{"id":106867,"date":"2026-09-18T13:00:44","date_gmt":"2026-09-18T05:00:44","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/106867.html"},"modified":"2026-09-18T13:00:44","modified_gmt":"2026-09-18T05:00:44","slug":"skill-state%e6%b7%b1%e5%85%a5%e8%a7%a3%e6%9e%90%ef%bc%9a%e6%8a%8aai-agent%e9%95%bf%e7%a8%8b%e6%89%a7%e8%a1%8c%e5%8f%98%e6%88%90%e5%8f%af%e6%a0%a1%e9%aa%8c%e7%9a%84%e7%8a%b6%e6%80%81%e8%bf%81%e7%a7%bb","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/106867.html","title":{"rendered":"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb"},"content":{"rendered":"

\u5199\u5728\u524d\u9762<\/h2>\n

\"\u56fe\u7247\"<\/p>\n

\u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u516c\u4f17\u53f7：WeThinkIn \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-Interview-Book，\u6b22\u8fce\u5927\u5bb6Star～<\/p>\n

Rocky\u6700\u65b0\u64b0\u5199\u768410\u4e07\u5b57AI Agent（AI\u667a\u80fd\u4f53）\u6df1\u5165\u6d45\u51fa\u5168\u7ef4\u5ea6\u89e3\u6790\u6587\u7ae0： \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AI Agent（AI\u667a\u80fd\u4f53）\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

AIGC\/LLM\/AI Agent\u7b97\u6cd5\u5c97\/\u5f00\u53d1\u5c97\u6c42\u804c\u9762\u8bd5\u5185\u63a8\u5b66\u4e60\u793e\u7fa4（\u6db5\u76d6AIGC\u3001LLM\u5927\u6a21\u578b\u3001AI Agent\u3001\u4f20\u7edf\u6df1\u5ea6\u5b66\u4e60\u3001\u81ea\u52a8\u9a7e\u9a76\u3001\u673a\u5668\u5b66\u4e60\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u5927\u6570\u636e\u6316\u6398\u3001\u5177\u8eab\u667a\u80fd\u3001\u5143\u5b87\u5b99\u3001AGI\u7b49AI\u884c\u4e1a\u6700\u65b0\u9762\u8bd5\u5e72\u8d27\u7ecf\u9a8c\u4e0e\u6838\u5fc3\u77e5\u8bc6）\u6b22\u8fce\u5927\u5bb6\u52a0\u5165：https:\/\/t.zsxq.com\/33pJ0<\/p>\n

\n

\u5927\u5bb6\u597d，\u6211\u662fRocky\u3002<\/p>\n

\u6838\u5fc3\u5bfc\u8bfb：Agent\u8fd0\u884c\u5f97\u8d8a\u4e45，\u4e3a\u4ec0\u4e48\u8d8a\u5bb9\u6613\u88ab\u81ea\u5df1\u7684\u5386\u53f2\u62d6\u7d2f？<\/h3>\n

\u4e00\u4e2a\u4ed3\u5e93 Agent \u5df2\u7ecf\u5904\u7406\u4e86\u4e0a\u767e\u6761\u6307\u4ee4：\u7269\u54c1\u5165\u5e93\u3001\u8d27\u67b6\u8c03\u6574\u3001\u8ba2\u5355\u51fa\u5e93，\u4ee5\u53ca\u7a7f\u63d2\u5176\u4e2d\u7684\u4f20\u611f\u5668\u65e5\u5fd7\u3002\u7b2c 101 \u6b65，\u5b83\u9700\u8981\u56de\u7b54\u7684\u53ef\u80fd\u53ea\u662f\u201citem_12 \u73b0\u5728\u5728\u54ea\u91cc\u201d\u3002\u5982\u679c\u7cfb\u7edf\u6bcf\u6b21\u90fd\u628a\u6b64\u524d\u7684\u5bf9\u8bdd\u3001\u5de5\u5177\u8f93\u51fa\u548c\u63a8\u7406\u91cd\u65b0\u4ea4\u7ed9\u6a21\u578b，\u8fd9\u4e2a\u7b80\u5355\u95ee\u9898\u5c31\u53d8\u6210\u4e86\u53e6\u4e00\u9879\u4efb\u52a1：\u4ece\u4e0d\u65ad\u589e\u957f\u7684\u6587\u5b57\u4e2d\u91cd\u5efa\u5f53\u524d\u4e16\u754c，\u5e76\u5224\u65ad\u54ea\u4e9b\u4fe1\u606f\u5df2\u7ecf\u5931\u6548\u3002<\/p>\n

\u957f\u4e0a\u4e0b\u6587\u53ef\u4ee5\u5bb9\u7eb3\u66f4\u591a\u5386\u53f2，\u5374\u4e0d\u4f1a\u81ea\u52a8\u5224\u65ad\u54ea\u6761\u65e7\u4e8b\u5b9e\u5df2\u7ecf\u88ab\u65b0\u4e8b\u5b9e\u8986\u76d6\u3002\u4e00\u6b21\u5931\u8d25\u547d\u4ee4\u3001\u4e00\u4e2a\u8fc7\u65f6\u4f4d\u7f6e\u3001\u4e00\u6bb5\u9519\u8bef\u5047\u8bbe，\u90fd\u53ef\u80fd\u7ee7\u7eed\u7559\u5728\u540e\u7eed\u8f93\u5165\u91cc\u3002\u8fd9\u91cc\u7684\u95ee\u9898\u5df2\u7ecf\u8fdb\u5165\u8fd0\u884c\u65f6\u8bbe\u8ba1：\u7cfb\u7edf\u7a76\u7adf\u628a\u4ec0\u4e48\u89c6\u4e3a\u6267\u884c\u7684\u4f9d\u636e，\u53c8\u5141\u8bb8\u54ea\u4e9b\u4fe1\u606f\u8de8\u6b65\u9aa4\u4fdd\u7559？<\/p>\n

Sanket Badhe\u3001Priyanka Tiwari \u548c Jonghyun Chung \u63d0\u51fa\u7684 SKILL.state: Scalable Long-Horizon Agent Skills，\u7ed9\u51fa\u4e00\u4e2a\u660e\u786e\u65b9\u6848：\u6bcf\u4e00\u6b65\u53ea\u8f93\u5165\u4e0d\u53ef\u53d8\u6280\u80fd\u89c4\u8303\u3001\u7ed3\u6784\u5316\u6267\u884c\u72b6\u6001\u548c\u6700\u65b0\u89c2\u5bdf；\u6a21\u578b\u5b8c\u6210\u672c\u6b65\u63a8\u7406\u540e，\u8fd0\u884c\u65f6\u6821\u9a8c\u5e76\u5e94\u7528\u72b6\u6001\u8865\u4e01，\u540e\u7eed\u6b65\u9aa4\u4e0d\u518d\u91cd\u653e\u6b64\u524d\u7684\u63a8\u7406\u8f68\u8ff9\u3002\u672c\u6587\u8ba8\u8bba arXiv:2608.26263 \u7684 v3 \u7248\u672c，\u4f5c\u8005\u673a\u6784\u4e3a Google LLC \u4e0e Purdue University\u3002<\/p>\n

Rocky\u8ba4\u4e3a，\u8fd9\u7bc7\u8bba\u6587\u6700\u6709\u4ef7\u503c\u7684\u5224\u65ad，\u662f\u628a\u201c\u5f53\u524d\u6267\u884c\u4e8b\u5b9e\u7684\u7ef4\u62a4\u201d\u4ece\u6a21\u578b\u53cd\u590d\u9605\u8bfb\u5386\u53f2\u7684\u9690\u542b\u80fd\u529b，\u63d0\u5347\u4e3a\u8fd0\u884c\u65f6\u7684\u663e\u5f0f\u8d23\u4efb\u3002 \u6a21\u578b\u4ecd\u8d1f\u8d23\u63a8\u7406，\u4f46\u8de8\u6b65\u9aa4\u4fdd\u7559\u7684\u4fe1\u606f\u9700\u8981\u8fdb\u5165\u6709\u7ed3\u6784\u3001\u6709\u66f4\u65b0\u8bed\u4e49\u3001\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u3002<\/p>\n

\u4f5c\u8005\u62a5\u544a，\u5728 Gemini-3-Flash \u7684\u4ed3\u5e93 100 \u6b65\u5b9e\u9a8c\u4e2d，SKILL.state \u7684\u52a8\u4f5c\u51c6\u786e\u7387\u4e3a 0.94，\u7d2f\u8ba1 Token \u4e3a 65,408；\u540c\u65f6\u4fdd\u7559\u5386\u53f2\u548c\u72b6\u6001\u7684 Stateful \u57fa\u7ebf\u5206\u522b\u4e3a 0.91 \u4e0e 1,062,387\u3002\u8fd9\u4e2a\u7ed3\u679c\u652f\u6301\u8ba4\u771f\u7814\u7a76\u663e\u5f0f\u72b6\u6001，\u4f46\u4e0d\u80fd\u76f4\u63a5\u63a8\u5e7f\u6210\u201c\u6240\u6709 Agent \u90fd\u80fd\u8282\u7701 16 \u500d\u6210\u672c\u201d\u3002\u6536\u76ca\u4f9d\u8d56\u9886\u57df\u662f\u5426\u9002\u5408\u72b6\u6001\u5316\u3001\u72b6\u6001\u662f\u5426\u5145\u5206\u3001\u5b57\u6bb5\u662f\u5426\u771f\u6b63\u6709\u754c，\u4ee5\u53ca\u6a21\u578b\u80fd\u5426\u7a33\u5b9a\u751f\u6210\u6b63\u786e\u8865\u4e01\u3002<\/p>\n

\u672c\u6587\u5148\u89e3\u91ca\u6267\u884c\u673a\u5236\u548c\u590d\u6742\u5ea6\u6761\u4ef6，\u518d\u4fdd\u7559\u5b8c\u6574\u5b9e\u9a8c\u8868\u683c，\u6700\u540e\u8ba8\u8bba\u8bc1\u636e\u8fb9\u754c\u4e0e\u5de5\u7a0b\u843d\u5730\u3002\u8bba\u6587\u7ed3\u679c\u3001\u6280\u672f\u89e3\u91ca\u548c Rocky \u7684\u7814\u7a76\u5efa\u8bae\u4f1a\u660e\u786e\u533a\u5206\u3002<\/p>\n

\u4e00\u3001\u95ee\u9898\u80cc\u666f：\u5bf9\u8bdd\u8bb0\u5f55\u3001\u6458\u8981\u548c\u6267\u884c\u72b6\u6001，\u5404\u81ea\u627f\u62c5\u4ec0\u4e48\u804c\u8d23？<\/h3>\n

1.1 \u5386\u53f2\u80fd\u591f\u89e3\u91ca\u8fc7\u7a0b，\u5374\u672a\u5fc5\u9002\u5408\u4f5c\u4e3a\u6bcf\u4e00\u6b65\u7684\u5de5\u4f5c\u533a<\/h4>\n

ReAct \u5f0f\u6267\u884c\u5c06\u63a8\u7406\u3001\u52a8\u4f5c\u3001\u89c2\u5bdf\u4e32\u8d77\u6765，\u4f18\u70b9\u662f\u7b80\u5355\u3001\u4fe1\u606f\u4fdd\u7559\u5145\u5206，\u4e5f\u4fbf\u4e8e\u8ffd\u8e2a\u6a21\u578b\u5982\u4f55\u5230\u8fbe\u5f53\u524d\u7ed3\u679c\u3002\u5728\u63a2\u7d22\u6027\u4efb\u52a1\u91cc，\u67d0\u4e2a\u65e9\u671f\u7ec6\u8282\u53ef\u80fd\u5230\u540e\u9762\u624d\u663e\u51fa\u4ef7\u503c，\u4fdd\u5b58\u5386\u53f2\u5177\u6709\u5b9e\u9645\u610f\u4e49\u3002<\/p>\n

\u95ee\u9898\u51fa\u73b0\u5728\u957f\u7a0b\u3001\u91cd\u590d\u3001\u72b6\u6001\u660e\u786e\u7684\u6d41\u7a0b\u4e2d\u3002\u6a21\u578b\u6bcf\u4e00\u6b65\u90fd\u8981\u91cd\u65b0\u6d88\u89e3\u201c\u7269\u54c1\u5148\u5728 A\u3001\u540e\u6765\u79fb\u5230 B\u3001\u53c8\u5c1d\u8bd5\u79fb\u5230 C \u4f46\u5931\u8d25\u201d\u8fd9\u79cd\u65f6\u5e8f\u5173\u7cfb\u3002\u8f93\u5165\u8d8a\u957f，\u91cd\u590d\u8bfb\u53d6\u8d8a\u591a；\u5386\u53f2\u4e2d\u7684\u9519\u8bef\u63a8\u7406\u4e0e\u65e7\u4e8b\u5b9e\u4e5f\u8d8a\u5bb9\u6613\u540c\u5f53\u524d\u89c2\u5bdf\u7ade\u4e89\u3002\u5c06\u5168\u90e8\u8fc7\u7a0b\u89c6\u4e3a\u5f53\u524d\u51b3\u7b56\u4e0a\u4e0b\u6587，\u4f1a\u628a\u72b6\u6001\u7ba1\u7406\u6210\u672c\u8f6c\u4ea4\u7ed9\u6a21\u578b\u3002<\/p>\n

\u6458\u8981\u5c1d\u8bd5\u538b\u7f29\u8fc7\u7a0b，\u4f46\u4e00\u4efd\u81ea\u7136\u8bed\u8a00\u6458\u8981\u540c\u6837\u9700\u8981\u51b3\u5b9a\u54ea\u4e9b\u5173\u7cfb\u4e0d\u53ef\u5220\u9664\u3002\u5982\u679c\u6458\u8981\u9057\u6f0f\u4e86\u7269\u54c1\u7f16\u53f7\u4e0e\u8d27\u67b6\u7f16\u53f7\u7684\u7ed1\u5b9a，\u5373\u4f7f\u8bed\u8a00\u6d41\u7545，\u4e5f\u53ef\u80fd\u65e0\u6cd5\u652f\u6301\u4e0b\u4e00\u6761\u51fa\u5e93\u547d\u4ee4\u3002\u72b6\u6001\u8868\u793a\u8981\u89e3\u51b3\u7684\u5173\u952e\u95ee\u9898\u662f：\u54ea\u4e9b\u4fe1\u606f\u5fc5\u987b\u7cbe\u786e\u4fdd\u7559，\u54ea\u4e9b\u53ef\u4ee5\u8986\u76d6，\u54ea\u4e9b\u5df2\u7ecf\u6ca1\u6709\u6267\u884c\u4ef7\u503c\u3002<\/p>\n

1.2 \u8bba\u6587\u6bd4\u8f83\u7684\u662f\u5177\u4f53\u8fd0\u884c\u65f6\u914d\u7f6e<\/h4>\n\n\u8fd0\u884c\u65b9\u5f0f\u6bcf\u4e00\u6b65\u4fdd\u7559\u4ec0\u4e48\u4e3b\u8981\u53d6\u820d<\/tr>\n\n\n\n\n\n
Prompt \/ ReAct-style<\/td>\n\u7d2f\u79ef\u89c2\u5bdf\u3001\u63a8\u7406\u4e0e\u52a8\u4f5c<\/td>\n\u4fdd\u7559\u8fc7\u7a0b\u5145\u5206，\u4f46\u8f93\u5165\u968f\u5386\u53f2\u589e\u957f<\/td>\n<\/tr>\n
Memory \/ Summary-style<\/td>\n\u81ea\u7136\u8bed\u8a00\u6458\u8981\u4e0e\u6700\u8fd1 3 \u6b65\u5bf9\u8bdd<\/td>\n\u538b\u7f29\u90e8\u5206\u5386\u53f2，\u4ecd\u4f9d\u8d56\u6458\u8981\u7684\u9009\u62e9\u4e0e\u66f4\u65b0<\/td>\n<\/tr>\n
Stateful \/ LangGraph-style<\/td>\n\u7ed3\u6784\u5316\u72b6\u6001\u4e0e\u5b8c\u6574\u5bf9\u8bdd\u5386\u53f2<\/td>\n\u72b6\u6001\u6709\u7ed3\u6784，\u4f46\u5386\u53f2\u8f93\u5165\u4ecd\u7136\u589e\u957f<\/td>\n<\/tr>\n
SKILL.state<\/td>\n\u6280\u80fd\u89c4\u8303\u3001\u5f53\u524d\u7ed3\u6784\u5316\u72b6\u6001\u3001\u6700\u65b0\u89c2\u5bdf<\/td>\n\u51cf\u5c11\u5386\u53f2\u91cd\u653e，\u8981\u6c42\u72b6\u6001\u8db3\u591f\u652f\u6301\u540e\u7eed\u51b3\u7b56<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u8fd9\u91cc\u7684 Stateful \u662f\u4f5c\u8005\u5b9e\u73b0\u7684 LangGraph-style \u57fa\u7ebf，\u5176\u9644\u5f55\u63d0\u793a\u660e\u786e\u5305\u542b\u5b8c\u6574\u5386\u53f2\u3002\u5b83\u4e0d\u80fd\u4ee3\u8868\u6240\u6709\u4f7f\u7528 LangGraph \u7684\u7cfb\u7edf，\u4e5f\u4e0d\u80fd\u636e\u6b64\u8ba4\u5b9a\u67d0\u4e2a\u6846\u67b6\u5fc5\u7136\u8981\u6c42\u65e0\u9650\u8ffd\u52a0\u5bf9\u8bdd\u3002\u672c\u6587\u5c06\u5b9e\u9a8c\u7ed3\u8bba\u9650\u5b9a\u5728\u8fd9\u4e9b\u5177\u4f53\u914d\u7f6e\u4e0a\u3002<\/p>\n

SKILL.state \u5173\u6ce8\u6280\u80fd\u5df2\u7ecf\u9009\u5b9a\u4e4b\u540e\u5982\u4f55\u6267\u884c\u3002\u5b83\u6ca1\u6709\u5728\u672c\u6587\u89e3\u51b3\u5927\u89c4\u6a21\u6280\u80fd\u68c0\u7d22\u3001\u8de8\u4efb\u52a1\u6280\u80fd\u53d1\u73b0\u6216\u6a21\u578b\u53c2\u6570\u5b66\u4e60\u3002\u628a\u8fd9\u4e9b\u95ee\u9898\u533a\u5206\u5f00，\u624d\u80fd\u5224\u65ad\u8fd9\u9879\u8fd0\u884c\u65f6\u6539\u9020\u9002\u5408\u653e\u5728\u7cfb\u7edf\u7684\u54ea\u4e00\u5c42\u3002<\/p>\n

\u4e8c\u3001\u65b9\u6cd5\u5c55\u5f00：\u53ea\u8ba9\u672a\u6765\u9700\u8981\u7684\u4fe1\u606f\u7a7f\u8fc7\u6267\u884c\u8fb9\u754c<\/h3>\n

\"Figure1<\/p>\n

\u56fe 1｜SKILL.state \u7684\u5b8c\u6574\u67b6\u6784\u56fe\u3002\u5de6\u4fa7\u5c06\u5386\u53f2\u6301\u7eed\u52a0\u5165\u4e0a\u4e0b\u6587，\u53f3\u4fa7\u5c06\u672c\u6b65\u63a8\u7406\u6295\u5f71\u6210\u72b6\u6001\u66f4\u65b0，\u518d\u4ee5\u5f53\u524d\u72b6\u6001\u7ee7\u7eed\u6267\u884c\u3002\u4e0b\u65b9 <\/p>\n

t<\/p>\n

=<\/p>\n

500<\/p>\n

t=500<\/p>\n

<\/span><\/span>t<\/span><\/span>=<\/span><\/span><\/span><\/span>500<\/span><\/span><\/span><\/span><\/span> \u7684\u6761\u5f62\u56fe\u662f\u589e\u957f\u8d8b\u52bf\u793a\u610f，\u4e0d\u662f\u672c\u6587\u62a5\u544a\u4e86 500 \u6b65\u5b9e\u6d4b；\u4ed3\u5e93\u6269\u5c55\u5b9e\u9a8c\u6700\u957f\u4e3a 200 \u6b65\u3002\u56fe\u4e2d\u7684\u5e38\u6570\u8f93\u5165\u7ed3\u8bba\u4e5f\u9700\u8981\u72b6\u6001\u4e0e\u89c2\u5bdf\u5927\u5c0f\u6709\u754c\u7684\u6761\u4ef6\u3002<\/p>\n

2.1 \u8f93\u5165\u4e09\u5143\u7ec4：\u89c4\u8303\u3001\u72b6\u6001\u4e0e\u6700\u65b0\u89c2\u5bdf<\/h4>\n

\u8bba\u6587\u5728\u5f15\u8a00\u4e2d\u4ee5\u516c\u5f0f（1）\u6982\u62ec\u6267\u884c\u8f93\u5165：<\/p>\n

<\/p>\n

A<\/p>\n

t<\/p>\n

=<\/p>\n

(<\/p>\n

P<\/p>\n

,<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

,<\/p>\n

O<\/p>\n

t<\/p>\n

)<\/p>\n

.<\/p>\n

A_t=(P,\\\\Sigma_t,O_t). <\/p>\n

<\/span><\/span>A<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>(<\/span>P<\/span>,<\/span><\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>O<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5176\u4e2d，<\/p>\n

P<\/p>\n

P<\/p>\n

<\/span><\/span>P<\/span><\/span><\/span><\/span><\/span> \u662f\u4e0d\u53ef\u53d8\u7684\u8fc7\u7a0b\u89c4\u8303，\u5305\u542b\u4efb\u52a1\u89c4\u5219\u3001\u53ef\u7528\u52a8\u4f5c\u548c\u884c\u4e3a\u8981\u6c42；<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

\\\\Sigma_t<\/p>\n

<\/span><\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u662f\u5f53\u524d\u7ed3\u6784\u5316\u6267\u884c\u72b6\u6001；<\/p>\n

O<\/p>\n

t<\/p>\n

O_t<\/p>\n

<\/span><\/span>O<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u662f\u6700\u65b0\u73af\u5883\u89c2\u5bdf\u3002\u8fd9\u91cc\u7684\u5927\u5199 <\/p>\n

A<\/p>\n

t<\/p>\n

A_t<\/p>\n

<\/span><\/span>A<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u8f93\u5165\u4e0a\u4e0b\u6587，\u540e\u6587\u7684\u5c0f\u5199 <\/p>\n

a<\/p>\n

t<\/p>\n

a_t<\/p>\n

<\/span><\/span>a<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u624d\u8868\u793a\u6267\u884c\u52a8\u4f5c，\u4e24\u8005\u4e0d\u80fd\u6df7\u7528\u3002<\/p>\n

