{"id":99151,"date":"2026-09-02T00:16:59","date_gmt":"2026-09-01T16:16:59","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/99151.html"},"modified":"2026-09-02T00:16:59","modified_gmt":"2026-09-01T16:16:59","slug":"%e6%a8%a1%e5%9e%8b%e5%90%8e%e8%ae%ad%e7%bb%83%ef%bc%9a%e4%bb%8e-sft-%e5%88%b0%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e7%9a%84%e5%b7%a5%e7%a8%8b%e5%ae%9e%e8%b7%b5","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/99151.html","title":{"rendered":"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5"},"content":{"rendered":"<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161655-6a96fa7748e8d.jpg\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h2>\u6a21\u578b\u540e\u8bad\u7ec3&#xff1a;\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5<\/h2>\n<p>Agent &#061; LLM &#043; \u4e0a\u4e0b\u6587 &#043; \u5de5\u5177\u3002\u5f53\u4e0a\u4e0b\u6587\u5de5\u7a0b\u548c\u5de5\u5177\u8bbe\u8ba1\u5df2\u7ecf\u505a\u5230\u4f4d&#xff0c;\u5269\u4e0b\u7684\u74f6\u9888\u5f80\u5f80\u5728 LLM \u672c\u8eab\u2014\u2014\u8fd9\u65f6\u9700\u8981\u901a\u8fc7\u540e\u8bad\u7ec3\u6539\u52a8\u6a21\u578b\u6743\u91cd&#xff0c;\u628a\u80fd\u529b\u6c89\u6dc0\u8fdb\u53c2\u6570\u3002\u672c\u7ae0\u7cfb\u7edf\u8ba8\u8bba\u4e86\u4ece\u9884\u8bad\u7ec3\u5230 RL \u7684\u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u6d41\u6c34\u7ebf&#xff0c;\u4ee5\u53ca SFT \u4e0e RL \u5728\u5de5\u7a0b\u5b9e\u8df5\u4e2d\u7684\u672c\u8d28\u533a\u522b\u3001\u6570\u636e\u5408\u6210\u65b9\u6cd5\u548c\u73af\u5883\u5de5\u7a0b\u8981\u70b9\u3002<\/p>\n<h3>\u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u5168\u666f<\/h3>\n<p>\u73b0\u4ee3\u6a21\u578b\u7684\u80fd\u529b\u5f00\u53d1\u901a\u5e38\u62c6\u6210\u56db\u4e2a\u73af\u8282&#xff1a;<\/p>\n<table>\n<tr>\u9636\u6bb5\u7528\u4ec0\u4e48\u6570\u636e\u4f18\u5316\u76ee\u6807\u5b66\u5230\u4ec0\u4e48\u5178\u578b\u4ee3\u4ef7<\/tr>\n<tbody>\n<tr>\n<td>\u9884\u8bad\u7ec3<\/td>\n<td>\u6d77\u91cf\u539f\u59cb\u4e92\u8054\u7f51\u6587\u672c<\/td>\n<td>\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd<\/td>\n<td>\u8bed\u8a00\u89c4\u5f8b\u3001\u4e16\u754c\u77e5\u8bc6\u3001\u57fa\u672c\u63a8\u7406<\/td>\n<td>\u6781\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>Mid-training<\/td>\n<td>\u76ee\u6807\u8bed\u8a00\/\u9886\u57df\u8bed\u6599<\/td>\n<td>\u7ee7\u7eed\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd<\/td>\n<td>\u8865\u9f50\u9886\u57df\u77e5\u8bc6\u3001\u57fa\u7840\u80fd\u529b<\/td>\n<td>\u4e2d\u5230\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>SFT<\/td>\n<td>\u51e0\u5343~\u51e0\u4e07\u6761\u793a\u8303\u5bf9<\/td>\n<td>\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd&#xff08;\u53ea\u5728\u56de\u7b54\u4e0a\u7b97\u635f\u5931&#xff09;<\/td>\n<td>\u6307\u4ee4\u9075\u5faa\u3001\u8f93\u51fa\u683c\u5f0f\u3001\u6d41\u7a0b\u534f\u8bae<\/td>\n<td>\u4f4e<\/td>\n<\/tr>\n<tr>\n<td>RL<\/td>\n<td>\u4efb\u52a1\u3001\u73af\u5883&#043;\u5956\u52b1\u4fe1\u53f7<\/td>\n<td>\u6700\u5927\u5316\u671f\u671b\u5956\u52b1<\/td>\n<td>\u53ef\u8fc1\u79fb\u7684\u51b3\u7b56\u7b56\u7565\u3001\u63a2\u7d22\u51fa\u7684\u65b0\u89e3\u6cd5<\/td>\n<td>\u9ad8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4e00\u4e2a\u76f4\u89c9\u7c7b\u6bd4&#xff1a;\u9884\u8bad\u7ec3\u662f&#034;\u901a\u8bc6\u6559\u80b2&#034;&#xff0c;Mid-training \u662f&#034;\u7814\u4fee\u4e13\u4e1a\u6559\u6750&#034;&#xff0c;SFT \u662f&#034;\u8001\u5e08\u793a\u8303\u7b54\u9898\u89c4\u8303&#034;&#xff0c;RL \u662f&#034;\u81ea\u5df1\u4e0b\u573a\u505a\u9898\u3001\u6309\u5f97\u5206\u53cd\u590d\u6253\u78e8&#034;\u3002<\/p>\n<h4>SFT \u7684\u672c\u8d28&#xff1a;\u6362\u4e86\u6570\u636e\u7684&#034;\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd&#034;<\/h4>\n<p>SFT \u5728\u6570\u5b66\u4e0a\u548c\u9884\u8bad\u7ec3\u662f\u540c\u4e00\u4e2a\u4efb\u52a1\u2014\u2014\u90fd\u662f\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u3001\u6700\u5c0f\u5316\u540c\u4e00\u4e2a\u635f\u5931\u51fd\u6570\u3002\u5dee\u522b\u53ea\u6709\u4e24\u70b9&#xff1a;<\/p>\n<li>\u6570\u636e\u4e0d\u540c&#xff1a;SFT \u7528\u4eba\u5de5\u51c6\u5907\u7684&#034;\u8f93\u5165\u2014\u8f93\u51fa&#034;\u793a\u8303\u5bf9<\/li>\n<li>\u635f\u5931\u5c4f\u853d&#xff08;loss masking&#xff09;&#xff1a;\u8ba1\u7b97\u635f\u5931\u65f6\u628a\u95ee\u9898\u90e8\u5206\u7684 token \u5c4f\u853d\u6389&#xff0c;\u53ea\u5bf9\u56de\u7b54\u90e8\u5206\u56de\u4f20\u68af\u5ea6<\/li>\n<p>SFT \u7684\u672c\u8d28\u662f\u7528\u6781\u9ad8\u7684\u6837\u672c\u6548\u7387&#xff0c;\u628a\u4e00\u5957\u7a33\u5b9a\u7684&#034;\u8f93\u5165\u2192\u8f93\u51fa&#034;\u6620\u5c04\u4e0e\u534f\u8bae\u56fa\u5316\u8fdb\u53c2\u6570\u3002\u5b83\u56fa\u5316\u7684\u662f&#034;\u683c\u5f0f\u3001\u98ce\u683c\u3001\u6d41\u7a0b&#034;\u8fd9\u7c7b\u534f\u8bae\u6027\u77e5\u8bc6&#xff0c;\u800c\u975e\u5927\u91cf\u4e8b\u5b9e\u6027\u77e5\u8bc6\u3002<\/p>\n<h4>SFT \u4e0e RL \u7684\u672c\u8d28\u533a\u522b<\/h4>\n<p>\u4e24\u6761\u8d2f\u7a7f\u5168\u7ae0\u7684\u4e3b\u7ebf\u4e4b\u4e00\u662f&#034;SFT \u8bb0\u5fc6&#xff0c;RL \u6cdb\u5316&#034;&#xff1a;<\/p>\n<ul>\n<li>SFT \u6700\u5927\u5316\u6807\u6ce8\u56de\u7b54\u7684\u6982\u7387\u2014\u2014\u63a8\u52a8\u6a21\u578b\u590d\u73b0\u793a\u8303\u3002\u793a\u8303\u7f3a\u4e4f\u591a\u6837\u6027\u65f6\u5bb9\u6613\u5bf9\u8868\u9762\u6a21\u5f0f\u8fc7\u62df\u5408\u3002<\/li>\n<li>RL \u6700\u5927\u5316\u671f\u671b\u5956\u52b1\u2014\u2014\u6a21\u578b\u63a2\u7d22\u591a\u6761\u8def\u5f84&#xff0c;\u63d0\u9ad8\u9ad8\u5956\u52b1\u8def\u5f84\u6982\u7387\u3002\u5f53\u5956\u52b1\u5fe0\u5b9e\u65f6\u53ef\u80fd\u53d1\u73b0\u793a\u8303\u4e2d\u6ca1\u6709\u7684\u53ef\u8fc1\u79fb\u7b56\u7565\u3002<\/li>\n<\/ul>\n<p>\u4ece\u6982\u7387\u5206\u5e03\u770b&#xff0c;SFT \u5e38\u8868\u73b0\u51fa\u8986\u76d6\u5f0f\u503e\u5411&#xff08;\u5c3d\u91cf\u8986\u76d6\u8bad\u7ec3\u6570\u636e\u4e2d\u7684\u591a\u4e2a\u6a21\u5f0f&#xff09;&#xff0c;RL \u914d\u5408\u53cd\u5411 KL \u7ea6\u675f\u66f4\u5bb9\u6613\u8868\u73b0\u51fa\u5bfb\u5cf0\u503e\u5411&#xff08;\u96c6\u4e2d\u5230\u5c11\u6570\u9ad8\u5956\u52b1\u5cf0\u4e0a&#xff09;\u3002<\/p>\n<table>\n<tr>\u7ef4\u5ea6SFTRL<\/tr>\n<tbody>\n<tr>\n<td>\u4f18\u5316\u76ee\u6807<\/td>\n<td>\u6700\u5927\u5316\u6807\u6ce8\u7b54\u6848\u6982\u7387<\/td>\n<td>\u6700\u5927\u5316\u671f\u671b\u5956\u52b1<\/td>\n<\/tr>\n<tr>\n<td>\u6837\u672c\u6548\u7387<\/td>\n<td>\u9ad8&#xff08;\u51e0\u5343\u6761\u89c1\u6548&#xff09;<\/td>\n<td>\u4f4e&#xff08;\u5e38\u662f SFT \u7684\u51e0\u5341~\u4e0a\u767e\u500d&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u8bad\u7ec3\u7a33\u5b9a\u6027<\/td>\n<td>\u9ad8\u3001\u6536\u655b\u5feb<\/td>\n<td>\u4f4e\u3001\u6613\u9707\u8361<\/td>\n<\/tr>\n<tr>\n<td>\u5206\u5e03\u6f02\u79fb\u4e0b<\/td>\n<td>\u53d6\u51b3\u4e8e\u793a\u8303\u8986\u76d6<\/td>\n<td>\u53d6\u51b3\u4e8e\u5956\u52b1\u4e0e\u63a2\u7d22<\/td>\n<\/tr>\n<tr>\n<td>\u6700\u9002\u5408<\/td>\n<td>\u56fa\u5316\u683c\u5f0f\/\u98ce\u683c\/\u6d41\u7a0b<\/td>\n<td>\u9700\u6cdb\u5316\u5230\u65b0\u573a\u666f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>Mid-training&#xff1a;\u8865\u77e5\u8bc6\u4e0e\u57fa\u7840\u80fd\u529b<\/h3>\n<p>Mid-training \u89e3\u51b3\u4e24\u7c7b\u7f3a\u53e3&#xff1a;\u77e5\u8bc6\u7f3a\u53e3&#xff08;\u901a\u7528\u9884\u8bad\u7ec3\u672a\u8986\u76d6\u76ee\u6807\u8bed\u8a00\u3001\u9886\u57df\u6587\u6863&#xff09;\u548c\u57fa\u7840\u80fd\u529b\u7f3a\u53e3&#xff08;\u957f\u4e0a\u4e0b\u6587\u3001\u4ee3\u7801\u6a21\u5f0f\u3001\u8de8\u6a21\u6001\u8868\u5f81&#xff09;\u3002<\/p>\n<p>\u4e3a\u4ec0\u4e48\u4e0d\u5e94\u628a