{"id":18411,"date":"2024-02-07T18:02:28","date_gmt":"2024-02-07T10:02:28","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=18411"},"modified":"2024-02-07T18:02:28","modified_gmt":"2024-02-07T10:02:28","slug":"%e6%96%af%e5%9d%a6%e7%a6%8f%e5%92%8copenai%e6%8f%90%e5%87%bameta-prompting%ef%bc%8c%e6%9c%80%e5%bc%ba%e9%9b%b6%e6%a0%b7%e6%9c%acprompting%e6%8a%80%e6%9c%af%e8%af%9e%e7%94%9f%e4%ba%86","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2024\/02\/07\/18411\/","title":{"rendered":"\u65af\u5766\u798f\u548cOpenAI\u63d0\u51fameta-prompting\uff0c\u6700\u5f3a\u96f6\u6837\u672cprompting\u6280\u672f\u8bde\u751f\u4e86"},"content":{"rendered":"
\u5728\u6211\u4eec\u7684\u5de5\u4f5c\u7fa4\u91cc\uff0c\u7ecf\u5e38\u4f1a\u6709\u4e00\u4f4d\u7ba1\u7406\u8005\u6765\u534f\u8c03\u6bcf\u4e2a\u4eba\u7684\u5de5\u4f5c\u5e76\u6c47\u603b\u5de5\u4f5c\u6210\u679c\u3002\u8fd1\u65e5\uff0c\u65af\u5766\u798f\u5927\u5b66\u7684 Mirac Suzgun \u548c OpenAI \u7684 Adam Tauman Kalai \u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684 prompting \u65b9\u6cd5\uff1ameta-prompting\u3002<\/div>\n
\n
\n

\u6700\u65b0\u4e00\u4ee3\u8bed\u8a00\u6a21\u578b\uff08\u5c24\u5176\u662f GPT-4\u3001PaLM \u548c LLaMa\uff09\u5df2\u7ecf\u6210\u529f\u62d3\u5c55\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u751f\u6210\u7684\u8fb9\u754c\u3002\u8fd9\u4e9b\u5927\u89c4\u6a21\u6a21\u578b\u53ef\u4ee5\u89e3\u51b3\u8bb8\u591a\u4e0d\u540c\u4efb\u52a1\uff0c\u4ece\u5199\u838e\u58eb\u6bd4\u4e9a\u98ce\u683c\u7684\u5341\u56db\u884c\u8bd7\u5230\u603b\u7ed3\u590d\u6742\u7684\u533b\u7597\u62a5\u544a\u548c\u89e3\u51b3\u7ade\u8d5b\u7ea7\u7684\u7f16\u7a0b\u95ee\u9898\u3002\u5c3d\u7ba1\u8fd9\u4e9b\u6a21\u578b\u53ef\u4ee5\u89e3\u51b3\u591a\u79cd\u591a\u6837\u7684\u95ee\u9898\uff0c\u4f46\u5b83\u4eec\u5e76\u975e\u603b\u662f\u6b63\u786e\u7684\uff0c\u6709\u65f6\u5019\u4e5f\u4f1a\u751f\u6210\u4e0d\u51c6\u786e\u3001\u8bef\u5bfc\u6027\u6216\u77db\u76fe\u7684\u54cd\u5e94\u7ed3\u679c\u3002<\/p>\n

\u968f\u7740\u8fd9\u4e9b\u6a21\u578b\u7684\u8fd0\u884c\u6210\u672c\u8d8a\u6765\u8d8a\u4f4e\uff0c\u4eba\u4eec\u81ea\u7136\u4f1a\u95ee\uff1a\u662f\u5426\u53ef\u4ee5\u4f7f\u7528\u811a\u624b\u67b6\u7cfb\u7edf\uff08scaffolding system\uff09\u5e76\u4f7f\u7528\u591a\u4e2a\u8bed\u8a00\u6a21\u578b\u67e5\u8be2\u6765\u4f18\u5316\u5e76\u4e14\u63d0\u5347\u8fd9\u4e9b\u6a21\u578b\u8f93\u51fa\u7684\u51c6\u786e\u5ea6\u548c\u7a33\u5065\u6027\u3002<\/p>\n

\u65af\u5766\u798f\u548c OpenAI \u7684\u8fd9\u9879\u7814\u7a76\u63d0\u51fa\u4e86\u4e00\u79cd\u53ef\u7528\u4e8e\u63d0\u5347\u8bed\u8a00\u6a21\u578b\u7684\u529f\u80fd\u548c\u6027\u80fd\u7684\u65b0\u6280\u672f\uff0c\u79f0\u4e3a meta-prompting\u3002<\/p>\n

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

    \n
  • \u8bba\u6587\u6807\u9898\uff1aMeta-Prompting: Enhancing Language Models with Task-Agnostic Scaffolding<\/li>\n
  • \u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2401.12954<\/li>\n
  • \u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/suzgunmirac\/meta-prompting<\/li>\n<\/ul>\n

    \u8fd9\u79cd\u6280\u672f\u6d89\u53ca\u6784\u5efa\u4e00\u4e2a\u9ad8\u5c42\u7ea7\u7684\u300c\u5143\u300d prompt\uff0c\u5176\u4f5c\u7528\u662f\u6307\u793a\u8bed\u8a00\u6a21\u578b\u505a\u5230\u4ee5\u4e0b\u51e0\u70b9\uff1a<\/p>\n

    1. \u5c06\u590d\u6742\u7684\u4efb\u52a1\u6216\u95ee\u9898\u5206\u89e3\u6210\u66f4\u5c0f\u7684\u5bb9\u6613\u89e3\u51b3\u7684\u5b50\u4efb\u52a1\uff1b<\/p>\n

