Strong general computer use — plus one demonstration when the instruction alone is not enough.</p>\n","updatedAt":"2026-08-18T03:17:30.946Z","author":{"_id":"6258a6455ea3a0a9b6de3f22","avatarUrl":"/avatars/6eeed72a97fb24465e5e65583fbe50cf.svg","fullname":"Lei Ke","name":"lkeab","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":16,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8474454879760742},"editors":["lkeab"],"editorAvatarUrls":["/avatars/6eeed72a97fb24465e5e65583fbe50cf.svg"],"reactions":[{"reaction":"👍","users":["Yingchen","speedinghzl","TberiusJunyao","leiwx52","dreamerdeo"],"count":5}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.15930","authors":[{"_id":"6a83ce55675db694db8cd53a","name":"Zihan Ding","hidden":false},{"_id":"6a83ce55675db694db8cd53b","user":{"_id":"6214e4ee1e35c843d42d1f88","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6214e4ee1e35c843d42d1f88/xEkzClh_6Zq1N9btiVPag.jpeg","isPro":false,"fullname":"Longxu Dou","user":"dreamerdeo","type":"user","name":"dreamerdeo"},"name":"Longxu Dou","status":"claimed_verified","statusLastChangedAt":"2026-08-18T08:45:05.183Z","hidden":false},{"_id":"6a83ce55675db694db8cd53c","name":"Qi Gao","hidden":false},{"_id":"6a83ce55675db694db8cd53d","name":"Xiangwu Guo","hidden":false},{"_id":"6a83ce55675db694db8cd53e","name":"Shengchao Hu","hidden":false},{"_id":"6a83ce55675db694db8cd53f","name":"Zilong Huang","hidden":false},{"_id":"6a83ce55675db694db8cd540","name":"Zihang Jiang","hidden":false},{"_id":"6a83ce55675db694db8cd541","name":"Lei Ke","hidden":false},{"_id":"6a83ce55675db694db8cd542","name":"Mengcheng Lan","hidden":false},{"_id":"6a83ce55675db694db8cd543","name":"Weixian Lei","hidden":false},{"_id":"6a83ce55675db694db8cd544","name":"Hanxuan Li","hidden":false},{"_id":"6a83ce55675db694db8cd545","name":"Honglin Li","hidden":false},{"_id":"6a83ce55675db694db8cd546","name":"Xiyun Li","hidden":false},{"_id":"6a83ce55675db694db8cd547","name":"Zaitang Li","hidden":false},{"_id":"6a83ce55675db694db8cd548","name":"Leowei Liang","hidden":false},{"_id":"6a83ce55675db694db8cd549","name":"Xin Luo","hidden":false},{"_id":"6a83ce55675db694db8cd54a","name":"Haozhe Ma","hidden":false},{"_id":"6a83ce55675db694db8cd54b","name":"Jiayi Mao","hidden":false},{"_id":"6a83ce55675db694db8cd54c","name":"Zhoujie Pan","hidden":false},{"_id":"6a83ce55675db694db8cd54d","name":"Can Qin","hidden":false},{"_id":"6a83ce55675db694db8cd54e","name":"Tianyuan Qu","hidden":false},{"_id":"6a83ce55675db694db8cd54f","name":"Weiqi Wang","hidden":false},{"_id":"6a83ce55675db694db8cd550","name":"Wenkai Wang","hidden":false},{"_id":"6a83ce55675db694db8cd551","name":"Yonglin Wang","hidden":false},{"_id":"6a83ce55675db694db8cd552","name":"Yuxin Wang","hidden":false},{"_id":"6a83ce55675db694db8cd553","name":"Chenxu Wu","hidden":false},{"_id":"6a83ce55675db694db8cd554","name":"Yingchen Yu","hidden":false},{"_id":"6a83ce55675db694db8cd555","name":"Chenyu Zhang","hidden":false},{"_id":"6a83ce55675db694db8cd556","name":"Yuhao Zheng","hidden":false}],"publishedAt":"2026-08-16T00:00:00.000Z","submittedOnDailyAt":"2026-08-18T00:00:00.000Z","title":"UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations","submittedOnDailyBy":{"_id":"6258a6455ea3a0a9b6de3f22","avatarUrl":"/avatars/6eeed72a97fb24465e5e65583fbe50cf.svg","isPro":false,"fullname":"Lei Ke","user":"lkeab","type":"user","name":"lkeab"},"summary":"Foundation GUI agents can automate complex digital tasks, but deployment is hindered by scarce and biased training data, ambiguous prompts, and unreliable execution. Routine workflows rely on user-specific tools and tacit conventions, so unstated instructions can produce arbitrary variations across runs. We present UI-Mate, a foundation GUI agent that