{"id":29732,"date":"2026-07-17T12:05:53","date_gmt":"2026-07-17T04:05:53","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=29732"},"modified":"2026-07-17T12:05:53","modified_gmt":"2026-07-17T04:05:53","slug":"%e5%89%8d-openai-%e7%a7%91%e5%ad%a6%e5%ae%b6%e8%b0%88-harness-%e5%b7%a5%e7%a8%8b%e7%9a%84%e8%87%aa%e6%88%91%e6%8f%90%e5%8d%87%e6%9e%b6%e6%9e%84%e8%ae%be%e8%ae%a1","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2026\/07\/17\/29732\/","title":{"rendered":"\u524d OpenAI \u79d1\u5b66\u5bb6\u8c08 Harness \u5de5\u7a0b\u7684\u81ea\u6211\u63d0\u5347\u67b6\u6784\u8bbe\u8ba1"},"content":{"rendered":"
\u672c\u6587\u805a\u7126 Harness \u5de5\u7a0b\u76f8\u5173\u7814\u7a76\uff0c\u4ee5\u53ca\u5b83\u5982\u4f55\u63a8\u52a8 RSI\u3002\u8fd1\u671f\u8bb8\u591a\u5173\u4e8e\u81ea\u52a8\u5316\u7814\u7a76\uff08auto-research\uff09\u3001\u81ea\u6211\u6539\u8fdb\u667a\u80fd\u4f53\u3001\u8fdb\u5316\u5f0f\u7a0b\u5e8f\u641c\u7d22\u7684\u5de5\u4f5c\uff0c\u90fd\u53ef\u4ee5\u56f4\u7ed5\u8fd9\u4e00\u95ee\u9898\u6765\u7ec4\u7ec7\u3002<\/div>\n
\n
\n

\u5927\u5bb6\u597d\uff0c\u6211\u662f\u7384\u59d0\u3002<\/p>\n

\u9012\u5f52\u81ea\u6211\u6539\u8fdb\uff08recursive self-improvement\uff0cRSI\uff09\u7684\u6982\u5ff5\u53ef\u8ffd\u6eaf\u81f3 I. J. Good\uff081965\uff09\uff0c\u4ed6\u5c06”\u8d85\u667a\u80fd\u673a\u5668\uff08ultraintelligent machine\uff09”\u5b9a\u4e49\u4e3a\u4e00\u79cd\u5728\u6240\u6709\u667a\u529b\u6d3b\u52a8\u4e0a\u90fd\u80fd\u8d85\u8d8a\u4eba\u7c7b\u3001\u5e76\u80fd\u8bbe\u8ba1\u51fa\u66f4\u597d\u7684\u673a\u5668\u6765\u6539\u8fdb\u81ea\u8eab\u7684\u7cfb\u7edf\u3002Yudkowsky\uff082008\uff09\u7528”\u9012\u5f52\u81ea\u6211\u6539\u8fdb”\u4e00\u8bcd\u63cf\u8ff0\u4e00\u4e2a\u7279\u5b9a\u7684\u53cd\u9988\u56de\u8def\uff1aAI \u5229\u7528\u5176\u5f53\u524d\u7684\u667a\u80fd\uff0c\u53bb\u6539\u8fdb\u4ea7\u751f\u5176\u667a\u80fd\u7684\u8ba4\u77e5\u673a\u5236\u3002<\/p>\n

\u8fd9\u4e00\u53cd\u9988\u56de\u8def\u5728\u73b0\u4ee3 AI \u4e2d\uff0c\u53ef\u80fd\u4f53\u73b0\u4e3a\u6a21\u578b\u76f4\u63a5\u91cd\u5199\u81ea\u5df1\u7684\u6743\u91cd\uff1b\u6216\u8005\u66f4\u5e7f\u4e49\u5730\u8bf4\uff0c\u6a21\u578b\u53bb\u6539\u8fdb\u8bad\u7ec3\u6d41\u6c34\u7ebf\uff08training pipeline\uff09\u4e0e\u90e8\u7f72\u7cfb\u7edf\uff08deployment system\uff09\uff0c\u4ece\u800c\u50ac\u751f\u51fa\u5728\u7ecf\u6d4e\u4ef7\u503c\u4efb\u52a1\u4e0a\u8868\u73b0\u66f4\u5f3a\u7684\u4e0b\u4e00\u4ee3\u6a21\u578b\u3002\u524d\u6cbf\u5b9e\u9a8c\u5ba4\u7684 AI \u7814\u53d1\u901f\u5ea6\u5df2\u88ab\u8bc1\u660e\u5728\u5927\u5e45\u52a0\u901f\uff08Anthropic\uff1bOpenAI\uff09\u3002<\/p>\n

\u6211\u7279\u610f\u63d0\u5230”\u90e8\u7f72\u7cfb\u7edf”\uff0c\u662f\u56e0\u4e3a\u4ecb\u4e8e\u539f\u59cb\u6a21\u578b\u4e0e\u771f\u5b9e\u4e16\u754c\u4e0a\u4e0b\u6587\u4e4b\u95f4\u7684\u8fd9\u4e00\u5c42\uff0c\u5176\u91cd\u8981\u6027\u4f3c\u4e4e\u4e0d\u4e9a\u4e8e\u6a21\u578b\u7684\u539f\u59cb\u667a\u80fd\uff08\u5373\u9884\u8bad\u7ec3\u540e\u7acb\u523b\u8dd1\u51fa\u7684\u8bc4\u6d4b\u6210\u7ee9\uff09\u3002Harness \u662f AI \u90e8\u7f72\u7684\u91cd\u8981\u7ec4\u6210\u90e8\u5206\uff0cClaude Code\u3001Codex \u7b49\u6210\u529f\u7684\u7f16\u7a0b\u667a\u80fd\u4f53\u4ea7\u54c1\u5df2\u7ecf\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9\u3002Harness\u00a0\u6307\u5305\u88f9\u5728\u57fa\u7840\u6a21\u578b\u5916\u56f4\u7684\u7cfb\u7edf\uff0c\u5b83\u8d1f\u8d23\u7f16\u6392\u6267\u884c\uff0c\u51b3\u5b9a\u6a21\u578b\u5982\u4f55\u601d\u8003\u4e0e\u89c4\u5212\u3001\u5982\u4f55\u8c03\u7528\u5de5\u5177\u4e0e\u884c\u52a8\u3001\u5982\u4f55\u611f\u77e5\u4e0e\u7ba1\u7406\u4e0a\u4e0b\u6587\u3001\u5982\u4f55\u5b58\u50a8\u4ea7\u7269\u3001\u4ee5\u53ca\u5982\u4f55\u8bc4\u4f30\u7ed3\u679c\u3002<\/p>\n

\u672c\u6587\u805a\u7126 Harness \u5de5\u7a0b\u76f8\u5173\u7814\u7a76\uff0c\u4ee5\u53ca\u5b83\u5982\u4f55\u63a8\u52a8 RSI\u3002\u8fd1\u671f\u8bb8\u591a\u5173\u4e8e\u81ea\u52a8\u5316\u7814\u7a76\uff08auto-research\uff09\u3001\u81ea\u6211\u6539\u8fdb\u667a\u80fd\u4f53\u3001\u8fdb\u5316\u5f0f\u7a0b\u5e8f\u641c\u7d22\u7684\u5de5\u4f5c\uff0c\u90fd\u53ef\u4ee5\u56f4\u7ed5\u8fd9\u4e00\u95ee\u9898\u6765\u7ec4\u7ec7\u3002\u6a21\u578b\u81ea\u535a\u5f08\uff08self-play\uff09\u3001\u5408\u6210\u6570\u636e\u3001\u6d4b\u8bd5\u65f6\u8bad\u7ec3\uff08test-time training\uff09\u4ee5\u53ca\u66f4\u5927\u7684\u6301\u7eed\u5b66\u4e60\uff08continual learning\uff09\u4e3b\u9898\u4e5f\u7b26\u5408 RSI \u7684\u613f\u666f\uff08\u5982 Yuan et al. 2024\u3001Chen et al. 2024\u3001Zhao et al. 2025\u3001Choi et al. 2026\uff09\uff0c\u4f46\u4e0d\u662f\u672c\u6587\u7684\u91cd\u70b9\u3002<\/p>\n

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

\u4e00\u3001Harness \u8bbe\u8ba1\u6a21\u5f0f<\/h2>\n

\u4e0e\u65e9\u671f\u667a\u80fd\u4f53\u6846\u67b6”\u667a\u80fd\u4f53 = LLM + \u8bb0\u5fc6 + \u5de5\u5177 + \u89c4\u5212 + \u884c\u52a8”\u76f8\u6bd4\uff0cHarness \u5de5\u7a0b\u8fd8\u989d\u5916\u5305\u62ec\uff1a\u5de5\u4f5c\u6d41\u8bbe\u8ba1\uff08\u5982\u5faa\u73af\u5de5\u7a0b\uff09\u3001\u8bc4\u4f30\u3001\u6743\u9650\u63a7\u5236\u4e0e\u6301\u4e45\u72b6\u6001\u7ba1\u7406\u3002\u5b83\u4e0d\u518d\u53ea\u662f\u63d0\u793a\u8bcd\u6a21\u677f\uff0c\u800c\u66f4\u63a5\u8fd1\u8fd0\u884c\u65f6\u4e0e\u8f6f\u4ef6\u7cfb\u7edf\u8bbe\u8ba1\uff1a\u6a21\u578b\u5982\u4f55\u89c2\u5bdf\u3001\u884c\u52a8\u3001\u8bb0\u5fc6\u3001\u81ea\u68c0\u4e0e\u6539\u8fdb\u3002<\/p>\n

\u8bbe\u8ba1\u4e0a\u5e94\u5f53\u523b\u610f\u4fdd\u6301\u7b80\u5355\u4e0e\u901a\u7528\uff0c\u4ee5\u5b9e\u73b0\u6cdb\u5316\uff1b\u540c\u65f6\u6700\u597d\u501f\u9274\u73b0\u6709\u8f6f\u4ef6\u5de5\u7a0b\u5b9e\u8df5\uff0c\u4ece\u800c\u53d7\u76ca\u4e8e\u9884\u8bad\u7ec3\u77e5\u8bc6\u3002\u64cd\u4f5c\u7cfb\u7edf\u4e0e Harness \u4e4b\u95f4\u5b58\u5728\u5f88\u5f3a\u7684\u7c7b\u6bd4\uff1a\u6b63\u5982\u64cd\u4f5c\u7cfb\u7edf\u4e00\u6837\uff0cHarness \u5e94\u5f53\u5c01\u88c5\u590d\u6742\u903b\u8f91\uff0c\u540c\u65f6\u4fdd\u6301\u63a5\u53e3\u7b80\u5355\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u914d\u7f6e\u3001\u5de5\u5177\u63a5\u53e3\u7b49\u534f\u8bae\u53ef\u80fd\u4f1a\u9010\u6e10\u5728\u884c\u4e1a\u5185\u6807\u51c6\u5316\u3002<\/p>\n

1.1 \u6a21\u5f0f\u4e00\uff1a\u5de5\u4f5c\u6d41\u81ea\u52a8\u5316<\/h3>\n

\u5b9a\u4e49\u4e00\u4e2a\u6a21\u578b\u53ef\u4ee5\u5728\u5176\u4e2d\u8fd0\u884c\u3001\u6d4b\u8bd5\u4e0e\u8fed\u4ee3\u7684\u5de5\u4f5c\u6d41\uff0c\u662f\u81ea\u52a8\u5316\u7684\u5173\u952e\u8bbe\u8ba1\u3002Karpathy \u7684 autoresearch \u4ed3\u5e93\uff08https:\/\/github.com\/karpathy\/autoresearch\uff09\u5c31\u662f\u6784\u5efa\u6b64\u7c7b\u5de5\u4f5c\u6d41\u7684\u4e00\u4e2a\u5e72\u51c0\u8303\u4f8b\u3002\u5e38\u89c1\u7684\u5de5\u4f5c\u6d41\u9075\u5faa\u76ee\u6807\u5bfc\u5411\u7684\u5faa\u73af\uff1a\u89c4\u5212\uff08plan\uff09\u2192 \u6267\u884c\uff08execute\uff09\u2192 \u89c2\u5bdf\/\u6d4b\u8bd5\uff08observe\/test\uff09\u2192 \u6539\u8fdb\uff08improve\uff09\u2192 \u518d\u6267\u884c\uff0c\u76f4\u5230\u76ee\u6807\u8fbe\u6210\u3002\u8fc7\u7a0b\u4e2d\u8fd8\u53ef\u80fd\u4e3b\u52a8\u5411\u7528\u6237\u53d1\u8d77\u8bf7\u6c42\uff0c\u4ee5\u6f84\u6e05\u4efb\u52a1\u89c4\u683c\u6216\u6267\u884c\u504f\u597d\u3002<\/p>\n

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

\uff08\u56fe1\uff1a\u7b80\u5316\u7684 Codex \u667a\u80fd\u4f53\u5faa\u73af\uff1a\u667a\u80fd\u4f53\u8c03\u7528\u5de5\u5177\uff0c\u5de5\u5177\u8fd4\u56de\u7ed3\u679c\u5f71\u54cd\u6a21\u578b\u7684\u4e0b\u4e00\u6b21\u751f\u6210\u3002\u56fe\u7247\u6765\u6e90\uff1aOpenAI codex agent \u6587\u7ae0\uff09<\/p>\n

\u8fd9\u5f20\u5de5\u4f5c\u6d41\u56fe\u8fd8\u5f3a\u8c03\u4e86\u4e00\u70b9\uff1a\u6a21\u578b\u901a\u8fc7”\u667a\u80fd\u4f53\u8fd0\u884c\u65f6\uff08agent runtime\uff09”\uff08\u800c\u975e\u9759\u6001\u63d0\u793a\u8bcd\u6a21\u677f\uff09\u5206\u6790\u81ea\u5df1\u7684\u8f68\u8ff9\u4e0e\u5931\u8d25\u6848\u4f8b\uff0c\u7136\u540e\u5728\u6b64\u57fa\u7840\u4e0a\u8fed\u4ee3\u3002<\/p>\n

1.2 \u6a21\u5f0f\u4e8c\uff1a\u6587\u4ef6\u7cfb\u7edf\u4f5c\u4e3a\u6301\u4e45\u8bb0\u5fc6<\/h3>\n

\u957f\u7a0b\u667a\u80fd\u4f53\u7cfb\u7edf\u4e2d\u53cd\u590d\u51fa\u73b0\u7684\u4e00\u4e2a\u6a21\u5f0f\u662f\uff1a\u7528\u7b80\u5355\u7684\u63a7\u5236\u65b9\u5f0f\u7ba1\u7406\u4e30\u5bcc\u7684\u72b6\u6001\u4e0e\u4ea7\u7269\u3002Harness \u4e0d\u5e94\u628a\u6574\u4e2a\u5de5\u4f5c\u6d41\u548c\u6240\u6709\u65e5\u5fd7\u90fd\u80cc\u5728\u4e0a\u4e0b\u6587\u91cc\uff0c\u800c\u5e94\u628a\u6301\u4e45\u72b6\u6001\u4fdd\u5b58\u5728\u6587\u4ef6\u4e2d\u3002\u5728\u957f\u7a0b\u667a\u80fd\u4f53\u5c55\u5f00\uff08rollout\uff09\u4e2d\uff0c\u5b9e\u9a8c\u65e5\u5fd7\u3001\u4ee3\u7801 diff\u3001\u8bba\u6587\u6458\u8981\u3001\u9519\u8bef\u5806\u6808\u3001\u5386\u53f2\u8f68\u8ff9\u7b49\u4ea7\u7269\uff0c\u5f80\u5f80\u4f1a\u589e\u957f\u5230\u8fdc\u8d85\u6a21\u578b\u8bad\u7ec3\u65f6\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u957f\u5ea6\u3002<\/p>\n