SFT \u5f53\u6210\u4e3b\u8981\u77e5\u8bc6\u6ce8\u5165\u5de5\u5177&#xff1f;\u5c11\u91cf QA \u5bf9\u53ea\u8986\u76d6\u6709\u9650\u95ee\u6cd5&#xff0c;\u66f4\u64c5\u957f\u8bad\u7ec3&#034;\u5982\u4f55\u8bbf\u95ee\u4e0e\u8868\u8fbe&#034;&#xff0c;\u4e0d\u9002\u5408\u627f\u8f7d\u5927\u89c4\u6a21\u3001\u76f8\u4e92\u5173\u8054\u7684\u539f\u59cb\u77e5\u8bc6\u3002Mid-training \u964d\u4f4e\u9886\u57df\u6587\u672c\u7684\u8bed\u8a00\u6a21\u578b\u635f\u5931&#xff0c;\u4e5f\u4e0d\u4fdd\u8bc1\u6a21\u578b\u4f1a\u81ea\u52a8\u6309\u7528\u6237\u95ee\u9898\u53d6\u51fa\u77e5\u8bc6\u2014\u2014\u7a33\u5065\u914d\u65b9\u901a\u5e38\u662f Mid-training \u5438\u6536\u77e5\u8bc6 \u2192 \u5c0f\u89c4\u6a21 SFT \u5efa\u7acb\u8f93\u51fa\u683c\u5f0f \u2192 \u6709\u975e\u96f6\u6210\u529f\u7387\u540e\u518d RL\u3002<\/p>\n<p>\u6570\u636e\u6784\u9020\u7684\u5173\u952e\u6b65\u9aa4\u5305\u62ec&#xff1a;\u4ece\u5931\u8d25\u5206\u5e03\u53cd\u63a8\u6570\u636e\u7f3a\u53e3\u3001\u6784\u9020\u9ad8\u5bc6\u5ea6\u76ee\u6807\u8bed\u6599\u3001\u6309\u80fd\u529b\u8fdb\u884c\u6570\u636e\u914d\u6bd4\u3001\u6bcf\u4e2a\u9636\u6bb5\u505a&#034;\u53cc\u91cd\u56de\u653e&#034;&#xff08;\u901a\u7528\u4fdd\u7559\u6570\u636e&#043;\u957f\u5ea6\u63d0\u5347\u540e\u7684\u65e7\u4efb\u52a1&#xff09;\u3001\u7528\u591a\u7ef4\u95e8\u7981\u51b3\u5b9a\u4f55\u65f6\u505c\u6b62\u3002<\/p>\n<h3>SFT \u7684\u5de5\u7a0b\u5b9e\u8df5<\/h3>\n<h4>LoRA \u53c2\u6570\u9ad8\u6548\u5fae\u8c03<\/h4>\n<p>\u5168\u53c2\u6570\u5fae\u8c03\u5bf9\u663e\u5b58\u8981\u6c42\u5f88\u9ad8&#xff0c;LoRA \u5728\u539f\u59cb\u6743\u91cd\u65c1\u6302\u4e00\u4e2a\u4f4e\u79e9&#034;\u8865\u4e01&#034;\u5b66\u4e60\u4efb\u52a1&#xff0c;\u53c2\u6570\u91cf\u4ec5\u5360 1%-5%\u3002\u5173\u952e\u5b9e\u8df5&#xff1a;<\/p>\n<ul>\n<li>\u5fc5\u987b\u5e94\u7528\u5230\u6240\u6709\u4e3b\u8981\u6743\u91cd\u77e9\u9635&#xff08;\u5c24\u5176 MLP \u5c42&#xff09;&#xff0c;\u53ea\u52a0\u5728\u6ce8\u610f\u529b\u5c42\u4f1a\u6389\u70b9<\/li>\n<li>\u6700\u4f18\u5b66\u4e60\u7387\u7ea6\u662f\u5168\u53c2\u5fae\u8c03\u7684 10 \u500d<\/li>\n<li>SFT \u7528\u4e2d\u9ad8 rank&#xff08;64-256&#xff09;&#xff0c;RL \u7528\u5c0f rank&#xff08;8-32&#xff09;<\/li>\n<\/ul>\n<h4>SFT \u6570\u636e\u4ece\u54ea\u6765<\/h4>\n<p>\u4e09\u6761\u8def\u5f84\u7ecf\u5e38\u7ec4\u5408\u4f7f\u7528&#xff1a;<\/p>\n<li>\u4eba\u5de5\u4e13\u5bb6\u793a\u8303\u2014\u2014\u8d28\u91cf\u6700\u9ad8\u4f46\u8d35\u4e14\u6162&#xff0c;\u9002\u5408\u5b9a\u4e49\u683c\u5f0f\u7684&#034;\u79cd\u5b50\u6570\u636e&#034;<\/li>\n<li>\u6559\u5e08\u6a21\u578b\u751f\u6210&#xff08;\u5408\u6210\u6570\u636e&#xff09;\u2014\u2014\u8ba9\u5f3a\u6a21\u578b\u6279\u91cf\u4ea7\u51fa&#xff0c;\u8fc7\u6ee4\u540e\u84b8\u998f\u7ed9\u5b66\u751f<\/li>\n<li>\u62d2\u7edd\u91c7\u6837\u2014\u2014\u6a21\u578b\u81ea\u5df1\u91c7\u6837\u591a\u6761\u5019\u9009&#xff0c;\u7528\u9a8c\u8bc1\u5668\u7b5b\u51fa\u6b63\u786e\u6837\u672c\u518d\u53cd\u8fc7\u6765\u8bad\u7ec3\u81ea\u5df1<\/li>\n<h4>\u84b8\u998f&#xff1a;Prompt \u84b8\u998f\u4e0e CoT \u84b8\u998f<\/h4>\n<p>Prompt \u84b8\u998f\u628a&#034;\u957f\u63d0\u793a &#043; \u601d\u8003\u578b\u6559\u5e08&#034;\u7684\u884c\u4e3a\u538b\u7f29\u5230&#034;\u77ed\u63d0\u793a\/\u65e0\u63d0\u793a &#043; \u975e\u601d\u8003\u5b66\u751f&#034;\u4e2d&#xff0c;\u53ef\u83b7\u5f97 20-30 \u500d\u7684\u54cd\u5e94\u901f\u5ea6\u63d0\u5347\u3002<\/p>\n<p>CoT \u84b8\u998f\u628a\u6559\u5e08\u6a21\u578b\u7684\u5b8c\u6574\u601d\u8003\u8f68\u8ff9\u8f6c\u79fb\u7ed9\u5b66\u751f\u6a21\u578b\u3002\u5728\u540c\u7b49\u53c2\u6570\u91cf\u4e0b\u53ef\u6062\u590d\u6559\u5e08 