[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84171-en":3,"doc-seo-84171-105":29,"detail-sidebar-cat-0-en-105":90},{"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":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84171,1374391974468,"Eden","https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0",8,"Research & Report","AnchorPrune Relevance-Anchored Contextual Expansion for Visual Token Pruning","Large vision-language models face high inference costs because high-resolution inputs generate thousands of visual tokens, many of which are redundant for a specific query. Existing pruning combines relevance and diversity but can conflict under strong compression, either overfocusing on correlated evidence or discarding indispensable tokens. AnchorPrune introduces a training-free, architecture-aware method that first builds a protected relevance anchor and then expands it using importance-weighted novelty to recover complementary, non-redundant context. It improves the accuracy–efficiency trade-off across image/video VLMs and benchmarks, notably preserving 97.6% performance on LLaVA-NeXT-7B with only 160/2,880 tokens.","arXiv :2607 .07033v2 [ cs .CV] 12 Jul 2026  \nAnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning  \nKyuan Oh and Bumsoo Kim†  \nChung-Ang University, Seoul, Korea  \n{oka04108,[bumsoo}@cau.ac.kr](bumsoo}@cau.ac.kr)  \nAbstract. Large vision-language models incur substantial inference costs because high-resolution inputs introduce thousands of visual tokens, many of which are redundant for a given query. Existing pruning methods often combine query relevance and token diversity, yet these objectives can conflict under aggressive compression: relevance-driven selection may overconcentrate the budget on correlated local evidence, while diversitydriven selection may suppress indispensable tokens or retain distinct but uninformative regions. We introduce AnchorPrune, a training-free framework that first constructs a protected relevance anchor and then expands it with complementary visual context. AnchorPrune adaptively determines the anchor size from the novelty profile of relevance-ranked tokens, preserving a compact set of query-critical evidence, and allocates the remaining budget through importance-weighted novelty to recover informative, non-redundant context relative to the anchor. This ordered design prevents contextual expansion from displacing indispensable query cues while improving overall visual coverage. AnchorPrune is lightweight, architecture-aware, and requires neither retraining nor model modification. Across image and video vision-language models and benchmarks, it consistently improves the accuracy–efficiency trade-off over training-free baselines, particularly under severe compression. On LLaVA-NeXT-7B, AnchorPrune preserves 97.6% of full-token performance using only 160 of 2,880 visual tokens. These results establish relevance-anchored contextual expansion as an effective principle for efficient multimodal inference.  \nCode is available at [https://github.com/MULTI-cau/AnchorPrune](https://github.com/MULTI-cau/AnchorPrune).  \nKeywords: Visual Token Pruning · Vision-Language Models · Efficient Multimodal Inference · Relevance-Anchored Contextual Expansion  \n1 Introduction  \nVision-language models (VLMs) achieve strong performance across image and video understanding tasks, but their long visual-token sequences incur substantial inference costs [2, 11 , 12 , 30] . High-resolution, multi-crop, and multi-frame inputs can introduce thousands of visual tokens, many of which are redundant for a given query yet must still be processed by the language model. This overhead is especially pronounced during prefilling, where visual tokens dominate  \n† : corresponding author  \n2 K. Oh and B. Kim  \nFig. 1: Qualitative comparison on LLaVA-1.5-7B with 64 retained visual tokens. VisionZip [24] uses query-agnostic saliency, DivPrune [1] emphasizes queryagnostic diversity, and CDPruner [28] combines relevance and diversity within a unified objective. Under severe compression, these strategies may miss, fragment, or suppress query-critical evidence. AnchorPrune instead protects a relevance anchor before expanding it with informative, non-redundant context, yielding correct predictions across diverse tasks.  \nsequence length, memory consumption, and attention computation. Visual token pruning therefore provides a direct route to efficient multimodal inference by retaining only the evidence needed for accurate prediction [1, 3 , 21 , 23 , 24 , 28 , 29] .  \nThe central challenge is not merely estimating token importance, but determining which evidence must be protected before the remaining budget is allocated. As illustrated in Fig. 1, saliency-based methods may retain redundant high-importance regions while overlooking query-specific cues [3, 21 , 23 , 24 , 29] . Diversity-based methods distribute selections across the visual feature space, but can fragment locally coherent evidence or preserve visually distinct yet queryirrelevant regions [1] . Query-aware methods improve instruction alignment, but relevance-dr","cbCait5YOJlFWeEJ","https://ap.wps.com/l/cbCait5YOJlFWeEJ","pdf",1140016,1,28,"English","en",105,"# Introduction\n## Motivation: Inference cost of long visual-token sequences\n## Limitations of existing relevance/diversity pruning\n## Proposed principle: protected relevance anchor then expansion","[{\"question\":\"Why do visual token pruning methods struggle under aggressive compression?\",\"answer\":\"Relevance-driven selection can overconcentrate on correlated local evidence, while diversity-driven selection may suppress indispensable tokens or keep distinct but query-irrelevant regions. Under strong compression, the coupling of relevance and diversity offers no explicit protection for critical evidence.\"},{\"question\":\"How does AnchorPrune decide what to keep during pruning?\",\"answer\":\"AnchorPrune first ranks tokens with query-conditioned priority scores and constructs a protected relevance anchor. It then expands the retained set using importance-weighted novelty to select tokens that are informative and non-redundant with respect to the anchor.\"},{\"question\":\"What are the reported efficiency and accuracy benefits of AnchorPrune?\",\"answer\":\"Across image and video VLMs and benchmarks, AnchorPrune improves the accuracy–efficiency trade-off, especially under severe compression. On LLaVA-NeXT-7B, it preserves 97.6% of full-token performance while using only 160 of 2,880 visual tokens.\"}]",1784193632,71,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"anchorprune-relevance-anchored-contextual-expansion-for-visual-token-pruning","",{"@graph":35,"@context":84},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/anchorprune-relevance-anchored-contextual-expansion-for-visual-token-pruning/84171/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"Why do visual token pruning methods struggle under aggressive compression?","Question",{"text":74,"@type":75},"Relevance-driven selection can overconcentrate on correlated local evidence, while diversity-driven selection may suppress indispensable tokens or keep distinct but query-irrelevant regions. Under strong compression, the coupling of relevance and diversity offers no explicit protection for critical evidence.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does AnchorPrune decide what to keep during pruning?",{"text":79,"@type":75},"AnchorPrune first ranks tokens with query-conditioned priority scores and constructs a protected relevance anchor. It then expands the retained set using importance-weighted novelty to select tokens that are informative and non-redundant with respect to the anchor.",{"name":81,"@type":72,"acceptedAnswer":82},"What are the reported efficiency and accuracy benefits of AnchorPrune?",{"text":83,"@type":75},"Across image and video VLMs and benchmarks, AnchorPrune improves the accuracy–efficiency trade-off, especially under severe compression. On LLaVA-NeXT-7B, it preserves 97.6% of full-token performance while using only 160 of 2,880 visual tokens.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]