\u5b66\u4f1a\u8bfb\u5199\u3001\u7f16\u8f91\u6587\u4ef6\u7cfb\u7edf\uff08\u901a\u5e38\u901a\u8fc7bash\u547d\u4ee4\uff09\u662f LLM \u7684\u4e00\u9879\u57fa\u7840\u6280\u80fd\uff0c\u56e0\u6b64\u7528\u6587\u4ef6\u8fd9\u79cd\u7b80\u5355\u5f62\u5f0f\u7ba1\u7406\u6301\u4e45\u8bb0\u5fc6\uff0c\u81ea\u7136\u4f1a\u968f\u6838\u5fc3\u6a21\u578b\u80fd\u529b\u7684\u63d0\u5347\u800c\u53d7\u76ca\u3002<\/p>\n

1.3 \u6a21\u5f0f\u4e09\uff1a\u5b50\u667a\u80fd\u4f53\u4e0e\u540e\u53f0\u4efb\u52a1<\/h3>\n

Harness \u53ef\u4ee5\u6d3e\u751f\u591a\u4e2a\u5b50\u667a\u80fd\u4f53\u5e76\u884c\u6267\u884c\uff0c\u5e76\u76d1\u63a7\u540e\u53f0\u4efb\u52a1\u3002\u5f53\u4e3b\u667a\u80fd\u4f53\u9700\u8981\u641c\u7d22\u591a\u4e2a\u5047\u8bbe\u3001\u5e76\u53d1\u8dd1\u5b9e\u9a8c\u3001\u6216\u628a\u9694\u79bb\u7684\u5b50\u4efb\u52a1\u59d4\u6d3e\u51fa\u53bb\u800c\u4e0d\u6c61\u67d3\u4e3b\u4e0a\u4e0b\u6587\u65f6\uff0c\u8fd9\u4e00\u70b9\u975e\u5e38\u6709\u7528\u3002\u7236\u667a\u80fd\u4f53\u968f\u540e\u9700\u8981\u4e00\u4e2a\u5c0f\u578b\u8fdb\u7a0b\u7ba1\u7406\u5668\uff1a\u542f\u52a8\u4efb\u52a1\u3001\u68c0\u67e5\u65e5\u5fd7\u3001\u53d6\u6d88\u5931\u8d25\u7684\u8fd0\u884c\u3001\u628a\u7ed3\u679c\u5408\u5e76\u56de\u4e3b\u667a\u80fd\u4f53\u7ebf\u7a0b\u3002<\/p>\n

\u5173\u952e\u8bbe\u8ba1\u9009\u62e9\u662f\u8ba9\u5e76\u884c\u663e\u5f0f\u5316\u4e14\u53ef\u68c0\u67e5\u3002\u5982\u679c\u5b50\u667a\u80fd\u4f53\u7684\u4ea7\u51fa\u53ea\u5b58\u5728\u4e8e\u4e34\u65f6\u7684\u804a\u5929\u4e0a\u4e0b\u6587\u4e2d\uff0c\u5b83\u4eec\u5f88\u5feb\u5c31\u4f1a\u8fc7\u65f6\u4e14\u4e0d\u53ef\u89c1\uff1b\u5982\u679c\u4ee5\u6587\u4ef6\u3001\u65e5\u5fd7\u548c\u72b6\u6001\u8bb0\u5f55\u7684\u5f62\u5f0f\u5b58\u50a8\uff0c\u6a21\u578b\u5c31\u80fd\u5728\u4e2d\u65ad\u540e\u6062\u590d\uff0c\u5e76\u5bf9\u81ea\u8eab\u7684\u6267\u884c\u5386\u53f2\u8fdb\u884c\u63a8\u7406\u3002<\/p>\n

1.4 \u6848\u4f8b\u7814\u7a76\uff1a\u7f16\u7a0b\u667a\u80fd\u4f53 Harness<\/h3>\n

\u4e3b\u6d41\u7f16\u7a0b\u667a\u80fd\u4f53\u7684\u6838\u5fc3\u63a5\u53e3\u5df2\u7ecf\u8d8b\u4e8e\u7a33\u5b9a\uff0cClaude Code\u3001Codex\u3001OpenCode \u4ee5\u53ca Cursor \u98ce\u683c\u7684\u667a\u80fd\u4f53\u666e\u904d\u91c7\u7528\u5982\u4e0b\u5faa\u73af\uff1a<\/p>\n

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

\uff08\u56fe2\uff1a\u89c2\u5bdf\u4ee3\u7801\u4ed3\u5e93 \u2192 \u89c4\u5212 \u2192 \u641c\u7d22\/\u9605\u8bfb\u6587\u4ef6 \u2192 \u7f16\u8f91\/\u5199\u8865\u4e01 \u2192 \u8dd1\u6d4b\u8bd5 \u2192 \u68c0\u67e5\u9519\u8bef \u2192 \u91cd\u590d\uff0c\u76f4\u81f3\u5b8c\u6210\uff09<\/p>\n

\u501f\u52a9\u4e00\u7ec4\u5de5\u5177\uff0c\u7f16\u7a0b\u667a\u80fd\u4f53\u53ef\u4ee5\u5728\u7ed9\u5b9a\u4ed3\u5e93\u4e2d\u5f00\u53d1\u548c\u8c03\u8bd5\u95ee\u9898\uff0c\u5c31\u50cf\u4eba\u7c7b\u5f00\u53d1\u8005\u914d\u5907 IDE \u4e00\u6837\u3002<\/p>\n

\uff08\u5e76\u975e\u5b8c\u6574\u6e05\u5355\uff0c\u4ec5\u4f5c\u6f14\u793a\u3002\uff09<\/p>\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n
\u5206\u7ec4<\/td>\n\u5de5\u5177\u5b9a\u4e49<\/td>\n<\/tr>\n
\u6587\u4ef6\u7cfb\u7edf<\/td>\n– \u6587\u4ef6\u53d1\u73b0\uff1aglob<\/code>\u3001grep<\/code>\u3001ls<\/code>
\n– \u6587\u4ef6\u8bfb\u53d6\uff1aread<\/code>\u3001read_many<\/code>
\n– \u6587\u4ef6\u4fee\u6539\uff1awrite<\/code>\uff08\u5199\u5165\u6574\u4e2a\u65b0\u6587\u4ef6\uff09\uff1bedit<\/code>\uff08\u5b57\u7b26\u4e32\u7cbe\u786e\u5339\u914d\u66ff\u6362\uff09\uff1bmulti_edit<\/code>\uff1bapply_patch<\/code>\uff08\u5e94\u7528\u7ed3\u6784\u5316 patch\/diff\uff09<\/td>\n<\/tr>\n
Shell \u6267\u884c<\/td>\n\u8fd0\u884c\u547d\u4ee4\uff1abash<\/code>\u3001PowerShell<\/code><\/td>\n<\/tr>\n
IO<\/td>\nlsp<\/code><\/p>\n

\uff0c\u4ee5\u53ca\u00a0git_status<\/code>\u3001git_diff<\/code>\u3001git_commit<\/code>\u00a0\u7b49 git \u5de5\u5177<\/td>\n<\/tr>\n

\u5916\u90e8\u4e0a\u4e0b\u6587<\/td>\nMCP \u5de5\u5177\u3001Skills<\/td>\n<\/tr>\n
\u7f51\u7edc\u641c\u7d22<\/td>\nweb_search<\/code><\/p>\n

\u3001web_fetch<\/code>\u3001\u6d4f\u89c8\u5668\u5de5\u5177<\/td>\n<\/tr>\n

\u4ea7\u7269\uff08Artifacts\uff09<\/td>\n\u8bfb\u53d6\u6587\u6863\u3001\u56fe\u7247\uff1b\u751f\u6210 HTML\u3001\u56fe\u7247<\/td>\n<\/tr>\n
\u540e\u53f0\u8fdb\u7a0b<\/td>\n\u5982\uff1aCronCreate<\/code>\u3001CronDelete<\/code>\u3001CronList<\/code><\/td>\n<\/tr>\n
\u667a\u80fd\u4f53\u59d4\u6d3e<\/td>\n\u5982\uff1aspawn_agent<\/code>\u3001resume_agent<\/code>\u3001wait_agent<\/code>\u3001list_agents<\/code>\u3001close_agent<\/code>\u3001interrupt_agent<\/code>\u00a0\u7b49<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

1.5 Harness \u5c42 vs \u6838\u5fc3\u667a\u80fd\uff1f<\/h3>\n

RSI \u7684\u672a\u6765\u5728\u591a\u5927\u7a0b\u5ea6\u4e0a\u4f9d\u8d56 Harness \u5de5\u7a0b\uff0c\u5f88\u96be\u9884\u6d4b\uff1b\u4f46 RSI \u7684\u8fd1\u671f\u8def\u5f84\u5927\u6982\u7387\u4e0d\u4f1a\u4ece”\u6a21\u578b\u76f4\u63a5\u91cd\u5199\u81ea\u8eab\u6743\u91cd”\u5f00\u59cb\u3002\u6211\u5bf9\u53ef\u884c\u7684\u8fd1\u671f\u8def\u5f84\u7684\u9884\u6d4b\u662f\uff1a<\/p>\n

Harness \u5de5\u7a0b\u5c06\u671d\u5143\u65b9\u6cd5\u8bba\uff08meta-methodology\uff09\u65b9\u5411\u6f14\u8fdb\uff08\u5373\u6539\u8fdb”\u83b7\u5f97\u66f4\u597d\u7b54\u6848\u7684\u673a\u5236”\uff0c\u800c\u4e0d\u53ea\u662f\u6539\u8fdb\u7b54\u6848\u672c\u8eab\uff09\u3002Harness \u7cfb\u7edf\u672c\u8eab\u6210\u4e3a\u4f18\u5316\u5bf9\u8c61\uff0c\u542f\u53d1\u5f0f\u89c4\u5219\u66f4\u5c11\uff0c\u901a\u7528\u673a\u5236\u66f4\u591a\u3002<\/p>\n

\u53cd\u8fc7\u6765\uff0c\u6210\u719f\u7684 Harness \u4f7f\u6a21\u578b\u81ea\u6211\u6539\u8fdb\u5faa\u73af\u7684\u81ea\u52a8\u5316\u7814\u7a76\u6210\u4e3a\u53ef\u80fd\uff1b\u800c\u66f4\u806a\u660e\u7684\u6a21\u578b\u80fd\u9632\u6b62 Harness \u8fc7\u5ea6\u8bbe\u8ba1\uff08overengineering\uff09\uff0c\u8ba9\u7cfb\u7edf\u4fdd\u6301\u53ef\u6301\u7eed\u3002<\/p>\n

\u6700\u7ec8\uff0c\u8bb8\u591a Harness \u6539\u8fdb\u5f88\u53ef\u80fd\u88ab\u5185\u5316\u4e3a\u6838\u5fc3\u6a21\u578b\u884c\u4e3a\uff0c\u4f46\u4e0e\u5916\u90e8\u4e0a\u4e0b\u6587\u548c\u5de5\u5177\u7684\u63a5\u53e3\u5e94\u5f53\u4f1a\u4fdd\u7559\u3002\u6211\u4eec\u5728\u63d0\u793a\u8bcd\u5de5\u7a0b\u4e0a\u5df2\u7ecf\u89c1\u8fc7\u8fd9\u4e00\u6a21\u5f0f\u7684\u6e29\u548c\u7248\u672c\uff1a\u968f\u7740\u6307\u4ee4\u5fae\u8c03\u4e0e\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u8fdb\u6b65\uff0c\u624b\u5de5\u63d0\u793a\u8bcd\u6280\u5de7\u53d8\u5f97\u4e0d\u518d\u90a3\u4e48\u6838\u5fc3\uff0c\u4f46\u5bf9\u76ee\u6807\u3001\u7ea6\u675f\u3001\u4e0a\u4e0b\u6587\u4e0e\u8bc4\u4f30\u7684\u663e\u5f0f\u6307\u5b9a\u7684\u9700\u6c42\u5e76\u6ca1\u6709\u6d88\u5931\u3002<\/p>\n

\u4e8c\u3001Harness \u4f18\u5316<\/h2>\n

Harness \u7cfb\u7edf\u4e2d\u88ab\u4f18\u5316\u5bf9\u8c61\u7684\u6f14\u8fdb\u8def\u5f84\u5927\u81f4\u662f\uff1a\u6307\u4ee4\u63d0\u793a\u8bcd\u2192 \u7ed3\u6784\u5316\u4e0a\u4e0b\u6587 \u2192 \u5de5\u4f5c\u6d41 \u2192 Harness \u4ee3\u7801 \u2192 \u4f18\u5316\u5668\u4ee3\u7801\u3002\u968f\u7740\u6a21\u578b\u53d8\u5f97\u66f4\u667a\u80fd\u3001\u66f4\u5f3a\u5927\uff0c\u6211\u4eec\u6b63\u8d70\u5411\u66f4\u590d\u6742\u7684\u4f18\u5316\u76ee\u6807\u4e0e\u66f4\u901a\u7528\u7684\u65b9\u6cd5\u3002<\/p>\n

2.1 \u4e0a\u4e0b\u6587\u5de5\u7a0b<\/h3>\n

\u7b80\u5355\u5730\u628a\u6240\u6709\u5de5\u5177\u54cd\u5e94\u548c\u6a21\u578b\u751f\u6210\u8ffd\u52a0\u8fdb\u4e0a\u4e0b\u6587\uff0c\u4f1a\u968f\u7740\u667a\u80fd\u4f53\u4efb\u52a1\u65f6\u7a0b\u7684\u663e\u8457\u62c9\u957f\u800c\u8fc5\u901f\u5931\u63a7\u3002\u4e0a\u4e0b\u6587\u7ba1\u7406\u662f\u8fd9\u6837\u4e00\u5c42\uff1a\u4e3a LLM \u6784\u5efa\u66f4\u7ed3\u6784\u5316\u3001\u66f4\u7cbe\u7b80\u7684\u4e0a\u4e0b\u6587\uff0c\u5e76\u7ba1\u7406\u6301\u4e45\u72b6\u6001\u3002\u6beb\u65e0\u7591\u95ee\uff0c\u957f\u4e0a\u4e0b\u6587\u7814\u7a76\u4f1a\u7ee7\u7eed\u8fdb\u6b65\uff0c\u4f46\u5f53\u4e0b\u957f\u4e0a\u4e0b\u6587\u667a\u80fd\u4e0e\u4e0a\u4e0b\u6587\u5de5\u7a0b\u6709\u65f6\u662f\u76f8\u4e92\u4ea4\u7ec7\u7684\u3002<\/p>\n

