IndusAgent: Reinforcing Open-Vocabulary Industrial Anomaly Detection with Agentic Tools</p>\n","updatedAt":"2026-05-21T05:00:30.252Z","author":{"_id":"693d4ec248332487069f8004","avatarUrl":"/avatars/c27ff0cf7683717cc86200d7d0d0ec3d.svg","fullname":"Zhenlong Yuan","name":"ZhenlongYuan","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.5820629000663757},"editors":["ZhenlongYuan"],"editorAvatarUrls":["/avatars/c27ff0cf7683717cc86200d7d0d0ec3d.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.20682","authors":[{"_id":"6a0e90e4164dbbc68a26c617","name":"Rongbin Tan","hidden":false},{"_id":"6a0e90e4164dbbc68a26c618","name":"Fangfang Lin","hidden":false},{"_id":"6a0e90e4164dbbc68a26c619","name":"Zhenlong Yuan","hidden":false},{"_id":"6a0e90e4164dbbc68a26c61a","name":"Min Qiu","hidden":false},{"_id":"6a0e90e4164dbbc68a26c61b","name":"Kejin Cui","hidden":false},{"_id":"6a0e90e4164dbbc68a26c61c","name":"Mengmeng Wang","hidden":false},{"_id":"6a0e90e4164dbbc68a26c61d","name":"Yi Wang","hidden":false},{"_id":"6a0e90e4164dbbc68a26c61e","name":"Zijian Song","hidden":false},{"_id":"6a0e90e4164dbbc68a26c61f","name":"Zhiyuan Wang","hidden":false},{"_id":"6a0e90e4164dbbc68a26c620","name":"Jiyuan Wang","hidden":false},{"_id":"6a0e90e4164dbbc68a26c621","name":"Yue Wang","hidden":false},{"_id":"6a0e90e4164dbbc68a26c622","name":"Shuhan Song§","hidden":false},{"_id":"6a0e90e4164dbbc68a26c623","name":"Huawei Cao","hidden":false}],"publishedAt":"2026-05-20T00:00:00.000Z","submittedOnDailyAt":"2026-05-21T00:00:00.000Z","title":"IndusAgent: Reinforcing Open-Vocabulary Industrial Anomaly Detection with Agentic Tools","submittedOnDailyBy":{"_id":"693d4ec248332487069f8004","avatarUrl":"/avatars/c27ff0cf7683717cc86200d7d0d0ec3d.svg","isPro":false,"fullname":"Zhenlong Yuan","user":"ZhenlongYuan","type":"user","name":"ZhenlongYuan"},"summary":"Multimodal large language models (MLLMs) have shown remarkable capability in bridging visual perception and textual reasoning, enabling zero-shot understanding across diverse industrial scenarios. However, their performance in open-vocabulary industrial anomaly detection (IAD) is often limited by domain-misaligned reasoning and hallucinated structural inferences. To address these challenges, we propose IndusAgent, a tool-augmented agentic framework for open-vocabulary IAD. Specifically, we first construct Indus-CoT, a structured dataset that integrates global visual observations, high-resolution local patches, and expert normalcy priors, providing supervision for fine-tuning the model on rigorous industrial inspection trajectories. Building on this, IndusAgent dynamically orchestrates a set of external tools, including dynamic region cropping, high-frequency feature enhancement, and prior retrieval, thus enabling the agent to actively resolve visual ambiguities and disentangle subtle anomalies. Furthermore, we introduce a gated reinforcement learning objective that jointly optimizes anomaly classification, localization accuracy, anomaly type reasoning, and efficient tool usage, ensuring that tool invocation occurs only when beneficial. Extensive evaluations on five industrial anomaly benchmarks, including MVTec-AD, VisA, MPDD, DTD, and SDD, demonstrate that IndusAgent achieves state-of-the-art zero-shot performance among all existing methods, validating our robustness and generalization capacity.","upvotes":33,"discussionId":"6a0e90e4164dbbc68a26c624","ai_summary":"IndusAgent is a tool-augmented agentic framework for open-vocabulary industrial anomaly detection that improves performance through structured visual reasoning and dynamic tool utilization.","ai_keywords":["multimodal large language models","open-vocabulary industrial anomaly detection","tool-augmented agentic framework","Indus-CoT","dynamic region cropping","high-frequency feature enhancement","prior retrieval","gated reinforcement learning objective","anomaly classification","localization accuracy","anomaly type reasoning"],"organization":{"_id":"68ef0ab704f0f03d81964936","name":"ict-cas","fullname":"Institute of Computing Technology, Chinese Academy of Sciences","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/656ad93853703dd78f3de7b8/MR5EsF33Ev3IdnMOHv1be.