Information Retrieval on NIAH (test)
99.6Average ScoreOriginal*
Evaluation Results
| Method | Links | |
|---|---|---|
| Original*SWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 99.6 | |
| Original*SWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 99.4 | |
| BOSCHSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 99.2 | |
| Original*SWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 99.1 | |
| APLSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 99 | |
| RAZORSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 98.9 | |
| BOSCHSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 98.9 | |
| FISHERSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 98.8 | |
| APLSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 98 | |
| PROXYSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 97.8 | |
| BOSCHSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 97.5 | |
| PROXYSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 97.3 | |
| DCAMSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 96.9 | |
| DCAMSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 95.8 | |
| QADASWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 95.2 | |
| FISHERSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 94.2 | |
| RAZORSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 94.1 | |
| BOSCHSWA ρ ratio=0.5, Model size=14B, Setting=Zero-shot2026.04 | 94 | |
| FISHERSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 92.8 | |
| Original*SWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 92.7 | |
| DCAMSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 92.7 | |
| BOSCHSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 91.8 | |
| BOSCHSWA ρ ratio=0.5, Model size=8B, Setting=Zero-shot2026.04 | 90.3 | |
| BOSCH-layerSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 90 | |
| BOSCH-multiSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 89.5 | |
| APLSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 89.1 | |
| PROXYSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 88.3 | |
| FISHERSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 87.2 | |
| BOSCH-singleSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 86.9 | |
| BOSCHSWA ρ ratio=0.5, Model size=30B, Setting=Zero-shot2026.04 | 86.3 | |
| RAZORSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 85.5 | |
| RAZORSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 85 | |
| QADASWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 83.9 | |
| BOSCHSWA ρ ratio=0.75, Model size=14B, Setting=Zero-shot2026.04 | 83.6 | |
| QADASWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 83.5 | |
| QADASWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 80.8 | |
| DCAMSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 80.7 | |
| RANDSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 79.6 | |
| RANDSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 79 | |
| BOSCHSWA ρ ratio=0.5, Model size=1.7B, Setting=Zero-shot2026.04 | 78.3 | |
| INTRSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 74.9 | |
| INTRSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 72.9 | |
| BOSCHSWA ρ ratio=0.75, Model size=8B, Setting=Zero-shot2026.04 | 72.7 | |
| INTRSWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 69.1 | |
| RANDSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 66.9 | |
| BOSCHSWA ρ ratio=0.75, Model size=1.7B, Setting=Zero-shot2026.04 | 58 | |
| PROXYSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 50.9 | |
| BOSCHSWA ρ ratio=0.75, Model size=30B, Setting=Zero-shot2026.04 | 50.2 | |
| BMESWA ρ ratio=0.25, Model size=30B, Setting=Zero-shot2026.04 | 47.5 | |
| BOSCHSWA ρ ratio=0.875, Model size=14B, Setting=Zero-shot2026.04 | 47.2 | |
| RANDSWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 45.9 | |
| BOSCHSWA ρ ratio=0.875, Model size=8B, Setting=Zero-shot2026.04 | 42.5 | |
| BMESWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 41.3 | |
| INTRSWA ρ ratio=0.25, Model size=14B, Setting=Zero-shot2026.04 | 41.1 | |
| BMESWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 40 | |
| BMESWA ρ ratio=0.25, Model size=8B, Setting=Zero-shot2026.04 | 30.8 | |
| BOSCHSWA ρ ratio=0.875, Model size=1.7B, Setting=Zero-shot2026.04 | 30.3 | |
| BOSCHSWA ρ ratio=0.875, Model size=30B, Setting=Zero-shot2026.04 | 26.9 | |
| APLSWA ρ ratio=0.25, Model size=1.7B, Setting=Zero-shot2026.04 | 22.9 |