integrates an environment-grounded training stack with in-context demonstration learning. UI-Mate makes three contributions: A Scalable Environment-Grounded Training Stack: A closed-loop data engine automates task generation, environment construction, rollout, filtering, capability balancing, SFT, and online RL across massively parallel environments via unified task-verifier bundles. In-Context Demonstration Learning: A mechanism that transforms multimodal demonstrations into flexible subtask-level workflows, follows relevant demonstrated steps, and re-plans from the live interface. OSWorkerBench Benchmark and Insights: A benchmark of 100 long-horizon office tasks across 41 applications that supports instruction-only and demonstration-guided evaluation. Its demonstration resources separate a 33-task self-demo setting, built from successful strong-agent rollouts of the same targets, from a 45-task variant-demo setting, built from human recordings of related but non-identical tasks. Experiments show that UI-Mate-27B sets a new open-weight state of the art on general computer-use benchmarks, scoring 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena. On OSWorkerBench, it reaches 41.0% strict success and 76.9% progress, outperforming its Qwen3.6-27B base by 17.7 and 24.5 points. On the 33-task self-demo subset, one demonstration raises strict success from 17.2% to 35.4% and progress from 67.9% to 81.1%, substantially improving long-horizon reliability. Project page: https://ui-mate.github.io.","upvotes":26,"discussionId":"6a83ce55675db694db8cd557","projectPage":"https://ui-mate.github.io/","githubRepo":"https://github.com/Tencent/UI-Mate","githubRepoAddedBy":"user","ai_summary":"UI-Mate is a foundation GUI agent that uses environment-grounded training and in-context demonstration learning to improve reliability on long-horizon office tasks, achieving state-of-the-art results on computer-use benchmarks.","ai_keywords":["foundation GUI agent","environment-grounded training","in-context demonstration learning","closed-loop data engine","SFT","online RL","task-verifier bundles","multimodal demonstrations","subtask-level workflows","OSWorkerBench","self-demo","variant-demo","OSWorld-Verified","WindowsAgentArena"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":6,"organization":{"_id":"6645f953c39288df638dbdd5","name":"Tencent-Hunyuan","fullname":"Tencent Hunyuan","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/62d22496c58f969c152bcefd/woKSjt2wXvBNKussyYPsa.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6258a6455ea3a0a9b6de3f22","avatarUrl":"/avatars/6eeed72a97fb24465e5e65583fbe50cf.svg","isPro":false,"fullname":"Lei Ke","user":"lkeab","type":"user"},{"_id":"638ee21e2cc490759feb433b","avatarUrl":"/avatars/2bd1068b6b9eb2447d2b761f69d8c7ae.svg","isPro":false,"fullname":"Yingchen Yu","user":"Yingchen","type":"user"},{"_id":"66a3b73ac44ffdaf67aa0a74","avatarUrl":"/avatars/ed73f39ea541ddcc35b7405af42485e5.svg","isPro":false,"fullname":"Xiyun Li","user":"xiyun98","type":"user"},{"_id":"63a430ec84a6a25c65bff20c","avatarUrl":"/avatars/3ab51214c491f28d8fe6b70c373d72aa.svg","isPro":false,"fullname":"Zilong Huang","user":"speedinghzl","type":"user"},{"_id":"62b9524a5ac8a1fccbce572d","avatarUrl":"/avatars/6c06788e69aec094df884b8ac25e4b16.svg","isPro":false,"fullname":"Stan Lei","user":"leiwx52","type":"user"},{"_id":"6a83d2adbe5309eb8b85e123","avatarUrl":"/avatars/2c2c221b0bbf043bf82155913f79a43c.svg","isPro":false,"fullname":"Hugo","user":"ycyu915","type":"user"},{"_id":"66f2432d5405e6677d430bbf","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/gxQWA3WTHGEQUL10fiIFh.png","isPro":false,"fullname":"GUO