{"id":777,"date":"2025-09-30T14:17:14","date_gmt":"2025-09-30T06:17:14","guid":{"rendered":"https:\/\/ferina.top\/?p=777"},"modified":"2025-09-30T14:17:16","modified_gmt":"2025-09-30T06:17:16","slug":"rlhf-and-ppo","status":"publish","type":"post","link":"https:\/\/ferina.top\/index.php\/2025\/09\/30\/rlhf-and-ppo\/","title":{"rendered":"RLHF and PPO"},"content":{"rendered":"\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u7ed9\u5b9a\u540c\u4e00\u4e2a\u8f93\u5165\uff0c\u6a21\u578b\u751f\u6210\u7684\u591a\u4e2a\u5019\u9009\u56de\u7b54\u4e2d\uff0c\u4eba\u7c7b\u4f1a\u5bf9\u8fd9\u4e9b\u56de\u7b54\u8fdb\u884c\u6392\u5e8f\u3002RM \u9700\u8981\u5b66\u4f1a\u9884\u6d4b\u8fd9\u4e2a\u6392\u5e8f\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u6570\u636e\u683c\u5f0f<\/h2>\n\n\n\n<p>\u5047\u8bbe\u5bf9\u4e8e\u540c\u4e00\u4e2a prompt\uff0c\u6211\u4eec\u6709\u4e24\u6bb5\u5019\u9009\u56de\u7b54\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>chosen (c)<\/strong> \u2192 \u4eba\u7c7b\u66f4\u559c\u6b22\u7684\u56de\u7b54<\/li>\n\n\n\n<li><strong>rejected (r)<\/strong> \u2192 \u4eba\u7c7b\u4e0d\u559c\u6b22\u7684\u56de\u7b54<\/li>\n<\/ul>\n\n\n\n<p>Reward Model \u4f1a\u7ed9\u5b83\u4eec\u5404\u81ea\u8f93\u51fa\u4e00\u4e2a\u5206\u6570\uff1a<\/p>\n\n\n\n<p>[<br>R_\\theta(x, y_c), \\quad R_\\theta(x, y_r)<br>]<\/p>\n\n\n\n<p>\u5176\u4e2d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>(x) \u662f\u8f93\u5165\uff08prompt\uff09\uff0c<\/li>\n\n\n\n<li>(y_c, y_r) \u662f\u56de\u7b54\uff0c<\/li>\n\n\n\n<li>(R_\\theta) \u662f RM \u7684\u6253\u5206\u51fd\u6570\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u76ee\u6807<\/h2>\n\n\n\n<p>\u6211\u4eec\u5e0c\u671b RM \u7684\u5206\u6570\u6ee1\u8db3\uff1a<\/p>\n\n\n\n<p>[<br>R_\\theta(x, y_c) &gt; R_\\theta(x, y_r)<br>]<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u635f\u5931\u51fd\u6570\uff08Reward Model loss\uff09<\/h2>\n\n\n\n<p>\u5e38\u7528\u7684\u662f <strong>pairwise ranking loss<\/strong>\uff08\u6210\u5bf9\u6392\u5e8f\u635f\u5931\uff09\uff0c\u516c\u5f0f\u662f\uff1a<\/p>\n\n\n\n<p>[<br>L(\\theta) = &#8211; \\mathbb{E}<em>{(x, y_c, y_r)} \\Big[ \\log \\sigma\\big( R<\/em>\\theta(x, y_c) &#8211; R_\\theta(x, y_r) \\big) \\Big]<br>]<\/p>\n\n\n\n<p>\u5176\u4e2d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>(\\sigma(z) = \\frac{1}{1 + e^{-z}}) \u662f sigmoid \u51fd\u6570\uff0c<\/li>\n\n\n\n<li>\u76f4\u89c2\u7406\u89e3\uff1aRM \u5b66\u4f1a\u628a \u201cchosen\u201d \u7684\u5206\u6570\u63a8\u9ad8\uff0c\u201crejected\u201d \u7684\u5206\u6570\u538b\u4f4e\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">4. \u7b80\u5355\u89e3\u91ca<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5982\u679c (R_\\theta(y_c) \\gg R_\\theta(y_r))\uff0c\u8bf4\u660e RM \u5224\u65ad\u6b63\u786e\uff0closs \u4f1a\u5f88\u5c0f\u3002<\/li>\n\n\n\n<li>\u5982\u679c (R_\\theta(y_c) \\approx R_\\theta(y_r))\uff0closs \u4f1a\u6bd4\u8f83\u5927\uff0c\u56e0\u4e3a\u6a21\u578b\u6ca1\u6709\u533a\u5206\u5f00\u4e24\u8005\u3002<\/li>\n\n\n\n<li>\u5982\u679c (R_\\theta(y_c) &lt; R_\\theta(y_r))\uff0closs \u6700\u5927\uff0c\u8bf4\u660e\u6a21\u578b\u5b8c\u5168\u5224\u65ad\u9519\u4e86\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u8f68\u8ff9\u6982\u7387\u516c\u5f0f<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-15.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"617\" height=\"118\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-15.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-781\"  sizes=\"auto, (max-width: 617px) 100vw, 617px\" \/><\/div><\/figure>\n\n\n\n<p>\u5bf9\u4e8e\u4e00\u4e2a\u8f68\u8ff9\u5728\u67d0\u4e2a\u7b56\u7565\u4e0b\u7684\u53ef\u80fd\u6027\uff0c\u5206\u6790\u5b83\u7684\u5f71\u54cd\u56e0\u7d20\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u521d\u59cb\u73af\u5883\uff1a\u03c10\uff08s0\uff09<\/li>\n\n\n\n<li>\u7b56\u7565\u9009\u62e9\u6982\u7387\uff1a\u03a0(at|st)\uff0c\u5373\u5728\u67d0\u72b6\u6001\u4e0b\uff0c\u8be5\u7b56\u7565\u9009\u62e9\u884c\u52a8at\u7684\u6982\u7387<\/li>\n\n\n\n<li>\u73af\u5883\u5f71\u54cd\u56e0\u7d20\uff1aP(st+1|st,at)\uff0c\u6709\u7684\u65f6\u5019\u4e8b\u60c5\u5e76\u4e0d\u90fd\u5982\u6211\u4eec\u6240\u6599\uff0c\u56e0\u4e3a\u73b0\u5b9e\u73af\u5883\u5177\u6709\u590d\u6742\u7684\u7279\u5f81\u3002\u5728\u4e00\u4e2a\u4e0b\u68cb\u7684\u6e38\u620f\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u8ba4\u4e3a\u8fd9\u4e00\u9879\u6052\u4e3a1\uff0c\u56e0\u4e3a\u8fd9\u662f\u4e00\u4e2a\u5f88\u786e\u5b9a\u7684\u573a\u666f\u3002\u4f46\u5047\u5982\uff0c\u6211\u4eec\u662f\u5728\u8ba8\u8bba\u4e00\u4e2a\u73b0\u5b9e\u573a\u666f\u4e2d\u7684\u673a\u5668\u4eba\uff0c\u5b83\u4e0d\u4e00\u5b9a\u4f1a\u6839\u636e\u9009\u62e9\u7684\u7b56\u7565\u505a\u51fa\u54cd\u5e94\u7684\u884c\u52a8\uff0c\u53ef\u80fd\u5173\u8282\u78e8\u635f\uff0c\u53ef\u80fd\u65ad\u7535\u4e86\u7b49\u7b49\u3002\u8fd9\u4e00\u9879\u7684\u6838\u5fc3\uff0c\u5728\u4e8e\u73af\u5883\u56e0\u7d20\u4f1a\u5f71\u54cdagent\u7684\u72b6\u6001\u8f6c\u79fb\uff0c\u7136\u540e\u6211\u4eec\u628a\u8fd9\u4e2a\u8f6c\u79fb\u7684\u6982\u7387\u5f15\u5165\u8f68\u8ff9\u7684\u8ba1\u7b97\u516c\u5f0f\u4e2d\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5bf9\u5e94\u7684discount