\u667a\u80fd\u4f53\u4e0a\u4e0b\u6587\u5de5\u7a0b\uff08Agentic Context Engineering\uff0cACE\uff1bZhang et al. 2025\uff09\u628a\u4e0a\u4e0b\u6587\u89c6\u4e3a\u4e00\u672c\u4e0d\u65ad\u6f14\u5316\u7684”\u4f5c\u6218\u624b\u518c\uff08playbook\uff09”\uff0c\u800c\u4e0d\u662f\u8d8a\u6765\u8d8a\u957f\u7684\u63d0\u793a\u8bcd\u3002\u5b83\u7528\u4e09\u4e2a\u7ec4\u4ef6\u7ef4\u62a4\u4e00\u4efd\u7531\u8981\u70b9\uff08bullet\uff09\u7ec4\u6210\u7684\u4e0a\u4e0b\u6587\u624b\u518c\uff0c\u6bcf\u4e2a\u8981\u70b9\u5e26\u6709\u6807\u8bc6\u7b26\u4e0e\u63cf\u8ff0\uff1a<\/p>\n

    \n
  • \u751f\u6210\u5668\uff08Generator\uff09\uff1a\u53c2\u8003\u8981\u70b9\uff0c\u4ea7\u51fa\u4efb\u52a1\u8f68\u8ff9\u3002<\/li>\n
  • \u53cd\u601d\u5668\uff08Reflector\uff09\uff1a\u4ece\u6210\u529f\u4e0e\u5931\u8d25\u7684\u8f68\u8ff9\u4e2d\u63d0\u70bc\u6d1e\u89c1\u3002<\/li>\n
  • \u7ba1\u7406\u5668\uff08Curator\uff09\uff1a\u4ee5\u589e\u91cf\u3001\u6761\u76ee\u5316\u7684\u65b9\u5f0f\u66f4\u65b0\u7ed3\u6784\u5316\u4e0a\u4e0b\u6587\u3002<\/li>\n<\/ul>\n

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

    \uff08\u56fe3\uff1aACE \u6846\u67b6\u3002\u56fe\u7247\u6765\u6e90\uff1aZhang et al. 2025\uff09<\/p>\n

    \u4e3a\u9632\u6b62\u8fed\u4ee3\u91cd\u5199\u8fc7\u7a0b\u4e2d\u7684\u4e0a\u4e0b\u6587\u574d\u7f29\uff08context collapse\uff09\u4e0e\u7b80\u6d01\u6027\u504f\u7f6e\uff08brevity bias\uff09\uff0cACE \u7684\u4e00\u4e2a\u5173\u952e\u8bbe\u8ba1\u662f\uff1a\u7ba1\u7406\u5668\u4e0d\u91cd\u5199\u6574\u6bb5\u63d0\u793a\u8bcd\uff0c\u800c\u662f\u8f93\u51fa\u4e00\u7ec4\uff08\u6807\u8bc6\u7b26\uff0c\u63cf\u8ff0\uff09\u5f62\u5f0f\u7684\u7ed3\u6784\u5316\u6761\u76ee\uff0c\u8fd9\u4e9b\u6761\u76ee\u901a\u8fc7\u786e\u5b9a\u6027\u903b\u8f91\u5408\u5e76\u8fdb\u7ed3\u6784\u5316\u7684\u4e0a\u4e0b\u6587\u65e5\u5fd7\u7c3f\u3002\u4e0a\u4e0b\u6587\u6761\u76ee\u4f1a\u88ab\u5b9a\u671f\u7cbe\u70bc\u4e0e\u53bb\u91cd\u3002<\/p>\n

    ACE \u4ece rollout \u4e2d\u5b66\u4e60\u6d1e\u89c1\uff0c\u5e2e\u52a9\u6211\u4eec\u8fc8\u5411\u81ea\u7ba1\u7406\u8bb0\u5fc6\uff1b\u4f46\u5176\u66f4\u65b0\u89c4\u5219\u4e0e\u6574\u4f53\u5de5\u4f5c\u6d41\u4ecd\u662f\u624b\u5de5\u6253\u9020\u7684\u3002\u4e3a\u4e86\u8d70\u5411\u66f4\u81ea\u6211\u6539\u8fdb\u7684\u5faa\u73af\uff0c\u5143\u4e0a\u4e0b\u6587\u5de5\u7a0b\uff08Meta Context Engineering\uff0cMCE\uff1bYe et al. 2026\uff09\u628a\u673a\u5236\uff08\u5982\u4f55\u7ba1\u7406\u4e0a\u4e0b\u6587\uff09\u4e0e\u4ea7\u7269\u5185\u5bb9\uff08\u4e0a\u4e0b\u6587\u91cc\u6709\u4ec0\u4e48\uff09\u5206\u79bb\u5f00\u6765\uff1a\u5728\u5143\u4f18\u5316\u5c42\u505a\u6280\u80fd\u8fdb\u5316\uff08skill evolution\uff09\uff0c\u5728\u57fa\u7840\u5c42\u505a\u4e0a\u4e0b\u6587\u4f18\u5316\u3002<\/p>\n

    \u4e00\u4e2a MCE \u6280\u80fd s \u2208 \ud835\udcae \u5b9a\u4e49\u4e00\u4e2a\u4e0a\u4e0b\u6587\u51fd\u6570 c_s = (\u03c1_s, F_s)\uff0c\u628a\u8f93\u5165 x \u6620\u5c04\u4e3a\u4e0a\u4e0b\u6587 c = F_s(x; \u03c1_s)\uff0c\u5176\u4e2d\uff1a<\/p>\n

      \n
    • \u03c1_s = {\u03c1_1, \u2026, \u03c1_m} \u4e3a\u9759\u6001\u7ec4\u4ef6\uff08\u63d0\u793a\u8bcd\u3001\u77e5\u8bc6\u5e93\u3001\u4ee3\u7801\u5e93\uff09\u3002<\/li>\n
    • F_s = {F_1, \u2026, F_k} \u4e3a\u52a8\u6001\u7b97\u5b50\uff08\u641c\u7d22\u3001\u9009\u62e9\u3001\u8fc7\u6ee4\u3001\u683c\u5f0f\u5316\uff09\u3002<\/li>\n<\/ul>\n

      \u53cc\u5c42\u4f18\u5316\uff08bi-level optimization\uff09\u662f\u6307\uff1a\u5185\u5c42\u5728\u7ed9\u5b9a\u6280\u80fd s \u7684\u8bad\u7ec3\u6570\u636e\u4e0a\u5bfb\u627e\u6700\u4f18\u4e0a\u4e0b\u6587 c_s*\uff0c\u5916\u5c42\u5728\u9a8c\u8bc1\u96c6\u4e0a\u5bfb\u627e\u8868\u73b0\u6700\u4f18\u7684\u6280\u80fd\uff1a<\/p>\n

        \n
      • \u5185\u5c42\uff1ac_s* = arg max_{c_s} J_train(c_s; s)<\/li>\n
      • \u5916\u5c42\uff1as* = arg max_{s \u2208 \ud835\udcae} J_val(c_s*)<\/li>\n<\/ul>\n

        \u6280\u80fd\u6570\u636e\u5e93\u8bb0\u5f55\u5386\u53f2\u6280\u80fd\u3001\u4e0a\u4e0b\u6587\u51fd\u6570\u4e0e\u8bc4\u4f30\u6307\u6807\uff1aH_{k-1} = {(s_i, c_i, J_i^train, J_i^val)}{i=1}^{k-1}\u3002\u4e00\u4e2a\u5143\u5c42\u667a\u80fd\u4f53\u9488\u5bf9\u4efb\u52a1 \u03c4\uff0c\u5bf9\u5386\u53f2\u6280\u80fd\u6267\u884c\u667a\u80fd\u4f53\u5f0f\u4ea4\u53c9\uff08agentic crossover\uff09\u4ee5\u521b\u5efa\u65b0\u6280\u80fd\uff1as_k = crossover(\u03c4, H{k-1})\u3002<\/p>\n

        \u968f\u540e\uff0c\u57fa\u7840\u5c42\u7684\u4e0a\u4e0b\u6587\u5de5\u7a0b\u5e08\u6267\u884c\u6280\u80fd s_k\uff0c\u5e76\u5728\u5f53\u524d\u6280\u80fd\u6307\u5bfc\u4e0b\u4ece rollout \u53cd\u9988 R_k \u4e2d\u5b66\u4e60\u4e0a\u4e0b\u6587\u51fd\u6570\uff1ac_k = engineer(\u03c4, s_k; c*_{k-1}, R_k)\u3002<\/p>\n

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

        \uff08\u56fe4\uff1aMCE \u6846\u67b6\uff1a\u5143\u5c42\u6280\u80fd\u8fdb\u5316\u5728\u4e0a\u4e0b\u6587\u7ba1\u7406\u673a\u5236\u4e0a\u641c\u7d22\uff0c\u57fa\u7840\u5c42\u4f18\u5316\u4efb\u52a1\u4e0a\u4e0b\u6587\u3002\u56fe\u7247\u6765\u6e90\uff1aYe et al. 2026\uff09<\/p>\n

        MCE \u4e0d\u50cf ACE \u90a3\u6837\u5f3a\u5236\u89c4\u5b9a”\u4e0a\u4e0b\u6587\u8be5\u5982\u4f55\u7ec4\u7ec7”\u7684\u542f\u53d1\u5f0f\u89c4\u5219\u3002\u5b83\u7528\u81ea\u7531\u5f62\u5f0f\u7684\u6280\u80fd\u5b58\u50a8\u4efb\u52a1\u6700\u91cd\u8981\u7684\u77e5\u8bc6\uff0c\u5e76\u8ba9\u6280\u80fd\u4e0e\u6280\u80fd\u6761\u4ef6\u4e0b\u7684\u4e0a\u4e0b\u6587\u4e00\u8d77\u8fed\u4ee3\u8fdb\u5316\u3002\u5b9e\u73b0\u4e0a\uff0c\u4e00\u4e2a\u4e0a\u4e0b\u6587\u51fd\u6570 c \u88ab\u5b9e\u4f8b\u5316\u4e3a\u4e13\u7528\u76ee\u5f55\u4e0b\u7684\u4e00\u7ec4\u6587\u4ef6\uff0c\u5305\u542b\u9759\u6001\u7ec4\u4ef6\uff08skill.md\uff09\u4e0e\u52a8\u6001\u7ec4\u4ef6\uff08\u4e0a\u4e0b\u6587\u4e0e\u6570\u636e rollout\uff09\u3002\u5143\u5c42\u4e0e\u57fa\u7840\u5c42\u4f18\u5316\u90fd\u5728\u667a\u80fd\u4f53\u7f16\u7a0b\u73af\u5883\u4e2d\u6267\u884c\uff0c\u4f7f\u7528\u6807\u51c6\u5de5\u5177\u96c6\uff1a<\/p>\n

        \ud835\udcaf = {Read, Write, Edit, Bash, Glob, Grep, TodoWrite}<\/p>\n

        Meta-Harness\uff08Lee et al. 2026\uff09\u53c8\u6df1\u5165\u4e86\u4e00\u5c42\uff1a\u88ab\u4f18\u5316\u7684\u5bf9\u8c61\u662f\u4ee3\u7801\u672c\u8eab\uff0c\u5373\u51b3\u5b9a”\u54ea\u4e9b\u4fe1\u606f\u5e94\u88ab\u5b58\u50a8\u3001\u68c0\u7d22\u5e76\u5448\u73b0\u7ed9\u6a21\u578b”\u5e76\u5bf9\u6b64\u8fdb\u884c\u4f18\u5316\u7684\u4ee3\u7801\u3002\u540d\u5b57\u91cc\u7684”Meta-“\u610f\u5473\u7740\uff1a\u5b83\u662f\u7528\u4e8e\u4f18\u5316 Harness \u7684 Harness\u3002<\/p>\n

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

        2.2 \u5de5\u4f5c\u6d41\u8bbe\u8ba1<\/h3>\n

        Harness \u5de5\u7a0b\u4e2d\u7684\u5de5\u4f5c\u6d41\u8bbe\u8ba1\u53ef\u4ee5\u7531\u9886\u57df\u4e13\u5bb6\u624b\u5de5\u5b8c\u6210\u3002\u4ee5\u81ea\u52a8\u5316\u7814\u7a76\u4e3a\u4f8b\uff0c\u5df2\u6709\u591a\u4e2a\u6846\u67b6\u88ab\u63d0\u51fa\u5e76\u9a8c\u8bc1\u3002AI Scientist\u00a0\u7cfb\u7edf\uff08Lu et al. 2026\uff09\u6784\u5efa\u4e86\u4e00\u6761\u6d41\u6c34\u7ebf\uff1a\u63d0\u51fa\u7814\u7a76\u60f3\u6cd5\u3001\u5199\u4ee3\u7801\u3001\u8dd1\u5b9e\u9a8c\u3001\u5206\u6790\u7ed3\u679c\u3001\u64b0\u5199\u8bba\u6587\u624b\u7a3f\u5e76\u6267\u884c\u540c\u884c\u8bc4\u5ba1\u3002Meng et al.\uff082026\uff09\u5219\u628a”\u53ef\u9a8c\u8bc1\u6027”\u4f5c\u4e3aScientistOne\u00a0\u7684\u6838\u5fc3\u8bbe\u8ba1\u7ea6\u675f\uff1a\u6bcf\u4e00\u6761\u8bba\u65ad\uff08\u5f15\u7528\u3001\u6570\u503c\u3001\u65b9\u6cd5\u3001\u7ed3\u8bba\uff09\u90fd\u5fc5\u987b\u53ef\u8ffd\u6eaf\u81f3\u8bc1\u636e\u6765\u6e90\uff0c\u5e76\u7531”\u8bc1\u636e\u94fe\uff08Chain-of-Evidence\uff09”\u68c0\u67e5\u8fdb\u884c\u5ba1\u8ba1\u3002<\/p>\n

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

        \uff08\u56fe5\uff1aAI Scientist \u7684\u6d41\u6c34\u7ebf\uff1a\u60f3\u6cd5\u751f\u6210\u3001\u5b9e\u9a8c\u3001\u8bba\u6587\u5199\u4f5c\u4e0e\u8bc4\u5ba1\u3002\u56fe\u7247\u6765\u6e90\uff1aLu et al. 2026\uff09<\/p>\n

        Autodata\u00a0\u667a\u80fd\u4f53\uff08Kulikov et al. 2026\uff09\u88ab\u8bbe\u8ba1\u4e3a\u4e00\u540d\u6570\u636e\u79d1\u5b66\u5bb6\uff0c\u8d1f\u8d23\u751f\u6210\u8bad\u7ec3\u4e0e\u8bc4\u4f30\u6570\u636e\u3002\u4e3b\u667a\u80fd\u4f53\u7ba1\u7406\u7740\u4e00\u4e2a\u63d0\u51fa\u95ee\u9898\u7684\u6311\u6218\u8005\uff08challenger\uff09\u3001\u4e00\u4e2a\u5f31\u6c42\u89e3\u5668\uff08weak solver\uff09\u3001\u4e00\u4e2a\u5f3a\u6c42\u89e3\u5668\uff08strong solver\uff09\u548c\u4e00\u4e2a\u9a8c\u8bc1\u5668\/\u88c1\u5224\uff08verifier\/judge\uff09\uff0c\u76ee\u6807\u662f\u5408\u6210\u96be\u5ea6”\u521a\u521a\u597d”\u7684\u6570\u636e\uff0c\u5373\u5f3a\u6c42\u89e3\u5668\u80fd\u89e3\u51fa\u3001\u800c\u5f31\u6c42\u89e3\u5668\u89e3\u4e0d\u51fa\u3002<\/p>\n

