[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-327547-105":59,"doc-detail-327547-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","homer-hierarchical-online-memory-exploration-and-reasoning-for-long-form-video-understanding","HOMER - Hierarchical Online Memory Exploration and Reasoning for Long-form Video Understanding","","Multimodal large language models excel on short clips but struggle with hour-long videos in online streaming settings, where limited memory forces incremental frame processing. Existing approaches either compress visuals without preserving semantic structure, or store temporal memories that omit explicit causal links, requiring the LLM to reconstruct multi-hop narrative reasoning at every query. HOMER introduces hierarchical online memory aligned to multi-scale video structure and an agentic reasoner that retrieves, verifies, and corrects evidence across multiple rounds, outperforming prior methods across long-video benchmarks.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/homer-hierarchical-online-memory-exploration-and-reasoning-for-long-form-video-understanding/327547/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/homer-hierarchical-online-memory-exploration-and-reasoning-for-long-form-video-understanding/327547.png","ImageObject",300,407,{"name":92,"@type":93},"Violet","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-29","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",16,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"Why do multimodal large language models struggle with hour-long online videos?","Question",{"text":113,"@type":114},"Because online streaming processes frames incrementally under limited memory, causing redundant visual content and dispersed evidence that is difficult for attention to locate and integrate within the model’s context.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"What gap do existing online memory methods leave for narrative reasoning?",{"text":118,"@type":114},"They often fail to explicitly model temporal and causal relationships between events, so multi-hop narrative reasoning must be reconstructed by the LLM at every query.",{"name":120,"@type":111,"acceptedAnswer":121},"How does HOMER improve long-form video understanding in online settings?",{"text":122,"@type":114},"HOMER builds hierarchical online memory with perceptual, entity, and event layers, then uses an agentic reasoner with multi-round retrieval plus a harness to verify and correct each step, enabling grounded multi-hop answers.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},327547,1790121882,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":145,"read_time":146,"preview_page_count":4},4398048950312,"https://ap-avatar.wpscdn.com/avatar/400002538284de19e3c?_k=1778320343897328908","HOMER: Understanding Long-form Videos with Hierarchical Memory and  \nAgentic Reasoning  \nYixin Ji 1 , Fanghua Ye2 , Juntao Li 1 Bo Zhao2 , Zexuan Qiu2 , Zhaopeng Tu2∗, Liefeng Bo2 , Min Zhang 1  \n1 Soochow University  \n2Tencent Hunyuan Multimodal Department  \n{jiyixin169, [fanghua.ye.21}@gmail.com](fanghua.ye.21}@gmail.com)[ ](fanghua.ye.21}@gmail.com){ljt, [minzhang}@suda.edu.cn](minzhang}@suda.edu.cn)  \narXiv :2607 .02588v 1 [ cs .CV] 1 Jul 2026  \nAbstract  \nMultimodal large language models excel on short clips but struggle on hour-long videos in an online setting, where frames are processed incrementally under limited memory.  \nExisting online methods either retain compact visual representations that lack semantic structure, or build higher-level memory stores organized around temporal proximity rather than explicit causal links, leaving multi-hop narrative reasoning to be reconstructed by the LLM at every query. We bridge this gap with HOMER, a Hierarchical Online Memory Exploration and Reasoning framework. HOMER’s memory mirrors the multi-scale structure of long videos, ranging from raw perception, to recurring entities, to events connected by explicit temporal and causal relations. Its agentic reasoner then explores this memory the way humans do, locating the relevant scene, looking up details, and composing the answer through multi-round memory retrieval, with a harness that verifies and corrects each step. HOMER outperforms the previous best agent method by +5 .5, +10 .8, and +4 .4 points on M3-Bench-robot, M3-Bench-web, and VideoMME-Long, and consistently lifts three various LLM backbones, indicating a modelagnostic structural capability for grounded retrieval over long videos.1  \n1 Introduction  \nReasoning over long-form videos is a basic requirement for multimodal large language models (MLLMs) used in embodied intelligence (Shenet al., 2026 ; Pai et al., 2025), autonomous agents (Chen et al., 2025a), and world modeling (Ren et al., 2025), yet MLLMs that handle short clips fluently still struggle on such long-form  \n∗ Corresponding authors.  \n1Code is available at [https://github.com/](https://github.com/)[ ](https://github.com/)Dereck0602/Homer  \ncontent (Zou et al., 2024 ; Bai et al., 2025) . Videos spanning tens of minutes to hours contain thousands of visually redundant frames, complex narrative structures, and dense interactions among characters. The resulting token count far exceeds the context window of current models, and even with sufficient context length, critical details are dispersed across the timeline, making it difficult for attention mechanisms to locate and integrate relevant evidence (Tang et al., 2025b ; Dou et al., 2026) . Long-video understanding therefore demands both compressing redundancy to fit the model’s context and pinpointing fine-grained evidence dispersed across that context.  \nExisting approaches address these demands from different angles. To reduce the token count, sparse frame selection methods (Wang et al., 2025b ; Tanget al., 2025a) subsample a small set of representative frames, trading temporal coverage for efficiency, while token compression methods (Liet al., 2024 ; Shen et al., 2025a) reduce the number of visual tokens per frame so that more frames fit within the context window, though aggressive compression may discard fine-grained details. To improve evidence localization, agentic reasoning methods (Wang et al., 2024 ; Shen et al., 2025b) equip MLLMs with visual tools to iteratively manipulate and reason over videos, achieving stronger multi-hop reasoning performance. Most of these methods, however, assume an offline setting where the full video is available at query time and segments can be re-retrieved on demand when the first evidence is insufficient.  \nMany real-world applications instead require an online setting where frames arrive as a continuous stream and storing the full video is infeasible, demanding to incrementally accumulate and organize information as the vi","cbCaiiKUBVScup5t","https://ap.wps.com/l/cbCaiiKUBVScup5t","pdf",6114959,18,"English","# Introduction\n## Challenges in online long-form video understanding\n## Prior work: token compression, sparse selection, and agentic methods\n## Online setting: memory efficiency vs retrieval difficulty\n## Proposed principles for HOMER\n## HOMER framework: hierarchical memory and agentic reasoner","[{\"question\":\"Why do multimodal large language models struggle with hour-long online videos?\",\"answer\":\"Because online streaming processes frames incrementally under limited memory, causing redundant visual content and dispersed evidence that is difficult for attention to locate and integrate within the model’s context.\"},{\"question\":\"What gap do existing online memory methods leave for narrative reasoning?\",\"answer\":\"They often fail to explicitly model temporal and causal relationships between events, so multi-hop narrative reasoning must be reconstructed by the LLM at every query.\"},{\"question\":\"How does HOMER improve long-form video understanding in online settings?\",\"answer\":\"HOMER builds hierarchical online memory with perceptual, entity, and event layers, then uses an agentic reasoner with multi-round retrieval plus a harness to verify and correct each step, enabling grounded multi-hop answers.\"}]","HOMER - Hierarchical Online Memory Exploration and Reasoning for Long-form Video Understanding | PDF",1789977488,45]