reward\uff08\u6298\u6263\u56de\u62a5\uff09<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e3a\u4e86\u9632\u6b62\u5956\u52b1\u7684\u53d1\u6563\uff0c\u540c\u65f6\u9f13\u52b1\u66f4\u77ed\u7684path\uff0c\u4e3a\u5956\u52b1\u7684\u7d2f\u52a0\u6dfb\u52a0\u4e86\u6839\u636e\u6b65\u6570\u7684\u6298\u6263\u56e0\u5b50<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-16.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"287\" height=\"131\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-16.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-783\"\/><\/div><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u8bed\u8a00\u6a21\u578b\u4e2d\u7684trajectories\uff1a\u4ece\u8fd9\u5f20\u56fe\u4e2d\u80fd\u5f88\u663e\u7136\u7684\u770b\u51fa\u6765state\u548caction\u3002state\u6307\u6bcf\u6b21\u90fd\u4f1a\u628a\u5f53\u524d\u5f97\u5230\u7684prompt\u8f93\u8fdb\u53bb\uff0caction\u6307\u6839\u636e\u8fd9\u4e2aprompt\u9009\u62e9\u4e00\u4e2atoken\uff0c\u7136\u540e\u91cd\u590d\u8fd9\u4e2a\u8fc7\u7a0b<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-17-1024x466.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"466\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-17-1024x466.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-785\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<p>\u68af\u5ea6\u4f18\u5316\u7b56\u7565\u7684\u516c\u5f0f\u63a8\u5bfc\u8fc7\u7a0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>1. \u671f\u671b\u5c55\u5f00<\/li>\n\n\n\n<li>2. \u68af\u5ea6\u7b26\u53f7\u79fb\u52a8<\/li>\n\n\n\n<li>3. \u4e00\u4e2a\u5bf9\u6570\u7684\u5c0f\u7b56\u7565\uff1a\u628a\u6982\u7387\u63d0\u53d6\u51fa\u6765<\/li>\n\n\n\n<li>4. \u628a\u63d0\u53d6\u51fa\u6765\u7684\u6982\u7387\u8f6c\u5316\u56de\u671f\u671b\u7684\u5f62\u5f0f<\/li>\n\n\n\n<li>5. \u4ee3\u5165\u8f68\u8ff9\u6982\u7387\u516c\u5f0f\uff0c\u53d1\u73b0\u9664\u4e86\u7b56\u7565\u9009\u62e9\u6982\u7387\u4e4b\u5916\uff0c\u521d\u59cb\u73af\u5883\u548c\u4e2d\u95f4\u7684\u73af\u5883\u5f71\u54cd\u548c \u03b8 \u65e0\u5173\u88ab\u6d88\u6389<\/li>\n\n\n\n<li>\u8865\u5145\uff1a\u8fd9\u91cc\u7684 \u03b8 \u5c31\u662f\u53c2\u6570\uff0c\u901a\u5e38\u662f\u795e\u7ecf\u7f51\u7edc\u6743\u91cd\u548c\u504f\u7f6e\u3002\u6839\u636e\u5e38\u8bc6\uff0c\u6211\u4eec\u7684\u795e\u7ecf\u7f51\u7edc\u56de\u6839\u636e\u8fd9\u4e9b\u53c2\u6570\uff0c\u5c06input\u8f6c\u5316\u4e3a\u4e00\u4e2a\u786e\u5b9a\u7684output\uff0c\u4e5f\u56e0\u6b64\uff0c\u8fd9\u4e2a\u53c2\u6570\u5b9e\u9645\u4e0a\u51b3\u5b9a\u4e86\u7b56\u7565\u9009\u62e9\u6982\u7387<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-18-1024x345.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"345\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-18-1024x345.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-787\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<p>\u5f53\u7136\uff0c\u6211\u4eec\u5728\u8ba1\u7b97\u8fd9\u6837\u7684\u671f\u671b\u7684\u65f6\u5019\uff0c\u4e0d\u53ef\u80fd\u771f\u7684\u8ba1\u7b97\u6240\u6709\u8f68\u8ff9\u7684\u60c5\u51b5\u3002\u56e0\u6b64\u5728\u8fd9\u91cc\uff0c\u91c7\u7528\u7684\u662f\u6837\u672c\u4f30\u8ba1\u7684\u65b9\u6cd5\u3002\u8fd9\u65f6\u5019\u6211\u4eec\u5f88\u5bb9\u6613\u5c31\u80fd\u628a\u5b83\u548c\u6240\u8c13\u7684\u201c\u8bad\u7ec3\u6570\u636e\u201d\u8054\u7cfb\u8d77\u6765\uff0c\u4e5f\u80fd\u7406\u89e3\u4e3a\u4ec0\u4e48\u6570\u636e\u6709\u8d28\u91cf\u7684\u5dee\u522b\uff0c\u56e0\u4e3a\u672c\u8d28\u4e0a\uff0c\u4ed6\u4eec\u90fd\u662f\u7528\u6765\u903c\u8fd1\u7406\u60f3\u671f\u671b\u7684\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-19-1024x279.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"279\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-19-1024x279.