70%-80% \u7684\u80fd\u529b\u3002DeepSeek-R1 \u5f00\u6e90\u7684\u84b8\u998f\u5c0f\u6a21\u578b\u6b63\u662f\u8fd9\u6761\u8def\u7ebf\u7684\u4ee3\u8868\u3002<\/p>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f&#xff1a;\u5bf9\u7edd\u5927\u591a\u6570\u505a\u540e\u8bad\u7ec3\u7684\u4eba\u6765\u8bf4&#xff0c;\u6839\u672c\u4e0d\u9700\u8981\u53bb\u84b8\u998f\u95ed\u6e90\u6a21\u578b\u7684\u601d\u7ef4\u94fe\u3002\u5f53\u524d\u6700\u5148\u8fdb\u7684\u5f00\u6e90\u6a21\u578b\u4e0e SOTA \u95ed\u6e90\u6a21\u578b\u5dee\u8ddd\u5e76\u6ca1\u6709\u60f3\u8c61\u4e2d\u5927&#xff1b;\u6559\u5e08\u6a21\u578b\u53ea\u9700\u8981&#034;\u660e\u663e\u9ad8\u4e8e\u5b66\u751f&#034;\u5373\u53ef\u3002<\/p>\n<h3>RL \u7b97\u6cd5&#xff1a;\u4ece rollout \u5230\u53c2\u6570\u66f4\u65b0<\/h3>\n<h4>GRPO \u7684\u56db\u6b65\u95ed\u73af<\/h4>\n<p>GRPO&#xff08;Group Relative Policy Optimization&#xff09;\u662f\u5f53\u524d\u6700\u5e38\u7528\u7684 RL \u7b97\u6cd5\u4e4b\u4e00\u3002\u4ee5 SWE-bench \u7684\u4e00\u6761\u4fee\u590d\u4efb\u52a1\u4e3a\u4f8b&#xff1a;<\/p>\n<li>\u8ba9\u7b56\u7565\u6a21\u578b\u91cd\u590d\u5c1d\u8bd5&#xff1a;\u540c\u4e00\u4efd\u4ee3\u7801\u548c\u95ee\u9898\u63cf\u8ff0\u590d\u5236\u5230 16 \u4e2a\u9694\u79bb\u6c99\u7bb1&#xff0c;\u8ba9\u6a21\u578b\u72ec\u7acb\u89e3\u51b3 16 \u6b21&#xff08;\u6bcf\u6b21\u53eb\u4e00\u4e2a rollout&#xff09;<\/li>\n<li>\u8ba1\u7b97\u5956\u52b1&#xff1a;\u9a8c\u8bc1\u5668\u5728\u5e72\u51c0\u73af\u5883\u4e2d\u5e94\u7528\u8865\u4e01\u5e76\u8fd0\u884c\u6d4b\u8bd5&#xff0c;\u901a\u8fc7\u5f97 1\u3001\u5931\u8d25\u5f97 0<\/li>\n<li>\u8ba1\u7b97\u76f8\u5bf9\u4f18\u52bf&#xff1a;\u7ec4\u5185\u5e73\u5747\u6210\u529f\u7387 4\/16&#xff0c;\u901a\u8fc7\u7684 4 \u6761\u5f97\u6b63\u4f18\u52bf\u3001\u5931\u8d25\u7684 12 \u6761\u5f97\u8d1f\u4f18\u52bf<\/li>\n<li>\u68af\u5ea6\u4e0b\u964d\u66f4\u65b0\u7b56\u7565&#xff1a;\u63d0\u9ad8\u6b63\u4f18\u52bf\u8f68\u8ff9\u4e2d\u9009\u62e9\u7684\u6982\u7387&#xff0c;\u964d\u4f4e\u8d1f\u4f18\u52bf\u7684<\/li>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161656-6a96fa784dca8.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>PPO \u4e0e GRPO \u7684\u533a\u522b\u4e3b\u8981\u5728&#034;\u62ff\u8c01\u6765\u6bd4\u8f83&#034;&#xff1a;GRPO \u76f4\u63a5\u6bd4\u8f83\u540c\u4e00\u95ee\u9898\u7684\u591a\u6761 rollout&#xff0c;\u4e0d\u9700\u8981\u989d\u5916\u4ef7\u503c\u6a21\u578b&#xff1b;PPO \u8bad\u7ec3\u4ef7\u503c\u6a21\u578b\u4f30\u8ba1\u6bcf\u6b65&#034;\u901a\u5e38\u80fd\u505a\u5230\u591a\u597d&#034;&#xff0c;\u66f4\u9002\u5408\u9700\u8981\u7ec6\u7c92\u5ea6\u4fe1\u7528\u5206\u914d\u7684\u957f\u8f68\u8ff9\u3002<\/p>\n<h4>On-Policy \u4e0e\u6570\u503c\u8bef\u5dee<\/h4>\n<p>LLM RL \u901a\u5e38\u4f18\u5148 on-policy\u2014\u2014\u8ba9\u5f53\u524d\u7b56\u7565\u81ea\u5df1\u751f\u6210 rollout&#xff0c;\u8bad\u7ec3\u96c6\u4e2d\u5728&#034;\u5f53\u524d\u6a21\u578b\u5b9e\u9645\u4f1a\u8fdb\u5165\u7684\u72b6\u6001&#034;\u3002\u4f46\u5927\u89c4\u6a21\u8bad\u7ec3\u4e2d&#xff0c;\u63a8\u7406\u5f15\u64ce&#xff08;vLLM\/SGLang&#xff09;\u548c\u8bad\u7ec3\u5f15\u64ce&#xff08;FSDP\/Megatron&#xff09;\u5373\u4f7f\u52a0\u8f7d\u540c\u4e00\u6743\u91cd&#xff0c;\u6d6e\u70b9\u7cbe\u5ea6\u3001\u5f52\u7ea6\u987a\u5e8f\u3001\u5f20\u91cf\u5e76\u884c\u65b9\u5f0f\u7684\u5dee\u5f02\u4e5f\u53ef\u80fd\u8ba9\u540c\u4e00 token \u7684 log probability \u4e0d\u540c&#xff0c;\u4f7f\u540d\u4e49\u4e0a\u7684 on-policy \u53d8\u6210\u9690\u5f0f off-policy\u3002<\/p>\n<p>\u654f\u611f\u6027\u6765\u81ea\u653e\u5927\u94fe\u6761&#xff1a;log probability \u5c0f\u8bef\u5dee \u2192 \u6307\u6570\u5316\u7684\u6982\u7387\u6bd4\u504f\u5dee \u2192 \u957f\u524d\u7f00\u4e0a\u7684\u7d2f\u79ef \u2192 clipping \u6539\u53d8 \u2192 \u68af\u5ea6\u65b9\u5411\u6539\u53d8\u3002\u82e5 4000 \u4e2a token \u7684 log ratio \u90fd\u6709 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">10\u2212310^{-3}<\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8141em\"><\/span><span class=\"mord\">1<\/span><span class=\"mord\"><span class=\"mord\">0<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8141em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mtight\">\u2212<\/span><span class=\"mord mtight\">3<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u540c\u5411\u504f\u5dee&#xff0c;\u8f68\u8ff9\u7ea7\u6bd4\u7387\u4f1a\u7d2f\u79ef\u5230 <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">e4\u224854.6e^4 \\\\approx 54.6<\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8141em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">e<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8141em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\">4<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2248<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.6444em\"><\/span><span class=\"mord\">54.6<\/span><\/span><\/span><\/span><\/span>\u3002<\/p>\n<h3>RL \u73af\u5883&#xff1a;\u4ece\u8bc4\u4f30\u5230\u4eff\u771f<\/h3>\n<h4>\u73af\u5883\u5de5\u7a0b\u662f RL \u7684\u524d\u7f6e\u6761\u4ef6<\/h4>\n<p>RL \u8bad\u7ec3\u7684\u74f6\u9888\u5f80\u5f80\u4e0d\u5728\u7b97\u6cd5&#xff0c;\u800c\u5728\u73af\u5883\u662f\u5426\u8db3\u591f\u771f\u5b9e\u3001\u53ef\u91cd\u7f6e\u3001\u53ef\u5e76\u884c\u3002\u771f\u5b9e API \u6709\u901f\u7387\u9650\u5236\u3001\u4f1a\u5c01\u53f7\u3001\u6709\u526f\u4f5c\u7528&#xff0c;\u6839\u672c\u65e0\u6cd5\u76f4\u63a5\u7528\u4e8e\u8bad\u7ec3\u3002<\/p>\n<p>\u73af\u5883\u5de5\u7a0b\u7684\u987a\u5e8f\u5e94\u662f&#xff1a;\u4efb\u52a1\u84dd\u56fe \u2192 \u53ef\u91cd\u7f6e\u6a21\u62df\u5668 \u2192 \u786e\u5b9a\u6027\u9a8c\u8bc1\u5668 \u2192 \u8bad\u7ec3\/\u8bc4\u4f30\u9694\u79bb \u2192 \u5c11\u91cf\u771f\u5b9e\u4ea4\u4e92\u6821\u51c6\u3002<\/p>\n<h4>\u9020\u4e0d\u51fa\u73af\u5883\u600e\u4e48\u529e&#xff1a;\u8ba9\u6a21\u578b\u626e\u6f14\u73af\u5883<\/h4>\n<p>\u5f53\u9ad8\u4fdd\u771f\u73af\u5883\u6839\u672c\u9020\u4e0d\u51fa\u6765\u65f6&#xff0c;\u53ef\u4ee5\u7528 LLM \u626e\u6f14\u73af\u5883&#xff1a;\u6a21\u578b\u5408\u6210\u5de5\u5177\u8c03\u7528\u7684\u8fd4\u56de\u503c&#xff08;\u5982 ZeroSearch \u7528 LLM \u626e\u6f14\u641c\u7d22\u5f15\u64ce&#xff09;&#xff0c;\u751a\u81f3\u4eff\u771f\u6574\u4e2a\u73af\u5883\u7684\u52a8\u6001&#xff08;\u5982 DreamGym \u84b8\u998f\u73af\u5883\u8f6c\u79fb&#xff09;\u3002\u4f46\u6a21\u62df\u5668\u7684\u4e16\u754c\u77e5\u8bc6\u5c31\u662f\u8bad\u7ec3\u7684\u5929\u82b1\u677f&#xff0c;\u7cfb\u7edf\u6027\u504f\u5dee\u4f1a\u88ab\u7b56\u7565\u7167\u5355\u5168\u6536\u2014\u2014\u7a33\u59a5\u505a\u6cd5\u662f\u6df7\u5408\u4f7f\u7528\u6a21\u578b\u6a21\u62df\u548c\u771f\u5b9e\u73af\u5883\u4ea4\u4e92\u3002<\/p>\n<h4>\u591a\u8f6e\u4efb\u52a1\u7684\u4fe1\u7528\u5206\u914d<\/h4>\n<p>\u4ece\u5355\u8f6e\u5230\u591a\u8f6e&#xff0c;\u590d\u6742\u6027\u53d1\u751f\u8d28\u53d8&#xff1a;\u7b56\u7565\u4e0d\u4ec5\u8981\u9009\u62e9\u5f53\u524d\u6700\u4f18\u52a8\u4f5c&#xff0c;\u8fd8\u8981\u5728\u5ef6\u8fdf\u5956\u52b1\u4e0b\u8fdb\u884c\u4fe1\u7528\u5206\u914d\u2014\u2014\u5224\u65ad\u591a\u6b65\u5e8f\u5217\u4e2d\u5230\u5e95\u54ea\u4e00\u6b65\u5bf9\u6700\u7ec8\u7ed3\u679c\u8d21\u732e\u6700\u5927\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161656-6a96fa789c982.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161656-6a96fa78dd25d.