[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83018-en":3,"doc-seo-83018-105":30,"detail-sidebar-cat-0-en-105":96},{"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},83018,7971461740909,"Levi","https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d",8,"Research & Report","SearchEyes Towards Frontier Multimodal Deep Search Intelligence via Search World Simulation","Training multimodal search agents for long-horizon multi-hop reasoning is limited by a structural disconnect: pipelines build training data, search environments, and reward signals separately, discarding intermediate metadata, relying on irreproducible external engines, and yielding sparse trajectory-level RL rewards. SearchEyes unifies these components using a typed knowledge-graph simulated search world. Perception-Knowledge Chains sample constrained visual-knowledge multi-hop paths while retaining hop-level entity metadata as reward anchors. Hop-Anchored Policy Optimization reuses these anchors for step-level credit assignment, achieving state-of-the-art results on six benchmarks, with SearchEyes-27B improving the best open-source baseline by 6.2 points.","arXiv :2607 .05943v 1 [ cs .AI ] 7 Jul 2026  \nSearchEyes: Towards Frontier Multimodal Deep Search Intelligence via Search World Simulation  \nZhengbo Jiao 1 , Yiming Cheng2 , Yilei Jiang 1 , Kaituo Feng 1 , Rui Huang3 , Tianyi Jiang4 , Juanxi Tian5 , Jiapeng Li2 , Qunzhong Wang 1 , Tailai Chen6 , Qianshan Wei6 , Chuan Xiao8 , Shanyu Rong4 , Yangfu Li7 , Yanhan Zhou2 ,  \nYunpu Ma9 , Yifan Zhang6 , Xiangyu Yue 1 ,†  \n1 MMLab, CUHK 2 THU 3 HKU 4 PKU 5 NTU 6 CASIA 7 ECNU 8 HIT 9 LMU  \nAbstract  \nTraining multimodal search agents to perform multi-hop reasoning remains challenging due toa fundamental structural disconnect: existing pipelines construct training data, search environments, and reward signals independently, causing synthesized structural metadata to be discarded, environments to rely on irreproducible external engines, and RL rewards to remain sparse at the trajectory level. We present SearchEyes, which uses a typed knowledge graph as the backbone of a simulated search world that unifies all three components. We propose Perception-Knowledge Chains (PKC) to sample constrained multi-hop paths over the visual-knowledge intersection of Wikidata5M, retaining hop-level entity metadata that simultaneously defines a self-contained search world and step-level reward anchors. We further propose Hop-Anchored Policy Optimization (HaPO), which reuses these anchors for step-level credit assignment without a separately trained process reward model. Experiments on six multimodal knowledge-intensive benchmarks show that SearchEyes achieves state-of-the-art performance among open-source multimodal search agents, with SearchEyes-27B improving over the strongest open-source baseline by 6.2 points on average.  \nDate: June 2026  \nCorrespondence: [Frostlin2005@gmail.com](Frostlin2005@gmail.com) ;[xyyue@ie.cuhk.edu.hk](xyyue@ie.cuhk.edu.hk)  \nProject Page: [https://github.com/Frostlinx/SearchEyes](https://github.com/Frostlinx/SearchEyes)  \n1 Introduction  \nReinforcement learning has become the dominant paradigm for training autonomous LLM agents [1– 4], enabling capabilities that span long-chain reasoning, tool use, and multi-step decision making. Among these, multi-step information retrieval and synthesis has emerged as a particularly impactful application, where agents iteratively search, read, and reason over external knowledge to solve complex questions. On the text side, systems such as Search-R1 [5], ASearcher [6], and REDSearcher [7] have demonstrated that RL-trained search agents can effectively perform long-horizon reasoning  \n| PKC |  | Construction |  |\n| --- | --- | --- | --- |\n| \u003Cbr>P/K alternation |  | \u003Cbr>anti-shortcut | \u003Cbr>hidden entities |\n\n\n| \u003Cbr>Mock Search World\u003Cbr> |  |\n| --- | --- |\n| (BM25+RRF) | Image Index\u003Cbr>search/lookup/crop |\n\n\n| Training |  |  |\n| --- | --- | --- |\n| SFT env\u003Cbr>enhancement\u003Cbr> | \u003Cbr>HaPO\u003Cbr>checkpoints | final + step\u003Cbr>reward\u003Cbr> |\n\nFigure 1 Overview of SearchEyes. A typed knowledge graph serves as the unified backbone for three tightly-coupled stages: PKC question synthesis (left), a self-contained search world for agent interaction (center), and hop-anchored reward signals for policy optimization (right) . By retaining structural metadata across all stages, SearchEyes eliminates the data–environment–reward disconnect that limits existing pipelines.  \nover retrieved evidence. In the multimodal domain, Vision-DeepResearch [8], OpenSearch-VL [9], MMSearch-R1 [10], and VSearcher [11] have begun equipping MLLMs with visual and textual search tools, coupled with dedicated training environments and agentic RL algorithms. While these efforts have yielded impressive results, training a strong long-horizon multimodal search agent remains an open challenge, as current pipelines exhibit fundamental bottlenecks in how training data, search environments, and reward signals are constructed and connected.  \nDespite the promising results above, existing pipelines suffer from a fundamental structural disconnect:","cbCaiior7PiAmDwC","https://ap.wps.com/l/cbCaiior7PiAmDwC","pdf",2113125,5,1,25,"English","en",105,"# Abstract\n# Introduction\n## Structural disconnect in existing pipelines\n## SearchEyes overview and unified search world\n## Perception-Knowledge Chains (PKC)\n## Hop-Anchored Policy Optimization (HaPO)","[{\"question\":\"What is the core problem addressed by SearchEyes?\",\"answer\":\"Existing pipelines construct training data, environments, and rewards independently, discarding intermediate structure, using irreproducible external engines, and producing sparse trajectory-level rewards.\"},{\"question\":\"How does SearchEyes unify data synthesis, environment simulation, and RL training?\",\"answer\":\"It uses a typed knowledge graph as a shared backbone to form a self-contained simulated search world, based on Wikidata5M typed triples.\"},{\"question\":\"What are Perception-Knowledge Chains (PKC) and what do they retain?\",\"answer\":\"PKC samples constrained multi-hop paths over the visual-knowledge intersection and keeps hop-level entity metadata, which later serves as anchors for reward signals.\"},{\"question\":\"How does Hop-Anchored Policy Optimization (HaPO) improve credit assignment?\",\"answer\":\"HaPO reuses the gold entity IDs retained by PKC as step-level anchors, enabling step-level credit assignment without a separately trained process reward model.\"}]",1784184698,63,{"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":91,"head_meta":93,"extra_data":95,"updated_unix":28},"searcheyes-towards-frontier-multimodal-deep-search-intelligence-via-search-world-simulation","",{"@graph":36,"@context":90},[37,54,69],{"@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":53},"https://docshare.wps.com/document/searcheyes-towards-frontier-multimodal-deep-search-intelligence-via-search-world-simulation/83018/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":24,"description":14,"dateModified":62,"datePublished":63,"encodingFormat":61,"isAccessibleForFree":64,"interactionStatistic":65},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-07-24","2026-07-16",true,{"@type":66,"interactionType":67,"userInteractionCount":20},"InteractionCounter",{"@type":68},"ViewAction",{"@type":70,"mainEntity":71},"FAQPage",[72,78,82,86],{"name":73,"@type":74,"acceptedAnswer":75},"What is the core problem addressed by SearchEyes?","Question",{"text":76,"@type":77},"Existing pipelines construct training data, environments, and rewards independently, discarding intermediate structure, using irreproducible external engines, and producing sparse trajectory-level rewards.","Answer",{"name":79,"@type":74,"acceptedAnswer":80},"How does SearchEyes unify data synthesis, environment simulation, and RL training?",{"text":81,"@type":77},"It uses a typed knowledge graph as a shared backbone to form a self-contained simulated search world, based on Wikidata5M typed triples.",{"name":83,"@type":74,"acceptedAnswer":84},"What are Perception-Knowledge Chains (PKC) and what do they retain?",{"text":85,"@type":77},"PKC samples constrained multi-hop paths over the visual-knowledge intersection and keeps hop-level entity metadata, which later serves as anchors for reward signals.",{"name":87,"@type":74,"acceptedAnswer":88},"How does Hop-Anchored Policy Optimization (HaPO) improve credit assignment?",{"text":89,"@type":77},"HaPO reuses the gold entity IDs retained by PKC as step-level anchors, enabling step-level credit assignment without a separately trained process reward model.","https://schema.org",{"og:url":52,"og:type":92,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":94,"canonical":52},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":97},[98,102,106,110,114,119,124,127,132,135,139],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":99,"show_sort_weight":100,"slug":101},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":103,"show_sort_weight":104,"slug":105},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},"Exam",70,"exam",{"id":20,"doc_module":4,"doc_module_name":46,"category_name":111,"show_sort_weight":112,"slug":113},"Comic",60,"comic",{"id":115,"doc_module":4,"doc_module_name":46,"category_name":116,"show_sort_weight":117,"slug":118},6,"Technology",50,"technology",{"id":120,"doc_module":4,"doc_module_name":46,"category_name":121,"show_sort_weight":122,"slug":123},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":125,"slug":126},30,"research-report",{"id":128,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":130,"slug":131},9,"Religion & Spirituality",20,"religion-spirituality",{"id":130,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":130,"slug":134},"World Cup","world-cup",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":136,"slug":138},10,"Lifestyle","lifestyle",{"id":140,"doc_module":4,"doc_module_name":46,"category_name":141,"show_sort_weight":20,"slug":142},19,"General","general"]