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-789\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u80fd\u5f97\u5230\u5982\u4e0a\u56fe\u7684\u5f3a\u5316\u5b66\u4e60\u57fa\u672c\u6d41\u7a0b<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>1. \u5b9a\u4e49\u7f51\u7edc\u548c\u7b56\u7565<\/li>\n\n\n\n<li>2. \u7528\u7b56\u7565\u5728\u7f51\u7edc\u4e0a\u8dd1\uff0c\u5f97\u5230\u4e00\u4e9b\u8f68\u8ff9\u5bf9\u5e94\u7684reward<\/li>\n\n\n\n<li>3. \u7528\u8fd9\u4e9b\u6837\u672c\u53bb\u8ba1\u7b97\u68af\u5ea6<\/li>\n\n\n\n<li>4. \u7528\u968f\u673a\u68af\u5ea6\u4e0a\u5347\u6cd5\uff0c\u66f4\u65b0\u7f51\u7edc\u91cc\u9762\u7684\u53c2\u6570<\/li>\n\n\n\n<li>5. \u56de\u5230\u7b2c\u4e8c\u6b65<\/li>\n<\/ul>\n\n\n\n<p>\u6240\u8c13\u7684\u5f3a\u5316\uff0c\u4ece\u8fd9\u4e2a\u6d41\u7a0b\u4e0a\u6765\u8bb2\uff0c\u5c31\u662f\u4e00\u4e2a\u8fd9\u6837\u8fed\u4ee3\u7684\u8fc7\u7a0b\uff0c\u76ee\u7684\u662f\u4f7fJ(\u03a0\u03b8 )\u6700\u5927\u5316\uff0c\u4e5f\u5373\u4f18\u5316\u7b56\u7565\u4f7fagent\u5728\u8fdb\u884caction\u7684\u65f6\u5019\uff0c\u5c3d\u53ef\u80fd\u83b7\u5f97\u66f4\u591a\u7684reward<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u5bf9\u4e8e\u8bed\u8a00\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\u6d41\u7a0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-20-1024x554.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"554\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-20-1024x554.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-791\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>logits<\/strong> \u662f\u4e00\u4e2a\u957f\u5ea6\u4e3a\u8bcd\u8868\u5927\u5c0f\u7684\u5411\u91cf\uff0c\u6bcf\u4e2a\u5206\u91cf\u662f\u201c\u5019\u9009\u4e0b\u4e00\u4e2a token \u7684\u672a\u5f52\u4e00\u5316\u5206\u6570\u201d\u3002<\/li>\n\n\n\n<li>\u628a logits \u7ecf\u8fc7 <strong>softmax \u2192 \u6982\u7387\u5206\u5e03<\/strong>\uff0c\u518d\u901a\u8fc7 <strong>\u8d2a\u5fc3\/\u91c7\u6837\/\u6e29\u5ea6\/Top-k\/Top-p<\/strong> \u7b49\u7b56\u7565\uff0c\u5c31\u80fd\u6700\u7ec8\u51b3\u5b9a\u4e0b\u4e00\u4e2a token\u3002<\/li>\n\n\n\n<li>\u5728\u4e0a\u56fe\u4e2d\u7684\u4e00\u4e2a\u5b9e\u4f8b\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u51fa\uff0c\u5728\u7ecf\u8fc7transformer\u5904\u7406\u4e4b\u540e\uff0c\u6bcf\u4e00\u4e2atoken\u5904\u90fd\u5bf9\u5e94\u751f\u6210\u4e86\u4e00\u4e2alogits\uff0c\u5bf9\u5e94\u4e86<strong>\u5f53\u8f93\u5165\u4e86\u8fd9\u4e2atoken\u53ca\u4e4b\u524d\u7684\u6240\u6709token\u65f6<\/strong>\uff0cmodel\u5bf9\u4e0b\u4e00\u4e2atoken\u7684\u5224\u65ad<\/li>\n\n\n\n<li>\u6839\u636e\u8f68\u8ff9\u7684\u77e5\u8bc6\uff0c\u5f88\u5bb9\u6613\u8fd9\u6837\u5f97\u5230\u6bcf\u4e2a\u4f4d\u7f6e\u5bf9\u5e94\u7684\u6982\u7387\u9009\u62e9\u5206\u6570\u3002\u6bd4\u5982\u5728\u201c\uff1f\u201d\u8fd9\u4e2astate\u4e0b\uff0c\u8fdb\u884caction\u201cShanghai\u201d\u7684\u6982\u7387\uff0c\u5176\u5b9e\u5c31\u662f\u8fd9\u4e2a\u4f4d\u7f6e\u7684logits\u5f52\u4e00\u5316\u4e4b\u540e\uff0c\u5411\u91cf\u5bf9\u5e94\u201cShanghai\u201d\u4f4d\u7f6e\u7684\u6570\u503c\u3002<\/li>\n\n\n\n<li>\u8fd9\u4e2a\u503c\u5c31\u53ef\u4ee5\u8fd9\u6837\u88ab\u8bb0\u5f55\u4e0b\u6765\uff0c\u4f5c\u4e3a\u540e\u7eed\u8ba1\u7b97\u7684\u4f9d\u636e<\/li>\n\n\n\n<li>\u4e00\u4e2a\u9700\u8981\u6ce8\u610f\u7684\u4e8b\uff1a\u672c\u8d28\u4e0a\u8fd9\u4e2ainput\u662f\u7531question\u548canswer\u62fc\u63a5\u8d77\u6765\u7684\uff0c\u800c\u6211\u4eec\u5176\u5b9e\u53ea\u9700\u8981anwer\u7684\u90a3\u90e8\u5206\u3002\u56e0\u6b64question\u53ef\u4ee5\u88ab\u89c6\u4f5c\u662f\u5f00\u59cb\u7684\u73af\u5883\uff0c\u800c\u9009\u62e9\u201cShanghai\u201d\uff0c\u624d\u662f\u7b2c\u4e00\u4e2aaction\u3002\u5b9e\u9645\u4e0a\u6211\u4eec\u9700\u8981\u8ba1\u7b97\u7684\u4e5f\u53ea\u6709\u56db\u6b65\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-21-1024x551.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"551\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-21-1024x551.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-793\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5bf9\u4e8eReward\u8ba1\u7b97\uff0c\u7c7b\u4f3c\u7684\u53ef\u4ee5\u8fd9\u6837\u8fdb\u884c\uff0c\u5bf9\u6bcf\u4e00\u4e2a\u9700\u8981\u8ba1\u7b97reward\u7684\u5730\u65b9\uff0c\u8fc7\u4e00\u4e2a\u7ebf\u6027\u5c42\u8f93\u51fa\u5206\u6570\u3002\uff08\u56de\u5fc6\u4e4b\u524d\u7684discount reward\uff0c\u8fd9\u91cc\u5b9e\u9645\u4e0a\u4e5f\u8981\u8fd9\u4e48\u8ba1\u7b97\uff0c\u56e0\u4e3a\u672c\u8d28\u4e0a\u751f\u6210\u4e00\u4e2atoken\u7c7b\u4f3c\u4e8ecat\u8d70\u4e00\u6b65\uff0c\u6240\u4ee5\u5728\u8ba1\u7b97reward\u65f6\uff0c\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u7d2f\u52a0\uff09<\/li>\n\n\n\n<li>\u7ea0\u6b63\u4e00\u4e2a\u601d\u7ef4\u8bef\u533a\uff1a\u6211\u4e4b\u524d\u597d\u50cf\u8ba4\u4e3a\u5bf9\u4e8e\u9879\uff0cR\uff08r\uff09\u8fd9\u4e2a\u503c\u662f\u5728<strong>\u8fd9\u4e00\u6b65\u4e4b\u524d\u6240\u6709\u7684\u5956\u52b1\u5206\u91cf\u7d2f\u52a0<\/strong>\uff0c\u4f46\u8fd9\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u9519\u8bef\uff0c\u56e0\u4e3a\u5728\u516c\u5f0f\u63a8\u5bfc\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5e76\u6ca1\u6709\u4efb\u4f55\u4e00\u6b65\u8868\u793a\u4e86\u8fd9\u4e2a\u610f\u601d\uff0c\u4ece\u5934\u5230\u5c3e\u8fd9\u4e00\u9879\u90fd\u4ee3\u8868\u4e86\u5bf9\u4e8e\u6574\u6761\u8f68\u8ff9\u6765\u8bf4\u7684<strong>\u5956\u52b1\u9879\u4e4b\u548c<\/strong>\uff0c\u4e5f\u5c31\u662f\u5bf9\u4e8e\u4e00\u6761\u8f68\u8ff9\u6765\u8bf4\uff0c\u8fd9\u4e00\u9879\u662f\u5b9a\u503c\uff0c\u53ea\u4e0d\u8fc7\u5728\u8ba1\u7b97\u7684\u65f6\u5019\u9700\u8981\u8003\u8651\u5728\u6bcf\u4e00\u6b65\u4e2d\u7684\u5956\u52b1\u5206\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-16.