XIANGWU","user":"SamuelGuo","type":"user"},{"_id":"625aa04e535747b1a15cc14b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/625aa04e535747b1a15cc14b/uhxpJprM4wW5m67OPIbRp.png","isPro":false,"fullname":"Weiqi Wang","user":"mightyweaver","type":"user"},{"_id":"641129818573c51c0458b793","avatarUrl":"/avatars/d4bc67c160a07146cf41c614678aa36b.svg","isPro":false,"fullname":"Tianqing Fang","user":"tqfang229","type":"user"},{"_id":"6587e349f8b453e1f54b1370","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6587e349f8b453e1f54b1370/K4I2G8v8MS6JC39GbKQlq.jpeg","isPro":false,"fullname":"zongqing","user":"zongqing0068","type":"user"},{"_id":"638f4354b0525fa370498bad","avatarUrl":"/avatars/4eb188a4df14b7029e3a32fbc29795e7.svg","isPro":false,"fullname":"lan mengcheng","user":"lmc22","type":"user"},{"_id":"68266b5261ed4d89177c3612","avatarUrl":"/avatars/e9d214c78916e3cfa554e545fd831446.svg","isPro":false,"fullname":"Kishan Panaganti","user":"kishanpb","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6645f953c39288df638dbdd5","name":"Tencent-Hunyuan","fullname":"Tencent Hunyuan","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/62d22496c58f969c152bcefd/woKSjt2wXvBNKussyYPsa.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.15930.md","query":{}}">
UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations
Published on Aug 16
· Submitted by Lei Ke on Aug 18 Abstract
UI-Mate is a foundation GUI agent that uses environment-grounded training and in-context demonstration learning to improve reliability on long-horizon office tasks, achieving state-of-the-art results on computer-use benchmarks.
Foundation GUI agents can automate complex digital tasks, but deployment is hindered by scarce and biased training data, ambiguous prompts, and unreliable execution. Routine workflows rely on user-specific tools and tacit conventions, so unstated instructions can produce arbitrary variations across runs. We present UI-Mate, a foundation GUI agent that integrates an environment-grounded training stack with in-context demonstration learning. UI-Mate makes three contributions: A Scalable Environment-Grounded Training Stack: A closed-loop data engine automates task generation, environment construction, rollout, filtering, capability balancing, SFT, and online RL across massively parallel environments via unified task-verifier bundles. In-Context Demonstration Learning: A mechanism that transforms multimodal demonstrations into flexible subtask-level workflows, follows relevant demonstrated steps, and re-plans from the live interface. OSWorkerBench Benchmark and Insights: A benchmark of 100 long-horizon office tasks across 41 applications that supports instruction-only and demonstration-guided evaluation. Its demonstration resources separate a 33-task self-demo setting, built from successful strong-agent rollouts of the same targets, from a 45-task variant-demo setting, built from human recordings of related but non-identical tasks. Experiments show that UI-Mate-27B sets a new open-weight state of the art on general computer-use benchmarks, scoring 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena. On OSWorkerBench, it reaches 41.0% strict success and 76.9% progress, outperforming its Qwen3.6-27B base by 17.7 and 24.5 points. On the 33-task self-demo subset, one demonstration raises strict success from 17.2% to 35.4% and progress from 67.9% to 81.1%, substantially improving long-horizon reliability. Project page: https://ui-mate.github.io.
Community
Strong general computer use — plus one demonstration when the instruction alone is not enough.
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/2608.15930 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.15930 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.15930 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.