Megadose AI progress, ranked and analyzed.

UEmbed: Unified Sparse and Dense Multimodal Embeddings

· HF Daily Papers ·
UEmbed produces sparse lexical and dense multimodal embeddings from a single decoder-only model pass.

The paper says it adds learnable special tokens, splits the vocabulary into subsets, and uses causal hidden states to build a full sparse vector alongside dense representations. The authors release 2B, 4B, and 9B versions trained on public data. UEmbed-9B scores 71.8 dense and 71.0 sparse on MMEB-v2, and is reported as competitive on BEIR. HF Daily Papers' note

score 5

Categories: Research