{"id":25536,"date":"2025-04-22T10:33:38","date_gmt":"2025-04-22T02:33:38","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=25536"},"modified":"2025-04-22T10:33:38","modified_gmt":"2025-04-22T02:33:38","slug":"sebastian-raschka%e9%95%bf%e6%96%87%ef%bc%9adeepseek-r1%e3%80%81o3%e8%83%8c%e5%90%8e%ef%bc%8crl%e6%8e%a8%e7%90%86%e8%ae%ad%e7%bb%83%e6%ad%a3%e6%82%84%e6%82%84%e7%aa%81%e7%a0%b4%e4%b8%8a%e9%99%90","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/04\/22\/25536\/","title":{"rendered":"Sebastian Raschka\u957f\u6587\uff1aDeepSeek-R1\u3001o3\u80cc\u540e\uff0cRL\u63a8\u7406\u8bad\u7ec3\u6b63\u6084\u6084\u7a81\u7834\u4e0a\u9650"},"content":{"rendered":"<div><img data-dominant-color=\"5092ac\" data-has-transparency=\"false\" style=\"--dominant-color: #5092ac;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-full wp-image-25539\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/9376b4685e26e7ec128605ddd80e5e436c334b-300x167-1.jpg\" width=\"300\" height=\"167\" alt=\"\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/9376b4685e26e7ec128605ddd80e5e436c334b-300x167-1.jpg 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/9376b4685e26e7ec128605ddd80e5e436c334b-300x167-1-150x84.jpg 150w\" sizes=\"auto, (max-width: 300px) 100vw, 300px\" \/><\/div>\n<div><\/div>\n<div class=\"article-desc\">\u53ea\u9760\u6a21\u578b\u5c3a\u5bf8\u53d8\u5927\u5df2\u7ecf\u4e0d\u884c\u4e86\uff1f\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u63a8\u7406\u9700\u8981\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6765\u300c\u52a0 buff\u300d\u3002<\/div>\n<div id=\"postspictures\" class=\"article-content\">\n<div id=\"container\" class=\"container am-engine\" data-v-1d7a5742=\"\" data-element=\"root\">\n<p>\u8457\u540d AI \u7814\u7a76\u8005\u548c\u535a\u4e3b Sebastian Raschka \u53c8\u53cc\u53d2\u53d5\u66f4\u65b0\u535a\u5ba2\u4e86\u3002<\/p>\n<p>\u8fd9\u6b21\u7684\u4e3b\u9898\u662f\u300aLLM \u63a8\u7406\u7684\u5f3a\u5316\u5b66\u4e60\u73b0\u72b6\u300b\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202504\/21\/2426f7550b8292c7ad49094c64694eff0ab17a.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u535a\u5ba2\u5730\u5740\uff1ahttps:\/\/magazine.sebastianraschka.com\/p\/the-state-of-llm-reasoning-model-training<\/p>\n<p>\u8fd9\u4e2a\u6708 AI \u793e\u533a\u5f88\u70ed\u95f9\uff0c\u5c24\u5176\u662f Llama 4 \u548c GPT-4.5 \u7b49\u65b0\u65d7\u8230\u6a21\u578b\u7684\u53d1\u5e03\u3002\u4f46\u4f60\u53ef\u80fd\u5df2\u7ecf\u6ce8\u610f\u5230\uff0c\u4eba\u4eec\u5bf9\u8fd9\u4e9b\u65b0\u6a21\u578b\u7684\u53cd\u5e94\u76f8\u5bf9\u5e73\u6de1\u3002\u539f\u56e0\u4e4b\u4e00\u53ef\u80fd\u662f Llama 4 \u548c GPT-4.5 \u4ecd\u7136\u662f\u4f20\u7edf\u7684\u6a21\u578b\uff0c\u8fd9\u610f\u5473\u7740\u5b83\u4eec\u7684\u8bad\u7ec3\u6ca1\u6709\u4f7f\u7528\u660e\u786e\u7684\u5f3a\u5316\u5b66\u4e60\u8fdb\u884c\u63a8\u7406\u3002<\/p>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0cxAI \u548c Anthropic \u7b49\u5f3a\u52b2\u5bf9\u624b\u5728\u5176\u6a21\u578b\u4e2d\u589e\u52a0\u4e86\u66f4\u591a\u63a8\u7406\u80fd\u529b\u548c\u529f\u80fd\u3002\u4f8b\u5982\uff0cGrok \u548c Claude \u7684\u754c\u9762\u73b0\u5728\u90fd\u4e3a\u67d0\u4e9b\u6a21\u578b\u6dfb\u52a0\u4e86\u4e00\u4e2a\u300c\u601d\u8003\u300d\uff08\u6216\u6269\u5c55\u601d\u8003\uff09\u6309\u94ae\uff0c\u53ef\u4ee5\u660e\u786e\u5207\u6362\u63a8\u7406\u529f\u80fd\u3002<\/p>\n<p>\u65e0\u8bba\u5982\u4f55\uff0cLlama 4 \u548c GPT-4.5\uff08\u975e\u63a8\u7406\uff09\u6a21\u578b\u7684\u4f4e\u8ff7\u53cd\u54cd\u8868\u660e\uff0c\u6211\u4eec\u6b63\u63a5\u8fd1\u4ec5\u9760\u6269\u5c55\u6a21\u578b\u89c4\u6a21\u548c\u6570\u636e\u6240\u80fd\u8fbe\u5230\u7684\u6781\u9650\u3002<\/p>\n<p>\u7136\u800c\uff0cOpenAI \u8fd1\u671f\u53d1\u5e03\u7684 o3 \u63a8\u7406\u6a21\u578b\u8868\u660e\uff0c\u5728\u6218\u7565\u6027\u6295\u5165\u8ba1\u7b97\u8d44\u6e90\u65b9\u9762\uff0c\u7279\u522b\u662f\u901a\u8fc7\u9488\u5bf9\u63a8\u7406\u4efb\u52a1\u91cf\u8eab\u5b9a\u5236\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u4ecd\u6709\u76f8\u5f53\u5927\u7684\u6539\u8fdb\u7a7a\u95f4\u3002\u636e OpenAI \u5458\u5de5\u5728\u76f4\u64ad\u4e2d\u4ecb\u7ecd\uff0co3 \u4f7f\u7528\u7684\u8bad\u7ec3\u8ba1\u7b97\u8d44\u6e90\u662f o1 \u7684 10 \u500d\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202504\/21\/a30a9b6122c753585c571784da0dbdd69db33f.webp\" alt=\"image.png\" data-type=\"inline\" \/>\u56fe\u6e90\uff1aOpenAI o3 \u4e0e o1 \u7684\u6027\u80fd\u4e0e\u7b97\u529b\u6bd4\u8f83\u3002<\/p>\n<p>\u867d\u7136\u5355\u9760\u63a8\u7406\u5e76\u975e\u7075\u4e39\u5999\u836f\uff0c\u4f46\u786e\u5b9e\u80fd\u63d0\u5347\u6a21\u578b\u5728\u6311\u6218\u6027\u4efb\u52a1\u4e0a\u7684\u51c6\u786e\u7387\u548c\u89e3\u51b3\u95ee\u9898\u7684\u80fd\u529b\uff08\u76ee\u524d\u4e3a\u6b62\uff09\u3002\u56e0\u6b64\uff0cSebastian \u9884\u8ba1\u4ee5\u63a8\u7406\u4e3a\u91cd\u70b9\u7684\u540e\u8bad\u7ec3\u5c06\u6210\u4e3a\u672a\u6765 LLM \u6d41\u7a0b\u7684\u6807\u51c6\u505a\u6cd5\u3002\u00a0\u672c\u6587\u5c06\u63a2\u8ba8\u5f3a\u5316\u5b66\u4e60\u5728\u63a8\u7406\u65b9\u9762\u7684\u6700\u65b0\u8fdb\u5c55\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s9.51cto.com\/oss\/202504\/21\/e201be079774775e4bf380f413a47adadb8d78.webp\" alt=\"image.png\" data-type=\"inline\" \/>\u56fe\u6e90\uff1a\u672c\u6587\u91cd\u70b9\u4ecb\u7ecd\u7528\u4e8e\u5f00\u53d1\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u65b9\u6cd5\u3002<\/p>\n<p>\u672c\u6587\u4e3b\u8981\u5185\u5bb9\u5305\u62ec\u4ee5\u4e0b\u51e0\u90e8\u5206\uff1a<\/p>\n<ul data-id=\"ua73dd4b-m1iFdK1m\">\n<li data-id=\"ld70c578-CKk6nV8o\">\u7406\u89e3\u63a8\u7406\u6a21\u578b\uff1b<\/li>\n<li data-id=\"ld70c578-0hEDZPPW\">RLHF\uff08Reinforcement Learning from Human Feedback\uff09\u57fa\u7840\uff1a\u4e00\u5207\u4ece\u4f55\u800c\u6765\uff1b<\/li>\n<li data-id=\"ld70c578-m78hJP3Z\">PPO\uff08Proximal Policy Optimization\uff09\u7b80\u4ecb\uff1a\u5f3a\u5316\u5b66\u4e60\u7684\u6838\u5fc3\u7b97\u6cd5\uff1b<\/li>\n<li data-id=\"ld70c578-KFVSnfWg\">RL \u7b97\u6cd5\uff1a\u4ece PPO \u5230 GRPO\uff08Generalized Return and Policy Optimization\uff09\uff1b<\/li>\n<li data-id=\"ld70c578-aRWJB0Ym\">RL \u5956\u52b1\u6a21\u578b\uff1a\u4ece RLHF \u5230 RLVR\uff08Reinforcement Learning with Verifiable Rewards\uff09\uff1b<\/li>\n<li data-id=\"ld70c578-oYkF1rsH\">DeepSeek-R1 \u63a8\u7406\u6a21\u578b\u7684\u8bad\u7ec3\u65b9\u6cd5\uff1b<\/li>\n<li data-id=\"ld70c578-kPWH1SMk\">\u4ece\u6700\u8fd1\u5173\u4e8e\u8bad\u7ec3\u63a8\u7406\u6a21\u578b\u7684 RL \u8bba\u6587\u4e2d\u6c72\u53d6\u7684\u6559\u8bad\uff1b<\/li>\n<li data-id=\"ld70c578-lrWuB0MC\">\u503c\u5f97\u5173\u6ce8\u7684\u63a8\u7406\u6a21\u578b\u8bad\u7ec3\u7814\u7a76\u8bba\u6587\u3002<\/li>\n<\/ul>\n<p>\u4e0b\u6587\u4ee5\u4f5c\u8005\u7b2c\u4e00\u4eba\u79f0\u53e3\u543b\u9648\u8ff0\u3002<\/p>\n<h4>\u7406\u89e3\u63a8\u7406\u6a21\u578b<\/h4>\n<p>\u6211\u4eec\u9996\u5148\u6765\u4e86\u89e3\u63a8\u7406\u7684\u5b9a\u4e49\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u63a8\u7406\uff08reasoning\uff09\u662f\u6307\u4f7f LLM \u80fd\u591f\u66f4\u597d\u5730\u5904\u7406\u590d\u6742\u4efb\u52a1\u7684\u63a8\u7406\uff08inference\uff09\u548c\u8bad\u7ec3\u6280\u5de7\u3002\u4e3a\u4e86\u66f4\u8be6\u7ec6\u5730\u89e3\u91ca\u5982\u4f55\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff08\u76ee\u524d\u4e3a\u6b62\uff09\uff0c\u6211\u5b9a\u4e49\u5982\u4e0b\uff1a\u5728 LLM \u7684\u8bed\u5883\u4e2d\uff0c\u63a8\u7406\u662f\u6307\u6a21\u578b\u5728\u63d0\u4f9b\u6700\u7ec8\u7b54\u6848\u4e4b\u524d\u751f\u6210\u4e2d\u95f4\u6b65\u9aa4\u7684\u80fd\u529b\u3002<\/p>\n<p>\u8fd9\u4e2a\u8fc7\u7a0b\u901a\u5e38\u88ab\u79f0\u4e3a\u601d\u7ef4\u94fe (CoT) \u63a8\u7406\u3002\u5728\u601d\u7ef4\u94fe\u63a8\u7406\u4e2d\uff0cLLM \u4f1a\u660e\u786e\u751f\u6210\u4e00\u4e2a\u7ed3\u6784\u5316\u7684\u8bed\u53e5\u6216\u8ba1\u7b97\u5e8f\u5217\uff0c\u4ee5\u8bf4\u660e\u5176\u5982\u4f55\u5f97\u51fa\u7ed3\u8bba\u3002\u5177\u4f53\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202504\/21\/38ae6c2180a359d51da21032147d0e3a8ca98d.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>LLM \u5982\u4f55\u5904\u7406\u591a\u6b65\u9aa4\u63a8\u7406\u4efb\u52a1\u7684\u7b80\u5355\u56fe\u4f8b\u3002\u6a21\u578b\u5e76\u975e\u4ec5\u4ec5\u56de\u5fc6\u4e00\u4e2a\u4e8b\u5b9e\uff0c\u800c\u662f\u9700\u8981\u7ed3\u5408\u591a\u4e2a\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u624d\u80fd\u5f97\u51fa\u6b63\u786e\u7684\u7ed3\u8bba\u3002\u6839\u636e\u5177\u4f53\u5b9e\u73b0\u65b9\u5f0f\uff0c\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u53ef\u80fd\u4f1a\u663e\u793a\u7ed9\u7528\u6237\uff0c\u4e5f\u53ef\u80fd\u4e0d\u4f1a\u663e\u793a\u3002<\/p>\n<p>\u5982\u679c\u4f60\u5bf9\u63a8\u7406\u6a21\u578b\u8fd8\u4e0d\u719f\u6089\uff0c\u5e0c\u671b\u770b\u5230\u66f4\u5168\u9762\u7684\u4ecb\u7ecd\uff0c\u63a8\u8350\u6211\u4e4b\u524d\u7684\u6587\u7ae0\uff1a<\/p>\n<ul data-id=\"ua73dd4b-LEWROM8U\">\n<li data-id=\"ld70c578-pjnBKvNs\">\u6587\u7ae0 1\uff1ahttps:\/\/magazine.sebastianraschka.com\/p\/first-look-at-reasoning-from-scratch<\/li>\n<li data-id=\"ld70c578-Siuc6dXH\">\u6587\u7ae0 2\uff1a\u300aSebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5\u300bhttps:\/\/magazine.sebastianraschka.com\/p\/understanding-reasoning-llms<\/li>\n<\/ul>\n<p>LLM \u7684\u63a8\u7406\u80fd\u529b\u53ef\u4ee5\u901a\u8fc7\u4e24\u79cd\u65b9\u5f0f\u5f97\u5230\u63d0\u9ad8\uff0cOpenAI \u535a\u5ba2\u6587\u7ae0\u4e2d\u7684\u4e00\u5f20\u56fe\u5f88\u597d\u5730\u8bf4\u660e\u4e86\u8fd9\u4e00\u70b9\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s9.51cto.com\/oss\/202504\/21\/445c381487402173ae06090acfd6bef959dd49.webp\" alt=\"image.png\" data-type=\"inline\" \/>\u51c6\u786e\u7387\u7684\u63d0\u5347\u53ef\u4ee5\u901a\u8fc7\u589e\u52a0\u8bad\u7ec3\u6216\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u6765\u5b9e\u73b0\uff0c\u5176\u4e2d\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u4e0e\u63a8\u7406\u65f6\u8ba1\u7b97\u3001\u63a8\u7406\u65f6\u6269\u5c55\u662f\u540c\u4e49\u8bcd\u3002\u56fe\u6e90\uff1ahttps:\/\/openai.com\/index\/learning-to-reason-with-llms\/<\/p>\n<p>\u6211\u6b64\u524d\u4ecb\u7ecd\u8fc7\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u65b9\u6cd5\uff0c\u672c\u6587\u60f3\u6df1\u5165\u63a2\u8ba8\u4e00\u4e0b\u8bad\u7ec3\u65b9\u6cd5\u3002<\/p>\n<h4>RLHF \u57fa\u7840\uff1a\u4e00\u5207\u4ece\u4f55\u800c\u6765<\/h4>\n<p>\u7528\u4e8e\u6784\u5efa\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u65b9\u6cd5\uff0c\u6216\u591a\u6216\u5c11\u4e0e\u7528\u4e8e\u5f00\u53d1\u548c\u5bf9\u9f50\u4f20\u7edf LLM \u7684 RLHF \u65b9\u6cd5\u8bba\u76f8\u5173\u3002\u56e0\u6b64\uff0c\u5728\u8ba8\u8bba\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u7684\u7279\u5b9a\u63a8\u7406\u4fee\u6539\uff08modification\uff09\u4e4b\u524d\uff0c\u6211\u60f3\u5148\u7b80\u5355\u56de\u987e\u4e00\u4e0b RLHF \u7684\u5de5\u4f5c\u539f\u7406\u3002<\/p>\n<p>\u4f20\u7edf LLM \u901a\u5e38\u7ecf\u5386\u4ee5\u4e0b\u4e09\u4e2a\u6b65\u9aa4\u7684\u8bad\u7ec3\u8fc7\u7a0b\uff1a<\/p>\n<ul data-id=\"ua73dd4b-OANAaYBS\">\n<li data-id=\"ld70c578-Se589f7o\">\u9884\u8bad\u7ec3\uff1b<\/li>\n<li data-id=\"ld70c578-BFPgGKT2\">\u76d1\u7763\u5fae\u8c03\uff1b<\/li>\n<li data-id=\"ld70c578-ftSwAhP0\">\u5bf9\u9f50\uff08\u901a\u5e38\u901a\u8fc7 RLHF \u8fdb\u884c\uff09\u3002<\/li>\n<\/ul>\n<p>\u300c\u539f\u59cb\u300d\u7684 LLM \u5bf9\u9f50\u65b9\u6cd5\u662f RLHF\uff0c\u5b83\u662f\u9075\u5faa InstructGPT \u8bba\u6587\u5f00\u53d1 LLM \u65f6\u7684\u6807\u51c6\u65b9\u6cd5\u4e4b\u4e00\uff0c\u8be5\u8bba\u6587\u63cf\u8ff0\u4e86\u7528\u4e8e\u5f00\u53d1\u7b2c\u4e00\u4e2a ChatGPT \u6a21\u578b\u7684\u65b9\u6cd5\u3002<\/p>\n<p>RLHF \u7684\u6700\u521d\u76ee\u6807\u662f\u4f7f LLM \u4e0e\u4eba\u7c7b\u504f\u597d\u4fdd\u6301\u4e00\u81f4\u3002\u4f8b\u5982\uff0c\u5047\u8bbe\u4f60\u591a\u6b21\u4f7f\u7528 LLM\uff0c\u5e76\u4e14 LLM \u4f1a\u9488\u5bf9\u7ed9\u5b9a\u7684\u63d0\u793a\u8bcd\u751f\u6210\u591a\u4e2a\u7b54\u6848\u3002RLHF \u4f1a\u5f15\u5bfc LLM \u751f\u6210\u66f4\u591a\u4f60\u504f\u597d\u7684\u7b54\u6848\u98ce\u683c\u3002RLHF \u901a\u5e38\u4e5f\u7528\u4e8e\u5bf9 LLM \u8fdb\u884c\u5b89\u5168\u8c03\u6574\uff1a\u907f\u514d\u5171\u4eab\u654f\u611f\u4fe1\u606f\u3001\u4f7f\u7528\u810f\u8bdd\u7b49\u7b49\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bb2\uff0cRLHF \u6d41\u7a0b\u91c7\u7528\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u5e76\u4ee5\u76d1\u7763\u65b9\u5f0f\u5bf9\u5176\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u79cd\u5fae\u8c03\u5c1a\u672a\u6210\u4e3a\u5f3a\u5316\u5b66\u4e60\u7684\u4e00\u90e8\u5206\uff0c\u4f46\u5b83\u4e3b\u8981\u662f\u4e00\u79cd\u5148\u51b3\u6761\u4ef6\u3002<\/p>\n<p>\u7136\u540e\uff0cRLHF \u4f7f\u7528\u4e00\u79cd\u79f0\u4e3a\u8fd1\u7aef\u7b56\u7565\u4f18\u5316 (PPO) \u7684\u7b97\u6cd5\u8fdb\u4e00\u6b65\u5bf9\u9f50 LLM\u3002\u8bf7\u6ce8\u610f\uff0c\u9664\u4e86 PPO \u4e4b\u5916\uff0c\u8fd8\u6709\u5176\u4ed6\u7b97\u6cd5\u53ef\u4ee5\u66ff\u4ee3\u3002\u6211\u7279\u610f\u63d0\u5230 PPO\uff0c\u662f\u56e0\u4e3a\u5b83\u662f RLHF \u6700\u521d\u4f7f\u7528\u7684\u7b97\u6cd5\uff0c\u5e76\u4e14\u81f3\u4eca\u4ecd\u662f\u6700\u6d41\u884c\u7684\u7b97\u6cd5\u3002\uff09<\/p>\n<p>\u4e3a\u7b80\u5355\u8d77\u89c1\uff0c\u6211\u4eec\u53ef\u4ee5\u5206\u4e09\u4e2a\u6b65\u9aa4\u6765\u4e86\u89e3 RLHF \u6d41\u7a0b\uff1a<\/p>\n<ul data-id=\"ua73dd4b-g68daebF\">\n<li data-id=\"ld70c578-cmlETvGd\">RLHF \u6b65\u9aa4 1\uff08\u5148\u51b3\u6761\u4ef6\uff09\uff1a\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\uff1b<\/li>\n<li data-id=\"ld70c578-WDRVvGkk\">RLHF \u6b65\u9aa4 2\uff1a\u521b\u5efa\u5956\u52b1\u6a21\u578b\uff08RM\uff09\uff1b<\/li>\n<li data-id=\"ld70c578-YQxxMh5r\">RLHF \u6b65\u9aa4 3\uff1a\u901a\u8fc7 PPO \u8fdb\u884c\u5fae\u8c03\u3002<\/li>\n<\/ul>\n<p>RLHF \u6b65\u9aa4 1\uff08\u5982\u4e0b\u56fe\u6240\u793a\uff09\u662f\u4e00\u4e2a\u76d1\u7763\u5fae\u8c03\u6b65\u9aa4\uff0c\u7528\u4e8e\u521b\u5efa\u57fa\u7840\u6a21\u578b\uff0c\u4ee5\u4fbf\u8fdb\u4e00\u6b65\u8fdb\u884c RLHF \u5fae\u8c03\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202504\/21\/53d03fd38ee9a78ec831330849c79b285be2cb.