        \u5728 Autodata \u4e2d\uff0c\u6311\u6218\u8005\u7684\u63d0\u793a\u8bcd\u4f1a\u6839\u636e\u6c42\u89e3\u5668\u4e0e\u9a8c\u8bc1\u5668\u7684\u53cd\u9988\u8fed\u4ee3\u66f4\u65b0\u3002\u5176\u5c40\u9650\u5728\u4e8e\uff1a\u5408\u6210\u4efb\u52a1\u88ab\u7528\u4e8e\u5fae\u8c03\u5f31\u6c42\u89e3\u5668\u800c\u975e\u5f3a\u6c42\u89e3\u5668\uff1b\u5982\u679c\u8fd9\u4e2a\u5faa\u73af\u65e0\u6cd5\u8fed\u4ee3\u6539\u8fdb\u5f3a\u6a21\u578b\uff0c\u90a3\u5b83\u66f4\u50cf\u662f\u5728\u751f\u6210\u7684\u63d0\u793a\u8bcd\u5206\u5e03\u4e0a\u505a\u95f4\u63a5\u84b8\u998f\uff0cRSI \u7684\u610f\u5473\u5c31\u6de1\u4e86\u4e9b\u3002<\/p>\n

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

        \uff08\u56fe6\uff1aAutodata \u7684\u667a\u80fd\u4f53\u5de5\u4f5c\u6d41\u8bbe\u8ba1\uff0c\u56f4\u7ed5\u6311\u6218\u8005\u3001\u6c42\u89e3\u5668\u4e0e\u9a8c\u8bc1\u5668\u89d2\u8272\u751f\u6210\u5408\u6210\u8bad\u7ec3\u4e0e\u8bc4\u4f30\u6570\u636e\u3002\u56fe\u7247\u6765\u6e90\uff1aKulikov et al. 2026\uff09<\/p>\n

        \u5de5\u4f5c\u6d41\u7684\u8bbe\u8ba1\u7a7a\u95f4\u662f\u5de8\u5927\u7684\u3002\u6211\u4eec\u81ea\u7136\u53ef\u4ee5\u628a\u5de5\u4f5c\u6d41\u8bbe\u8ba1\u770b\u4f5c\u4e00\u4e2a\u641c\u7d22\u95ee\u9898\uff0c\u4ece\u800c\u5e94\u5f53\u80fd\u7528\u7b97\u6cd5\u627e\u5230\u597d\u7684\u65b9\u6848\uff0c\u800c\u975e\u53ea\u80fd\u624b\u5de5\u6253\u9020\u3002\u6cbf\u8fd9\u4e00\u65b9\u5411\uff0c\u667a\u80fd\u4f53\u7cfb\u7edf\u81ea\u52a8\u5316\u8bbe\u8ba1\uff08Automated Design of Agentic Systems\uff0cADAS\uff1bHu et al. 2025\uff09\u628a\u667a\u80fd\u4f53\u8bbe\u8ba1\u672c\u8eab\u5f62\u5f0f\u5316\u4e3a\u4e00\u4e2a\u4f18\u5316\u95ee\u9898\uff0c”\u5143\u667a\u80fd\u4f53\u641c\u7d22\uff08meta-agent search\uff09”\uff1a\u7531\u4e00\u4e2a\u5143\u667a\u80fd\u4f53\u63d0\u51fa\u65b0\u7684\u667a\u80fd\u4f53\u5de5\u4f5c\u6d41\u8bbe\u8ba1\u3002<\/p>\n

          \n
        • \u7528 CoT\u3001self-refine \u7b49\u7b80\u5355\u667a\u80fd\u4f53\u521d\u59cb\u5316\u4e00\u4e2a\u5de5\u4f5c\u6d41\u6863\u6848\u5e93\u3002<\/li>\n
        • \u8ba9\u5143\u667a\u80fd\u4f53\u4ee5\u4ee3\u7801\u5f62\u5f0f\u7f16\u5199\u65b0\u667a\u80fd\u4f53\uff0c\u7075\u611f\u6765\u81ea\u6863\u6848\u5e93\u4e2d\u7684\u73b0\u6709\u65b9\u6848\u3002<\/li>\n<\/ul>\n
            \n
          • \u5143\u667a\u80fd\u4f53\u5148\u751f\u6210\u65b0\u5de5\u4f5c\u6d41\u7684\u9ad8\u5c42\u63cf\u8ff0\uff0c\u518d\u7528\u4ee3\u7801\u5b9e\u73b0\u3002<\/li>\n<\/ul>\n
              \n
            • \u8349\u7a3f\u7a0b\u5e8f\u518d\u7ecf\u8fc7\u4e24\u8f6e\u81ea\u6211\u7cbe\u70bc\uff08self-refine\uff0c\u5373\u8ba9\u6a21\u578b\u7ed9\u51fa\u53cd\u9988\uff0c\u518d\u8ba9\u540c\u4e00\u6a21\u578b\u57fa\u4e8e\u53cd\u9988\u6539\u8fdb\u4e4b\u524d\u7684\u8f93\u51fa\uff1bMadaan et al. 2023\uff09\uff0c\u7531\u5143\u667a\u80fd\u4f53\u68c0\u67e5\u5176\u65b0\u9896\u6027\u3002<\/li>\n
            • \u8bc4\u4f30\u6bcf\u4e2a\u65b0\u5019\u9009\uff0c\u628a\u6210\u529f\u8005\u52a0\u56de\u6863\u6848\u5e93\u3002<\/li>\n
            • \u91cd\u590d\u6b65\u9aa4 2\u20133\uff0c\u76f4\u5230\u8fbe\u5230\u6700\u5927\u8fed\u4ee3\u6b21\u6570\u3002
              \n\"\u56fe\u7247\"<\/li>\n<\/ul>\n

              \uff08\u56fe7\uff1aADAS \u793a\u610f\u3002\u56fe\u7247\u6765\u6e90\uff1aHu et al. 2025\uff09<\/p>\n

              AFlow\uff08Zhang et al. 2025\uff09\u628a\u667a\u80fd\u4f53\u5de5\u4f5c\u6d41\u8868\u793a\u4e3a\u4e00\u5f20\u56fe\uff1a\u8282\u70b9\u4ee3\u8868\u8c03\u7528 LLM \u7684\u52a8\u4f5c\uff0c\u8fb9\u7528\u4ee3\u7801\u5b9e\u73b0\u903b\u8f91\u64cd\u4f5c\u3002\u5176\u5de5\u4f5c\u6d41\u4f18\u5316\u4f9d\u8d56\u00a0MCTS\uff08\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff09\uff1a<\/p>\n

                \n
              • \u7528\u6a21\u677f\u521d\u59cb\u5316\u8d77\u59cb\u5de5\u4f5c\u6d41 W_0\uff0c\u4f5c\u4e3a\u6811\u7684\u6839\u3002<\/li>\n
              • \u6309”\u5f97\u5206 + \u5747\u5300\u63a2\u7d22”\u7684\u8f6f\u6df7\u5408\u7b56\u7565\u9009\u62e9\u4e00\u4e2a\u5de5\u4f5c\u6d41\u8282\u70b9\u3002<\/li>\n
              • \u8ba9 LLM \u57fa\u4e8e\u8be5\u8282\u70b9\u7684\u8bc4\u4f30\u8868\u73b0\u751f\u6210\u4e00\u4e2a\u4fee\u6539\u540e\u7684\u5de5\u4f5c\u6d41\uff0c\u5b8c\u6210\u6269\u5c55\u3002<\/li>\n
              • \u6267\u884c\u5e76\u8bc4\u4f30\u65b0\u5de5\u4f5c\u6d41\u3002<\/li>\n
              • \u82e5\u65b0\u5de5\u4f5c\u6d41\u5728 N \u8f6e\u9884\u7b97\u5185\u6709\u6539\u8fdb\uff0c\u5219\u52a0\u56de\u6811\u4e2d\u3002<\/li>\n
              • \u91cd\u590d\u6b65\u9aa4 2\u20135\uff0c\u5f53 top-k \u5e73\u5747\u5206\u8fdb\u5165\u5e73\u53f0\u671f\u6216\u9884\u7b97\u8017\u5c3d\u65f6\u505c\u6b62\u3002
                \n\"\u56fe\u7247\"<\/li>\n<\/ul>\n

                \uff08\u56fe8\uff1aAFlow \u5728\u5de5\u4f5c\u6d41\u5019\u9009\u6811\u4e0a\u7684\u4f18\u5316\u8fc7\u7a0b\u3002\u56fe\u7247\u6765\u6e90\uff1aZhang et al. 2025\uff09<\/p>\n

                AFlow \u5728\u95ee\u7b54\u3001\u4ee3\u7801\u4e0e\u6570\u5b66\u4efb\u52a1\u4e0a\u7684\u5b9e\u9a8c\u663e\u793a\uff0c\u5b83\u76f8\u6bd4\u4eba\u5de5\u8bbe\u8ba1\u7684\u5de5\u4f5c\u6d41\u548c ADAS \u90fd\u6709\u53ef\u89c2\u7684\u63d0\u5347\u3002<\/p>\n

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

                \uff08\u56fe9\uff1aAFlow \u4e0e\u4eba\u5de5\u65b9\u6cd5\u3001ADAS \u7684\u5b9e\u9a8c\u5bf9\u6bd4\u3002\u56fe\u7247\u6765\u6e90\uff1aZhang et al. 2025\uff09<\/p>\n

                2.3 \u81ea\u6211\u6539\u8fdb\u7684 Harness<\/h3>\n

                \u65e0\u8bba\u4e0a\u4e0b\u6587\u5de5\u7a0b\u8fd8\u662f\u5de5\u4f5c\u6d41\u8bbe\u8ba1\uff0c\u90fd\u53ea\u662f Harness \u7684\u4e00\u90e8\u5206\u3002\u6211\u4eec\u9700\u8981\u5728\u6574\u4e2a\u8bbe\u8ba1\u7a7a\u95f4\u4e2d\u641c\u7d22\uff0c\u628a\u4e0a\u4e0b\u6587\u7ba1\u7406\u903b\u8f91\u3001\u5de5\u4f5c\u6d41\u3001\u6743\u9650\u4ee5\u53ca\u8bb8\u591a\u5176\u4ed6 Harness \u7ec4\u4ef6\u653e\u5728\u4e00\u8d77\u4f18\u5316\u3002\u6b63\u5982 Meta-Harness\u3001ADAS\u3001AFlow \u7b49\u5de5\u4f5c\u4e2d\u6240\u89c1\uff0c\u2728\u4ee3\u7801\u2728\u662f\u5b9a\u4e49\u7a0b\u5e8f\u4e0e\u7cfb\u7edf\u7684\u901a\u7528\u8bed\u8a00\u3002\u7b80\u8a00\u4e4b\uff0cHarness \u5c31\u662f\u7528\u4ee3\u7801\u6765\u7f16\u6392\u63d0\u793a\u8bcd\u3001\u5de5\u5177\u8c03\u7528\u3001\u5b50\u667a\u80fd\u4f53\u3001\u63a7\u5236\u6d41\u3001\u8bb0\u5fc6\u4e0e\u5de5\u4f5c\u6d41\u903b\u8f91\u5982\u4f55\u534f\u540c\u3002\u5982\u679c LLM \u80fd\u4f18\u5316\u6267\u884c\u667a\u80fd\u4f53\u7684\u4ee3\u7801\uff0c\u5b83\u5c31\u80fd\u89e6\u53ca\u6bd4\u624b\u5199\u63d0\u793a\u8bcd\u5927\u5f97\u591a\u7684\u8bbe\u8ba1\u7a7a\u95f4\u3002<\/p>\n

                \u81ea\u6559\u81ea\u5b66\u4f18\u5316\u5668\uff08Self-Taught Optimizer\uff0cSTOP\uff1bZelikman et al. 2023\uff09\u662f\u9012\u5f52\u5f0f\u811a\u624b\u67b6\uff08scaffolding\uff09\u6539\u8fdb\u7684\u65e9\u671f\u4f8b\u5b50\u3002\u5728 t=0 \u65f6\uff0c\u79cd\u5b50\u6539\u8fdb\u5668 I_0 \u63a5\u6536\u521d\u59cb\u89e3 s\u3001\u6548\u7528\u51fd\u6570 u \u548c\u9ed1\u76d2\u8bed\u8a00\u6a21\u578b M\uff0c\u8fd4\u56de\u6539\u8fdb\u540e\u7684\u89e3 s\u2032\uff0c\u5373 s\u2032 = I(u, s; M)\u3002STOP \u7684\u76ee\u6807\u4e0d\u662f\u76f4\u63a5\u6539\u8fdb s\uff0c\u800c\u662f\u6539\u8fdb\u6539\u8fdb\u5668 I \u672c\u8eab\u3002<\/p>\n

                \u9996\u5148\uff0c\u628a\u5143\u6548\u7528\uff08meta-utility\uff09\u5b9a\u4e49\u4e3a\u7ed9\u5b9a\u6539\u8fdb\u5668\u51fd\u6570 I \u5728\u4e00\u7ec4\u4e0b\u6e38\u4efb\u52a1 \ud835\udc9f \u4e0a\u7684\u5e73\u5747\u6548\u7528\uff1a<\/p>\n

                \u00fb(I) \u225c (1\/|\ud835\udc9f|) \u00b7 \ud835\udd3c_{(u,s)~\ud835\udc9f}[u(I(u, s; M))]<\/p>\n

                \u7531\u4e8e\u6539\u8fdb”\u6539\u8fdb\u5668\u51fd\u6570”\u672c\u8eab\u4e5f\u662f\u4e00\u4e2a\u4f18\u5316\u95ee\u9898\uff0c\u6211\u4eec\u53ef\u4ee5\u57fa\u4e8e I_{t-1} \u5728\u5143\u6548\u7528\u4e0b\u7684\u8868\u73b0\uff0c\u9012\u5f52\u5730\u5f97\u5230 I \u7684\u65b0\u7248\u672c\uff0c\u5373\u81ea\u6211\u6539\u8fdb\u66f4\u65b0\uff1a<\/p>\n

                I_t = I_{t-1}(\u00fb, I_{t-1}; M)<\/p>\n

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

                \uff08\u56fe10\uff1aSTOP \u7b97\u6cd5\u3002\u56fe\u7247\u6765\u6e90\uff1aZelikman et al. 2023\uff09<\/p>\n