    2. \u4f7f\u7528\u9002\u5f53\u4e14\u8be6\u7ec6\u7684\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\u5c06\u8fd9\u4e9b\u5b50\u4efb\u52a1\u5206\u914d\u7ed9\u4e13\u4e1a\u7684\u300c\u4e13\u5bb6\u300d\u6a21\u578b\uff1b<\/p>\n

    3. \u76d1\u7763\u8fd9\u4e9b\u4e13\u5bb6\u6a21\u578b\u4e4b\u95f4\u7684\u901a\u4fe1\uff1b<\/p>\n

    4. \u901a\u8fc7\u8fd9\u4e2a\u8fc7\u7a0b\u5e94\u7528\u5176\u81ea\u5df1\u7684\u6279\u5224\u6027\u601d\u7ef4\u3001\u63a8\u7406\u548c\u9a8c\u8bc1\u6280\u80fd\u3002<\/p>\n

    \u5bf9\u4e8e\u53ef\u4f7f\u7528 meta-prompting \u6709\u6548\u8c03\u7528\u7684\u8bed\u8a00\u6a21\u578b\uff0c\u5f53\u5bf9\u5176\u8fdb\u884c\u67e5\u8be2\u65f6\uff0c\u8be5\u6a21\u578b\u7684\u4f5c\u7528\u662f\u4f5c\u4e3a\u6307\u6325\u5458\uff08conductor\uff09\u3002\u5b83\u4f1a\u8f93\u51fa\u4e00\u7ec4\u6d88\u606f\u5386\u53f2\uff08\u6216\u8005\u79f0\u4e3a\u53d9\u8ff0\uff08narrative\uff09\uff09\uff0c\u5176\u7531\u591a\u4e2a\u4e13\u5bb6\u6a21\u578b\u7684\u54cd\u5e94\u6784\u6210\u3002\u8fd9\u4e2a\u8bed\u8a00\u6a21\u578b\u9996\u5148\u4f1a\u8d1f\u8d23\u751f\u6210\u6d88\u606f\u5386\u53f2\u4e2d\u7684\u6307\u6325\u5458\u90e8\u5206\uff0c\u8fd9\u5176\u4e2d\u5305\u62ec\u4e13\u5bb6\u7684\u9009\u53d6\u4ee5\u53ca\u4e3a\u5b83\u4eec\u6784\u5efa\u7279\u5b9a\u7684\u6307\u4ee4\u3002\u4f46\u662f\uff0c\u540c\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b\u672c\u8eab\u4e5f\u4f1a\u4f5c\u4e3a\u72ec\u7acb\u4e13\u5bb6\uff0c\u5176\u4f1a\u57fa\u4e8e\u4e13\u4e1a\u77e5\u8bc6\u4ee5\u53ca\u6307\u6325\u5458\u4e3a\u6bcf\u6761\u5177\u4f53\u67e5\u8be2\u9009\u53d6\u7684\u4fe1\u606f\u751f\u6210\u8f93\u51fa\u3002<\/p>\n

    \u8fd9\u79cd\u65b9\u6cd5\u53ef\u8ba9\u5355\u4e2a\u7edf\u4e00\u7684\u8bed\u8a00\u6a21\u578b\u7ef4\u6301\u8fde\u8d2f\u4e00\u81f4\u7684\u63a8\u7406\u8def\u7ebf\uff0c\u540c\u65f6\u8fd8\u53ef\u4ee5\u5229\u7528\u5404\u79cd\u4e0d\u540c\u7684\u4e13\u5bb6\u89d2\u8272\u3002\u901a\u8fc7\u52a8\u6001\u5730\u4e3a prompting \u9009\u62e9\u4e0a\u4e0b\u6587\uff0c\u8fd9\u4e9b\u4e13\u5bb6\u80fd\u4e3a\u8be5\u8fc7\u7a0b\u5f15\u5165\u5168\u65b0\u7684\u89c6\u89d2\uff0c\u540c\u65f6\u6307\u6325\u5458\u6a21\u578b\u8fd8\u80fd\u4fdd\u6301\u5bf9\u5b8c\u6574\u5386\u53f2\u7684\u9e1f\u77b0\u89c6\u89d2\u5e76\u7ef4\u6301\u534f\u8c03\u3002<\/p>\n

    \u56e0\u6b64\uff0c\u8fd9\u79cd\u65b9\u6cd5\u80fd\u8ba9\u5355\u4e2a\u9ed1\u7bb1\u8bed\u8a00\u6a21\u578b\u65e2\u6709\u6548\u4f5c\u4e3a\u4e2d\u5fc3\u6307\u6325\u5458\uff0c\u540c\u65f6\u53c8\u5145\u5f53\u4e00\u7cfb\u5217\u4e0d\u540c\u4e13\u5bb6\uff0c\u8fd9\u6837\u4fbf\u53ef\u4ee5\u5f97\u5230\u66f4\u52a0\u51c6\u786e\u3001\u53ef\u9760\u548c\u8fde\u8d2f\u4e00\u81f4\u7684\u54cd\u5e94\u3002<\/p>\n

    \u8fd9\u91cc\u65b0\u63d0\u51fa\u7684 meta-prompting \u6280\u672f\u7ec4\u5408\u5e76\u6269\u5c55\u4e86\u8fd1\u671f\u7814\u7a76\u63d0\u51fa\u7684\u591a\u79cd\u4e0d\u540c\u7684 prompting \u601d\u60f3\uff0c\u5305\u62ec\u9ad8\u5c42\u7ea7\u89c4\u5212\u548c\u51b3\u7b56\u3001\u52a8\u6001\u4eba\u8bbe\u5206\u914d\u3001\u591a\u667a\u80fd\u4f53\u8fa9\u8bba\u3001\u81ea\u6211\u8c03\u8bd5\u548c\u81ea\u6211\u53cd\u601d\u3002<\/p>\n

    meta-prompting \u7684\u4e00\u4e2a\u5173\u952e\u65b9\u9762\u662f\u5176\u5177\u6709\u4e00\u4e2a\u6027\u8d28\uff1a\u4e0d\u53d7\u5177\u4f53\u4efb\u52a1\u5f71\u54cd\u3002<\/p>\n