\u8bba\u6587\u5728\u65b9\u6cd5\u90e8\u5206\u4ee5\u516c\u5f0f（2）\u518d\u6b21\u7ed9\u51fa\u540c\u4e00\u5b9a\u4e49：<\/p>\n

<\/p>\n

A<\/p>\n

t<\/p>\n

=<\/p>\n

(<\/p>\n

P<\/p>\n

,<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

,<\/p>\n

O<\/p>\n

t<\/p>\n

)<\/p>\n

.<\/p>\n

A_t=(P,\\\\Sigma_t,O_t). <\/p>\n

<\/span><\/span>A<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>(<\/span>P<\/span>,<\/span><\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>O<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u8fd9\u4e24\u4e2a\u7f16\u53f7\u516c\u5f0f\u5185\u5bb9\u76f8\u540c，\u5206\u522b\u627f\u62c5\u5f15\u5165\u601d\u8def\u548c\u6b63\u5f0f\u5b9a\u4e49\u7684\u4f5c\u7528\u3002\u771f\u6b63\u6539\u53d8\u6267\u884c\u65b9\u5f0f\u7684\u7ea6\u675f\u662f：\u540e\u7eed Prompt \u4e0d\u518d\u5305\u542b\u5148\u524d\u89c2\u5bdf\u3001\u52a8\u4f5c\u548c\u63a8\u7406\u7684\u5b8c\u6574\u6587\u672c；\u6b64\u524d\u6709\u4ef7\u503c\u7684\u4fe1\u606f\u5fc5\u987b\u5df2\u7ecf\u8fdb\u5165\u72b6\u6001\u3002<\/p>\n

\u8fd8\u8981\u533a\u5206 Agent \u4fdd\u5b58\u7684\u6267\u884c\u72b6\u6001 \u548c \u73af\u5883\u771f\u5b9e\u72b6\u6001\u3002\u524d\u8005\u7531\u6a21\u578b\u63d0\u51fa\u66f4\u65b0，\u540e\u8005\u5b58\u5728\u4e8e\u4ed3\u5e93\u3001\u6570\u636e\u5e93\u6216\u5de5\u5177\u7cfb\u7edf\u4e2d\u3002\u4e00\u4e2a JSON \u5bf9\u8c61\u7ed3\u6784\u6b63\u786e，\u4e0d\u4ee3\u8868\u5176\u4e2d\u8bb0\u5f55\u7684\u4e8b\u5b9e\u4e00\u5b9a\u771f\u5b9e；\u663e\u5f0f\u5316\u4f7f\u5dee\u5f02\u66f4\u5bb9\u6613\u68c0\u67e5，\u5e76\u4e0d\u4f1a\u8ba9\u5dee\u5f02\u81ea\u52a8\u6d88\u5931\u3002<\/p>\n

2.2 Schema \u6309\u9886\u57df\u8bbe\u8ba1，\u800c\u4e0d\u662f\u4e3a\u6bcf\u9053\u9898\u5b9a\u5236<\/h4>\n

\u4f5c\u8005\u4f7f\u7528\u9886\u57df\u7ea7\u72b6\u6001 Schema\u3002\u4f8b\u5982 InterCode CTF \u7684 100 \u4e2a\u4efb\u52a1\u590d\u7528\u4e94\u4e2a\u5b57\u6bb5：discovered_flags\u3001tested_hypotheses\u3001active_files\u3001working_dir \u548c cmd_summary\u3002\u8fd9\u4e9b\u5b57\u6bb5\u5206\u522b\u4fdd\u5b58\u5df2\u53d1\u73b0\u7684\u7ed3\u679c\u3001\u5df2\u6d4b\u8bd5\u5047\u8bbe\u3001\u6d3b\u8dc3\u6587\u4ef6\u3001\u5de5\u4f5c\u76ee\u5f55\u4ee5\u53ca\u547d\u4ee4\u76f8\u5173\u6458\u8981\u3002<\/p>\n

\u8fd9\u79cd\u8bbe\u8ba1\u5177\u6709\u5408\u7406\u7684\u5de5\u7a0b\u76f4\u89c9：\u4e0b\u4e00\u6b65\u9700\u8981\u77e5\u9053\u67d0\u4e2a\u5047\u8bbe\u5df2\u7ecf\u5931\u8d25，\u672a\u5fc5\u9700\u8981\u91cd\u8bfb\u63d0\u51fa\u5047\u8bbe\u65f6\u7684\u5168\u90e8\u6587\u5b57\u3002\u4f46 tested_hypotheses \u662f\u4e00\u4e2a\u56fa\u5b9a\u5b57\u6bb5，\u5e76\u4e0d\u610f\u5473\u7740\u5b83\u7684\u5217\u8868\u957f\u5ea6\u56fa\u5b9a；cmd_summary \u662f\u4e00\u4e2a\u56fa\u5b9a\u5b57\u6bb5，\u4e5f\u4e0d\u610f\u5473\u7740\u5176\u4e2d\u7684\u5b57\u7b26\u4e32\u6c38\u8fdc\u4e0d\u589e\u957f\u3002\u5b57\u6bb5\u6570\u91cf\u6709\u754c\u4e0e\u5e8f\u5217\u5316\u540e\u7684\u72b6\u6001\u5927\u5c0f\u6709\u754c，\u662f\u4e24\u9879\u4e0d\u540c\u7ea6\u675f\u3002<\/p>\n

\u4e00\u4e2a\u53ef\u6267\u884c\u7684 Schema \u8fd8\u9700\u8981\u7ea6\u5b9a\u8eab\u4efd\u3001\u5355\u4f4d\u3001\u751f\u547d\u5468\u671f\u548c\u672a\u77e5\u503c\u3002\u4f8b\u5982，\u201c\u5c1a\u672a\u68c0\u67e5\u8d27\u67b6\u201d\u548c\u201c\u8d27\u67b6\u5df2\u7ecf\u786e\u8ba4\u4e3a\u7a7a\u201d\u80fd\u5426\u7528\u540c\u4e00\u4e2a null \u8868\u793a？\u67d0\u4e2a\u5047\u8bbe\u662f\u672a\u6d4b\u8bd5\u3001\u5df2\u5426\u5b9a，\u8fd8\u662f\u56e0\u4e3a\u5de5\u5177\u5931\u8d25\u800c\u65e0\u6cd5\u5224\u65ad？\u8fd9\u4e9b\u533a\u5206\u51b3\u5b9a\u538b\u7f29\u540e\u7684\u72b6\u6001\u662f\u5426\u4ecd\u80fd\u652f\u6301\u53ef\u9760\u52a8\u4f5c\u3002<\/p>\n

2.3 \u8f93\u51fa\u4e09\u5143\u7ec4：\u4e34\u65f6\u63a8\u7406\u3001\u72b6\u6001\u8865\u4e01\u4e0e\u52a8\u4f5c<\/h4>\n

\u8bba\u6587\u516c\u5f0f（3）：<\/p>\n

<\/p>\n

(<\/p>\n

R<\/p>\n

t<\/p>\n

,<\/p>\n

\u0394<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

,<\/p>\n

a<\/p>\n

t<\/p>\n

)<\/p>\n

.<\/p>\n

(R_t,\\\\Delta\\\\Sigma_t,a_t). <\/p>\n

<\/span><\/span>(<\/span>R<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>\u0394<\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>a<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

<\/p>\n

R<\/p>\n

t<\/p>\n

R_t<\/p>\n

<\/span><\/span>R<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u662f\u672c\u6b65\u4e2d\u95f4\u63a8\u7406，<\/p>\n

\u0394<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

\\\\Delta\\\\Sigma_t<\/p>\n

<\/span><\/span>\u0394<\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u662f\u63cf\u8ff0\u952e\u4fee\u6539\u4e0e\u5220\u9664\u7684 JSON \u5b57\u5178，<\/p>\n

a<\/p>\n

t<\/p>\n

a_t<\/p>\n

<\/span><\/span>a<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u662f\u5c06\u8981\u6267\u884c\u7684\u547d\u4ee4\u3002SKILL.state \u5e76\u672a\u53d6\u6d88\u5355\u6b65\u591a\u6b65\u63a8\u7406；\u5b83\u6539\u53d8\u7684\u662f\u63a8\u7406\u751f\u6210\u4e4b\u540e\u7684\u4fe1\u606f\u4fdd\u7559\u65b9\u5f0f\u3002\u72b6\u6001\u66f4\u65b0\u901a\u8fc7\u9a8c\u8bc1\u540e，\u672c\u6b65\u63a8\u7406\u4e0d\u8fdb\u5165\u540e\u7eed Prompt\u3002<\/p>\n

\u8bba\u6587\u516c\u5f0f（4）：<\/p>\n

<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

+<\/p>\n

1<\/p>\n

=<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

\u2295<\/p>\n

\u0394<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

.<\/p>\n

\\\\Sigma_{t+1}=\\\\Sigma_t\\\\oplus\\\\Delta\\\\Sigma_t. <\/p>\n

<\/span><\/span>\u03a3<\/span><\/span>t<\/span>+<\/span>1<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2295<\/span><\/span><\/span><\/span>\u0394<\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

<\/p>\n

\u2295<\/p>\n

\\\\oplus<\/p>\n

<\/span><\/span>\u2295<\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u5e26\u6709 null \u5220\u9664\u8bed\u4e49\u7684\u5b57\u5178\u5408\u5e76\u64cd\u4f5c\u3002\u8fd9\u4e2a\u5b9a\u4e49\u5f3a\u8c03\u8865\u4e01\u4f5c\u7528\u4e8e\u5df2\u6709\u72b6\u6001，\u800c\u975e\u6bcf\u4e00\u6b65\u91cd\u5199\u6574\u4e2a\u72b6\u6001\u3002\u56fe\u4e2d\u51fa\u73b0\u7684 \u201cJSON Patch\u201d \u5e94\u6309\u8bba\u6587\u7684\u5b57\u5178\u8865\u4e01\u8bed\u4e49\u7406\u89e3，\u4e0d\u80fd\u76f4\u63a5\u5f53\u6210\u5df2\u7ecf\u58f0\u660e\u7b26\u5408\u67d0\u4e2a JSON Patch \u6807\u51c6\u534f\u8bae\u3002<\/p>\n

\u9644\u5f55\u7684\u4ed3\u5e93\u793a\u4f8b\u4e2d，\u6700\u65b0\u8bf7\u6c42\u662f\u51fa\u5e93 item_12，\u72b6\u6001\u8bb0\u5f55\u8be5\u7269\u54c1\u4f4d\u4e8e shelf_42\u3002\u6a21\u578b\u8f93\u51fa：<\/p>\n

{<\/span>
\n "state_patch"<\/span>:<\/span> {<\/span>
\n "inventory"<\/span>:<\/span> {<\/span>
\n "shelf_42"<\/span>:<\/span> null<\/span>
\n }<\/span>
\n }<\/span>,<\/span>
\n "action"<\/span>:<\/span> "Ship item_12 shelf_42"<\/span>
\n}<\/span><\/p>\n

\u8bfb\u8005\u5e94\u6ce8\u610f：\u5220\u9664\u7684\u542b\u4e49\u3001\u5d4c\u5957\u5b57\u5178\u5982\u4f55\u5408\u5e76\u3001\u7f3a\u5931\u952e\u662f\u5426\u4fdd\u7559，\u9700\u8981\u7531\u8fd0\u884c\u65f6\u660e\u786e\u5b9a\u4e49\u3002\u53ea\u77e5\u9053\u9876\u5c42\u6267\u884c\u4e00\u6b21 dict.update，\u4e0d\u8db3\u4ee5\u63a8\u51fa\u6df1\u5c42\u952e\u4e5f\u4f1a\u6309\u9884\u671f\u5408\u5e76\u6216\u5220\u9664\u3002\u8bba\u6587\u63d0\u4f9b\u4e86\u4e0a\u8ff0\u793a\u4f8b，\u4f46\u6ca1\u6709\u5728\u6b63\u6587\u7ed9\u51fa\u8db3\u4ee5\u8986\u76d6\u6240\u6709\u5d4c\u5957\u60c5\u51b5\u7684\u5408\u5e76\u5b9e\u73b0\u3002<\/p>\n

2.4 \u7b97\u6cd5 1：\u8fd0\u884c\u65f6\u5982\u4f55\u63a5\u7ba1\u72b6\u6001\u66f4\u65b0<\/h4>\n

\u4ee5\u4e0b\u4e2d\u6587\u4f2a\u4ee3\u7801\u4fdd\u7559\u8bba\u6587 Algorithm 1 \u7684\u6267\u884c\u987a\u5e8f；\u65e0\u6548\u8865\u4e01\u7684\u56de\u6eda\u91cd\u8bd5\u6765\u81ea\u8bba\u6587\u9650\u5236\u90e8\u5206\u7684\u8865\u5145\u8bf4\u660e\u3002<\/p>\n

\u8f93\u5165：\u4e0d\u53ef\u53d8\u6280\u80fd\u89c4\u8303 P，\u521d\u59cb\u72b6\u6001 Sigma_0
\n\u5bf9\u6267\u884c\u6b65\u9aa4 t = 0, …, T：
\n \u63a5\u6536\u6700\u65b0\u89c2\u5bdf O_t
\n \u6784\u5efa\u8f93\u5165 (P, Sigma_t, O_t)
\n \u8c03\u7528\u6a21\u578b，\u751f\u6210 (\u672c\u6b65\u63a8\u7406 R_t, \u72b6\u6001\u8865\u4e01 DeltaSigma_t, \u52a8\u4f5c a_t)
\n \u6821\u9a8c\u72b6\u6001\u8865\u4e01
\n \u82e5\u8865\u4e01\u65e0\u6548：\u4e0d\u63d0\u4ea4\u8be5\u8865\u4e01，\u8fdb\u5165\u56de\u6eda\u4e0e\u91cd\u8bd5\u5904\u7406
\n \u82e5\u8865\u4e01\u6709\u6548：Sigma_(t+1) = Merge(Sigma_t, DeltaSigma_t)
\n \u6267\u884c\u52a8\u4f5c a_t
\n \u672c\u6b65\u63a8\u7406\u4e0d\u8fdb\u5165\u4e0b\u4e00\u6b21\u8f93\u5165<\/p>\n

\u8bba\u6587\u4f2a\u4ee3\u7801\u91c7\u7528 <\/p>\n

t<\/p>\n

=<\/p>\n

0<\/p>\n

,<\/p>\n

\u2026<\/p>\n

,<\/p>\n

T<\/p>\n

t=0,\\\\ldots,T<\/p>\n

<\/span><\/span>t<\/span><\/span>=<\/span><\/span><\/span><\/span>0<\/span>,<\/span><\/span>\u2026<\/span><\/span>,<\/span><\/span>T<\/span><\/span><\/span><\/span><\/span> \u7684\u7d22\u5f15，\u5b9e\u9a8c\u4e2d\u7684 <\/p>\n

T<\/p>\n

T<\/p>\n

<\/span><\/span>T<\/span><\/span><\/span><\/span><\/span> \u5219\u8868\u793a\u6267\u884c\u957f\u5ea6\u3002\u5de5\u7a0b\u5b9e\u73b0\u5e94\u7edf\u4e00\u8f6e\u6570\u7ea6\u5b9a，\u907f\u514d\u628a\u7d22\u5f15\u4e0a\u754c\u4e0e\u5b9e\u9645\u8c03\u7528\u6b21\u6570\u6df7\u7528\u3002<\/p>\n

\u66f4\u503c\u5f97\u5173\u6ce8\u7684\u662f\u63d0\u4ea4\u987a\u5e8f：\u8bba\u6587\u7b97\u6cd5\u5148\u66f4\u65b0\u5185\u90e8\u72b6\u6001，\u518d\u6267\u884c\u52a8\u4f5c\u3002\u5bf9\u4e8e\u4f1a\u5931\u8d25\u6216\u53ea\u90e8\u5206\u6210\u529f\u7684\u5916\u90e8\u5de5\u5177，\u8fd9\u4f1a\u51fa\u73b0\u72b6\u6001\u4e0e\u73af\u5883\u4e0d\u4e00\u81f4\u7684\u7a97\u53e3\u3002\u4f8b\u5982\u5185\u90e8\u5df2\u5220\u9664\u5e93\u5b58\u8bb0\u5f55，\u51fa\u5e93\u547d\u4ee4\u5374\u5931\u8d25\u4e86\u3002\u8bba\u6587\u8bf4\u660e\u4e86\u65e0\u6548\u8865\u4e01\u7684\u56de\u6eda\u91cd\u8bd5，\u5374\u6ca1\u6709\u5728\u8be5\u7b97\u6cd5\u4e2d\u5b8c\u6574\u5c55\u5f00\u5de5\u5177\u5931\u8d25\u540e\u7684\u4e8b\u52a1\u534f\u8bae\u3002<\/p>\n

Rocky\u7684\u5de5\u7a0b\u5efa\u8bae\u662f，\u5c06\u62df\u8bae\u72b6\u6001\u4e0e\u5df2\u786e\u8ba4\u4e8b\u5b9e\u5206\u5f00，\u5e76\u4f7f\u5de5\u5177\u8fd4\u56de\u80fd\u591f\u9a71\u52a8\u786e\u8ba4\u3001\u8865\u507f\u6216\u56de\u6eda\u3002 \u5177\u4f53\u9009\u62e9\u5148\u6267\u884c\u540e\u63d0\u4ea4，\u8fd8\u662f\u4f7f\u7528\u5f85\u786e\u8ba4\u72b6\u6001\u548c\u4e8b\u52a1\u65e5\u5fd7，\u8981\u7ed3\u5408\u5de5\u5177\u7684\u5e42\u7b49\u6027\u4e0e\u4e1a\u52a1\u7ea6\u675f\u3002\u8fd9\u91cc\u662f\u5728\u8865\u5145\u751f\u4ea7\u7cfb\u7edf\u8bbe\u8ba1，\u5e76\u975e\u8bba\u6587\u5df2\u7ecf\u5b8c\u6210\u4e86\u5206\u5e03\u5f0f\u4e8b\u52a1\u9a8c\u8bc1\u3002<\/p>\n

\u4e09\u3001\u590d\u6742\u5ea6\u5206\u6790：\u4ece\u5e73\u65b9\u589e\u957f\u5230\u7ebf\u6027\u589e\u957f，\u9700\u8981\u54ea\u4e9b\u524d\u63d0？<\/h3>\n

3.1 \u5386\u53f2\u91cd\u653e\u4e3a\u4ec0\u4e48\u4ea7\u751f\u5e73\u65b9\u7ea7\u7d2f\u8ba1\u8f93\u5165<\/h4>\n

\u8bba\u6587\u516c\u5f0f（5）：<\/p>\n

<\/p>\n

\u2211<\/p>\n

t<\/p>\n

=<\/p>\n

1<\/p>\n

T<\/p>\n

\u2223<\/p>\n

C<\/p>\n

t<\/p>\n

\u2223<\/p>\n

=<\/p>\n

O<\/p>\n

(<\/p>\n

T<\/p>\n

2<\/p>\n

)<\/p>\n

.<\/p>\n

\\\\sum_{t=1}^{T}|C_t|=\\\\mathcal O(T^2). <\/p>\n

<\/span><\/span><\/span>t<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>C<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span>=<\/span><\/span><\/span><\/span>O<\/span>(<\/span>T<\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

<\/p>\n

C<\/p>\n

t<\/p>\n

C_t<\/p>\n

<\/span><\/span>C<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u5386\u53f2\u7d2f\u79ef\u578b\u8fd0\u884c\u65f6\u7b2c <\/p>\n

t<\/p>\n

t<\/p>\n

<\/span><\/span>t<\/span><\/span><\/span><\/span><\/span> \u6b65\u7684\u4e0a\u4e0b\u6587\u3002\u82e5\u6bcf\u6b65\u8ffd\u52a0\u7684\u4fe1\u606f\u91cf\u8fd1\u4f3c\u56fa\u5b9a，\u4e14\u6ca1\u6709\u622a\u65ad\u6216\u786c\u4e0a\u9650，\u5355\u6b65\u8f93\u5165\u957f\u5ea6\u968f <\/p>\n

t<\/p>\n

t<\/p>\n

<\/span><\/span>t<\/span><\/span><\/span><\/span><\/span> \u7ebf\u6027\u589e\u957f\u3002\u7d2f\u8ba1\u8bfb\u53d6\u7684\u5386\u53f2\u76f8\u5f53\u4e8e <\/p>\n

1<\/p>\n

+<\/p>\n

2<\/p>\n

+<\/p>\n

\u22ef<\/p>\n

+<\/p>\n

T<\/p>\n

1+2+\\\\cdots+T<\/p>\n

<\/span><\/span>1<\/span><\/span>+<\/span><\/span><\/span><\/span>2<\/span><\/span>+<\/span><\/span><\/span><\/span>\u22ef<\/span><\/span>+<\/span><\/span><\/span><\/span>T<\/span><\/span><\/span><\/span><\/span>，\u4e8e\u662f\u5f97\u5230\u5e73\u65b9\u7ea7\u8f93\u5165\u89c4\u6a21\u3002<\/p>\n

\u8fd9\u4e2a\u7ed3\u8bba\u63cf\u8ff0\u7684\u662f\u4e0a\u8ff0\u8ffd\u52a0\u6a21\u5f0f\u3002\u5177\u6709\u56fa\u5b9a\u7a97\u53e3\u3001\u4e25\u683c\u6458\u8981\u4e0a\u9650\u6216\u5176\u4ed6\u6709\u754c\u4e0a\u4e0b\u6587\u7b56\u7565\u7684\u8fd0\u884c\u65f6，\u4e0d\u5fc5\u6ee1\u8db3\u540c\u6837\u7684\u589e\u957f\u5f62\u5f0f\u3002\u8bba\u6587\u540e\u6765\u52a0\u5165\u9884\u7b97\u5339\u914d\u5b9e\u9a8c，\u6b63\u662f\u4e3a\u4e86\u533a\u5206\u201c\u8f93\u5165\u53d8\u77ed\u201d\u548c\u201c\u4fdd\u7559\u4e86\u6b63\u786e\u72b6\u6001\u201d\u4e24\u79cd\u4f5c\u7528\u3002<\/p>\n

3.2 \u72b6\u6001\u8f93\u5165\u4e3a\u4ec0\u4e48\u53ef\u80fd\u4fdd\u6301\u6709\u754c<\/h4>\n

