Information Retrieval on LongMemEval-S (challenge subset n=150)
59.4Recall@1RRF(BM25, late)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| RRF(BM25, late)Encoder=all-MiniLM-L6-v2, Fusion=Reciprocal Rank Fusion2026.06 | 59.4 | 95.6 | 93.2 | |
| BM25 ⊕ lateEncoder=all-MiniLM-L6-v2, Fusion=Weighted/Late Interaction Fusion2026.06 | 59.4 | 95.9 | 93 | |
| BM25Encoder=all-MiniLM-L6-v22026.06 | 58.9 | 94.8 | 91.6 | |
| Dense, late (top-3)Encoder=all-MiniLM-L6-v2, Mechanism=top-32026.06 | 56.9 | 95.1 | 90.8 | |
| Dense, late (max-sim)Encoder=all-MiniLM-L6-v2, Mechanism=max-sim2026.06 | 56.8 | 95.3 | 90.6 | |
| Dense, early (mean)Encoder=all-MiniLM-L6-v2, Pooling=mean2026.06 | 54.5 | 92 | 87.6 |