Mixture-of-experts models improve efficiency of multimodal embeddings
MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models
Machine LearningArtificial Intelligence
Summary
Multimodal embeddings help computers understand and link different types of information like images and text together. Scaling these embeddings to handle many tasks usually means making models very large and slow. The authors propose MoEMB, a method that uses a mix of expert components to smartly grow the model’s capacity without slowing it down much. Their approach shows better performance than previous large models but with less computation. They also explore ways to make the model even more efficient by adapting how much computation it uses depending on the task.
multimodal embeddingmixture-of-experts (MoE)encodercontrastive learningscalabilityadaptive computationretrieval systemsrepresentation sizelatent reasoninglarge-scale recommendation
Authors
Xuanming Cui, Shlok Kumar Mishra, Wentao Bao, Aashu Singh, Zihao Wang, Xiangjun Fan, Jun Xiao, Ser-Nam Lim, Jianpeng Cheng
Abstract
Universal multimodal embedding (UME) increasingly demands encoder's capacity for handling a broad range of tasks and modalities with increased complexity. Prior scaling methods either increase the representation size, retrieval effort, or scales the encoder into a heavy multimodal LLM. Recent works, such as Think-Then-Embed (TTE), explore scaling via reasoning tokens. However, embedding models are hard to scale up: increasing parameters directly tradeoffs for the large training batch size that contrastive learning needs, and retrieval has to be served under tight latency. Moreover, UME tasks are diverse in complexity, where scaling up embedders can bring significant redundant computation. In this work, we propose MOEMB, which instead scales UME along the expert axis through mixture-of-experts (MoE), growing encoder capacity while preserving single-vector, non-autoregressive encoding. Through a systematic study of the design space and training recipes for MoE-based UME, MoEMB sets a new state of the art on both MMEB-V2 and MRMR among models trained on public MMEB-family data: with only 3B active parameters, MoEMB surpasses TTE-based methods with >4x active parameters, using significantly less computes. To further improve the scalability and efficiency, we conduct the first comprehensive study of adaptive computation for MoE-based embedding, spanning diverse strategies across training-based and inference-only methods. Together, these results support expert scaling as an effective and efficient direction for UME, with adaptive computation further improving efficiency for MLLM-based embedding models towards large-scale retrieval and recommendation systems.