[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-1-en-105":3,"doc-seo-237363-105":53,"doc-detail-237363-en":127},{"code":4,"msg":5,"data":6},0,"success",[7,14,19,24,29,34,39,44,49],{"id":8,"doc_module":9,"doc_module_name":10,"category_name":11,"show_sort_weight":12,"slug":13},11,1,"Template","Presentations",90,"presentations",{"id":15,"doc_module":9,"doc_module_name":10,"category_name":16,"show_sort_weight":17,"slug":18},12,"Resumes",80,"resumes",{"id":20,"doc_module":9,"doc_module_name":10,"category_name":21,"show_sort_weight":22,"slug":23},14,"Invoices",70,"invoices",{"id":25,"doc_module":9,"doc_module_name":10,"category_name":26,"show_sort_weight":27,"slug":28},15,"Posters",60,"posters",{"id":30,"doc_module":9,"doc_module_name":10,"category_name":31,"show_sort_weight":32,"slug":33},16,"Social Media",50,"social-media",{"id":35,"doc_module":9,"doc_module_name":10,"category_name":36,"show_sort_weight":37,"slug":38},17,"Forms",40,"forms",{"id":40,"doc_module":9,"doc_module_name":10,"category_name":41,"show_sort_weight":42,"slug":43},18,"Letters",30,"letters",{"id":45,"doc_module":9,"doc_module_name":10,"category_name":46,"show_sort_weight":47,"slug":48},21,"Paper Templates",5,"papers-templates",{"id":50,"doc_module":9,"doc_module_name":10,"category_name":51,"show_sort_weight":4,"slug":52},158,"General","general-158",{"code":4,"msg":54,"data":55},"ok",{"site_id":56,"language":57,"slug":58,"title":59,"keywords":60,"description":61,"schema_data":62,"social_meta":120,"head_meta":122,"extra_data":124,"updated_unix":126},105,"en","resum-unlocking-long-horizon-search-intelligence-via-context-summarization","ReSum - Unlocking Long-Horizon Search Intelligence via Context Summarization","","Large Language Model (LLM)-based web agents excel at knowledge-intensive tasks but confront a core tension: extensive exploration is required while limited context windows restrict what can be retained. Existing methods often modify architectures with internal memory tokens, reducing compatibility with preexisting agents and requiring costly retraining. ReSum addresses this via plug-and-play context summarization using an external tool, enabling unbounded exploration without training. To align reasoning over compressed contexts, ReSum-GRPO adapts GRPO through advantage broadcasting, improving long-horizon credit assignment and delivering gains over ReAct.",{"@graph":63,"@context":119},[64,80,102],{"@type":65,"itemListElement":66},"BreadcrumbList",[67,71,74,77],{"item":68,"name":69,"@type":70,"position":9},"https://docshare.wps.com","Home","ListItem",{"item":72,"name":10,"@type":70,"position":73},"https://docshare.wps.com/template/",2,{"item":75,"name":11,"@type":70,"position":76},"https://docshare.wps.com/template/presentations/",3,{"item":78,"name":59,"@type":70,"position":79},"https://docshare.wps.com/template/resum-unlocking-long-horizon-search-intelligence-via-context-summarization/237363/",4,{"url":78,"name":59,"@type":81,"image":82,"author":87,"headline":59,"publisher":90,"fileFormat":93,"inLanguage":57,"description":61,"dateModified":94,"datePublished":95,"encodingFormat":93,"isAccessibleForFree":96,"interactionStatistic":97},"DigitalDocument",{"url":83,"@type":84,"width":85,"height":86},"https://docshare.wps.com/thumbnails/resum-unlocking-long-horizon-search-intelligence-via-context-summarization/237363.png","ImageObject",442,249,{"name":88,"@type":89},"Jiven","Person",{"url":68,"name":91,"@type":92},"DocShare","Organization","application/pdf","2026-11-02","2026-09-11",true,{"@type":98,"interactionType":99,"userInteractionCount":101},"InteractionCounter",{"@type":100},"ViewAction",13,{"@type":103,"mainEntity":104},"FAQPage",[105,111,115],{"name":106,"@type":107,"acceptedAnswer":108},"What problem does ReSum address for LLM-based web agents?","Question",{"text":109,"@type":110},"ReSum targets the conflict between the need for extensive long-horizon exploration and the limited context windows of standard LLM agents.","Answer",{"name":112,"@type":107,"acceptedAnswer":113},"How does ReSum enable unbounded exploration without architectural retraining?",{"text":114,"@type":110},"ReSum periodically invokes an external tool to condense interaction histories into compact summaries, then resumes exploration from these compressed states.",{"name":116,"@type":107,"acceptedAnswer":117},"Why is ReSum-GRPO proposed, and how does it work?",{"text":118,"@type":110},"ReSum-GRPO improves alignment by adapting GRPO to the ReSum paradigm using advantage broadcasting across segmented trajectories, enabling better credit assignment over long horizons.","https://schema.org",{"og:url":78,"og:type":121,"og:title":59,"og:site_name":91,"og:description":61},"article",{"robots":123,"canonical":78},"index,follow",{"doc_id":125,"site_id":56},237363,1791749698,{"code":4,"msg":5,"data":128},{"doc_id":125,"user_id":129,"nickname":88,"user_avatar":130,"doc_module":9,"category_id":8,"category_name":11,"doc_title":59,"doc_description":61,"doc_content":131,"file_id":132,"file_url":133,"file_type":134,"file_size":135,"view_count":101,"is_deleted":4,"is_public":9,"is_downloadable":9,"audit_status":9,"page_count":136,"language":137,"language_code":57,"site_id":56,"html_lang":57,"table_of_contents":138,"faqs":139,"seo_title":140,"seo_description":61,"update_tm":141,"read_time":142},1099513958607,"https://ap-avatar.wpscdn.com/avatar/100002390cf8733938c?x-image-process=image/resize,m_fixed,w_180,h_180&k=1778829742770036399","ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization  \nXixi Wu * 1 2 Kuan Li * 2 3 Yida Zhao 2 Liwen Zhang 2 Litu Ou 2 Huifeng Yin 2 Zhongwang Zhang 2 Xinmiao Yu 2 Dingchu Zhang 2 Yong Jiang 2 Pengjun Xie 2 Fei Huang 2 Minhao Cheng 4 Shuai Wang 3  \nHong Cheng 1 Jingren Zhou 2  \narXiv :2509 . 