                \u5b9e\u9a8c\u4e2d\uff0c\u8fdb\u5316\u540e\u7684\u6539\u8fdb\u5668\u53d1\u73b0\u4e86\u591a\u79cd\u7b56\u7565\uff0c\u5982\u9057\u4f20\u7b97\u6cd5\u3001\u5206\u89e3\u5e76\u6539\u8fdb\u5c40\u90e8\u3001\u591a\u81c2\u63d0\u793a\u8bcd\u8001\u864e\u673a\uff08multi-armed prompt bandits\uff09\u3001\u6a21\u62df\u9000\u706b\u3001\u8c03\u8282\u6e29\u5ea6\u3001\u675f\u641c\u7d22\/\u6811\u641c\u7d22\u7b49\u3002\u8fd9\u7c7b\u4f3c\u4e8e\u628a Harness \u5de5\u4f5c\u6d41\u8868\u793a\u4e3a\u53ef\u4f18\u5316\u7684\u5bf9\u8c61\u3002<\/p>\n

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

                \uff08\u56fe11\uff1aSTOP \u53d1\u73b0\u7684\u81ea\u6211\u6539\u8fdb\u7b56\u7565\u793a\u4f8b\u3002\u56fe\u7247\u6765\u6e90\uff1aZelikman et al. 2023\uff09<\/p>\n

                Zelikman et al.\uff082023\uff09\u7684\u4e00\u4e2a\u8b66\u793a\u6027\u7ed3\u679c\u662f\uff1aSTOP \u5728 GPT-4 \u4e0b\u80fd\u9010\u8f6e\u63d0\u5347\u4e0b\u6e38\u5e73\u5747\u8868\u73b0\uff0c\u4f46\u5728 GPT-3.5\u3001Mixtral \u7b49\u8f83\u5f31\u6a21\u578b\u4e0b\u53cd\u800c\u9000\u5316\u3002\u4ec5\u6709\u9012\u5f52\u7ed3\u6784\u662f\u4e0d\u591f\u7684\uff0c\u57fa\u7840\u6a21\u578b\u5fc5\u987b\u8db3\u591f\u6709\u80fd\u529b\u624d\u80fd\u6539\u8fdb\u673a\u5236\u3002\u8fd9\u610f\u5473\u7740\uff1aHarness \u6539\u8fdb\u80fd\u8ba9\u6a21\u578b\u90e8\u7f72\u5f97\u66f4\u597d\uff0c\u4f46\u667a\u80fd\u4ecd\u662f\u6838\u5fc3\u3002<\/p>\n

                Lin et al.\uff082026\uff09\u66f4\u7ec6\u81f4\u5730\u8003\u5bdf\u4e86 Harness \u8fdb\u5316\u5bf9\u6a21\u578b\u80fd\u529b\u7684\u4f9d\u8d56\u3002\u4ed6\u4eec\u62c6\u89e3\u51fa\u4e24\u4e2a\u8f74\uff1a(1)Harness \u66f4\u65b0\u80fd\u529b\uff08harness-updating\uff09\uff0c\u4ea7\u51fa\u6709\u7528 Harness \u4fee\u6539\u7684\u80fd\u529b\uff1b(2)Harness \u53d7\u76ca\u80fd\u529b\uff08harness-benefit\uff09\uff0c\u5229\u7528\u66f4\u65b0\u540e\u7684 Harness \u66f4\u597d\u5730\u89e3\u51b3\u4efb\u52a1\u7684\u80fd\u529b\u3002\u6709\u8da3\u7684\u662f\uff0c\u5b9e\u9a8c\u4e2d\u4ece Qwen3.5-9B \u5230 Claude Opus 4.6\u3001\u4e0d\u540c\u89c4\u6a21\u4e0e\u6838\u5fc3\u667a\u80fd\u7684\u4e00\u7cfb\u5217\u6a21\u578b\uff0c\u8868\u73b0\u51fa\u76f8\u8fd1\u7684 Harness \u66f4\u65b0\u80fd\u529b\uff1a9B \u7684 Harness \u63d0\u8bae\u8005\/\u8fdb\u5316\u8005\u5199\u51fa\u7684\u6280\u80fd\uff0c\u4e0e Opus \u5728\u8fc7\u7a0b\u4e0a\u51e0\u4e4e\u540c\u6784\u3002\u800c\u8981\u5145\u5206\u5229\u7528 Harness\uff0c\u6a21\u578b\u9700\u8981\u6b63\u786e\u4e14\u9002\u65f6\u5730\u8c03\u7528\u6280\u80fd\/\u5de5\u5177\uff0c\u5e76\u64c5\u957f\u957f\u7a0b\u6307\u4ee4\u9075\u5faa\u3002<\/p>\n

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

                \uff08\u56fe12\uff1a\u4e3b\u8981\u7ed3\u679c (A) \u4ece Qwen2-32B \u5230 Opus 4.6\uff0cHarness \u66f4\u65b0\u80fd\u529b\u5927\u4f53\u6301\u5e73\uff1b(B) Harness \u53d7\u76ca\u80fd\u529b\u5448\u975e\u5355\u8c03\uff0c\u4e2d\u95f4\u6863\u6a21\u578b\u53d7\u76ca\u6700\u5927\u3002\u56fe\u7247\u6765\u6e90\uff1aLin et al. 2026\uff09<\/p>\n

                \u66f4\u8fd1\u7684\u5de5\u4f5c\u00a0Self-Harness\uff08Zhang et al. 2026\uff09\u4f9d\u9760 LLM \u667a\u80fd\u4f53\u901a\u8fc7”\u63d0\u8bae\u2014\u8bc4\u4f30\u2014\u63a5\u53d7\uff08propose-evaluate-accept\uff09”\u5faa\u73af\u6765\u6539\u8fdb\u81ea\u5df1\u7684 Harness\u3002<\/p>\n

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

                \uff08\u56fe13\uff1aSelf-Harness \u4f7f\u7528”\u5f31\u70b9\u6316\u6398\u2014\u6709\u754c Harness \u63d0\u8bae\u2014\u9a8c\u8bc1”\u7684\u5faa\u73af\u6765\u66f4\u65b0 Harness\u3002\u56fe\u7247\u6765\u6e90\uff1aZhang et al. 2026\uff09<\/p>\n

                Self-Harness \u7684\u5faa\u73af\u6709\u4e09\u4e2a\u9636\u6bb5\uff1a<\/p>\n

                  \n
                • \u5f31\u70b9\u6316\u6398\uff08Weakness mining\uff09\uff1a\u628a\u5931\u8d25\u805a\u7c7b\u4e3a”\u6709\u9a8c\u8bc1\u5668\u4f9d\u636e”\u7684\u5931\u8d25\u6a21\u5f0f\u3002<\/li>\n<\/ul>\n

                  a.\u7528\u5f53\u524d Harness h_t \u5728\u4efb\u52a1\u4e0a\u8bc4\u4f30\uff0c\u6536\u96c6\u6267\u884c\u8f68\u8ff9\u7528\u4e8e\u5206\u6790\u3002<\/p>\n

                  b.\u6ce8\u610f\uff1a\u4e24\u6b21\u8fd0\u884c\u5728\u9519\u8bef\u65e5\u5fd7\u8868\u9762\u4e0a\u53ef\u80fd\u6709\u7740\u76f8\u540c\u7684\u9a8c\u8bc1\u5668\u7ed3\u679c\uff08\u5982\u8d85\u65f6\u3001\u4ea7\u7269\u7f3a\u5931\uff09\uff0c\u5374\u6709\u7740\u4e0d\u540c\u7684\u56e0\u679c\u673a\u5236\u3002\u56e0\u6b64\u9700\u8981\u4fe1\u606f\u4e30\u5bcc\u7684\u5931\u8d25\u8bb0\u5f55\uff0c\u5305\u542b\u6700\u7ec8\u7684\u9a8c\u8bc1\u5668\u5c42\u539f\u56e0\u3001\u76f8\u5173\u667a\u80fd\u4f53\u884c\u4e3a\u7684\u56e0\u679c\u72b6\u6001\u3001\u4ee5\u53ca\u8f68\u8ff9\u66b4\u9732\u51fa\u7684\u62bd\u8c61\u667a\u80fd\u4f53\u673a\u5236\uff0c\u4ee5\u63ed\u793a\u6839\u672c\u539f\u56e0\u3002<\/p>\n

                    \n
                  • Harness \u63d0\u8bae\uff08Harness proposal\uff09\uff1a\u57fa\u4e8e\u6316\u6398\u51fa\u7684\u5931\u8d25\u6a21\u5f0f\uff0c\u63d0\u51fa\u6709\u754c\u7684 Harness \u4fee\u6539\u3002<\/li>\n<\/ul>\n

                    a.\u540c\u4e00\u6a21\u578b\u5728 h_t \u4e0b\u88ab\u8c03\u7528\uff0c\u5145\u5f53\u63d0\u8bae\u8005\u3002<\/p>\n

                    b.\u6a21\u578b\u83b7\u5f97\u4e00\u4e2a\u6709\u754c\u7684\u63d0\u8bae\u4e0a\u4e0b\u6587\uff1a(1) \u5f53\u524d Harness \u7684\u53ef\u7f16\u8f91\u9762\uff1b(2) \u6765\u81ea\u8bc4\u4f30\u7cfb\u7edf\u7684\u3001\u6709\u9a8c\u8bc1\u5668\u4f9d\u636e\u7684\u5931\u8d25\u6a21\u5f0f\uff1b(3) \u5e94\u5f53\u4fdd\u7559\u7684\u901a\u8fc7\u884c\u4e3a\u8bb0\u5f55\uff1b(4) \u8fc7\u5f80\u5df2\u5c1d\u8bd5\u4fee\u6539\u7684\u6458\u8981\u3002<\/p>\n

                    c.Harness \u4fee\u6539\u5e94\u4f18\u5148\u9488\u5bf9\u53cd\u590d\u51fa\u73b0\u3001\u4e14\u53ef\u89e3\u51b3\u7684\u9519\u8bef\u6a21\u5f0f\uff08\u800c\u975e\u4efb\u52a1\u672c\u8eab\u56fa\u6709\u7684\u96be\u5ea6\uff09\uff0c\u4e14\u80fd\u901a\u8fc7\u7a84\u6539\u52a8\u89e3\u51b3\u3002<\/p>\n

                    d.Harness \u4fee\u6539\u5019\u9009\u4e4b\u95f4\u5e94\u5f53\u6709\u533a\u5206\u5ea6\u3001\u6709\u591a\u6837\u6027\u3002<\/p>\n

                      \n
                    • \u63d0\u8bae\u9a8c\u8bc1\uff08Proposal validation\uff09\uff1a\u9a8c\u8bc1\u5e76\u5408\u5e76\u5408\u683c\u4fee\u6539\uff0c\u751f\u6210\u65b0 Harness h_{t+1}\u3002<\/li>\n<\/ul>\n

                      a.\u5019\u9009\u4fee\u6539\u5728 held-in \u96c6 D_in\uff08\u6d4b\u8bd5\u5f31\u70b9\u662f\u5426\u89e3\u51b3\uff09\u4e0e held-out \u96c6 D_out\uff08\u68c0\u67e5\u662f\u5426\u5f15\u5165\u5176\u4ed6\u672a\u77e5\u95ee\u9898\uff09\u4e0a\u505a\u56de\u5f52\u6d4b\u8bd5\u3002<\/p>\n

                      b.\u53ea\u6709\u5728 held-in \u4e0e held-out \u4e0a\u5747\u65e0\u56de\u5f52\u7684\u5019\u9009\u624d\u4f1a\u88ab\u63a5\u53d7\u3002<\/p>\n

                      c.\u88ab\u63a5\u53d7\u7684\u5019\u9009\u88ab\u5408\u5e76\u4ee5\u66f4\u65b0 Harness \u81f3 h_{t+1}\uff1b\u88ab\u62d2\u7edd\u7684\u5019\u9009\u53ea\u8bb0\u5f55\u65e5\u5fd7\uff0c\u4e0d\u6539\u53d8\u5f53\u524d\u751f\u6548\u7684 Harness\u3002<\/p>\n

                      \u5728 Terminal-Bench-2 \u4e0a\u8fd0\u884c MiniMax M2.5\u3001Qwen3.5-35B-A3B\u3001GLM-5 \u65f6\uff0cSelf-Harness \u80fd\u5b66\u5230\u9488\u5bf9\u7279\u5b9a\u6a21\u578b\u7684 Harness \u6307\u4ee4\uff0c\u7784\u51c6\u4e0d\u540c\u57fa\u7840\u6a21\u578b\u7684\u4e0d\u540c\u5f31\u70b9\uff0c\u5e76\u63d0\u5347 held-out \u901a\u8fc7\u7387\u3002<\/p>\n

                      Self-Harness \u8fd9\u7c7b\u5de5\u4f5c\u786e\u5b9e\u5f15\u53d1\u6211\u7684\u62c5\u5fe7\uff1a\u5982\u679c\u4e00\u4e2a\u7a0b\u5e8f\u88ab\u5141\u8bb8\u7f16\u8f91\u64cd\u4f5c\u7cfb\u7edf\uff0c\u62bd\u8c61\u8fb9\u754c\u5c31\u88ab\u6253\u7834\u4e86\u3002\u53ef\u7f16\u8f91\u9762\u9700\u8981\u88ab\u59a5\u5584\u8bbe\u8ba1\uff0c\u6743\u9650\u63a7\u5236\u4e0e\u5b89\u5168\u5c42\u9700\u8981\u7f6e\u4e8e\u8fd9\u4e00\u5faa\u73af\u4e4b\u5916\u3002\u56f4\u7ed5\u5956\u52b1\u9ed1\u5ba2\uff08reward hacking\uff09\u7684\u6240\u6709\u6311\u6218\u4f9d\u7136\u5b58\u5728\u3002<\/p>\n

                      \u667a\u80fd\u4f53 Harness \u5de5\u7a0b\uff08Agentic Harness Engineering\uff0cAHE\uff1bLin et al. 2026\uff09\u8ba4\u4e3a\uff0cHarness \u8fdb\u5316\u7684\u74f6\u9888\u5728\u4e8e\u53ef\u89c2\u6d4b\u6027\uff08observability\uff09\uff0c\u5373\u5f53\u4e00\u6b21 rollout \u5931\u8d25\u65f6\uff0c\u6211\u4eec\u9700\u8981\u77e5\u9053\u662f\u54ea\u4e2a\u7ec4\u4ef6\u5bfc\u81f4\u7684\uff0c\u4e14\u6bcf\u6b21\u4fee\u6539\u90fd\u8981\u6709\u8bc1\u636e\u652f\u6491\u3002<\/p>\n

                      \u8be5\u6846\u67b6\u4ee5\u4e09\u6839\u53ef\u89c2\u6d4b\u6027\u652f\u67f1\u6784\u6210\u95ed\u73af\uff1a<\/p>\n

                        \n
                      • \u7ec4\u4ef6\u53ef\u89c2\u6d4b\u6027\uff1a\u6bcf\u4e2a\u53ef\u7f16\u8f91\u7684 Harness \u7ec4\u4ef6\u90fd\u5728\u6587\u4ef6\u7cfb\u7edf\u4e2d\u6709\u5bf9\u5e94\u8868\u793a\uff0c\u4f7f\u52a8\u4f5c\u7a7a\u95f4\u663e\u5f0f\u4e14\u53ef\u8ffd\u8e2a\u3002<\/li>\n<\/ul>\n