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"287\" height=\"131\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-16.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-783\"\/><\/div><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8865\u5145\u4e00\u4e2a\u8bad\u7ec3\u4e0a\u7684\u7591\u60d1\uff1a\u4e4b\u524d\u4e00\u76f4\u5728\u60f3\uff0c\u5982\u679c\u8fd9\u6837\u8f93\u5165\u7684\u8bdd\uff0c\u90a3\u201d\u6bd4\u8f83\u201c\u8fd9\u4e2a\u6982\u5ff5\u5c31\u4e0d\u77e5\u9053\u600e\u6837\u4f53\u73b0\u4e86\u3002\u5b9e\u9645\u4e0a\uff0c\u771f\u6b63\u8bad\u7ec3\u65f6\u5019\uff0c\u662f\u4ee5batch\u4e3a\u5355\u4f4d\uff0c\u6bcf\u4e00\u4e2abatch\u5bf9\u5e94\u4e00\u4e2aQ\u548c\u82e5\u5e72A\uff0cA\u5e26\u6709\u4eba\u4e3a\u7684\u6807\u6ce8\u3002\u8fd9\u6837\u7684\u4e00\u4e2abatch\u624d\u662f\u4e00\u4e2a\u6709\u6548\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u53ef\u4ee5\u8fdb\u884c\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p><strong>\u66f4\u597d\u7684\u5956\u52b1\u51fd\u6570\uff1aReward-to-go<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e4b\u524d\u63d0\u5230\uff0c\u5728\u8ba1\u7b97\u6bcf\u4e00\u9879\u7684\u65f6\u5019\uff0c\u5956\u52b1\u9879\u662f\u4e00\u4e2a\u5b9a\u503c\u3002\u4f46\u4ece\u5e38\u8bc6\u7684\u89d2\u5ea6\u4e0a\u6765\u8bb2\uff0c\u672a\u6765\u7684\u52a8\u4f5c\u4e0d\u4f1a\u5f71\u54cd\u8fc7\u53bb\u7684\u5956\u52b1\uff0c\u4e5f\u56e0\u6b64\uff0c\u5bf9\u4e8e\u6bd4\u8f83\u540e\u9762\u7684\u52a8\u4f5c\u6765\u8bf4\uff0c\u7531\u4e8e\u5b83\u5b9e\u9645\u4e0a\u6839\u672c\u4e0d\u4f1a\u6539\u53d8\u524d\u9762\u7684\u5956\u52b1\u9879\uff0c\u8fd9\u4e9b\u9879\u5bf9\u8fd9\u91cc\u7684\u68af\u5ea6\u662f\u6ca1\u6709\u8d21\u732e\u7684\u3002\u76f8\u5f53\u4e8e\u566a\u58f0\u3002<\/li>\n\n\n\n<li>\u56e0\u6b64\u5c06\u539f\u672c\u7684\u5168\u8f68\u8ff9\u56de\u62a5\u6539\u4e3areward-to-go\uff0c\u53ea\u8ba1\u7b97\u8fd9\u4e4b\u540e\u7684\u5956\u52b1\u9879\u7684\u5f71\u54cd<\/li>\n\n\n\n<li><strong>\u5168\u8f68\u8ff9\u56de\u62a5<\/strong>\uff1aR(\u03c4)=1+2+3=6<\/li>\n\n\n\n<li><strong>reward-to-go<\/strong>\uff1a<\/li>\n\n\n\n<li>G0=1+2+3=6<\/li>\n\n\n\n<li>G1=2+3=5<\/li>\n\n\n\n<li>G2=3<\/li>\n\n\n\n<li><strong>Reward-to-go<\/strong> \u7684\u6838\u5fc3\u539f\u7406\u662f <strong>\u56e0\u679c\u6027<\/strong>\uff1a\u5f53\u524d\u52a8\u4f5c\u53ea\u5f71\u54cd\u672a\u6765\u5956\u52b1<\/li>\n\n\n\n<li>\u516c\u5f0f\u6210\u7acb\u7684\u539f\u56e0\uff1a\u5bf9\u4e8e t \u65f6\u523b\u4e4b\u524d\u7684\u5956\u52b1\uff0c\u5b83\u5bf9 \u2207\u03b8log\u2061\u03c0\u03b8(at\u2223st)\\nabla_\\theta \\log \\pi_\\theta(a_t|s_t)\u2207\u03b8\u200blog\u03c0\u03b8\u200b(at\u200b\u2223st\u200b) \u7684\u671f\u671b\u4e3a 0<\/li>\n\n\n\n<li>\u597d\u5904\uff1a\u964d\u4f4e\u68af\u5ea6\u65b9\u5dee\uff0c\u4f7f\u7b56\u7565\u68af\u5ea6\u66f4\u7a33\u5b9a<\/li>\n<\/ul>\n\n\n\n<p><strong>\u53e6\u4e00\u79cd\u65b9\u5f0f\uff1abaseline<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-22-1024x247.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"247\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-22-1024x247.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-801\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728\u5956\u52b1\u9879\u4e2d\u5f15\u5165\u4e00\u4e2abaseline\uff0c\u53ef\u4ee5\u662f\u5e38\u6570\uff0c\u4e5f\u53ef\u4ee5\u662f\u548cs\u548c\u03a0\u6709\u5173\u7684\u4e00\u4e2a\u51fd\u6570\uff0c\u79f0\u4e3aValue\u51fd\u6570<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li>V\u03a0(s)\uff0c\u5728\u8fd9\u91cc\u7684\u610f\u4e49\u548c\u201c\u672a\u6765\u7684\u671f\u671b\u5956\u52b1\u201d\u6709\u5173\uff0c\u8868\u793a\u5728\u5f53\u524d\u7b56\u7565\u548c\u72b6\u6001\u4e0b\uff0c\u5bf9\u672a\u6765\u80fd\u5f97\u5230\u5956\u52b1\u7684\u9884\u671f<\/li>\n\n\n\n<li>\u548cReward\u7c7b\u4f3c\u7684\uff0c\u8fd9\u4e2a\u51fd\u6570\u4e5f\u53ef\u4ee5\u7528\u90a3\u79cd\u6a21\u5f0f\u6765\u9884\u6d4b<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p><strong>\u5f15\u5165\u8fd9\u4e24\u9879\u540e\uff0c\u5f97\u5230\u4e86\u6700\u540e\u7684\u5f62\u6001<\/strong>\uff1aA\u03c0(s,a) &#8212;<strong>Advantage