    \u4f20\u7edf\u7684\u811a\u624b\u67b6\u65b9\u6cd5\u9700\u8981\u9488\u5bf9\u6bcf\u4e2a\u4efb\u52a1\u8c03\u6574\u5177\u4f53\u7684\u6307\u4ee4\u6216\u793a\u4f8b\uff0c\u800c meta-prompting \u5219\u4e0d\u540c\uff0c\u5176\u5728\u591a\u79cd\u4efb\u52a1\u548c\u8f93\u5165\u4e0a\u90fd\u91c7\u7528\u4e86\u540c\u4e00\u5957\u9ad8\u5c42\u7ea7\u6307\u4ee4\u3002\u5bf9\u6015\u9ebb\u70e6\u7684\u7528\u6237\u6765\u8bf4\uff0c\u8fd9\u79cd\u901a\u7528\u6027\u5c24\u5176\u6709\u76ca\uff0c\u56e0\u4e3a\u8fd9\u6837\u5c31\u4e0d\u5fc5\u4e3a\u6bcf\u4e2a\u5177\u4f53\u4efb\u52a1\u63d0\u4f9b\u8be6\u7ec6\u7684\u793a\u4f8b\u6216\u5177\u4f53\u6307\u793a\u4e86\u3002<\/p>\n

    \u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5bf9\u4e8e\u300c\u5199\u4e00\u9996\u5173\u4e8e\u81ea\u62cd\u7684\u838e\u58eb\u6bd4\u4e9a\u5f0f\u5341\u56db\u884c\u8bd7\u300d\u8fd9\u6837\u7684\u4e00\u6b21\u6027\u8bf7\u6c42\uff0c\u7528\u6237\u65e0\u9700\u8865\u5145\u9ad8\u8d28\u91cf\u7684\u65b0\u53e4\u5178\u4e3b\u4e49\u8bd7\u6b4c\u793a\u4f8b\u3002<\/p>\n

    meta-prompting \u65b9\u6cd5\u80fd\u63d0\u4f9b\u4e00\u79cd\u5e7f\u6cdb\u3001\u7075\u6d3b\u7684\u6846\u67b6\uff0c\u800c\u53c8\u4e0d\u4f1a\u5f71\u54cd\u5176\u7279\u5b9a\u6027\u6216\u76f8\u5173\u6027\uff0c\u4ece\u800c\u53ef\u4ee5\u63d0\u5347\u8bed\u8a00\u6a21\u578b\u7684\u5b9e\u7528\u6027\u3002\u6b64\u5916\uff0c\u4e3a\u4e86\u5c55\u73b0 meta-prompting \u65b9\u6cd5\u7684\u901a\u7528\u6027\u548c\u6574\u5408\u80fd\u529b\uff0c\u8be5\u56e2\u961f\u8fd8\u5bf9\u5176\u7cfb\u7edf\u8fdb\u884c\u4e86\u589e\u5f3a\uff0c\u4f7f\u5176\u53ef\u4ee5\u8c03\u7528 Python \u89e3\u91ca\u5668\u3002\u5982\u6b64\u4e00\u6765\uff0c\u8be5\u6280\u672f\u5c31\u80fd\u652f\u6301\u66f4\u52a0\u52a8\u6001\u548c\u5168\u9762\u7684\u5e94\u7528\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u63d0\u5347\u5176\u6709\u6548\u5904\u7406\u591a\u79cd\u4efb\u52a1\u548c\u67e5\u8be2\u7684\u6f5c\u529b\u3002<\/p>\n

    \u56fe 2 \u5c55\u793a\u4e86\u4e00\u4e2a meta-prompting \u7684\u4f1a\u8bdd\u6d41\u7a0b\u793a\u4f8b\u3002<\/p>\n

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

    \u5176\u63cf\u7ed8\u4e86\u5143\u6a21\u578b\uff08Meta Model\uff0c\u5373\u6307\u6325\u5458\u6a21\u578b\uff09\u4f7f\u7528\u8f93\u5165\u548c\u6765\u81ea\u591a\u4e2a\u4e0d\u540c\u7684\u4e13\u4e1a\u4e13\u5bb6\u6a21\u578b\u6216\u4ee3\u7801\u6267\u884c\u7684\u8f93\u51fa\u89e3\u8bfb\u5176\u81ea\u8eab\u8f93\u51fa\u7684\u8fc7\u7a0b\u3002\u8fd9\u6837\u7684\u914d\u7f6e\u8ba9 meta-prompting \u6210\u4e3a\u4e86\u4e00\u4e2a\u8fd1\u4e4e\u901a\u7528\u7684\u5de5\u5177\u3002\u5176\u5141\u8bb8\u5c06\u591a\u4e2a\u8bed\u8a00\u6a21\u578b\u7684\u4ea4\u4e92\u548c\u8ba1\u7b97\u805a\u5408\u6210\u5355\u4e00\u4e14\u8fde\u8d2f\u7684\u53d9\u8ff0\u3002meta-prompting \u7684\u4e0d\u540c\u4e4b\u5904\u5728\u4e8e\u5176\u8ba9\u8bed\u8a00\u6a21\u578b\u81ea\u5df1\u51b3\u5b9a\u8981\u4f7f\u7528\u54ea\u4e9b prompt \u6216\u4f7f\u7528\u54ea\u4e9b\u4ee3\u7801\u6bb5\u3002<\/p>\n

