ReLoop-UME：基于可学习检索寄存器的循环深度通用多模态嵌入
原标题：ReLoop-UME: Recurrent Depth with Learnable Retrieval Registers for Universal Multimodal Embedding
AI 摘要
arXiv 上发表了一篇关于通用多模态嵌入（UME）的论文，提出 ReLoop-UME 模型。该模型通过循环复用参数共享的检索形成块，并引入可学习的检索寄存器，在不增加 token 工作区的情况下沿模型深度扩展计算。实验表明，ReLoop-UME 在 MMEB-V2 和 MRMR 基准上持续提升检索性能，且运行速度比 UME-R1 快 44.9 倍，比 PLUME 快 1.5 倍。
正文节选
Computer Science > Computer Vision and Pattern Recognition Title:ReLoop-UME: Recurrent Depth with Learnable Retrieval Registers for Universal Multimodal Embedding View PDF HTML (experimental) Abstract:Universal multimodal embedding (UME) maps heterogeneous multimodal inputs into a shared embedding space. Existing UME models either form embeddings through single forward encoding or add computation through explicit rationale tokens and latent autoregressive states. Although token expan