function<\/strong><\/p>\n\n\n\n<p>\u4f18\u52bf\u51fd\u6570 A\u03c0(s,a)A^\\pi(s,a)A\u03c0(s,a) \u63cf\u8ff0\u7684\u662f <strong>\u5728\u72b6\u6001 s \u4e0b\uff0c\u91c7\u53d6\u52a8\u4f5c a \u76f8\u5bf9\u4e8e\u5e73\u5747\u6c34\u5e73\u7684\u597d\u574f<\/strong>\uff1a A\u03c0(s,a)=Q\u03c0(s,a)\u2212V\u03c0(s)<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Q\u03c0(s,a)\uff1a\u7b56\u7565 \u03c0 \u4e0b\uff0c\u4ece\u72b6\u6001 s \u91c7\u53d6\u52a8\u4f5c a \u7684\u671f\u671b\u7d2f\u79ef\u56de\u62a5<\/li>\n\n\n\n<li>V\u03c0(s)\uff1a\u7b56\u7565 \u03c0 \u4e0b\uff0c\u4ece\u72b6\u6001 s \u5f00\u59cb\u7684\u5e73\u5747\u671f\u671b\u56de\u62a5\uff08\u5373\u91c7\u53d6\u52a8\u4f5c\u7684\u5e73\u5747\u6c34\u5e73\uff09<\/li>\n<\/ul>\n\n\n\n<p><strong>\u76f4\u89c2\u89e3\u91ca<\/strong>\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>Advantage function \u8861\u91cf\u67d0\u4e2a action \u5728\u5f53\u524d\u72b6\u6001\u4e0b\uff0c\u6bd4\u201c\u5e73\u5747\u52a8\u4f5c\u201d\u591a\u8d5a\u4e86\u591a\u5c11\u5956\u52b1\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>A(s,a)&gt;0\uff1a\u8fd9\u4e2a\u52a8\u4f5c\u6bd4\u5e73\u5747\u6c34\u5e73\u597d<\/li>\n\n\n\n<li>A(s,a)&lt;0\uff1a\u8fd9\u4e2a\u52a8\u4f5c\u6bd4\u5e73\u5747\u6c34\u5e73\u5dee<\/li>\n\n\n\n<li>A(s,a)=0\uff1a\u52a8\u4f5c\u4e0e\u5e73\u5747\u6c34\u5e73\u4e00\u6837\u597d<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<p>\u800c\u672c\u8d28\u4e0a\u6211\u4eec\u8981\u505a\u7684\u68af\u5ea6\u4e0a\u5347\u7684\u8fc7\u7a0b\uff0c\u5c31\u662f\u8ba9\u66f4\u597d\u7684\u52a8\u4f5c\uff0c\u66f4\u53ef\u80fd\u53d1\u751f\u3002<\/p>\n\n\n\n<p>PS\uff1a\u8fd9\u4e2a\u51cf\u53bb\u5e73\u5747\u6c34\u5e73\u7684\u6bd4\u65b9\u771f\u7684\u5f88\u5f62\u8c61\uff0c\u4e00\u4e0b\u611f\u89c9\u56de\u5230\u4e86\u9ad8\u4e2d\u6570\u5b66\uff0c\u505a\u7edf\u8ba1\u8868\u7684\u65f6\u5019<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p><strong>\u89e3\u51b3\u8ba1\u7b97\u95ee\u9898\uff1a\u600e\u6837\u8ba1\u7b97A\u03c0(s,a) <\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-24-1024x509.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"509\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-24-1024x509.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-810\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e0a\u9762\u8fd9\u4e9b\u8ba1\u7b97\u4f53\u73b0\u51fa\u4e00\u4e9b\u95ee\u9898<\/li>\n\n\n\n<li>\u7b2c\u4e00\u79cd\u65b9\u5f0f\uff1a\u4f7f\u7528\u4e86\u4e00\u4e2a\u5e73\u5747\u503c\u4ee3\u66ff\u4e86\u540e\u9762\u6240\u6709\u7684\u5177\u4f53\u884c\u52a8\u503c\uff0chigh bias<\/li>\n\n\n\n<li>\u968f\u7740\u5c55\u5f00\u7684\u8d8a\u6765\u8d8a\u591a\uff0cbias\u964d\u4f4e\u4e86\uff0c\u4f46\u662f\u65b9\u5deevariance\u63d0\u9ad8\u4e86<\/li>\n\n\n\n<li>\u8fd9\u662f\u4e00\u4e2a\u5728\u4e24\u8005\u4e4b\u95f4\u5e73\u8861\u7684\u95ee\u9898<\/li>\n\n\n\n<li>\u56e0\u6b64\u6700\u540e\uff0c\u901a\u8fc7\u4e0b\u9762\u8fd9\u4e2a\u9012\u5f52\u7684\u65b9\u5f0f\u8ba1\u7b97<strong>A\u03c0(s,a) <\/strong><\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">1 Off-Policy Learning\uff08\u79bb\u7b56\u7565\u5b66\u4e60\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5b9a\u4e49<\/strong>\uff1a\u4f7f\u7528 <strong>\u884c\u4e3a\u7b56\u7565 (\\mu(a|s))<\/strong> \u751f\u6210\u6570\u636e\uff0c\u4f46\u8bad\u7ec3\/\u4f18\u5316\u7684\u662f <strong>\u76ee\u6807\u7b56\u7565 (\\pi_\\theta(a|s))<\/strong><\/li>\n\n\n\n<li><strong>\u4f18\u70b9<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u53ef\u4ee5\u590d\u7528\u5386\u53f2\u6570\u636e\u3001\u65e7\u6a21\u578b\u6570\u636e\u3001\u4eba\u7c7b\u6f14\u793a<\/li>\n\n\n\n<li>\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\uff0c\u4e0d\u5fc5\u6bcf\u6b21\u90fd\u7528\u6700\u65b0\u7b56\u7565\u751f\u6210<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u95ee\u9898<\/strong>\uff1a\u76f4\u63a5\u4f7f\u7528 (\\mu) \u7684\u6570\u636e\u66f4\u65b0 (\\pi_\\theta) \u4f1a\u6709\u504f\u5dee \u2192 \u9700\u8981\u4fee\u6b63<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">2  Importance Sampling\uff08\u91cd\u8981\u6027\u91c7\u6837\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u76ee\u7684<\/strong>\uff1a\u5728\u5206\u5e03\u4e0d\u4e00\u81f4\u65f6\u505a\u671f\u671b\u4f30\u8ba1<\/li>\n\n\n\n<li><strong>\u6838\u5fc3\u516c\u5f0f<\/strong>\uff1a<br>[<br>\\mathbb{E}<em>{a\\sim \\pi}[f(a)] = \\mathbb{E}<\/em>{a\\sim \\mu} \\Big[ \\frac{\\pi(a|s)}{\\mu(a|s)} f(a) \\Big]<br>]<\/li>\n\n\n\n<li><strong>\u542b\u4e49<\/strong>\uff1a\u7528\u6743\u91cd (\\frac{\\pi}{\\mu}) \u8c03\u6574\u65e7\u6570\u636e\uff0c\u4f7f\u68af\u5ea6\u66f4\u65b0\u65e0\u504f<\/li>\n\n\n\n<li><strong>\u6ce8\u610f\u4e8b\u9879<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6743\u91cd\u8fc7\u5927 \u2192 \u68af\u5ea6\u65b9\u5dee\u5927<\/li>\n\n\n\n<li>\u5b9e\u8df5\u4e2d\u5e38\u7528 <strong>\u622a\u65ad \/ Clipped IS<\/strong> \u964d\u4f4e\u65b9\u5dee<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">3 PPO Loss\uff08Proximal Policy Optimization\uff09<\/h3>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-25-1024x652.