                        a.\u4e00\u4e2a Harness \u5305\u542b 7 \u4e2a\u7ec4\u4ef6\uff1a\u7cfb\u7edf\u63d0\u793a\u8bcd\u3001\u5de5\u5177\u63cf\u8ff0\u3001\u5de5\u5177\u5b9e\u73b0\u3001\u4e2d\u95f4\u4ef6\u3001\u6280\u80fd\u3001\u5b50\u667a\u80fd\u4f53\u914d\u7f6e\u3001\u957f\u671f\u8bb0\u5fc6\u3002<\/p>\n

                        b.\u6bcf\u79cd\u5931\u8d25\u6a21\u5f0f\u88ab\u6620\u5c04\u5230\u4e00\u4e2a\u7ec4\u4ef6\uff0c\u4f7f\u4fee\u6539\u66f4\u6709\u9488\u5bf9\u6027\u3002<\/p>\n

                          \n
                        • \u7ecf\u9a8c\u53ef\u89c2\u6d4b\u6027\uff1a\u628a\u5927\u91cf\u539f\u59cb\u8f68\u8ff9\u5206\u6790\u3001\u6c47\u603b\u4e3a”\u8bc1\u636e\u4e0e\u5931\u8d25\u6a21\u5f0f”\u7684\u5c42\u7ea7\u7ed3\u6784\u3002<\/li>\n<\/ul>\n

                          a.\u6bcf\u4e2a Harness \u751f\u6210 k \u6761\u8f68\u8ff9\u3002<\/p>\n

                          b.\u7528\u4e00\u4e2a\u667a\u80fd\u4f53\uff08”Agent debugger”\uff09\u5206\u6790\u6bcf\u6761\u5b58\u50a8\u4e3a\u72ec\u7acb\u6587\u4ef6\u7684\u8f68\u8ff9\uff0c\u751f\u6210\u9010\u4efb\u52a1\u7684\u6210\u8d25\u6839\u56e0\u5206\u6790\u62a5\u544a\u3002<\/p>\n

                          c.\u6240\u6709\u9010\u4efb\u52a1\u62a5\u544a\u88ab\u805a\u5408\u4e3a\u57fa\u51c6\u603b\u89c8\u4f9b\u4e0b\u4e00\u6b65\u4f7f\u7528\uff0c\u539f\u59cb\u8f68\u8ff9\u5728\u9700\u8981\u65f6\u53ef\u8bbf\u95ee\u3002\u8fd9\u79cd\u5206\u5c42\u8bbf\u95ee\u7ed3\u6784\u66f4\u7701 token\u3002<\/p>\n

                            \n
                          • \u51b3\u7b56\u53ef\u89c2\u6d4b\u6027\uff1a\u6bcf\u6b21\u4fee\u6539\u90fd\u914d\u5bf9\u4e00\u6761”\u5bf9\u4e0b\u4e00\u8f6e\u7684\u9884\u6d4b”\uff0c\u4ee5\u4f9b\u9a8c\u8bc1\u3002<\/li>\n<\/ul>\n

                            a.\u4e00\u4e2a\u667a\u80fd\u4f53\uff08”Evolve agent”\uff09\u9605\u8bfb\u4ed3\u5e93\u3001\u51b3\u5b9a\u7f16\u8f91\u54ea\u4e2a\u7ec4\u4ef6\uff0c\u7136\u540e\u4ea7\u51fa\u4fee\u6539\u53ca\u5176\u7406\u7531\u3002<\/p>\n

                            b.\u6bcf\u6b21\u4fee\u6539\u90fd\u662f\u4e00\u6761\u6587\u4ef6\u7ea7\u7684\u3001\u53ef\u8bc1\u4f2a\u7684\u58f0\u660e\uff0c\u53ef\u5728\u4e0b\u4e00\u8f6e\u88ab\u9a8c\u8bc1\uff0c\u5e76\u53d7\u4e24\u6761\u7ea6\u675f\uff1a<\/p>\n

                            i.\u4fee\u6539\u4ec5\u5e94\u7528\u4e8e Harness \u5de5\u4f5c\u533a\uff1bruns \u76ee\u5f55\u3001tracer\u3001\u9a8c\u8bc1\u5668\u4e0e LLM \u914d\u7f6e\u5747\u4e3a\u53ea\u8bfb\u3002\u8fd9\u7981\u6389\u4e86\u4e00\u7c7b\u5956\u52b1\u9ed1\u5ba2\u884c\u4e3a\uff08\u5982\u7981\u7528\u9a8c\u8bc1\u5668\u3001\u66f4\u6362\u6a21\u578b\u3001\u62ac\u9ad8\u63a8\u7406\u9884\u7b97\uff09\uff0c\u4ece\u800c\u4fdd\u8bc1\u8bb0\u5f55\u7684\u6bcf\u4e00\u70b9\u6536\u76ca\u90fd\u53ef\u5f52\u56e0\u4e8e Harness \u4fee\u6539\u3002<\/p>\n

                            ii.\u4fee\u6539\u5fc5\u987b\u7531\u8bc1\u636e\u9a71\u52a8\uff0c\u5e76\u9644\u4e00\u6761”\u5ba3\u8a00\uff08manifesto\uff09”\u6761\u76ee\uff1a\u5931\u8d25\u8bc1\u636e\u7684\u540d\u79f0\u3001\u63a8\u65ad\u7684\u6839\u56e0\u3001\u9488\u5bf9\u6027\u4fee\u590d\u3001\u4ee5\u53ca\u9884\u6d4b\u5f71\u54cd\uff08\u5305\u542b\u9884\u671f\u4fee\u590d\u9879\u4e0e\u5b58\u5728\u56de\u5f52\u98ce\u9669\u9879\uff09\u3002<\/p>\n

                            \u5728 Terminal-Bench-2 \u4e0a\uff0cAHE \u8d85\u8fc7\u4e86\u4eba\u5de5\u8bbe\u8ba1\u7684 Harness\uff08OpenCode\u3001Terminus-2\u3001Codex\uff09\uff08Hard \u6863\u9664\u5916\uff09\u4ee5\u53ca\u82e5\u5e72\u81ea\u6211\u8fdb\u5316\u57fa\u7ebf\uff08ACE\u3001TF-GRPO\uff09\u3002\u540c\u4e00\u5957\u51bb\u7ed3 Harness \u4e0d\u7ecf\u7ee7\u7eed\u8fdb\u5316\u5373\u53ef\u8fc1\u79fb\u5230 SWE-bench-verified\uff0c\u8bf4\u660e\u8fdb\u5316\u51fa\u7684 Harness \u80fd\u628a\u5de5\u7a0b\u7ecf\u9a8c\u7f16\u7801\u8fdb Harness \u7ec4\u4ef6\uff0c\u800c\u975e\u505a\u9488\u5bf9\u7279\u5b9a\u57fa\u51c6\u7684\u8fc7\u62df\u5408\u4f18\u5316\u3002<\/p>\n

                            2.4 \u8fdb\u5316\u641c\u7d22<\/h3>\n

                            \u8fdb\u5316\u641c\u7d22\u662f\u4e00\u79cd\u53d7\u81ea\u7136\u9009\u62e9\u542f\u53d1\u7684\u4f18\u5316\u65b9\u6cd5\uff08\u53c2\u89c1\u6211\u6b64\u524d\u5173\u4e8e\u8fdb\u5316\u7b97\u6cd5\u7684\u65e7\u6587\uff09\u3002\u5b83\u901a\u8fc7\u53d8\u5f02\u6765\u8fdb\u5316\u4e00\u4e2a\u89e3\u7684\u79cd\u7fa4\uff0c\u53ea\u4fdd\u7559\u7fa4\u4f53\u4e2d”\u9002\u5e94\u5ea6”\u9ad8\u7684\u4e2a\u4f53\u3002\u8fdb\u5316\u641c\u7d22\u5728\u4ee5\u4e0b\u60c5\u5f62\u5f88\u597d\u7528\uff1a(1) \u641c\u7d22\u7a7a\u95f4\u5e9e\u5927\u6216\u5f62\u72b6\u602a\u5f02\uff1b(2) \u96be\u4ee5\u76f4\u63a5\u7528\u68af\u5ea6\u4f18\u5316\u3001\u4f46\u5bb9\u6613\u8bc4\u4f30\u89e3\u3002Harness \u641c\u7d22\u770b\u8d77\u6765\u975e\u5e38\u9002\u5408\u3002<\/p>\n

                            \u8fc7\u53bb\u7684\u7814\u7a76\u5df2\u5c06\u8fdb\u5316\u641c\u7d22\u7528\u4e8e\u63d0\u793a\u8bcd\u5de5\u7a0b\u3002Promptbreeder\uff08Fernando et al. 2023\uff09\u901a\u8fc7\u4e30\u5bcc\u7684\u53d8\u5f02\u64cd\u4f5c\u4f18\u5316\u4efb\u52a1\u63d0\u793a\u8bcd\uff1b\u6709\u8da3\u7684\u662f\uff0c\u53d8\u5f02\u63d0\u793a\u8bcd\u672c\u8eab\uff08\u5373\u6307\u793a LLM \u53bb\u53d8\u5f02\u4efb\u52a1\u63d0\u793a\u8bcd\u7684\u6307\u4ee4\uff09\u4e5f\u901a\u8fc7\u8fdb\u5316\u6765\u6539\u8fdb\u3002GEPA\uff08Agrawal et al. 2025\uff09\u628a\u57fa\u4e8e\u53cd\u601d\u7684\u63d0\u793a\u4e0e\u8fdb\u5316\u641c\u7d22\u7ed3\u5408\uff0c\u7528\u5bf9\u8bd5\u9519\u8f68\u8ff9\u7684\u81ea\u7136\u8bed\u8a00\u53cd\u601d\u6765\u63d0\u51fa\u63d0\u793a\u8bcd\u66f4\u65b0\u3002<\/p>\n

                            Novikov et al.\uff082025\uff09\u63d0\u51fa\u4e86\u00a0AlphaEvolve\uff0c\u4e00\u4e2a\u7f16\u7a0b\u667a\u80fd\u4f53\u8fdb\u5316\u641c\u7d22\u7cfb\u7edf\uff1a\u5b83\u7ef4\u62a4\u4e00\u4e2a\u5019\u9009\u7a0b\u5e8f\u6c60\uff0c\u63d0\u793a\u51bb\u7ed3\u7684 LLM \u751f\u6210\u7528\u4e8e\u6539\u8fdb\u7684 diff\u3002\u968f\u7740\u7cfb\u7edf\u53cd\u590d\u8bc4\u4f30\u5b50\u7a0b\u5e8f\u5e76\u4fdd\u7559\u6210\u529f\u8005\uff0c\u5b83\u4f1a\u968f\u65f6\u95f4\u53d1\u73b0\u66f4\u597d\u7684\u89e3\u3002<\/p>\n

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

                            \uff08\u56fe14\uff1aAlphaEvolve \u7684\u5de5\u4f5c\u65b9\u5f0f\u3002\u56fe\u7247\u6765\u6e90\uff1aNovikov et al. 2025\uff09<\/p>\n

                            AlphaEvolve \u7684\u8bbe\u8ba1\u4e2d\u6709\u51e0\u4e2a\u5173\u952e\u7ec6\u8282\uff1a<\/p>\n

                              \n
                            • \u63d0\u793a\u8bcd\u4e2d\u5305\u542b\u7236\u7a0b\u5e8f\u3001\u7ed3\u679c\u3001\u6307\u4ee4\uff0c\u6709\u65f6\u8fd8\u6709\u5143\u4fe1\u606f\u3002<\/li>\n
                            • \u7f16\u7a0b\u667a\u80fd\u4f53\u53ef\u8bbf\u95ee\u5b8c\u6574\u4ed3\u5e93\uff0c\u4f46\u5f85\u6539\u8fdb\u7684\u4ee3\u7801\u533a\u57df\u7528\u00a0#EVOLVE-BLOCK-START \u4e0e\u00a0# EVOLVE-BLOCK-END\u00a0\u663e\u5f0f\u6807\u8bb0\u3002<\/li>\n
                            • \u5143\u63d0\u793a\u8bcd\uff08meta-prompt\uff09\u6309 LLM \u7684\u5efa\u8bae\u4e0e\u6307\u4ee4\u3001\u4e0a\u4e0b\u6587\u5171\u540c\u8fdb\u5316\uff0c\u65b9\u5f0f\u4e0e\u6211\u4eec\u8fdb\u5316\u89e3\u7a0b\u5e8f\u7c7b\u4f3c\u3002<\/li>\n
                            • \u6d88\u878d\u5b9e\u9a8c\u663e\u793a\u4e86\u8fdb\u5316\u6d41\u7a0b\u3001\u63d0\u793a\u8bcd\u4e2d\u7684\u4e0a\u4e0b\u6587\u3001\u5143\u63d0\u793a\u8bcd\u3001\u6574\u6587\u4ef6\u8fdb\u5316\u4ee5\u53ca\u4f7f\u7528\u66f4\u5f3a LLM \u7684\u4ef7\u503c\u3002
                              \n\"\u56fe\u7247\"<\/li>\n<\/ul>\n

                              \uff08\u56fe15\uff1a\u6d88\u878d\u5b9e\u9a8c\u8bc1\u660e AlphaEvolve \u82e5\u5e72\u8bbe\u8ba1\u7684\u4ef7\u503c\u3002\u56fe\u7247\u6765\u6e90\uff1aNovikov et al. 2025\uff09<\/p>\n

                              \u8fd1\u671f\u7684\u53d8\u4f53\u5982\u00a0ThetaEvolve\uff08Wang et al. 2025\uff09\u628a\u8fdb\u5316\u641c\u7d22\u4e0e RL\u3001\u4e0a\u4e0b\u6587\u5185\u5b66\u4e60\u7ed3\u5408\uff1bDemoEvolve\uff08Che et al. 2026\uff09\u7528\u4eba\u7c7b\u4e13\u5bb6\u6f14\u793a\u589e\u5f3a\u81ea\u6211 rollout \u6863\u6848\u5e93\uff0c\u4f5c\u4e3a Harness \u5c42\u8bca\u65ad\u4e0e\u7f16\u8f91\u7684\u53c2\u8003\u7ecf\u9a8c\u3002ShinkaEvolve\uff08Lange et al. 2025\uff09\u5219\u5f15\u5165\u4e86\u4e09\u4e2a\u63d0\u5347 LLM \u91c7\u6837\u6548\u7387\u7684\u65b0\u7ec4\u4ef6\uff1a<\/p>\n

                                \n
                              • \u66f4\u5177\u91c7\u6837\u6548\u7387\u7684\u63a2\u7d22\uff1a\u8bbe\u8ba1\u7236\u4ee3\u91c7\u6837\uff0c\u5728\u8868\u73b0\u6392\u540d\u4e0e\u540e\u4ee3\u6570\u91cf\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/li>\n
                              • \u4ee3\u7801\u65b0\u9896\u6027\u62d2\u7edd\u91c7\u6837\uff1a\u57fa\u4e8e\u5d4c\u5165\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff0c\u4e22\u5f03\u4e0e\u73b0\u6709\u79cd\u7fa4\u8fc7\u4e8e\u76f8\u4f3c\u7684\u5019\u9009\u3002<\/li>\n
                              • \u5728”\u5143\u8349\u7a3f\u672c\uff08meta-scratchpad\uff09”\u4e2d\u8bc6\u522b\u6210\u529f\u89e3\u91cc\u7684\u597d\u6a21\u5f0f\uff0c\u4ee5\u5f15\u5bfc\u672a\u6765\u7684\u53d8\u5f02\u3002<\/li>\n<\/ul>\n

