[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-86070-en":3,"doc-seo-86070-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},86070,687197207057,"Sage","https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0",8,"Research & Report","Stream-aware Side Adaptation for Large Pre-trained Multimodal Embedding Models in Sequential Recommendation","Large pretrained multimodal embedding models such as Qwen3-VL Embedding have strong potential for sequential recommendation by providing reusable semantic item representations across modalities and domains. Directly using such embeddings often underperforms due to domain misalignment, while efficient adaptation is difficult because existing side adapters degrade with depth. The work proposes Stresa, a frozen-backbone stream-aware side-adaptation framework using SHAF and ReSA for effective fusion and selective residual updates, achieving consistent gains over baselines on public datasets.","Stream-aware Side Adaptation for Large Pre-trained Multimodal Embedding Models in Sequential Recommendation  \nJunchen Fu  \nUniversity of Glasgow United Kingdom [j.fu.3@research.gla.ac.uk](j.fu.3@research.gla.ac.uk)  \nKaiwen Zheng  \nUniversity of Glasgow United Kingdom [k.zheng.1@research.gla.ac.uk](k.zheng.1@research.gla.ac.uk)  \nIoannis Arapakis  \nTelefónica Scientific Research, Telefónica Innovación Digital Barcelona, Spain [arapakis.ioannis@gmail.com](arapakis.ioannis@gmail.com)  \nWenhao Deng  \nUniversity of Glasgow United Kingdom [w.deng.1@research.gla.ac.uk](w.deng.1@research.gla.ac.uk)  \nXin Xin  \nShandong University Qingdao, China [xinxin@sdu.edu.cn](xinxin@sdu.edu.cn)  \nJoemon M. Jose  \nUniversity of Glasgow Glasgow, United Kingdom [joemon.jose@glasgow.ac.uk](joemon.jose@glasgow.ac.uk)  \narXiv :2607 . 10909v 1 [ cs .IR] 12 Jul 2026  \nXuri Ge∗ Shandong University Jinan, China [xuri.ge@sdu.edu.cn](xuri.ge@sdu.edu.cn)  \nAbstract  \nRecently, large pretrained multimodal embedding models such as Qwen3-VL Embedding have shown strong promise for sequential recommendation, as they provide reusable semantic item representations across modalities and domains. However, directly using these embeddings often leads to suboptimal performance because of domain misalignment. Efficient side adaptation is therefore an attractive solution. Although adapting all backbone layers should help, existing side adapters often degrade with depth, prompting layer dropping despite the loss of useful hidden states. This is due to two major challenges: (1) the lack of modeling in selecting fused representations during residual addition, and (2) the insufficient preservation of earlier representations during progressive sigmoid fusion. This paper therefore asks a practical question: How can we design a side adaptation approach that effectively unlocks the potential of large pre-trained multimodal embedding models?  \nTo address this question, we propose Stresa, a stream-aware side-adaptation framework for frozen large pre-trained multimodal embedding models in sequential recommendation. Stresa introduces Stream-aware Hidden-Adapter Fusion (SHAF) to preserve historical side memory during fusion and Residual Stream Adapter (ReSA) to produce selective residual updates across layers. Empirically, Stresa consistently outperforms standard side adapters and stateof-the-art baselines on public datasets across multiple backbone embedding models. These results highlight the promise of adapting large embedding models for sequential recommendation. Our code is publicly available at [https://github.com/GAIR-Lab/Stresa](https://github.com/GAIR-Lab/Stresa).  \nCCS Concepts  \n• Information systems → Recommender systems.  \nKeywords  \nMultimodal Recommendation, Sequential Recommendation, Large Embedding Models, Side Adapter  \n∗ Corresponding Author.  \n| Method | Residual |  |  |  | Fusion |\n| --- | --- | --- | --- | --- | --- |\n| Stresa |  | Selective residual\u003Cbr>update ( ReSA) |  |  | Earlier Representation Preserved Fusion (SHAF) |\n| Side Adapter |  | Direct Addition\u003Cbr>(Residual) |  |  | Direct Sigmoid Fusion\u003Cbr>(Sigmoid) |\n\nFigure 1: Conventional Side Adapter vs. Stresa. Using the frozen Qwen3-VL-8B-Embedding backbone, Dropped 18 layers and Full 36 layers mean that the item encoder extracts hidden states from 18 or 36 backbone layers for adapter computation. Left: HR@10 under both settings. Right: Side Adapter uses residual addition and sigmoid fusion, whereas Stresa uses selective residual update (ReSA) and earlierrepresentation-preserved fusion (SHAF). HSP and HSA denote hidden states from the pretrained model and previous adapter layer, respectively.  \n1 Introduction  \nLarge-scale pretrained multimodal embedding models are increasingly becoming reusable infrastructure for retrieval, ranking, and recommendation, as they project heterogeneous item content into transferable semantic spaces [14, 19, 23, 35, 40] . A recent stateof-the-art example is Qwen3-VL-Embedding, w","cbCaivjqSbgMRWBn","https://ap.wps.com/l/cbCaivjqSbgMRWBn","pdf",1872501,4,1,10,"English","en",105,"# Abstract\n# Introduction\n## Background and Motivation\n## Challenges of Direct Use and Practical Constraints\n# Method: Stresa\n## Stream-aware Hidden-Adapter Fusion (SHAF)\n## Residual Stream Adapter (ReSA)\n# Experiments and Results\n## Empirical Performance on Public Datasets","[{\"question\":\"Why do large pretrained multimodal embedding models often perform suboptimally in sequential recommendation?\",\"answer\":\"Direct application can cause domain misalignment because these models target general semantic representation rather than recommendation-specific objectives like user transitions and relevance.\"},{\"question\":\"What problem do existing side adapters face when adapting to deeper layers?\",\"answer\":\"Side adapters often degrade with depth, which leads to layer dropping and can discard useful hidden states.\"},{\"question\":\"How does Stresa improve side adaptation while keeping the backbone frozen?\",\"answer\":\"Stresa introduces SHAF to preserve historical side memory during fusion and ReSA to generate selective residual updates across layers, enabling better use of pretrained representations.\"}]",1784208308,25,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"stream-aware-side-adaptation-for-large-pre-trained-multimodal-embedding-models-in-sequential-recommendation","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":20},"https://docshare.wps.com/document/stream-aware-side-adaptation-for-large-pre-trained-multimodal-embedding-models-in-sequential-recommendation/86070/",{"url":52,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-26","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why do large pretrained multimodal embedding models often perform suboptimally in sequential recommendation?","Question",{"text":75,"@type":76},"Direct application can cause domain misalignment because these models target general semantic representation rather than recommendation-specific objectives like user transitions and relevance.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What problem do existing side adapters face when adapting to deeper layers?",{"text":80,"@type":76},"Side adapters often degrade with depth, which leads to layer dropping and can discard useful hidden states.",{"name":82,"@type":73,"acceptedAnswer":83},"How does Stresa improve side adaptation while keeping the backbone frozen?",{"text":84,"@type":76},"Stresa introduces SHAF to preserve historical side memory during fusion and ReSA to generate selective residual updates across layers, enabling better use of pretrained representations.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,134],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":22,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":22,"slug":133},"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]