{"id":7388,"date":"2025-09-08T05:07:00","date_gmt":"2025-09-07T20:07:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7388"},"modified":"2025-09-06T21:16:11","modified_gmt":"2025-09-06T12:16:11","slug":"ui-tars-2-technical-report-advancing-gui-agent-with-multi-turn-reinforcement-learning","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7388","title":{"rendered":"UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning\u00a0<\/strong>[151.0]<br>\u30b0\u30e9\u30d5\u30a3\u30ab\u30eb\u30e6\u30fc\u30b6\u30a4\u30f3\u30bf\u30d5\u30a7\u30fc\u30b9\u306e\u305f\u3081\u306e\u81ea\u5f8b\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u306e\u958b\u767a\u306f\u3001\u4eba\u5de5\u77e5\u80fd\u306b\u304a\u3051\u308b\u5927\u304d\u306a\u8ab2\u984c\u3092\u793a\u3057\u3066\u3044\u308b\u3002 \u672c\u7a3f\u3067\u306f,GUI\u4e2d\u5fc3\u306e\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u30e2\u30c7\u30eb\u3067\u3042\u308bUI-TARS-2\u3092\u63d0\u6848\u3059\u308b\u3002 \u5b9f\u8a3c\u7684\u306a\u8a55\u4fa1\u3067\u306f\u3001UI-TARS-2\u306f\u4ee5\u524d\u306eUI-TARS-1.5\u3088\u308a\u3082\u5927\u5e45\u306b\u6539\u5584\u3055\u308c\u3066\u3044\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2509.02544v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2509.02544v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Tue, 02 Sep 2025 17:44:45 GMT)<\/li>\n\n\n\n<li><a href=\"https:\/\/devneko.jp\/wordpress\/?p=6136\">UI-TARS: Pioneering Automated GUI Interaction with Native Agents \u2013 arXiv\u6700\u65b0\u8ad6\u6587\u306e\u7d39\u4ecb<\/a>, <a href=\"https:\/\/devneko.jp\/wordpress\/?p=6625\">UFO2: The Desktop AgentOS\u00a0, UI-TARS-1.5 \u2013 arXiv\u6700\u65b0\u8ad6\u6587\u306e\u7d39\u4ecb<\/a>\u306e\u30a2\u30c3\u30d7\u30c7\u30fc\u30c8\u3002\u300cEmpirical evaluation shows that UI-TARS-2 delivers significant improvements over UI-TARS-1.5 [56], achieving strong results in both GUI-based interaction and game environments. On GUI benchmarks, the model reaches 88.2 on Online-Mind2Web [77], 47.5 on OSWorld [75], 50.6 on WindowsAgentArena [10], and 73.3 on AndroidWorld [52], representing clear gains over the previous generation and outperforming strong baselines such as Claude and OpenAI agents in multiple cases.\u300d\u3068\u524d\u56de\u30e2\u30c7\u30eb\u306b\u6bd4\u3079\u5927\u304d\u306a\u6539\u5584\u3092\u4e3b\u5f35\u3002\u4e0b\u8a18\u304c\u6539\u5584\u70b9\u3068\u3044\u3046\u3053\u3068\u3067\u306f\u3042\u308b\u304c\u3001\u6700\u521d\u306e\u30d0\u30fc\u30b8\u30e7\u30f3\u304b\u3089\u3084\u308c\u308b\u3053\u3068\u306f\u5168\u90e8\u3084\u308b\u3068\u3044\u3046\u96f0\u56f2\u6c17\u304c\u3059\u3054\u3044\n<ul class=\"wp-block-list\">\n<li>First, to mitigate data scarcity, we design a scalable Data Flywheel that co-evolves the model and its training corpus through continual pretraining, supervised fine-tuning, rejection sampling, and multiturn RL<\/li>\n\n\n\n<li>Second, to overcome the difficulties of scalable multi-turn RL, we design a training framework that stabilizes optimization in long-horizon settings. <\/li>\n\n\n\n<li>Third, to move beyond the limitations of pure GUI interaction, we construct a hybrid GUI-centered environment that augments on-screen actions with access to complementary resources such as file systems, terminals, and other external tools, enabling agents to solve a broader spectrum of realistic workflows. <\/li>\n\n\n\n<li>Fourth, to support large-scale training and evaluation, we build a unified sandbox platform capable of orchestrating heterogeneous environments\u2014ranging from cloud VMs for GUI interaction to browser-based sandboxes for games\u2014under a consistent API. <\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/bytedance\/ui-tars\">GitHub &#8211; bytedance\/UI-TARS<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[181],"class_list":["post-7388","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-gui-agent"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7388","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7388"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7388\/revisions"}],"predecessor-version":[{"id":7390,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7388\/revisions\/7390"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7388"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7388"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7388"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}