\u8bba\u6587\u516c\u5f0f（6）：<\/p>\n

<\/p>\n

\u2223<\/p>\n

P<\/p>\n

t<\/p>\n

\u2223<\/p>\n

=<\/p>\n

O<\/p>\n

(<\/p>\n

\u2223<\/p>\n

P<\/p>\n

\u2223<\/p>\n

+<\/p>\n

\u2223<\/p>\n

\u03a3<\/p>\n

\u2223<\/p>\n

+<\/p>\n

\u2223<\/p>\n

O<\/p>\n

\u2223<\/p>\n

)<\/p>\n

.<\/p>\n

|P_t|=\\\\mathcal O(|P|+|\\\\Sigma|+|O|). <\/p>\n

<\/span><\/span>\u2223<\/span>P<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span>=<\/span><\/span><\/span><\/span>O<\/span>(<\/span>\u2223<\/span>P<\/span>\u2223<\/span><\/span>+<\/span><\/span><\/span><\/span>\u2223\u03a3\u2223<\/span><\/span>+<\/span><\/span><\/span><\/span>\u2223<\/span>O<\/span>\u2223<\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u8fd9\u91cc\u7684 <\/p>\n

P<\/p>\n

t<\/p>\n

P_t<\/p>\n

<\/span><\/span>P<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u7b2c <\/p>\n

t<\/p>\n

t<\/p>\n

<\/span><\/span>t<\/span><\/span><\/span><\/span><\/span> \u6b65\u5b8c\u6574 Prompt，\u9700\u4e0e\u4e0d\u53ef\u53d8\u89c4\u8303 <\/p>\n

P<\/p>\n

P<\/p>\n

<\/span><\/span>P<\/span><\/span><\/span><\/span><\/span> \u533a\u5206\u3002\u8bba\u6587\u516c\u5f0f（7）\u8fdb\u4e00\u6b65\u7ed9\u51fa：<\/p>\n

<\/p>\n

\u2211<\/p>\n

t<\/p>\n

=<\/p>\n

1<\/p>\n

T<\/p>\n

\u2223<\/p>\n

P<\/p>\n

t<\/p>\n

\u2223<\/p>\n

=<\/p>\n

O<\/p>\n

(<\/p>\n

T<\/p>\n

)<\/p>\n

.<\/p>\n

\\\\sum_{t=1}^{T}|P_t|=\\\\mathcal O(T). <\/p>\n

<\/span><\/span><\/span>t<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>P<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span>=<\/span><\/span><\/span><\/span>O<\/span>(<\/span>T<\/span>)<\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u63a8\u5bfc\u7684\u5173\u952e\u5728\u4e8e：\u89c4\u8303\u5927\u5c0f\u3001\u6267\u884c\u72b6\u6001\u5927\u5c0f\u548c\u6700\u65b0\u89c2\u5bdf\u5927\u5c0f，\u5fc5\u987b\u53d7\u5230\u4e0d\u968f\u6267\u884c\u957f\u5ea6\u589e\u957f\u7684\u4e0a\u754c\u7ea6\u675f\u3002\u4e3a\u4e86\u628a\u6761\u4ef6\u5199\u6e05\u695a，\u53ef\u4ee5\u5c06\u8bba\u6587\u5206\u6790\u5c55\u5f00\u4e3a\u4e0b\u9762\u8fd9\u4e2a\u89e3\u91ca\u5f0f：<\/p>\n

<\/p>\n

\u2211<\/p>\n

t<\/p>\n

=<\/p>\n

1<\/p>\n

T<\/p>\n

\u2223<\/p>\n

P<\/p>\n

t<\/p>\n

\u2223<\/p>\n

=<\/p>\n

O
\n \u2009\u2063<\/p>\n

(<\/p>\n

T<\/p>\n

\u2223<\/p>\n

P<\/p>\n

\u2223<\/p>\n

+<\/p>\n

\u2211<\/p>\n

t<\/p>\n

=<\/p>\n

1<\/p>\n

T<\/p>\n

\u2223<\/p>\n

\u03a3<\/p>\n

t<\/p>\n

\u2223<\/p>\n

+<\/p>\n

\u2211<\/p>\n

t<\/p>\n

=<\/p>\n

1<\/p>\n

T<\/p>\n

\u2223<\/p>\n

O<\/p>\n

t<\/p>\n

\u2223<\/p>\n

)<\/p>\n

.<\/p>\n

\\\\sum_{t=1}^{T}|P_t|=\\\\mathcal O\\\\!\\\\left(T|P|+\\\\sum_{t=1}^{T}|\\\\Sigma_t|+\\\\sum_{t=1}^{T}|O_t|\\\\right). <\/p>\n

<\/span><\/span><\/span>t<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>P<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span>=<\/span><\/span><\/span><\/span>O<\/span><\/span><\/span>(<\/span><\/span>T<\/span>\u2223<\/span>P<\/span>\u2223<\/span><\/span>+<\/span><\/span><\/span>t<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>\u03a3<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span>+<\/span><\/span><\/span>t<\/span>=<\/span>1<\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span><\/span><\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>O<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span>)<\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5982\u679c\u72b6\u6001\u4e0d\u65ad\u8ffd\u52a0\u65b0\u5b9e\u4f53\u3001\u5df2\u6d4b\u8bd5\u5047\u8bbe\u548c\u547d\u4ee4\u6458\u8981，\u6216\u8005\u6bcf\u6b65\u5de5\u5177\u8fd4\u56de\u8d8a\u6765\u8d8a\u957f，\u90a3\u4e48\u4ec5\u4ec5\u5220\u9664\u5bf9\u8bdd\u5386\u53f2\u4e0d\u8db3\u4ee5\u4fdd\u8bc1 <\/p>\n

O<\/p>\n

(<\/p>\n

T<\/p>\n

)<\/p>\n

\\\\mathcal O(T)<\/p>\n

<\/span><\/span>O<\/span>(<\/span>T<\/span>)<\/span><\/span><\/span><\/span><\/span>\u3002\u9886\u57df Schema \u7ed9\u51fa\u4e86\u7ed3\u6784\u7ea6\u675f，\u8fd8\u9700\u8981\u5bb9\u91cf\u3001\u5f52\u6863\u3001\u9009\u62e9\u6027\u8bfb\u53d6\u6216\u89c2\u6d4b\u538b\u7f29\u6765\u7ea6\u675f\u5b9e\u9645\u5927\u5c0f\u3002<\/p>\n

3.3 \u8f93\u5165\u957f\u5ea6\u3001\u7d2f\u8ba1 Token\u3001\u63a8\u7406\u5ef6\u8fdf\u4e0e\u8d39\u7528\u4e0d\u80fd\u753b\u7b49\u53f7<\/h4>\n

\u516c\u5f0f（5）\u2014（7）\u76f4\u63a5\u5206\u6790\u7d2f\u8ba1 Prompt \u89c4\u6a21\u3002\u5c06\u5b83\u63a8\u5e7f\u5230\u5168\u90e8 Token \u6d88\u8017，\u8fd8\u9700\u8981\u7ea6\u675f\u6bcf\u6b65\u8f93\u51fa\u3001\u63a8\u7406\u957f\u5ea6\u4ee5\u53ca\u56de\u6eda\u91cd\u8bd5\u6b21\u6570\u3002\u4e00\u4e2a\u5b57\u6bb5\u9519\u8bef\u5bfc\u81f4\u6a21\u578b\u53cd\u590d\u91cd\u8bd5，\u5373\u4f7f\u6bcf\u6b21 Prompt \u56fa\u5b9a，\u5b9e\u9645\u6267\u884c\u6210\u672c\u4ecd\u53ef\u80fd\u5347\u9ad8\u3002<\/p>\n

\u8ba1\u8d39\u8fd8\u53d7\u5230\u8f93\u5165\u4e0e\u8f93\u51fa\u5355\u4ef7\u3001\u7f13\u5b58\u547d\u4e2d\u548c\u4f9b\u5e94\u5546\u8ba1\u8d39\u65b9\u5f0f\u5f71\u54cd；\u5ef6\u8fdf\u53d7\u5230\u6a21\u578b\u8ba1\u7b97\u3001\u5de5\u5177\u7b49\u5f85\u3001\u5e76\u53d1\u548c\u91cd\u8bd5\u5f71\u54cd\u3002\u672c\u6587\u5f15\u7528\u7684\u201c16.2 \u500d\u201d\u662f\u7279\u5b9a\u8868\u683c\u7684\u7d2f\u8ba1 Token \u6bd4\u503c，\u4e0d\u662f\u5ef6\u8fdf\u52a0\u901f\u6bd4，\u4e5f\u4e0d\u662f\u5b9e\u6d4b\u8d39\u7528\u6536\u76ca\u3002\u8bba\u6587\u6ca1\u6709\u7ed9\u51fa\u8db3\u4ee5\u66ff\u4ee3\u8fd9\u4e9b\u72ec\u7acb\u6d4b\u91cf\u7684\u7aef\u5230\u7aef\u6210\u672c\u62a5\u544a\u3002<\/p>\n

Rocky\u8ba4\u4e3a，\u663e\u5f0f\u72b6\u6001\u964d\u4f4e\u7684\u662f\u6bcf\u4e00\u6b65\u91cd\u5efa\u5386\u53f2\u7684\u8d1f\u62c5\u3002\u5b83\u7684\u957f\u671f\u4ef7\u503c，\u9700\u8981\u540c\u65f6\u901a\u8fc7\u6b63\u786e\u7387\u3001\u6267\u884c\u6210\u672c\u548c\u72b6\u6001\u7ef4\u62a4\u6210\u672c\u6765\u68c0\u9a8c\u3002<\/p>\n

\u56db\u3001\u8bc4\u6d4b\u8bbe\u7f6e：\u5148\u5f04\u6e05\u4efb\u52a1\u3001\u6307\u6807\u4e0e\u57fa\u7ebf\u7684\u8fb9\u754c<\/h3>\n

4.1 \u4e24\u7c7b\u5408\u6210\u73af\u5883\u4e0e\u4e24\u7c7b\u516c\u5f00\u57fa\u51c6<\/h4>\n

SkillExecBench \u5305\u542b\u4ed3\u5e93\u7ba1\u7406\u4e0e\u8f6f\u4ef6\u4ed3\u5e93\u4e24\u4e2a\u53d7\u63a7\u73af\u5883\u3002\u4ed3\u5e93\u73af\u5883\u7ef4\u62a4 500 \u4e2a\u8d27\u67b6，\u901a\u8fc7 Store\u3001Ship\u3001Move \u548c Wait \u5904\u7406\u7269\u54c1\u4e0e\u8ba2\u5355；\u8f6f\u4ef6\u73af\u5883\u6a21\u62df\u5206\u652f\u3001\u63d0\u4ea4\u3001PR \u4e0e CI \u72b6\u6001，\u4ee5\u4f9d\u8d56\u5173\u7cfb\u53d8\u5316\u8003\u5bdf\u6267\u884c\u80fd\u529b\u3002\u8fd9\u4e9b\u4efb\u52a1\u6709\u786e\u5b9a\u6027\u7684\u73af\u5883\u8f6c\u79fb，\u9002\u5408\u8bca\u65ad\u72b6\u6001\u8ffd\u8e2a，\u5374\u4e0d\u662f\u5b8c\u6574\u5de5\u4e1a\u4ed3\u50a8\u6216\u771f\u5b9e\u8f6f\u4ef6\u5f00\u53d1\u7cfb\u7edf\u3002<\/p>\n

InterCode CTF \u5305\u542b 100 \u4e2a\u7ec8\u7aef\u6311\u6218，Agent \u5728\u5bb9\u5668\u4e2d\u901a\u8fc7\u547d\u4ee4\u8bd5\u63a2\u5047\u8bbe\u5e76\u5bfb\u627e flag；\u8bba\u6587\u6309\u7cbe\u786e\u7ed3\u679c\u5339\u914d\u8861\u91cf pass@1\u3002Sierra <\/p>\n

\u03c4<\/p>\n

\\\\tau<\/p>\n

<\/span><\/span>\u03c4<\/span><\/span><\/span><\/span><\/span>-Bench \u5305\u542b Retail \u4e0e Airline \u5ba2\u670d\u6d41\u7a0b，\u7531\u6a21\u62df\u7528\u6237\u3001\u5de5\u5177\u8c03\u7528\u548c\u6570\u636e\u5e93\u72b6\u6001\u53d8\u5316\u7ec4\u6210\u3002\u5176\u6210\u7ee9\u4f9d\u8d56\u7a0b\u5e8f\u5316\u8bc4\u4f30\u5668，\u4e0d\u80fd\u76f4\u63a5\u7b49\u540c\u4e8e\u771f\u5b9e\u7528\u6237\u6ee1\u610f\u5ea6\u6216\u5bf9\u6240\u6709\u4e1a\u52a1\u89c4\u5219\u7684\u5b8c\u6574\u5f62\u5f0f\u5316\u8bc1\u660e\u3002<\/p>\n

\u8f6f\u4ef6\u4ed3\u5e93\u52a8\u4f5c\u5b9a\u4e49\u5b58\u5728\u4e00\u5904\u9700\u8981\u590d\u73b0\u8005\u5904\u7406\u7684\u5dee\u5f02：\u6b63\u6587\u5217\u4e3e CherryPick\u3001Merge\u3001RunTests\u3001CreateRelease \u548c Rollback，\u9644\u5f55\u5217\u4e3e Commit\u3001CreatePR\u3001Merge\u3001FixCI \u548c Wait\u3002\u4e24\u8005\u4e0d\u5b8c\u5168\u4e00\u81f4\u3002\u672c\u6587\u4fdd\u7559\u201c\u6a21\u62df\u8f6f\u4ef6\u72b6\u6001\u6267\u884c\u201d\u8fd9\u4e2a\u5171\u540c\u5b9a\u4e49，\u4e0d\u5c06\u5176\u5199\u6210\u5df2\u786e\u5b9a\u7684\u7edf\u4e00 API \u89c4\u8303\u3002<\/p>\n

4.2 \u5206\u6570\u3001\u5b57\u7b26\u4e0e Token \u7684\u4e0d\u540c\u542b\u4e49<\/h4>\n

\u4ed3\u5e93\u4efb\u52a1\u7684\u8bc4\u5206\u53ef\u5199\u6210：<\/p>\n

<\/p>\n

S<\/p>\n

c<\/p>\n

o<\/p>\n

r<\/p>\n

e<\/p>\n

=<\/p>\n

S<\/p>\n

u<\/p>\n

c<\/p>\n

c<\/p>\n

e<\/p>\n

s<\/p>\n

s<\/p>\n

f<\/p>\n

u<\/p>\n

l<\/p>\n

\u00a0<\/p>\n

A<\/p>\n

c<\/p>\n

t<\/p>\n

i<\/p>\n

o<\/p>\n

n<\/p>\n

s<\/p>\n

T<\/p>\n

o<\/p>\n

t<\/p>\n

a<\/p>\n

l<\/p>\n

\u00a0<\/p>\n

A<\/p>\n

c<\/p>\n

t<\/p>\n

i<\/p>\n

o<\/p>\n

n<\/p>\n

a<\/p>\n

b<\/p>\n

l<\/p>\n

e<\/p>\n

\u00a0<\/p>\n

E<\/p>\n

v<\/p>\n

e<\/p>\n

n<\/p>\n

t<\/p>\n

s<\/p>\n

.<\/p>\n

\\\\mathrm{Score}=\\\\frac{\\\\mathrm{Successful\\\\ Actions}}{\\\\mathrm{Total\\\\ Actionable\\\\ Events}}. <\/p>\n

<\/span><\/span>Score<\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>Total<\/span>\u00a0<\/span>Actionable<\/span>\u00a0<\/span>Events<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>Successful<\/span>\u00a0<\/span>Actions<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>.<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u6240\u4ee5 0.94 \u662f\u52a8\u4f5c\u5c42\u9762\u7684\u51c6\u786e\u7387，\u4e0d\u662f\u201c94% \u7684\u5b8c\u6574 100 \u6b65\u4efb\u52a1\u5168\u90e8\u6210\u529f\u201d\u3002\u8f6f\u4ef6\u4ed3\u5e93\u9644\u5f55\u66f4\u5177\u4f53\u5730\u4ee5\u6b63\u786e\u89e3\u51b3\u5e76\u5408\u5165\u4e3b\u5206\u652f\u4e14\u4e0d\u7834\u574f CI \u7684\u8bf7\u6c42\u6bd4\u4f8b\u5b9a\u4e49\u6210\u529f\u3002\u516c\u5f00\u4ea4\u4e92\u57fa\u51c6\u5219\u4f7f\u7528\u4efb\u52a1\u6210\u529f\u7387\u3002\u4e0d\u540c\u6307\u6807\u4e0d\u80fd\u76f4\u63a5\u6df7\u5408\u6210\u4e00\u4e2a\u603b\u5e73\u5747\u3002<\/p>\n

\u8bba\u6587\u6307\u6807\u5b9a\u4e49\u5c06 Average Prompt Size \u89c4\u5b9a\u4e3a\u6bcf\u6b21\u8c03\u7528\u7684\u5e73\u5747\u5b57\u7b26\u4e32\u5b57\u7b26\u6570；Total Tokens \u662f\u7d2f\u8ba1 Token \u6d88\u8017\u3002\u540e\u6587\u6709\u4e2a\u522b\u6bb5\u843d\u548c\u8868\u6ce8\u628a Prompt \u9884\u7b97\u5199\u6210 Token，\u5b58\u5728\u53e3\u5f84\u4e0d\u4e00\u81f4\u3002\u672c\u6587\u4fdd\u7559\u539f\u6570\u503c，\u5e76\u5728\u6d89\u53ca\u8fd9\u4e9b\u53e3\u5f84\u7684\u5730\u65b9\u660e\u786e\u6307\u51fa，\u907f\u514d\u81ea\u884c\u9009\u62e9\u4e00\u4e2a\u5355\u4f4d\u540e\u5f53\u6210\u5df2\u6838\u5b9e\u4e8b\u5b9e\u3002<\/p>\n

\u5e95\u5c42\u6a21\u578b\u4e3a Gemini-3-Flash\u3001Gemma-4-31B-it \u548c Qwen-3-8B-it\u3002\u4f5c\u8005\u8bbe\u7f6e temperature=0\u3001top-p=1，\u5e76\u5bf9\u5408\u6210\u5b9e\u9a8c\u91c7\u7528 5 \u4e2a\u751f\u6210\u79cd\u5b50，\u62a5\u544a\u5747\u503c\u4e0e\u6837\u672c\u6807\u51c6\u5dee\u3002\u96f6\u6e29\u5ea6\u6709\u52a9\u4e8e\u51cf\u5c11\u91c7\u6837\u5dee\u5f02，\u4f46\u4e0d\u5355\u72ec\u4fdd\u8bc1\u6a21\u578b\u670d\u52a1\u3001\u5de5\u5177\u73af\u5883\u4e0e\u5168\u90e8\u8fd0\u884c\u8fc7\u7a0b\u5b8c\u5168\u786e\u5b9a\u3002<\/p>\n

\u4f5c\u8005\u62a5\u544a\u957f\u7a0b\u6761\u4ef6\u4e0b\u7684\u914d\u5bf9 t \u68c0\u9a8c <\/p>\n

p<\/p>\n

<<\/p>\n

0.01<\/p>\n

p<0.01<\/p>\n

<\/span><\/span>p<\/span><\/span><<\/span><\/span><\/span><\/span>0.01<\/span><\/span><\/span><\/span><\/span>\u3002\u6ca1\u6709\u9010\u79cd\u5b50\u7ed3\u679c\u4e0e\u5b8c\u6574\u6bd4\u8f83\u660e\u7ec6，\u5c31\u65e0\u6cd5\u72ec\u7acb\u590d\u7b97\u6bcf\u4e2a\u68c0\u9a8c，\u4e5f\u4e0d\u80fd\u628a\u8fd9\u53e5\u603b\u4f53\u9648\u8ff0\u5957\u5230\u6240\u6709\u6a21\u578b\u3001\u4efb\u52a1\u548c\u8fd0\u884c\u65f6\u7ec4\u5408\u4e0a；\u540e\u6587\u5b58\u5728\u5206\u6570\u76f8\u540c\u548c\u77ed\u7a0b\u8868\u73b0\u4e0d\u5360\u4f18\u7684\u7ec4\u5408\u3002<\/p>\n

4.3 \u9644\u5f55\u7b97\u6cd5：\u540c\u4e00\u968f\u673a\u8fc7\u7a0b\u751f\u6210\u53ef\u6bd4\u4e8b\u4ef6<\/h4>\n

\u8bba\u6587\u9644\u5f55\u7684\u7b2c\u4e8c\u4e2a\u7b97\u6cd5\u7ed9\u51fa\u4e86\u4ed3\u5e93\u4efb\u52a1\u751f\u6210\u8fc7\u7a0b，\u6838\u5fc3\u662f\u56fa\u5b9a\u968f\u673a\u79cd\u5b50，\u8ba9\u4e0d\u540c\u8fd0\u884c\u65f6\u9762\u5bf9\u76f8\u540c\u5916\u90e8\u4e8b\u4ef6\u5e8f\u5217：<\/p>\n

seed = 42 # \u9644\u5f55\u793a\u4f8b\u79cd\u5b50
\nrng = Random(seed)
\ngt_shelves = 500 \u4e2a\u521d\u59cb\u4e3a\u7a7a\u7684\u8d27\u67b6
\nevents = []
\n\u91cd\u590d Horizon \u6b21：
\n available = \u67e5\u8be2\u7a7a\u8d27\u67b6
\n occupied = \u67e5\u8be2\u5360\u7528\u8d27\u67b6
\n possible_events = [Receive]
\n \u82e5 occupied \u975e\u7a7a：\u52a0\u5165 Order \u548c Maintenance
\n event_type = rng \u4ece possible_events \u4e2d\u91c7\u6837
\n events \u8ffd\u52a0\u7531 event_type \u6784\u9020\u7684\u89c2\u5bdf
\n \u6309 event_type \u66f4\u65b0\u751f\u6210\u5668\u7684\u771f\u5b9e\u72b6\u6001
\n\u8fd4\u56de events<\/p>\n

\u8fd9\u91cc\u7684 42 \u662f\u751f\u6210\u4f2a\u4ee3\u7801\u793a\u4f8b，\u4e0d\u4ee3\u8868\u5168\u90e8\u5b9e\u9a8c\u4ec5\u4f7f\u7528\u8fd9\u4e00\u79cd\u5b50\u3002\u9700\u8981\u590d\u73b0\u65f6，\u8fd8\u5e94\u83b7\u5f97\u4e94\u4e2a\u5b9e\u9a8c\u79cd\u5b50\u7684\u5177\u4f53\u53d6\u503c\u3001\u4e8b\u4ef6\u6784\u9020\u89c4\u5219，\u4ee5\u53ca Agent \u52a8\u4f5c\u504f\u79bb\u9884\u671f\u540e\u73af\u5883\u5982\u4f55\u7ee7\u7eed\u53cd\u9988\u3002\u56fa\u5b9a\u5916\u90e8\u4e8b\u4ef6\u6d41\u80fd\u591f\u589e\u5f3a\u53ef\u6bd4\u6027，\u4f46\u771f\u5b9e\u4ea4\u4e92\u4ecd\u53ef\u80fd\u53d7\u5404\u8fd0\u884c\u65f6\u81ea\u8eab\u52a8\u4f5c\u5f71\u54cd\u3002<\/p>\n