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u56fe\u6e90\uff1aInstructGPT \u8bba\u6587\u3002arXiv\uff1ahttps:\/\/arxiv.org\/abs\/2203.02155<\/p>\n<p>\u5728 RLHF \u6b65\u9aa4 1 \u4e2d\uff0c\u6211\u4eec\u521b\u5efa\u6216\u91c7\u6837\u63d0\u793a\u8bcd\uff08\u4f8b\u5982\u4ece\u6570\u636e\u5e93\u4e2d\u83b7\u53d6\uff09\uff0c\u5e76\u8981\u6c42\u4eba\u7c7b\u7528\u6237\u64b0\u5199\u9ad8\u8d28\u91cf\u7684\u56de\u590d\u3002\u7136\u540e\uff0c\u6211\u4eec\u4f7f\u7528\u8be5\u6570\u636e\u96c6\u4ee5\u76d1\u7763\u65b9\u5f0f\u5fae\u8c03\u9884\u8bad\u7ec3\u7684\u57fa\u7840\u6a21\u578b\u3002\u5982\u524d\u6240\u8ff0\uff0c\u8fd9\u5728\u6280\u672f\u4e0a\u5e76\u975e RL \u8bad\u7ec3\u7684\u4e00\u90e8\u5206\uff0c\u800c\u4ec5\u4ec5\u662f\u4e00\u4e2a\u5148\u51b3\u6761\u4ef6\u3002<\/p>\n<p>\u5728 RLHF \u6b65\u9aa4 2 \u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u76d1\u7763\u5fae\u8c03\u5f97\u5230\u7684\u6a21\u578b\u6765\u521b\u5efa\u5956\u52b1\u6a21\u578b\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202504\/21\/25235391935794e1b486681d0a1fa93e8372c5.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u63d0\u793a\u8bcd\uff0c\u6211\u4eec\u4f1a\u6839\u636e\u4e0a\u4e00\u6b65\u521b\u5efa\u7684\u5fae\u8c03\u540e\u7684 LLM \u751f\u6210\u56db\u4e2a\u56de\u590d\u3002\u7136\u540e\uff0c\u4eba\u5de5\u6807\u6ce8\u5458\u4f1a\u6839\u636e\u4ed6\u4eec\u7684\u504f\u597d\u5bf9\u8fd9\u4e9b\u56de\u590d\u8fdb\u884c\u6392\u5e8f\u3002\u867d\u7136\u8fd9\u4e2a\u6392\u5e8f\u8fc7\u7a0b\u6bd4\u8f83\u8017\u65f6\uff0c\u4f46\u4e0e\u521b\u5efa\u7528\u4e8e\u76d1\u7763\u5fae\u8c03\u7684\u6570\u636e\u96c6\u76f8\u6bd4\uff0c\u5b83\u53ef\u80fd\u8017\u8d39\u66f4\u5c11\u7684\u4eba\u529b\u3002\u8fd9\u662f\u56e0\u4e3a\u5bf9\u56de\u590d\u8fdb\u884c\u6392\u5e8f\u53ef\u80fd\u6bd4\u7f16\u5199\u56de\u590d\u66f4\u7b80\u5355\u3002<\/p>\n<p>\u5728\u7f16\u8bd1\u5305\u542b\u8fd9\u4e9b\u6392\u5e8f\u7684\u6570\u636e\u96c6\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u8bbe\u8ba1\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u4f1a\u8f93\u51fa\u5956\u52b1\u5206\u6570\uff0c\u7528\u4e8e RLHF \u6b65\u9aa4 3 \u4e2d\u7684\u540e\u7eed\u4f18\u5316\u9636\u6bb5\u3002\u8fd9\u91cc\u7684\u601d\u8def\u662f\uff1a\u5956\u52b1\u6a21\u578b\u53ef\u4ee5\u53d6\u4ee3\u5e76\u81ea\u52a8\u5316\u8017\u8d39\u4eba\u529b\u7684\u4eba\u5de5\u6392\u5e8f\uff0c\u4ece\u800c\u4f7f\u5927\u578b\u6570\u636e\u96c6\u4e0a\u7684\u8bad\u7ec3\u53d8\u5f97\u53ef\u884c\u3002<\/p>\n<p>\u5956\u52b1\u6a21\u578b\u901a\u5e38\u6e90\u81ea\u4e0a\u4e00\u6b65\u76d1\u7763\u5fae\u8c03\u6b65\u9aa4\u4e2d\u521b\u5efa\u7684 LLM\u3002\u4e3a\u4e86\u5c06 RLHF \u6b65\u9aa4 1 \u4e2d\u7684\u6a21\u578b\u8f6c\u6362\u4e3a\u5956\u52b1\u6a21\u578b\uff0c\u5176\u8f93\u51fa\u5c42\uff08\u4e0b\u4e00\u4e2a token \u5206\u7c7b\u5c42\uff09\u5c06\u88ab\u66ff\u6362\u4e3a\u4e00\u4e2a\u5177\u6709\u5355\u4e2a\u8f93\u51fa\u8282\u70b9\u7684\u56de\u5f52\u5c42\u3002<\/p>\n<p>RLHF \u6d41\u7a0b\u7684\u7b2c\u4e09\u6b65\u662f\u4f7f\u7528\u5956\u52b1\u6a21\u578b\u5bf9\u4e4b\u524d\u7684\u76d1\u7763\u5fae\u8c03\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s6.51cto.com\/oss\/202504\/21\/b42b02441c929c26afb81161053da2728f6e1a.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u5728 RLHF \u6b65\u9aa4 3\uff08\u6700\u540e\u9636\u6bb5\uff09\u4e2d\uff0c\u6211\u4eec\u6839\u636e\u5728 RLHF \u6b65\u9aa4 2 \u4e2d\u6240\u521b\u5efa\u5956\u52b1\u6a21\u578b\u7684\u5956\u52b1\u5206\u6570\uff0c\u4f7f\u7528 PPO \u6765\u66f4\u65b0 SFT \u6a21\u578b\u3002<\/p>\n<h4>PPO \u7b80\u4ecb\uff1a\u5f3a\u5316\u5b66\u4e60\u7684\u6838\u5fc3\u7b97\u6cd5<\/h4>\n<p>\u5982\u524d\u6240\u8ff0\uff0c\u539f\u59cb RLHF \u65b9\u6cd5\u4f7f\u7528\u4e86\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5 PPO\u3002PPO \u7684\u5f00\u53d1\u662f\u4e3a\u4e86\u63d0\u9ad8\u7b56\u7565\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u548c\u6548\u7387\u3002\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u7b56\u7565\u4ec5\u6307\u6211\u4eec\u60f3\u8981\u8bad\u7ec3\u7684\u6a21\u578b\uff1b\u5728\u672c\u4f8b\u4e2d\uff0c\u7b56\u7565 = LLM\u3002<\/p>\n<p>PPO \u80cc\u540e\u7684\u4e00\u4e2a\u5173\u952e\u601d\u60f3\u662f\uff1a\u5b83\u9650\u5236\u4e86\u7b56\u7565\u5728\u6bcf\u4e2a\u66f4\u65b0\u6b65\u9aa4\u4e2d\u5141\u8bb8\u7684\u66f4\u6539\u91cf\u3002\u8fd9\u662f\u901a\u8fc7\u622a\u65ad\u635f\u5931\u51fd\u6570\u6765\u5b9e\u73b0\u7684\uff0c\u6709\u52a9\u4e8e\u9632\u6b62\u6a21\u578b\u8fdb\u884c\u8fc7\u5927\u7684\u66f4\u65b0\uff0c\u4ece\u800c\u907f\u514d\u7834\u574f\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u6b64\u5916\uff0cPPO \u8fd8\u5728\u635f\u5931\u51fd\u6570\u4e2d\u5305\u542b\u4e86 KL \u6563\u5ea6\u60e9\u7f5a\u3002\u8be5\u60e9\u7f5a\u9879\u5c06\u5f53\u524d\u7b56\u7565\uff08\u6b63\u5728\u8bad\u7ec3\u7684\u6a21\u578b\uff09\u4e0e\u539f\u59cb SFT \u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\uff0c\u9f13\u52b1\u66f4\u65b0\u7ed3\u679c\u4fdd\u6301\u5408\u7406\u7684\u63a5\u8fd1\u3002\u6bd5\u7adf\uff0c\u8fd9\u6837\u505a\u7684\u76ee\u7684\u662f\u6839\u636e\u504f\u597d\u8c03\u6574\u6a21\u578b\uff0c\u800c\u4e0d\u662f\u5b8c\u5168\u91cd\u65b0\u8bad\u7ec3\u3002<\/p>\n<p>\u8fd9\u5c31\u662f PPO \u4e2d\u300cproximal\u300d\uff08\u8fd1\u7aef\uff09\u4e00\u8bcd\u7684\u7531\u6765\uff1a\u8be5\u7b97\u6cd5\u8bd5\u56fe\u4f7f\u66f4\u65b0\u7ed3\u679c\u63a5\u8fd1\u73b0\u6709\u6a21\u578b\uff0c\u540c\u65f6\u4ecd\u5141\u8bb8\u6539\u8fdb\u3002\u4e3a\u9f13\u52b1\u63a2\u7d22\uff0cPPO \u8fd8\u6dfb\u52a0\u4e86\u71b5\u5956\u52b1\uff0c\u4ece\u800c\u9f13\u52b1\u6a21\u578b\u5728\u8bad\u7ec3\u671f\u95f4\u6539\u53d8\u8f93\u51fa\u3002<\/p>\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u60f3\u4ecb\u7ecd\u4e00\u4e9b\u672f\u8bed\uff0c\u4ee5\u4fbf\u5728\u76f8\u5bf9\u8f83\u9ad8\u7684\u5c42\u6b21\u4e0a\u89e3\u91ca PPO\u3002\u8fd9\u91cc\u6d89\u53ca\u5f88\u591a\u4e13\u4e1a\u672f\u8bed\uff0c\u56e0\u6b64\u5728\u7ee7\u7eed\u4e4b\u524d\uff0c\u6211\u5c1d\u8bd5\u5728\u4e0b\u56fe\u4e2d\u603b\u7ed3\u4e00\u4e0b\u5173\u952e\u672f\u8bed\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202504\/21\/c6cb8cf40a59d1b2b360760a3fab24a59e09e6.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>RLHF \u4e2d\u5173\u952e\u672f\u8bed\u7684\u8bf4\u660e\u3002<\/p>\n<p>\u4e0b\u9762\u6211\u5c06\u901a\u8fc7\u4f2a\u4ee3\u7801\u6765\u8bf4\u660e PPO \u4e2d\u7684\u5173\u952e\u6b65\u9aa4\u3002\u6b64\u5916\uff0c\u4e3a\u4e86\u66f4\u76f4\u89c2\u5730\u8bf4\u660e\uff0c\u6211\u8fd8\u5c06\u4f7f\u7528\u4e00\u4e2a\u6bd4\u55bb\uff1a\u5047\u8bbe\u4f60\u662f\u4e00\u4f4d\u7ecf\u8425\u5c0f\u578b\u5916\u5356\u670d\u52a1\u7684\u53a8\u5e08\u3002\u4f60\u4e0d\u65ad\u5c1d\u8bd5\u65b0\u7684\u83dc\u8c31\u53d8\u5316\u4ee5\u63d0\u9ad8\u5ba2\u6237\u6ee1\u610f\u5ea6\u3002\u4f60\u7684\u603b\u4f53\u76ee\u6807\u662f\u6839\u636e\u5ba2\u6237\u53cd\u9988\uff08\u5956\u52b1\uff09\u8c03\u6574\u83dc\u8c31\uff08\u7b56\u7565\uff09\u3002<\/p>\n<p>1\u3001\u8ba1\u7b97\u65b0\u65e7\u7b56\u7565\u4e2d\u4e0b\u4e00\u4e2a token \u6982\u7387\u7684\u6bd4\u7387\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_0\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">ratio <span class=\"token operator\">=<\/span> new_policy_prob <span class=\"token operator\">\/<\/span>old_policy_prob<\/code><\/pre>\n<ul id=\"code_id_0\" class=\"pre-numbering\">\n<li>1.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u7b80\u800c\u8a00\u4e4b\uff0c\u8fd9\u4f1a\u68c0\u67e5\u65b0\u65e7\u7b56\u7565\u7684\u5dee\u5f02\u3002 \u9644\u6ce8\uff1a\u5173\u4e8e\u300cnew_policy_prob\u300d\uff0c\u6211\u4eec\u5c1a\u672a\u4f7f\u7528\u6700\u7ec8\u66f4\u65b0\u540e\u7684\u7b56\u7565\u3002\u6211\u4eec\u4f7f\u7528\u7684\u662f\u5f53\u524d\u7684\u7b56\u7565\u7248\u672c\uff08\u5373\u6211\u4eec\u6b63\u5728\u8bad\u7ec3\u7684\u6a21\u578b\uff09\u3002\u4e0d\u8fc7\uff0c\u6309\u7167\u60ef\u4f8b\uff0c\u6211\u4eec\u5c06\u5176\u79f0\u4e3a\u300c\u65b0\u300d\u3002\u56e0\u6b64\uff0c\u5373\u4f7f\u4f60\u4ecd\u5728\u8fdb\u884c\u5b9e\u9a8c\uff0c\u6211\u4eec\u4e5f\u4f1a\u6309\u7167\u60ef\u4f8b\u5c06\u4f60\u5f53\u524d\u7684\u8349\u7a3f\u79f0\u4e3a\u300c\u65b0\u7b56\u7565\u300d\u3002<\/p>\n<p>2\u3001\u5c06\u8be5\u6bd4\u7387\u4e58\u4ee5\u884c\u52a8\uff08action\uff09\u7684\u4f18\u52a3\u7a0b\u5ea6\uff08\u79f0\u4e3a\u4f18\u52bf\uff09\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_1\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">raw_score <span class=\"token operator\">=<\/span> ratio <span class=\"token operator\">*<\/span> advantage<\/code><\/pre>\n<ul id=\"code_id_1\" class=\"pre-numbering\">\n<li>1.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u8fd9\u91cc\uff0c\u4e3a\u4e86\u7b80\u5355\u8d77\u89c1\uff0c\u6211\u4eec\u53ef\u4ee5\u5047\u8bbe\u4f18\u52bf\u662f\u6839\u636e\u5956\u52b1\u4fe1\u53f7\u8ba1\u7b97\u7684\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_2\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">advantage <span class=\"token operator\">=<\/span> actual_reward <span class=\"token operator\">-<\/span> expected_reward<\/code><\/pre>\n<ul id=\"code_id_2\" class=\"pre-numbering\">\n<li>1.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u5728\u53a8\u5e08\u7684\u7c7b\u6bd4\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u4f18\u52bf\u89c6\u4e3a\u65b0\u83dc\u7684\u8868\u73b0\u5982\u4f55\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_3\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">advantage <span class=\"token operator\">=<\/span> customer_rating <span class=\"token operator\">-<\/span> expected_rating<\/code><\/pre>\n<ul id=\"code_id_3\" class=\"pre-numbering\">\n<li>1.