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u5de5\u5177\u8c03\u7528 RL \u5f15\u5165\u4e86\u4e00\u4e2a\u5173\u952e\u5b9e\u73b0\u7ec6\u8282&#xff1a;\u73af\u5883\u8fd4\u56de\u7684 token \u4e0d\u662f\u7b56\u7565\u751f\u6210\u7684&#xff0c;\u8ba1\u7b97\u7b56\u7565\u68af\u5ea6\u65f6\u5e94\u5c4f\u853d\u8fd9\u4e9b\u53cd\u9988 token&#xff0c;\u53ea\u5bf9\u6a21\u578b\u81ea\u5df1\u7684\u601d\u8003\u548c\u5de5\u5177\u8c03\u7528\u53c2\u6570\u56de\u4f20\u68af\u5ea6\u3002\u5426\u5219\u6a21\u578b\u4f1a\u88ab\u8bad\u7ec3\u53bb\u9884\u6d4b\u6c99\u76d2\u8f93\u51fa&#xff0c;\u800c\u4e0d\u662f\u5b66\u4f1a\u5982\u4f55\u4f7f\u7528\u5de5\u5177\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161657-6a96fa79217a6.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h3>\u4f55\u65f6\u9009\u62e9 Mid-training\u3001SFT \u4e0e RL<\/h3>\n<p>\u4e66\u4e2d\u7ed9\u51fa\u4e86\u5b9e\u64cd\u8bca\u65ad\u6846\u67b6&#xff1a;<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161657-6a96fa7973f77.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<table>\n<tr>\u89c2\u5bdf\u5230\u7684\u73b0\u8c61\u4e3b\u8981\u7f3a\u53e3\u4f18\u5148\u65b9\u6cd5<\/tr>\n<tbody>\n<tr>\n<td>pass&#064;k \u4ecd\u63a5\u8fd1 0<\/td>\n<td>\u77e5\u8bc6\u4e0e\u80fd\u529b\u4e0d\u5728\u57fa\u6a21\u4e2d<\/td>\n<td>Mid-training<\/td>\n<\/tr>\n<tr>\n<td>\u5076\u5c14\u505a\u5bf9\u4f46\u683c\u5f0f\u4e0d\u7a33\u5b9a<\/td>\n<td>\u884c\u4e3a\u534f\u8bae\u6ca1\u6709\u56fa\u5316<\/td>\n<td>SFT<\/td>\n<\/tr>\n<tr>\n<td>\u6709\u975e\u96f6\u6210\u529f\u7387\u4f46\u597d\u7b56\u7565\u6982\u7387\u4f4e<\/td>\n<td>\u6982\u7387\u5206\u914d\u4e0e\u7b56\u7565\u4f18\u5316<\/td>\n<td>RL<\/td>\n<\/tr>\n<tr>\n<td>\u53ea\u6709\u5c11\u91cf\u793a\u8303&#xff0c;\u65e0\u5728\u7ebf\u73af\u5883<\/td>\n<td>\u53ef\u6a21\u4eff\u6570\u636e\u6709\u3001\u5728\u7ebf\u53cd\u9988\u65e0<\/td>\n<td>SFT\/RFT<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6838\u5fc3\u5224\u65ad\u903b\u8f91\u662f&#xff1a;\u5148\u6392\u9664\u4e0d\u9700\u8981\u6539\u6743\u91cd\u7684\u65b9\u6848&#xff08;prompt\u3001\u5de5\u5177\u3001\u7ea6\u675f\u80fd\u89e3\u51b3\u5c31\u4e0d\u8bad\u7ec3&#xff09;&#xff1b;\u518d\u6d4b\u80fd\u529b\u652f\u6301&#xff08;pass&#064;k \u8fd1\u96f6\u5148\u8865 Mid-training&#xff09;&#xff1b;\u7528 SFT \u7acb\u534f\u8bae\u4e0d\u585e\u77e5\u8bc6&#xff1b;\u53ea\u5728\u6709\u63a2\u7d22\u7a7a\u95f4\u65f6\u4e0a RL\u3002<\/p>\n<h3>\u7ecf\u5178 RL \u4e0e LLM Agent \u7684\u6839\u672c\u5dee\u5f02<\/h3>\n<p>\u4e66\u4e2d\u7528\u540c\u4e00\u4e2a\u5bfb\u5b9d\u6e38\u620f\u5bf9\u6bd4\u4e86 Q-learning \u4e0e LLM Agent&#xff1a;<\/p>\n<ul>\n<li>Q-learning \u9700\u8981\u8fd1 10000 \u5c40\u8bd5\u9519&#xff08;10 \u79d2&#xff09;&#xff0c;\u4f46\u7406\u89e3&#034;\u95e8&#034;\u201c\u94a5\u5319\u201d&#034;\u5251&#034;\u53ea\u662f\u65e0\u610f\u4e49\u7b26\u53f7<\/li>\n<li>LLM Agent \u7b2c\u4e00\u5c40\u5c31\u5728 18 \u6b65\u5185\u901a\u5173&#xff0c;\u6bcf\u4e00\u6b65\u90fd\u6709\u76ee\u7684\u548c\u903b\u8f91\u652f\u6491<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161657-6a96fa79aeb7f.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u59da\u987a\u96e8\u5728\u300aThe Second Half\u300b\u4e2d\u603b\u7ed3\u7684 OpenAI \u8ba4\u77e5\u6f14\u53d8\u503c\u5f97\u6df1\u601d&#xff1a;\u4e0d\u662f\u7b97\u6cd5 &gt; \u73af\u5883 &gt; \u5148\u9a8c&#xff0c;\u800c\u662f\u5148\u9a8c &gt; \u73af\u5883 &gt; \u7b97\u6cd5\u3002\u9884\u8bad\u7ec3\u8d4b\u4e88\u7684\u5148\u9a8c\u77e5\u8bc6\u4f7f LLM Agent \u80fd\u591f\u7406\u89e3\u672a\u89c1\u8fc7\u7684\u6307\u4ee4&#xff08;\u96f6\u6837\u672c\u6cdb\u5316&#xff09;&#xff0c;\u7528\u5c11\u91cf\u793a\u8303\u9002\u5e94\u65b0\u4efb\u52a1\u2014\u2014\u8fd9\u4e0e\u65e0\u5148\u9a8c\u77e5\u8bc6\u3001\u9700\u5927\u91cf\u8bd5\u9519\u7684\u7ecf\u5178 RL \u5f62\u6210\u9c9c\u660e\u5bf9\u6bd4\u3002<\/p>\n<h3>\u5c0f\u7ed3<\/h3>\n<p>\u672c\u7ae0\u7684\u4e24\u6761\u4e3b\u7ebf\u8d2f\u7a7f\u59cb\u7ec8&#xff1a;\u4e3b\u7ebf\u4e00\u662f&#034;SFT \u8bb0\u5fc6&#xff0c;RL \u6cdb\u5316&#034;\u2014\u2014\u5728\u53d7\u63a7\u5b9e\u9a8c\u6761\u4ef6\u4e0b&#xff0c;SFT \u5bf9\u8bad\u7ec3\u6570\u636e\u8fc7\u62df\u5408&#xff0c;RL \u5728\u5206\u5e03\u53d8\u5316\u65f6\u66f4\u5bb9\u6613\u5b66\u5230\u53ef\u8fc1\u79fb\u7b56\u7565&#xff1b;\u4f46\u8fd9\u662f\u7279\u5b9a\u5b9e\u9a8c\u6761\u4ef6\u4e0b\u7684\u7ed3\u679c&#xff0c;\u4e0d\u662f SFT \u4e0e RL \u7684\u666e\u904d\u5c5e\u6027\u3002\u4e3b\u7ebf\u4e8c\u662f&#034;\u6570\u636e\u548c\u73af\u5883&#xff0c;\u6bd4\u7b97\u6cd5\u66f4\u91cd\u8981&#034;\u2014\u2014\u638c\u63e1\u73b0\u6210 RL \u7b97\u6cd5&#xff08;PPO\u3001GRPO&#xff09;\u7684\u57fa\u672c\u7528\u6cd5\u5373\u53ef&#xff0c;\u771f\u6b63\u51b3\u5b9a\u6210\u8d25\u7684\u662f Mid-training \u8bed\u6599\u6709\u6ca1\u6709\u8865\u9f50\u5e95\u5ea7\u3001\u793a\u8303\u6570\u636e\u6709\u6ca1\u6709\u5efa\u7acb\u884c\u4e3a\u534f\u8bae\u3001\u4eff\u771f\u73af\u5883\u4e0e\u5956\u52b1\u80fd\u4e0d\u80fd\u63d0\u4f9b\u53ef\u9760\u53cd\u9988\u3002\u5f88\u591a\u573a\u666f\u4e0b&#xff0c;\u53ea\u8981\u524d\u4e24\u7c7b\u6570\u636e\u8d28\u91cf\u5230\u4f4d&#xff0c;\u751a\u81f3\u6839\u672c\u4e0d\u9700\u8981\u505a RL\u3002<\/p>\n<p>\u672c\u6587\u5185\u5bb9\u6574\u7406\u81ea\u5f00\u6e90\u6280\u672f\u4e66\u300a\u6df1\u5165\u7406\u89e3 AI Agent\u300b(bojieli\/ai-agent-book)&#xff0c;\u91c7\u7528 Apache 2.0 \u8bb8\u53ef\u8bc1<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6a21\u578b\u540e\u8bad\u7ec3&#xff1a;\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5<br \/>\nAgent  LLM  \u4e0a\u4e0b\u6587  \u5de5\u5177\u3002\u5f53\u4e0a\u4e0b\u6587\u5de5\u7a0b\u548c\u5de5\u5177\u8bbe\u8ba1\u5df2\u7ecf\u505a\u5230\u4f4d&#xff0c;\u5269\u4e0b\u7684\u74f6\u9888\u5f80\u5f80\u5728 LLM \u672c\u8eab\u2014\u2014\u8fd9\u65f6\u9700\u8981\u901a\u8fc7\u540e\u8bad\u7ec3\u6539\u52a8\u6a21\u578b\u6743\u91cd&#xff0c;\u628a\u80fd\u529b\u6c89\u6dc0\u8fdb\u53c2\u6570\u3002\u672c\u7ae0\u7cfb\u7edf\u8ba8\u8bba\u4e86\u4ece\u9884\u8bad\u7ec3\u5230 RL \u7684\u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u6d41\u6c34\u7ebf&#xff0c;\u4ee5\u53ca SFT \u4e0e RL \u5728\u5de5\u7a0b\u5b9e\u8df5\u4e2d\u7684\u672c\u8d28\u533a\u522b\u3001\u6570\u636e\u5408\u6210\u65b9\u6cd5\u548c\u73af\u5883\u5de5\u7a0b\u8981\u70b9\u3002<br \/>\n\u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u5168\u666f<br \/>\n\u73b0\u4ee3\u6a21\u578b\u7684\u80fd\u529b\u5f00\u53d1\u901a\u5e38\u62c6\u6210\u56db\u4e2a\u73af\u8282&#xff1a;\u9636\u6bb5\u7528\u4ec0\u4e48\u6570\u636e\u4f18<\/p>\n","protected":false},"author":2,"featured_media":99144,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[1699,52,50,188,51],"topic":[],"class_list":["post-99151","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-agi","tag-aigc","tag-50","tag-188","tag-51"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/99151.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6a21\u578b\u540e\u8bad\u7ec3&#xff1a;\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 Agent LLM \u4e0a\u4e0b\u6587 \u5de5\u5177\u3002\u5f53\u4e0a\u4e0b\u6587\u5de5\u7a0b\u548c\u5de5\u5177\u8bbe\u8ba1\u5df2\u7ecf\u505a\u5230\u4f4d&#xff0c;\u5269\u4e0b\u7684\u74f6\u9888\u5f80\u5f80\u5728 LLM \u672c\u8eab\u2014\u2014\u8fd9\u65f6\u9700\u8981\u901a\u8fc7\u540e\u8bad\u7ec3\u6539\u52a8\u6a21\u578b\u6743\u91cd&#xff0c;\u628a\u80fd\u529b\u6c89\u6dc0\u8fdb\u53c2\u6570\u3002\u672c\u7ae0\u7cfb\u7edf\u8ba8\u8bba\u4e86\u4ece\u9884\u8bad\u7ec3\u5230 RL \u7684\u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u6d41\u6c34\u7ebf&#xff0c;\u4ee5\u53ca SFT \u4e0e RL \u5728\u5de5\u7a0b\u5b9e\u8df5\u4e2d\u7684\u672c\u8d28\u533a\u522b\u3001\u6570\u636e\u5408\u6210\u65b9\u6cd5\u548c\u73af\u5883\u5de5\u7a0b\u8981\u70b9\u3002 \u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u5168\u666f \u73b0\u4ee3\u6a21\u578b\u7684\u80fd\u529b\u5f00\u53d1\u901a\u5e38\u62c6\u6210\u56db\u4e2a\u73af\u8282&#xff1a;\u9636\u6bb5\u7528\u4ec0\u4e48\u6570\u636e\u4f18\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/99151.