
{"id":9603,"date":"2026-06-25T15:38:46","date_gmt":"2026-06-25T07:38:46","guid":{"rendered":"https:\/\/infernews.com\/blog\/skillharness\/"},"modified":"2026-06-25T15:39:36","modified_gmt":"2026-06-25T07:39:36","slug":"skillharness","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/skillharness\/","title":{"rendered":"SkillHarness\uff1a\u5e6b CUA \u5b78\u5f97\u66f4\u5b89\u5168"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-b92e6b6b8487.jpg\" alt=\"Repository image for YurunChen\/SkillHarness\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u662f\u4e00\u500b\u7814\u7a76\u539f\u578b\uff0c\u73fe\u6642 GitHub \u5132\u5b58\u5eab\u4e3b\u8981\u63d0\u4f9b SkillHarness \u8ad6\u6587 PDF\u3002\u5b83\u8981\u89e3\u6c7a\u7684\u662f Computer-Use Agents\uff08CUAs\uff09\u5728\u52d5\u614b\u96fb\u8166\u4ecb\u9762\u4e2d\u5b78\u7fd2\u8207\u91cd\u7528\u6280\u80fd\u6642\uff0c\u5bb9\u6613\u53d7 prompt injections\u3001\u5f48\u7a97\u8207\u74b0\u5883\u8b8a\u5316\u5f71\u97ff\uff0c\u4ee4\u5df2\u5b78\u5230\u7684\u6280\u80fd\u8b8a\u5f97\u5371\u96aa\u6216\u4e0d\u7a69\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u505a\u6cd5\u591a\u6578\u6cbf\u7528\u300c\u5f9e\u6210\u529f\u8ecc\u8de1\u62bd\u53d6\u53ef\u91cd\u7528\u6280\u80fd\u300d\u9019\u500b\u7bc4\u5f0f\uff0c\u5e38\u898b\u8868\u9054\u5f62\u5f0f\u5305\u62ec\u51fd\u5f0f\u6216 API\uff0c\u4f46\u4f5c\u8005\u8a8d\u70ba\u9019\u985e\u65b9\u6cd5\u9810\u8a2d\u74b0\u5883\u662f static and safe\u3002SkillHarness \u6539\u7528 safety-constrained interaction process \u53bb\u770b\u5f85\u6280\u80fd\u7684\u5b78\u7fd2\u8207\u4f7f\u7528\uff0c\u6838\u5fc3\u4e0d\u662f\u591a\u5b78\u5e7e\u500b\u6280\u80fd\uff0c\u800c\u662f\u5148\u5224\u65b7\u54ea\u4e9b\u6280\u80fd\u5728\u7576\u4e0b\u60c5\u5883\u4ecd\u7136\u5b89\u5168\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ad6\u6587\u63d0\u51fa\u5169\u500b\u8fa8\u8b58\u5ea6\u5f88\u9ad8\u7684\u8a2d\u8a08\uff1a\u4e00\u662f <strong>skill boundary<\/strong>\uff0c\u7528 multi-source supervision signals \u5f9e\u4e92\u52d5\u8ecc\u8de1\u4e2d\u627e\u51fa safe skills\uff1b\u4e8c\u662f <strong>selective skill reuse<\/strong>\uff0c\u6309\u7576\u524d\u60c5\u5883\u62c6\u89e3\u4efb\u52d9\uff0c\u53ea\u555f\u52d5\u90e8\u5206\u6280\u80fd\uff0c\u800c\u4e0d\u662f\u6574\u5305\u7167\u642c\u3002\u9019\u7a2e\u53d6\u5411\u7684\u4ee3\u50f9\uff0c\u662f\u7cfb\u7d71\u8a2d\u8a08\u6703\u6bd4\u55ae\u7d14\u6536\u96c6\u6210\u529f\u8ecc\u8de1\u66f4\u8907\u96dc\uff0c\u4f46\u63db\u4f86\u7684\u662f\u5728\u52d5\u614b\u74b0\u5883\u4e0b\u66f4\u7a69\u5b9a\u7684\u884c\u70ba\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u5132\u5b58\u5eab\u672a\u9644\u7a0b\u5f0f\u78bc\uff0c\u6240\u4ee5\u66ab\u6642\u4e0d\u80fd\u76f4\u63a5\u90e8\u7f72\u6216\u91cd\u8dd1\u5be6\u9a57\uff1b\u8f03\u5408\u7406\u7684\u7406\u89e3\u65b9\u5f0f\uff0c\u662f\u5148\u628a\u5b83\u7576\u6210\u4e00\u5957 CUA \u5b89\u5168\u6280\u80fd\u6846\u67b6\u4f86\u8b80\u3002\u82e5\u4e4b\u5f8c\u4f5c\u8005\u91cb\u51fa\u5be6\u4f5c\uff0c\u6700\u9700\u8981\u89c0\u5bdf\u7684\u6703\u662f\u5b83\u600e\u6a23\u63a5\u5165\u4ee3\u7406\u7684\u8ecc\u8de1\u8cc7\u6599\u3001\u600e\u6a23\u5efa\u7acb safety constraints\uff0c\u4ee5\u53ca\u80fd\u5426\u5728 OSWorld \u4e00\u985e\u96fb\u8166\u64cd\u4f5c\u57fa\u6e96\u4ee5\u5916\u7dad\u6301\u6548\u679c\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u985e\u578b\u5c6c\u65bc<strong>\u6846\u67b6\uff0f\u7814\u7a76\u8ad6\u6587\u9805\u76ee<\/strong>\uff0c\u91cd\u9ede\u5728\u5b89\u5168\u6280\u80fd\u5b78\u7fd2\uff0c\u4e0d\u662f\u5373\u88dd\u5373\u7528\u5de5\u5177<\/li>\n\n\n\n<li>\u4e3b\u8981\u6279\u8a55\u820a\u65b9\u6cd5\u4f9d\u8cf4 static and safe environment \u5047\u8a2d\uff0c\u653e\u5230\u52d5\u614b\u5834\u666f\u6703\u5b78\u5230\u4e0d\u5b89\u5168\u6280\u80fd<\/li>\n\n\n\n<li>\u8ad6\u6587\u8072\u7a31 learned skills \u7684 unsafe rate \u964d\u4f4e <strong>57.1%<\/strong>\uff0c\u4e26\u63d0\u5347\u52d5\u614b\u74b0\u5883\u4e0b\u7684 execution stability<\/li>\n\n\n\n<li>\u8f03\u9069\u5408\u7814\u7a76 Computer-Use Agents\u3001\u4ee3\u7406\u5b89\u5168\u3001\u684c\u9762\u81ea\u52d5\u5316\u8207\u9577\u6d41\u7a0b\u4efb\u52d9\u7684\u5718\u968a\u7559\u610f<\/li>\n\n\n\n<li>\u76f8\u95dc\u8108\u7d61\u6a21\u578b\u8207\u65b9\u6cd5\u5305\u62ec Computer-Use Agents\uff08CUAs\uff09\u3001Voyager\u3001ASI\uff0c\u4ee5\u53ca\u4ee5\u51fd\u5f0f\uff0fAPI \u5f62\u5f0f\u5c01\u88dd\u6280\u80fd\u7684\u8def\u7dda<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/YurunChen\/SkillHarness\" rel=\"noopener noreferrer\">https:\/\/github.com\/YurunChen\/SkillHarness<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.20636\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.20636<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u7bc7\u805a\u7126 Computer-Use Agents \u5b89\u5168\u6280\u80fd\u5b78\u7fd2\u7684\u7814\u7a76\u539f\u578b\u3002\u5b83\u91dd\u5c0d\u52d5\u614b\u4ecb\u9762\u8207 prompt injections\uff0c\u91cd\u505a\u6280\u80fd\u5b78\u7fd2\u8207\u91cd\u7528\u65b9\u5f0f\u3002<\/p>\n","protected":false},"author":8,"featured_media":9602,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,116,137,38,170,150,76,197,201],"tags":[],"class_list":["post-9603","post","type-post","status-publish","format-standard","hentry","category-133","category-agentic","category-api","category-38","category-170","category-150","category-76","category-framework","category-skill-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9603","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=9603"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9603\/revisions"}],"predecessor-version":[{"id":9605,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9603\/revisions\/9605"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9602"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9603"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9603"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9603"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}