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"693d4ec248332487069f8004","avatarUrl":"/avatars/c27ff0cf7683717cc86200d7d0d0ec3d.svg","isPro":false,"fullname":"Zhenlong Yuan","user":"ZhenlongYuan","type":"user"},{"_id":"6983188c7f99218c71ee1f64","avatarUrl":"/avatars/12d0df0cdfeeaa2373e86e5079947215.svg","isPro":false,"fullname":"Rongbin Tan","user":"survivours","type":"user"},{"_id":"6a0e91f55241b99cabf849f2","avatarUrl":"/avatars/643e62282c9331d836a5824b9020c309.svg","isPro":false,"fullname":"Fangfang Lin","user":"FangfangLin","type":"user"},{"_id":"6a0e9264147d922b7ec5d101","avatarUrl":"/avatars/f4317cc5a590c8ac8931db5a57f21a13.svg","isPro":false,"fullname":"Kevin Choi","user":"clarkgogogo","type":"user"},{"_id":"65755c08b238c76bbaed1c4c","avatarUrl":"/avatars/17790ad1a0940ad8450b7a967133bce8.svg","isPro":false,"fullname":"WangYuji","user":"YujiWang132","type":"user"},{"_id":"64faa838da4429f25dd19471","avatarUrl":"/avatars/635f33920a05f63ba7783c3ecc75a462.svg","isPro":false,"fullname":"WENHAO WANG","user":"wwh0411","type":"user"},{"_id":"67a2482de9c7f4628c86780a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/67a2482de9c7f4628c86780a/OwcZ24jSvPjpH6X0zo7Xh.jpeg","isPro":false,"fullname":"HAO XIURAN","user":"eggtartking","type":"user"},{"_id":"69251e48bcc86e2036f66e28","avatarUrl":"/avatars/cd2e12f9334d8d3cc73529b1192e9e30.svg","isPro":false,"fullname":"wang","user":"bswy","type":"user"},{"_id":"68ca4a4103a9520d2f6cac75","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/68ca4a4103a9520d2f6cac75/c_kHs2sXcUpRC7Biyp_K1.jpeg","isPro":false,"fullname":"ZhidongYang","user":"ZhidongYangRyan","type":"user"},{"_id":"69a68bc32820bb13bff6ddfe","avatarUrl":"/avatars/46cfc1bb672c26f5c328fb89c9330e77.svg","isPro":false,"fullname":"Shuhan Song","user":"Shuhannn97","type":"user"},{"_id":"69cccc9bef66c52b6c96cb89","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/uIe5B59uYyY4YvY9ROPis.png","isPro":false,"fullname":"Aria Perez","user":"hudsonadams94","type":"user"},{"_id":"69ccd0d9eb9cdf88f2a3124f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/1WwI27M4GYPcourBnWpaX.png","isPro":false,"fullname":"Lucas Rodriguez","user":"michael-hernand","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"68ef0ab704f0f03d81964936","name":"ict-cas","fullname":"Institute of Computing Technology, Chinese Academy of Sciences","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/656ad93853703dd78f3de7b8/MR5EsF33Ev3IdnMOHv1be.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.20682.md"}">
IndusAgent: Reinforcing Open-Vocabulary Industrial Anomaly Detection with Agentic Tools
Authors: ,
,
,
,
,
,
,
,
,
,
,
,
Abstract
IndusAgent is a tool-augmented agentic framework for open-vocabulary industrial anomaly detection that improves performance through structured visual reasoning and dynamic tool utilization.
AI-generated summary
Multimodal large language models (MLLMs) have shown remarkable capability in bridging visual perception and textual reasoning, enabling zero-shot understanding across diverse industrial scenarios. However, their performance in open-vocabulary industrial anomaly detection (IAD) is often limited by domain-misaligned reasoning and hallucinated structural inferences. To address these challenges, we propose IndusAgent, a tool-augmented agentic framework for open-vocabulary IAD. Specifically, we first construct Indus-CoT, a structured dataset that integrates global visual observations, high-resolution local patches, and expert normalcy priors, providing supervision for fine-tuning the model on rigorous industrial inspection trajectories. Building on this, IndusAgent dynamically orchestrates a set of external tools, including dynamic region cropping, high-frequency feature enhancement, and prior retrieval, thus enabling the agent to actively resolve visual ambiguities and disentangle subtle anomalies. Furthermore, we introduce a gated reinforcement learning objective that jointly optimizes anomaly classification, localization accuracy, anomaly type reasoning, and efficient tool usage, ensuring that tool invocation occurs only when beneficial. Extensive evaluations on five industrial anomaly benchmarks, including MVTec-AD, VisA, MPDD, DTD, and SDD, demonstrate that IndusAgent achieves state-of-the-art zero-shot performance among all existing methods, validating our robustness and generalization capacity.
Community
IndusAgent: Reinforcing Open-Vocabulary Industrial Anomaly Detection with Agentic Tools
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2605.20682 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.20682 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.20682 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.