{"id":12778,"date":"2026-10-01T00:01:23","date_gmt":"2026-09-30T16:01:23","guid":{"rendered":"https:\/\/infernews.com\/blog\/hybridcua\/"},"modified":"2026-10-01T00:17:56","modified_gmt":"2026-09-30T16:17:56","slug":"hybridcua","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/hybridcua\/","title":{"rendered":"HybridCUA \u8b93 Agent \u61c2\u5f97\u4f55\u6642\u7528\u547d\u4ee4\u5217"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

\u6d59\u6c5f\u5927\u5b78\u3001\u5317\u4eac\u5927\u5b78\u8207\u6e05\u83ef\u5927\u5b78\u7684\u7814\u7a76\u5718\u968a\uff0c\u5c07 CUA (Computer-use agents) \u5f9e\u55ae\u9760\u6ed1\u9f20\u9375\u76e4\u64cd\u4f5c\uff0c\u5e36\u5230 GUI \u8207 command line interface\uff08CLI\uff09\u5354\u4f5c\u7684\u5de5\u4f5c\u65b9\u5f0f\u3002HybridCUA \u5c6c\u65bc\u8a13\u7df4\u6846\u67b6\u53ca\u6a21\u578b\u9805\u76ee\uff0c\u8655\u7406\u7684\u662f\u4ee3\u7406\u77e5\u9053\u53ef\u4ee5\u7528 shell \u4e4b\u5f8c\uff0c\u4ecd\u672a\u80fd\u5224\u65b7\u4f55\u6642\u4f7f\u7528\u3001\u5982\u4f55\u4f7f\u7528\u7684\u554f\u984c\u3002<\/p>\n\n\n\n

\u9023\u4e32\u7684 GUI \u64cd\u4f5c\u53ef\u4ee5\u7531\u4e00\u689d\u547d\u4ee4\u53d6\u4ee3\uff0c\u4f46 CLI \u66b4\u9732\u7d66\u672a\u53d7\u8a13\u7df4\u7684\u4ee3\u7406\uff0c\u53cd\u800c\u4ee4 OSWorld \u6e96\u78ba\u7387\u4e0b\u964d 2.5 \u81f3 11.5 \u500b\u767e\u5206\u9ede\u3002HybridCUA-8K \u6536\u96c6 GUI only\u3001CLI only\uff0c\u4ee5\u53ca\u4ea4\u932f\u4f7f\u7528\u5169\u8005\u7684\u8ecc\u8de1\uff0c\u518d\u914d\u5408 supervised fine-tuning \u548c CLI-aware rewards\uff0c\u8a13\u7df4 HybridCUA-9B \u9078\u64c7\u8f03\u5408\u9069\u7684\u64cd\u4f5c\u65b9\u5f0f\u3002<\/p>\n\n\n\n

\"HybridCUA<\/figure>\n\n\n\n

\u2022 HybridCUA-9B \u5728 OSWorld \u9054\u5230 53.6%\uff0c\u8f03\u57fa\u790e\u6a21\u578b\u9ad8 14.8 \u500b\u767e\u5206\u9ede
\n\u2022 \u5e73\u5747\u6bcf\u9805\u4efb\u52d9\u4f7f\u7528 14.0 \u6b65\uff0c\u4e26\u53ef\u8f49\u79fb\u81f3 OSWorld-MCP \u53ca Windows
\n\u2022 HybridCUA-8K \u5305\u542b 5K \u689d\u6df7\u5408\u8ecc\u8de1\u53ca 3K \u500b\u5df2\u9a57\u8b49 RLVR \u4efb\u52d9
\n\u2022 CLI reward \u5206\u70ba trajectory level \u8207 step level\uff0c\u5206\u5225\u9f13\u52f5\u6210\u529f\u5b8c\u6210\u53ca\u6e1b\u5c11\u5931\u6557\u6307\u4ee4\u6210\u672c<\/p>\n\n\n\n

\u5132\u5b58\u5eab\u4ee5 env_infra\u3001online-rl \u548c site \u5206\u958b\u74b0\u5883\u5e73\u53f0\u3001GRPO \u8a13\u7df4\u6d41\u7a0b\u53ca\u9805\u76ee\u7db2\u7ad9\uff1b\u524d\u5169\u8005\u900f\u904e HTTP \u7684 \/v1\/sessions \u5354\u4f5c\uff0c\u8a13\u7df4\u5668\u6bcb\u9808\u81ea\u884c\u555f\u52d5 VM \u6216\u5bb9\u5668\u3002README \u63d0\u4f9b\u5b89\u88dd\u8173\u672c\u53ca GPU\u3001nvcc\u3001Ray \u7b49\u74b0\u5883\u8981\u6c42\uff0c\u4f46\u6a21\u578b\u4ecd\u6a19\u793a\u70ba coming soon\uff0c\u56e0\u6b64\u76ee\u524d\u8f03\u9069\u5408\u7814\u7a76 Computer-use agents\u3001GUI automation \u6216\u8de8\u5e73\u53f0\u4ee3\u7406\u8a13\u7df4\u7684\u5718\u968a\u95b1\u8b80\u8cc7\u6599\u96c6\u8207\u8a13\u7df4\u67b6\u69cb\u3002<\/p>\n\n\n\n

\u7d50\u679c\u5728 WindowsAgentArena \u63d0\u5347 4.0 \u500b\u767e\u5206\u9ede\uff0cOSWorld-MCP \u9054 47.1%\uff0cWindows \u9054 36.0%\uff0c\u986f\u793a GUI \u52a0 CLI \u7684\u65b9\u6cd5\u5177\u5099\u4e00\u5b9a\u8de8\u5e73\u53f0\u80fd\u529b\uff1b\u4e0d\u904e\u9805\u76ee\u4ecd\u4f9d\u8cf4\u8907\u96dc\u7684\u5206\u6563\u5f0f\u8a13\u7df4\u74b0\u5883\uff0c\u672a\u63d0\u4f9b\u5373\u6642\u53ef\u4e0b\u8f09\u6a21\u578b\uff0c\u8ddd\u96e2\u4e00\u822c\u4f7f\u7528\u8005\u76f4\u63a5\u5b89\u88dd\u4ecd\u6709\u8ddd\u96e2\u3002<\/p>\n\n\n\n

\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 GitHub<\/strong><\/a> \u00b7 \u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

\u6d59\u6c5f\u5927\u5b78\u3001\u5317\u4eac\u5927\u5b78\u8207\u6e05\u83ef\u5927\u5b78\u5718\u968a\uff0c\u8b93 Computer-use agents \u5728 GUI \u64cd\u4f5c\u8207 CLI \u6307\u4ee4\u4e4b\u9593\u4f5c\u51fa\u9078\u64c7\uff0c\u6e1b\u5c11\u5197\u9577\u9ede\u64ca\u6d41\u7a0b\u53ca\u9023\u9396\u932f\u8aa4\u3002<\/p>\n","protected":false},"author":8,"featured_media":12777,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,195,196,116,151,127,150,199],"tags":[],"class_list":["post-12778","post","type-post","status-publish","format-standard","hentry","category-133","category-195","category-196","category-agentic","category-mcp","category-127","category-150","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12778","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=12778"}],"version-history":[{"count":3,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12778\/revisions"}],"predecessor-version":[{"id":12787,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12778\/revisions\/12787"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12777"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12778"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12778"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12778"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}