Llama-3.1-8B [43], Context=32k2026.05 | 50 | 37.5 | |
| MII-pooledBackbone=Mistral-7B [44], Context=32k2026.05 | 41.7 | 0 |