Candidate-ranking Retrieval on Diagnostic Probes 1024-token context
96.7Needle AccuracyFull
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| FullArchitecture=Full2026.06 | 96.7 | 98.9 | 98.5 | 90.6 | 94.6 | 31.4 | 54.9 | |
| SWAArchitecture=SWA2026.06 | 51.3 | 99.1 | 54.1 | 44.4 | 49.3 | 28.2 | 27.8 | |
| PBBArchitecture=PBB2026.06 | 43.8 | 84.2 | 48.1 | 42 | 45.1 | 23.7 | 29.2 | |
| BridgeArchitecture=Bridge2026.06 | 43.7 | 81.2 | 46.2 | 37.8 | 42 | 23.5 | 27.6 | |
| SE-BridgeArchitecture=SE-Bridge2026.06 | 43.6 | 98.3 | 56 | 47.1 | 51.6 | 22.3 | 29 | |
| BlockArchitecture=Block2026.06 | 42.9 | 41.5 | 44.9 | 33.4 | 39.2 | 20.4 | 27.1 |