Learned Sparse Retrieval
Properties
created
28.06.2026, 11:55
modified
28.06.2026, 11:57
published
Empty
sources
Empty
topics
Learned Sparse Retrieval, SPLADE, Sparse Embeddings, BGE M3
authors
Jakub
ai-assisted
No
Learned Sparse Retrieval (LSR) is an information retrieval paradigm that uses neural networks to map queries and documents into high-dimensional, sparse vectors. It combines the deep semantic understanding of neural embeddings with the efficiency and transparency of traditional keyword-based search.
# List
- SPLADE
- Transformer-based model that outputs sparse vectors (vocab-weighted) — like “semantic BM25”. Scores come from a neural model trained to activate relevant tokens ( Splade_PP_en_v1, Making Sparse Neural IR Models More Effective, SparseEmbed).
- SPLARE
- LSR approach from NAVER that replaces vocabulary-space projection with representations from pre-trained Sparse Autoencoders (SAEs) — making it monosemantic, language-agnostic, and scalable ( arxiv)
- Best LSR model on MMTEB (Multilingual, v2); uses frozen Llama Scope SAE + SPLADE-style max-pooling + LoRA fine-tuning
- BGE M3
- Has capabilities in Multi-Linguality, Multi-Functionality, and Multi-Granularity. Capable of supporting over 100 languages, BGE-M3 and supports dense retrieval, multi-vector retrieval, and sparse retrieval within a single framework makes it an ideal choice for a wide range of information retrieval (IR) applications ( Source).
- TODO: miniCOIL: on the Road to Usable Sparse Neural Retrieval