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u4f8b\u5982\uff0c\u5982\u679c\u4e00\u4f4d\u987e\u5ba2\u7ed9\u65b0\u83dc\u54c1\u6253\u4e86 9 \u5206\uff08\u6ee1\u5206 10 \u5206\uff09\uff0c\u800c\u987e\u5ba2\u901a\u5e38\u7ed9\u6211\u4eec 7 \u5206\uff08\u6ee1\u5206 10 \u5206\uff09\uff0c\u90a3\u4e48\u8fd9\u5c31\u662f +2 \u7684\u4f18\u52bf\u3002<\/p>\n<p>\u9700\u8981\u6ce8\u610f\uff0c\u8fd9\u53ea\u662f\u4e00\u4e2a\u7b80\u5316\u7684\u63cf\u8ff0\u3002\u5b9e\u9645\u4e0a\uff0c\u8fd9\u6d89\u53ca\u5230\u5e7f\u4e49\u4f18\u52bf\u4f30\u8ba1\uff08Generalized advantage estimation\uff0cGAE\uff09\uff0c\u8fd9\u91cc\u4e0d\u518d\u8d58\u8ff0\u3002\u7136\u800c\uff0c\u9700\u8981\u63d0\u53ca\u7684\u4e00\u4e2a\u91cd\u8981\u7ec6\u8282\u662f\uff1a\u9884\u671f\u5956\u52b1\u662f\u7531\u6240\u8c13\u7684\u300c\u8bc4\u4ef7\u8005\u300d\uff08\u6709\u65f6\u4e5f\u79f0\u4e3a\u4ef7\u503c\u6a21\u578b\uff09\u8ba1\u7b97\u7684\uff0c\u800c\u5956\u52b1\u6a21\u578b\u5219\u8ba1\u7b97\u5b9e\u9645\u5956\u52b1\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u4f18\u52bf\u8ba1\u7b97\u6d89\u53ca\u53e6\u5916\u4e24\u4e2a\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u7684\u89c4\u6a21\u901a\u5e38\u4e0e\u6211\u4eec\u6b63\u5728\u5fae\u8c03\u7684\u539f\u59cb\u6a21\u578b\u76f8\u540c\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u6211\u4eec\u53ef\u4ee5\u628a\u8fd9\u4e2a\u8bc4\u4ef7\u8005\u6216\u4ef7\u503c\u6a21\u578b\u60f3\u8c61\u6210\u4e00\u4f4d\u670b\u53cb\uff0c\u5728\u5c06\u65b0\u83dc\u54c1\u7aef\u4e0a\u684c\u4e4b\u524d\uff0c\u6211\u4eec\u4f1a\u9080\u8bf7\u4ed6\u54c1\u5c1d\u3002\u6211\u4eec\u8fd8\u4f1a\u8bf7\u8fd9\u4f4d\u670b\u53cb\u4f30\u8ba1\u987e\u5ba2\u4f1a\u5982\u4f55\u8bc4\u4ef7\u8fd9\u9053\u83dc\uff08\u8fd9\u5c31\u662f\u9884\u671f\u5956\u52b1\uff09\u3002\u5956\u52b1\u6a21\u578b\u5c31\u662f\u7ed9\u51fa\u53cd\u9988\uff08\u5373\u5b9e\u9645\u5956\u52b1\uff09\u7684\u5b9e\u9645\u987e\u5ba2\u3002<\/p>\n<p>3\u3001\u8ba1\u7b97\u622a\u65ad\u5206\u6570\uff1a<\/p>\n<p>\u5982\u679c\u65b0\u7b56\u7565\u53d8\u5316\u8fc7\u5927\uff08\u4f8b\u5982\u6bd4\u4f8b &gt; 1.2 \u6216 &lt; 0.8\uff09\uff0c\u6211\u4eec\u4f1a\u6309\u5982\u4e0b\u65b9\u5f0f\u622a\u65ad\u8be5\u6bd4\u4f8b\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_4\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">clipped_ratio <span class=\"token operator\">=<\/span> <span class=\"token function\">clamp<\/span> <span class=\"token punctuation\">(<\/span>ratio<span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.8<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">1.2<\/span><span class=\"token punctuation\">)<\/span>\r\nclipped_score <span class=\"token operator\">=<\/span> clipped_ratio <span class=\"token operator\">*<\/span> advantage<\/code><\/pre>\n<ul id=\"code_id_4\" class=\"pre-numbering\">\n<li>1.<\/li>\n<li>2.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5047\u8bbe\u65b0\u83dc\u8c31\u7684\u8bc4\u4ef7\u7279\u522b\u597d\uff08\u6216\u7279\u522b\u5dee\uff09\uff0c\u6211\u4eec\u53ef\u80fd\u4f1a\u5fcd\u4e0d\u4f4f\u9a6c\u4e0a\u5c31\u5f7b\u5e95\u4fee\u6539\u6574\u4e2a\u83dc\u5355\u3002\u4f46\u8fd9\u6837\u505a\u98ce\u9669\u5f88\u5927\uff0c\u56e0\u6b64\u6211\u4eec\u6682\u65f6\u9650\u5236\u83dc\u8c31\u53ef\u4ee5\u4fee\u6539\u7684\u8303\u56f4\u3002\u6bd4\u5982\uff0c\u6211\u4eec\u53ef\u80fd\u628a\u83dc\u505a\u5f97\u66f4\u8fa3\u4e86\uff0c\u800c\u90a3\u4f4d\u987e\u5ba2\u78b0\u5de7\u559c\u6b22\u5403\u8fa3\uff0c\u4f46\u8fd9\u5e76\u4e0d\u610f\u5473\u7740\u5176\u4ed6\u4eba\u4e5f\u559c\u6b22\u3002<\/p>\n<p>4\u3001\u7136\u540e\uff0c\u6211\u4eec\u53d6\u539f\u59cb\u5f97\u5206\u548c\u88c1\u526a\u5f97\u5206\u4e2d\u8f83\u5c0f\u7684\u4e00\u4e2a\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_5\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\"><span class=\"token keyword\">if<\/span> advantage <span class=\"token operator\">&gt;=<\/span> <span class=\"token number\">0<\/span><span class=\"token operator\">:<\/span>\r\n    final_score <span class=\"token operator\">=<\/span> <span class=\"token function\">min<\/span> <span class=\"token punctuation\">(<\/span>raw_score<span class=\"token punctuation\">,<\/span> clipped_score<span class=\"token punctuation\">)<\/span>\r\n<span class=\"token keyword\">else<\/span><span class=\"token operator\">:<\/span>\r\n    final_score <span class=\"token operator\">=<\/span> <span class=\"token function\">max<\/span> <span class=\"token punctuation\">(<\/span>raw_score<span class=\"token punctuation\">,<\/span> clipped_score<span class=\"token punctuation\">)<\/span><\/code><\/pre>\n<ul id=\"code_id_5\" class=\"pre-numbering\">\n<li>1.<\/li>\n<li>2.<\/li>\n<li>3.<\/li>\n<li>4.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u8fd9\u540c\u6837\u4e0e\u4fdd\u6301\u8c28\u614e\u6709\u5173\u3002\u4f8b\u5982\uff0c\u5982\u679c\u4f18\u52bf\u662f\u6b63\u7684\uff08\u65b0\u884c\u4e3a\u66f4\u597d\uff09\uff0c\u6211\u4eec\u9650\u5236\u5956\u52b1\u3002\u8fd9\u662f\u56e0\u4e3a\u6211\u4eec\u4e0d\u60f3\u8fc7\u5206\u4fe1\u4efb\u4e00\u4e2a\u53ef\u80fd\u53ea\u662f\u5de7\u5408\u6216\u8fd0\u6c14\u7684\u597d\u7ed3\u679c\u3002<\/p>\n<p>\u5982\u679c\u4f18\u52bf\u662f\u8d1f\u7684\uff08\u65b0\u884c\u4e3a\u66f4\u5dee\uff09\uff0c\u6211\u4eec\u9650\u5236\u60e9\u7f5a\u3002\u5176\u7406\u5ff5\u662f\u76f8\u4f3c\u7684\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u9664\u975e\u6211\u4eec\u975e\u5e38\u786e\u5b9a\uff0c\u5426\u5219\u6211\u4eec\u4e0d\u60f3\u5bf9\u4e00\u4e2a\u4e0d\u597d\u7684\u7ed3\u679c\u53cd\u5e94\u8fc7\u5ea6\u3002<\/p>\n<p>\u7b80\u800c\u8a00\u4e4b\uff0c\u5982\u679c\u4f18\u52bf\u662f\u6b63\u7684\uff0c\u6211\u4eec\u53d6\u4e24\u4e2a\u5f97\u5206\u4e2d\u8f83\u5c0f\u7684\u4e00\u4e2a\uff08\u4ee5\u907f\u514d\u8fc7\u5ea6\u5956\u52b1\uff09\uff0c\u800c\u5982\u679c\u4f18\u52bf\u662f\u8d1f\u7684\uff0c\u5219\u53d6\u8f83\u5927\u7684\u4e00\u4e2a\uff08\u4ee5\u907f\u514d\u8fc7\u5ea6\u60e9\u7f5a\uff09\u3002<\/p>\n<p>\u5728\u8fd9\u4e2a\u6bd4\u55bb\u4e2d\uff0c\u8fd9\u53ef\u4ee5\u786e\u4fdd\u5982\u679c\u4e00\u4e2a\u83dc\u8c31\u7684\u8868\u73b0\u597d\u4e8e\u9884\u671f\uff0c\u6211\u4eec\u4e0d\u4f1a\u5728\u4e0d\u81ea\u4fe1\u7684\u60c5\u51b5\u4e0b\u8fc7\u5ea6\u5956\u52b1\u5b83\u3002\u5982\u679c\u8868\u73b0\u4e0d\u4f73\uff0c\u9664\u975e\u5b83\u4e00\u76f4\u4e0d\u597d\uff0c\u5426\u5219\u6211\u4eec\u4e5f\u4e0d\u4f1a\u8fc7\u5ea6\u60e9\u7f5a\u5b83\u3002<\/p>\n<p>5\u3001\u8ba1\u7b97\u635f\u5931\uff1a<\/p>\n<p>\u8fd9\u4e2a\u6700\u7ec8\u5f97\u5206\u5c31\u662f\u6211\u4eec\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u8981\u6700\u5927\u5316\u7684\uff08\u901a\u8fc7\u68af\u5ea6\u4e0b\u964d\u6765\u6700\u5c0f\u5316\u8fd9\u4e2a\u5f97\u5206\u7684\u7b26\u53f7\u53cd\u8f6c\u503c\uff09\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u6dfb\u52a0\u4e86\u4e00\u4e2a KL \u60e9\u7f5a\u9879\uff0c\u5176\u4e2d \u03b2 \u662f\u60e9\u7f5a\u5f3a\u5ea6\u7684\u8d85\u53c2\u6570\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_6\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">loss <span class=\"token operator\">=<\/span> <span class=\"token operator\">-<\/span>final_score <span class=\"token operator\">+<\/span> \u03b2 <span class=\"token operator\">*<\/span> <span class=\"token constant\">KL<\/span> <span class=\"token punctuation\">(<\/span>new_policy <span class=\"token operator\">||<\/span> reference_policy<span class=\"token punctuation\">)<\/span><\/code><\/pre>\n<ul id=\"code_id_6\" class=\"pre-numbering\">\n<li>1.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u5728\u8fd9\u4e2a\u6bd4\u55bb\u4e2d\uff0c\u6211\u4eec\u6dfb\u52a0\u60e9\u7f5a\u9879\u662f\u4e3a\u4e86\u786e\u4fdd\u65b0\u83dc\u8c31\u4e0e\u6211\u4eec\u539f\u6765\u7684\u98ce\u683c\u4e0d\u4f1a\u5dee\u522b\u592a\u5927\u3002\u8fd9\u53ef\u4ee5\u9632\u6b62\u4f60\u6bcf\u5468\u90fd\u300c\u5f7b\u5e95\u6539\u9020\u53a8\u623f\u300d\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u4e0d\u5e0c\u671b\u7a81\u7136\u628a\u4e00\u5bb6\u610f\u5927\u5229\u9910\u5385\u6539\u9020\u6210\u70e7\u70e4\u5e97\u3002<\/p>\n<p>\u8fd9\u4e9b\u4fe1\u606f\u91cf\u5f88\u5927\uff0c\u56e0\u6b64\u6211\u901a\u8fc7\u4e0b\u9762\u7684\u56fe\uff0c\u7528\u4e00\u4e2a\u5177\u4f53\u7684\u3001\u6570\u503c\u5316\u7684\u4f8b\u5b50\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u80cc\u666f\u4e0b\u8fdb\u884c\u4e86\u603b\u7ed3\u3002\u4f46\u5982\u679c\u8fd9\u592a\u590d\u6742\u4e86\uff0c\u4f60\u53ef\u4ee5\u76f4\u63a5\u8df3\u8fc7\uff1b\u5373\u4f7f\u4e0d\u770b\uff0c\u4e5f\u4e0d\u4f1a\u5f71\u54cd\u4f60\u7406\u89e3\u6587\u7ae0\u7684\u5176\u4ed6\u90e8\u5206\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s6.51cto.com\/oss\/202504\/21\/68cc25b6983b656c799660f384db49acadc8d2.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u6211\u627f\u8ba4\u6211\u5728\u4ecb\u7ecd PPO \u65f6\u53ef\u80fd\u6709\u70b9\u8fc7\u4e8e\u8be6\u7ec6\u4e86\u3002\u4f46\u65e2\u7136\u5df2\u7ecf\u5199\u51fa\u6765\u4e86\uff0c\u5220\u6389\u4e5f\u633a\u96be\u7684\u3002\u5e0c\u671b\u4f60\u4eec\u4e2d\u7684\u4e00\u4e9b\u4eba\u4f1a\u89c9\u5f97\u5b83\u6709\u7528\uff01<\/p>\n<p>\u8bdd\u867d\u5982\u6b64\uff0c\u4e0b\u4e00\u8282\u4e2d\u76f8\u5173\u7684\u8981\u70b9\u662f PPO \u4e2d\u6d89\u53ca\u591a\u4e2a\u6a21\u578b\uff1a<\/p>\n<p>1. \u7b56\u7565\u6a21\u578b\uff1a\u8fd9\u662f\u6211\u4eec\u5e0c\u671b\u901a\u8fc7\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u8fdb\u884c\u8bad\u7ec3\u5e76\u8fdb\u4e00\u6b65\u5bf9\u9f50\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u3002<\/p>\n<p>2. \u5956\u52b1\u6a21\u578b\uff1a\u8fd9\u662f\u4e00\u4e2a\u7ecf\u8fc7\u8bad\u7ec3\u4ee5\u9884\u6d4b\u5956\u52b1\u7684\u6a21\u578b\uff08\u53c2\u89c1\u5f3a\u5316\u5b66\u4e60\u4e0e\u4eba\u7c7b\u53cd\u9988\uff08RLHF\uff09\u7684\u7b2c 2 \u6b65\uff09\u3002<\/p>\n<p>3. \u8bc4\u8bba\u5bb6\u6a21\u578b\uff1a\u8fd9\u662f\u4e00\u4e2a\u53ef\u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u7528\u4e8e\u4f30\u8ba1\u5956\u52b1\u3002<\/p>\n<p>4. \u53c2\u8003\u6a21\u578b\uff08\u539f\u59cb\u7b56\u7565\uff09\uff1a\u6211\u4eec\u4f7f\u7528\u5b83\u6765\u786e\u4fdd\u7b56\u7565\u4e0d\u4f1a\u504f\u79bb\u592a\u8fdc\u3002<\/p>\n<p>\u987a\u4fbf\u8bf4\u4e00\u4e0b\uff0c\u4f60\u53ef\u80fd\u4f1a\u597d\u5947\u4e3a\u4ec0\u4e48\u6211\u4eec\u9700\u8981\u5956\u52b1\u6a21\u578b\u548c\u8bc4\u8bba\u5bb6\u6a21\u578b\u3002\u5956\u52b1\u6a21\u578b\u901a\u5e38\u662f\u5728\u4f7f\u7528 PPO \u8bad\u7ec3\u7b56\u7565\u4e4b\u524d\u8bad\u7ec3\u7684\u3002\u5b83\u901a\u8fc7\u81ea\u52a8\u5316\u4eba\u7c7b\u88c1\u5224\u7684\u504f\u597d\u6807\u8bb0\u6765\u7ed9\u7b56\u7565 LLM \u751f\u6210\u7684\u5b8c\u6574\u54cd\u5e94\u6253\u5206\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u8bc4\u8bba\u5bb6\u6a21\u578b\u5219\u7528\u4e8e\u8bc4\u5224\u90e8\u5206\u54cd\u5e94\u3002\u6211\u4eec\u7528\u5b83\u6765\u521b\u5efa\u6700\u7ec8\u54cd\u5e94\u3002\u867d\u7136\u5956\u52b1\u6a21\u578b\u901a\u5e38\u4fdd\u6301\u51bb\u7ed3\u72b6\u6001\uff0c\u4f46\u8bc4\u8bba\u5bb6\u6a21\u578b\u5728\u8bad\u7ec3\u671f\u95f4\u4f1a\u4e0d\u65ad\u66f4\u65b0\uff0c\u4ee5\u66f4\u597d\u5730\u4f30\u8ba1\u5956\u52b1\u6a21\u578b\u6240\u521b\u5efa\u7684\u5956\u52b1\u3002<\/p>\n<p>PPO \u7684\u66f4\u591a\u7ec6\u8282\u8d85\u51fa\u4e86\u672c\u6587\u7684\u8303\u56f4\uff0c\u4f46\u6709\u5174\u8da3\u7684\u8bfb\u8005\u53ef\u4ee5\u5728\u4ee5\u4e0b\u56db\u7bc7\u65e9\u4e8e InstructGPT \u8bba\u6587\u7684\u6587\u7ae0\u4e2d\u627e\u5230\u6570\u5b66\u7ec6\u8282\uff1a<\/p>\n<p>1. \u300a\u5f02\u6b65\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff082016\uff09\u300b\uff1a\u7531 Mnih\u3001Badia\u3001Mirza\u3001Graves\u3001Lillicrap\u3001Harley\u3001Silver \u548c Kavukcuoglu \u64b0\u5199\uff0c\u4ecb\u7ecd\u4e86\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\uff0c\u4f5c\u4e3a\u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u7684\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u4e2d Q \u5b66\u4e60\u7684\u66ff\u4ee3\u65b9\u6848\u3002<\/p>\n<p>2. \u300a\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\uff082017\uff09\u300b\uff1a\u7531 Schulman\u3001Wolski\u3001Dhariwal\u3001Radford \u548c Klimov \u64b0\u5199\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u6539\u8fdb\u7684\u8fd1\u7aef\u7b56\u7565\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u6bd4\u666e\u901a\u7684\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\u66f4\u5177\u6570\u636e\u6548\u7387\u548c\u53ef\u6269\u5c55\u6027\u3002<\/p>\n<p>3. \u300a\u4ece\u4eba\u7c7b\u504f\u597d\u5fae\u8c03\u8bed\u8a00\u6a21\u578b\uff082020\uff09\u300b\uff1a\u7531 Ziegler\u3001Stiennon\u3001Wu\u3001Brown\u3001Radford\u3001Amodei\u3001Christiano \u548c Irving \u64b0\u5199\uff0c\u9610\u8ff0\u4e86 PPO \u548c\u5956\u52b1\u5b66\u4e60\u7684\u6982\u5ff5\uff0c\u5e76\u5c06\u5176\u5e94\u7528\u4e8e\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c\u5305\u62ec KL \u6b63\u5219\u5316\uff0c\u4ee5\u9632\u6b62\u7b56\u7565\u4e0e\u81ea\u7136\u8bed\u8a00\u504f\u5dee\u8fc7\u5927\u3002<\/p>\n<p>4. \u300a\u4ece\u4eba\u7c7b\u53cd\u9988\u5b66\u4e60\u603b\u7ed3\uff082022\uff09\u300b\uff1a\u7531 Stiennon\u3001Ouyang\u3001Wu\u3001Ziegler\u3001Lowe\u3001Voss\u3001Radford\u3001Amodei \u548c Christiano \u64b0\u5199\uff0c\u4ecb\u7ecd\u4e86\u540e\u6765\u5728 InstructGPT \u8bba\u6587\u4e2d\u4e5f\u4f7f\u7528\u7684\u6d41\u884c\u7684 RLHF \u4e09\u6b65\u6d41\u7a0b\u3002<\/p>\n<h4>RL \u7b97\u6cd5\uff1a\u4ece PPO \u5230 