Candidate-ranking retrieval on 8192-token synthetic probes Semantic paraphrase+distractor
16.7AccuracySWA+Full
Evaluation Results
| Method | Links | |
|---|---|---|
| SWA+FullBackbone=Qwen2.5-7B, N (Sample Size)=482026.06 | 16.7 | |
| PBB+FullBackbone=Qwen2.5-7B, N (Sample Size)=482026.06 | 14.6 | |
| SE-Bridge+FullBackbone=Qwen2.5-7B, N (Sample Size)=482026.06 | 10.4 | |
| Block+FullBackbone=Qwen2.5-7B, N (Sample Size)=482026.06 | 0 |