                                \u4e0e\u4e0a\u8ff0\u805a\u7126”\u89e3\u7684\u6539\u8fdb”\u7684\u65b9\u6cd5\u4e0d\u540c\uff0c\u8fbe\u5c14\u6587-\u54e5\u5fb7\u5c14\u673a\uff08Darwin G\u00f6del Machine\uff0cDGM\uff1bZhang et al. 2025\uff09\u660e\u786e\u4ee5”\u53ef\u7f16\u8f91 Harness \u4ee3\u7801\u4ed3\u5e93”\u7684\u8fdb\u5316\u4e3a\u76ee\u6807\uff0c\u4f7f\u7528\u57fa\u4e8e LLM \u7684\u7f16\u7a0b\u667a\u80fd\u4f53\uff0c\u51c6\u786e\u8bf4\uff0c\u8fd9\u4e2a\u667a\u80fd\u4f53\u88ab\u5141\u8bb8\u4fee\u6539\u81ea\u5df1\u7684 Harness\u3002\u540e\u7eed\u5de5\u4f5c Hyperagents\uff08Zhang et al. 2026\uff09\u5f15\u5165\u4e86\u4e00\u4e2a\u5143\u667a\u80fd\u4f53\uff0c\u6765\u63a7\u5236\u5982\u4f55\u4fee\u6539\u73b0\u6709\u4efb\u52a1\u667a\u80fd\u4f53\u4ee5\u521b\u5efa\u65b0\u667a\u80fd\u4f53\u3002<\/p>\n

                                  \n
                                • \u6c60\u4e2d\u4ece\u4e00\u4e2a\u7f16\u7a0b\u667a\u80fd\u4f53\u5f00\u59cb\u3002<\/li>\n
                                • \u6bcf\u8f6e\u8fed\u4ee3\u6309”\u4e0e\u8868\u73b0\u6210\u6b63\u6bd4\u3001\u4e0e\u5176\u5df2\u6709\u5b50\u4ee3\u6570\u6210\u53cd\u6bd4”\u7684\u6982\u7387\u6311\u9009\u4e00\u4e2a\u7236\u4ee3\uff0c\u5bf9\u5176\u8fdb\u884c\u4fee\u6539\u5e76\u5206\u53c9\u4ea7\u751f\u65b0\u667a\u80fd\u4f53\u3002<\/li>\n
                                • \u88ab\u9009\u4e2d\u7684\u7236\u667a\u80fd\u4f53\u68c0\u67e5\u81ea\u5df1\u7684\u57fa\u51c6\u8bc4\u4f30\u65e5\u5fd7\uff0c\u7136\u540e\u5bf9\u81ea\u8eab\u7684 Harness \u4ee3\u7801\u5e93\u63d0\u51fa\u6539\u8fdb\uff0c\u751f\u6210\u7f16\u7a0b\u667a\u80fd\u4f53\u7684\u65b0\u7248\u672c\u3002\u4ee3\u7801\u7f16\u8f91\u901a\u8fc7\u4e24\u4e2a\u57fa\u7840\u5de5\u5177\u5b9e\u73b0\uff1a(1) bash\uff08\u53c2\u6570\uff1a\uff09\uff1b(2) editor\uff08\u53c2\u6570\uff1aview\/create\/edit\uff09\u3002<\/li>\n
                                • \u65b0\u7f16\u7a0b\u667a\u80fd\u4f53\u88ab\u8bc4\u4f30\uff0c\u53ea\u6709\u8868\u73b0\u8db3\u591f\u9ad8\u7684\u624d\u4f1a\u88ab\u52a0\u56de\u6c60\u4e2d\u3002<\/li>\n
                                • \u91cd\u590d\u6b65\u9aa4 2\u20134\uff0c\u76f4\u5230\u6ee1\u8db3\u505c\u6b62\u6761\u4ef6\u3002<\/li>\n<\/ul>\n

                                  DGM \u662f\u56fa\u5b9a\u6a21\u578b\u4e0b\u7684 Harness \u8fdb\u5316\u3002\u5728\u4ee5 Claude 3.5 Sonnet \u4e3a\u57fa\u7840 LLM\u3001\u521d\u59cb Harness \u914d\u7f6e\u5f88\u7b80\u5355\u7684\u5b9e\u9a8c\u4e2d\uff0cDGM \u53d1\u73b0\u7684\u667a\u80fd\u4f53\u5728 SWE-bench Verified\uff0820% \u2192 50%\uff09\u4e0e Polyglot\uff0814.2% \u2192 30.7%\uff09\u4e0a\u4e0e\u624b\u5de5\u6253\u9020\u7684\u667a\u80fd\u4f53\u76f8\u5f53\u6216\u66f4\u4f18\u3002<\/p>\n

                                  \u8fd9\u4e00\u65cf\u65b9\u6cd5\u5728\u5019\u9009\u89e3\u53ef\u81ea\u52a8\u8bc4\u4f30\u3001\u9002\u5e94\u5ea6\u6613\u91cf\u5316\u7684\u573a\u666f\u6548\u679c\u5f88\u597d\uff0c\u5982\u77e9\u9635\u4e58\u6cd5\u3001GPU kernel \u4f18\u5316\u3001\u7b97\u6cd5\u7ade\u8d5b\u3001\u6570\u636e\u4e2d\u5fc3\u8c03\u5ea6\u3002\u5728\u8bc4\u4f30\u7f13\u6162\u3001\u6a21\u7cca\u6216\u4e3b\u8981\u4f9d\u8d56\u542f\u53d1\u5f0f\u7684\u9886\u57df\u5219\u5f88\u5403\u529b\u3002\u8fdb\u5316\u7684\u7b97\u529b\u6548\u7387\u4e0e\u6709\u6548\u6027\u4e5f\u662f\u9690\u5fe7\u3002<\/p>\n

                                  2.5 \u4e0e\u6a21\u578b\u6743\u91cd\u7684\u8054\u5408\u4f18\u5316<\/h3>\n

                                  Harness \u8fdb\u5316\u6539\u53d8\u7684\u662f\u6a21\u578b\u5916\u56f4\u7684\u975e\u53c2\u6570\u7cfb\u7edf\u3002\u8981\u5b9e\u73b0\u5b8c\u5168\u7684\u81ea\u6211\u6539\u8fdb\uff0c\u6a21\u578b\u5f53\u7136\u4e5f\u5b8c\u5168\u53ef\u4ee5\u88ab\u5141\u8bb8\u540c\u65f6\u66f4\u65b0\u81ea\u5df1\u7684\u6743\u91cd\u3002\u6743\u91cd\u66f4\u65b0\u53ef\u4ee5\u901a\u8fc7\u6539\u8fdb\u6a21\u578b\u8bad\u7ec3\u6d41\u6c34\u7ebf\uff0c\u6216\u5728\u6d4b\u8bd5\u65f6\u8fdb\u884c\u6301\u7eed\u5b66\u4e60\u6765\u5b9e\u73b0\u3002\u6301\u7eed\u5b66\u4e60\u8fd9\u4e2a\u4e3b\u9898\u503c\u5f97\u672a\u6765\u5355\u72ec\u5199\u4e00\u7bc7\u6587\u7ae0\u3002<\/p>\n

                                  SIA\uff08Hebbar et al. 2026\uff09\u662f\u628a Harness \u6539\u8fdb\u4e0e\u6a21\u578b\u53c2\u6570\u66f4\u65b0\u653e\u8fdb\u540c\u4e00\u4f18\u5316\u5faa\u73af\u7684\u65e9\u671f\u5c1d\u8bd5\uff0c\u8bbe\u8ba1\u4e0a\u6709\u4e09\u4e2a\u7ec4\u4ef6\uff1a<\/p>\n

                                    \n
                                  • \u5143\u667a\u80fd\u4f53\uff08Meta-Agent\uff09\uff1a\u63d0\u51fa\u521d\u59cb Harness\u3002<\/li>\n
                                  • \u4efb\u52a1\u667a\u80fd\u4f53\uff08Task-Specific Agent\uff09\uff1a\u6267\u884c\u4efb\u52a1\u3002<\/li>\n
                                  • \u53cd\u9988\u667a\u80fd\u4f53\uff08Feedback-Agent\uff09\uff1a\u57fa\u4e8e\u8fd1\u671f\u8f68\u8ff9\uff0c\u51b3\u5b9a\u4e0b\u4e00\u6b65\u66f4\u65b0 Harness \u8fd8\u662f\u66f4\u65b0\u6a21\u578b\u6743\u91cd\u3002
                                    \n\"\u56fe\u7247\"<\/li>\n<\/ul>\n

                                    \uff08\u56fe16\uff1aSIA \u4e2d\u7684\u53cd\u9988\u667a\u80fd\u4f53\u51b3\u5b9a\u4e0b\u4e00\u6b21\u8fed\u4ee3\u7684\u7c7b\u578b\u3002\u56fe\u7247\u6765\u6e90\uff1aHebbar et al. 2026\uff09<\/p>\n

                                    SIA \u7684\u5b9e\u9a8c\u4e2d\u5b58\u5728\u4e00\u4e9b\u6df7\u6742\u56e0\u7d20\uff0c\u4f7f\u7ed3\u679c\u96be\u4ee5\u89e3\u8bfb\u3002\u4f8b\u5982\uff0c\u4efb\u52a1\u667a\u80fd\u4f53\u8fdc\u5f31\u4e8e\u5143\u667a\u80fd\u4f53\u4e0e\u53cd\u9988\u667a\u80fd\u4f53\u6240\u7528\u7684\u6a21\u578b\uff08gpt-oss-120b vs Claude Sonnet 4.6\uff09\uff0c\u57fa\u7ebf\u4e5f\u592a\u5f31\uff0c\u65e0\u6cd5\u4e0e\u76f8\u5173\u65b9\u6cd5\u5e72\u51c0\u5730\u4ea4\u53c9\u5bf9\u7167\u3002\u6211\u8ba4\u4e3a\u8fd9\u4e2a\u65b9\u5411\u6709\u8da3\uff0c\u4f46\u8bc1\u636e\u5c1a\u5c5e\u521d\u6b65\u3002\u8bad\u7ec3\u7a33\u5b9a\u6027\u3001\u53e4\u5fb7\u54c8\u7279\u6548\uff08Goodhart effect\uff09\u7b49\u8bb8\u591a\u6311\u6218\u4ecd\u7136\u60ac\u800c\u672a\u51b3\u3002<\/p>\n

                                    Continual Harness\uff08Karten et al. 2026\uff09\u5728\u957f\u7a0b\u6e38\u620f\u573a\u666f\u4e2d\u5b9e\u9a8c\u4e86 Harness \u66f4\u65b0\uff0c\u5e76\u901a\u8fc7\u5728\u4f4e\u5956\u52b1\u8f68\u8ff9\u4e0a\u84b8\u998f\u5f3a\u6559\u5e08\u6a21\u578b\u7684\u6807\u7b7e\u6765\u5171\u540c\u5b66\u4e60\u4e00\u4e2a\u7b56\u7565\u6a21\u578b\u3002<\/p>\n

                                    \u4e09\u3001\u672a\u6765\u6311\u6218<\/h2>\n

                                    AI Scientist \u8fd9\u4e00\u7cfb\u5217\u5de5\u4f5c\u6709\u529b\u5730\u8bc1\u660e\uff1a\u4e13\u5bb6\u8bbe\u8ba1\u7684 Harness \u80fd\u534f\u8c03\u81ea\u52a8\u5316\u7814\u7a76\u5faa\u73af\u7684\u5927\u90e8\u5206\u73af\u8282\uff0c\u5176\u5b9e\u9a8c\u5f62\u5f0f\u662f\u64b0\u5199\u7814\u7a76\u8bba\u6587\u3002\u4f46”\u8bba\u6587\u4ea7\u51fa”\u5e76\u4e0d\u7b49\u540c\u4e8e”\u79d1\u5b66\u53d1\u73b0”\u3002\u4e00\u4e2a\u7cfb\u7edf\u53ef\u4ee5\u5199\u51fa\u770b\u4f3c\u53ef\u4fe1\u7684\u624b\u7a3f\uff0c\u540c\u65f6\u4ecd\u5b58\u5728\u4f2a\u9020\u5f15\u7528\u3001\u5b9e\u73b0\u6f02\u79fb\u6216\u5b9e\u9a8c\u7ed3\u679c\u8584\u5f31\u7684\u95ee\u9898\u3002<\/p>\n

                                    Trehan & Chopra\uff082026\uff09\u6d4b\u8bd5\u4e86 LLM \u80fd\u5426\u5728\u6781\u7b80\u811a\u624b\u67b6\u548c\u57fa\u7840\u5de5\u5177\uff08\u5373read_file\u3001write_file\u3001llm_search\u3001list_files\uff09\u4e0b\u4ece\u4e00\u4e2a\u7814\u7a76\u60f3\u6cd5\u8d70\u5230\u4e00\u7bc7\u8bba\u6587\u3002\u6bcf\u4e2a\u60f3\u6cd5\u6709\u72ec\u7acb\u5de5\u4f5c\u533a\uff0c\u667a\u80fd\u4f53\u53ef\u4ee5\u751f\u6210\u5e76\u9605\u8bfb\u6587\u6863\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u7684\u4e00\u90e8\u5206\u3002\u4ed6\u4eec\u5728\u4e09\u4e2a\u9886\u57df\uff08\u4e16\u754c\u6a21\u578b\u3001\u591a\u667a\u80fd\u4f53 RL\u3001AI \u5b89\u5168\u4e0e\u5bf9\u9f50\uff09\u505a\u4e86\u5b9e\u9a8c\uff0c\u6bcf\u4e2a\u9886\u57df\u542b 45\u201350 \u7bc7\u9ad8\u8d28\u91cf\u79cd\u5b50\u6587\u6863\u4ee5\u6fc0\u53d1\u65b0\u60f3\u6cd5\u3002\u53ea\u6709\u56db\u4e2a\u60f3\u6cd5\u88ab\u4eba\u7c7b\u4e13\u5bb6\u9009\u4e2d\u8dd1\u5b8c\u6574\u6d41\u6c34\u7ebf\uff0c\u4e14\u53ea\u6709\u4e00\u4e2a\u88ab\u5b8c\u6574\u6267\u884c\u6210\u8bba\u6587\u3002\u4ed6\u4eec\u89c2\u5bdf\u5230\u516d\u79cd\u53cd\u590d\u51fa\u73b0\u7684\u5931\u8d25\u6a21\u5f0f\uff1a<\/p>\n