GRPO<\/h4>\n<p>\u5982\u524d\u6240\u8ff0\uff0cPPO \u662f RLHF \u6700\u521d\u4f7f\u7528\u7684\u7b97\u6cd5\u3002\u4ece\u6280\u672f\u89d2\u5ea6\u770b\uff0c\u5b83\u5728\u7528\u4e8e\u5f00\u53d1\u63a8\u7406\u6a21\u578b\u7684 RL pipeline \u4e2d\u8fd0\u884c\u5f97\u975e\u5e38\u597d\u3002\u4e0d\u8fc7\uff0cDeepSeek-R1 \u5728\u4ed6\u4eec\u7684 RL pipeline \u4e2d\u4f7f\u7528\u7684\u662f\u4e00\u79cd\u540d\u4e3a\u300c\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09\u300d\u7684\u7b97\u6cd5\uff0c\u8be5\u7b97\u6cd5\u5728\u4ed6\u4eec\u65e9\u671f\u7684\u4e00\u7bc7\u8bba\u6587\u4e2d\u5df2\u6709\u4ecb\u7ecd\uff1a<\/p>\n<p>\u300aDeepSeekMath\uff1a \u7a81\u7834\u5f00\u653e\u8bed\u8a00\u6a21\u578b\u4e2d\u6570\u5b66\u63a8\u7406\u7684\u6781\u9650\u300b\uff082024 \u5e74\uff09https:\/\/arxiv.org\/abs\/2402.03300<\/p>\n<blockquote><p>DeepSeek \u56e2\u961f\u5f15\u5165\u4e86 GRPO\uff0c\u8fd9\u662f\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09\u7b97\u6cd5\u7684\u4e00\u4e2a\u53d8\u4f53\uff0c\u65e8\u5728\u589e\u5f3a\u6570\u5b66\u63a8\u7406\u80fd\u529b\uff0c\u540c\u65f6\u4f18\u5316 PPO \u7684\u5185\u5b58\u4f7f\u7528\u3002<\/p><\/blockquote>\n<p>\u56e0\u6b64\uff0c\u8fd9\u91cc\u7684\u4e3b\u8981\u7814\u7a76\u52a8\u673a\u662f\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u3002\u6548\u7387\u7684\u63d0\u5347\u662f\u901a\u8fc7\u653e\u5f03\u300c\u6279\u8bc4\u5bb6\u300d\uff08\u4ef7\u503c\u6a21\u578b\uff09\u5373\u8ba1\u7b97\u4ef7\u503c\u51fd\u6570\uff08\u9884\u671f\u672a\u6765\u5956\u52b1\uff09\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u6765\u5b9e\u73b0\u7684\u3002\u4e0e\u5176\u4f9d\u8d56\u8fd9\u4e2a\u989d\u5916\u7684\u6a21\u578b\u6765\u8ba1\u7b97\u4f30\u8ba1\u5956\u52b1\u4ee5\u786e\u5b9a\u4f18\u52bf\uff0cGRPO \u91c7\u53d6\u4e86\u4e00\u79cd\u66f4\u7b80\u5355\u7684\u65b9\u6cd5\uff1a\u5b83\u4ece\u7b56\u7565\u6a21\u578b\u672c\u8eab\u91c7\u6837\u591a\u4e2a\u7b54\u6848\uff0c\u5e76\u4f7f\u7528\u5b83\u4eec\u7684\u76f8\u5bf9\u8d28\u91cf\u6765\u8ba1\u7b97\u4f18\u52bf\u3002<\/p>\n<p>\u4e3a\u4e86\u8bf4\u660e PPO \u548c GRPO \u4e4b\u95f4\u7684\u5dee\u5f02\uff0c\u4ece DeepSeekMath \u8bba\u6587\u4e2d\u501f\u7528\u4e86\u4e00\u5f20\u5f88\u6709\u7528\u7684\u56fe\u8868\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202504\/21\/e61a2fd50cc498980882757043f73b09e1dbee.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<h4>RL \u5956\u52b1\u5efa\u6a21\uff1a\u4ece RLHF \u5230 RLVR<\/h4>\n<p>\u5230\u76ee\u524d\u4e3a\u6b62\uff0c\u6211\u4eec\u5c06 RLHF \u89c6\u4e3a\u4e00\u79cd\u7a0b\u5e8f\uff0c\u5e76\u4ecb\u7ecd\u4e86\u4e24\u79cd\u5e38\u7528\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff1aPPO \u548c GRPO\u3002<\/p>\n<p>\u4f46\u662f\uff0c\u5982\u679c RLHF \u5df2\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5bf9\u9f50\u5de5\u5177\u5305\u7684\u6838\u5fc3\u90e8\u5206\uff0c\u90a3\u4e48\u8fd9\u4e0e\u63a8\u7406\u6709\u4f55\u5173\u8054\u5462\uff1f<\/p>\n<p>DeepSeek \u56e2\u961f\u5c06\u7c7b\u4f3c\u7684\u57fa\u4e8e RL \u7684\u65b9\u6cd5\uff08\u642d\u914d GRPO\uff09\u5e94\u7528\u4e8e\u8bad\u7ec3\u5176 R1 \u548c R1-Zero \u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u7531\u6b64\u5efa\u7acb\u4e86 RLHF \u4e0e\u63a8\u7406\u4e4b\u95f4\u7684\u8054\u7cfb\u3002<\/p>\n<p>\u4e0d\u540c\u4e4b\u5904\u5728\u4e8e\uff0c\u4e0e\u4f9d\u8d56\u4eba\u7c7b\u504f\u597d\u5e76\u8bad\u7ec3\u5956\u52b1\u6a21\u578b\u4e0d\u540c\uff0cDeepSeek-R1 \u56e2\u961f\u91c7\u7528\u4e86\u53ef\u9a8c\u8bc1\u5956\u52b1\u3002\u8fd9\u79cd\u65b9\u6cd5\u88ab\u79f0\u4e3a\u5e26\u6709\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLVR\uff09\u3002<\/p>\n<p>\u9700\u8981\u518d\u6b21\u5f3a\u8c03\u7684\u662f\uff1a\u4e0e\u6807\u51c6\u7684 RLHF \u76f8\u6bd4\uff0cRLVR \u89c4\u907f\u4e86\u5bf9\u5956\u52b1\u6a21\u578b\u7684\u9700\u6c42\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u6a21\u578b\u5e76\u975e\u4ece\u4eba\u7c7b\u6807\u6ce8\u6837\u672c\u4e2d\u5b66\u4e60\u4f55\u4e3a\u300c\u4f18\u8d28\u300d\u7b54\u6848\uff0c\u800c\u662f\u4ece\u786e\u5b9a\u6027\u5de5\u5177\uff08\u5982\u7b26\u53f7\u9a8c\u8bc1\u5668\u6216\u57fa\u4e8e\u89c4\u5219\u7684\u5de5\u5177\uff09\u5904\u83b7\u53d6\u76f4\u63a5\u7684\u4e8c\u5143\u53cd\u9988\uff08\u6b63\u786e\u6216\u9519\u8bef\uff09\u3002\u53ef\u5c06\u4e4b\u7406\u89e3\u4e3a\u7528\u8ba1\u7b97\u5668\u89e3\u51b3\u6570\u5b66\u95ee\u9898\u6216\u7528\u7f16\u8bd1\u5668\u6765\u751f\u6210\u4ee3\u7801\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s8.51cto.com\/oss\/202504\/21\/e7b5d7b98adf29e0eff957932317befee3c8db.webp\" alt=\"image.png\" data-type=\"inline\" \/>\u5177\u6709\u53ef\u9a8c\u8bc1\u5956\u52b1\uff08RLVR\uff09\u7684\u5f3a\u5316\u5b66\u4e60\u793a\u4f8b\uff1a\u6a21\u578b\u63a5\u6536\u63d0\u793a\u4ee5\u89e3\u51b3\u6570\u5b66\u95ee\u9898\u5e76\u751f\u6210\u7b54\u6848\u3002\u6b64\u65f6\u4e0d\u4f7f\u7528\u5b66\u4e60\u5956\u52b1\u6a21\u578b\uff0c\u800c\u662f\u7531\u7b26\u53f7\u9a8c\u8bc1\u5668\uff08\u4f8b\u5982\u8ba1\u7b97\u5668\uff09\u68c0\u67e5\u8f93\u51fa\uff0c\u5e76\u57fa\u4e8e\u6b63\u786e\u6027\u63d0\u4f9b\u4e8c\u5143\u53cd\u9988\u3002<\/p>\n<p>\u8fd9\u6837\u505a\u7684\u4e00\u4e2a\u52a8\u673a\u662f\uff0c\u5728\u5f3a\u5316\u5b66\u4e60\u671f\u95f4\uff0c\u901a\u8fc7\u81ea\u52a8\u6b63\u786e\u6027\u68c0\u67e5\u4f5c\u4e3a\u76d1\u7763\u4fe1\u53f7\uff0c\u4ece\u800c\u907f\u514d\u4f7f\u7528\u566a\u58f0\u5927\u6216\u6210\u672c\u9ad8\u6602\u7684\u4eba\u7c7b\u53cd\u9988\u6216 \u5b66\u4e60\u5956\u52b1\u3002\u53e6\u4e00\u4e2a\u52a8\u673a\u662f\uff0c\u501f\u52a9\u8ba1\u7b97\u5668\u8fd9\u7c7b\u300c\u5ec9\u4ef7\u300d\u5de5\u5177\uff0c\u6211\u4eec\u53ef\u4ee5\u66ff\u4ee3\u6210\u672c\u9ad8\u6602\u7684\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u4ee5\u53ca\u5956\u52b1\u6a21\u578b\u672c\u8eab\u3002\u7531\u4e8e\u5956\u52b1\u6a21\u578b\u901a\u5e38\u662f\u6574\u4e2a\u9884\u8bad\u7ec3\u6a21\u578b\uff08\u53ea\u662f\u5e26\u6709\u56de\u5f52\u5934\u90e8\uff09\uff0c\u56e0\u6b64 RLVR \u7684\u6548\u7387\u8981\u9ad8\u5f97\u591a\u3002<\/p>\n<p>\u603b\u4e4b\uff0cDeepSeek-R1 \u7ed3\u5408\u4e86 RLVR \u4e0e GRPO\uff0c\u4ece\u800c\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6dd8\u6c70\u4e86\u4e24\u4e2a\u6210\u672c\u9ad8\u6602\u7684\u6a21\u578b\uff1a\u5956\u52b1\u6a21\u578b\u548c\u4ef7\u503c\u6a21\u578b\uff08\u8bc4\u8bba\u5bb6\uff09\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s5.51cto.com\/oss\/202504\/21\/928368c729edb0ba60b8514b3aa1b0b7e5fb74.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u8bad\u7ec3\u4e2d\u5f3a\u5316\u5b66\u4e60\u8bbe\u7f6e\u7684\u6bd4\u8f83\u3002\u4f20\u7edf\u7684\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u4e0e\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09\u7ed3\u5408\u65f6\uff0c\u4f1a\u540c\u65f6\u4f7f\u7528\u57fa\u4e8e\u4eba\u7c7b\u504f\u597d\u7684\u5956\u52b1\u6a21\u578b\u548c\u8bc4\u8bba\u5bb6\uff08\u4ef7\u503c\u6a21\u578b\uff09\u6765\u6307\u5bfc\u5b66\u4e60\u3002\u800c\u5e7f\u4e49\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09\u5219\u6dd8\u6c70\u4e86\u8bc4\u8bba\u5bb6\u6a21\u578b\u3002\u66f4\u6709\u751a\u8005\uff0c\u5f53 GRPO \u4e0e\u5e26\u6709\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLVR\uff09\u7ed3\u5408\u65f6\uff0c\u8fd8\u53bb\u6389\u4e86\u5956\u52b1\u6a21\u578b\uff0c\u8f6c\u800c\u4f9d\u9760\u6765\u81ea\u7b26\u53f7\u5de5\u5177\uff08\u5982\u8ba1\u7b97\u5668\u6216\u7f16\u8bd1\u5668\uff09\u7684\u53ef\u9a8c\u8bc1\u5956\u52b1\u3002<\/p>\n<p>\u5728\u4e0b\u4e00\u8282\u4e2d\uff0c\u6211\u6253\u7b97\u7b80\u8981\u4ecb\u7ecd DeepSeek-R1 \u7684\u8bad\u7ec3\u6d41\u7a0b\uff0c\u5e76\u63a2\u8ba8 DeepSeek \u56e2\u961f\u6240\u4f7f\u7528\u7684\u4e0d\u540c\u53ef\u9a8c\u8bc1\u5956\u52b1\u65b9\u6cd5\u3002<\/p>\n<h4>DeepSeek-R1 \u63a8\u7406\u6a21\u578b\u7684\u8bad\u7ec3\u65b9\u6cd5<\/h4>\n<p>\u73b0\u5728\u6211\u4eec\u5df2\u7ecf\u9610\u660e\u4e86 RLHF \u548c RLVR \u4ee5\u53ca PPO \u548c GRPO \u7684\u6982\u5ff5\uff0c\u63a5\u4e0b\u6765\u5c06\u7b80\u8981\u56de\u987e DeepSeek-R1 \u8bba\u6587\u4e2d\u5173\u4e8e\u5f3a\u5316\u5b66\u4e60\u548c\u63a8\u7406\u7684\u4e3b\u8981\u89c1\u89e3\u3002<\/p>\n<p>\u9996\u5148\uff0c\u5b58\u5728\u4e09\u79cd\u7c7b\u578b\u7684\u6a21\u578b\uff1a<\/p>\n<p>1. \u4ec5\u901a\u8fc7\u7eaf\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8bad\u7ec3\u7684 DeepSeek-R1-Zero\u3002<\/p>\n<p>2. \u901a\u8fc7\u6307\u4ee4\u5fae\u8c03\uff08SFT\uff09\u548c\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u5171\u540c\u8bad\u7ec3\u7684 DeepSeek-R1\u3002<\/p>\n<p>3. \u901a\u8fc7\u6307\u4ee4\u5fae\u8c03\uff08SFT\uff09\u4e14\u672a\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u521b\u5efa\u7684 DeepSeek-Distill \u53d8\u4f53\u3002<\/p>\n<p>\u6211\u7ed8\u5236\u4e86\u4e00\u4e2a DeepSeek-R1 \u6d41\u7a0b\u56fe\uff0c\u7528\u4ee5\u5c55\u793a\u8fd9\u4e9b\u6a21\u578b\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u5982\u4e0b\u6240\u793a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202504\/21\/475557363e4e3e9addf53184608a519a3f7c81.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>DeepSeek-R1 \u7cfb\u5217\u7684\u8bad\u7ec3 pipeline<\/p>\n<p>DeepSeek-R1-Zero \u662f\u4f7f\u7528\u53ef\u9a8c\u8bc1\u5956\u52b1\uff08RLVR\uff09\u4e0e GRPO \u8bad\u7ec3\u7684\uff0c\u4e8b\u5b9e\u8bc1\u660e\uff0c\u8fd9\u8db3\u4ee5\u4f7f\u6a21\u578b\u901a\u8fc7\u751f\u6210\u4e2d\u95f4\u6b65\u9aa4\u5c55\u73b0\u63a8\u7406\u80fd\u529b\uff0c\u4e5f\u8bc1\u660e\u4e86\u8df3\u8fc7\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u9636\u6bb5\u662f\u53ef\u80fd\u7684\u3002\u8be5\u6a21\u578b\u901a\u8fc7\u63a2\u7d22\u800c\u975e\u4ece\u793a\u4f8b\u4e2d\u5b66\u4e60\u6765\u63d0\u5347\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p>DeepSeek-R1 \u662f\u65d7\u8230\u578b\u53f7\uff0c\u6027\u80fd\u6700\u4f73\u7684\u4e3b\u529b\u6a21\u578b\uff0c\u4e0e DeepSeek-R1-Zero \u7684\u533a\u522b\u5728\u4e8e\uff0c\u5b83\u4ea4\u66ff\u4f7f\u7528\u4e86\u6307\u4ee4\u5fae\u8c03\u3001RLVR \u548c RLHF\u3002<\/p>\n<p>DeepSeek-Distill \u53d8\u4f53\u65e8\u5728\u6210\u4e3a\u66f4\u5c0f\u4e14\u66f4\u6613\u90e8\u7f72\u7684\u6a21\u578b\uff0c\u5b83\u4eec\u662f\u901a\u8fc7\u4f7f\u7528 DeepSeek-R1 \u6a21\u578b\u7684\u6307\u4ee4\u6570\u636e\u5bf9 Llama 3 \u548c Qwen 2.5 \u6a21\u578b\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03\u751f\u6210\u7684\u3002\u8fd9\u79cd\u505a\u6cd5\u5728\u63a8\u7406\u90e8\u5206\u672a\u4f7f\u7528\u4efb\u4f55\u5f3a\u5316\u5b66\u4e60\uff08\u4e0d\u8fc7\uff0cLlama 3 \u548c Qwen 2.5 \u57fa\u7840\u6a21\u578b\u7684\u521b\u5efa\u7528\u4e86 RLHF\uff09\u3002<\/p>\n<p>\u5173\u4e8e DeepSeek-R1 \u6d41\u7a0b\u7684\u66f4\u591a\u8bb2\u89e3\uff0c\u53ef\u53c2\u9605\u6211\u4e4b\u524d\u7684\u300aUnderstanding Reasoning LLMs\u300b\u4e00\u6587\uff1a<\/p>\n<p>(https:\/\/magazine.sebastianraschka.com\/p\/understanding-reasoning-llms)<\/p>\n<p>\u9700\u8981\u5f3a\u8c03\u7684\u662f\uff0cDeepSeek \u56e2\u961f\u8bad\u7ec3 DeepSeek-R1-Zero \u65f6\u672a\u4f7f\u7528\u57fa\u4e8e LLM \u7684\u5956\u52b1\u6a21\u578b\uff0c\u800c\u662f\u5bf9 DeepSeek-R1-Zero \u548c DeepSeek-R1 \u7684\u63a8\u7406\u8bad\u7ec3\u91c7\u7528\u4e86\u57fa\u4e8e\u89c4\u5219\u7684\u5956\u52b1\uff1a<\/p>\n<p>\u5728\u5f00\u53d1 DeepSeek-R1-Zero \u65f6\uff0c\u6211\u4eec\u6ca1\u6709\u5e94\u7528\u7ed3\u679c\u6216\u8fc7\u7a0b\u795e\u7ecf\u5956\u52b1\u6a21\u578b\uff0c\u56e0\u4e3a\u6211\u4eec\u53d1\u73b0\uff0c\u5728\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u8fc7\u7a0b\u4e2d\uff0c\u795e\u7ecf\u5956\u52b1\u6a21\u578b\u53ef\u80fd\u4f1a\u51fa\u73b0\u5956\u52b1\u52ab\u6301\u95ee\u9898\u3002<\/p>\n<p>\u4e3a\u4e86\u8bad\u7ec3 DeepSeek-R1-Zero\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u4e00\u4e2a\u4e3b\u8981\u7531\u4e24\u79cd\u5956\u52b1\u7ec4\u6210\u7684\u57fa\u4e8e\u89c4\u5219\u7684\u5956\u52b1\u7cfb\u7edf\uff1a<\/p>\n<p>\uff081\uff09\u51c6\u786e\u6027\u5956\u52b1\uff1a\u51c6\u786e\u6027\u5956\u52b1\u6a21\u578b\u7528\u4e8e\u8bc4\u4f30\u56de\u7b54\u662f\u5426\u6b63\u786e\u3002\u4f8b\u5982\uff0c\u5728\u6570\u5b66\u95ee\u9898\u6709\u786e\u5b9a\u6027\u7ed3\u679c\u7684\u60c5\u51b5\u4e0b\uff0c\u6a21\u578b\u9700\u8981\u4ee5\u6307\u5b9a\u683c\u5f0f\uff08\u4f8b\u5982\uff0c\u5728\u65b9\u6846\u5185\uff09\u63d0\u4f9b\u6700\u7ec8\u7b54\u6848\uff0c\u4ee5\u4fbf\u53ef\u9760\u5730\u8fdb\u884c\u57fa\u4e8e\u89c4\u5219\u7684\u6b63\u786e\u6027\u9a8c\u8bc1\u3002\u540c\u6837\uff0c\u5bf9\u4e8e LeetCode \u95ee\u9898\uff0c\u53ef\u4ee5\u4f7f\u7528\u7f16\u8bd1\u5668\u6839\u636e\u9884\u5b9a\u4e49\u7684\u6d4b\u8bd5\u7528\u4f8b\u751f\u6210\u53cd\u9988\u3002<\/p>\n<p>\uff082\uff09\u683c\u5f0f\u5956\u52b1\uff1a\u9664\u4e86\u51c6\u786e\u6027\u5956\u52b1\u6a21\u578b\u5916\uff0c\u6211\u4eec\u8fd8\u91c7\u7528\u4e86\u683c\u5f0f\u5956\u52b1\u6a21\u578b\uff0c\u8981\u6c42\u6a21\u578b\u5c06\u5176\u601d\u8003\u8fc7\u7a0b\u7f6e\u4e8e\u300e&lt;think&gt;\u300f\u548c \u300e&lt;\/think&gt;\u300f\u6807\u7b7e\u4e4b\u95f4\u3002<\/p>\n<h4>\u4ece\u6700\u8fd1\u5173\u4e8e\u8bad\u7ec3\u63a8\u7406\u6a21\u578b\u7684 RL \u8bba\u6587\u4e2d\u6c72\u53d6\u7684\u6559\u8bad<\/h4>\n<p>\u6211\u610f\u8bc6\u5230\u5f15\u8a00\uff08\u5373\u5230\u6b64\u4e3a\u6b62\u7684\u6240\u6709\u5185\u5bb9\uff09\u6bd4\u6211\u9884\u60f3\u7684\u8981\u957f\u5f97\u591a\u3002\u5c3d\u7ba1\u5982\u6b64\uff0c\u6211\u8fd8\u662f\u8ba4\u4e3a\u6709\u5fc5\u8981\u7528\u8fd9\u4e48\u957f\u7684\u7bc7\u5e45\u6765\u4ecb\u7ecd\u4e0b\u9762\u7684\u7ecf\u9a8c\u6559\u8bad\u3002<\/p>\n<p>\u4e0a\u4e2a\u6708\uff0c\u6211\u9605\u8bfb\u4e86\u5927\u91cf\u6709\u5173\u63a8\u7406\u6a21\u578b\u7684\u6700\u65b0\u8bba\u6587\uff0c\u5e76\u5c06\u5176\u4e2d\u6700\u6709\u8da3\u7684\u89c2\u70b9\u548c\u89c1\u89e3\u5f52\u7eb3\u5728\u672c\u8282\u4e2d\u3002<\/p>\n<p><strong>1\u3001\u5f3a\u5316\u5b66\u4e60\u8fdb\u4e00\u6b65\u6539\u8fdb\u84b8\u998f\u6a21\u578b<\/strong><\/p>\n<p>DeepSeek-R1 \u8bba\u6587\u6e05\u695a\u5730\u8868\u660e\uff0c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u540e\u7684\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u4f18\u4e8e\u5355\u72ec\u7684\u5f3a\u5316\u5b66\u4e60\u3002<\/p>\n<p>\u9274\u4e8e\u8fd9\u4e00\u89c2\u5bdf\uff0c\u76f4\u89c2\u5730\u8bf4\uff0c\u989d\u5916\u7684\u5f3a\u5316\u5b66\u4e60\u5e94\u80fd\u8fdb\u4e00\u6b65\u6539\u8fdb\u84b8\u998f\u6a21\u578b\uff08\u56e0\u4e3a\u84b8\u998f\u6a21\u578b\u672c\u8d28\u4e0a\u4ee3\u8868\u4e86\u901a\u8fc7\u4f7f\u7528\u5927\u6a21\u578b\u751f\u6210\u7684\u63a8\u7406\u6837\u672c\u8fdb\u884c SFT \u8bad\u7ec3\u7684\u6a21\u578b\uff09\u3002<\/p>\n<p>\u4e8b\u5b9e\u4e0a\uff0cDeepSeek \u56e2\u961f\u660e\u786e\u89c2\u5bdf\u5230\u4e86\u8fd9\u4e00\u73b0\u8c61\uff1a<\/p>\n<blockquote><p>\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u53d1\u73b0\uff0c\u5c06 RL \u5e94\u7528\u4e8e\u8fd9\u4e9b\u7ecf\u8fc7\u84b8\u998f\u7684\u6a21\u578b\u8fd8\u80fd\u4ea7\u751f\u663e\u8457\u7684\u8fdb\u4e00\u6b65\u6536\u76ca\u3002\u6211\u4eec\u8ba4\u4e3a\u8fd9\u503c\u5f97\u8fdb\u4e00\u6b65\u63a2\u8ba8\uff0c\u56e0\u6b64\u5728\u6b64\u4ec5\u4ecb\u7ecd\u7b80\u5355\u7684 SFT \u84b8\u998f\u6a21\u578b\u7684\u7ed3\u679c\u3002<\/p><\/blockquote>\n<p>\u591a\u4e2a\u56e2\u961f\u72ec\u7acb\u5df2\u9a8c\u8bc1\u4e86\u8fd9\u4e9b\u89c2\u5bdf\u7ed3\u679c\uff1a<\/p>\n<ul data-id=\"ua73dd4b-A6qXWnID\">\n<li data-id=\"ld70c578-DoVBSFvV\">[8] \u7814\u7a76\u4eba\u5458\u4f7f\u7528 1.5B \u7684 DeepSeek-R1-Distill-Qwen \u6a21\u578b\uff0c\u4ec5\u7528 7,000 \u4e2a\u6837\u672c\u548c 42 \u7f8e\u5143\u7684\u8ba1\u7b97\u9884\u7b97\uff0c\u5c31\u8bc1\u660e\u4e86 RL \u5fae\u8c03\u5e26\u6765\u7684\u6027\u80fd\u5927\u5e45\u63d0\u5347\u3002\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u662f\uff0c\u8fd9\u4e2a\u5c0f\u6a21\u578b\u5728 AIME24 \u6570\u5b66\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8d85\u8fc7\u4e86 OpenAI \u7684 o1-preview\u3002<\/li>\n<li data-id=\"ld70c578-VgVYaa8o\">[15] \u4e0d\u8fc7\uff0c\u53e6\u4e00\u4e2a\u7814\u7a76\u5c0f\u7ec4\u63d0\u9192\u8bf4\uff0c\u8fd9\u4e9b\u6536\u76ca\u5728\u7edf\u8ba1\u5b66\u4e0a\u5e76\u4e0d\u603b\u662f\u663e\u8457\u7684\u3002\u8fd9\u8868\u660e\uff0c\u5c3d\u7ba1 RL \u53ef\u4ee5\u6539\u8fdb\u8f83\u5c0f\u7684\u84b8\u998f\u6a21\u578b\uff0c\u4f46\u57fa\u51c6\u7ed3\u679c\u6709\u65f6\u53ef\u80fd\u5938\u5927\u4e86\u6539\u8fdb\u6548\u679c\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202504\/21\/74e6b5954fb283dc604765ad06c695bd58bd7f.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u6ce8\u91ca\u56fe\u6765\u81ea\u300a\u51b7\u9759\u770b\u5f85\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7684\u8fdb\u5c55\uff1a \u9677\u9631\u4e0e\u590d\u73b0\u4e4b\u8def\u300b\uff0chttps:\/\/arxiv.org\/abs\/2504.07086<\/p>\n<p><strong>2\u3001\u5197\u957f\u9519\u8bef\u7b54\u6848\u7684\u95ee\u9898<\/strong><\/p>\n<p>\u6211\u4e4b\u524d\u63d0\u5230\u8fc7\uff0c\u6709\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u63a8\u7406\uff08RLVR\uff09\u5e76\u4e0d\u4e25\u683c\u8981\u6c42\u4f7f\u7528 GRPO \u7b97\u6cd5\uff1bDeepSeek \u7684 GRPO \u53ea\u662f\u78b0\u5de7\u6548\u7387\u9ad8\u3001\u6027\u80fd\u597d\u800c\u5df2\u3002<\/p>\n<p>\u7136\u800c\uff0c\u6587\u732e [12] \u8868\u660e\uff0c\u666e\u901a PPO \u642d\u914d\u57fa\u672c\u7684\u4e8c\u8fdb\u5236\u6b63\u786e\u6027\u5956\u52b1\u8db3\u4ee5\u6269\u5c55\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u548c\u54cd\u5e94\u957f\u5ea6\u3002<\/p>\n<p>\u66f4\u6709\u8da3\u7684\u662f\uff0cPPO \u548c GRPO \u90fd\u5b58\u5728\u957f\u5ea6\u504f\u5dee\u3002\u6709\u51e0\u7bc7\u8bba\u6587\u63a2\u8ba8\u4e86\u5904\u7406\u8fc7\u957f\u9519\u8bef\u7b54\u6848\u7684\u65b9\u6cd5\uff1a<\/p>\n<p>[14] \u63d0\u4f9b\u4e86\u4e00\u9879\u5206\u6790\uff0c\u8bf4\u660e\u4e86\u7531\u4e8e\u635f\u5931\u8ba1\u7b97\u4e2d\u7684\u6570\u5b66\u504f\u5dee\uff0cPPO \u5982\u4f55\u65e0\u610f\u4e2d\u504f\u5411\u4e8e\u8f83\u957f\u7684\u56de\u7b54\uff1bGRPO \u53ef\u80fd\u4e5f\u5b58\u5728\u540c\u6837\u7684\u95ee\u9898\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202504\/21\/07e3bb97181e697743b524a701dac51b50933f.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u6458\u81ea\u300a\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8fdb\u884c\u7b80\u660e\u63a8\u7406\u300b\uff0chttps:\/\/arxiv.org\/abs\/2504.05185<\/p>\n<p>\u4f5c\u4e3a\u4e0a\u8ff0\u58f0\u660e\u7684\u540e\u7eed\uff0c[7] [10] \u7279\u522b\u6307\u51fa\u4e86 GRPO \u4e2d\u7684\u957f\u5ea6\u548c\u96be\u5ea6\u504f\u5dee\u3002\u4fee\u6539\u540e\u7684\u53d8\u4f53\u300cDr. GRPO\u300d\u901a\u8fc7\u53bb\u9664\u957f\u5ea6\u548c\u6807\u51c6\u5dee\u5f52\u4e00\u5316\uff0c\u7b80\u5316\u4e86\u4f18\u52bf\u8ba1\u7b97\uff0c\u63d0\u4f9b\u4e86\u66f4\u6e05\u6670\u7684\u8bad\u7ec3\u4fe1\u53f7\u3002<\/p>\n<ul data-id=\"ua73dd4b-MNqAgiNI\">\n<li data-id=\"ld70c578-3vtIacJy\">[1] GRPO \u4e2d\u660e\u786e\u60e9\u7f5a\u5197\u957f\u7684\u9519\u8bef\u7b54\u6848\uff0c\u540c\u65f6\u5956\u52b1\u7b80\u6d01\u6b63\u786e\u7684\u7b54\u6848\u3002<\/li>\n<li data-id=\"ld70c578-9X18b9Rg\">[3] [6] \u5728 GRPO \u4e2d\u6ca1\u6709\u76f4\u63a5\u63a7\u5236\u7b54\u6848\u957f\u5ea6\uff0c\u4f46\u53d1\u73b0 token \u7ea7\u5956\u52b1\u662f\u6709\u76ca\u7684\uff0c\u53ef\u4ee5\u8ba9\u6a21\u578b\u66f4\u597d\u5730\u4e13\u6ce8\u4e8e\u5173\u952e\u63a8\u7406\u6b65\u9aa4\u3002<\/li>\n<li data-id=\"ld70c578-M3RnNV5w\">[5] \u5728 GRPO \u4e2d\u5bf9\u8d85\u8fc7\u7279\u5b9a\u957f\u5ea6\u7684\u56de\u7b54\u5f15\u5165\u660e\u786e\u7684\u60e9\u7f5a\u63aa\u65bd\uff0c\u4ece\u800c\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5b9e\u73b0\u7cbe\u786e\u7684\u957f\u5ea6\u63a7\u5236\u3002<\/li>\n<\/ul>\n<p><strong>3\u3001\u4ece RL \u4e2d\u4ea7\u751f\u7684\u80fd\u529b<\/strong><\/p>\n<p>\u9664\u4e86 DeepSeek-R1 \u8bba\u6587\u4e2d\u63d0\u5230\u7684\u987f\u609f\u65f6\u523b\uff0cRL \u8fd8\u88ab\u8bc1\u660e\u80fd\u591f\u5728\u6a21\u578b\u4e2d\u8bf1\u5bfc\u51fa\u5b9d\u8d35\u7684\u81ea\u6211\u9a8c\u8bc1\u548c\u53cd\u601d\u63a8\u7406\u80fd\u529b [2][9]\u3002\u6709\u8da3\u7684\u662f\uff0c\u4e0e\u987f\u609f\u65f6\u523b\u7c7b\u4f3c\uff0c\u8fd9\u4e9b\u80fd\u529b\u662f\u5728\u6ca1\u6709\u660e\u786e\u6307\u4ee4\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u81ea\u7136\u51fa\u73b0\u7684\u3002<\/p>\n<p>[1] \u8868\u660e\uff0c\u6269\u5c55\u4e0a\u4e0b\u6587\u957f\u5ea6\uff08\u6700\u591a 128k tokens\uff09\u53ef\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6a21\u578b\u7684\u81ea\u6211\u53cd\u7701\u548c\u81ea\u6211\u4fee\u6b63\u80fd\u529b\u3002<\/p>\n<p><strong>4\u3001\u8d85\u8d8a\u7279\u5b9a\u9886\u57df\u7684\u6cdb\u5316<\/strong><\/p>\n<p>\u8fc4\u4eca\u4e3a\u6b62\uff0c\u5927\u591a\u6570\u7814\u7a76\u5de5\u4f5c\u90fd\u96c6\u4e2d\u5728\u6570\u5b66\u6216\u7f16\u7801\u60c5\u5883\u4e2d\u7684\u63a8\u7406\u4efb\u52a1\u4e0a\u3002\u7136\u800c\uff0c[4] \u901a\u8fc7\u5728\u903b\u8f91\u8c1c\u9898\u4e0a\u8bad\u7ec3\u6a21\u578b\uff0c\u8bc1\u660e\u4e86\u6210\u529f\u7684\u6cdb\u5316\u3002\u5728\u903b\u8f91\u8c1c\u9898\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u5728\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e2d\u4e5f\u53d6\u5f97\u4e86\u5f88\u597d\u7684\u8868\u73b0\u3002\u8fd9\u8bc1\u660e\u4e86 RL \u80fd\u591f\u8bf1\u5bfc\u51fa\u72ec\u7acb\u4e8e\u7279\u5b9a\u9886\u57df\u77e5\u8bc6\u7684\u901a\u7528\u63a8\u7406\u884c\u4e3a\u3002<\/p>\n<p><strong>5\u3001\u6269\u5c55\u5230\u66f4\u5e7f\u6cdb\u7684\u9886\u57df<\/strong><\/p>\n<p>\u4f5c\u4e3a\u4e0a\u8ff0\u90e8\u5206\u7684\u540e\u7eed\uff0c\u53e6\u4e00\u4e2a\u6709\u8da3\u7684\u89c1\u89e3 [11] \u662f\uff0c\u63a8\u7406\u80fd\u529b\u53ef\u4ee5\u81ea\u7136\u5730\u6269\u5c55\u5230\u6570\u5b66\u3001\u4ee3\u7801\u548c\u903b\u8f91\u7b49\u7ed3\u6784\u5316\u9886\u57df\u4e4b\u5916\u3002\u6a21\u578b\u5df2\u7ecf\u6210\u529f\u5730\u5e94\u7528\u4e8e\u533b\u5b66\u3001\u5316\u5b66\u3001\u5fc3\u7406\u5b66\u3001\u7ecf\u6d4e\u5b66\u548c\u6559\u80b2\u7b49\u9886\u57df\uff0c\u5229\u7528\u751f\u6210\u5f0f soft-scoring \u65b9\u6cd5\u6709\u6548\u5730\u5904\u7406\u81ea\u7531\u5f62\u5f0f\u7684\u7b54\u6848\u3002<\/p>\n<p>\u63a8\u7406\u6a21\u578b\u4e0b\u4e00\u6b65\u7684\u91cd\u8981\u5de5\u4f5c\u5305\u62ec\uff1a<\/p>\n<ul data-id=\"ua73dd4b-AgkVV675\">\n<li data-id=\"ld70c578-2B2W2a95\">\u5c06\u73b0\u6709\u7684\u63a8\u7406\u6a21\u578b\uff08\u5982 o1\u3001DeepSeek-R1\uff09\u4e0e\u5916\u90e8\u5de5\u5177\u4f7f\u7528\u548c\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u7b49\u529f\u80fd\u76f8\u7ed3\u5408\uff1bOpenAI \u521a\u521a\u5b9e\u73b0\u7684 o3 \u6a21\u578b\u5728\u8fd9\u65b9\u9762\u94fa\u5e73\u4e86\u9053\u8def\uff1b<\/li>\n<li data-id=\"ld70c578-yXfA03Q1\">\u8bf4\u5230\u5de5\u5177\u4f7f\u7528\u548c\u641c\u7d22\uff0c[9] \u7814\u7a76\u8868\u660e\uff0c\u8d4b\u4e88\u63a8\u7406\u6a21\u578b\u641c\u7d22\u80fd\u529b\uff0c\u53ef\u8bf1\u5bfc\u51fa\u81ea\u6211\u4fee\u6b63\u548c\u8de8\u57fa\u51c6\u5f3a\u6cdb\u5316\u7b49\u884c\u4e3a\uff0c\u5c3d\u7ba1\u8bad\u7ec3\u6570\u636e\u96c6\u6781\u5c11\uff1b<\/li>\n<li data-id=\"ld70c578-NuMh87ev\">\u57fa\u4e8e DeepSeek-R1 \u56e2\u961f\u5728\u4fdd\u6301\u57fa\u4e8e\u77e5\u8bc6\u7684\u4efb\u52a1\u6027\u80fd\u65b9\u9762\u6240\u7ecf\u5386\u7684\u8270\u8f9b\uff0c\u6211\u8ba4\u4e3a\u4e3a\u63a8\u7406\u6a21\u578b\u6dfb\u52a0\u641c\u7d22\u80fd\u529b\u51e0\u4e4e\u662f\u4e0d\u8d39\u5439\u7070\u4e4b\u529b\u7684\u4e8b\u3002<\/li>\n<\/ul>\n<p><strong>6\u3001\u63a8\u7406\u662f\u5426\u5b8c\u5168\u5f52\u529f\u4e8e RL\uff1f<\/strong><\/p>\n<p>DeepSeek-R1\uff08\u548c R1-Zero\uff09\u80cc\u540e\u7684\u57fa\u672c\u89c2\u70b9\u662f\uff0cRLVR \u660e\u786e\u8bf1\u5bfc\u63a8\u7406\u80fd\u529b\u3002\u7136\u800c\uff0c\u6700\u8fd1\u7684\u7814\u7a76\u7ed3\u679c [10] \u8868\u660e\uff0c\u63a8\u7406\u884c\u4e3a\uff08\u5305\u62ec \u300c\u987f\u609f\u65f6\u523b\u300d\uff09\u53ef\u80fd\u5df2\u7ecf\u5b58\u5728\u4e8e\u57fa\u7840\u6a21\u578b\u4e2d\uff0c\u8fd9\u662f\u56e0\u4e3a\u5bf9\u5927\u91cf\u601d\u7ef4\u94fe\u6570\u636e\u8fdb\u884c\u4e86\u9884\u8bad\u7ec3\u3002<\/p>\n<p>\u6211\u6700\u8fd1\u5bf9 DeepSeek V3 \u57fa\u7840\u6a21\u578b\u548c R1 \u6a21\u578b\u8fdb\u884c\u7684\u6bd4\u8f83\u5f3a\u5316\u4e86\u8fd9\u4e00\u89c2\u70b9\uff0c\u56e0\u4e3a\u66f4\u65b0\u540e\u7684\u57fa\u7840\u6a21\u578b\u4e5f\u8868\u73b0\u51fa\u4e86\u7c7b\u4f3c\u63a8\u7406\u7684\u884c\u4e3a\u3002\u4f8b\u5982\uff0c\u539f\u59cb V3 \u6a21\u578b\u548c R1 \u6a21\u578b\u4e4b\u95f4\u7684\u6bd4\u8f83\u6e05\u695a\u5730\u663e\u793a\u4e86\u975e\u63a8\u7406\u6a21\u578b\u548c\u63a8\u7406\u6a21\u578b\u4e4b\u95f4\u7684\u533a\u522b\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202504\/21\/a20b38372e8508d17a5391811216742bd50a2d.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u4e0d\u8fc7\uff0c\u5982\u679c\u5c06\u66f4\u65b0\u540e\u7684 V3 \u57fa\u672c\u578b\u53f7\u4e0e R1 \u8fdb\u884c\u6bd4\u8f83\uff0c\u60c5\u51b5\u5c31\u4e0d\u4e00\u6837\u4e86\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202504\/21\/e15512025f72b61bc06537c6578a470479c82a.