SAIL-Embedding Technical Report: Omni-modal Embedding Foundation Model

  • 2025-10-14 16:43:22
  • Lin Lin, Jiefeng Long, Zhihe Wan, Yuchi Wang, Dingkang Yang, Shuang Yang, Yueyang Yao, Xu Chen, Zirui Guo, Shengqiang Li, Weiran Li, Hanyu Li, Yaling Mou, Yan Qiu, Haiyang Yu, Xiao Liang, Hongsheng Li, Chao Feng
  • 0

Abstract

Multimodal embedding models aim to yield informative unified representationsthat empower diverse cross-modal tasks. Despite promising developments in theevolution from CLIP-based dual-tower architectures to large vision-languagemodels, prior works still face unavoidable challenges in real-worldapplications and business scenarios, such as the limited modality support,unstable training mechanisms, and industrial domain gaps. In this work, weintroduce SAIL-Embedding, an omni-modal embedding foundation model thataddresses these issues through tailored training strategies and architecturaldesign. In the optimization procedure, we propose a multi-stage training schemeto boost the multifaceted effectiveness of representation learning.Specifically, the content-aware progressive training aims to enhance themodel's adaptability to diverse downstream tasks and master enrichedcross-modal proficiency. The collaboration-aware recommendation enhancementtraining further adapts multimodal representations for recommendation scenariosby distilling knowledge from sequence-to-item and ID-to-item embeddings whilemining user historical interests. Concurrently, we develop the stochasticspecialization and dataset-driven pattern matching to strengthen model trainingflexibility and generalizability. Experimental results show that SAIL-Embeddingachieves SOTA performance compared to other methods in different retrievaltasks. In online experiments across various real-world scenarios integratedwith our model, we observe a significant increase in Lifetime (LT), which is acrucial indicator for the recommendation experience. For instance, the modeldelivers the 7-day LT gain of +0.158% and the 14-day LT gain of +0.144% in theDouyin-Selected scenario. For the Douyin feed rank model, the match featuresproduced by SAIL-Embedding yield a +0.08% AUC gain.

 

Quick Read (beta)

loading the full paper ...