\u4e94\u3001\u4ed3\u5e93\u957f\u7a0b\u6267\u884c：\u6700\u6e05\u695a\u7684\u6536\u76ca\u6765\u81ea\u8f93\u5165\u89c4\u6a21\u63a7\u5236<\/h3>\n

\u8bba\u6587\u8868 1｜Gemini-3-Flash \u7684\u4ed3\u5e93\u957f\u5ea6\u6269\u5c55，\u5747\u503c \u00b1 \u6807\u51c6\u5dee，5 \u4e2a\u79cd\u5b50\u3002Avg Prompt \u7684\u5355\u4f4d\u5728\u8be5\u8868\u660e\u786e\u4e3a\u5b57\u7b26；Score \u662f\u52a8\u4f5c\u51c6\u786e\u7387\u3002<\/p>\n\nHorizon (T)RuntimeScore (Accuracy)Avg Prompt (Chars)Total Tokens<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
10<\/td>\nPrompt (ReAct)<\/td>\n0.90 \u00b1 0.02<\/td>\n3,249 \u00b1 94<\/td>\n9,438 \u00b1 371<\/td>\n<\/tr>\n
10<\/td>\nMemory (Summary)<\/td>\n1.00 \u00b1 0.00<\/td>\n3,300 \u00b1 123<\/td>\n9,972 \u00b1 204<\/td>\n<\/tr>\n
10<\/td>\nStateful (LangGraph)<\/td>\n1.00 \u00b1 0.00<\/td>\n3,430 \u00b1 42<\/td>\n10,337 \u00b1 299<\/td>\n<\/tr>\n
10<\/td>\nSKILL.state<\/td>\n1.00 \u00b1 0.00<\/td>\n1,775 \u00b1 74<\/td>\n5,870 \u00b1 131<\/td>\n<\/tr>\n
25<\/td>\nPrompt (ReAct)<\/td>\n0.92 \u00b1 0.02<\/td>\n6,052 \u00b1 192<\/td>\n42,689 \u00b1 2,238<\/td>\n<\/tr>\n
25<\/td>\nMemory (Summary)<\/td>\n0.99 \u00b1 0.00<\/td>\n6,357 \u00b1 203<\/td>\n43,067 \u00b1 1,948<\/td>\n<\/tr>\n
25<\/td>\nStateful (LangGraph)<\/td>\n1.00 \u00b1 0.00<\/td>\n5,858 \u00b1 301<\/td>\n41,238 \u00b1 3,196<\/td>\n<\/tr>\n
25<\/td>\nSKILL.state<\/td>\n1.00 \u00b1 0.00<\/td>\n1,736 \u00b1 49<\/td>\n14,714 \u00b1 564<\/td>\n<\/tr>\n
50<\/td>\nPrompt (ReAct)<\/td>\n0.88 \u00b1 0.04<\/td>\n11,931 \u00b1 346<\/td>\n171,658 \u00b1 6,978<\/td>\n<\/tr>\n
50<\/td>\nMemory (Summary)<\/td>\n0.93 \u00b1 0.03<\/td>\n7,582 \u00b1 283<\/td>\n131,455 \u00b1 6,841<\/td>\n<\/tr>\n
50<\/td>\nStateful (LangGraph)<\/td>\n0.94 \u00b1 0.00<\/td>\n11,594 \u00b1 438<\/td>\n170,992 \u00b1 7,918<\/td>\n<\/tr>\n
50<\/td>\nSKILL.state<\/td>\n0.96 \u00b1 0.01<\/td>\n1,773 \u00b1 53<\/td>\n30,151 \u00b1 1,231<\/td>\n<\/tr>\n
100<\/td>\nPrompt (ReAct)<\/td>\n0.84 \u00b1 0.07<\/td>\n36,362 \u00b1 1,304<\/td>\n1,245,413 \u00b1 53,241<\/td>\n<\/tr>\n
100<\/td>\nMemory (Summary)<\/td>\n0.87 \u00b1 0.05<\/td>\n29,607 \u00b1 978<\/td>\n1,082,154 \u00b1 83,212<\/td>\n<\/tr>\n
100<\/td>\nStateful (LangGraph)<\/td>\n0.91 \u00b1 0.02<\/td>\n31,354 \u00b1 831<\/td>\n1,062,387 \u00b1 53,839<\/td>\n<\/tr>\n
100<\/td>\nSKILL.state<\/td>\n0.94 \u00b1 0.01<\/td>\n1,905 \u00b1 93<\/td>\n65,408 \u00b1 5,431<\/td>\n<\/tr>\n
200<\/td>\nPrompt (ReAct)<\/td>\n0.74 \u00b1 0.14<\/td>\n48,007 \u00b1 2,092<\/td>\n2,608,755 \u00b1 102,415<\/td>\n<\/tr>\n
200<\/td>\nMemory (Summary)<\/td>\n0.84 \u00b1 0.09<\/td>\n84,364 \u00b1 3,446<\/td>\n6,175,509 \u00b1 294,089<\/td>\n<\/tr>\n
200<\/td>\nStateful (LangGraph)<\/td>\n0.88 \u00b1 0.03<\/td>\n72,305 \u00b1 3,096<\/td>\n5,041,164 \u00b1 346,925<\/td>\n<\/tr>\n
200<\/td>\nSKILL.state<\/td>\n0.94 \u00b1 0.02<\/td>\n1,811 \u00b1 184<\/td>\n122,384 \u00b1 4,522<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4ece 10 \u6b65\u5230 200 \u6b65，SKILL.state \u7684\u5e73\u5747 Prompt \u5728\u7ea6 1,736\u20141,905 \u5b57\u7b26\u4e4b\u95f4\u53d8\u5316，\u6574\u4f53\u63a5\u8fd1\u7a33\u5b9a\u3002\u76f8\u5bf9\u5730，Memory \u5728 200 \u6b65\u65f6\u5e73\u5747 Prompt \u8fbe\u5230 84,364 \u5b57\u7b26，\u7d2f\u8ba1 Token \u8fbe\u5230 6,175,509\u3002\u6458\u8981\u673a\u5236\u7684\u5b58\u5728，\u5e76\u6ca1\u6709\u5728\u8fd9\u4e00\u5b9e\u73b0\u4e2d\u81ea\u52a8\u5f62\u6210\u4e25\u683c\u5bb9\u91cf\u4e0a\u9650\u3002<\/p>\n

100 \u6b65\u65f6，Stateful \u4e0e SKILL.state \u7684\u7d2f\u8ba1 Token \u6bd4\u4e3a <\/p>\n

1,062,387<\/p>\n

\/<\/p>\n

65,408<\/p>\n

\u2248<\/p>\n

16.24<\/p>\n

1{,}062{,}387\/65{,}408\\\\approx16.24<\/p>\n

<\/span><\/span>1<\/span>,<\/span><\/span>062<\/span>,<\/span><\/span>387\/65<\/span>,<\/span><\/span>408<\/span><\/span>\u2248<\/span><\/span><\/span><\/span>16.24<\/span><\/span><\/span><\/span><\/span>；\u540e\u8005\u76f8\u5bf9\u51cf\u5c11\u7ea6 93.84%\u3002\u5206\u6570\u4ece 0.91 \u63d0\u9ad8\u5230 0.94，\u662f 3 \u4e2a\u767e\u5206\u70b9\u7684\u52a8\u4f5c\u51c6\u786e\u7387\u5dee\u3002\u76f8\u5bf9 Prompt \u57fa\u7ebf，Token \u6bd4\u7ea6\u4e3a 19.04，\u51c6\u786e\u7387\u5dee\u4e3a 10 \u4e2a\u767e\u5206\u70b9\u3002\u9009\u62e9\u4e0d\u540c\u57fa\u7ebf\u4f1a\u5f97\u5230\u4e0d\u540c\u500d\u7387，\u5f15\u7528\u65f6\u5fc5\u987b\u540c\u65f6\u7ed9\u51fa\u5206\u6bcd\u3002<\/p>\n

200 \u6b65\u65f6，SKILL.state \u4ecd\u4e3a <\/p>\n

0.94<\/p>\n

\u00b1<\/p>\n

0.02<\/p>\n

0.94\\\\pm0.02<\/p>\n

<\/span><\/span>0.94<\/span><\/span>\u00b1<\/span><\/span><\/span><\/span>0.02<\/span><\/span><\/span><\/span><\/span>，\u7d2f\u8ba1 122,384 Token\u3002\u8fd9\u652f\u6301\u201c\u5728\u7ed9\u5b9a\u4ed3\u5e93\u73af\u5883\u4e2d，\u663e\u5f0f\u72b6\u6001\u6709\u6548\u6291\u5236\u5386\u53f2\u8f93\u5165\u589e\u957f\u201d\u7684\u7ed3\u8bba\u3002\u4f46\u5b83\u6ca1\u6709\u8bc1\u660e\u4efb\u610f\u5b9e\u4f53\u89c4\u6a21\u3001\u4efb\u610f\u9886\u57df\u6216\u65e0\u9650\u957f\u5ea6\u4e0b\u4ecd\u80fd\u7ef4\u6301\u540c\u7b49\u6b63\u786e\u7387\u3002500 \u4e2a\u8d27\u67b6\u662f\u73af\u5883\u5bb9\u91cf，200 \u6b65\u662f\u5b9e\u6d4b\u6267\u884c\u957f\u5ea6，\u4e24\u8005\u4e5f\u4e0d\u80fd\u4e92\u6362\u3002<\/p>\n

\u516d\u3001\u566a\u58f0\u9c81\u68d2\u6027：\u5b9e\u9a8c\u9a8c\u8bc1\u7684\u662f\u65e0\u5173\u5e72\u6270\u8fc7\u6ee4<\/h3>\n

\u8bba\u6587\u8868 2｜\u4ed3\u5e93\u566a\u58f0\u5b9e\u9a8c，50 \u6b65，Gemini-3-Flash\u3002<\/p>\n\nNoise LevelRuntimeScore<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n\n
5 Events (Low)<\/td>\nPrompt<\/td>\n0.68<\/td>\n<\/tr>\n
5 Events (Low)<\/td>\nMemory<\/td>\n1.00<\/td>\n<\/tr>\n
5 Events (Low)<\/td>\nStateful<\/td>\n1.00<\/td>\n<\/tr>\n
5 Events (Low)<\/td>\nSKILL.state<\/td>\n1.00<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nPrompt<\/td>\n0.61<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nMemory<\/td>\n1.00<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nStateful<\/td>\n0.98<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nSKILL.state<\/td>\n0.97<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nPrompt<\/td>\n0.53<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nMemory<\/td>\n0.96<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nStateful<\/td>\n0.98<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nSKILL.state<\/td>\n0.98<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u5728\u6bcf\u6b65 50 \u6761\u566a\u58f0\u4e0b，Prompt \u7684\u5206\u6570\u4e3a 0.53，SKILL.state \u4e3a 0.98\u3002\u628a\u65e0\u5173\u4fe1\u606f\u8fc7\u6ee4\u5728\u72b6\u6001\u8865\u4e01\u751f\u6210\u73af\u8282\u4e4b\u5916，\u53ef\u4ee5\u51cf\u5c11\u5b83\u4eec\u5728\u672a\u6765\u6b65\u9aa4\u91cd\u590d\u51fa\u73b0，\u8fd9\u662f\u7ed3\u679c\u652f\u6301\u7684\u4e00\u79cd\u673a\u5236\u89e3\u91ca\u3002<\/p>\n

\u4f46 SKILL.state \u5e76\u672a\u5728\u6240\u6709\u566a\u58f0\u6761\u4ef6\u4e0b\u72ec\u5360\u6700\u4f18\u300220 \u6761\u566a\u58f0\u65f6，Memory \u4e3a 1.00\u3001Stateful \u4e3a 0.98，\u800c SKILL.state \u4e3a 0.97；50 \u6761\u566a\u58f0\u65f6，SKILL.state \u4e0e Stateful \u540c\u4e3a 0.98\u3002\u7ed3\u679c\u66f4\u9002\u5408\u8868\u8ff0\u4e3a\u76f8\u5bf9\u76f4\u63a5\u7d2f\u79ef\u5386\u53f2\u5177\u6709\u660e\u663e\u9c81\u68d2\u6027，\u800c\u975e\u5bf9\u5168\u90e8\u8bb0\u5fc6\u7b56\u7565\u7684\u5168\u9762\u80dc\u51fa\u3002<\/p>\n

\u9644\u5f55\u5c06\u5b9e\u9645\u8bc4\u6d4b\u9650\u5b9a\u4e3a Condition 1：\u968f\u673a\u751f\u6210\u3001\u4e0e\u4e3b\u4efb\u52a1\u4e25\u683c\u65e0\u5173\u3001\u4e14\u4e0d\u4f1a\u6539\u53d8\u771f\u5b9e\u73af\u5883\u72b6\u6001\u7684\u80cc\u666f\u9065\u6d4b\u3002\u5c3d\u7ba1\u6b63\u6587\u4f7f\u7528\u4e86\u8f83\u5bbd\u6cdb\u7684\u201ccontext corruption\u201d\u63aa\u8f9e，\u8fd9\u7ec4\u5b9e\u9a8c\u4e0d\u80fd\u8bc1\u660e\u5bf9\u6297\u63d0\u793a\u6ce8\u5165\u3001\u4f2a\u9020\u4e1a\u52a1\u6307\u4ee4\u3001\u6076\u610f\u5de5\u5177\u8f93\u51fa\u6216\u771f\u6b63\u72b6\u6001\u7a81\u53d8\u4e0b\u7684\u5b89\u5168\u6027\u3002\u65b0\u89c2\u5bdf\u4f9d\u7136\u76f4\u63a5\u8fdb\u5165\u672c\u6b65 Prompt，\u6076\u610f\u5185\u5bb9\u4ecd\u53ef\u80fd\u5f71\u54cd\u6a21\u578b\u63d0\u51fa\u7684\u8865\u4e01\u3002<\/p>\n

\u4e03\u3001\u72b6\u6001\u6062\u590d：\u96f6\u989d\u5916\u6ede\u540e\u9700\u8981\u7ea0\u6b63\u4fe1\u606f\u5230\u8fbe<\/h3>\n

\u8bba\u6587\u8868 3｜\u4ed3\u5e93\u72b6\u6001\u6062\u590d，Gemini-3-Flash\u3002<\/p>\n\nScenarioRuntimeSuccessRecovery Steps<\/tr>\n\n\n\n\n\n\n\n\n
A: Secret Audit<\/td>\nPrompt \/ Memory \/ Stateful<\/td>\nYes<\/td>\n5\u20138<\/td>\n<\/tr>\n
A: Secret Audit<\/td>\nSKILL.state<\/td>\nYes<\/td>\n0<\/td>\n<\/tr>\n
B: Secret Barcode<\/td>\nPrompt \/ Memory \/ Stateful<\/td>\nYes<\/td>\n6\u20138<\/td>\n<\/tr>\n
B: Secret Barcode<\/td>\nSKILL.state<\/td>\nYes<\/td>\n0<\/td>\n<\/tr>\n
C: Secret Move<\/td>\nPrompt \/ Memory \/ Stateful<\/td>\nYes<\/td>\n5\u20138<\/td>\n<\/tr>\n
C: Secret Move<\/td>\nSKILL.state<\/td>\nYes<\/td>\n0<\/td>\n<\/tr>\n
D: Canceled Order<\/td>\nAll Runtimes<\/td>\nNo<\/td>\nN\/A<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Secret Audit\u3001Secret Barcode \u548c Secret Move \u4e09\u79cd\u573a\u666f\u4e2d，SKILL.state \u62a5\u544a 0 \u4e2a\u6062\u590d\u6b65\u9aa4，\u5386\u53f2\u57fa\u7ebf\u51fa\u73b0\u7ea6 5\u20148 \u6216 6\u20148 \u6b65\u6ede\u540e\u3002\u8bba\u6587\u89e3\u91ca\u662f：\u73af\u5883\u63d0\u4f9b\u7ea0\u6b63\u544a\u8b66\u540e，\u72b6\u6001\u7acb\u5373\u66f4\u65b0，\u8fc7\u65f6\u4e8b\u5b9e\u4e0d\u518d\u901a\u8fc7\u957f\u5386\u53f2\u53cd\u590d\u5f71\u54cd\u5224\u65ad\u3002<\/p>\n

\u201c0\u201d\u5e94\u7406\u89e3\u4e3a\u8be5\u6062\u590d\u6307\u6807\u4e0b\u6ca1\u6709\u989d\u5916\u6ede\u540e，\u4e0d\u80fd\u7406\u89e3\u4e3a\u7cfb\u7edf\u65e0\u9700\u89c2\u5bdf\u5c31\u80fd\u53d1\u73b0\u5916\u90e8\u53d8\u5316\u3002\u663e\u5f0f\u72b6\u6001\u6709\u5229\u4e8e\u8986\u76d6\u65e7\u4e8b\u5b9e，\u72b6\u6001\u66f4\u65b0\u4ecd\u7136\u9700\u8981\u53ef\u4fe1\u7684\u65b0\u8bc1\u636e\u3002 \u5982\u679c\u73b0\u5b9e\u4e2d\u6709\u4eba\u79fb\u52a8\u7269\u54c1\u4f46\u5de5\u5177\u7cfb\u7edf\u6ca1\u6709\u63d0\u4f9b\u4efb\u4f55\u53d8\u5316\u4fe1\u53f7，\u4e09\u5143\u7ec4\u4e2d\u540c\u6837\u7f3a\u5c11\u6062\u590d\u4f9d\u636e\u3002<\/p>\n

\u6b64\u5916，Canceled Order \u573a\u666f\u4e0b\u6240\u6709\u8fd0\u884c\u65f6\u90fd\u5931\u8d25\u3002\u8bba\u6587\u6ca1\u6709\u5728\u8be5\u8868\u63d0\u4f9b\u8db3\u4ee5\u5b9a\u4f4d\u5171\u540c\u5931\u8d25\u6839\u56e0\u7684\u5b8c\u6574\u8f68\u8ff9，\u56e0\u6b64\u5e94\u4fdd\u7559\u8fd9\u4e2a\u53cd\u4f8b，\u800c\u4e0d\u662f\u5c06\u6062\u590d\u673a\u5236\u6982\u62ec\u4e3a\u6240\u6709\u6f02\u79fb\u90fd\u80fd\u81ea\u52a8\u5904\u7406\u3002<\/p>\n

\u516b\u3001\u516c\u5f00\u4ea4\u4e92\u57fa\u51c6：\u6210\u529f\u7387\u4e0e\u7d2f\u8ba1 Token \u7684\u6539\u5584\u5e76\u4e0d\u8981\u6c42\u6bcf\u6b65\u66f4\u77ed<\/h3>\n

\u8bba\u6587\u8868 4｜Gemini-3-Flash \u5728\u4e09\u4e2a\u516c\u5f00\u4ea4\u4e92\u4efb\u52a1\u8bbe\u7f6e\u4e2d\u7684\u7ed3\u679c\u3002k \u4e0e M \u5206\u522b\u8868\u793a\u5343\u4e0e\u767e\u4e07；Prompt \u5217\u6cbf\u7528\u8bba\u6587\u540d\u79f0，\u5176\u5355\u4f4d\u95ee\u9898\u89c1\u4e0b\u6587\u3002<\/p>\n\nRuntimeCTF Pass@1CTF PromptCTF TokensRetail Pass RateRetail PromptRetail TokensAirline Pass RateAirline PromptAirline Tokens<\/tr>\n\n\n\n\n\n
Prompt (ReAct)<\/td>\n43.2%<\/td>\n1,909<\/td>\n977k<\/td>\n48.2%<\/td>\n2,819<\/td>\n4.48M<\/td>\n21.8%<\/td>\n5,100<\/td>\n4.85M<\/td>\n<\/tr>\n
Memory (Summary)<\/td>\n46.4%<\/td>\n1,797<\/td>\n1.03M<\/td>\n29.9%<\/td>\n2,737<\/td>\n4.24M<\/td>\n23.6%<\/td>\n4,700<\/td>\n4.65M<\/td>\n<\/tr>\n
Stateful (LangGraph)<\/td>\n41.8%<\/td>\n1,946<\/td>\n1.13M<\/td>\n51.7%<\/td>\n3,065<\/td>\n3.92M<\/td>\n28.1%<\/td>\n5,400<\/td>\n5.28M<\/td>\n<\/tr>\n
SKILL.state<\/td>\n54.2%<\/td>\n813<\/td>\n387k<\/td>\n58.3%<\/td>\n3,325<\/td>\n3.47M<\/td>\n32.4%<\/td>\n2,800<\/td>\n2.88M<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

8.1 InterCode CTF：\u663e\u5f0f\u4fdd\u5b58\u5047\u8bbe，\u51cf\u5c11\u65e0\u6548\u91cd\u590d<\/h4>\n

SKILL.state \u7684 pass@1 \u4e3a 54.2%，\u6bd4\u6700\u5f3a\u57fa\u7ebf Memory \u7684 46.4% \u9ad8 7.8 \u4e2a\u767e\u5206\u70b9，\u6bd4 Stateful \u7684 41.8% \u9ad8 12.4 \u4e2a\u767e\u5206\u70b9\u3002\u7d2f\u8ba1 Token \u4ece ReAct \u7684 977k \u964d\u4e3a 387k，\u6309\u8868\u4e2d\u820d\u5165\u6570\u8ba1\u7b97，\u51cf\u5c11\u7ea6 60.4%\u3002<\/p>\n