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u6b64\u5916\uff0c[13] \u8fd8\u53d1\u73b0\uff0c\u5728\u4e0d\u540c\u9886\u57df\u548c\u6a21\u578b\u5927\u5c0f\u7684\u9884\u8bad\u7ec3\u4e2d\uff0c\u81ea\u6211\u53cd\u601d\u548c\u81ea\u6211\u7ea0\u6b63\u884c\u4e3a\u4f1a\u9010\u6e10\u51fa\u73b0\u3002\u8fd9\u4f7f\u5f97\u5c06\u63a8\u7406\u80fd\u529b\u5b8c\u5168\u5f52\u56e0\u4e8e RL \u65b9\u6cd5\u53d8\u5f97\u66f4\u52a0\u590d\u6742\u3002<\/p>\n<p>\u4e5f\u8bb8\u7ed3\u8bba\u662f\uff0cRL \u7edd\u5bf9\u80fd\u5c06\u7b80\u5355\u7684\u57fa\u7840\u6a21\u578b\u8f6c\u5316\u4e3a\u63a8\u7406\u6a21\u578b\u3002\u7136\u800c\uff0c\u8fd9\u5e76\u4e0d\u662f\u8bf1\u5bfc\u6216\u63d0\u9ad8\u63a8\u7406\u80fd\u529b\u7684\u552f\u4e00\u65b9\u6cd5\u3002\u6b63\u5982 DeepSeek-R1 \u56e2\u961f\u6240\u5c55\u793a\u7684\uff0c\u84b8\u998f\u4e5f\u80fd\u63d0\u9ad8\u63a8\u7406\u80fd\u529b\u3002\u7531\u4e8e\u672c\u6587\u4e2d\u7684\u84b8\u998f\u6307\u7684\u662f\u5728\u601d\u7ef4\u94fe\u6570\u636e\u4e0a\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03\uff0c\u56e0\u6b64\u5728\u5305\u542b\u601d\u7ef4\u94fe\u6570\u636e\u7684\u6570\u636e\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u5f88\u53ef\u80fd\u4e5f\u4f1a\u8bf1\u53d1\u8fd9\u4e9b\u80fd\u529b\u3002<\/p>\n<p>\uff08\u6b63\u5982\u6211\u5728\u4e66\u4e2d\u901a\u8fc7\u5b9e\u8df5\u4ee3\u7801\u6240\u89e3\u91ca\u7684\uff0c\u9884\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u6bd5\u7adf\u662f\u57fa\u4e8e\u76f8\u540c\u7684\u4e0b\u4e00\u4e2a token \u9884\u6d4b\u4efb\u52a1\u548c\u635f\u5931\u51fd\u6570\uff09\u3002<\/p>\n<h4>\u503c\u5f97\u5173\u6ce8\u7684\u63a8\u7406\u6a21\u578b\u8bad\u7ec3\u7814\u7a76\u8bba\u6587<\/h4>\n<p>\u5728\u4e0a\u4e2a\u6708\u9605\u8bfb\u4e86\u5927\u91cf\u63a8\u7406\u8bba\u6587\u4e4b\u540e\uff0c\u6211\u8bd5\u56fe\u5728\u4e0a\u4e00\u8282\u4e2d\u603b\u7ed3\u51fa\u6700\u6709\u8da3\u7684\u6536\u83b7\u3002\u4e0d\u8fc7\uff0c\u5bf9\u4e8e\u90a3\u4e9b\u5bf9\u66f4\u8be6\u7ec6\u7684\u8d44\u6599\u6765\u6e90\u611f\u5230\u597d\u5947\u7684\u4eba\uff0c\u6211\u8fd8\u5728\u672c\u8282\u4e0b\u9762\u5217\u51fa\u4e86 15 \u7bc7\u76f8\u5173\u8bba\u6587\uff0c\u4f5c\u4e3a\u9009\u8bfb\u3002(\u4e3a\u7b80\u5355\u8d77\u89c1\uff0c\u4ee5\u4e0b\u6458\u8981\u6309\u65e5\u671f\u6392\u5e8f\uff09\u3002<\/p>\n<p>\u8bf7\u6ce8\u610f\uff0c\u8fd9\u4efd\u6e05\u5355\u5e76\u4e0d\u5168\u9762\uff08\u6211\u7684\u4e0a\u9650\u662f 15 \u7bc7\uff09\uff0c\u56e0\u4e3a\u672c\u6587\u5df2\u7ecf\u592a\u957f\u4e86\uff01<\/p>\n<p>[1] \u6269\u5c55\u5f3a\u5316\u5b66\u4e60\uff08\u548c\u4e0a\u4e0b\u6587\u957f\u5ea6\uff09<\/p>\n<p>22 Jan, Kimi k1.5: Scaling Reinforcement Learning with LLMs, https:\/\/arxiv.org\/abs\/2501.12599<\/p>\n<p>\u6709\u8da3\u7684\u662f\uff0c\u8fd9\u7bc7\u8bba\u6587\u4e0e DeepSeek-R1 \u8bba\u6587\u5728\u540c\u4e00\u5929\u53d1\u8868\uff01\u5728\u8fd9\u91cc\uff0c\u4f5c\u8005\u5c55\u793a\u4e86\u7528 RL \u8bad\u7ec3\u7684\u591a\u6a21\u6001 LLM\u3002\u4e0e DeepSeek-R1 \u7c7b\u4f3c\uff0c\u4ed6\u4eec\u6ca1\u6709\u4f7f\u7528\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09\uff0c\u800c\u662f\u91c7\u7528\u4e86\u53ef\u9a8c\u8bc1\u5956\u52b1\u3002PRM \u662f RL \u4e2d\u4f7f\u7528\u7684\u4e00\u79cd\u5956\u52b1\u6a21\u578b\uff08\u5c24\u5176\u662f\u5728 LLM \u8bad\u7ec3\u4e2d\uff09\uff0c\u5b83\u4e0d\u4ec5\u8bc4\u4f30\u6700\u7ec8\u7b54\u6848\uff0c\u8fd8\u8bc4\u4f30\u5f97\u51fa\u7b54\u6848\u7684\u63a8\u7406\u6b65\u9aa4\u3002<\/p>\n<p>\u8fd9\u91cc\u7684\u53e6\u4e00\u4e2a\u5173\u952e idea \u662f\uff0c\u6269\u5c55\u4e0a\u4e0b\u6587\u957f\u5ea6\uff08\u6700\u591a 128k \u4e2a token\uff09\u6709\u52a9\u4e8e\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u8fdb\u884c\u89c4\u5212\u3001\u53cd\u601d\u548c\u81ea\u6211\u4fee\u6b63\u3002\u56e0\u6b64\uff0c\u9664\u4e86\u4e0e DeepSeek-R1 \u7c7b\u4f3c\u7684\u6b63\u786e\u6027\u5956\u52b1\u5916\uff0c\u5b83\u4eec\u8fd8\u6709\u957f\u5ea6\u5956\u52b1\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u4ed6\u4eec\u63d0\u5021\u8f83\u77ed\u7684\u6b63\u786e\u7b54\u6848\uff0c\u800c\u4e0d\u6b63\u786e\u7684\u957f\u7b54\u6848\u5219\u4f1a\u53d7\u5230\u66f4\u591a\u60e9\u7f5a\u3002<\/p>\n<p>\u4ed6\u4eec\u8fd8\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a long2short \u7684\u65b9\u6cd5\uff0c\u7528\u4e8e\u5c06\u8fd9\u4e9b\u957f\u601d\u7ef4\u94fe\u6280\u80fd\u63d0\u70bc\u4e3a\u66f4\u9ad8\u6548\u7684 short-CoT \u6a21\u578b\u3002(\u5b83\u901a\u8fc7\u4f7f\u7528\u6a21\u578b\u5408\u5e76\u3001\u6700\u77ed\u62d2\u7edd\u91c7\u6837\u3001DPO \u548c\u7b2c\u4e8c\u8f6e\u5177\u6709\u66f4\u5f3a\u957f\u5ea6\u60e9\u7f5a\u7684 RL \u7b49\u65b9\u6cd5\uff0c\u4ece long-CoT \u6a21\u578b\u4e2d\u63d0\u70bc\u51fa\u66f4\u77ed\u7684\u6b63\u786e\u7b54\u6848\uff09\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s9.51cto.com\/oss\/202504\/21\/978f29b70d0a54affb7460ce132f15297de994.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[2] \u5927\u578b\u63a8\u7406\u6a21\u578b\u7684\u7ade\u4e89\u6027\u7f16\u7a0b<\/p>\n<p>3 Feb, Competitive Programming with Large Reasoning Models, https:\/\/arxiv.org\/abs\/2502.06807<\/p>\n<p>\u8fd9\u7bc7 OpenAI \u7684\u8bba\u6587\u8bc4\u4f30\u4e86\u4ed6\u4eec\u7684 o \u7cfb\u5217\u6a21\u578b\uff08\u5982 o1\u3001o1-ioi \u548c o3\uff09\u5728\u7ade\u4e89\u6027\u7f16\u7a0b\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002\u867d\u7136\u6ca1\u6709\u6df1\u5165\u63a2\u8ba8\u5982\u4f55\u5e94\u7528 RL \u7684\u6280\u672f\u7ec6\u8282\uff0c\u4f46\u5b83\u4ecd\u7136\u63d0\u4f9b\u4e86\u4e00\u4e9b\u6709\u8da3\u7684\u542f\u793a\u3002<\/p>\n<p>\u9996\u5148\uff0c\u8fd9\u4e9b\u6a21\u578b\u662f\u4f7f\u7528\u57fa\u4e8e\u7ed3\u679c\u7684 RL \u8bad\u7ec3\u51fa\u6765\u7684\uff0c\u800c\u4e0d\u662f\u57fa\u4e8e\u8fc7\u7a0b\u7684\u5956\u52b1\u6a21\u578b\u3002\u8fd9\u4e0e DeepSeek-R1 \u548c Kimi \u7b49\u65b9\u6cd5\u7c7b\u4f3c\u3002<\/p>\n<p>\u5176\u4e2d\u4e00\u4e2a\u6709\u8da3\u7684\u53d1\u73b0\u662f\uff0co3 \u53ef\u4ee5\u5b66\u4e60\u81ea\u5df1\u7684\u6d4b\u8bd5\u65f6\u95f4\uff08\u5373\u63a8\u7406\u65f6\u95f4\u6269\u5c55\uff09\u7b56\u7565\u3002\u4f8b\u5982\uff0c\u5b83\u7ecf\u5e38\u7f16\u5199\u4e00\u4e2a\u95ee\u9898\u7684\u7b80\u5355\u7c97\u66b4\u7248\u672c\uff08\u7528\u6548\u7387\u6362\u53d6\u6b63\u786e\u6027\uff09\uff0c\u7136\u540e\u7528\u5b83\u6765\u9a8c\u8bc1\u66f4\u4f18\u5316\u89e3\u51b3\u65b9\u6848\u7684\u8f93\u51fa\u3002\u8fd9\u79cd\u7b56\u7565\u4e0d\u662f\u624b\u5de5\u7f16\u7801\u7684\uff0c\u800c\u662f\u6a21\u578b\u81ea\u5df1\u60f3\u51fa\u6765\u7684\u3002<\/p>\n<p>\u603b\u7684\u6765\u8bf4\uff0c\u8bba\u6587\u8ba8\u8bba\u4e86\u6269\u5c55\u901a\u7528 RL \u5141\u8bb8\u6a21\u578b\u5f00\u53d1\u81ea\u5df1\u7684\u63a8\u7406\u548c\u9a8c\u8bc1\u65b9\u6cd5\uff0c\u800c\u4e0d\u9700\u8981\u4efb\u4f55\u4eba\u7c7b\u542f\u53d1\u5f0f\u65b9\u6cd5\u6216\u7279\u5b9a\u9886\u57df\u7684\u63a8\u7406 pipeline\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0co1-ioi \u7b49\u5176\u4ed6\uff08\u65e9\u671f\uff09\u6a21\u578b\u5219\u4f9d\u8d56\u4e8e\u624b\u5de5\u5236\u4f5c\u7684\u6d4b\u8bd5\u65f6\u95f4\u7b56\u7565\uff0c\u4f8b\u5982\u5bf9\u6210\u5343\u4e0a\u4e07\u7684\u6837\u672c\u8fdb\u884c\u805a\u7c7b\u548c\u91cd\u6392\uff0c\u8fd9\u9700\u8981\u5927\u91cf\u7684\u4eba\u5de5\u8bbe\u8ba1\u548c\u8c03\u6574\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s8.51cto.com\/oss\/202504\/21\/b4abd11464b420966f21325158c3b72165a318.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[3] \u63a2\u7d22\u7ed3\u679c\u5956\u52b1\u7684\u6781\u9650<\/p>\n<p>10 Feb, Exploring the Limit of Outcome Reward for Learning Mathematical Reasoning, https:\/\/arxiv.org\/abs\/2502.06781<\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u63a2\u8ba8\u4e86\u4ec5\u6709\u4e8c\u8fdb\u5236\u300c\u6b63\u786e\u300d\u6216\u300c\u9519\u8bef\u300d\u53cd\u9988\u7684 RL\uff08\u5982 DeepSeek-R1 \u4e2d\u7684 RL\uff09\u5728\u89e3\u51b3\u6570\u5b66\u95ee\u9898\u65b9\u9762\u80fd\u8d70\u591a\u8fdc\u3002\u4e3a\u6b64\uff0c\u4ed6\u4eec\u9996\u5148\u4f7f\u7528 Best-of-N \u91c7\u6837\u6765\u6536\u96c6\u6b63\u9762\u6837\u672c\uff0c\u5e76\u5bf9\u8fd9\u4e9b\u6837\u672c\u5e94\u7528\u884c\u4e3a\u514b\u9686\uff0c\u7ed3\u679c\u8868\u660e\u7406\u8bba\u4e0a\u8fd9\u8db3\u4ee5\u4f18\u5316\u7b56\u7565\u3002<\/p>\n<p>\u4e3a\u4e86\u5e94\u5bf9\u5956\u52b1\u7a00\u758f\u7684\u6311\u6218\uff08\u5c24\u5176\u662f\u5f53\u957f\u601d\u7ef4\u94fe\u5305\u542b\u90e8\u5206\u6b63\u786e\u6b65\u9aa4\u65f6\uff09\uff0c\u4ed6\u4eec\u6dfb\u52a0\u4e86\u4e00\u4e2a token \u7ea7\u5956\u52b1\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u53ef\u5b66\u4e60\u4e3a\u63a8\u7406\u7684\u4e0d\u540c\u90e8\u5206\u5206\u914d\u91cd\u8981\u6027\u6743\u91cd\u3002\u8fd9\u6709\u52a9\u4e8e\u6a21\u578b\u5728\u5b66\u4e60\u65f6\u4e13\u6ce8\u4e8e\u6700\u5173\u952e\u7684\u6b65\u9aa4\uff0c\u5e76\u63d0\u9ad8\u6574\u4f53\u6027\u80fd\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202504\/21\/2707c0130833b594f4a8608d5666ae745304f6.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[4] \u57fa\u4e8e\u89c4\u5219\u5f3a\u5316\u7684 LLM \u63a8\u7406\uff08\u5173\u4e8e\u903b\u8f91\u6570\u636e\uff09<\/p>\n<p>20 Feb, Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning, https:\/\/arxiv.org\/abs\/2502.14768<\/p>\n<p>DeepSeek-R1 \u4e13\u6ce8\u4e8e\u6570\u5b66\u548c\u4ee3\u7801\u4efb\u52a1\u3002\u672c\u6587\u4f7f\u7528\u903b\u8f91\u8c1c\u9898\u4f5c\u4e3a\u4e3b\u8981\u8bad\u7ec3\u6570\u636e\uff0c\u8bad\u7ec3\u4e86\u4e00\u4e2a 7B \u6a21\u578b\u3002<\/p>\n<p>\u7814\u7a76\u4eba\u5458\u91c7\u7528\u4e86\u4e0e DeepSeek-R1 \u7c7b\u4f3c\u7684\u57fa\u4e8e\u89c4\u5219\u7684 RL \u8bbe\u7f6e\uff0c\u4f46\u505a\u4e86\u4e00\u4e9b\u8c03\u6574\uff1a<\/p>\n<p>1. \u4ed6\u4eec\u5f15\u5165\u4e86\u4e25\u683c\u7684\u683c\u5f0f\u5956\u52b1\uff0c\u60e9\u7f5a\u8d70\u6377\u5f84\u7684\u884c\u4e3a\uff0c\u5e76\u786e\u4fdd\u6a21\u578b\u4f7f\u7528 &lt;think&gt; \u548c &lt;answer&gt; \u6807\u8bb0\u5c06\u63a8\u7406\u4e0e\u6700\u7ec8\u7b54\u6848\u5206\u5f00\u3002<\/p>\n<p>2. \u4ed6\u4eec\u8fd8\u4f7f\u7528\u4e86\u7cfb\u7edf\u63d0\u793a\uff0c\u660e\u786e\u544a\u8bc9\u6a21\u578b\u5728\u7ed9\u51fa\u6700\u7ec8\u7b54\u6848\u4e4b\u524d\u8981\u5148\u4e00\u6b65\u4e00\u6b65\u5730\u601d\u8003\u95ee\u9898\u3002<\/p>\n<p>\u5373\u4f7f\u53ea\u6709 5K \u4e2a\u5408\u6210\u903b\u8f91\u95ee\u9898\uff0c\u6a21\u578b\u4e5f\u80fd\u53d1\u5c55\u51fa\u826f\u597d\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5e76\u80fd\u5f88\u597d\u5730\u63a8\u5e7f\u5230 AIME \u548c AMC \u7b49\u66f4\u96be\u7684\u6570\u5b66\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u3002<\/p>\n<p>\u8fd9\u4e00\u70b9\u7279\u522b\u6709\u8da3\uff0c\u56e0\u4e3a\u5b83\u8868\u660e\u57fa\u4e8e\u903b\u8f91\u7684 RL \u8bad\u7ec3\u53ef\u4ee5\u6559\u4f1a\u6a21\u578b\u63a8\u7406\u7684\u65b9\u5f0f\uff0c\u5e76\u5c06\u5176\u5e94\u7528\u5230\u539f\u59cb\u9886\u57df\u4e4b\u5916\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202504\/21\/a71bcbf22e9b5d8c13e6843f5cef0e470804db.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[5] \u63a7\u5236\u63a8\u7406\u6a21\u578b\u7684\u601d\u8003\u65f6\u95f4<\/p>\n<p>6 Mar, L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning, https:\/\/arxiv.org\/abs\/2503.04697<\/p>\n<p>\u63a8\u7406\u6a21\u578b\u7684\u4e00\u4e2a\u7279\u70b9\u662f\uff0c\u7531\u4e8e\u601d\u7ef4\u94fe\u5f0f\u63a8\u7406\uff0c\u5b83\u4eec\u5f80\u5f80\u4f1a\u4ea7\u751f\u8f83\u957f\u7684\u8f93\u51fa\u3002\u4f46\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0c\u6ca1\u6709\u660e\u786e\u7684\u65b9\u6cd5\u6765\u63a7\u5236\u54cd\u5e94\u7684\u957f\u5ea6\u3002<\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u4ecb\u7ecd\u4e86\u957f\u5ea6\u63a7\u5236\u7b56\u7565\u4f18\u5316\uff08LCPO\uff09\uff0c\u8fd9\u662f\u4e00\u79cd\u7b80\u5355\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff0c\u53ef\u5e2e\u52a9\u6a21\u578b\u5728\u4f18\u5316\u51c6\u786e\u6027\u7684\u540c\u65f6\u9075\u5b88\u7528\u6237\u6307\u5b9a\u7684\u957f\u5ea6\u9650\u5236\u3002<\/p>\n<p>\u7b80\u800c\u8a00\u4e4b\uff0cLCPO \u7c7b\u4f3c\u4e8e GRPO\uff0c\u5373\u300cGRPO + \u957f\u5ea6\u63a7\u5236\u81ea\u5b9a\u4e49\u5956\u52b1\u300d\uff0c\u5176\u5b9e\u73b0\u65b9\u5f0f\u4e3a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_7\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">reward <span class=\"token operator\">=<\/span> reward_correctness <span class=\"token operator\">-<\/span> \u03b1 <span class=\"token operator\">*<\/span> <span class=\"token operator\">|<\/span>target_length <span class=\"token operator\">-<\/span> actual_length<span class=\"token operator\">|<\/span><\/code><\/pre>\n<ul id=\"code_id_7\" class=\"pre-numbering\">\n<li>1.