ReLoop-UME: Recurrent Depth with Learnable Retrieval Registers for Universal Multimodal Embedding
Authors: Shijie Wang, Xiangzhao Hao, Yueti Li, Guangyu Cao, Xinyu Tang, Haiyun Guo
Submitted: 30 July 2026
Subject: Computer Vision and Pattern Recognition (cs.CV)
arXiv ID: 2607.28751
Abstract
Universal multimodal embedding (UME) maps heterogeneous multimodal inputs into a shared embedding space, enabling cross-modal retrieval and understanding. Existing UME models typically form embeddings either through single-pass forward encoding or by augmenting computation with explicit rationale tokens and latent autoregressive states. While token expansion can improve complex matching, serial generation increases retrieval latency and makes the final embedding dependent on intermediate generated states. This raises a key question: can useful computation be expanded along model depth while keeping the token workspace fixed?
We analyze positive-negative similarity separation at every layer of independently trained UME models and observe a shared progression: early layers contextualize multimodal inputs, a contiguous middle-to-late stage forms retrieval-discriminative features, and the final layers map these features into the embedding space. Based on this finding, we propose ReLoop-UME, which executes the early layers once, recurrently reuses a parameter-shared retrieval-forming block, and applies the final mapping layers after the last loop. Learnable Retrieval Registers provide persistent retrieval-specific states that accumulate and exchange evidence across loops, with the final register serving as the embedding readout.
On MMEB-V2 and MRMR benchmarks, ReLoop-UME consistently improves retrieval across different backbones while running 44.9× faster than UME-R1 and 1.5× faster than PLUME. This work offers a new perspective on depth-wise computation allocation in multimodal embedding, with significant implications for real-time retrieval systems in 2026 and beyond.
via ArXiv CV