[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273632-105":59,"doc-detail-273632-en":130},{"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":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","compliance-capability-and-conflict-benchmarking-multimodal-llms-under-system-messages","Compliance, Capability, and Conflict - Benchmarking Multimodal LLMs under System Messages","","Production deployments of Multimodal Large Language Models (MLLMs) increasingly depend on system messages to control model behavior. However, prior benchmarks either test constraints only in text or inject them into the user turn, leaving system-message adherence in multimodal settings largely unmeasured and unclear about trade-offs with core vision-language ability. This work introduces VSYSBENCH, built on MMVet-v2, covering 5 categories and 22 sub-categories of system-level constraints, and evaluates compliance and correctness via JSR and CCS.",{"@graph":69,"@context":122},[70,84,105],{"@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/compliance-capability-and-conflict-benchmarking-multimodal-llms-under-system-messages/273632/",{"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/compliance-capability-and-conflict-benchmarking-multimodal-llms-under-system-messages/273632.png","ImageObject",300,407,{"name":92,"@type":93},"Levi","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-25","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":44},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What problem does VSYSBENCH address in system-message evaluation for MLLMs?","Question",{"text":112,"@type":113},"It measures how well system-level constraints are followed in genuinely multimodal contexts, including under deliberately conflicting user requests, while jointly checking both compliance and answer correctness.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"How is VSYSBENCH constructed and categorized?",{"text":117,"@type":113},"VSYSBENCH is built on MMVet-v2 with 2,258 human-verified samples organized into 5 main categories and 22 sub-categories of constraints, each paired with misaligned user conflicts to stress-test instructional hierarchy.",{"name":119,"@type":110,"acceptedAnswer":120},"How does VSYSBENCH score responses across two axes?",{"text":121,"@type":113},"It scores constraint compliance and answer correctness jointly using the Joint Satisfaction Rate (JSR) and Cross-Constraint Sensitivity (CCS), so that passing one axis cannot hide failures on the other.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},273632,1789472323,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":44,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":129,"read_time":144,"preview_page_count":4},7971461740909,"https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d","Compliance, Capability, and Conflict: Benchmarking Multimodal LLMs  \nunder System Messages  \nJuan Yeo∗‡  \nGraduate School of Data Science Seoul National University [juanyeo@snu.ac.kr](juanyeo@snu.ac.kr)  \nGeewook Kim∗†  \nNAVER Cloud AI KAIST AI  \n[gwkim.rsrch@gmail.com](gwkim.rsrch@gmail.com)  \narXiv : 2608 . 19207v 1 [cs .CL] 12 Jun 2026  \nAbstract  \nProduction deployments of Multimodal Large Language Models (MLLMs) increasingly rely on system messages to govern model behavior. Yet existing benchmarks either evaluate constraints in text only or embed them into the user turn, leaving system-message adherence in multimodal contexts largely unmeasured; they also leave open whether compliance comes at the cost of foundational visionlanguage capabilities. We introduce VSYSBENCH, a benchmark built on MMVet-v2 that organizes constraints into 5 main categories and 22 sub-categories, ranging from textual directives in visual contexts to fully visiongrounded ones, each paired with a misaligned counterpart that stress-tests the instructional hierarchy. VSYSBENCH scores each response jointly along two axes, constraint compliance and answer correctness, via the Joint Satisfaction Rate (JSR) and Cross-Constraint Sensitivity (CCS) . Across 16 MLLMs, we find that imposing system messages substantially erodes base task accuracy, that compliance collapses under user conflict for open-weight models while remaining stable for top proprietary ones, and that vision-grounded constraints are the hardest category for every model. Our benchmark is publicly available at [https://github](https://github). com/naver-ai/VSysBench.  \n1 Introduction  \nProduction deployments of Multimodal Large Language Models (MLLMs) are increasingly governed not by user queries alone, but by the system messages that precede them. In practice, behavior is steered by system messages rather than fine-tuning: prompt-level control is cheaper, faster to iterate, and lets a single base model serve many downstream products. The system message therefore  \n∗ Juan Yeo and Geewook Kim contributed equally to this work and share first authorship.  \n‡ Work done at NAVER Cloud AI.  \n† Corresponding author.  \nSystem Message  \nQuery  \nGround Truth  \nOrange  \nFigure 1: Two representative cases from VSYSBENCH. (Top) The model returns the correct answer (orange) but violates the JSON schema, emitting only 2 of the 4 required bbox coordinates—answer correct, constraint failed. (Bottom) The model faithfully follows the location-tag format but misidentifies the fruit (banana instead of orange)—constraint satisfied, answer wrong. A benchmark that scored only one of the two axes would mark each of these as a success. VSYSBENCH scores both jointly on the same instance.  \nsits above the user query, governing overall behavior across interactions. This role takes on additional weight in MLLMs, where the system message must also shape how the model interprets visual inputs.  \nDespite this central role, most multimodal benchmarks evaluate MLLMs under a default helpfulassistant prompt, leaving system-message adherence largely unmeasured. A growing body of instruction-following work has begun to address  \nthis gap, yet each existing line leaves the visuallygrounded system message setting untouched. SysBench (Qin et al., 2024) directly probes systemmessage following but is text-only, providing no mechanism to examine how visual inputs interact with system-level constraints. MIA-Bench (Qian et al., 2025) and MM-IFEval (Ding et al., 2025) extend instruction following to image-text pairs but embed constraints into the user turn, and thus cannot test whether a system message holds firm against a conflicting user request. VC-IFEval (He et al., 2026) introduces vision-dependent constraints with LLM-as-judge scoring but, again, places them at the user level and does not measure how system-level enforcement interacts with the model’s core visual reasoning. What remains missing is a benchmark that (i) places constrai","cbCaifQ5ryb7qICR","https://ap.wps.com/l/cbCaifQ5ryb7qICR","pdf",1495123,17,"English","# Introduction\n## System-message control in production MLLMs\n## Limitations of existing multimodal benchmarks\n## VSYSBENCH overview and scoring\n## Experimental findings and trade-offs","[{\"question\":\"What problem does VSYSBENCH address in system-message evaluation for MLLMs?\",\"answer\":\"It measures how well system-level constraints are followed in genuinely multimodal contexts, including under deliberately conflicting user requests, while jointly checking both compliance and answer correctness.\"},{\"question\":\"How is VSYSBENCH constructed and categorized?\",\"answer\":\"VSYSBENCH is built on MMVet-v2 with 2,258 human-verified samples organized into 5 main categories and 22 sub-categories of constraints, each paired with misaligned user conflicts to stress-test instructional hierarchy.\"},{\"question\":\"How does VSYSBENCH score responses across two axes?\",\"answer\":\"It scores constraint compliance and answer correctness jointly using the Joint Satisfaction Rate (JSR) and Cross-Constraint Sensitivity (CCS), so that passing one axis cannot hide failures on the other.\"}]","Compliance, Capability, and Conflict - Benchmarking Multimodal LLMs under System Messages | PDF",43]