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u5176\u4e2d\u76ee\u6807\u957f\u5ea6\u662f\u7528\u6237\u63d0\u793a\u7684\u4e00\u90e8\u5206\u3002\u4e0a\u8ff0 LCPO \u65b9\u6cd5\u9f13\u52b1\u6a21\u578b\u5b8c\u5168\u9075\u5b88\u6240\u63d0\u4f9b\u7684\u76ee\u6807\u957f\u5ea6\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u5f15\u5165\u4e86\u4e00\u4e2a LCPO-Max \u53d8\u4f53\uff0c\u5b83\u4e0d\u662f\u9f13\u52b1\u6a21\u578b\u5b8c\u5168\u5339\u914d\u76ee\u6807\u957f\u5ea6\uff0c\u800c\u662f\u9f13\u52b1\u6a21\u578b\u4fdd\u6301\u4f4e\u4e8e\u6700\u5927 token \u957f\u5ea6\uff1a<\/p>\n<div>\n<div class=\"hljs-cto\">\n<div class=\"hljs-cto\"><button class=\"copy_btn disable\" data-clipboard-target=\"#code_id_8\">\u590d\u5236<\/button><\/p>\n<div class=\"code-toolbar\">\n<pre class=\"has-pre-numbering language-javascript\" tabindex=\"0\"><code class=\"language-javascript\">reward <span class=\"token operator\">=<\/span> reward_correctness <span class=\"token operator\">*<\/span> <span class=\"token function\">clip<\/span> <span class=\"token punctuation\">(<\/span>\u03b1 <span class=\"token operator\">*<\/span> <span class=\"token punctuation\">(<\/span>target_length <span class=\"token operator\">-<\/span> actual_length<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">+<\/span> \u03b4<span class=\"token punctuation\">,<\/span> <span class=\"token number\">0<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><\/code><\/pre>\n<ul id=\"code_id_8\" class=\"pre-numbering\">\n<li>1.<\/li>\n<\/ul>\n<div class=\"toolbar\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u4f5c\u8005\u4f7f\u7528 LCPO \u8bad\u7ec3\u4e86\u4e00\u4e2a\u540d\u4e3a L1 \u7684 1.5B \u6a21\u578b\uff0c\u5b83\u53ef\u4ee5\u6839\u636e prompt \u8c03\u6574\u8f93\u51fa\u957f\u5ea6\u3002\u8fd9\u6837\uff0c\u7528\u6237\u5c31\u53ef\u4ee5\u6839\u636e\u4efb\u52a1\u5728\u51c6\u786e\u6027\u548c\u8ba1\u7b97\u91cf\u4e4b\u95f4\u8fdb\u884c\u6743\u8861\u3002\u6709\u8da3\u7684\u662f\uff0c\u8bba\u6587\u8fd8\u53d1\u73b0\uff0c\u8fd9\u4e9b\u957f\u94fe\u6a21\u578b\u5728\u77ed\u63a8\u7406\u65b9\u9762\u7684\u8868\u73b0\u4e5f\u51fa\u4eba\u610f\u6599\u5730\u597d\uff0c\u5728\u76f8\u540c\u7684 token \u957f\u5ea6\u4e0b\uff0c\u751a\u81f3\u8d85\u8fc7\u4e86 GPT-4o \u7b49\u66f4\u5927\u7684\u6a21\u578b\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s5.51cto.com\/oss\/202504\/21\/130854d353333f4b5b79200e57e8bceaa74bd7.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[6] \u5728 LLM \u4e2d\u6fc0\u52b1\u641c\u7d22\u80fd\u529b<\/p>\n<p>10 Mar, R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning, https:\/\/arxiv.org\/abs\/2503.05592<\/p>\n<p>\u50cf DeepSeek-R1 \u8fd9\u6837\u7684\u63a8\u7406\u6a21\u578b\u662f\u901a\u8fc7 RL \u8bad\u7ec3\u51fa\u6765\u7684\uff0c\u5b83\u4eec\u4f9d\u8d56\u4e8e\u81ea\u5df1\u7684\u5185\u90e8\u77e5\u8bc6\u3002\u672c\u6587\u4f5c\u8005\u7684\u7814\u7a76\u91cd\u70b9\u662f\u901a\u8fc7\u589e\u52a0\u5916\u90e8\u641c\u7d22\u7cfb\u7edf\u7684\u8bbf\u95ee\u6743\u9650\uff0c\u5728\u9700\u8981\u66f4\u591a\u65f6\u95f4\u654f\u611f\u4fe1\u606f\u6216\u6700\u65b0\u4fe1\u606f\u7684\u77e5\u8bc6\u578b\u4efb\u52a1\u4e2d\u6539\u8fdb\u8fd9\u4e9b\u6a21\u578b\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u672c\u6587\u901a\u8fc7\u6559\u8fd9\u4e9b\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u5916\u90e8\u641c\u7d22\u7cfb\u7edf\u6765\u6539\u8fdb\u5b83\u4eec\u3002\u4f5c\u8005\u6ca1\u6709\u4f9d\u8d56\u6d4b\u8bd5\u65f6\u95f4\u7b56\u7565\u6216\u76d1\u7763\u8bad\u7ec3\uff0c\u800c\u662f\u91c7\u7528\u4e86\u4e24\u9636\u6bb5\u5f3a\u5316\u5b66\u4e60\u6cd5\uff0c\u5e2e\u52a9\u6a21\u578b\u5b66\u4e60\u5982\u4f55\u4ee5\u53ca\u4f55\u65f6\u81ea\u884c\u641c\u7d22\u3002\u6a21\u578b\u9996\u5148\u5b66\u4e60\u641c\u7d22\u683c\u5f0f\uff0c\u7136\u540e\u5b66\u4e60\u5982\u4f55\u4f7f\u7528\u641c\u7d22\u7ed3\u679c\u627e\u5230\u6b63\u786e\u7b54\u6848\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202504\/21\/6677a6031e3f42d141a7540ddad132e0b38cfc.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[7] \u5f00\u6e90 LLM \u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60<\/p>\n<p>18 Mar, DAPO: An Open-Source LLM Reinforcement Learning System at Scale, https:\/\/arxiv.org\/abs\/2503.14476<\/p>\n<p>\u867d\u7136\u672c\u6587\u4e3b\u8981\u8ba8\u8bba\u5f00\u53d1\u7c7b\u4f3c DeepSeek-R1 \u7684\u8bad\u7ec3\u6d41\u7a0b\u5e76\u5c06\u5176\u5f00\u6e90\uff0c\u4f46\u5b83\u4e5f\u5bf9 DeepSeek-R1 \u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684 GRPO \u7b97\u6cd5\u63d0\u51fa\u4e86\u6709\u8da3\u7684\u6539\u8fdb\u3002<\/p>\n<p>1. Clip-higher\uff1a\u589e\u52a0 PPO \u526a\u679d\u8303\u56f4\u7684\u4e0a\u9650\uff0c\u4ee5\u9f13\u52b1\u63a2\u7d22\u5e76\u9632\u6b62\u8bad\u7ec3\u671f\u95f4\u71b5\u5d29\u6e83\u3002<\/p>\n<p>2. \u52a8\u6001\u91c7\u6837\uff1a\u901a\u8fc7\u8fc7\u6ee4\u6389\u6240\u6709\u91c7\u6837\u54cd\u5e94\u59cb\u7ec8\u6b63\u786e\u6216\u59cb\u7ec8\u9519\u8bef\u7684 prompt \u6765\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n<p>3. token \u7ea7\u7b56\u7565\u68af\u5ea6\u635f\u5931\uff1a\u4ece\u6837\u672c\u7ea7\u8f6c\u79fb\u5230 token \u7ea7\u68af\u5ea6\u635f\u5931\u8ba1\u7b97\uff0c\u4ee5\u4fbf\u66f4\u957f\u7684\u54cd\u5e94\u80fd\u591f\u5bf9\u68af\u5ea6\u66f4\u65b0\u4ea7\u751f\u66f4\u5927\u7684\u5f71\u54cd\u3002<\/p>\n<p>4. \u8fc7\u957f\u5956\u52b1\u5851\u9020\uff1a\u5bf9\u56e0\u8fc7\u957f\u800c\u88ab\u622a\u65ad\u7684\u54cd\u5e94\u6dfb\u52a0\u8f6f\u60e9\u7f5a\uff0c\u4ee5\u51cf\u5c11\u5956\u52b1\u566a\u97f3\u5e76\u6709\u52a9\u4e8e\u7a33\u5b9a\u8bad\u7ec3\u3002<\/p>\n<p>\u6807\u51c6 GRPO \u91c7\u7528\u6837\u672c\u7ea7\u635f\u5931\u8ba1\u7b97\u3002\u8fd9\u9996\u5148\u9700\u8981\u5bf9\u6bcf\u4e2a\u6837\u672c\u7684 token \u635f\u5931\u6c42\u5e73\u5747\u503c\uff0c\u7136\u540e\u518d\u5bf9\u6240\u6709\u6837\u672c\u7684\u635f\u5931\u6c42\u5e73\u5747\u503c\u3002\u7531\u4e8e\u6837\u672c\u7684\u6743\u91cd\u76f8\u7b49\uff0c\u56e0\u6b64\u54cd\u5e94\u8f83\u957f\u7684\u6837\u672c\u4e2d\u7684 token \u5bf9\u6574\u4f53\u635f\u5931\u7684\u8d21\u732e\u53ef\u80fd\u4f1a\u4e0d\u6210\u6bd4\u4f8b\u5730\u8f83\u5c0f\u3002\u540c\u65f6\uff0c\u7814\u7a76\u4eba\u5458\u89c2\u5bdf\u5230\uff0c\u8f83\u957f\u7684\u54cd\u5e94\u901a\u5e38\u5728\u6700\u7ec8\u7b54\u6848\u4e4b\u524d\u5305\u542b\u4e00\u4e9b\u4e71\u7801\uff0c\u800c\u8fd9\u4e9b\u4e71\u7801\u5728\u539f\u59cb GRPO \u6837\u672c\u7ea7\u635f\u5931\u8ba1\u7b97\u4e2d\u4e0d\u4f1a\u53d7\u5230\u8db3\u591f\u7684\u60e9\u7f5a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202504\/21\/835ed4200e4b775c10a4817ce03d53875b8672.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[8] \u5f3a\u5316\u5b66\u4e60\u5728\u5c0f\u578b LLM \u4e2d\u7684\u63a8\u7406<\/p>\n<p>20 Mar, Reinforcement Learning for Reasoning in Small LLMs: What Works and What Doesn&#8217;t, https:\/\/arxiv.org\/abs\/2503.16219<\/p>\n<p>DeepSeek-R1 \u7684\u539f\u59cb\u8bba\u6587\u8868\u660e\uff0c\u5728\u5f00\u53d1\u5c0f\u578b\uff08\u8f83\u5c0f\uff09\u63a8\u7406\u6a21\u578b\u65f6\uff0c\u84b8\u998f\u6bd4\u7eaf\u5f3a\u5316\u5b66\u4e60\u80fd\u53d6\u5f97\u66f4\u597d\u7684\u7ed3\u679c\u3002\u5728\u672c\u6587\u4e2d\uff0c\u7814\u7a76\u4eba\u5458\u5bf9\u6b64\u8fdb\u884c\u4e86\u8ddf\u8fdb\u7814\u7a76\uff0c\u5e76\u63a2\u7d22\u4e86\u5982\u4f55\u5229\u7528\u5f3a\u5316\u5b66\u4e60\u8fdb\u4e00\u6b65\u6539\u8fdb\u5c0f\u578b\u84b8\u998f\u63a8\u7406\u6a21\u578b\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u4ed6\u4eec\u4f7f\u7528 1.5B \u7684 DeepSeek-R1-Distill-Qwen \u6a21\u578b\uff0c\u53d1\u73b0\u4ec5\u9700 7000 \u4e2a\u8bad\u7ec3\u6837\u672c\u548c 42 \u7f8e\u5143\u7684\u8ba1\u7b97\u9884\u7b97\uff0c\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u5c31\u80fd\u5e26\u6765\u663e\u8457\u7684\u63d0\u5347\u3002\u4f8b\u5982\uff0c\u5728 AIME24 \u6570\u5b66\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u8fd9\u4e9b\u6539\u8fdb\u8db3\u4ee5\u8d85\u8d8a OpenAI \u7684 o1-preview\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u8be5\u8bba\u6587\u8fd8\u6709 3 \u4e2a\u6709\u8da3\u7684\u53d1\u73b0\uff1a<\/p>\n<p>1. \u5c0f\u578b LLM \u53ef\u4ee5\u5728\u4f7f\u7528\u7d27\u51d1\u3001\u9ad8\u8d28\u91cf\u6570\u636e\u96c6\u7684\u524d 50-100 \u4e2a\u8bad\u7ec3\u6b65\u5185\u5b9e\u73b0\u5feb\u901f\u63a8\u7406\u63d0\u5347\u3002\u4f46\u662f\uff0c\u5982\u679c\u8bad\u7ec3\u65f6\u95f4\u8fc7\u957f\uff0c\u6027\u80fd\u4f1a\u8fc5\u901f\u4e0b\u964d\uff0c\u8fd9\u4e3b\u8981\u662f\u7531\u4e8e\u957f\u5ea6\u9650\u5236\u548c\u8f93\u51fa\u4e0d\u7a33\u5b9a\uff1b<\/p>\n<p>2. \u5c06\u8f83\u6613\u548c\u8f83\u96be\u7684\u95ee\u9898\u6df7\u5408\u5728\u4e00\u8d77\uff0c\u6709\u52a9\u4e8e\u6a21\u578b\u5728\u8bad\u7ec3\u65e9\u671f\u4ea7\u751f\u66f4\u77ed\u3001\u66f4\u7a33\u5b9a\u7684\u54cd\u5e94\u3002\u7136\u800c\uff0c\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\uff0c\u6027\u80fd\u4ecd\u7136\u4f1a\u4e0b\u964d\uff1b<\/p>\n<p>3. \u4f7f\u7528\u4f59\u5f26\u5f62\u5956\u52b1\u51fd\u6570\u6709\u52a9\u4e8e\u66f4\u6709\u6548\u5730\u63a7\u5236\u8f93\u51fa\u957f\u5ea6\uff0c\u5e76\u63d0\u9ad8\u8bad\u7ec3\u4e00\u81f4\u6027\u3002\u4f46\u4e0e\u57fa\u4e8e\u51c6\u786e\u7387\u7684\u6807\u51c6\u5956\u52b1\u76f8\u6bd4\uff0c\u8fd9\u4f1a\u7565\u5fae\u964d\u4f4e\u5cf0\u503c\u6027\u80fd\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202504\/21\/134fd97752186dcd2c01797c47600a8dcb3c8a.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[9] \u5b66\u4e60\u901a\u8fc7\u641c\u7d22\u8fdb\u884c\u63a8\u7406<\/p>\n<p>25 Mar, ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning, https:\/\/arxiv.org\/abs\/2503.19470<\/p>\n<p>\u672c\u6587\u63d0\u51fa\u7684 ReSearch \u6846\u67b6\u6269\u5c55\u4e86 DeepSeek-R1 \u8bba\u6587\u4e2d\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff0c\u5c06\u641c\u7d22\u7ed3\u679c\u7eb3\u5165\u63a8\u7406\u8fc7\u7a0b\u3002\u8be5\u6a21\u578b\u4f1a\u6839\u636e\u5176\u6b63\u5728\u8fdb\u884c\u7684\u63a8\u7406\u94fe\u5b66\u4e60\u4f55\u65f6\u4ee5\u53ca\u5982\u4f55\u8fdb\u884c\u641c\u7d22\uff0c\u7136\u540e\u5c06\u68c0\u7d22\u5230\u7684\u4fe1\u606f\u7528\u4e8e\u540e\u7eed\u7684\u63a8\u7406\u6b65\u9aa4\u3002<\/p>\n<p>\u8fd9\u4e00\u5207\u90fd\u662f\u5728\u63a8\u7406\u6b65\u9aa4\u4e2d\u65e0\u9700\u76d1\u7763\u6570\u636e\u7684\u60c5\u51b5\u4e0b\u5b8c\u6210\u7684\u3002\u7814\u7a76\u4eba\u5458\u8fd8\u8868\u660e\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u4ea7\u751f\u8bf8\u5982\u81ea\u6211\u4fee\u6b63\u548c\u53cd\u601d\u7b49\u6709\u7528\u7684\u884c\u4e3a\uff0c\u5e76\u4e14\u5c3d\u7ba1\u53ea\u5728\u4e00\u4e2a\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f46\u5b83\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s6.51cto.com\/oss\/202504\/21\/4222500067eff354def435bb3126078c8bea94.