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"652\" data-original=\"https:\/\/ferina.top\/wp-content\/uploads\/2025\/09\/image-25-1024x652.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-812\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 Clipped Surrogate Loss<\/h3>\n\n\n\n<p>[<br>L^{CLIP}(\\theta) = \\mathbb{E}_t \\Big[ \\min \\big( r_t(\\theta) A_t, \\ \\text{clip}(r_t(\\theta), 1-\\epsilon,1+\\epsilon) A_t \\big) \\Big]<br>]<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>(r_t(\\theta) = \\frac{\\pi_\\theta(a_t|s_t)}{\\pi_{\\theta_\\text{old}}(a_t|s_t)}) \u2192 <strong>\u91cd\u8981\u6027\u91c7\u6837\u6bd4\u7387<\/strong><\/li>\n\n\n\n<li>(A_t = G_t &#8211; V(s_t)) \u2192 <strong>Advantage function<\/strong><\/li>\n\n\n\n<li>clip\/min \u2192 \u9650\u5236\u7b56\u7565\u66f4\u65b0\u5e45\u5ea6\uff0c\u4fdd\u8bc1\u7a33\u5b9a\u6027<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 Value Loss\uff08\u72b6\u6001\u4ef7\u503c\u635f\u5931\uff09<\/h3>\n\n\n\n<p>[<br>L^{VF}(\\phi) = (V_\\phi(s_t) &#8211; G_t)^2<br>]<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7528\u4e8e\u8f85\u52a9 Advantage \u4f30\u8ba1\uff0c\u964d\u4f4e\u65b9\u5dee<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 Entropy Bonus<\/h3>\n\n\n\n<p>[<br>L^S = -\\beta H[\\pi_\\theta(\\cdot|s_t)]<br>]<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4fdd\u6301\u7b56\u7565\u63a2\u7d22\uff0c\u9632\u6b62\u8fc7\u65e9\u6536\u655b<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3.4 \u603b Loss<\/h3>\n\n\n\n<p>[<br>L^{PPO} = L^{CLIP} &#8211; c_1 L^{VF} + c_2 L^S<br>]<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6838\u5fc3\u601d\u60f3<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5229\u7528\u65e7\u7b56\u7565\u6570\u636e\u66f4\u65b0\u65b0\u7b56\u7565\uff08off-policy + IS\uff09<\/li>\n\n\n\n<li>\u7528 Advantage \u6307\u5bfc\u52a8\u4f5c\u9009\u62e9<\/li>\n\n\n\n<li>Clip \u4fdd\u8bc1\u68af\u5ea6\u7a33\u5b9a\uff0cEntropy \u4fdd\u6301\u63a2\u7d22<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">Reward Hacking<\/h3>\n\n\n\n<p>\u521a\u521a\u63d0\u5230\u7684\u6240\u6709\u7b56\u7565\u90fd\u662f\u4e3a\u4e86\u8ba9model\u751f\u6210\u66f4\u4ee4\u4eba\u6ee1\u610f\u7684\u7ed3\u679c\uff0c\u4f46\u662f\u8fd9\u4e9b\u7ed3\u679c\u4e0d\u4e00\u5b9a\u662fmake sense\u7684\u3002\u6bd4\u5982\u5982\u679c\u6211\u60f3\u8ba9\u4e00\u4e2amodel\u66f4\u52a0\u7684polite\uff0c\u7136\u540e\u628a\u5b83\u8bad\u6210\u4e86\u53ea\u4f1athank you\u7684model\uff0c\u8fd9\u663e\u7136\u662f\u4e0d\u597d\u7684<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u5e0c\u671b\u6700\u7ec8\u751f\u6210\u7684\u6a21\u578b\uff0c\u4e0d\u4f1a\u4e0e\u539f\u672c\u7684\u6a21\u578b\u53d1\u751f\u8f83\u5927\u7684\u504f\u79fb\uff0c\u4ee5\u6b64\u6765\u4fdd\u8bc1make sense<\/p>\n\n\n\n<p>\u6211\u4eec\u7684\u505a\u6cd5\u662f\uff0c\u4fdd\u5b58\u4e00\u4e2afrozen model\uff0c\u5f53\u4f7f\u7528Reward model\u53bb\u8bc4\u5224\u4e00\u4e9btoken\u7684\u65f6\u5019\uff0c\u5728\u5956\u52b1\u51fd\u6570\u4e0a\u51cf\u53bb\u4e00\u4e2aterm\uff0c\u4ee3\u8868frozen model\u548ccurrent model\u5bf9\u4e8e\u540c\u6837\u7684\u4e00\u4e2ahidden states\u751f\u6210\u7684probability\u5dee\u5f02\u3002<\/p>\n\n\n\n<p>\u901a\u8fc7\u8fd9\u6837\u7684\u65b9\u5f0f\uff0c\u6765\u60e9\u7f5a\u5bf9\u4e8e\u539f\u59cbmodel\u8fc7\u5927\u7684\u504f\u79fb\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Code<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">\u9636\u6bb51\uff1a\u6570\u636e\u51c6\u5907\u4e0e\u524d\u5411\u4f20\u64ad<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">1. <code>_step_safety_checker<\/code> &#8211; \u8f93\u5165\u5b89\u5168\u68c0\u67e5 &#8211; line 647<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>queries<\/code>: List[torch.LongTensor] &#8211; \u63d0\u793a\u6587\u672c\u7684token IDs<\/li>\n\n\n\n<li><code>responses<\/code>: List[torch.LongTensor] &#8211; \u65e7\u6a21\u578b\u751f\u6210\u54cd\u5e94\u7684token IDs<\/li>\n\n\n\n<li><code>scores<\/code>: List[torch.FloatTensor] &#8211; \u6bcf\u4e2aresponse\u5bf9\u5e94\u7684\u5956\u52b1\u5206\u6570<\/li>\n\n\n\n<li><code>response_masks<\/code>: Optional &#8211; \u54cd\u5e94token\u7684\u63a9\u7801<\/li>\n<\/ul>\n\n\n\n<p><strong>\u8f93\u51fa\uff1a<\/strong> \u7ecf\u8fc7\u9a8c\u8bc1\u548c\u683c\u5f0f\u5316\u7684\u76f8\u540c\u7ed3\u6784\u6570\u636e<br><strong>\u4f5c\u7528\uff1a<\/strong> \u786e\u4fdd\u8f93\u5165\u6570\u636e\u683c\u5f0f\u6b63\u786e\uff0c\u907f\u514d\u8fd0\u884c\u65f6\u9519\u8bef<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. <code>prepare_model_inputs<\/code> &#8211; \u6784\u5efa\u6a21\u578b\u8f93\u5165 &#8211; line 691<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> queries, responses<br><strong>\u8f93\u51fa\uff1a<\/strong> Dict\u5305\u542b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>input_ids<\/code>: (batch_size, seq_len) &#8211; \u62fc\u63a5\u540e\u7684query+response<\/li>\n\n\n\n<li><code>attention_mask<\/code>: (batch_size, seq_len) &#8211; padding\u63a9\u7801<br><strong>\u4f5c\u7528\uff1a<\/strong> \u5c06query\u548cresponse\u62fc\u63a5\u6210\u7edf\u4e00\u7684\u6a21\u578b\u8f93\u5165\u683c\u5f0f<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">!!! 3. <code>batched_forward_pass<\/code> &#8211; \u7b56\u7565\u6a21\u578b\u524d\u5411\u4f20\u64ad &#8211; line 730 \u6838\u5fc3\u51fd\u6570\uff1a\u7528\u4e8e\u8ba1\u7b97log\u6982\u7387\u548c\u4ef7\u683c\u51fd\u6570\u4f30\u8ba1\uff0c\u53c2\u8003\u57fa\u7840\u77e5\u8bc6\uff0c\u8fd9\u4e9b\u662f\u8ba1\u7b97\u68af\u5ea6\u7684\u91cd\u8981\u53c2\u6570\u90e8\u5206<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> self.model, queries, responses, model_inputs<br><strong>\u8f93\u51fa\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>all_logprobs<\/code>: (Batch_Size, Seq_Len-1) &#8211; \u6bcf\u4e2atoken\u7684log\u6982\u7387<\/li>\n\n\n\n<li><code>logits_or_none<\/code>: logits\u6216None<\/li>\n\n\n\n<li><code>values<\/code>: (Batch_Size, Seq_Len-1) &#8211; \u4ef7\u503c\u51fd\u6570\u4f30\u8ba1<\/li>\n\n\n\n<li><code>masks<\/code>: (Batch_Size, Seq_Len-1) &#8211; \u6709\u6548token\u63a9\u7801<br><strong>\u4f5c\u7528\uff1a<\/strong> \u8ba1\u7b97\u5f53\u524d\u7b56\u7565\u4e0b\u5404token\u7684\u6982\u7387\u5206\u5e03\u548c\u4ef7\u503c\u4f30\u8ba1<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">4. <code>batched_forward_pass<\/code> &#8211; \u53c2\u8003\u6a21\u578b\u524d\u5411\u4f20\u64ad &#8211; line 741<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> self.ref_model, queries, responses, model_inputs<br><strong>\u8f93\u51fa\uff1a<\/strong> \u540c\u4e0a\uff0c\u4f46\u57fa\u4e8e\u53c2\u8003\u6a21\u578b<br><strong>\u4f5c\u7528\uff1a<\/strong> \u8ba1\u7b97\u53c2\u8003\u7b56\u7565\u7684log\u6982\u7387\uff0c\u7528\u4e8eKL\u6563\u5ea6\u8ba1\u7b97<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">!!! 5. <code>compute_rewards<\/code> &#8211; \u5956\u52b1\u8ba1\u7b97 &#8211; line 758  \u5b9a\u4e49\u5728 line 1139<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> scores, all_logprobs, ref_logprobs, masks <\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>scores\uff1a\u662f\u4ece\u6570\u636e\u96c6\u4e2d\u5f97\u5230\u7684\u539f\u59cb\u6570\u636e\u7684\u8bc4\u5206\uff0c\u6210\u4e3areward\u9700\u8981\u52a0\u4e0aKL\u6563\u5ea6<\/li>\n<\/ul>\n\n\n\n<p><br><strong>\u8f93\u51fa\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>rewards<\/code>: (Batch_Size, Seq_Len-1) &#8211; \u6bcf\u4e2atoken\u7684\u5956\u52b1<\/li>\n\n\n\n<li><code>non_score_reward<\/code>: \u975e\u5206\u6570\u5956\u52b1\u90e8\u5206<\/li>\n\n\n\n<li><code>kls<\/code>: KL\u6563\u5ea6\u503c<br><strong>\u4f5c\u7528\uff1a<\/strong> \u7ed3\u5408\u5956\u52b1\u5206\u6570\u548cKL\u60e9\u7f5a\u8ba1\u7b97\u6bcf\u4e2atoken\u7684\u6700\u7ec8\u5956\u52b1<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">!!! 6. <code>compute_advantages<\/code> &#8211; \u4f18\u52bf\u51fd\u6570\u8ba1\u7b97 line 772 \u5b9a\u4e49\u5728line 1196<\/h4>\n\n\n\n<p>\u8fd9\u4e00\u6b65\u7684\u6839\u636e\u662f\u4e4b\u524d\u63d0\u5230\u7684\u9012\u63a8\u516c\u5f0f<\/p>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> values, rewards, masks<br><strong>\u8f93\u51fa\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>values<\/code>: \u4ef7\u503c\u4f30\u8ba1<\/li>\n\n\n\n<li><code>advantages<\/code>: (Batch_Size, Seq_Len-1) &#8211; \u4f18\u52bf\u51fd\u6570<\/li>\n\n\n\n<li><code>returns<\/code>: (Batch_Size, Seq_Len-1) &#8211; \u5b9e\u9645\u4e0a\u662fA + V = Q\uff0c\u662f\u7528\u6765\u505a\u4f30\u8ba1\u7528\u7684<br><strong>\u4f5c\u7528\uff1a<\/strong> \u4f7f\u7528GAE(\u5e7f\u4e49\u4f18\u52bf\u4f30\u8ba1)\u8ba1\u7b97\u4f18\u52bf\u51fd\u6570<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u9636\u6bb52\uff1aPPO\u4f18\u5316<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">7. Mini-batch\u8bad\u7ec3\u5faa\u73af<\/h4>\n\n\n\n<p>\u5728\u6574\u4e2abatch\u4e0a\u6267\u884c\u591a\u8f6e(ppo_epochs)\u4f18\u5316\uff1a<\/p>\n\n\n\n<p><strong>\u5916\u5c42\u5faa\u73af\uff1a<\/strong> \u968f\u673a\u6253\u4e71\u6837\u672c\u987a\u5e8f<br><strong>\u4e2d\u5c42\u5faa\u73af\uff1a<\/strong> \u6309backward_batch_size\u5206\u5757<br><strong>\u5185\u5c42\u5faa\u73af\uff1a<\/strong> \u6309mini_batch_size\u8fdb\u4e00\u6b65\u5206\u5757<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">8. <code>batched_forward_pass<\/code> &#8211; \u5728\u7ebf\u6a21\u578b\u524d\u5411\u4f20\u64ad line 828<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> \u5f53\u524d\u6a21\u578b\u53c2\u6570 + mini-batch\u6570\u636e<br><strong>\u8f93\u51fa\uff1a<\/strong> \u5f53\u524d\u7b56\u7565\u7684logprobs, logits, vpreds\uff08\u8fd9\u4e2a\u5c31\u662fvalue\uff09<br><strong>\u4f5c\u7528\uff1a<\/strong> \u83b7\u53d6\u5f53\u524d\u7b56\u7565\u7684\u8f93\u51fa\u7528\u4e8e\u635f\u5931\u8ba1\u7b97<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">9. <code>train_minibatch<\/code> &#8211; \u6838\u5fc3\u4f18\u5316\u6b65\u9aa4  line 837<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u65e7\u7b56\u7565\u7684logprobs, values<\/li>\n\n\n\n<li>\u65b0\u7b56\u7565\u7684logprobs, logits, vpreds<\/li>\n\n\n\n<li>advantages, returns, masks<\/li>\n<\/ul>\n\n\n\n<p><strong>\u8f93\u51fa\uff1a<\/strong> \u8bad\u7ec3\u7edf\u8ba1\u4fe1\u606fdict<br><strong>\u4f5c\u7528\uff1a<\/strong> \u8ba1\u7b97PPO\u635f\u5931\u5e76\u6267\u884c\u68af\u5ea6\u66f4\u65b0<\/p>\n\n\n\n<p>\uff01\uff01\uff01<strong>\u6838\u5fc3\u51fd\u6570PPO Loss\u8ba1\u7b97<\/strong> <strong>\u5b9a\u4e49\u5728 line 1224<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53c2\u8003PPO Loss\u7684\u5b9a\u4e49\uff0c\u8fdb\u884c\u8ba1\u7b97<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u9636\u6bb53\uff1a\u7edf\u8ba1\u8bb0\u5f55\u4e0e\u66f4\u65b0<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">10. <code>record_step_stats<\/code> &#8211; \u8bb0\u5f55\u8bad\u7ec3\u7edf\u8ba1<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> \u6240\u6709\u524d\u5411\u4f20\u64ad\u548c\u8bad\u7ec3\u8fc7\u7a0b\u7684\u8f93\u51fa<br><strong>\u8f93\u51fa\uff1a<\/strong> \u6c47\u603b\u7684\u8bad\u7ec3\u7edf\u8ba1\u4fe1\u606f<br><strong>\u4f5c\u7528\uff1a<\/strong> \u8bb0\u5f55KL\u6563\u5ea6\u3001\u5956\u52b1\u3001\u6bd4\u7387\u7b49\u5173\u952e\u6307\u6807<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">11. <code>kl_ctl.update<\/code> &#8211; KL\u63a7\u5236\u5668\u66f4\u65b0<\/h4>\n\n\n\n<p><strong>\u8f93\u5165\uff1a<\/strong> \u5f53\u524dKL\u6563\u5ea6\u503c, batch_size<br><strong>\u8f93\u51fa\uff1a<\/strong> \u66f4\u65b0KL\u60e9\u7f5a\u7cfb\u6570<br><strong>\u4f5c\u7528\uff1a<\/strong> \u81ea\u9002\u5e94\u8c03\u6574KL\u60e9\u7f5a\u5f3a\u5ea6<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u51c6\u5907\u9636\u6bb5<\/strong>\uff1a\u9a8c\u8bc1\u8f93\u5165 \u2192 \u6784\u5efa\u6a21\u578b\u8f93\u5165<\/li>\n\n\n\n<li><strong>\u8bc4\u4f30\u9636\u6bb5<\/strong>\uff1a\u7b56\u7565\u6a21\u578b\u524d\u5411 \u2192 \u53c2\u8003\u6a21\u578b\u524d\u5411 \u2192 \u5956\u52b1\u8ba1\u7b97 \u2192 \u4f18\u52bf\u4f30\u8ba1<\/li>\n\n\n\n<li><strong>\u4f18\u5316\u9636\u6bb5<\/strong>\uff1a\u591a\u8f6emini-batch\u8bad\u7ec3 \u2192 \u635f\u5931\u8ba1\u7b97 \u2192 \u53c2\u6570\u66f4\u65b0<\/li>\n\n\n\n<li><strong>\u8bb0\u5f55\u9636\u6bb5<\/strong>\uff1a\u7edf\u8ba1\u6536\u96c6 \u2192 KL\u63a7\u5236\u66f4\u65b0 \u2192 \u5b66\u4e60\u7387\u8c03\u6574<\/li>\n\n\n\n<li><\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u7ed9\u5b9a\u540c\u4e00\u4e2a\u8f93\u5165\uff0c\u6a21\u578b\u751f\u6210\u7684\u591a\u4e2a\u5019\u9009\u56de\u7b54\u4e2d\uff0c\u4eba\u7c7b\u4f1a\u5bf9\u8fd9\u4e9b\u56de\u7b54\u8fdb\u884c\u6392\u5e8f\u3002RM \u9700\u8981\u5b66\u4f1a\u9884\u6d4b\u8fd9\u4e2a\u6392\u5e8f\u3002 1. \u6570\u636e\u683c [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[31,30],"class_list":["post-777","post","type-post","status-publish","format-standard","hentry","category-8","tag-ai","tag-rlhf"],"_links":{"self":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/777","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/comments?post=777"}],"version-history":[{"count":35,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/777\/revisions"}],"predecessor-version":[{"id":824,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/posts\/777\/revisions\/824"}],"wp:attachment":[{"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/media?parent=777"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/categories?post=777"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ferina.top\/index.php\/wp-json\/wp\/v2\/tags?post=777"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}