Your Embedding Model is SMARTer Than You Think

Jianrui Zhang, Hyun Jung Lee, Sukanta Ganguly, Tae-Eui Kam, Donghyun Kim, Yong Jae Lee

arXiv:2605.24938 · 2026-05-26 공개 · arXiv · PDF

post-training multimodal-retrieval multi-vector single-vector contrastive-training visual-document fine-grained-retrieval inference-enhancement

Abstract

Multimodal retrieval relies heavily on single-vector retrievers, which compress rich, sequential token sequences into one single global representation. While efficient, they discard fine-grained, local evidence critical for dense retrieval tasks. Multi-vector approaches were introduced as a solution, but they strictly require training and many ignore the necessity of a globally summarizing representation. To address this, we introduce SMART, a framework that unlocks the latent multi-vector capabilities of standard single-vector models. We first demonstrate that standard contrastive training on the pooled embedding implicitly shapes the retrieval geometry of preceding hidden states via gradient flow. By applying direct late-interaction over these frozen hidden states during inference, SMART acts as a plug-and-play upgrade that consistently improves performance across diverse modalities, improving even the state-of-the-art models further on MMEB-V2. We also reveal SMART's superior performance, as simple lightweight post-training not only saves time and compute, but also brings forth further improvement on Visual Document retrieval, allowing a single-vector model to outperform SoTA multi-vector counterparts. Ultimately, SMART offers both a highly efficient inference enhancement and a powerful finetuning technique for multimodal retrieval. We open source our code and weights at https://github.com/HanSolo9682/SMART.

한국어 요약

📋 한 줄 요약

**[Multimodal Retrieval / Late Interaction]** SMART가 단일 벡터 모델의 hidden state에 frozen late-interaction을 적용 — 학습 없이 plug-and-play 향상, MMEB-V2 SOTA 추가 향상·visual document retrieval에서 단일 벡터가 multi-vector SOTA 능가.

🎯 핵심 기여도

💡 핵심 아이디어

Single-vector contrastive 학습이 이미 preceding hidden state에 multi-vector retrieval geometry를 implicitly shaping하며, 별도 학습 없이 frozen hidden state에 late-interaction을 적용하는 plug-and-play 인퍼런스만으로 multi-vector 능력을 unlock해 단일 벡터 모델이 multi-vector SOTA를 능가할 수 있다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: 단일 벡터 모델의 잠재 multi-vector 능력에 대한 통찰, 학습 없는 plug-and-play 효율, multi-vector SOTA 능가로 효율·품질 양립, MMEB-V2 SOTA 추가 향상의 일반화 가능성. **한계**: Hidden state late-interaction의 메모리·연산 cost, post-training이 효과 극대화에 필요한 경우, 어떤 hidden layer를 사용할지의 선택 문제.

🚀 실용적 활용