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-01T16:16:59+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161655-6a96fa7748e8d.jpg\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"2 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/99151.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/99151.html\",\"name\":\"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-01T16:16:59+00:00\",\"dateModified\":\"2026-09-01T16:16:59+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/99151.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/99151.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/99151.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/99151.html","og_locale":"zh_CN","og_type":"article","og_title":"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6a21\u578b\u540e\u8bad\u7ec3&#xff1a;\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 Agent LLM \u4e0a\u4e0b\u6587 \u5de5\u5177\u3002\u5f53\u4e0a\u4e0b\u6587\u5de5\u7a0b\u548c\u5de5\u5177\u8bbe\u8ba1\u5df2\u7ecf\u505a\u5230\u4f4d&#xff0c;\u5269\u4e0b\u7684\u74f6\u9888\u5f80\u5f80\u5728 LLM \u672c\u8eab\u2014\u2014\u8fd9\u65f6\u9700\u8981\u901a\u8fc7\u540e\u8bad\u7ec3\u6539\u52a8\u6a21\u578b\u6743\u91cd&#xff0c;\u628a\u80fd\u529b\u6c89\u6dc0\u8fdb\u53c2\u6570\u3002\u672c\u7ae0\u7cfb\u7edf\u8ba8\u8bba\u4e86\u4ece\u9884\u8bad\u7ec3\u5230 RL \u7684\u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u6d41\u6c34\u7ebf&#xff0c;\u4ee5\u53ca SFT \u4e0e RL \u5728\u5de5\u7a0b\u5b9e\u8df5\u4e2d\u7684\u672c\u8d28\u533a\u522b\u3001\u6570\u636e\u5408\u6210\u65b9\u6cd5\u548c\u73af\u5883\u5de5\u7a0b\u8981\u70b9\u3002 \u56db\u9636\u6bb5\u80fd\u529b\u5f00\u53d1\u5168\u666f \u73b0\u4ee3\u6a21\u578b\u7684\u80fd\u529b\u5f00\u53d1\u901a\u5e38\u62c6\u6210\u56db\u4e2a\u73af\u8282&#xff1a;\u9636\u6bb5\u7528\u4ec0\u4e48\u6570\u636e\u4f18","og_url":"https:\/\/www.wsisp.com\/helps\/99151.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-01T16:16:59+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260901161655-6a96fa7748e8d.jpg"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"2 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/99151.html","url":"https:\/\/www.wsisp.com\/helps\/99151.html","name":"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-01T16:16:59+00:00","dateModified":"2026-09-01T16:16:59+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/99151.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/99151.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/99151.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u4ece SFT \u5230\u5f3a\u5316\u5b66\u4e60\u7684\u5de5\u7a0b\u5b9e\u8df5"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/99151","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=99151"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/99151\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/99144"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=99151"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=99151"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=99151"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=99151"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}