Long-context retrieval on NarrativeQA LongEmbed
71.5nDCG@10BM25
Evaluation results
| Method | Links | |
|---|---|---|
| BM25MLDR Train=false2026.06 | 71.5 | |
| MaxPMLDR Train=true, Backbone=Qwen3-Embedding-0.6B, Inference chunk size=642026.06 | 60.4 | |
| MPE Fixed-64MLDR Train=true, Backbone=Qwen3-Embedding-0.6B, Inference chunk size=642026.06 | 58 | |
| MPE-Rand[32, 1024]MLDR Train=true, Backbone=Qwen3-Embedding-0.6B, Inference chunk size=642026.06 | 57.9 | |
| MaxP-TrainMLDR Train=true, Backbone=Qwen3-Embedding-0.6B, Inference chunk size=642026.06 | 55.7 | |
| BGE-M3mvMLDR Train=true, Retrieval Style=ColBERT-style multi-vector2026.06 | 55.4 | |
| Single-vectorMLDR Train=true, Backbone=Qwen3-Embedding-0.6B2026.06 | 54.3 | |
| BGE-M3MLDR Train=true2026.06 | 48.7 | |
| E5-Mistral-7BMLDR Train=false2026.06 | 44.9 | |
| OpenAI-Ada-0022026.06 | 41.1 | |
| jina-v2-base-enMLDR Train=false2026.06 | 39.4 |