FEATUREDAI HOT (Curated Pool)· aihot-apiZH00:00 · 08·03
→UEmbed: One decoder-only model that outputs both sparse and dense multimodal embeddings
UEmbed is a decoder-only multimodal embedding model that produces both sparse lexical vectors and dense semantic vectors in a single causal forward pass. It appends N learnable special tokens and partitions the vocabulary into N disjoint subsets; each token predicts sparse weights over its assigned subset, and the N subsets are concatenated into the full sparse vector. The authors release UEmbed at 2B, 4B, and 9B scales, all trained on public data. UEmbed-9B hits 71.8 (dense) and 71.0 (sparse) on MMEB-v2, outperforming RzenEmbed, and stays competitive with strong baselines on BEIR. The paper also demonstrates utility across effectiveness, efficiency, and agentic applications. The post doesn't disclose inference latency or memory footprint, so real-world cost is still an open question.
#Embedding#Tingyu Song#Mingxin Li#Yanzhao Zhang
why featured
Featured · importance 72 · hook + knowledge
editor take
One model outputs both sparse and dense vectors in a single pass, but latency and memory aren't disclosed.
sharp
This caught my eye because it tackles a real engineering headache: most production RAG or search stacks run sparse retrieval (keyword-level, like BM25) and dense retrieval (semantic vectors) as two separate models, doubling inference cost. UEmbed's trick is neat—it appends N learnable tokens to the input, partitions the vocabulary into N chunks, and each token predicts sparse weights over its assigned chunk. Concatenate them and you get a full sparse vector, plus the dense vector, all in one causal forward pass.
The 9B model hits 71.8 dense and 71.0 sparse on MMEB-v2, beating RzenEmbed, and stays competitive on BEIR. They also demo it in agentic workflows, so it's not just a benchmark exercise.
I'd discount it a bit though. The paper doesn't disclose inference latency or memory footprint—those are the numbers that matter for real deployment. The vocabulary-splitting step for sparse generation could end up slower than running two small separate models, and we just don't know. Training data is described as public but the exact composition isn't spelled out, which makes reproduction and fair comparison tricky. Treat this as a promising architectural direction, not a drop-in production tool yet.
HKR breakdown
hook ✓knowledge ✓resonance —