13313v3 [ cs .CL] 26 Mar 2026  \nAbstract  \nLarge Language Model (LLM)-based web agents excel at knowledge-intensive tasks but face a fundamental conflict between the need for extensive exploration and the constraints of limited context windows. Current solutions typically rely on architectural modifications, e.g., internal memory tokens, which break compatibility with preexisting agents and necessitate costly end-to-end retraining. To overcome these limitations, we introduce ReSum, a lightweight, plug-and-play paradigm that enables unbounded exploration by periodically invoking an external tool to condense interaction histories into compact summaries. Although this paradigm functions without training, standard agents are not inherently aligned to reason over such compressed contexts. To bridge this gap, we propose ReSum-GRPO, which adapts Group Relative Policy Optimization (GRPO) via advantage broadcasting to propagate final rewards across segmented trajectories, enabling credit assignments over long-horizons. Extensive experiments show that ReSum achieves a 4 .5% improvement over ReAct in training-free settings, with ReSum-GRPO yielding a further 8.2% gain.  \nNotably, with only 1K training samples, a ReSumenhanced 30B agent achieves competitive performance with leading open-source models, showing ReSum’s effectiveness.  \n1. Introduction  \nLarge Language Model (LLM)-based agents have demonstrated strong performance on complex, knowledge-  \n*Equal contribution 1The Chinese University of Hong Kong 2Tongyi Lab, Alibaba Group 3The Hong Kong University of Science and Technology 4Penn State University. Correspondence to: Liwen Zhang \u003C[guiwuqiansha@gmail.com](guiwuqiansha@gmail.com) >, Yong Jiang \u003C [jiangyong.ml@gmail.com](jiangyong.ml@gmail.com) >.  \nPreprint. March 27, 2026.  \nintensive tasks (Yao et al., 2023 ; Wang et al., 2024 ; Jin et al., 2025 ; Xi et al., 2025 ; Gao et al., 2025) . Among these, web agents are particularly critical: they actively search and browse the open web, extract and ground facts from diverse sources, and synthesize answers that are both user-specific and up-to-date (Wu et al., 2025a ; Li et al., 2025a) .  \nHowever, answering complex queries is nontrivial. Consider this example: A painter, whose father died of heart disease, had an elder sister and five children with his wife. Later, his marriage broke down and he had three more relationships. What is the name of the literary work based on this person? This query exemplifies the challenges of web search: it involves multiple entities, intertwined relationships, and fragmentary information with high uncertainty. Such tasks cannot be resolved with a few simple search calls; instead, they demand extended cycles of targeted querying, browsing, and cross-verification to progressively assemble a complete evidence chain (Gao et al., 2025) .  \nThis need for long-horizon exploration faces a fundamental barrier: context constraints. Most LLMs have limited context windows (Yang et al., 2024 ; Jiang et al., 2023), and the popular ReAct paradigm (Yao et al., 2023), which appends every observation and thought to the history, quickly exhausts this budget (Figure 1) . Current solutions, such as MEM1 (Zhou et al., 2025b) and MemAgent (Yu et al., 2025a), often rely on architectural modifications, e.g., generating internal memory tokens. While effective, these approaches break compatibility with pre-existing agents and necessitate end-to-end retraining.  \nTo overcome these limitations, we introduce ReSum, a lightweight paradigm that enables unbounded exploration through context summarization. The core insight is to periodically invoke an external tool to condense","cbCain9lrE5Oe5Fb","https://ap.wps.com/l/cbCain9lrE5Oe5Fb","pdf",1360337,29,"English","# Introduction\n## Problem: context constraints in long-horizon web agents\n## ReSum: plug-and-play context summarization\n## ReSumTool-30B: summary tool for evidence extraction\n## ReSum-GRPO: GRPO adaptation for long-horizon credit assignment","[{\"question\":\"What problem does ReSum address for LLM-based web agents?\",\"answer\":\"ReSum targets the conflict between the need for extensive long-horizon exploration and the limited context windows of standard LLM agents.\"},{\"question\":\"How does ReSum enable unbounded exploration without architectural retraining?\",\"answer\":\"ReSum periodically invokes an external tool to condense interaction histories into compact summaries, then resumes exploration from these compressed states.\"},{\"question\":\"Why is ReSum-GRPO proposed, and how does it work?\",\"answer\":\"ReSum-GRPO improves alignment by adapting GRPO to the ReSum paradigm using advantage broadcasting across segmented trajectories, enabling better credit assignment over long horizons.\"}]","ReSum - Unlocking Long-Horizon Search Intelligence via Context Summarization | PDF",1789111881,10]