CTF \u72b6\u6001\u4fdd\u5b58\u5df2\u6d4b\u8bd5\u5047\u8bbe\u4e0e\u5f53\u524d\u5de5\u4f5c\u76ee\u5f55，\u56e0\u800c\u5177\u6709\u51cf\u5c11\u91cd\u590d\u63a2\u7d22\u7684\u5408\u7406\u673a\u5236\u3002\u5b83\u4ecd\u9700\u8981\u6a21\u578b\u9009\u62e9\u503c\u5f97\u6d4b\u8bd5\u7684\u4e0b\u4e00\u6761\u5047\u8bbe；\u72b6\u6001\u7ef4\u62a4\u4e0d\u80fd\u66ff\u4ee3\u89e3\u9898\u63a8\u7406\u3002\u672c\u5b9e\u9a8c\u63d0\u4f9b\u7684\u662f\u4f5c\u8005\u62a5\u544a\u7684\u4efb\u52a1\u7ea7\u7ed3\u679c，\u6ca1\u6709\u901a\u8fc7\u66f4\u7ec6\u6d88\u878d\u5355\u72ec\u8bc6\u522b\u6bcf\u4e2a\u5b57\u6bb5\u7684\u8d21\u732e\u3002<\/p>\n

\u8fd8\u6709\u4e00\u4e2a\u7edf\u8ba1\u53e3\u5f84\u9700\u8981\u6f84\u6e05：\u8868\u5934\u5199 100 \u4e2a\u4efb\u52a1，\u4e14\u6307\u6807\u88ab\u5b9a\u4e49\u4e3a\u4e8c\u5143 pass@1，\u4f46 54.2%\u300143.2% \u8fd9\u6837\u7684\u6bd4\u4f8b\u4e0d\u80fd\u76f4\u63a5\u5bf9\u5e94\u201c100 \u4e2a\u4efb\u52a1\u5404\u8fd0\u884c\u4e00\u6b21\u201d\u7684\u6574\u6570\u6210\u529f\u6570\u3002\u5b83\u53ef\u80fd\u6d89\u53ca\u91cd\u590d\u8bc4\u4f30\u6216\u5176\u4ed6\u805a\u5408\u65b9\u5f0f，\u672c\u6587\u6838\u5bf9\u7684\u63cf\u8ff0\u4e0d\u8db3\u4ee5\u786e\u5b9a\u3002\u56e0\u6b64\u4fdd\u7559\u62a5\u544a\u503c，\u4e0d\u628a\u5b83\u6539\u5199\u4e3a\u201c54.2 \u9053\u9898\u6210\u529f\u201d，\u4e5f\u4e0d\u64c5\u81ea\u8865\u5145\u91cd\u590d\u6b21\u6570\u3002<\/p>\n

8.2 Retail：\u5e73\u5747 Prompt \u66f4\u957f，\u4f46\u7d2f\u8ba1 Token \u66f4\u5c11<\/h4>\n

Retail \u4e0a，SKILL.state \u6210\u529f\u7387\u4e3a 58.3%，\u9ad8\u4e8e Stateful \u7684 51.7%，\u7d2f\u8ba1 Token \u4e3a 3.47M，\u4e5f\u662f\u8868\u4e2d\u6700\u4f4e\u3002\u7136\u800c\u5176 Prompt \u503c\u4e3a 3,325，\u53cd\u800c\u9ad8\u4e8e ReAct \u7684 2,819\u3001Memory \u7684 2,737 \u4e0e Stateful \u7684 3,065\u3002<\/p>\n

\u8fd9\u4e2a\u53cd\u4f8b\u5f88\u91cd\u8981：\u9ad8\u6548\u6267\u884c\u4e0d\u7b49\u4e8e\u6bcf\u4e00\u6b21\u6a21\u578b\u8c03\u7528\u90fd\u4f7f\u7528\u6700\u77ed\u8f93\u5165\u3002 \u66f4\u5145\u5206\u7684\u72b6\u6001\u53ef\u80fd\u589e\u52a0\u5355\u6b65\u8f93\u5165，\u5374\u5e2e\u52a9\u51cf\u5c11\u65e0\u6548\u4ea4\u4e92\u3002\u8bba\u6587\u6ca1\u6709\u5728\u8fd9\u91cc\u7ed9\u51fa\u8c03\u7528\u6b21\u6570\u3001\u8f93\u51fa Token \u548c\u5931\u8d25\u91cd\u8bd5\u7684\u5b8c\u6574\u5206\u89e3，\u6240\u4ee5\u201c\u6b65\u9aa4\u51cf\u5c11\u201d\u5c5e\u4e8e\u5408\u7406\u89e3\u91ca，\u4e0d\u80fd\u5199\u6210\u5df2\u7ecf\u9010\u9879\u6d4b\u91cf\u7684\u786e\u5b9a\u5f52\u56e0\u3002<\/p>\n

8.3 Airline：\u66f4\u4f4e\u7d2f\u8ba1\u6d88\u8017，\u6210\u529f\u7387\u4ecd\u53ea\u6709\u4e09\u6210\u5de6\u53f3<\/h4>\n

Airline \u4e0a，SKILL.state \u7684\u6210\u529f\u7387\u4e3a 32.4%，\u6bd4 Stateful \u7684 28.1% \u9ad8 4.3 \u4e2a\u767e\u5206\u70b9；\u7d2f\u8ba1 Token \u4e3a 2.88M，\u4f4e\u4e8e ReAct \u7684 4.85M \u548c Stateful \u7684 5.28M\u3002\u6309\u8868\u4e2d\u5df2\u7ecf\u820d\u5165\u7684\u6570\u503c\u91cd\u65b0\u8ba1\u7b97，\u964d\u5e45\u5206\u522b\u7ea6\u4e3a 40.6% \u4e0e 45.5%；\u8bba\u6587\u6b63\u6587\u7ed9\u51fa 40.5% \u4e0e 45.4%，\u53ef\u80fd\u6d89\u53ca\u672a\u5c55\u793a\u7cbe\u5ea6，\u672c\u6587\u4e0d\u5c06\u4e8c\u8005\u5f3a\u884c\u89c6\u4e3a\u5b8c\u5168\u4e00\u81f4\u3002<\/p>\n

32.4% \u540c\u65f6\u8bf4\u660e\u4efb\u52a1\u4ecd\u7136\u56f0\u96be，\u4e0d\u80fd\u53ea\u5f15\u7528\u76f8\u5bf9\u63d0\u5347\u800c\u7565\u53bb\u7edd\u5bf9\u6210\u529f\u7387\u3002\u8bba\u6587\u6b63\u6587\u5c06\u7ea6 2,800 \u7684 Prompt \u63cf\u8ff0\u6210 tokens\/step，\u800c\u6307\u6807\u5b9a\u4e49\u4f7f\u7528\u5b57\u7b26\u6570\u3002\u7f3a\u5c11\u65e5\u5fd7\u4e0e\u7edf\u4e00\u8ba1\u6570\u8bf4\u660e\u65f6，\u65e0\u6cd5\u636e\u6b64\u7cbe\u786e\u91cd\u5efa\u5355\u6b65 Token \u9884\u7b97\u3002<\/p>\n

\u4e5d\u3001\u9884\u7b97\u5339\u914d：\u77ed\u8f93\u5165\u672c\u8eab\u4e0d\u8db3\u4ee5\u89e3\u91ca\u5168\u90e8\u6536\u76ca<\/h3>\n

\u8bba\u6587\u8868 5｜\u4ed3\u5e93 100 \u6b65，Gemini-3-Flash，\u9884\u7b97\u5339\u914d\u63a7\u5236\u3002\u8868\u6ce8\u79f0\u7ea6 1,800 tokens，\u6b63\u6587\u540c\u4e00\u5b9e\u9a8c\u53c8\u79f0\u7ea6 1,800 characters；\u4e0b\u9762\u4fdd\u7559 Avg Prompt \u540d\u79f0\u548c\u539f\u6570\u503c，\u4e0d\u81ea\u884c\u7edf\u4e00\u5355\u4f4d\u3002<\/p>\n\nRuntime \/ ConfigurationScoreAvg PromptTotal Tokens<\/tr>\n\n\n\n\n\n\n
Full ReAct (Unbounded)<\/td>\n0.84<\/td>\n36,362<\/td>\n1,245,413<\/td>\n<\/tr>\n
Truncated (Sliding Window)<\/td>\n0.18<\/td>\n1,800<\/td>\n62,100<\/td>\n<\/tr>\n
Summary-capped<\/td>\n0.52<\/td>\n1,840<\/td>\n63,400<\/td>\n<\/tr>\n
ReAct + LLMLingua<\/td>\n0.22<\/td>\n1,810<\/td>\n62,350<\/td>\n<\/tr>\n
SKILL.state (Structured)<\/td>\n0.94<\/td>\n1,905<\/td>\n65,408<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u6ed1\u52a8\u7a97\u53e3\u3001\u53d7\u9650\u6458\u8981\u548c LLMLingua \u7684\u7d2f\u8ba1 Token \u90fd\u5728\u7ea6 6.2 \u4e07\u20146.3 \u4e07，\u63a5\u8fd1 SKILL.state \u7684 65,408\u3002\u4f46\u5206\u6570\u5206\u522b\u4e3a 0.18\u30010.52 \u548c 0.22，\u663e\u8457\u4f4e\u4e8e SKILL.state \u7684 0.94\u3002\u8fd9\u4e2a\u6bd4\u8f83\u652f\u6301\u4e00\u4e2a\u5b9e\u7528\u5224\u65ad：\u6267\u884c\u4f9d\u8d56\u7cbe\u786e\u5b9e\u4f53\u5173\u7cfb\u65f6，\u4fdd\u7559\u7684\u4fe1\u606f\u7c7b\u578b\u6bd4\u5355\u7eaf\u51cf\u5c11\u8f93\u5165\u957f\u5ea6\u66f4\u5173\u952e\u3002<\/p>\n

\u4f8b\u5982，\u8d27\u67b6\u7f16\u53f7\u770b\u8d77\u6765\u91cd\u590d，\u5b9e\u9645\u4e0a\u627f\u62c5\u201c\u54ea\u4e2a\u7269\u54c1\u5728\u54ea\u91cc\u201d\u7684\u6620\u5c04\u5173\u7cfb\u3002\u7edf\u8ba1\u538b\u7f29\u5982\u679c\u7834\u574f\u8fd9\u79cd\u5173\u8054，\u5373\u4f7f\u7559\u4e0b\u4e86\u4efb\u52a1\u5927\u610f，\u4e5f\u4e0d\u80fd\u6b63\u786e\u751f\u6210\u4e0b\u4e00\u6761\u52a8\u4f5c\u3002\u4f5c\u8005\u5c06 LLMLingua \u7684\u9000\u5316\u5f52\u56e0\u4e3a\u5173\u952e\u69fd\u4f4d\u4fe1\u606f\u4e22\u5931；\u672c\u6587\u5c06\u5176\u9650\u5b9a\u4e3a\u8be5\u914d\u7f6e\u4e0e\u8be5\u4efb\u52a1\u7684\u89e3\u91ca，\u4e0d\u63a8\u5e7f\u6210\u6240\u6709\u538b\u7f29\u5668\u90fd\u4f1a\u5931\u6548\u3002<\/p>\n

\u8fd9\u7ec4\u5b9e\u9a8c\u4e0d\u662f\u4e25\u683c\u610f\u4e49\u4e0a\u7684\u5b8c\u5168\u7b49\u9884\u7b97\u3002SKILL.state \u7684 Avg Prompt \u4e3a 1,905，\u6ed1\u52a8\u7a97\u53e3\u4e3a 1,800，\u76f8\u5dee\u7ea6 5.8%；\u7d2f\u8ba1 Token \u4e5f\u7565\u6709\u4e0d\u540c，\u518d\u52a0\u4e0a\u5355\u4f4d\u4e0d\u4e00\u81f4\u3001\u538b\u7f29\u5668\u5177\u4f53\u914d\u7f6e\u672a\u5145\u5206\u5c55\u5f00，\u4e0d\u80fd\u5ba3\u79f0\u5df2\u6392\u9664\u6240\u6709\u8d44\u6e90\u4e0e\u5b9e\u73b0\u5dee\u5f02\u3002\u66f4\u7a33\u59a5\u7684\u7ed3\u8bba\u662f：\u5728\u4f5c\u8005\u62a5\u544a\u7684\u8fd1\u4f3c\u9884\u7b97\u5bf9\u7167\u4e2d，\u663e\u5f0f\u72b6\u6001\u4fdd\u7559\u4e86\u5bf9\u4ed3\u5e93\u4efb\u52a1\u66f4\u6709\u7528\u7684\u7ed3\u6784\u3002<\/p>\n

\u5341\u3001\u8f6f\u4ef6\u4ed3\u5e93\u6269\u5c55：\u5173\u7cfb\u72b6\u6001\u6bd4\u72ec\u7acb\u5e93\u5b58\u66f4\u96be\u7ef4\u62a4<\/h3>\n

\u8bba\u6587\u8868 6｜Gemini-3-Flash \u7684\u6a21\u62df\u8f6f\u4ef6\u4ed3\u5e93\u957f\u5ea6\u6269\u5c55\u3002\u6309\u8bba\u6587\u6307\u6807\u5b9a\u4e49，Avg Prompt Size \u4e3a\u5e73\u5747\u5b57\u7b26\u957f\u5ea6\u3002<\/p>\n\nHorizonRuntimeScoreAvg Prompt SizeTotal Tokens Consumed<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
10<\/td>\nPrompt<\/td>\n0.89 \u00b1 0.11<\/td>\n3,411 \u00b1 197<\/td>\n11,670 \u00b1 841<\/td>\n<\/tr>\n
10<\/td>\nMemory<\/td>\n0.93 \u00b1 0.09<\/td>\n4,379 \u00b1 234<\/td>\n15,732 \u00b1 562<\/td>\n<\/tr>\n
10<\/td>\nStateful<\/td>\n1.00 \u00b1 0.00<\/td>\n4,200 \u00b1 321<\/td>\n14,120 \u00b1 318<\/td>\n<\/tr>\n
10<\/td>\nSKILL.state<\/td>\n1.00 \u00b1 0.00<\/td>\n2,298 \u00b1 134<\/td>\n7,608 \u00b1 149<\/td>\n<\/tr>\n
25<\/td>\nPrompt<\/td>\n0.84 \u00b1 0.05<\/td>\n11,754 \u00b1 608<\/td>\n111,970 \u00b1 3,314<\/td>\n<\/tr>\n
25<\/td>\nMemory<\/td>\n0.89 \u00b1 0.07<\/td>\n9,399 \u00b1 317<\/td>\n94,629 \u00b1 2,839<\/td>\n<\/tr>\n
25<\/td>\nStateful<\/td>\n0.94 \u00b1 0.03<\/td>\n14,016 \u00b1 586<\/td>\n128,702 \u00b1 3,863<\/td>\n<\/tr>\n
25<\/td>\nSKILL.state<\/td>\n0.88 \u00b1 0.08<\/td>\n2,545 \u00b1 556<\/td>\n21,920 \u00b1 431<\/td>\n<\/tr>\n
50<\/td>\nPrompt<\/td>\n0.71 \u00b1 0.14<\/td>\n23,136 \u00b1 911<\/td>\n462,118 \u00b1 13,764<\/td>\n<\/tr>\n
50<\/td>\nMemory<\/td>\n0.65 \u00b1 0.12<\/td>\n35,550 \u00b1 2,412<\/td>\n688,182 \u00b1 23,539<\/td>\n<\/tr>\n
50<\/td>\nStateful<\/td>\n0.74 \u00b1 0.08<\/td>\n31,166 \u00b1 3,231<\/td>\n577,027 \u00b1 27,293<\/td>\n<\/tr>\n
50<\/td>\nSKILL.state<\/td>\n0.86 \u00b1 0.04<\/td>\n2,545 \u00b1 63<\/td>\n45,100 \u00b1 894<\/td>\n<\/tr>\n
100<\/td>\nPrompt<\/td>\n0.53 \u00b1 0.16<\/td>\n46,270 \u00b1 1,847<\/td>\n1,848,500 \u00b1 55,391<\/td>\n<\/tr>\n
100<\/td>\nMemory<\/td>\n0.57 \u00b1 0.05<\/td>\n71,100 \u00b1 5,836<\/td>\n2,752,700 \u00b1 82,467<\/td>\n<\/tr>\n
100<\/td>\nStateful<\/td>\n0.63 \u00b1 0.10<\/td>\n62,330 \u00b1 2,488<\/td>\n2,308,000 \u00b1 35,183<\/td>\n<\/tr>\n
100<\/td>\nSKILL.state<\/td>\n0.78 \u00b1 0.08<\/td>\n2,545 \u00b1 471<\/td>\n90,200 \u00b1 2,792<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

100 \u6b65\u65f6，SKILL.state \u4e3a <\/p>\n

0.78<\/p>\n

\u00b1<\/p>\n

0.08<\/p>\n

0.78\\\\pm0.08<\/p>\n

<\/span><\/span>0.78<\/span><\/span>\u00b1<\/span><\/span><\/span><\/span>0.08<\/span><\/span><\/span><\/span><\/span>，Stateful \u4e3a <\/p>\n

0.63<\/p>\n

\u00b1<\/p>\n

0.10<\/p>\n

0.63\\\\pm0.10<\/p>\n

<\/span><\/span>0.63<\/span><\/span>\u00b1<\/span><\/span><\/span><\/span>0.10<\/span><\/span><\/span><\/span><\/span>，\u7d2f\u8ba1 Token \u5206\u522b\u4e3a 90,200 \u548c 2,308,000\u3002\u663e\u5f0f\u72b6\u6001\u7684\u76f8\u5bf9\u4f18\u52bf\u4ecd\u5728，\u4f46 SKILL.state \u81ea\u8eab\u4ece 10 \u6b65\u7684 1.00 \u964d\u5230\u4e86 100 \u6b65\u7684 0.78，\u8bf4\u660e\u63a7\u5236\u4e0a\u4e0b\u6587\u589e\u957f\u5e76\u4e0d\u80fd\u5b8c\u5168\u6d88\u9664\u957f\u7a0b\u9519\u8bef\u3002<\/p>\n

25 \u6b65\u65f6，SKILL.state \u4e3a <\/p>\n

0.88<\/p>\n

\u00b1<\/p>\n

0.08<\/p>\n

0.88\\\\pm0.08<\/p>\n

<\/span><\/span>0.88<\/span><\/span>\u00b1<\/span><\/span><\/span><\/span>0.08<\/span><\/span><\/span><\/span><\/span>，\u4f4e\u4e8e Stateful \u7684 <\/p>\n

0.94<\/p>\n

\u00b1<\/p>\n

0.03<\/p>\n

0.94\\\\pm0.03<\/p>\n

<\/span><\/span>0.94<\/span><\/span>\u00b1<\/span><\/span><\/span><\/span>0.03<\/span><\/span><\/span><\/span><\/span>，\u4e5f\u7565\u4f4e\u4e8e Memory \u7684 <\/p>\n

0.89<\/p>\n

\u00b1<\/p>\n

0.07<\/p>\n

0.89\\\\pm0.07<\/p>\n

<\/span><\/span>0.89<\/span><\/span>\u00b1<\/span><\/span><\/span><\/span>0.07<\/span><\/span><\/span><\/span><\/span>\u3002\u8fd9\u8fdb\u4e00\u6b65\u9650\u5236\u4e86\u201c\u4efb\u4f55\u957f\u5ea6\u90fd\u6700\u597d\u201d\u7684\u8868\u8ff0\u3002\u5bf9\u5206\u652f\u3001PR \u4e0e CI \u6784\u6210\u7684\u5173\u7cfb\u7f51\u7edc，\u4fee\u6539\u4e00\u4e2a\u5b9e\u4f53\u53ef\u80fd\u4f7f\u591a\u4e2a\u4f9d\u8d56\u72b6\u6001\u5931\u6548；\u7ef4\u62a4\u5c40\u90e8\u5b57\u6bb5\u4e0e\u7ef4\u62a4\u5173\u7cfb\u4e00\u81f4\u6027\u4e4b\u95f4，\u4ecd\u6709\u660e\u663e\u8ddd\u79bb\u3002<\/p>\n

Rocky\u8ba4\u4e3a，\u590d\u6742\u6d41\u7a0b\u7684\u72b6\u6001\u8bbe\u8ba1，\u5e94\u660e\u786e\u54ea\u4e9b\u5b57\u6bb5\u7531\u5de5\u5177\u4e8b\u5b9e\u76f4\u63a5\u66f4\u65b0，\u54ea\u4e9b\u5b57\u6bb5\u662f\u6a21\u578b\u63a8\u65ad，\u4ee5\u53ca\u54ea\u4e9b\u5173\u8054\u5fc5\u987b\u4e00\u8d77\u8fc1\u79fb\u3002 \u8fd9\u4f1a\u5f71\u54cd\u8865\u4e01\u6821\u9a8c\u80fd\u5426\u8bc6\u522b\u201c\u5b57\u6bb5\u7c7b\u578b\u6b63\u786e，\u4f46\u4f9d\u8d56\u5173\u7cfb\u5df2\u7ecf\u77db\u76fe\u201d\u7684\u60c5\u51b5\u3002<\/p>\n

\u5341\u4e00\u3001\u5f00\u653e\u6743\u91cd\u6a21\u578b：\u8282\u7701\u4e0a\u4e0b\u6587，\u4e0d\u4ee3\u8868\u81ea\u52a8\u83b7\u5f97\u53ef\u9760\u6267\u884c\u80fd\u529b<\/h3>\n