                                      \n
                                    • \u504f\u5411\u8bad\u7ec3\u6570\u636e\u9ed8\u8ba4\u503c\uff1a\u4f7f\u7528\u65e7\u5e93\u3001\u8fc7\u65f6\u547d\u4ee4\u3001\u6807\u51c6\u683c\u5f0f\uff0c\u6216\u672a\u7ecf\u5b9e\u9645\u4ed3\u5e93\/\u6570\u636e\u96c6\u9a8c\u8bc1\u7684\u5047\u8bbe\u3002<\/li>\n
                                    • \u6267\u884c\u538b\u529b\u4e0b\u7684\u5b9e\u73b0\u6f02\u79fb\uff1a\u5f53\u5b9e\u73b0\u53d8\u5f97\u6280\u672f\u4e0a\u590d\u6742\u65f6\uff0c\u6a21\u578b\u53ef\u80fd\u8f6c\u5411\u5e38\u89c1\u7684\u66f4\u7b80\u5355\u65b9\u6848\uff0c\u800c\u975e\u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u3002<\/li>\n
                                    • \u8bb0\u5fc6\u4e0e\u4e0a\u4e0b\u6587\u9000\u5316\uff1a\u957f\u7a0b\u9879\u76ee\u4f1a\u4e22\u5931\u5173\u952e\u7ec6\u8282\uff0c\u9664\u975e\u65e5\u5fd7\u88ab\u5199\u6210\u6301\u4e45\u4ea7\u7269\u3002<\/li>\n
                                    • \u8fc7\u5ea6\u4e50\u89c2\uff1a\u5c3d\u7ba1\u5b9e\u9a8c\u5145\u6ee1\u566a\u58f0\u6216\u5df2\u5931\u8d25\uff0c\u6a21\u578b\u4ecd\u5ba3\u5e03\u6210\u529f\uff1bBubeck et al.\uff082025\uff09\u4e5f\u89c2\u5bdf\u5230\u7c7b\u4f3c\u7684”p \u503c\u64cd\u7eb5\u4e0e Eureka \u65f6\u523b”\u6a21\u5f0f\uff0c\u6a21\u578b\u4f1a\u5f15\u5165”\u6570\u503c\u521b\u53ef\u8d34”\uff0c\u5728\u4fe1\u53f7\u4ecd\u662f\u566a\u58f0\u65f6\u5c31\u5ba3\u5e03\u80dc\u5229\u3002<\/li>\n
                                    • \u9886\u57df\u667a\u80fd\u4e0d\u8db3\uff1a\u6a21\u578b\u7f3a\u4e4f\u9ed8\u4f1a\u7684\u6280\u827a\u6027\u77e5\u8bc6\uff0c\u5982\u9884\u4f30\u5b9e\u73b0\u590d\u6742\u5ea6\u3001\u5224\u65ad\u5b9e\u9a8c\u7ed3\u679c\u662f\u5426\u5408\u7406\u3001\u77e5\u9053\u54ea\u4e9b\u57fa\u7ebf\u91cd\u8981\u3002<\/li>\n
                                    • \u79d1\u5b66\u54c1\u5473\u8584\u5f31\uff1a\u5b9e\u9a8c\u4e5f\u8bb8\u53ef\u4ee5\u6267\u884c\uff0c\u4f46\u6ca1\u80fd\u56de\u7b54\u6b63\u786e\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n

                                      \u8fc8\u5411\u5b8c\u5168\u7684 RSI\uff0c\u7814\u7a76\u8005\u4eec\u5df2\u53d6\u5f97\u5b9e\u8d28\u8fdb\u5c55\uff0c\u4f46\u4ecd\u6709\u82e5\u5e72\u74f6\u9888\uff1a<\/p>\n

                                      1. \u8bc4\u4f30\u5668\u5f31\u4e14\u6a21\u7cca\u3002\u8bb8\u591a\u7814\u7a76\u8bba\u65ad\u6ca1\u6709\u5feb\u901f\u7cbe\u786e\u7684\u9a8c\u8bc1\u5668\uff0c\u8bb8\u591a\u771f\u5b9e\u4e16\u754c\u4efb\u52a1\u4e5f\u662f\u5982\u6b64\u3002\u5f53\u524d\u7684\u81ea\u6211\u6539\u8fdb\u5faa\u73af\u5728\u8bc4\u4f30\u6307\u6807\u53ef\u6d4b\u91cf\u3001\u5ba2\u89c2\u7684\u4efb\u52a1\u4e0a\u6548\u679c\u6700\u597d\uff0c\u8fd9\u4e0e RL \u7684\u594f\u6548\u6761\u4ef6\u76f8\u4f3c\u3002\u800c\u7814\u7a76\u54c1\u5473\u3001\u65b0\u9896\u6027\u4e0e\u957f\u671f\u79d1\u5b66\u4ef7\u503c\u8981\u96be\u8861\u91cf\u5f97\u591a\u3002\u4f8b\u5982\uff0c\u7814\u7a76\u54c1\u5473\u5f80\u5f80\u6df7\u5408\u4e86\u95ee\u9898\u6846\u5b9a\u3001\u5b9e\u9a8c\u8bbe\u8ba1\uff0c\u4ee5\u53ca”\u54ea\u4e9b\u610f\u5916\u7ed3\u679c\u503c\u5f97\u8ffd\u3001\u54ea\u4e9b\u5931\u8d25\u6848\u4f8b\u503c\u5f97\u91cd\u8bd5”\u7684\u5224\u65ad\u3002<\/p>\n

                                      2. \u4e0a\u4e0b\u6587\u4e0e\u8bb0\u5fc6\u7684\u751f\u547d\u5468\u671f\u3002\u968f\u7740 AI \u667a\u80fd\u4f53\u53d8\u5f97\u66f4\u81ea\u4e3b\u72ec\u7acb\uff0c\u8bb0\u5fc6\u4f1a\u4e0d\u65ad\u589e\u957f\u3002\u4e00\u4e2a\u6709\u7528\u7684 Harness \u9700\u8981\u7ba1\u7406\u4e0a\u4e0b\u6587\u4e0e\u8bb0\u5fc6\uff0c\u5728\u5f25\u8865\u957f\u4e0a\u4e0b\u6587\u751f\u6210\u73b0\u6709\u5c40\u9650\u7684\u540c\u65f6\uff0c\u6700\u5927\u5316\u957f\u7a0b\u4efb\u52a1\u7684\u6210\u529f\u7387\u3002\u65e2\u7136\u4eba\u7c7b\u80fd\u7ec8\u8eab\u7ef4\u6301\u8bb0\u5fc6\uff0c\u6211\u5728\u6b64\u770b\u5230\u4e00\u4e2a\u7c7b\u6bd4\uff1a\u4e0a\u4e0b\u6587\u5de5\u7a0b\u5c06\u4f1a\u3001\u4e5f\u5e94\u5f53\u6210\u4e3a\u667a\u80fd\u7684\u6838\u5fc3\u7ec4\u6210\u90e8\u5206\uff0c\u800c\u4e0d\u662f\u505c\u7559\u5728\u8f6f\u4ef6\u7cfb\u7edf\u5c42\u3002<\/p>\n

                                      3. \u9634\u6027\u7ed3\u679c\uff08negative results\uff09\u3002\u7814\u7a76\u8005\u7684\u6fc0\u52b1\u662f\u53d1\u8868\u6210\u529f\u7ed3\u679c\uff0c\u56e0\u6b64\u6587\u732e\u504f\u5411\u6210\u529f\u3002\u8bad\u7ec3\u4e8e\u6d77\u91cf\u6570\u636e\uff08\u81f3\u5c11\u76ee\u524d mostly \u662f\u4eba\u7c7b\u521b\u9020\u7684\uff0c\u7b11\uff09\u7684 LLM\uff0c\u53ef\u80fd\u4e0d\u64c5\u957f\u51b3\u5b9a\u4f55\u65f6\u653e\u5f03\u4e00\u4e2a\u5047\u8bbe\u3001\u62a5\u544a\u4e00\u4e2a\u9634\u6027\u7ed3\u679c\u3001\u751a\u81f3\u627f\u8ba4\u4e00\u6b21\u5931\u8d25\uff0c\u56e0\u4e3a\u6570\u636e\u4e2d\u6210\u529f\u4e0e\u5931\u8d25\u6848\u4f8b\u4e0d\u5e73\u8861\u3002\u4e00\u4e2a\u7814\u7a76\u578b Harness \u5e94\u8ba9\u5931\u8d25\u7684\u5c1d\u8bd5\u6613\u4e8e\u4fdd\u5b58\uff0c\u56e0\u4e3a\u4ece\u5931\u8d25\u4e2d\u5b66\u4e60\u662f\u524a\u51cf\u4efb\u52a1\u641c\u7d22\u7a7a\u95f4\u7684\u6700\u4f73\u65b9\u5f0f\u3002<\/p>\n

                                      4. \u591a\u6837\u6027\u574d\u7f29\u3002\u8fdb\u5316\u4e0e RL \u5faa\u73af\u503e\u5411\u4e8e\u5229\u7528\u5df2\u77e5\u7684\u9ad8\u5956\u52b1\u6a21\u5f0f\u3002\u6211\u4eec\u9700\u8981\u673a\u5236\u6765\u9632\u6b62\u79cd\u7fa4\u574d\u7f29\u6210\u540c\u4e00\u4e2a\u89e3\u7684\u53d8\u4f53\u3002\u8fd9\u5bf9\u5f00\u653e\u5f0f\u7814\u7a76\u5c24\u5176\u5173\u952e\uff0c\u6700\u597d\u7684\u8def\u5f84\u5728\u5f53\u524d\u8bc4\u4f30\u5668\u4e0b\u53ef\u80fd\u8d77\u521d\u770b\u8d77\u6765\u66f4\u5dee\u3002<\/p>\n

                                      5. \u5956\u52b1\u9ed1\u5ba2\uff08reward hacking\uff09\u3002\u81ea\u6211\u6539\u8fdb\u5faa\u73af\u4f18\u5316\u7684\u662f\u4efb\u4f55\u88ab\u7ed9\u4e88\u7684\u4fe1\u53f7\u3002\u5982\u679c\u5956\u52b1\u6765\u81ea\u5355\u5143\u6d4b\u8bd5\uff0c\u667a\u80fd\u4f53\u53ef\u80fd\u5bf9\u6d4b\u8bd5\u8fc7\u62df\u5408\uff1b\u5982\u679c\u6765\u81ea\u88c1\u5224\u6a21\u578b\uff0c\u5b83\u53ef\u80fd\u5b66\u4f1a\u9488\u5bf9\u8be5\u88c1\u5224\u7684\u5956\u52b1\u9ed1\u5ba2\u6280\u5de7\uff1b\u5982\u679c\u6765\u81ea\u57fa\u51c6\u5206\u6570\uff0c\u5b83\u53ef\u80fd\u5229\u7528\u57fa\u51c6\u7684\u6f0f\u6d1e\u3002<\/p>\n

                                      \u8bc4\u4f30\u5668\u4e0e\u6743\u9650\u63a7\u5236\u5f88\u53ef\u80fd\u5e94\u5f53\u7f6e\u4e8e\u8fdb\u5316 Harness \u7684\u5faa\u73af\u4e4b\u5916\uff0c\u8f85\u4ee5 held-out \u6d4b\u8bd5\u3001\u8f68\u8ff9\u5ba1\u8ba1\u3001\u4ee5\u53ca\u5728\u5173\u952e\u51b3\u7b56\u70b9\u7684\u4eba\u7c7b\u8bc4\u5ba1\uff0c”\u591a\u5927\u7a0b\u5ea6\u7684\u76d1\u7763\u53ef\u4ee5\u88ab\u89c4\u6a21\u5316\u4e0e\u81ea\u52a8\u5316”\u4ecd\u662f\u4e00\u4e2a\u5f00\u653e\u7684\u7814\u7a76\u9886\u57df\u3002<\/p>\n

                                      6. \u957f\u671f\u6210\u529f\u3002\u5916\u5728\u4f18\u5316\u5faa\u73af\u4f5c\u7528\u4e8e\u5355\u6b21 rollout \u4e4b\u5916\u7684\u5956\u52b1\uff0c\u800c\u6211\u4eec\u53ea\u80fd\u5728\u8bad\u7ec3\u6c99\u7bb1\u4e2d\u6a21\u62df\u8fd9\u4e9b rollout\u3002<\/p>\n

                                      \u4ee5\u7f16\u7a0b\u667a\u80fd\u4f53\u4e3a\u4f8b\u3002\u7f16\u7a0b\u667a\u80fd\u4f53\u5df2\u7ecf\u63d0\u5347\u4e86\u8f6f\u4ef6\u5de5\u7a0b\u7684\u65e5\u5e38\u751f\u4ea7\u529b\uff0c\u4f46\u8bb8\u591a\u4f18\u5316\u76ee\u6807\u4ecd\u8fc7\u4e8e\u77ed\u671f\u3002\u5b83\u5e38\u5e38\u80fd\u5b8c\u6210\u624b\u5934\u4efb\u52a1\uff0c\u4f46\u5bf9\u4e8e”\u5982\u4f55\u7ef4\u62a4\u4e00\u4e2a\u7531\u6570\u767e\u4e0a\u5343\u540d\u5de5\u7a0b\u5e08\u5171\u540c\u7ef4\u62a4\u7684\u4ed3\u5e93\u7684\u957f\u671f\u5065\u5eb7”\uff0c\u7b54\u6848\u5c31\u4e0d\u90a3\u4e48\u660e\u663e\u4e86\u3002\u6807\u51c6\u7684\u6c99\u7bb1\u5f0f RLVR \u8bad\u7ec3\u5f88\u5c11\u6355\u6349\u5230\u53ef\u7ef4\u62a4\u6027\u3001\u6240\u6709\u6743\u8fb9\u754c\u3001\u8fc1\u79fb\u6210\u672c\u3001\u5411\u540e\u517c\u5bb9\u6027\u6216\u672a\u6765\u7684\u8c03\u8bd5\u8d1f\u62c5\u3002<\/p>\n

                                      7. \u4eba\u7684\u89d2\u8272\u3002\u4eba\u7c7b\u5e94\u5f53\u5411\u6808\u7684\u4e0a\u5c42\u79fb\u52a8\uff0c\u800c\u4e0d\u662f\u88ab\u79fb\u51fa\u5faa\u73af\uff0c\u5373\u4eba\u7c7b\u5e94\u5728\u6070\u5f53\u7684\u65f6\u95f4\u3001\u6070\u5f53\u7684\u62bd\u8c61\u5c42\u7ea7\u63d0\u4f9b\u76d1\u7763\uff0c\u6211\u4eec\u7684\u7cfb\u7edf\u8bbe\u8ba1\u5e94\u8003\u8651\u4f55\u65f6\u3001\u5982\u4f55\u8bbe\u7f6e\u8fd9\u6837\u7684\u63a5\u89e6\u70b9\u3002<\/p>\n

                                      \u4e0a\u9762\u5217\u51fa\u7684\u8bb8\u591a\u6311\u6218\u90fd\u9700\u8981\u4eba\u7c7b\u7684\u53cd\u9988\u4e0e\u5f15\u5bfc\u3002\u6bd5\u7adf\uff0c\u6211\u4eec\u6784\u5efa\u8fd9\u9879\u6280\u672f\u662f\u4e3a\u4e86\u4eba\u7c7b\u66f4\u597d\u7684\u672a\u6765\uff0c\u800c\u4e0d\u662f\u53cd\u8fc7\u6765\u3002<\/p>\n

                                      \u672c\u6587\u5f15\u7528\uff1a<\/p>\n

                                      \n
                                      \n