Candidate-ranking retrieval on 8192-token synthetic probes Weighted average
50.6AccuracySWA+Full
Evaluation Results
| Method | Links | |
|---|---|---|
| SWA+FullBackbone=Qwen2.5-7B, N (Sample Size)=8322026.06 | 50.6 | |
| PBB+FullBackbone=Qwen2.5-7B, N (Sample Size)=8322026.06 | 50.2 | |
| SE-Bridge+FullBackbone=Qwen2.5-7B, N (Sample Size)=8322026.06 | 49.4 | |
| Block+FullBackbone=Qwen2.5-7B, N (Sample Size)=8322026.06 | 38.6 |