    \u8be5\u56e2\u961f\u4f7f\u7528 GPT-4 \u4f5c\u4e3a\u57fa\u7840\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u4e86\u5168\u9762\u7684\u5b9e\u9a8c\uff0c\u6bd4\u8f83\u4e86 meta-prompting \u4e0e\u5176\u5b83\u65e0\u5173\u4efb\u52a1\u578b\u811a\u624b\u67b6\u65b9\u6cd5\u3002<\/p>\n

    \u5b9e\u9a8c\u53d1\u73b0\uff0cmeta-prompting \u4e0d\u4ec5\u80fd\u63d0\u5347\u6574\u4f53\u6027\u80fd\uff0c\u800c\u4e14\u5728\u591a\u4e2a\u4e0d\u540c\u4efb\u52a1\u4e0a\u4e5f\u5f80\u5f80\u80fd\u5b9e\u73b0\u65b0\u7684\u6700\u4f73\u7ed3\u679c\u3002\u5176\u7075\u6d3b\u6027\u5c24\u5176\u503c\u5f97\u79f0\u9053\uff1a\u6307\u6325\u5458\u6a21\u578b\u6709\u80fd\u529b\u8c03\u7528\u4e13\u5bb6\u6a21\u578b\uff08\u57fa\u672c\u4e0a\u5c31\u662f\u5176\u672c\u8eab\uff0c\u53ea\u662f\u6307\u4ee4\u4e0d\u4e00\u6837\uff09\u6267\u884c\u591a\u79cd\u4e0d\u540c\u7684\u529f\u80fd\u3002\u8fd9\u4e9b\u529f\u80fd\u53ef\u80fd\u5305\u62ec\u70b9\u8bc4\u4e4b\u524d\u7684\u8f93\u51fa\u3001\u4e3a\u7279\u5b9a\u4efb\u52a1\u9009\u53d6\u7279\u5b9a AI \u4eba\u8bbe\u3001\u4f18\u5316\u751f\u6210\u7684\u5185\u5bb9\u3001\u786e\u4fdd\u6700\u7ec8\u8f93\u51fa\u5728\u5b9e\u8d28\u548c\u5f62\u5f0f\u4e0a\u90fd\u6ee1\u8db3\u6240\u9700\u6807\u51c6\u3002<\/p>\n

    \u5982\u56fe 1 \u6240\u793a\uff0c\u76f8\u6bd4\u4e4b\u524d\u7684\u591a\u79cd\u65b9\u6cd5\uff0c\u65b0\u65b9\u6cd5\u7684\u63d0\u5347\u5f88\u660e\u663e\u3002<\/p>\n

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

    meta-prompting<\/h4>\n

    \u76f4\u89c9\u77e5\u8bc6\u548c\u62bd\u8c61\u6982\u8ff0\u3002meta-prompting \u7684\u5de5\u4f5c\u65b9\u6cd5\u662f\u4f7f\u7528\u4e00\u4e2a\u6a21\u578b\u6765\u534f\u8c03\u548c\u6267\u884c\u591a\u4e2a\u72ec\u7acb\u67e5\u8be2\uff0c\u7136\u540e\u5c06\u5b83\u4eec\u7684\u54cd\u5e94\u7efc\u5408\u8d77\u6765\uff0c\u8fdb\u800c\u6e32\u67d3\u5f97\u5230\u4e00\u4e2a\u6700\u7ec8\u54cd\u5e94\u3002\u4ece\u539f\u7406\u4e0a\u8bb2\uff0c\u8be5\u673a\u5236\u91c7\u7528\u4e86\u4e00\u79cd\u96c6\u6210\u65b9\u6cd5\uff0c\u5373\u501f\u7528\u72ec\u7acb\u4e13\u4e1a\u6a21\u578b\u7684\u529b\u91cf\u548c\u591a\u6837\u6027\u6765\u534f\u4f5c\u89e3\u51b3\u548c\u5904\u7406\u6d89\u53ca\u591a\u65b9\u9762\u7684\u4efb\u52a1\u6216\u95ee\u9898\u3002<\/p>\n

    meta-prompting \u7b56\u7565\u7684\u6838\u5fc3\u662f\u5176\u6d45\u5c42\u7684\u7ed3\u6784\uff0c\u5176\u4e2d\u4f7f\u7528\u4e00\u4e2a\u5355\u4e00\u6a21\u578b\uff08\u79f0\u4e3a\u5143\u6a21\u578b\uff09\u4f5c\u4e3a\u6743\u5a01\u7684\u4e3b\u5b9e\u4f53\u3002<\/p>\n

    \u8fd9\u79cd prompting \u7ed3\u6784\u7c7b\u4f3c\u4e8e\u7ba1\u5f26\u4e50\u961f\uff0c\u5176\u4e2d\u6307\u6325\u5bb6\u7684\u89d2\u8272\u5c31\u7531\u5143\u6a21\u578b\u5145\u5f53\uff0c\u6bcf\u4f4d\u97f3\u4e50\u6f14\u594f\u8005\u90fd\u5bf9\u5e94\u4e00\u4e2a\u4e0d\u540c\u7684\u7279\u5b9a\u9886\u57df\u7684\u6a21\u578b\u3002\u6b63\u5982\u6307\u6325\u5bb6\u53ef\u4ee5\u8ba9\u591a\u79cd\u4e50\u5668\u534f\u8c03\u5f39\u594f\u51fa\u548c\u8c10\u7684\u65cb\u5f8b\u4e00\u6837\uff0c\u5143\u6a21\u578b\u4e5f\u53ef\u4ee5\u5c06\u591a\u4e2a\u6a21\u578b\u7684\u89e3\u7b54\u548c\u89c1\u89e3\u7ec4\u5408\u8d77\u6765\uff0c\u4e3a\u590d\u6742\u7684\u95ee\u9898\u6216\u4efb\u52a1\u7ed9\u51fa\u51c6\u786e\u4e14\u5168\u9762\u7684\u89e3\u7b54\u3002<\/p>\n

