[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-327367-105":59,"doc-detail-327367-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","mmbench-live-a-continuously-evolving-benchmark-for-multimodal-models","MMBench-Live - A Continuously Evolving Benchmark for Multimodal Models","","Evaluation benchmarks are crucial for vision-language models, yet most multimodal benchmarks remain static and therefore degrade under temporal staleness, data contamination, and high maintenance cost. MMBench-Live presents a continuously evolving multimodal benchmark built via a multi-agent automated pipeline. It frames benchmark evolution as task-guided dataset construction, using structured specifications, feedback-controlled real-time acquisition, and verifiable QA generation with executable reasoning. A distribution-consistent update strategy preserves cross-version comparability and supports stable model rankings.",{"@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/mmbench-live-a-continuously-evolving-benchmark-for-multimodal-models/327367/",{"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/mmbench-live-a-continuously-evolving-benchmark-for-multimodal-models/327367.png","ImageObject",300,407,{"name":92,"@type":93},"Aurora","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-30","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",15,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"Why do static multimodal benchmarks become problematic over time?","Question",{"text":113,"@type":114},"Static benchmarks can suffer from temporal staleness, data contamination, and high maintenance cost, making frequent updates impractical at scale.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does MMBench-Live enable continuous benchmark updates?",{"text":118,"@type":114},"It uses a multi-agent-driven automated pipeline that treats evolution as task-guided dataset construction, combining structured benchmark specifications, feedback-controlled real-time data acquisition, and verifiable QA generation with executable reasoning.",{"name":120,"@type":111,"acceptedAnswer":121},"How does MMBench-Live maintain comparability across different benchmark versions?",{"text":122,"@type":114},"It introduces a distribution-consistent update strategy that extracts task-related visual patterns from the original benchmark to guide data collection and filtering, helping preserve task semantics and distributional characteristics.","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},327367,1790500394,{"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},4810365810221,"https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d","# MMBench-Live:A Continuously Evolving Benchmark for Multimodal Models\n\nYuanzhi Liu¹Shousheng Zhao¹Bo Zhou¹Kongming Liang¹Zhanyu Ma¹  \narXiv:2607.01813vl[cs.CV] 2 Jul 2026  \n## Abstract\n\nEvaluation benchmarks are essential for assessingvision-language models(VLMs),but most mul-timodal benchmarks are static,making them vul-nerable to temporal staleness,data contamination,and costly maintenance.We present MMBench-Live,a continuously evolving multimodal bench-mark built by a multi-agent-driven automatedpipeline.Our framework treats benchmark evolu-tion as task-guided dataset construction,integrat-ing structured benchmark specification,feedback-controlled real-time data acquisition,and veri-fiable QA generation with executable reasoning.To maintain cross-version comparability,we intro-duce a distribution-consistent update strategy thatextracts task-related visual patterns from the orig-inal benchmark to guide data collection and fil-tering.Instantiated from MMBench,MMBench-Live contains 5.9K newly generated evaluation in-stances with a high answer correctness rate,whileeach update costs about $30 and takes 1-2 hours.Extensive evaluations show that MMBench-Livepreserves stable model rankings,maintains seman-tic alignment with the original benchmark,andexhibits weaker contamination-related memoriza-tion signals,suggesting a practical and scalableparadigm for sustainable multimodal benchmarkevolution.The project is available at https://github.com/PRIS-CV/MMBench-Live.  \n## 1.Introduction\n\nMost existing vision-language evaluation benchmarks(Luet al.,2022;Fu et al.,2023;Liet al.,2024;Liu et al.,2024b)adopt a static evaluation paradigm,implicitly assumingthata fixed test set can serve as a long-term proxy for modelcapability.This assumption becomes increasingly fragileas modern vision-language models(VLMs)scale rapidly  \nand are trained on ever-growing web-scale corpora.Staticbenchmarks are vulnerable to temporal staleness,data con-tamination(Touvron et al.,2023a;Chen et al.,2024;Songet al.,2025b),and high maintenance costs,making frequentupdates impractical at scale.However,dynamically updat-ing a benchmark is not simply a matter of replacing oldsamples with new ones:a reliable live benchmark mustintroduce temporally fresh instances while preserving theoriginal task semantics,capability coverage,data character-istics,and cross-version comparability.  \nPrior work has explored several strategies for updating mul-timodal evaluation datasets.Self-evolving methods basedon visual perturbations or linguistic rewrites(Yang et al.,2025b;Zhang et al.,2026)may suffer from semantic driftand degraded comparability,while generation-based meth-ods(Wang et al.,2026;Zhang et al.,2025a;Wen et al.,2025)can introduce distributional gaps under complex semanticconstraints.Real-world data updating (Jiang et al.,2025;Shabtay et al,2025;Jain et al.,2025;White et al.,2025)isa promising direction because it naturally improves tempo-ral freshness and reduces direct reuse of static benchmarksamples.Nevertheless,existing pipelines are often tied tofixed data sources,predefined task formats,or task-specificconstruction procedures.Consequently,continuously incor-porating fresh real-world data across diverse multimodaltasks while preserving semantic coherence,distributionalalignment,and evaluation stability remains challenging.  \nTo address this challenge,we propose MMBench-Live,acontinuously evolving multimodal benchmark built througha multi-agent-driven automated pipeline.Our frameworkmodels benchmark evolution as task-guided dataset con-struction.Starting from the original benchmark,it convertsevaluation objectives,task hierarchy,and atomic tasks intostructured benchmark descriptions,which serve as the se-mantic basis for data acquisition and QA generation.To pre-serve cross-version comparability,MMBench-Live identi-fies task-related visual patterns from the original benchmarkand uses them to guide feedback-controlled real-time dataacquisition,enabling the pipeline","cbCaieDgxmIHWxk9","https://ap.wps.com/l/cbCaieDgxmIHWxk9","pdf",11592936,35,"English","# Abstract\n# 1. Introduction\n## 2.1 Data Contamination in Model Evaluation\n# 2. Related Work","[{\"question\":\"Why do static multimodal benchmarks become problematic over time?\",\"answer\":\"Static benchmarks can suffer from temporal staleness, data contamination, and high maintenance cost, making frequent updates impractical at scale.\"},{\"question\":\"How does MMBench-Live enable continuous benchmark updates?\",\"answer\":\"It uses a multi-agent-driven automated pipeline that treats evolution as task-guided dataset construction, combining structured benchmark specifications, feedback-controlled real-time data acquisition, and verifiable QA generation with executable reasoning.\"},{\"question\":\"How does MMBench-Live maintain comparability across different benchmark versions?\",\"answer\":\"It introduces a distribution-consistent update strategy that extracts task-related visual patterns from the original benchmark to guide data collection and filtering, helping preserve task semantics and distributional characteristics.\"}]","MMBench-Live - A Continuously Evolving Benchmark for Multimodal Models | PDF",1789977078,88]