ReLoop-UME boosts multimodal AI 44.9x with new embedding method

A team led by Shijie Wang and Xiangzhao Hao from Tsinghua University has unveiled ReLoop-UME, a novel universal multimodal embedding (UME) model that dramatically improves efficiency and performance in cross-modal retrieval tasks. Published on arXiv (arXiv:2607.28751), the paper proposes a recurrent depth architecture with learnable retrieval registers, enabling persistent retrieval-specific states that accumulate evidence across loops without expanding token workspace. This design contrasts with prior UME systems that either use single-pass encoding or rely on serial generation with intermediate states—both of which increase latency and dependency on unstable intermediate outputs.

The researchers analyzed the internal dynamics of UME models and observed a consistent three-stage progression: early layers contextualize inputs, a middle-to-late stage forms retrieval-discriminative features, and final layers map these into a shared embedding space. ReLoop-UME exploits this structure by processing early layers once, then recurrently applying a parameter-shared retrieval-forming block with learnable registers that act as persistent memory states. The final register output serves as the embedding readout. On MMEB-V2 and MRMR benchmarks, ReLoop-UME delivers consistent retrieval improvements across multiple backbones while achieving a 44.9x speedup over UME-R1 and 1.5x faster than PLUME.