    \u4ece\u6982\u5ff5\u4e0a\u8bb2\uff0c\u5728\u8fd9\u4e2a\u6846\u67b6\u5185\uff0c\u7279\u5b9a\u9886\u57df\u7684\u4e13\u5bb6\u53ef\u4ee5\u6709\u591a\u79cd\u591a\u6837\u7684\u5f62\u5f0f\uff0c\u6bd4\u5982\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u5fae\u8c03\u8fc7\u7684\u8bed\u8a00\u6a21\u578b\u3001\u7528\u4e8e\u5904\u7406\u7279\u5b9a\u7c7b\u578b\u67e5\u8be2\u7684\u4e13\u7528 API\uff0c\u751a\u81f3\u8fd8\u53ef\u4ee5\u662f\u8ba1\u7b97\u5668\u8fd9\u6837\u7684\u8ba1\u7b97\u5de5\u5177\u6216\u7528\u4e8e\u6267\u884c\u4ee3\u7801\u7684 Python \u89e3\u91ca\u5668\u7b49\u4ee3\u7801\u5de5\u5177\u3002\u8fd9\u4e9b\u529f\u80fd\u5404\u5f02\u7684\u4e13\u5bb6\u90fd\u5728\u5143\u6a21\u578b\u7684\u76d1\u7763\u4e0b\u63a5\u53d7\u6307\u793a\u548c\u7edf\u4e00\uff0c\u65e0\u6cd5\u76f4\u63a5\u76f8\u4e92\u4e92\u52a8\u6216\u4ea4\u6d41\u3002<\/p>\n

    Algorithmic Procedure. \u7b97\u6cd5 1 \u7ed9\u51fa\u4e86\u65b0\u63d0\u51fa\u7684 meta-prompting \u65b9\u6cd5\u7684\u4f2a\u4ee3\u7801\u3002<\/p>\n

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

    \u7b80\u5355\u603b\u7ed3\u4e00\u4e0b\uff0c\u9996\u5148\u662f\u5bf9\u8f93\u5165\u6267\u884c\u53d8\u6362\uff0c\u4f7f\u5176\u7b26\u5408\u9002\u5f53\u7684\u6a21\u677f\uff1b\u7136\u540e\u6267\u884c\u4ee5\u4e0b\u5faa\u73af\uff1a(a) \u5411\u5143\u6a21\u578b\u63d0\u4ea4 prompt\uff0c(b) \u5982\u6709\u9700\u8981\uff0c\u4f7f\u7528\u7279\u5b9a\u9886\u57df\u7684\u4e13\u5bb6\u6a21\u578b\uff0c(c) \u8fd4\u56de\u6700\u7ec8\u54cd\u5e94\uff0c(d) \u5904\u7406\u9519\u8bef\u3002<\/p>\n

    \u9700\u8981\u6307\u51fa\uff0c\u8be5\u56e2\u961f\u5728\u5b9e\u9a8c\u4e2d\u91c7\u7528\u7684\u5143\u6a21\u578b\u548c\u4e13\u5bb6\u6a21\u578b\u90fd\u662f GPT-4\u3002\u5b83\u4eec\u7684\u89d2\u8272\u5dee\u5f02\u662f\u7531\u5404\u81ea\u6536\u5230\u7684\u6307\u4ee4\u786e\u5b9a\u7684\uff1b\u5176\u4e2d\u5143\u6a21\u578b\u9075\u5faa\u56fe 3 \u63d0\u4f9b\u7684\u4e00\u7ec4\u6307\u4ee4\uff0c\u800c\u4e13\u5bb6\u6a21\u578b\u5219\u9075\u4ece\u5143\u6a21\u578b\u5728\u63a8\u7406\u65f6\u95f4\u52a8\u6001\u786e\u5b9a\u7684\u6307\u4ee4\u3002<\/p>\n

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

    \u5b9e\u9a8c\u8bbe\u7f6e<\/h4>\n

    \u57fa\u51c6<\/strong><\/p>\n

    \u8be5\u56e2\u961f\u6bd4\u8f83\u4e86 meta-prompting \u4e0e\u4ee5\u4e0b prompting \u65b9\u6cd5\u7684\u65e0\u5173\u4efb\u52a1\u578b\u96f6\u6837\u672c\u5f0f\u7248\u672c\uff1a<\/p>\n

      \n
    • \u6807\u51c6 prompting<\/li>\n
    • \u96f6\u6837\u672c\u601d\u7ef4\u94fe prompting<\/li>\n
    • \u4e13\u5bb6 prompting<\/li>\n
    • \u591a\u4eba\u8bbe prompting<\/li>\n<\/ul>\n

      \u6570\u636e\u96c6\u548c\u4efb\u52a1<\/strong><\/p>\n

      \u8be5\u56e2\u961f\u5728\u5b9e\u9a8c\u4e2d\u91c7\u7528\u4e86\u591a\u79cd\u4efb\u52a1\u548c\u6570\u636e\u96c6\uff0c\u5b83\u4eec\u9700\u8981\u591a\u79cd\u4e0d\u540c\u80fd\u529b\uff0c\u6bd4\u5982\u6570\u5b66\u548c\u7b97\u6cd5\u63a8\u7406\u3001\u7279\u5b9a\u9886\u57df\u77e5\u8bc6\u548c\u6587\u5b66\u521b\u9020\u529b\u3002\u8fd9\u4e9b\u6570\u636e\u96c6\u548c\u4efb\u52a1\u5305\u62ec\uff1a<\/p>\n