\u8bba\u6587\u8868 7｜Gemma-4-31B-it \u7684\u4ed3\u5e93\u6269\u5c55\u3002\u8be5\u8868\u5c06\u5206\u6570\u5199\u6210\u5c0f\u6570，\u5c06\u5206\u6570\u6807\u51c6\u5dee\u5199\u6210\u767e\u5206\u6570；\u4f8b\u5982 0.42 \u00b1 4.1% \u5bf9\u5e94\u7ea6 0.42 \u00b1 0.041，\u8fd9\u91cc\u4fdd\u7559\u8bba\u6587\u8bb0\u6cd5\u3002<\/p>\n\nHorizonRuntimeScore \u00b1 SDAvg Prompt \u00b1 SDTotal Tokens \u00b1 SD<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
10 Steps<\/td>\nPrompt<\/td>\n0.90 \u00b1 3.1%<\/td>\n3,145 \u00b1 242<\/td>\n9,092 \u00b1 1,231<\/td>\n<\/tr>\n
10 Steps<\/td>\nMemory<\/td>\n0.85 \u00b1 4.2%<\/td>\n2,611 \u00b1 138<\/td>\n7,330 \u00b1 838<\/td>\n<\/tr>\n
10 Steps<\/td>\nStateful<\/td>\n0.90 \u00b1 2.8%<\/td>\n3,191 \u00b1 149<\/td>\n9,144 \u00b1 518<\/td>\n<\/tr>\n
10 Steps<\/td>\nSKILL.state<\/td>\n0.98 \u00b1 1.5%<\/td>\n2,116 \u00b1 212<\/td>\n6,814 \u00b1 875<\/td>\n<\/tr>\n
25 Steps<\/td>\nPrompt<\/td>\n0.64 \u00b1 5.4%<\/td>\n5,720 \u00b1 182<\/td>\n41,697 \u00b1 1,610<\/td>\n<\/tr>\n
25 Steps<\/td>\nMemory<\/td>\n0.72 \u00b1 4.8%<\/td>\n3,990 \u00b1 362<\/td>\n28,933 \u00b1 412<\/td>\n<\/tr>\n
25 Steps<\/td>\nStateful<\/td>\n0.76 \u00b1 4.1%<\/td>\n5,714 \u00b1 282<\/td>\n39,314 \u00b1 585<\/td>\n<\/tr>\n
25 Steps<\/td>\nSKILL.state<\/td>\n0.84 \u00b1 3.6%<\/td>\n2,080 \u00b1 114<\/td>\n16,302 \u00b1 2,190<\/td>\n<\/tr>\n
50 Steps<\/td>\nPrompt<\/td>\n0.31 \u00b1 6.2%<\/td>\n10,809 \u00b1 352<\/td>\n151,845 \u00b1 2,150<\/td>\n<\/tr>\n
50 Steps<\/td>\nMemory<\/td>\n0.41 \u00b1 5.8%<\/td>\n7,217 \u00b1 316<\/td>\n114,113 \u00b1 1,620<\/td>\n<\/tr>\n
50 Steps<\/td>\nStateful<\/td>\n0.55 \u00b1 5.1%<\/td>\n11,083 \u00b1 262<\/td>\n155,164 \u00b1 2,210<\/td>\n<\/tr>\n
50 Steps<\/td>\nSKILL.state<\/td>\n0.68 \u00b1 3.9%<\/td>\n2,113 \u00b1 176<\/td>\n33,762 \u00b1 1,385<\/td>\n<\/tr>\n
100 Steps<\/td>\nPrompt<\/td>\n0.21 \u00b1 4.7%<\/td>\n27,686 \u00b1 412<\/td>\n923,164 \u00b1 12,400<\/td>\n<\/tr>\n
100 Steps<\/td>\nMemory<\/td>\n0.24 \u00b1 4.2%<\/td>\n18,537 \u00b1 229<\/td>\n701,954 \u00b1 9,850<\/td>\n<\/tr>\n
100 Steps<\/td>\nStateful<\/td>\n0.42 \u00b1 4.5%<\/td>\n20,210 \u00b1 822<\/td>\n557,968 \u00b1 8,100<\/td>\n<\/tr>\n
100 Steps<\/td>\nSKILL.state<\/td>\n0.42 \u00b1 4.1%<\/td>\n2,105 \u00b1 216<\/td>\n65,480 \u00b1 1,258<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

100 \u6b65\u65f6，Gemma \u7684 SKILL.state \u4e0e Stateful \u5206\u6570\u540c\u4e3a 0.42，\u7d2f\u8ba1 Token \u5206\u522b\u4e3a 65,480 \u4e0e 557,968\u3002\u8fd9\u4e2a\u7ec4\u5408\u6700\u76f4\u63a5\u652f\u6301\u6548\u7387\u6539\u5584，\u4e0d\u80fd\u5199\u6210\u51c6\u786e\u7387\u4e5f\u5f97\u5230\u63d0\u5347\u3002SKILL.state \u4ece 10 \u6b65\u7684 0.98 \u964d\u4e3a 100 \u6b65\u7684 0.42，\u8868\u660e\u957f\u7a0b\u53ef\u9760\u6027\u95ee\u9898\u5e76\u672a\u89e3\u51b3\u3002<\/p>\n

\u8bba\u6587\u8868 8｜Qwen-3-8B-it \u7684\u4ed3\u5e93\u6269\u5c55，\u5206\u6570\u6807\u51c6\u5dee\u5355\u4f4d\u540c\u8868 7\u3002<\/p>\n\nHorizonRuntimeScore \u00b1 SDAvg Prompt \u00b1 SDTotal Tokens \u00b1 SD<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
10 Steps<\/td>\nPrompt<\/td>\n0.84 \u00b1 3.8%<\/td>\n3,150 \u00b1 245<\/td>\n9,150 \u00b1 1,250<\/td>\n<\/tr>\n
10 Steps<\/td>\nMemory<\/td>\n0.80 \u00b1 4.5%<\/td>\n2,640 \u00b1 145<\/td>\n7,420 \u00b1 860<\/td>\n<\/tr>\n
10 Steps<\/td>\nStateful<\/td>\n0.84 \u00b1 3.4%<\/td>\n3,210 \u00b1 155<\/td>\n9,210 \u00b1 540<\/td>\n<\/tr>\n
10 Steps<\/td>\nSKILL.state<\/td>\n0.94 \u00b1 2.1%<\/td>\n2,120 \u00b1 215<\/td>\n6,920 \u00b1 890<\/td>\n<\/tr>\n
25 Steps<\/td>\nPrompt<\/td>\n0.54 \u00b1 6.1%<\/td>\n5,790 \u00b1 195<\/td>\n42,450 \u00b1 1,680<\/td>\n<\/tr>\n
25 Steps<\/td>\nMemory<\/td>\n0.62 \u00b1 5.4%<\/td>\n4,050 \u00b1 375<\/td>\n29,640 \u00b1 440<\/td>\n<\/tr>\n
25 Steps<\/td>\nStateful<\/td>\n0.66 \u00b1 4.8%<\/td>\n5,780 \u00b1 295<\/td>\n39,950 \u00b1 610<\/td>\n<\/tr>\n
25 Steps<\/td>\nSKILL.state<\/td>\n0.76 \u00b1 4.2%<\/td>\n2,088 \u00b1 120<\/td>\n16,680 \u00b1 2,240<\/td>\n<\/tr>\n
50 Steps<\/td>\nPrompt<\/td>\n0.24 \u00b1 6.5%<\/td>\n10,950 \u00b1 365<\/td>\n154,200 \u00b1 2,280<\/td>\n<\/tr>\n
50 Steps<\/td>\nMemory<\/td>\n0.33 \u00b1 6.1%<\/td>\n7,320 \u00b1 330<\/td>\n116,400 \u00b1 1,710<\/td>\n<\/tr>\n
50 Steps<\/td>\nStateful<\/td>\n0.44 \u00b1 5.7%<\/td>\n11,210 \u00b1 280<\/td>\n158,300 \u00b1 2,340<\/td>\n<\/tr>\n
50 Steps<\/td>\nSKILL.state<\/td>\n0.58 \u00b1 4.5%<\/td>\n2,118 \u00b1 185<\/td>\n34,510 \u00b1 1,420<\/td>\n<\/tr>\n
100 Steps<\/td>\nPrompt<\/td>\n0.15 \u00b1 5.1%<\/td>\n28,150 \u00b1 430<\/td>\n941,500 \u00b1 12,800<\/td>\n<\/tr>\n
100 Steps<\/td>\nMemory<\/td>\n0.18 \u00b1 4.6%<\/td>\n18,840 \u00b1 245<\/td>\n718,200 \u00b1 10,200<\/td>\n<\/tr>\n
100 Steps<\/td>\nStateful<\/td>\n0.31 \u00b1 4.9%<\/td>\n20,580 \u00b1 850<\/td>\n569,400 \u00b1 8,450<\/td>\n<\/tr>\n
100 Steps<\/td>\nSKILL.state<\/td>\n0.34 \u00b1 4.6%<\/td>\n2,110 \u00b1 220<\/td>\n66,850 \u00b1 1,310<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Qwen \u5728 100 \u6b65\u65f6，SKILL.state \u5206\u6570\u4e3a 0.34，\u9ad8\u4e8e Stateful \u7684 0.31，\u7d2f\u8ba1 Token \u5219\u4ece 569,400 \u964d\u81f3 66,850；\u4f46 0.34 \u7684\u7edd\u5bf9\u52a8\u4f5c\u51c6\u786e\u7387\u4ecd\u7136\u5f88\u4f4e\u3002\u4e00\u4e2a\u4fbf\u5b9c\u4f46\u7ecf\u5e38\u9519\u8bef\u7684\u6267\u884c\u5668，\u53ef\u80fd\u5c06\u6210\u672c\u8f6c\u79fb\u5230\u4eba\u5de5\u7ea0\u9519\u3001\u56de\u6eda\u548c\u4e1a\u52a1\u635f\u5931\u4e0a\u3002<\/p>\n

\u4f5c\u8005\u5bf9 Gemma \u7684\u5931\u8d25\u65e5\u5fd7\u7ed9\u51fa\u4e09\u7c7b\u9519\u8bef\u5360\u6bd4：\u8fc7\u65e9\u8986\u76d6\u6216\u5220\u9664\u72b6\u6001 68%，Schema \u7406\u89e3\u6216\u7c7b\u578b\u8f6c\u6362\u95ee\u9898 20%，JSON \u8bed\u6cd5\u683c\u5f0f\u95ee\u9898 12%\u3002\u8fd9\u63d0\u793a\u8865\u4e01\u751f\u6210\u548c\u5408\u5e76\u7ea6\u5b9a\u503c\u5f97\u4f18\u5148\u6539\u8fdb\u3002\u4e0d\u8fc7\u8bba\u6587\u6ca1\u6709\u5728\u8be5\u6bb5\u63d0\u4f9b\u9519\u8bef\u6837\u672c\u6570\u3001\u5206\u7c7b\u4e00\u81f4\u6027\u8bc4\u4f30\u6216\u6d88\u9664\u8fd9\u4e9b\u9519\u8bef\u540e\u7684\u5bf9\u7167，\u56e0\u6b64\u4e0d\u80fd\u4ec5\u51ed\u5206\u7c7b\u5c31\u8bc1\u660e\u201c\u6a21\u578b\u63a8\u7406\u80fd\u529b\u5b8c\u5168\u4e0d\u662f\u74f6\u9888\u201d\u3002<\/p>\n

\u53d7\u7ea6\u675f\u89e3\u7801\u53ef\u4ee5\u51cf\u5c11\u8bed\u6cd5\u9519\u8bef，\u4e5f\u53ef\u5e2e\u52a9\u6ee1\u8db3\u7ed3\u6784\u7ea6\u675f；\u5b83\u65e0\u6cd5\u5355\u72ec\u4fdd\u8bc1\u7269\u54c1\u4f4d\u7f6e\u6b63\u786e\u3001\u5220\u9664\u65f6\u673a\u6b63\u786e\u6216\u67d0\u4e2a\u4e1a\u52a1\u52a8\u4f5c\u5408\u6cd5\u3002\u5bf9\u4e8e\u5360\u6bd4\u6700\u5927\u7684\u8986\u76d6\/\u5220\u9664\u9519\u8bef，\u8fd8\u9700\u8981\u6838\u5b9e\u6a21\u578b\u751f\u6210\u7684\u662f\u8865\u4e01\u8fd8\u662f\u5b8c\u6574\u66ff\u6362，\u4ee5\u53ca\u5d4c\u5957\u5b57\u5178\u7684\u5408\u5e76\u8bed\u4e49\u3002<\/p>\n

\u5341\u4e8c\u3001\u8f6f\u4ef6\u73af\u5883\u4e2d\u7684\u566a\u58f0\u4e0e\u6062\u590d：\u4fdd\u7559\u5177\u4f53\u6210\u529f，\u4e5f\u4fdd\u7559\u5931\u8d25<\/h3>\n

\u8bba\u6587\u8868 9｜\u6a21\u62df\u8f6f\u4ef6\u4ed3\u5e93\u566a\u58f0\u9c81\u68d2\u6027，50 \u6b65，Gemini-3-Flash；\u566a\u58f0\u4e3a\u4e0d\u6539\u53d8\u73af\u5883\u72b6\u6001\u7684\u65e0\u5173 syslog\u3002<\/p>\n\nNoise LevelRuntimeScore<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
0 Events (Baseline)<\/td>\nPrompt<\/td>\n0.76<\/td>\n<\/tr>\n
0 Events (Baseline)<\/td>\nMemory<\/td>\n0.85<\/td>\n<\/tr>\n
0 Events (Baseline)<\/td>\nStateful<\/td>\n0.88<\/td>\n<\/tr>\n
0 Events (Baseline)<\/td>\nSKILL.state<\/td>\n0.90<\/td>\n<\/tr>\n
5 Events (Low)<\/td>\nPrompt<\/td>\n0.62<\/td>\n<\/tr>\n
5 Events (Low)<\/td>\nMemory<\/td>\n0.85<\/td>\n<\/tr>\n
5 Events (Low)<\/td>\nStateful<\/td>\n0.86<\/td>\n<\/tr>\n
5 Events (Low)<\/td>\nSKILL.state<\/td>\n0.88<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nPrompt<\/td>\n0.48<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nMemory<\/td>\n0.83<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nStateful<\/td>\n0.85<\/td>\n<\/tr>\n
20 Events (Medium)<\/td>\nSKILL.state<\/td>\n0.86<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nPrompt<\/td>\n0.11<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nMemory<\/td>\n0.74<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nStateful<\/td>\n0.78<\/td>\n<\/tr>\n
50 Events (High)<\/td>\nSKILL.state<\/td>\n0.80<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4ece\u65e0\u566a\u58f0\u5230\u6bcf\u6b65 50 \u6761\u566a\u58f0，SKILL.state \u4ece 0.90 \u964d\u5230 0.80，Prompt \u4ece 0.76 \u964d\u5230 0.11\u3002\u4e24\u8005\u5dee\u8ddd\u5728\u9ad8\u566a\u58f0\u4e0b\u6269\u5927，\u4f46 SKILL.state \u4ecd\u6709\u4e0b\u964d\u3002\u6b63\u786e\u8868\u8ff0\u5e94\u662f\u66f4\u7a33\u5065，\u800c\u975e\u5bf9\u566a\u58f0\u4e0d\u654f\u611f\u3002<\/p>\n

\u8bba\u6587\u8868 10｜\u6a21\u62df\u8f6f\u4ef6\u4ed3\u5e93\u7684\u72b6\u6001\u6062\u590d，\u6bd4\u8f83\u6536\u5230\u975e\u7ed3\u6784\u5316\u544a\u8b66\u540e\u7684\u6062\u590d\u6ede\u540e\u3002<\/p>\n\nScenarioRuntimeSuccessRecovery Steps<\/tr>\n\n\n\n\n\n\n\n\n\n\n
A: Force Push<\/td>\nPrompt<\/td>\nYes<\/td>\n12<\/td>\n<\/tr>\n
A: Force Push<\/td>\nMemory<\/td>\nYes<\/td>\n8<\/td>\n<\/tr>\n
A: Force Push<\/td>\nStateful<\/td>\nYes<\/td>\n10<\/td>\n<\/tr>\n
A: Force Push<\/td>\nSKILL.state<\/td>\nYes<\/td>\n0<\/td>\n<\/tr>\n
B: Flaky CI Test<\/td>\nPrompt<\/td>\nYes<\/td>\n14<\/td>\n<\/tr>\n
B: Flaky CI Test<\/td>\nMemory<\/td>\nYes<\/td>\n9<\/td>\n<\/tr>\n
B: Flaky CI Test<\/td>\nStateful<\/td>\nYes<\/td>\n11<\/td>\n<\/tr>\n
B: Flaky CI Test<\/td>\nSKILL.state<\/td>\nYes<\/td>\n0<\/td>\n<\/tr>\n
C: PR Closed<\/td>\nAll Runtimes<\/td>\nNo<\/td>\nN\/A<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Force Push \u548c Flaky CI Test \u4e0b，SKILL.state \u62a5\u544a\u96f6\u989d\u5916\u6062\u590d\u6b65\u9aa4，\u5176\u4ed6\u8fd0\u884c\u65f6\u9700\u8981 8\u201414 \u6b65\u3002PR Closed \u573a\u666f\u5219\u5168\u90e8\u5931\u8d25\u3002\u8fd9\u4e0e\u4ed3\u5e93 Canceled Order \u7684\u5171\u540c\u5931\u8d25\u5f62\u6210\u547c\u5e94：\u4e00\u90e8\u5206\u5f02\u5e38\u53ef\u4ee5\u901a\u8fc7\u8986\u76d6\u65e7\u4e8b\u5b9e\u5904\u7406，\u53e6\u4e00\u4e9b\u5f02\u5e38\u6d89\u53ca\u4efb\u52a1\u662f\u5426\u4ecd\u7136\u53ef\u6267\u884c\u3001\u76ee\u6807\u662f\u5426\u9700\u8981\u91cd\u65b0\u89e3\u91ca，\u5355\u7eaf\u66f4\u65b0\u72b6\u6001\u672a\u5fc5\u8db3\u591f\u3002<\/p>\n

\u4ece\u8fd9\u4e9b\u7ed3\u679c\u5f97\u5230\u7684\u5de5\u7a0b\u542f\u53d1，\u662f\u4e3a\u5f02\u5e38\u8bbe\u7f6e\u660e\u786e\u7684\u5904\u7406\u5206\u652f：\u7ee7\u7eed\u6267\u884c\u3001\u91cd\u65b0\u89c2\u5bdf\u3001\u4fee\u590d\u524d\u7f6e\u6761\u4ef6\u3001\u8f6c\u5165\u4eba\u5de5\u5904\u7406\u6216\u7ec8\u6b62\u4efb\u52a1\u3002\u8fd9\u4e2a\u6269\u5c55\u9700\u8981\u5355\u72ec\u8bc4\u6d4b，\u4e0d\u80fd\u628a\u8bba\u6587\u6ca1\u6709\u63d0\u4f9b\u7684\u5f02\u5e38\u7b56\u7565\u7b97\u4f5c\u65e2\u6709\u80fd\u529b\u3002<\/p>\n

\u5341\u4e09\u3001\u9884\u7b97\u968f\u957f\u5ea6\u6269\u5c55：\u9519\u8bef\u4fdd\u7559\u7b56\u7565\u4f1a\u5728\u957f\u4efb\u52a1\u91cc\u653e\u5927<\/h3>\n

\u8bba\u6587\u8868 11｜Gemini-3-Flash \u4ed3\u5e93\u9884\u7b97\u5339\u914d\u7684\u591a\u957f\u5ea6\u7ed3\u679c，\u5747\u503c \u00b1 \u6807\u51c6\u5dee，5 \u4e2a\u79cd\u5b50\u3002\u7ea6 1,800 \u9884\u7b97\u7684\u5355\u4f4d\u95ee\u9898\u4e0e\u8868 5 \u76f8\u540c\u3002<\/p>\n\nHorizon (T)SKILL.stateSummary-cappedTruncated (Window)ReAct + LLMLinguaReAct (Full)<\/tr>\n\n\n\n\n\n
10<\/td>\n1.00 \u00b1 0.00<\/td>\n0.92 \u00b1 0.03<\/td>\n0.90 \u00b1 0.04<\/td>\n0.88 \u00b1 0.04<\/td>\n0.90 \u00b1 0.02<\/td>\n<\/tr>\n
25<\/td>\n1.00 \u00b1 0.00<\/td>\n0.76 \u00b1 0.04<\/td>\n0.62 \u00b1 0.05<\/td>\n0.60 \u00b1 0.05<\/td>\n0.92 \u00b1 0.02<\/td>\n<\/tr>\n
50<\/td>\n0.96 \u00b1 0.01<\/td>\n0.64 \u00b1 0.05<\/td>\n0.35 \u00b1 0.06<\/td>\n0.38 \u00b1 0.06<\/td>\n0.88 \u00b1 0.04<\/td>\n<\/tr>\n
100<\/td>\n0.94 \u00b1 0.01<\/td>\n0.52 \u00b1 0.05<\/td>\n0.18 \u00b1 0.05<\/td>\n0.22 \u00b1 0.05<\/td>\n0.84 \u00b1 0.07<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

10 \u6b65\u65f6，\u5404\u538b\u7f29\u7b56\u7565\u5206\u6570\u5904\u4e8e 0.88\u20140.92，SKILL.state \u4e3a 1.00；100 \u6b65\u65f6，\u6ed1\u52a8\u7a97\u53e3\u4e0e LLMLingua \u5206\u522b\u964d\u81f3 0.18 \u548c 0.22，SKILL.state \u4ecd\u4e3a 0.94\u3002\u968f\u7740\u8ddd\u79bb\u62c9\u957f，\u65e9\u671f\u5b9e\u4f53\u6620\u5c04\u88ab\u4e22\u5f03\u7684\u540e\u679c\u66f4\u5bb9\u6613\u663e\u73b0\u3002<\/p>\n

\u81ea\u7136\u8bed\u8a00\u6458\u8981\u5728 100 \u6b65\u8fbe\u5230 0.52，\u660e\u663e\u597d\u4e8e\u53e6\u5916\u4e24\u79cd\u538b\u7f29\u63a7\u5236，\u4e5f\u8bf4\u660e\u538b\u7f29\u65b9\u6cd5\u4e0d\u80fd\u4e00\u6982\u800c\u8bba\u3002\u503c\u5f97\u7ee7\u7eed\u7814\u7a76\u7684\u662f，\u600e\u6837\u628a\u7ed3\u6784\u5316\u7ea6\u675f\u3001\u6458\u8981\u548c\u6309\u9700\u8bc1\u636e\u8bfb\u53d6\u7ed3\u5408\u8d77\u6765，\u5e76\u5728\u540c\u4e00\u8ba1\u6570\u53e3\u5f84\u4e0b\u6bd4\u8f83\u6b63\u786e\u7387\u3001\u8c03\u7528\u6570\u3001\u8f93\u5165\u8f93\u51fa Token \u4e0e\u7ef4\u62a4\u5f00\u9500\u3002<\/p>\n

\u5341\u56db\u3001\u8fb9\u754c\u4e0e\u53ef\u590d\u73b0\u6027：\u5145\u5206\u72b6\u6001\u662f\u4e00\u9879\u5047\u8bbe，\u9700\u8981\u6301\u7eed\u68c0\u9a8c<\/h3>\n

14.1 \u4e09\u79cd\u60c5\u51b5\u4e0b，\u4e22\u5f03\u5386\u53f2\u53ef\u80fd\u4e22\u6389\u672a\u6765\u5fc5\u9700\u7684\u4fe1\u606f<\/h4>\n

