Research map / Architectures and efficiency / RNNs and normalization
LEMUR: Learned Multi-Vector Retrieval
Elias Jääsaari, Ville Hyvönen, Teemu Roos
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LEMUR learned single-vector reduction instead of best baseline (MUVERA/IGP/DESSERT/PLAID) | multi-vector similarity search · BEIR datasets, ColBERTv2, k=100 | QPS at >=80% recall: 5-16x faster than the best-performing baseline | Improved |
| increase LEMUR hidden layer size d' (1024, 2048, 4096) (ablation) | LEMUR · BEIR, Recall100@k' | recall | Improved |
| increase d' from 2048 to 4096 (ablation) | LEMUR · BEIR end-to-end latency | end-to-end latency | No clear effect |
| increase d' from 1024 to 2048 (ablation) | LEMUR · BEIR end-to-end latency | latency | Improved |
| use 1024-dim LEMUR embeddings instead of 10240-dim MUVERA FDEs (ablation) | single-vector candidate retrieval · BEIR, Recall100@k' (seven of eight datasets) | recall | Improved |
| increase MLP depth beyond two layers (ablation) | LEMUR · MaxSim estimation | performance | No clear effect |
| use random (untrained) hidden-layer weights instead of trained feature encoder (ablation) | LEMUR · training ablation (Appendix H.3) | retrieval performance | Hurt |
| train on actual queries instead of corpus documents encoded with the query encoder (ablation) | LEMUR · training set selection | retrieval performance | Improved |
| use LEMUR instead of MUVERA on non-ColBERTv2 embeddings | multi-vector search · Quora, answerai-colbert-small, GTE-ModernColBERT, LFM2-ColBERT | recall: MUVERA fails to exceed 60% recall | Improved |
| use LEMUR on visual document retrieval models | multi-vector search · ViDoRe, ColModernVBERT and ColQwen2 | QPS vs recall | Improved |
Closest papers by meaning
- MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encoding (NeurIPS 2024 · similarity 0.91)
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.85)
- MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction (ICLR 2026 · similarity 0.84)
- Adaptive Retrieval and Scalable Indexing for k-NN Search with Cross-Encoders (ICLR 2024 · similarity 0.84)
- Your Embedding Model Is SMARTer Than You Think (NeurIPS 2026 · similarity 0.84)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