        \n
      • Game of 24\uff1a\u76ee\u6807\u662f\u4f7f\u7528\u56db\u4e2a\u7ed9\u5b9a\u6570\u503c\uff08\u6bcf\u4e2a\u53ea\u80fd\u4f7f\u7528\u4e00\u6b21\uff09\u6784\u5efa\u4e00\u4e2a\u7ed3\u679c\u4e3a 24 \u7684\u7b97\u672f\u8868\u8fbe\u5f0f\u3002<\/li>\n
      • \u4e09\u4e2a BIG-Bench Hard\uff08BBH\uff09\u4efb\u52a1\uff1aGeometric Shapes\u3001MultiStep Arithmetic Two \u548c Word Sorting\uff1b\u53e6\u5916\u8fd8\u6709\u4e00\u4e2a\u76f4\u63a5\u4ece BIG-Bench \u5957\u4ef6\u83b7\u53d6\u7684\u63a8\u7406\u4efb\u52a1 Checkmate-in-One\u3002<\/li>\n
      • Python Programming Puzzles\uff08P3\uff09\uff0c\u5373 Python \u7f16\u7a0b\u9898\uff0c\u5305\u542b\u591a\u4e2a\u96be\u5ea6\u3002<\/li>\n
      • Multilingual Grade School Math\uff0c\u5373\u591a\u8bed\u8a00\u5c0f\u5b66\u6570\u5b66\uff0c\u8fd9\u662f GSM8K \u6570\u636e\u96c6\u7684\u4e00\u4e2a\u591a\u8bed\u8a00\u7248\u672c\uff0c\u5305\u542b\u5b5f\u52a0\u62c9\u8bed\u3001\u65e5\u8bed\u548c\u65af\u74e6\u5e0c\u91cc\u8bed\u7b49\u8bed\u8a00\u3002<\/li>\n
      • Shakespearean Sonnet Writing\uff0c\u5373\u838e\u58eb\u6bd4\u4e9a\u5f0f\u5341\u56db\u884c\u8bd7\u5199\u4f5c\uff0c\u8fd9\u662f\u8be5\u56e2\u961f\u521b\u5efa\u7684\u4e00\u4e2a\u65b0\u4efb\u52a1\uff0c\u76ee\u6807\u662f\u5199\u51fa\u6309\u300cABAB CDCD EFEF GG\u300d\u4e25\u683c\u62bc\u97f5\u7684\u5341\u56db\u884c\u8bd7\uff0c\u5176\u4e2d\u5e94\u4e00\u8bcd\u4e0d\u5dee\u5730\u5305\u542b\u6240\u63d0\u4f9b\u7684\u4e09\u4e2a\u8bcd\u3002<\/li>\n<\/ul>\n

        \u7b54\u6848\u63d0\u53d6\u548c\u8bc4\u4f30\u534f\u8bae<\/strong><\/p>\n

        \u5982\u56fe 3 \u6240\u793a\uff0c\u5bf9\u4e8e\u65b0\u63d0\u51fa\u7684 meta-prompting \u65b9\u6cd5\uff0c\u7cfb\u7edf\u6307\u4ee4\u4f1a\u9f13\u52b1\u5143\u6a21\u578b\u4ee5\u7279\u5b9a\u683c\u5f0f\u7ed9\u51fa\u6700\u7ec8\u7b54\u6848\u3002<\/p>\n

        \u81f3\u4e8e\u8bc4\u4f30\uff0c\u5219\u4f1a\u6839\u636e\u4efb\u52a1\u7684\u6027\u8d28\u548c\u5f62\u5f0f\uff0c\u91c7\u7528\u4ee5\u4e0b\u4e09\u4e2a\u6307\u6807\u4e4b\u4e00\uff1a<\/p>\n

          \n
        • Exact Match (EM)\uff0c\u7cbe\u786e\u5339\u914d<\/li>\n
        • Soft Match (SM)\uff0c\u8f6f\u5339\u914d<\/li>\n
        • Functionally Correct (FC)\uff0c\u529f\u80fd\u6b63\u786e\u6027<\/li>\n<\/ul>\n

          \u6a21\u578b\u548c\u63a8\u7406<\/strong><\/p>\n

          \u8be5\u56e2\u961f\u7684\u4e3b\u8981\u5b9e\u9a8c\u90fd\u4f7f\u7528\u4e86 GPT-4\uff08gpt-4-32k\uff09\u3002\u4e00\u4e9b\u8865\u5145\u5b9e\u9a8c\u5219\u4f7f\u7528\u4e86 GPT-3.5\uff08gpt-35-turbo\uff09\u3002\u4e0d\u7ba1\u662f GPT-3.5 \u8fd8\u662f GPT-4\uff0c\u90fd\u4f7f\u7528\u4e86\u4ee5\u4e0b\u6307\u4ee4\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n

          \u5728\u5168\u90e8\u5b9e\u9a8c\u4e2d\uff0c\u5143\u6a21\u578b\u4f7f\u7528\u7684\u53c2\u6570\u548c\u7cfb\u7edf\u6307\u4ee4\u90fd\u662f\u4e00\u6837\u7684\u3002\u6e29\u5ea6\u503c\u8bbe\u7f6e\u4e3a 0\uff0ctop-p \u503c\u8bbe\u7f6e\u4e3a 0.95\uff0c\u6700\u5927 token \u6570\u4e3a 1024\u3002<\/p>\n

          \u4e3b\u8981\u7ed3\u679c\u548c\u8ba8\u8bba<\/h4>\n

          \u8868 1 \u603b\u7ed3\u4e86\u5b9e\u9a8c\u7ed3\u679c\uff0c\u65b0\u63d0\u51fa\u7684 meta-prompting \u7684\u4f18\u8d8a\u6027\u5f97\u5230\u4e86\u4f53\u73b0\u3002<\/p>\n

