[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85546-en":3,"doc-seo-85546-105":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},85546,8796095360427,"Lucas Martin","https://ap-avatar.wpscdn.com/davatar_994ba38a5ba835b3df7d355c54d3ed8d",8,"Research & Report","EvoGuard: An Extensible Agentic RL-based Framework for Practical and Evolving AI-Generated Image Detection","The rapid proliferation of AI-Generated Images (AIGIs) creates severe misinformation risks, making detection both critical and difficult. EvoGuard reframes AIGI detection as learned, reasoning-based evidence synthesis over a pool of heterogeneous off-the-shelf detectors, rather than building a single stronger model. A capability-aware selection mechanism profiles detectors and collects complementary evidence, while a dynamic orchestration mechanism performs multi-round cross-validation of conflicting or low-confidence signals. A GRPO-based agentic reinforcement learning training uses only low-cost binary labels to avoid fine-grained annotations. Experiments show superior accuracy, reduced positive–negative bias, and plug-and-play integration of new detectors without retraining, yielding a practical long-term defense as threats evolve.","EvoGuard: An Extensible Agentic RL-based Framework for Practical and Evolving AI-Generated Image Detection  \nChenyang Zhu The University of Tokyo Tokyo, Japan National Institute of Informatics Tokyo, Japan  \n[chenyangzhu@g.ecc.u-tokyo.ac.jp](chenyangzhu@g.ecc.u-tokyo.ac.jp)  \nMaorong Wang National Institute of Informatics  \nTokyo, Japan [maorong@nii.ac.jp](maorong@nii.ac.jp)  \narXiv :2603 . 17343v2 [ cs .CV] 12 Jul 2026  \nJun Liu  \nNational Institute of Informatics Tokyo, Japan  \n[csjunliu@nii.ac.jp](csjunliu@nii.ac.jp)  \nChing-Chun Chang National Institute of Informatics Tokyo, Japan  \n[ccchang@nii.ac.jp](ccchang@nii.ac.jp)  \nIsao Echizen  \nNational Institute of Informatics Tokyo, Japan The University of Tokyo Tokyo, Japan  \n[iechizen@nii.ac.jp](iechizen@nii.ac.jp)  \nAbstract  \nThe rapid proliferation of AI-Generated Images (AIGIs) poses severe misinformation risks, making AIGI detection critical yet challenging. Traditional detection paradigms mainly rely on low-level features, whereas recent research increasingly focuses on leveraging the general understanding ability of Multimodal Large Language Models (MLLMs) to achieve better generalization, yet it still suffers from limited extensibility and expensive data annotations. Instead of building yet another detector, we recast AIGI detection as learned, reasoning-based evidence synthesis over a pool of heterogeneous off-the-shelf detectors, realized through EvoGuard, a novel agentic framework. A capability-aware selection mechanism profiles each detector and gathers complementary evidence per sample; a dynamic orchestration mechanism then reasons over heterogeneous outputs across multiple rounds, cross-validating conflicting or lowconfidence signals before concluding. This design exploits the complementary strengths among heterogeneous detectors, transcending the limits of any single model. Furthermore, optimized by a GRPO-based Agentic Reinforcement Learning algorithm using only low-cost binary labels, it eliminates the reliance on fine-grained annotations. Ex-  \ntensive experiments demonstrate that this learned reasoning paradigm outperforms single-detector and static ensembling, achieving SOTA accuracy while mitigating the bias between positive and negative samples. More importantly, it allows the plug-and-play integration of new detectors to boost overall performance in a train-free manner, offering a highly practical, long-term solution to ever-evolving AIGI threats. Source code will be publicly available upon acceptance.  \n1. Introduction  \nFrom GANs [19], Diffusion models [60] to autoregressive models [15, 45], the rapid evolution of AI generative models has spurred unprecedented growth in the creative industry. However, the proliferation of hyper-realistic AI-Generated Images (AIGIs) has introduced severe risks of misinformation [22], making AIGI detection a critical research area. Existing methods include capturing generator artifacts in spatial and frequency domains [69, 76], leveraging pretrained vision encoders [50, 88], and exploiting diffusion reconstruction errors [78] . However, due to the rapid iteration of generative models, image degradations on social me-  \n-> We need ever-evolving AI-generated detection to keep up  \nSeamless Extension  \nAdd new detectors without retrain  \nDynamic Orchestration  \nExploit strengths of diverse tools  \nAgentic RL Training Low-cost data annotation  \nFigure 1 . Motivation of EvoGuard. As generative models rapidly evolve, AIGI detection must evolve accordingly. Prior work focuses on building ever-stronger single detectors; we instead let an MLLM agent reason over a pool of off-the-shelf detectors, cross-validating their outputs rather than routing to one or fusing all by fixed rules. This reasoning-based synthesis exploits complementary strengths from heterogeneous detectors, enables training-free extensibility, and removes the need for fine-grained annotations.  \ndia [42], and sophisticated adversarial attacks [68], existing det","cbCaikT3wuRCZZBE","https://ap.wps.com/l/cbCaikT3wuRCZZBE","pdf",1840723,1,13,"English","en",105,"# Introduction\n## Motivation and challenges\n## Role of multimodal large language models","[{\"question\":\"Why is AI-generated image detection considered challenging today?\",\"answer\":\"AI generative models evolve quickly and produce highly realistic images, while existing detectors often rely on low-level artifacts that may not generalize to real-world variations. The detection problem remains difficult in complex and ever-changing environments.\"},{\"question\":\"How does EvoGuard differ from traditional single-detector approaches?\",\"answer\":\"EvoGuard uses an agentic MLLM to reason over outputs from multiple heterogeneous off-the-shelf detectors, performing multi-round cross-validation of conflicting or low-confidence signals rather than routing to one detector or using fixed fusion rules.\"},{\"question\":\"How does EvoGuard achieve extensibility without retraining?\",\"answer\":\"EvoGuard includes a plug-and-play capability to integrate new detectors to improve overall performance in a train-free manner. This is supported by the capability-aware selection and dynamic orchestration that leverage complementary strengths across detectors.\"}]",1784204349,33,{"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},"evoguard-an-extensible-agentic-rl-based-framework-for-practical-and-evolving-ai-generated-image-detection","",{"@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/evoguard-an-extensible-agentic-rl-based-framework-for-practical-and-evolving-ai-generated-image-detection/85546/",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 is AI-generated image detection considered challenging today?","Question",{"text":74,"@type":75},"AI generative models evolve quickly and produce highly realistic images, while existing detectors often rely on low-level artifacts that may not generalize to real-world variations. The detection problem remains difficult in complex and ever-changing environments.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does EvoGuard differ from traditional single-detector approaches?",{"text":79,"@type":75},"EvoGuard uses an agentic MLLM to reason over outputs from multiple heterogeneous off-the-shelf detectors, performing multi-round cross-validation of conflicting or low-confidence signals rather than routing to one detector or using fixed fusion rules.",{"name":81,"@type":72,"acceptedAnswer":82},"How does EvoGuard achieve extensibility without retraining?",{"text":83,"@type":75},"EvoGuard includes a plug-and-play capability to integrate new detectors to improve overall performance in a train-free manner. This is supported by the capability-aware selection and dynamic orchestration that leverage complementary strengths across detectors.","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"]