MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interactions

Abstract

Universal multimodal embedding models have achieved great success in capturing semantic relevance between queries and candidates. However, current methods either condense queries and candidates into a single vector, potentially losing fine-grained details, or produce too many vectors that are prohibitive for multimodal retrieval. In this work, we introduce MetaEmbed, a new paradigm for multimodal retrieval that rethinks how multimodal embeddings are constructed and interacted with at scale. During training, a fixed number of learnable Meta Tokens is appended to the input sequence, and their last-layer contextualized representations serve as compact yet expressive multi- vector embeddings. Through the proposed Matryoshka Multi-Vector Retrieval training, MetaEmbed learns to organize information by granularity across multiple vectors. Upon that, we enable test-time scaling in multimodal retrieval, where users can balance retrieval quality against efficiency demands by selecting the number of tokens used for indexing and retrieval interactions. Extensive evaluations on Massive Multimodal Embedding Benchmark (MMEB) and Visual Document Retrieval Benchmark (ViDoRe) confirm that MetaEmbed achieves state-of-the-art retrieval performance while scaling robustly to large VLMs with 32B parameters.