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

          \u89c2\u5bdf\u8fd9\u4e9b\u65b9\u6cd5\u5728\u6240\u6709\u4efb\u52a1\u4e0a\u7684\u603b\u4f53\u6027\u80fd\uff0c\u53ef\u4ee5\u770b\u5230 meta-prompting \u4e3a\u51c6\u786e\u5ea6\u5e26\u6765\u7684\u663e\u8457\u63d0\u5347\uff0c\u5c24\u5176\u662f\u4f7f\u7528\u4e86 Python \u89e3\u91ca\u5668\u5de5\u5177\u8f85\u52a9\u65f6\u3002<\/p>\n

          \u5177\u4f53\u6765\u8bf4\uff0cmeta-prompting \u65b9\u6cd5\u80dc\u8fc7\u6807\u51c6 prompting \u65b9\u6cd5 17.1%\uff0c\u8d85\u8fc7\u4e13\u5bb6\uff08\u52a8\u6001\uff09 prompting 17.3%\uff0c\u4e5f\u6bd4\u591a\u4eba\u8bbe prompting \u4f18\u79c0 15.2%\u3002<\/p>\n

          \u53e6\u5916\u4ece\u56fe 4 \u548c 5 \u53ef\u4ee5\u770b\u5230\uff0c\u76f8\u6bd4\u4e8e\u4e0d\u4f7f\u7528 Python \u89e3\u91ca\u5668\u7684 meta-prompting\uff0c\u6574\u5408 Python \u89e3\u91ca\u5668\u65f6\uff0c\u5728\u4e0d\u540c\u4efb\u52a1\u4e0a\u7684\u6574\u4f53\u6027\u80fd\u53ef\u83b7\u5f97 11.5% \u7684\u63d0\u5347\u3002<\/p>\n

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

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

          \u8be5\u56e2\u961f\u8fd8\u5728\u8bba\u6587\u4e2d\u6df1\u5165\u8ba8\u8bba\u4e86\u4ece\u5b9e\u9a8c\u4e2d\u5f97\u5230\u4e86\u5173\u952e\u89c1\u89e3\uff0c\u5305\u62ec meta-prompting \u7684\u6027\u80fd\u4f18\u8d8a\u6027\u3001\u96f6\u6837\u672c\u5206\u89e3\u80fd\u529b\u3001\u9519\u8bef\u68c0\u6d4b\u3001\u4fe1\u606f\u805a\u5408\u548c\u4ee3\u7801\u6267\u884c\u7b49\u3002\u8fd9\u91cc\u6211\u4eec\u5c31\u4e0d\u8be6\u7ec6\u8bf4\u660e\u4e86\uff0c\u4f46 Fresh Eyes \u8fd9\u4e00\u6982\u5ff5\u5012\u662f\u503c\u5f97\u4ecb\u7ecd\u4e00\u756a\u3002<\/p>\n

          Fresh Eyes \u4e5f\u5c31\u662f\u7528\u53e6\u4e00\u53cc\u773c\u775b\u770b\uff0c\u8fd9\u6709\u52a9\u4e8e\u7f13\u89e3\u8bed\u8a00\u6a21\u578b\u7684\u4e00\u4e2a\u4f17\u6240\u5468\u77e5\u7684\u95ee\u9898\uff1a\u72af\u9519\u65f6\u4f1a\u4e00\u8def\u9519\u5230\u5e95\u5e76\u4e14\u4f1a\u8868\u73b0\u51fa\u8fc7\u5ea6\u81ea\u4fe1\u3002<\/p>\n

          Fresh Eyes \u662f meta-prompting \u4e0e\u591a\u4eba\u8bbe prompting \u7684\u4e00\u5927\u5173\u952e\u5dee\u5f02\uff0c\u5e76\u4e14\u5b9e\u9a8c\u7ed3\u679c\u4e5f\u8bc1\u660e\u4e86\u5176\u5177\u6709\u4f18\u52bf\u3002\u5728 meta-prompting \u4e2d\uff0c\u53ef\u4ee5\u4f7f\u7528\u4e13\u5bb6\uff08\u6216\u4eba\u8bbe\uff09\u6765\u5bf9\u95ee\u9898\u8fdb\u884c\u91cd\u65b0\u8bc4\u4f30\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u673a\u4f1a\u5f97\u5230\u5168\u65b0\u7684\u89c1\u89e3\uff0c\u4ece\u800c\u6709\u671b\u53d1\u73b0\u4e4b\u524d\u672a\u88ab\u53d1\u73b0\u6709\u8bef\u7684\u89e3\u7b54\u3002<\/p>\n

          \u57fa\u4e8e\u8ba4\u77e5\u5fc3\u7406\u5b66\uff0cFresh Eyes \u53ef\u4ee5\u5e26\u6765\u66f4\u5177\u521b\u9020\u6027\u7684\u95ee\u9898\u6c42\u89e3\u548c\u9519\u8bef\u68c0\u6d4b\u7ed3\u679c\u3002<\/p>\n

          \u4e0b\u9762\u7684\u4f8b\u5b50\u5c55\u793a\u4e86 Fresh Eyes \u5728\u5b9e\u8df5\u4e2d\u7684\u597d\u5904\u3002\u5047\u8bbe\u4efb\u52a1\u662f Game of 24\uff0c\u63d0\u4f9b\u7684\u6570\u503c\u662f 6\u300111\u300112 \u548c 13\uff0c\u8981\u6c42\u6784\u5efa\u4e00\u4e2a\u80fd\u8ba9\u7ed3\u679c\u4e3a 24 \u7684\u7b97\u672f\u8868\u8fbe\u5f0f\u5e76\u4e14\u6bcf\u4e2a\u6570\u53ea\u80fd\u7528\u4e00\u6b21\u3002\u5176\u5386\u53f2\u8fc7\u7a0b\u53ef\u80fd\u4f1a\u662f\u8fd9\u6837\uff1a<\/p>\n