\u8bba\u6587\u660e\u786e\u627f\u8ba4，\u72b6\u6001\u4f5c\u4e3a\u672a\u6765\u6267\u884c\u7684\u5145\u5206\u7edf\u8ba1\u91cf\u662f\u65b9\u6cd5\u7684\u524d\u63d0\u3002\u7b2c\u4e00，\u4efb\u52a1\u5f00\u59cb\u65f6\u65e0\u6cd5\u9884\u5148\u786e\u5b9a Schema，\u72b6\u6001\u7ed3\u6784\u9700\u8981\u5728\u6267\u884c\u4e2d\u53d1\u73b0\u3002\u7b2c\u4e8c，\u67d0\u6761\u65e7\u89c2\u5bdf\u5f53\u65f6\u770b\u4f3c\u65e0\u5173，\u540e\u7eed\u624d\u66b4\u9732\u91cd\u8981\u6027，\u4f46\u6b64\u524d\u6ca1\u6709\u88ab\u4fdd\u5b58\u3002\u7b2c\u4e09，\u4efb\u52a1\u76ee\u6807\u672c\u8eab\u5c31\u662f\u5ba1\u8ba1\u5386\u53f2\u3001\u8ffd\u8e2a\u6765\u6e90\u6216\u89e3\u91ca\u8fc7\u53bb\u52a8\u4f5c，\u5386\u53f2\u4e0d\u80fd\u88ab\u5f53\u4f5c\u6267\u884c\u8d1f\u62c5\u76f4\u63a5\u4e22\u5f03\u3002<\/p>\n

\u5145\u5206\u6027\u5e76\u975e\u201cJSON \u91cc\u6709\u51e0\u4e2a\u5408\u9002\u5b57\u6bb5\u201d\u5c31\u80fd\u5f97\u5230\u7684\u6570\u5b66\u4fdd\u8bc1\u3002\u5b83\u6d89\u53ca\u4fe1\u606f\u9009\u62e9\u662f\u5426\u6b63\u786e\u3001\u4efb\u52a1\u662f\u5426\u53d8\u5316\u3001\u72b6\u6001\u662f\u5426\u4fdd\u5b58\u4e0d\u786e\u5b9a\u6027，\u4ee5\u53ca\u65b0\u76ee\u6807\u662f\u5426\u9700\u8981\u91cd\u65b0\u8bbf\u95ee\u8fc7\u53bb\u8bc1\u636e\u3002\u8bba\u6587\u6ca1\u6709\u8bc1\u660e\u4efb\u610f\u9886\u57df\u90fd\u5b58\u5728\u8db3\u591f\u5c0f\u7684\u5145\u5206\u72b6\u6001\u3002<\/p>\n

14.2 \u6821\u9a8c\u53ef\u4ee5\u4fdd\u62a4\u683c\u5f0f\u8fb9\u754c，\u8bed\u4e49\u6b63\u786e\u6027\u9700\u8981\u989d\u5916\u673a\u5236<\/h4>\n

\u8fd0\u884c\u65f6\u8d1f\u8d23 Schema \u4e0e\u8865\u4e01\u6821\u9a8c，\u4f7f\u65e0\u6548\u8f93\u51fa\u4e0d\u4f1a\u76f4\u63a5\u5199\u5165\u6301\u4e45\u72b6\u6001，\u8fd9\u662f\u6e05\u695a\u7684\u8d23\u4efb\u5212\u5206\u3002\u4f46\u7c7b\u578b\u5408\u6cd5\u7684\u9519\u8bef\u4e8b\u5b9e\u4ecd\u53ef\u80fd\u901a\u8fc7\u7ed3\u6784\u68c0\u67e5\u3002\u8ba2\u5355\u6807\u8bc6\u662f\u5b57\u7b26\u4e32，\u5e76\u4e0d\u80fd\u8bc1\u660e\u6a21\u578b\u9009\u4e2d\u4e86\u6b63\u786e\u8ba2\u5355\u3002<\/p>\n

\u751f\u4ea7\u73af\u5883\u9700\u8981\u5c06\u6a21\u578b\u63d0\u8bae\u4e0e\u53ef\u6838\u9a8c\u4e8b\u5b9e\u8fde\u63a5\u8d77\u6765：\u91cd\u8981\u5b57\u6bb5\u9644\u5e26\u5de5\u5177\u8bc1\u636e\u6216\u7248\u672c\u53f7；\u6d89\u53ca\u5916\u90e8\u526f\u4f5c\u7528\u65f6\u68c0\u67e5\u524d\u7f6e\u6761\u4ef6；\u5931\u8d25\u540e\u91cd\u65b0\u540c\u6b65\u6743\u5a01\u6570\u636e\u5e93\u3002\u5ba1\u8ba1\u65e5\u5fd7\u53ef\u4ee5\u72ec\u7acb\u4fdd\u5b58\u5de5\u5177\u52a8\u4f5c\u3001\u72b6\u6001\u5dee\u5f02\u548c\u7ed3\u679c，\u4e14\u65e0\u9700\u5728\u6bcf\u4e00\u6b65\u5168\u91cf\u6ce8\u5165 Prompt\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u53ef\u501f\u9274\u8bba\u6587\u7684\u4fe1\u606f\u9694\u79bb\u601d\u60f3，\u4f46\u5c5e\u4e8e\u6269\u5c55\u65b9\u6848\u3002<\/p>\n

\u591a Agent \u573a\u666f\u8fdb\u4e00\u6b65\u5f15\u5165\u5e76\u53d1\u5199\u3002\u8bba\u6587\u5f53\u524d\u5b9e\u73b0\u805a\u7126\u5355 Agent，\u6ca1\u6709\u9a8c\u8bc1\u51b2\u7a81\u68c0\u6d4b\u3001\u7248\u672c\u6bd4\u8f83\u3001\u5199\u5165\u987a\u5e8f\u6216\u591a\u8282\u70b9\u4e00\u81f4\u6027\u3002\u5171\u4eab\u4e00\u4e2a JSON \u6587\u4ef6\u5e76\u4e0d\u81ea\u52a8\u6210\u4e3a\u53ef\u9760\u534f\u8c03\u534f\u8bae\u3002<\/p>\n

14.3 \u590d\u73b0\u5e94\u4f18\u5148\u6f84\u6e05\u7684\u8bc1\u636e\u7f3a\u53e3<\/h4>\n

\u8bba\u6587\u516c\u5f00\u4e86\u8fd0\u884c\u65f6\u516c\u5f0f\u3001\u7b97\u6cd5\u3001\u56db\u7c7b\u4e3b\u8981 Prompt \u6a21\u677f\u3001\u73af\u5883\u8bf4\u660e\u4e0e\u5b9e\u9a8c\u8868\u683c，\u8db3\u4ee5\u7406\u89e3\u8bbe\u8ba1\u65b9\u5411\u3002arXiv \u6e90\u5305\u63d0\u4f9b\u8bba\u6587 LaTeX \u548c\u56fe\u5f62\u8d44\u4ea7；\u5728\u672c\u6587\u6838\u5bf9\u7684\u8bba\u6587\u4e0e\u5143\u6570\u636e\u4e2d，\u672a\u627e\u5230\u4f5c\u8005\u7ed9\u51fa\u7684\u5b98\u65b9\u5b9e\u73b0\u4ed3\u5e93\u94fe\u63a5\u3002\u4e0d\u80fd\u628a\u8bba\u6587\u6e90\u5305\u5f53\u6210\u53ef\u76f4\u63a5\u8fd0\u884c\u7684 Agent \u5b9e\u73b0，\u4e5f\u4e0d\u80fd\u636e\u6b64\u65ad\u8a00\u5176\u4ed6\u6e20\u9053\u7edd\u5bf9\u4e0d\u5b58\u5728\u4ee3\u7801\u3002<\/p>\n

\u590d\u73b0\u8005\u9996\u5148\u9700\u8981\u786e\u8ba4：\u9884\u7b97\u6309\u5b57\u7b26\u8fd8\u662f Token \u8ba1\u7b97；InterCode \u7684\u5c0f\u6570\u6210\u529f\u7387\u91c7\u7528\u4ec0\u4e48\u91cd\u590d\u4e0e\u805a\u5408\u53e3\u5f84；\u516c\u5f00\u57fa\u51c6\u7684\u5177\u4f53\u7248\u672c\u3001\u4efb\u52a1\u5b50\u96c6\u4e0e\u505c\u6b62\u6761\u4ef6；\u6bcf\u79cd\u8fd0\u884c\u65f6\u7684\u8c03\u7528\u3001\u8f93\u51fa\u3001\u91cd\u8bd5\u4e0e\u538b\u7f29\u5f00\u9500\u5982\u4f55\u7edf\u8ba1；\u5d4c\u5957\u8865\u4e01\u548c null \u5220\u9664\u7684\u7cbe\u786e\u5b9a\u4e49；\u4ee5\u53ca\u8f6f\u4ef6\u4ed3\u5e93\u5b9e\u9645\u91c7\u7528\u54ea\u5957\u52a8\u4f5c\u63a5\u53e3\u3002<\/p>\n

\u672c\u6587\u4fdd\u7559\u8bba\u6587\u62a5\u544a\u7684\u5168\u90e8\u5b9e\u9a8c\u8868\u683c，\u5e76\u6838\u5bf9\u6570\u503c\u4e0e\u8ba1\u7b97\u5173\u7cfb，\u6ca1\u6709\u91cd\u8dd1\u6a21\u578b\u57fa\u51c6，\u4e5f\u6ca1\u6709\u628a\u4f5c\u8005\u7684\u7edf\u8ba1\u663e\u8457\u6027\u58f0\u660e\u5f53\u4f5c\u72ec\u7acb\u590d\u73b0\u7ed3\u679c\u3002\u516c\u5f00\u6570\u636e\u91cc\u7684\u6539\u5584\u503c\u5f97\u8fdb\u4e00\u6b65\u7814\u7a76，\u53e3\u5f84\u4e0d\u5b8c\u6574\u7684\u90e8\u5206\u5219\u9700\u8981\u8865\u5145\u8fd0\u884c\u8bc1\u636e\u3002<\/p>\n

\u5341\u4e94\u3001Rocky\u7684\u7814\u7a76\u5224\u65ad：\u4ece\u72b6\u6001\u8868\u793a\u7ee7\u7eed\u8d70\u5411\u72b6\u6001\u6cbb\u7406<\/h3>\n

15.1 \u9996\u5148\u843d\u5730\u5728\u72b6\u6001\u8fb9\u754c\u6e05\u6670\u7684\u6d41\u7a0b<\/h4>\n

\u5e93\u5b58\u5904\u7406\u3001\u5de5\u5355\u6d41\u8f6c\u3001\u53d7\u7ea6\u675f\u5ba2\u670d\u548c\u90e8\u7f72\u6d41\u7a0b，\u901a\u5e38\u5177\u6709\u53ef\u8bc6\u522b\u5b9e\u4f53\u3001\u76f8\u5bf9\u7a33\u5b9a\u7684\u52a8\u4f5c\u7a7a\u95f4\u4e0e\u53ef\u6821\u9a8c\u5de5\u5177\u8fd4\u56de，\u9002\u5408\u5148\u505a\u72b6\u6001\u5316\u8bd5\u9a8c\u3002\u5f00\u653e\u7814\u7a76\u3001\u9700\u6c42\u63a2\u7d22\u548c\u957f\u6587\u521b\u4f5c\u4e2d，\u672a\u6765\u76f8\u5173\u4fe1\u606f\u66f4\u96be\u63d0\u524d\u5224\u65ad，\u8fc7\u65e9\u4e22\u5f03\u5386\u53f2\u53ef\u80fd\u635f\u5931\u4ef7\u503c\u3002<\/p>\n

\u5bf9 Agent \u56e2\u961f，\u8d77\u70b9\u53ef\u4ee5\u662f\u4e00\u4e2a\u8f83\u5c0f\u4f46\u5b8c\u6574\u7684\u4e1a\u52a1\u6d41\u7a0b：\u5b9a\u4e49\u6700\u5c0f\u6267\u884c\u72b6\u6001，\u660e\u786e\u6bcf\u4e2a\u5b57\u6bb5\u7684\u8bc1\u636e\u6765\u6e90\u548c\u66f4\u65b0\u89c4\u5219，\u518d\u540c\u73b0\u6709\u8fd0\u884c\u65f6\u505a\u7b49\u4efb\u52a1\u3001\u7b49\u5de5\u5177\u3001\u7edf\u4e00\u8ba1\u6570\u53e3\u5f84\u7684\u5bf9\u7167\u3002\u4e0d\u8981\u4ece\u201c\u6240\u6709\u5386\u53f2\u90fd\u5e94\u5220\u9664\u201d\u8fd9\u4e2a\u53e3\u53f7\u5f00\u59cb\u3002<\/p>\n

15.2 \u5c06\u72b6\u6001\u5bb9\u91cf\u3001\u5145\u5206\u6027\u4e0e\u66f4\u65b0\u6b63\u786e\u6027\u4e00\u8d77\u8bc4\u4f30<\/h4>\n

\u672a\u6765\u7814\u7a76\u53ef\u4ee5\u56f4\u7ed5\u4e09\u4e2a\u76f8\u4e92\u7275\u5236\u7684\u76ee\u6807\u5c55\u5f00：\u72b6\u6001\u8d8a\u77ed，\u6210\u672c\u8d8a\u5bb9\u6613\u63a7\u5236；\u72b6\u6001\u8d8a\u5b8c\u6574，\u672a\u6765\u4fe1\u606f\u7f3a\u5931\u98ce\u9669\u8d8a\u4f4e；\u66f4\u65b0\u8d8a\u590d\u6742，\u6a21\u578b\u8d8a\u5bb9\u6613\u51fa\u9519\u3002\u53ea\u62a5\u544a\u5e73\u5747 Prompt \u957f\u5ea6，\u65e0\u6cd5\u53cd\u6620\u8fd9\u4e9b\u5173\u7cfb\u3002<\/p>\n

\u503c\u5f97\u589e\u52a0\u7684\u6307\u6807\u5305\u62ec：\u5173\u952e\u5b57\u6bb5\u9057\u6f0f\u7387\u3001\u4e8b\u5b9e\u8fc7\u671f\u7387\u3001\u8865\u4e01\u62d2\u7edd\u7387\u3001\u8bed\u4e49\u9519\u8bef\u7387\u3001\u5de5\u5177\u5931\u8d25\u540e\u7684\u72b6\u6001\u504f\u79bb\u65f6\u95f4，\u4ee5\u53ca\u6062\u590d\u5230\u6743\u5a01\u72b6\u6001\u6240\u9700\u7684\u8c03\u7528\u6570\u3002\u518d\u9010\u6b65\u6269\u5927\u5b9e\u4f53\u6570\u91cf\u3001\u4efb\u52a1\u957f\u5ea6\u548c\u76ee\u6807\u53d8\u5316\u7a0b\u5ea6，\u624d\u80fd\u5224\u65ad\u201c\u6709\u754c\u72b6\u6001\u201d\u662f\u5728\u771f\u5b9e\u6269\u5c55\u4e0b\u6210\u7acb，\u8fd8\u662f\u53ea\u5728\u56fa\u5b9a\u5b9e\u9a8c\u5bb9\u91cf\u91cc\u6210\u7acb\u3002<\/p>\n

15.3 \u533a\u5206\u5de5\u4f5c\u72b6\u6001\u3001\u8bc1\u636e\u5b58\u50a8\u4e0e\u5f85\u9a8c\u8bc1\u63a8\u65ad<\/h4>\n

\u4e00\u4e2a\u66f4\u5b8c\u6574\u7684\u7cfb\u7edf\u53ef\u4ee5\u8ba9\u7d27\u51d1\u72b6\u6001\u670d\u52a1\u5f53\u524d\u6267\u884c，\u8ba9\u5916\u90e8\u8bc1\u636e\u5b58\u50a8\u670d\u52a1\u6309\u9700\u8ffd\u6eaf，\u5e76\u4e3a\u6a21\u578b\u5047\u8bbe\u8bbe\u7f6e\u5355\u72ec\u7684\u53ef\u4fe1\u5ea6\u4e0e\u5931\u6548\u6761\u4ef6\u3002\u8fd9\u6837\u65e2\u4e0d\u4f1a\u628a\u6240\u6709\u8fc7\u7a0b\u65e0\u5dee\u522b\u52a0\u5165\u6bcf\u6b65\u8f93\u5165，\u4e5f\u4e0d\u4f1a\u628a\u6682\u65f6\u7684\u731c\u6d4b\u6c38\u4e45\u56fa\u5316\u6210\u4e8b\u5b9e\u3002<\/p>\n

\u8fd9\u79cd\u7ec4\u5408\u4e0d\u662f\u8bba\u6587\u5df2\u7ecf\u9a8c\u8bc1\u7684\u67b6\u6784\u3002\u5b83\u662f\u4ece\u8bba\u6587\u5145\u5206\u6027\u8fb9\u754c\u51fa\u53d1\u7684\u4e00\u6761\u7814\u7a76\u65b9\u5411：\u5f53\u72b6\u6001\u53d1\u73b0\u7f3a\u53e3\u65f6，\u662f\u5426\u80fd\u6709\u63a7\u5236\u5730\u68c0\u7d22\u8bc1\u636e\u5e76\u4fee\u6b63，\u800c\u4e0d\u662f\u8ba9\u6458\u8981\u6216\u65e5\u5fd7\u518d\u6b21\u65e0\u9650\u81a8\u80c0\u3002<\/p>\n

15.4 \u5546\u4e1a\u4ef7\u503c\u6700\u7ec8\u8981\u843d\u5230\u6bcf\u4e2a\u6210\u529f\u4efb\u52a1\u7684\u603b\u6210\u672c<\/h4>\n

\u5bf9\u4ea7\u54c1\u56e2\u961f，\u5e94\u5173\u6ce8\u957f\u4efb\u52a1\u80fd\u5426\u7a33\u5b9a\u5b8c\u6210\u3001\u5f02\u5e38\u662f\u5426\u53ef\u6062\u590d\u3001\u4eba\u5de5\u63a5\u7ba1\u662f\u5426\u51cf\u5c11\u3002\u5bf9\u7b97\u6cd5\u5de5\u7a0b\u5e08，\u5e94\u5173\u6ce8\u6a21\u578b\u80fd\u5426\u6b63\u786e\u7ef4\u62a4\u5173\u7cfb\u4e0e\u8865\u4e01，\u800c\u975e\u53ea\u8f93\u51fa\u5408\u6cd5 JSON\u3002\u5bf9\u521b\u4e1a\u8005\u4e0e\u6295\u8d44\u4eba，\u4e00\u5957\u663e\u5f0f\u72b6\u6001\u683c\u5f0f\u672c\u8eab\u5e76\u4e0d\u8db3\u4ee5\u5f62\u6210\u58c1\u5792；\u9886\u57df\u72b6\u6001\u5efa\u6a21\u3001\u5de5\u5177\u4e00\u81f4\u6027\u3001\u53ef\u9760\u4ea4\u4ed8\u548c\u6301\u7eed\u8bc4\u6d4b\u624d\u51b3\u5b9a\u5176\u79ef\u7d2f\u4ef7\u503c\u3002<\/p>\n

Rocky\u8ba4\u4e3a，SKILL.state\u5c55\u793a\u4e86\u4e00\u6761\u503c\u5f97\u9a8c\u8bc1\u7684Agent\u5de5\u7a0b\u65b9\u5411：\u628a\u8fd0\u884c\u8fc7\u7a0b\u4e2d\u7684\u77ed\u671f\u63a8\u7406，\u8f6c\u5316\u4e3a\u4e0b\u4e00\u6b65\u53ef\u4f7f\u7528\u3001\u53ef\u68c0\u67e5\u3001\u53ef\u4fee\u6b63\u7684\u6267\u884c\u72b6\u6001\u3002 \u5f53\u72b6\u6001\u7684\u5145\u5206\u6027\u3001\u5bb9\u91cf\u548c\u4e8b\u5b9e\u4e00\u81f4\u6027\u90fd\u6709\u6e05\u6670\u8fb9\u754c\u65f6，\u4e0a\u4e0b\u6587\u7ba1\u7406\u624d\u5f00\u59cb\u6210\u4e3a\u53ef\u5ea6\u91cf\u7684\u7cfb\u7edf\u80fd\u529b\u3002<\/p>\n

\u672f\u8bed\u4e0e\u6982\u5ff5\u901f\u67e5<\/h3>\n\n\u672f\u8bed\u672c\u6587\u4e2d\u7684\u542b\u4e49<\/tr>\n\n\n\n\n\n\n\n\n\n\n
Procedural specification<\/td>\n\u4e0d\u53ef\u53d8\u6280\u80fd\u89c4\u8303，\u5b9a\u4e49\u4efb\u52a1\u89c4\u5219\u4e0e\u52a8\u4f5c\u7ea6\u675f<\/td>\n<\/tr>\n
Execution state<\/td>\nAgent \u8de8\u6b65\u9aa4\u4fdd\u7559\u7684\u7ed3\u6784\u5316\u5de5\u4f5c\u72b6\u6001，\u4e0d\u5929\u7136\u7b49\u4e8e\u73af\u5883\u771f\u76f8<\/td>\n<\/tr>\n
State patch<\/td>\n\u63cf\u8ff0\u952e\u66f4\u65b0\u548c\u5220\u9664\u7684\u8865\u4e01，\u9700\u660e\u786e\u5d4c\u5957\u5408\u5e76\u8bed\u4e49<\/td>\n<\/tr>\n
Ephemeral reasoning<\/td>\n\u672c\u6b65\u7528\u4e8e\u751f\u6210\u52a8\u4f5c\u548c\u8865\u4e01\u3001\u540e\u7eed\u4e0d\u91cd\u653e\u7684\u4e34\u65f6\u63a8\u7406<\/td>\n<\/tr>\n
Sufficient statistic<\/td>\n\u5728\u4efb\u52a1\u5047\u8bbe\u4e0b\u5305\u542b\u672a\u6765\u6267\u884c\u6240\u9700\u5386\u53f2\u4fe1\u606f\u7684\u72b6\u6001\u8868\u793a<\/td>\n<\/tr>\n
Bounded prompt<\/td>\n\u8f93\u5165\u5927\u5c0f\u5177\u6709\u72ec\u7acb\u4e8e\u6267\u884c\u957f\u5ea6\u7684\u4e0a\u754c，\u9700\u8981\u5404\u7ec4\u6210\u90e8\u5206\u6709\u754c<\/td>\n<\/tr>\n
Recovery steps<\/td>\n\u7ea0\u6b63\u4fe1\u606f\u5230\u8fbe\u540e\u6062\u590d\u6267\u884c\u6240\u9700\u7684\u989d\u5916\u6b65\u9aa4，\u975e\u5916\u90e8\u53d8\u5316\u53d1\u73b0\u80fd\u529b<\/td>\n<\/tr>\n
Budget-matched control<\/td>\n\u5728\u76f8\u8fd1\u8f93\u5165\u6216\u8d44\u6e90\u9884\u7b97\u4e0b\u6bd4\u8f83\u4fe1\u606f\u4fdd\u7559\u7b56\u7565，\u5fc5\u987b\u7edf\u4e00\u5355\u4f4d<\/td>\n<\/tr>\n
State consistency<\/td>\nAgent \u4fdd\u5b58\u7684\u72b6\u6001\u3001\u5de5\u5177\u8fd4\u56de\u4e0e\u73af\u5883\u771f\u5b9e\u72b6\u6001\u4e4b\u95f4\u7684\u4e00\u81f4\u5173\u7cfb<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u63a8\u8350\u9605\u8bfb<\/h2>\n