          1. \u5143\u6a21\u578b\u63d0\u8bae\u54a8\u8be2\u89e3\u7b54\u6570\u5b66\u95ee\u9898\u7684\u4e13\u5bb6\u6a21\u578b\u548c\u4f7f\u7528 Python \u7f16\u7a0b\u3002\u5b83\u5f3a\u8c03\u4e86\u5bf9\u51c6\u786e\u5ea6\u548c\u9075\u5b88\u7ea6\u675f\u6761\u4ef6\u7684\u5fc5\u8981\u6027\uff0c\u5e76\u5efa\u8bae\u5982\u6709\u9700\u8981\u53ef\u8ba9\u53e6\u4e00\u4e2a\u4e13\u5bb6\u53c2\u4e0e\u8fdb\u6765\u3002<\/p>\n

          2. \u4e00\u4e2a\u4e13\u5bb6\u7ed9\u51fa\u4e86\u4e00\u4e2a\u89e3\u7b54\uff0c\u800c\u53e6\u4e00\u4e2a\u4e13\u5bb6\u5219\u8ba4\u4e3a\u5176\u4e0d\u5bf9\uff0c\u4e8e\u662f\u5143\u6a21\u578b\u5efa\u8bae\u5199\u4e00\u4e2a Python \u7a0b\u5e8f\u6765\u5bfb\u627e\u6709\u6548\u7684\u89e3\u3002<\/p>\n

          3. \u54a8\u8be2\u4e00\u4e2a\u7f16\u7a0b\u4e13\u5bb6\uff0c\u8ba9\u5176\u5199\u4e00\u4e2a\u7a0b\u5e8f\u3002<\/p>\n

          4. \u53e6\u4e00\u4e2a\u7f16\u7a0b\u4e13\u5bb6\u5728\u811a\u672c\u4e2d\u53d1\u73b0\u4e86\u4e00\u4e2a\u9519\u8bef\uff0c\u7136\u540e\u5bf9\u5176\u8fdb\u884c\u4fee\u6539\u5e76\u6267\u884c\u4fee\u6539\u540e\u7684\u811a\u672c\u3002<\/p>\n

          5. \u518d\u54a8\u8be2\u4e00\u4e2a\u6570\u5b66\u4e13\u5bb6\uff0c\u8ba9\u5176\u9a8c\u8bc1\u8be5\u7a0b\u5e8f\u8f93\u51fa\u7684\u89e3\u3002<\/p>\n

          6. \u9a8c\u8bc1\u5b8c\u6210\u540e\uff0c\u7531\u5143\u6a21\u578b\u5c06\u5176\u8f93\u51fa\u4f5c\u4e3a\u6700\u7ec8\u7b54\u6848\u3002<\/p>\n

          \u8fd9\u4e2a\u793a\u4f8b\u5c55\u73b0\u4e86 meta-prompting \u5982\u4f55\u5728\u6bcf\u4e00\u6b65\u7eb3\u5165\u65b0\u89c2\u70b9\uff0c\u8fd9\u6837\u4e0d\u4ec5\u80fd\u627e\u5230\u89e3\u7b54\uff0c\u800c\u4e14\u8fd8\u80fd\u6709\u6548\u8bc6\u522b\u548c\u7ea0\u6b63\u9519\u8bef\u3002<\/p>\n

          \u8be5\u56e2\u961f\u6700\u540e\u8ba8\u8bba\u4e86\u4e00\u4e9b\u4e0e meta-prompting \u6709\u5173\u7684\u5176\u5b83\u95ee\u9898\uff0c\u5305\u62ec\u5bf9\u6240\u4f7f\u7528\u7684\u4e13\u5bb6\u7c7b\u578b\u7684\u5206\u6790\u3001\u83b7\u5f97\u6700\u7ec8\u7ed3\u679c\u6240\u9700\u7684\u5bf9\u8bdd\u8f6e\u6570\u4ee5\u53ca\u5982\u4f55\u5e94\u5bf9\u65e0\u89e3\u95ee\u9898\u7b49\u60c5\u51b5\u3002\u8be6\u60c5\u8bf7\u53c2\u9605\u539f\u8bba\u6587\u3002<\/p>\n<\/div>\n<\/div>\n

          <\/div>\n

          <\/g><\/svg><\/i> \"Loading\"<\/p>\n

          <\/div>\n","protected":false},"excerpt":{"rendered":"

          \u5728\u6211\u4eec\u7684\u5de5\u4f5c\u7fa4\u91cc\uff0c\u7ecf\u5e38\u4f1a\u6709\u4e00\u4f4d\u7ba1\u7406\u8005\u6765\u534f\u8c03\u6bcf\u4e2a\u4eba\u7684\u5de5\u4f5c\u5e76\u6c47\u603b\u5de5\u4f5c\u6210\u679c\u3002\u8fd1\u65e5\uff0c\u65af\u5766\u798f\u5927\u5b66\u7684 Mirac Suzg […]<\/p>\n

          <\/div>\n

          <\/g><\/svg><\/i> \"Loading\"<\/p>\n

          <\/div>\n","protected":false},"author":56,"featured_media":18412,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,80],"tags":[124,887],"class_list":["post-18411","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-80","tag-ai","tag-887"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18411","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=18411"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18411\/revisions"}],"predecessor-version":[{"id":18420,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/18411\/revisions\/18420"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/18412"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=18411"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=18411"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=18411"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}