Rocky\u4e00\u76f4\u5728\u8fd0\u8425\u6280\u672f\u4ea4\u6d41\u7fa4（WeThinkIn-\u6280\u672f\u4ea4\u6d41\u7fa4），\u8fd9\u4e2a\u7fa4\u7684\u521d\u5fc3\u4e3b\u8981\u805a\u7126\u4e8e\u6280\u672f\u8bdd\u9898\u7684\u8ba8\u8bba\u4e0e\u5b66\u4e60，\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e\u7b97\u6cd5\u3001\u5f00\u53d1\u3001\u7ade\u8d5b\u3001\u79d1\u7814\u4ee5\u53ca\u5de5\u4f5c\u6c42\u804c\u7b49\u3002\u7fa4\u91cc\u6709\u5f88\u591a\u4eba\u5de5\u667a\u80fd\u884c\u4e1a\u7684\u5927\u725b，\u6b22\u8fce\u5927\u5bb6\u5165\u7fa4\u4e00\u8d77\u5b66\u4e60\u4ea4\u6d41～（\u8bf7\u6dfb\u52a0\u5c0f\u52a9\u624b\u5fae\u4fe1Jarvis8866，\u62c9\u4f60\u8fdb\u7fa4～）<\/p>\n

1. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AI Agent（AI\u667a\u80fd\u4f53）\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

2025\u5e74\u53ef\u4ee5\u8bf4\u662fAI Agent\u5168\u9762\u843d\u5730\u5e94\u7528\u7684\u5143\u5e74，\u56e0\u6b64Rocky\u5728\u6301\u7eed\u64b0\u5199\u5bf9AI Agent\u7684\u5168\u7ef4\u5ea6\u89e3\u6790\u6587\u7ae0：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AI Agent（AI\u667a\u80fd\u4f53）\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

2. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790\u6269\u6563\u6a21\u578bDDPM\u3001DDIM\u3001Score-Based\u3001SDE\u3001LDM\u3001Classifier\/Classifier-Free Guidance\u3001Rectified Flow\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9\u6269\u6563\u6a21\u578b\u7684\u672c\u8d28\u539f\u7406\u4e0e\u548c\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6\u8fdb\u884c\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u6df1\u5165\u6d45\u51fa\u5206\u6790\u8bb2\u89e3，\u540c\u65f6\u4e0d\u65ad\u8ddf\u8fdb\u8865\u5145\u6269\u6563\u6a21\u578b\u7684\u6700\u65b0\u6280\u672f\u53d1\u5c55，\u5e0c\u671b\u80fd\u7ed9\u5927\u5bb6\u5e26\u6765\u5e2e\u52a9：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790\u6269\u6563\u6a21\u578bDDPM\u3001DDIM\u3001Score-Based\u3001SDE\u3001LDM\u3001Classifier\/Classifier-Free Guidance\u3001Rectified Flow\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

3. \u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790FLUX.2\u3001Seedream（\u5373\u68a6）\u3001Z-image\u3001GLM-Image\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9AIGC\u65f6\u4ee3\u201c\u4e2d\u573a\u65f6\u523b\u201d\u4e4b\u540e\u7684\u4e3b\u6d41AIGC\u521b\u4f5c\u5927\u6a21\u578b\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6\u8fdb\u884c\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u6df1\u5165\u6d45\u51fa\u5206\u6790\u8bb2\u89e3，\u529b\u6c42\u8ba9\u5927\u5bb6\u901a\u4fd7\u6613\u61c2\u7406\u89e3AIGC\u65f6\u4ee3\u7684\u6280\u672f\u6d6a\u6f6e\u7684\u672c\u8d28\u4ef7\u503c：<\/p>\n

\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790FLUX.2\u3001Seedream（\u5373\u68a6）\u3001Z-image\u3001GLM-Image\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

4. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790FLUX.1 Kontext\u548cFLUX.1 Krea\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9FLUX.1 Kontext\u548cFLUX.1 Krea\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6\u4f5c\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u68b3\u7406\u4e0e\u89e3\u6790：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790FLUX.1 Kontext\u548cFLUX.1 Krea\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

5. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790DeepSeek\u7cfb\u5217\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9DeepSeek\u7cfb\u5217\u6a21\u578b\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6\u4f5c\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u68b3\u7406\u4e0e\u89e3\u6790：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790DeepSeek\u7cfb\u5217\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

6. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion 3（SD 3）\u548cFLUX.1\u7cfb\u5217\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9Stable Diffusion 3\u548cFLUX.1\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6\u4f5c\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u68b3\u7406\u4e0e\u89e3\u6790：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion 3（SD 3）\u548cFLUX.1\u7cfb\u5217\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

7. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion XL（SDXL）\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9Stable Diffusion XL\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6\u4f5c\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u68b3\u7406\u4e0e\u89e3\u6790：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion XL（SDXL）\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

8. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion（SD）\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9Stable Diffusion 1.x-2.x\u7cfb\u5217\u6a21\u578b\u7684\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6\u505a\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u68b3\u7406\u4e0e\u89e3\u6790：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion（SD）\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

9. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion\u4e2dU-Net\u7684\u524d\u4e16\u4eca\u751f\u4e0e\u6838\u5fc3\u77e5\u8bc6<\/h4>\n

Rocky\u5bf9Stable Diffusion\u4e2d\u6700\u4e3a\u5173\u952e\u7684U-Net\u7ed3\u6784\u8fdb\u884c\u4e86\u6df1\u5165\u6d45\u51fa\u7684\u5168\u9762\u89e3\u6790，\u5305\u62ec\u5176\u5728\u4f20\u7edf\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u4ef7\u503c\u548c\u5728AIGC\u4e2d\u7684\u4ef7\u503c：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Stable Diffusion\u4e2dU-Net\u7684\u524d\u4e16\u4eca\u751f\u4e0e\u6838\u5fc3\u77e5\u8bc6<\/p>\n

10. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790LoRA（Low-Rank Adaptation）\u6a21\u578b\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

\u5bf9\u4e8eAIGC\u65f6\u4ee3\u4e2d\u7684\u201cResNet\u201d\u2014\u2014LoRA\u6a21\u578b，Rocky\u8fdb\u884c\u4e86\u6df1\u5165\u6d45\u51fa\u7684\u5168\u9762\u8bb2\u89e3：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790LoRA（Low-Rank Adaptation）\u6a21\u578b\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

11. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790ControlNet\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

AIGC\u56fe\u50cf\u521b\u4f5c\u5f00\u6e90\u793e\u533a\u5df2\u7ecf\u5f62\u6210\u4ee5Stable Difffusion\/FLUX\u4e3a\u6838\u5fc3，ConrtolNet\u548cLoRA\u4f5c\u4e3a\u9996\u8981AI\u8f85\u52a9\u5de5\u5177\u7684\u53d8\u5316\u4e07\u5343\u7684AIGC\u56fe\u50cf\u521b\u4f5c\u5de5\u4f5c\u6d41\u3002<\/p>\n

ControlNet\u6b63\u662f\u8ba9AI\u56fe\u50cf\u521b\u4f5c\u793e\u533a\u65e0\u6bd4\u7e41\u8363\u7684\u5173\u952e\u4e00\u73af，\u5b83\u8ba9AIGC\u56fe\u50cf\u521b\u4f5c\u8fc7\u7a0b\u66f4\u52a0\u7684\u53ef\u63a7，\u66f4\u6709\u52a9\u4e8e\u5e7f\u6cdb\u5730\u5c06AIGC\u7b97\u6cd5\u89e3\u51b3\u65b9\u6848\u5e94\u7528\u5230\u5404\u884c\u5404\u4e1a\u4e2d：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790ControlNet\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

12. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Sora\u3001Seedance\u3001keling\u7b49AI\u89c6\u9891\u5927\u6a21\u578b\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

AI\u7ed8\u753b\u548cAI\u89c6\u9891\u662f\u4e24\u4e2a\u4e92\u76f8\u4fc3\u8fdb\u3001\u76f8\u4e92\u4ea4\u878d\u7684\u9886\u57df，2024\u5e74\u65e0\u7591\u662fAI\u89c6\u9891\u9886\u57df\u7684\u7206\u53d1\u4e4b\u5e74，Rocky\u5bf9AI\u89c6\u9891\u9886\u57df\u6838\u5fc3\u7684Sora\u3001Seedance\u3001Keling\u7b49\u5927\u6a21\u578b\u8fdb\u884c\u4e86\u5168\u9762\u7cfb\u7edf\u7684\u68b3\u7406\u4e0e\u89e3\u6790：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790Sora\u3001Seedance\u3001keling\u7b49AI\u89c6\u9891\u5927\u6a21\u578b\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

13. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AIGC\u65f6\u4ee3Transformer\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

\u5728AIGC\u65f6\u4ee3\u4e2d，Transformer\u4e3aAI\u884c\u4e1a\u5e26\u6765\u4e86\u6df1\u523b\u7684\u53d8\u9769\u3002Transformer\u67b6\u6784\u6b63\u5728\u4e00\u6b65\u4e00\u6b65\u91cd\u6784\u6240\u6709\u7684AI\u6280\u672f\u65b9\u5411，\u6210\u4e3aAI\u6280\u672f\u67b6\u6784\u5927\u4e00\u7edf\u4e0e\u591a\u6a21\u6001\u6574\u5408\u7684\u5173\u952e\u6838\u5fc3\u57fa\u5ea7，\u5927\u6709\u4e00\u7edf\u201cAI\u6c5f\u6e56\u201d\u4e4b\u52bf\u3002Rocky\u4e5f\u5bf9Transformer\u6a21\u578b\u8fdb\u884c\u6301\u7eed\u7684\u6df1\u5165\u6d45\u51fa\u68b3\u7406\u4e0e\u89e3\u6790：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AIGC\u65f6\u4ee3Transformer\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

14. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790ComfyUI\u3001Diffusers\u3001Stable Diffusion WebUI\u7b49\u4e3b\u6d41AIGC\u521b\u4f5c\u6846\u67b6\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

AIGC\u521b\u4f5c\u6846\u67b6\u6b63\u662fAIGC\u7b97\u6cd5\u5de5\u4f5c\u6d41\u7684\u8fd0\u884c\u8f7d\u4f53，\u76ee\u524d\u4e3b\u6d41\u7684AIGC\u521b\u4f5c\u6846\u67b6\u6709ComfyUI\u3001Diffusers\u3001Stable Diffusion WebUI\u7b49\u3002\u5728\u4f20\u7edf\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3，PyTorch\u3001TensorFlow\u4ee5\u53caCaffe\u662f\u4f20\u7edf\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u57fa\u7840\u8fd0\u884c\u6846\u67b6，\u5230\u4e86AIGC\u65f6\u4ee3，Rocky\u76f8\u4fe1ComfyUI\u5c31\u662fAIGC\u65f6\u4ee3\u7684\u201cPyTorch\u201d\u3001Stable Diffusion WebUI\u5c31\u662fAIGC\u65f6\u4ee3\u7684\u201cTensorFlow\u201d\u3001Diffusers\u5c31\u662fAIGC\u65f6\u4ee3\u7684\u201cCaffe\u201d：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790ComfyUI\u3001Diffusers\u3001Stable Diffusion WebUI\u7b49\u4e3b\u6d41AIGC\u521b\u4f5c\u6846\u67b6\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n

15. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790ComfyUI\u3001Diffusers\u3001Stable Diffusion WebUI\u7b49\u4e3b\u6d41AIGC\u521b\u4f5c\u6846\u67b6\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

\u5728AIGC\u65f6\u4ee3\u4e2d，\u5982\u4f55\u5feb\u901f\u8f6c\u8eab，\u5165\u5c40AIGC\u4ea7\u4e1a？\u5982\u4f55\u6210\u4e3aAIGC\/LLM\/AI Agent\u7b97\u6cd5\/\u5f00\u53d1\u5de5\u7a0b\u5e08？\u5982\u4f55\u5728\u5b66\u6821\u4e2d\u7cfb\u7edf\u6027\u5b66\u4e60AIGC\/LLM\/AI Agent\u77e5\u8bc6，\u65a9\u83b7\u5fc3\u4eea\u7684AIGC\/LLM\/AI Agent\u7b97\u6cd5\/\u5f00\u53d1offer？<\/p>\n

Don\u2018t worry，Rocky\u4e3a\u5927\u5bb6\u603b\u7ed3\u6574\u7406\u4e86\u5168\u9762\u7684AIGC\/LLM\/AI Agent\u7b97\u6cd5\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6210\u957f\u79d8\u7c4d，\u4e3a\u5927\u5bb6\u7b54\u7591\u89e3\u60d1，\u5e0c\u671b\u80fd\u7ed9\u5927\u5bb6\u5e26\u6765\u5e2e\u52a9：<\/p>\n

\u624b\u628a\u624b\u6559\u4f60\u6210\u4e3aAIGC\/LLM\/AI Agent\u7b97\u6cd5\/\u5f00\u53d1\u5de5\u7a0b\u5e08，\u65a9\u83b7AIGC\/LLM\/AI Agent\u7b97\u6cd5\/\u5f00\u53d1offer！<\/p>\n

16. AIGC\u4ea7\u4e1a\u7684\u6df1\u5ea6\u601d\u8003\u4e0e\u5206\u6790<\/h4>\n

2023\u5e743\u670821\u65e5，\u5fae\u8f6f\u521b\u59cb\u4eba\u6bd4\u5c14\u00b7\u76d6\u8328\u5728\u5176\u535a\u5ba2\u6587\u7ae0\u300aThe Age of AI has begun\u300b\u4e2d\u8868\u793a，\u81ea\u4ece1980\u5e74\u9996\u6b21\u770b\u5230\u56fe\u5f62\u7528\u6237\u754c\u9762（graphical user interface）\u4ee5\u6765，\u4ee5OpenAI\u4e3a\u4ee3\u8868\u7684\u79d1\u6280\u516c\u53f8\u53d1\u5e03\u7684AIGC\u6a21\u578b\u662f\u4ed6\u6240\u89c1\u8fc7\u7684\u6700\u5177\u9769\u547d\u6027\u7684\u6280\u672f\u8fdb\u6b65\u3002<\/p>\n

Rocky\u4e5f\u8ba4\u4e3a，AIGC\u53ca\u5176\u751f\u6001，\u4f1a\u6210\u4e3aAI\u884c\u4e1a\u91cd\u5927\u53d8\u9769\u7684\u4e3b\u5bfc\u529b\u91cf\u3002AIGC\u4f1a\u5e26\u6765\u4e00\u4e2a\u5168\u65b0\u7684\u7ea2\u5229\u671f，\u672a\u6765\u968f\u7740AIGC\u7684\u5168\u9762\u843d\u5730\u548c\u6df1\u5ea6\u5546\u7528，\u4f1a\u6df1\u523b\u6539\u53d8\u6211\u4eec\u7684\u5de5\u4f5c\u3001\u751f\u6d3b\u3001\u5b66\u4e60\u4ee5\u53ca\u4ea4\u6d41\u65b9\u5f0f，\u5404\u884c\u5404\u4e1a\u90fd\u5c06\u88ab\u91cd\u65b0\u5b9a\u4e49，\u8fc7\u7a0b\u4f1a\u975e\u5e38\u6709\u8da3\u3002<\/p>\n

\u90a3\u4e48，\u5728\u6b64\u57fa\u7840\u4e0a，\u6211\u4eec\u8be5\u5982\u4f55\u66f4\u597d\u7684\u5ba1\u89c6AIGC\u7684\u672a\u6765？\u6211\u4eec\u8be5\u5982\u4f55\u66f4\u597d\u5730\u62e5\u62b1AIGC\u5f15\u9886\u7684\u9769\u65b0？Rocky\u51c6\u5907\u4ece\u6280\u672f\u3001\u4ea7\u54c1\u3001\u5546\u4e1a\u6a21\u5f0f\u3001\u957f\u671f\u4e3b\u4e49\u7b49\u7ef4\u5ea6\u6301\u7eed\u5206\u4eab\u4e00\u4e9b\u4e2a\u4eba\u7684\u6838\u5fc3\u601d\u8003\u4e0e\u89c2\u70b9，\u5e0c\u671b\u80fd\u5e2e\u52a9\u5404\u4f4d\u8bfb\u8005\u5bf9AIGC\u6709\u4e00\u4e2a\u5168\u9762\u7684\u4e86\u89e3：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5168\u9762\u89e3\u6790AIGC\u65f6\u4ee3\u6838\u5fc3\u4ef7\u503c\u4e0e\u53d1\u5c55\u8d8b\u52bf（2025\u5e74\u7248）<\/p>\n

17. AI\u7b97\u6cd5\u5de5\u7a0b\u5e08\u7684\u72ec\u5b64\u4e5d\u5251\u79d8\u7c4d<\/h4>\n

\u4e3a\u4e86\u65b9\u4fbf\u5927\u5bb6\u5b9e\u4e60\u3001\u6821\u62db\u4ee5\u53ca\u793e\u62db\u7684\u9762\u8bd5\u51c6\u5907，\u540c\u65f6\u5e2e\u52a9\u5927\u5bb6\u63d0\u5347\u6269\u5c55\u6280\u672f\u57fa\u672c\u9762，Rocky\u5c06\u7b26\u5408\u5927\u5382\u548cAI\u72ec\u89d2\u517d\u4ef7\u503c\u7684\u7b97\u6cd5\u9ad8\u9891\u9762\u8bd5\u77e5\u8bc6\u70b9\u64b0\u5199\u603b\u7ed3\u6210\u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300b\u4e4b\u72ec\u5b64\u4e5d\u5251\u79d8\u7c4d:<\/p>\n

\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011AIGC\u65f6\u4ee3\u7684\u7b97\u6cd5\u5de5\u7a0b\u5e08\u7684\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d（\u6301\u7eed\u66f4\u65b0\u4e2d）<\/p>\n

18. \u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AIGC\u65f6\u4ee3\u4e2dGAN（Generative Adversarial Network）\u7cfb\u5217\u6a21\u578b\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/h4>\n

GAN\u7cfb\u5217\u6a21\u578b\u4f5c\u4e3a\u4f20\u7edf\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3\u7684\u6700\u70ed\u95e8\u751f\u6210\u5f0fAl\u6a21\u578b，\u5728AIGC\u65f6\u4ee3\u7ee7\u7eed\u7e41\u8363，\u4f5c\u4e3aStable Diffusion\/FLUX\u7cfb\u5217\u5927\u6a21\u578b\u7684\u201c\u5f97\u529b\u52a9\u624b\u201d，\u5e7f\u6cdb\u6d3b\u8dc3\u4e8eAlGC\u56fe\u50cf\u521b\u4f5c\u7684\u4ea7\u54c1\u4e0e\u5de5\u4f5c\u6d41\u4e2d：<\/p>\n

\u6df1\u5165\u6d45\u51fa\u5b8c\u6574\u89e3\u6790AIGC\u65f6\u4ee3\u4e2dGAN（Generative Adversarial Network）\u7cfb\u5217\u6a21\u578b\u6838\u5fc3\u57fa\u7840\u77e5\u8bc6<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5199\u5728\u524d\u9762 \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u516c\u4f17\u53f7：WeThinkIn \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-Interview-Book，\u6b22\u8fce\u5927\u5bb6Star～ Rocky\u6700\u65b0\u64b0\u5199\u768410\u4e07\u5b57AI Agent（AI\u667a\u80fd\u4f53）\u6df1\u5165\u6d45\u51fa\u5168\u7ef4<\/p>\n","protected":false},"author":2,"featured_media":106865,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[52,50,207,86,3546],"topic":[],"class_list":["post-106867","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-aigc","tag-50","tag-207","tag-86","tag-3546"],"yoast_head":"\nSKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/106867.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u5199\u5728\u524d\u9762 \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u516c\u4f17\u53f7：WeThinkIn \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-Interview-Book，\u6b22\u8fce\u5927\u5bb6Star～ Rocky\u6700\u65b0\u64b0\u5199\u768410\u4e07\u5b57AI Agent（AI\u667a\u80fd\u4f53）\u6df1\u5165\u6d45\u51fa\u5168\u7ef4\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/106867.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-18T05:00:44+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260918050042-6aacc57a63134.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"13 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/106867.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/106867.html\",\"name\":\"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-18T05:00:44+00:00\",\"dateModified\":\"2026-09-18T05:00:44+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/106867.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/106867.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/106867.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/106867.html","og_locale":"zh_CN","og_type":"article","og_title":"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u5199\u5728\u524d\u9762 \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u516c\u4f17\u53f7：WeThinkIn \u6b22\u8fce\u5927\u5bb6\u5173\u6ce8Rocky\u7684\u77e5\u4e4e：Rocky Ding \u300a\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u300bAIGC\/LLM\/AI Agent\u7b97\u6cd5\u5de5\u7a0b\u5e08\/\u5f00\u53d1\u5de5\u7a0b\u5e08\u6c42\u804c\u9762\u8bd5\u79d8\u7c4d\u72ec\u5bb6\u8d44\u6e90：\u3010\u4e09\u5e74\u9762\u8bd5\u4e94\u5e74\u6a21\u62df\u3011WeThinkIn\/AIGC-Interview-Book，\u6b22\u8fce\u5927\u5bb6Star～ Rocky\u6700\u65b0\u64b0\u5199\u768410\u4e07\u5b57AI Agent（AI\u667a\u80fd\u4f53）\u6df1\u5165\u6d45\u51fa\u5168\u7ef4","og_url":"https:\/\/www.wsisp.com\/helps\/106867.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-18T05:00:44+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260918050042-6aacc57a63134.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"13 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/106867.html","url":"https:\/\/www.wsisp.com\/helps\/106867.html","name":"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-18T05:00:44+00:00","dateModified":"2026-09-18T05:00:44+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/106867.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/106867.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/106867.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"SKILL.state\u6df1\u5165\u89e3\u6790\uff1a\u628aAI Agent\u957f\u7a0b\u6267\u884c\u53d8\u6210\u53ef\u6821\u9a8c\u7684\u72b6\u6001\u8fc1\u79fb"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/106867","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=106867"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/106867\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/106865"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=106867"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=106867"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=106867"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=106867"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}