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>PS\uff1a\u8fd9\u4e2a\u65b9\u6cd5\u548c\u4e4b\u524d\u8ba8\u8bba\u7684 R1-Searcher \u6709\u4ec0\u4e48\u4e0d\u540c\uff1f<\/p>\n<p>R1-Searcher \u91c7\u7528\u4e24\u9636\u6bb5\u3001\u57fa\u4e8e\u7ed3\u679c\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u3002\u5728\u7b2c\u4e00\u9636\u6bb5\uff0c\u5b83\u6559\u4f1a\u6a21\u578b\u5982\u4f55\u8c03\u7528\u5916\u90e8\u68c0\u7d22\uff1b\u5728\u7b2c\u4e8c\u9636\u6bb5\uff0c\u5b83\u5b66\u4e60\u5982\u4f55\u5229\u7528\u68c0\u7d22\u5230\u7684\u4fe1\u606f\u6765\u56de\u7b54\u95ee\u9898\u3002<\/p>\n<p>\u76f8\u6bd4\u4e4b\u4e0b\uff0cReSearch \u5c06\u641c\u7d22\u76f4\u63a5\u96c6\u6210\u5230\u63a8\u7406\u8fc7\u7a0b\u4e2d\u3002\u5b83\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\u5bf9\u6a21\u578b\u8fdb\u884c\u7aef\u5230\u7aef\u8bad\u7ec3\uff0c\u800c\u65e0\u9700\u5bf9\u63a8\u7406\u6b65\u9aa4\u8fdb\u884c\u4efb\u4f55\u76d1\u7763\u3002\u8bf8\u5982\u53cd\u601d\u9519\u8bef\u67e5\u8be2\u5e76\u8fdb\u884c\u7ea0\u6b63\u7b49\u884c\u4e3a\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u81ea\u7136\u800c\u7136\u5730\u51fa\u73b0\u3002<\/p>\n<p>[10] \u7406\u89e3 R1-Zero \u7c7b\u8bad\u7ec3<\/p>\n<p>26 Mar, Understanding R1-Zero-Like Training: A Critical Perspective, https:\/\/arxiv.org\/abs\/2503.20783<\/p>\n<p>\u672c\u6587\u65e8\u5728\u63a2\u7a76 DeepSeek-R1-Zero \u6a21\u578b\u91c7\u7528\u7eaf\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff08pure RL\uff09\u4e3a\u4f55\u80fd\u591f\u63d0\u5347\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p>\u4f5c\u8005\u53d1\u73b0\uff0c\u4e00\u4e9b\u57fa\u7840\u6a21\u578b\uff08\u5982 Qwen2.5\uff09\u5728\u672a\u7ecf\u4efb\u4f55\u5f3a\u5316\u5b66\u4e60\u7684\u60c5\u51b5\u4e0b\uff0c\u4fbf\u5df2\u7ecf\u8868\u73b0\u51fa\u8f83\u5f3a\u7684\u63a8\u7406\u80fd\u529b\uff0c\u751a\u81f3\u80fd\u591f\u5448\u73b0\u51fa\u6240\u8c13\u7684\u300c\u987f\u609f\u65f6\u523b\u300d\uff08Aha moment\uff09\u3002\u56e0\u6b64\uff0c\u8fd9\u79cd\u300c\u987f\u609f\u300d\u53ef\u80fd\u5e76\u975e\u5f3a\u5316\u5b66\u4e60\u6240\u8bf1\u53d1\uff0c\u800c\u662f\u6e90\u4e8e\u9884\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u7ee7\u627f\u3002\u8fd9\u4e00\u53d1\u73b0\u5bf9\u300c\u6df1\u5c42\u63a8\u7406\u884c\u4e3a\u662f\u7531\u5f3a\u5316\u5b66\u4e60\u5355\u72ec\u5f15\u5bfc\u5f62\u6210\u300d\u7684\u89c2\u70b9\u63d0\u51fa\u4e86\u8d28\u7591\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u672c\u6587\u8fd8\u63ed\u793a\u4e86 GRPO \u65b9\u6cd5\u4e2d\u5b58\u5728\u7684\u4e24\u4e2a\u504f\u5dee\u95ee\u9898\uff1a<\/p>\n<ul data-id=\"ua73dd4b-1KSI0S2k\">\n<li data-id=\"ld70c578-NU6GFxew\">\u56de\u7b54\u957f\u5ea6\u504f\u5dee\uff08Response-length bias\uff09\uff1aGRPO \u5728\u8ba1\u7b97\u4f18\u52bf\u503c\uff08advantage\uff09\u65f6\u5bf9\u56de\u7b54\u957f\u5ea6\u8fdb\u884c\u4e86\u5f52\u4e00\u5316\u3002\u8fd9\u5bfc\u81f4\u8f83\u957f\u7684\u9519\u8bef\u7b54\u6848\u6240\u53d7\u5230\u7684\u60e9\u7f5a\u53d8\u5c0f\uff0c\u4ece\u800c\u4fc3\u4f7f\u6a21\u578b\u503e\u5411\u4e8e\u751f\u6210\u5197\u957f\u4f46\u9519\u8bef\u7684\u56de\u7b54\u3002<\/li>\n<li data-id=\"ld70c578-MdXb2JBf\">\u9898\u76ee\u96be\u5ea6\u504f\u5dee\uff08Difficulty-level bias\uff09\uff1aGRPO \u540c\u65f6\u4f9d\u636e\u6bcf\u9053\u9898\u5bf9\u5e94\u5956\u52b1\u7684\u6807\u51c6\u5dee\u8fdb\u884c\u5f52\u4e00\u5316\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u65e0\u8bba\u9898\u76ee\u672c\u8eab\u662f\u7b80\u5355\u8fd8\u662f\u56f0\u96be\uff0c\u53ea\u8981\u5176\u5bf9\u5e94\u7684\u5956\u52b1\u65b9\u5dee\u8f83\u5c0f\uff0c\u5c31\u5bb9\u6613\u5728\u4f18\u5316\u4e2d\u88ab\u8d4b\u4e88\u8f83\u9ad8\u6743\u91cd\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u6539\u8fdb\u7684 GRPO \u53d8\u4f53 \u2014\u2014Dr. GRPO\u3002\u8be5\u65b9\u6cd5\u5728\u4f18\u52bf\u51fd\u6570\u8ba1\u7b97\u4e2d\u53d6\u6d88\u4e86\u5bf9\u56de\u7b54\u957f\u5ea6\u7684\u5f52\u4e00\u5316\u5904\u7406\uff0c\u540c\u65f6\u4e5f\u4e0d\u518d\u5728\u9898\u76ee\u7ea7\u522b\u4e0a\u5bf9\u5956\u52b1\u6807\u51c6\u5dee\u8fdb\u884c\u5f52\u4e00\u5316\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u53d8\u5f97\u66f4\u9ad8\u6548\uff0c\u540c\u65f6\u4e5f\u53ef\u4ee5\u6709\u6548\u907f\u514d\u751f\u6210\u5197\u957f\u4f46\u9519\u8bef\u7684\u7b54\u6848\u3002\u7279\u522b\u662f\u5728\u6a21\u578b\u7ed9\u51fa\u9519\u8bef\u56de\u7b54\u65f6\uff0c\u8be5\u65b9\u6cd5\u4e0d\u518d\u9f13\u52b1\u5176\u751f\u6210\u957f\u7bc7\u5e45\u7684\u4f4e\u8d28\u91cf\u6587\u672c\u3002<\/p>\n<p>[11] \u5728\u591a\u6837\u5316\u9886\u57df\u4e2d\u6269\u5c55\u57fa\u4e8e\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60<\/p>\n<p>31 Mar, Crossing the Reward Bridge: Expanding RL with Verifiable Rewards Across Diverse Domains, https:\/\/arxiv.org\/abs\/2503.23829<\/p>\n<p>DeepSeek-R1 \u53ca\u540e\u7eed\u591a\u6570\u63a8\u7406\u6a21\u578b\u4e3b\u8981\u5173\u6ce8\u6613\u4e8e\u9a8c\u8bc1\u9886\u57df\uff08\u5982\u7f16\u7a0b\u548c\u6570\u5b66\uff09\u7684\u5956\u52b1\u4fe1\u53f7\u3002\u672c\u6587\u5219\u63a2\u8ba8\u4e86\u5982\u4f55\u5c06\u8fd9\u4e9b\u65b9\u6cd5\u62d3\u5c55\u81f3\u66f4\u590d\u6742\u7684\u9886\u57df\uff0c\u5982\u533b\u5b66\u3001\u5316\u5b66\u3001\u5fc3\u7406\u5b66\u3001\u7ecf\u6d4e\u5b66\u53ca\u6559\u80b2\u5b66\u7b49 \u2014\u2014 \u8fd9\u4e9b\u9886\u57df\u7684\u56de\u7b54\u901a\u5e38\u4e3a\u81ea\u7531\u5f62\u5f0f\uff0c\u96be\u4ee5\u7b80\u5355\u5730\u7528\u300c\u6b63\u786e \/ \u9519\u8bef\u300d\u4e8c\u5143\u6807\u51c6\u8fdb\u884c\u5224\u65ad\u3002<\/p>\n<p>\u7814\u7a76\u53d1\u73b0\uff0c\u5373\u4fbf\u5728\u4e0a\u8ff0\u590d\u6742\u9886\u57df\u4e2d\uff0c\u82e5\u91c7\u7528\u4e13\u5bb6\u64b0\u5199\u7684\u53c2\u8003\u7b54\u6848\uff0c\u8bc4\u4f30\u53ef\u884c\u6027\u4ecd\u8d85\u51fa\u9884\u671f\u3002\u4e3a\u6784\u5efa\u5956\u52b1\u4fe1\u53f7\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65e0\u9700\u5bc6\u96c6\u9886\u57df\u6807\u6ce8\u7684\u300c\u751f\u6210\u5f0f\u8f6f\u8bc4\u5206\u65b9\u6cd5\u300d\uff08generative, soft-scoring method\uff09\uff0c\u4ece\u800c\u663e\u8457\u964d\u4f4e\u5bf9\u7279\u5b9a\u9886\u57df\u6807\u6ce8\u6570\u636e\u7684\u4f9d\u8d56\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202504\/21\/864e8811860a3f139ed943fa0d73770b15bb2b.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[12] \u5728\u7b80\u5316\u8bbe\u5b9a\u4e0b\u6269\u5c55\u5f3a\u5316\u5b66\u4e60\u89c4\u6a21<\/p>\n<p>31 Mar, Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model, https:\/\/arxiv.org\/abs\/2503.24290<\/p>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u4f5c\u8005\u63a2\u7d22\u4e86\u4e00\u79cd\u7528\u4e8e\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u4efb\u52a1\u8bad\u7ec3\u7684\u6781\u7b80\u5f3a\u5316\u5b66\u4e60\u8bbe\u7f6e\u3002\u4ed6\u4eec\u91c7\u7528\u4e86\u57fa\u7840\u7248\u672c\u7684\u6982\u7387\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\uff08vanilla PPO\uff09\uff0c\u800c\u975e DeepSeek-R1-Zero \u4e2d\u6240\u4f7f\u7528\u7684\u5e7f\u4e49\u968f\u673a\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\uff08GRPO\uff09\uff0c\u5e76\u7701\u7565\u4e86\u5f3a\u5316\u5b66\u4e60\u5bf9\u4eba\u7c7b\u53cd\u9988\uff08RLHF\uff09\u6d41\u7a0b\u4e2d\u5e38\u89c1\u7684 Kullback-Leibler \u6b63\u5219\u5316\u9879\uff08KL regularization\uff09\u3002<\/p>\n<p>\u6709\u8da3\u7684\u662f\uff0c\u7814\u7a76\u53d1\u73b0\uff0c\u8be5\u7b80\u5316\u8bbe\u7f6e\uff08\u5373 vanilla PPO \u8bad\u7ec3\u5668\u52a0\u4e0a\u57fa\u4e8e\u7b54\u6848\u6b63\u786e\u6027\u7684\u7b80\u5355\u4e8c\u5143\u5956\u52b1\u51fd\u6570\uff09\u5bf9\u4e8e\u8bad\u7ec3\u51fa\u5728\u63a8\u7406\u6027\u80fd\u548c\u751f\u6210\u957f\u5ea6\u4e24\u65b9\u9762\u90fd\u6709\u6240\u63d0\u5347\u7684\u6a21\u578b\u800c\u8a00\uff0c\u5df2\u8db3\u591f\u6709\u6548\u3002<\/p>\n<p>\u5728\u4f7f\u7528\u4e0e DeepSeek-R1-Zero \u76f8\u540c\u7684 Qwen-32B \u57fa\u5ea7\u6a21\u578b\u7684\u524d\u63d0\u4e0b\uff0c\u4f5c\u8005\u6240\u6784\u5efa\u7684\u6a21\u578b\u5728\u591a\u4e2a\u63a8\u7406\u57fa\u51c6\u4efb\u52a1\u4e0a\u5b9e\u73b0\u4e86\u8d85\u8d8a\u540e\u8005\u7684\u6027\u80fd\u8868\u73b0\uff0c\u540c\u65f6\u8bad\u7ec3\u6b65\u9aa4\u4ec5\u4e3a\u5176\u5341\u5206\u4e4b\u4e00\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202504\/21\/78f63d394fdc10b2492945a0aee0458958de07.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[13] \u91cd\u65b0\u5ba1\u89c6\u9884\u8bad\u7ec3\u9636\u6bb5\u4e2d\u7684\u53cd\u601d\u673a\u5236<\/p>\n<p>5 Apr, Rethinking Reflection in Pre-Training, https:\/\/arxiv.org\/abs\/2504.04022<\/p>\n<p>\u57fa\u4e8e DeepSeek-R1 \u8bba\u6587\u63d0\u51fa\u7684\u91cd\u8981\u53d1\u73b0 \u2014\u2014 \u5373\u5728\u57fa\u7840\u6a21\u578b\u4e0a\u76f4\u63a5\u5e94\u7528\u7eaf\u5f3a\u5316\u5b66\u4e60\uff08pure RL\uff09\u4ee5\u6fc0\u53d1\u63a8\u7406\u80fd\u529b \u2014\u2014 \u6211\u4eec\u66fe\u63a8\u6d4b\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u4e3b\u8981\u6e90\u81ea\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002\u7136\u800c\uff0c\u672c\u8bba\u6587\u7ed9\u51fa\u4e86\u4e00\u4e2a\u51fa\u4eba\u610f\u6599\u7684\u8f6c\u6298\uff1a\u7814\u7a76\u53d1\u73b0\uff0c\u81ea\u6211\u7ea0\u6b63\uff08self-correction\uff09\u80fd\u529b\u5b9e\u9645\u4e0a\u5728\u9884\u8bad\u7ec3\uff08pre-training\uff09\u9636\u6bb5\u5c31\u5df2\u5f00\u59cb\u663e\u73b0\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u4f5c\u8005\u901a\u8fc7\u5728\u4efb\u52a1\u4e2d\u5f15\u5165\u4eba\u4e3a\u8bbe\u8ba1\u7684\u9519\u8bef\u601d\u7ef4\u94fe\uff0c\u6765\u8861\u91cf\u6a21\u578b\u662f\u5426\u5177\u5907\u8bc6\u522b\u5e76\u4fee\u6b63\u8fd9\u4e9b\u9519\u8bef\u7684\u80fd\u529b\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u65e0\u8bba\u662f\u663e\u5f0f\u53cd\u601d\uff08explicit reflection\uff09\u8fd8\u662f\u9690\u5f0f\u4fee\u6b63\uff08implicit correction\uff09\uff0c\u8fd9\u4e9b\u80fd\u529b\u90fd\u4f1a\u5728\u9884\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u9010\u6b65\u81ea\u7136\u6d8c\u73b0\u3002\u8fd9\u4e00\u73b0\u8c61\u5728\u591a\u79cd\u4efb\u52a1\u9886\u57df\u548c\u4e0d\u540c\u6a21\u578b\u89c4\u6a21\u4e2d\u5747\u6709\u4f53\u73b0\uff0c\u751a\u81f3\u5728\u9884\u8bad\u7ec3\u65e9\u671f\u7684\u6a21\u578b\u68c0\u67e5\u70b9\uff08checkpoint\uff09\u4e0a\u4e5f\u53ef\u89c2\u5bdf\u5230\u521d\u6b65\u7684\u81ea\u6211\u7ea0\u6b63\u8ff9\u8c61\uff0c\u4e14\u968f\u7740\u9884\u8bad\u7ec3\u8ba1\u7b97\u91cf\u7684\u589e\u52a0\uff0c\u8be5\u80fd\u529b\u6301\u7eed\u589e\u5f3a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s5.51cto.com\/oss\/202504\/21\/844650a256b379bb296805497e7efb184ae003.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[14] \u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8fdb\u884c\u7b80\u6d01\u63a8\u7406<\/p>\n<p>7 Apr, Concise Reasoning via Reinforcement Learning, https:\/\/arxiv.org\/abs\/2504.05185<\/p>\n<p>\u4f17\u6240\u5468\u77e5\uff0c\u5177\u5907\u63a8\u7406\u80fd\u529b\u7684\u8bed\u8a00\u6a21\u578b\u901a\u5e38\u4f1a\u751f\u6210\u66f4\u957f\u7684\u8f93\u51fa\u6587\u672c\uff0c\u8fd9\u65e0\u7591\u63d0\u9ad8\u4e86\u8ba1\u7b97\u6210\u672c\u3002\u6700\u65b0\u8fd9\u7bc7\u8bba\u6587\u6307\u51fa\uff0c\u8fd9\u79cd\u884c\u4e3a\u5e76\u975e\u6e90\u4e8e\u66f4\u957f\u56de\u7b54\u672c\u8eab\u5bf9\u4e8e\u63d0\u9ad8\u51c6\u786e\u6027\u6709\u6240\u5e2e\u52a9\uff0c\u800c\u662f\u5728\u4e8e\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u5185\u5728\u504f\u5411\u6240\u81f4\u3002<\/p>\n<p>\u7814\u7a76\u8868\u660e\uff0c\u5728 RL \u8fc7\u7a0b\u4e2d\uff0c\u5f53\u667a\u80fd\u4f53\uff08Agent\uff09\u56e0\u9519\u8bef\u56de\u7b54\u800c\u83b7\u5f97\u8d1f\u5956\u52b1\u65f6\uff0c\u7b56\u7565\u4f18\u5316\u7b97\u6cd5 \u2014\u2014 \u7279\u522b\u662f\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\uff08PPO\uff09\u2014\u2014 \u503e\u5411\u4e8e\u9f13\u52b1\u6a21\u578b\u751f\u6210\u66f4\u957f\u7684\u56de\u7b54\u3002\u5177\u4f53\u800c\u8a00\uff0cPPO \u7684\u635f\u5931\u8ba1\u7b97\u673a\u5236\u4f7f\u5f97\u5728\u8d1f\u5956\u52b1\u7684\u60c5\u5f62\u4e0b\uff0c\u968f\u7740\u751f\u6210\u6587\u672c\u957f\u5ea6\u7684\u589e\u52a0\uff0c\u5e73\u5747\u6bcf\u4e2a token \u7684\u635f\u5931\u4f1a\u9010\u6e10\u53d8\u5c0f\u3002\u56e0\u6b64\uff0c\u5373\u4fbf\u6a21\u578b\u4f9d\u65e7\u7ed9\u51fa\u9519\u8bef\u7b54\u6848\uff0c\u4f46\u751f\u6210\u66f4\u957f\u7684\u56de\u590d\u5f62\u5f0f\u4f1a\u5728\u6570\u5b66\u4e0a\u300c\u7a00\u91ca\u300d\u6bcf\u4e2a token \u6240\u627f\u53d7\u7684\u60e9\u7f5a\u3002<\/p>\n<p>\u6362\u53e5\u8bdd\u8bf4\uff0c\u957f\u6587\u672c\u7684\u7ed3\u6784\u63a9\u76d6\u4e86\u6574\u4f53\u9519\u8bef\uff0c\u4ece\u800c\u5728\u635f\u5931\u51fd\u6570\u4e0a\u8868\u73b0\u4e3a\u4f18\u5316\u65b9\u5411\u4e0a\u7684\u300c\u6539\u8fdb\u300d\uff0c\u5373\u4f7f\u8fd9\u4e9b\u989d\u5916 token \u5bf9\u5b9e\u9645\u63a8\u7406\u7ed3\u679c\u6ca1\u6709\u5e2e\u52a9\u3002\u4e8e\u662f\u6a21\u578b\u300c\u5b66\u4f1a\u300d\u4e86\u901a\u8fc7\u62c9\u957f\u8f93\u51fa\u957f\u5ea6\u6765\u51cf\u8f7b\u60e9\u7f5a\uff0c\u800c\u975e\u771f\u6b63\u63d0\u5347\u7b54\u6848\u7684\u6b63\u786e\u6027\u3002<\/p>\n<p>\u4e0d\u8fc7\u9700\u8981\u5f3a\u8c03\u7684\u662f\uff0c\u8fd9\u4e00\u73b0\u8c61\u4ec5\u5728\u4f7f\u7528 PPO \u7684\u8bad\u7ec3\u6d41\u7a0b\u4e0b\u88ab\u89c2\u5bdf\u5230\uff1a \u300c\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5f53\u524d\u7684\u5206\u6790\u5e76\u4e0d\u9002\u7528\u4e8e GRPO\uff0c\u5bf9\u8be5\u7c7b\u65b9\u6cd5\u7684\u4e25\u683c\u5206\u6790\u5c06\u5728\u672a\u6765\u5c55\u5f00\u3002\u300d<\/p>\n<p>\u6b64\u5916\uff0c\u7814\u7a76\u8005\u8fd8\u53d1\u73b0\uff0c\u901a\u8fc7\u5f15\u5165\u7b2c\u4e8c\u9636\u6bb5\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff08\u5373\u5bf9\u90e8\u5206\u53ef\u89e3\u7b54\u7684\u95ee\u9898\u8fdb\u884c\u5c11\u91cf RL \u5fae\u8c03\uff09\uff0c\u4e0d\u4ec5\u53ef\u4ee5\u7f29\u77ed\u8f93\u51fa\u957f\u5ea6\uff0c\u800c\u4e14\u6709\u65f6\u8fd8\u80fd\u7ef4\u6301\u751a\u81f3\u63d0\u5347\u6a21\u578b\u7684\u51c6\u786e\u7387\u3002\u8fd9\u4e00\u53d1\u73b0\u5bf9\u4e8e\u6a21\u578b\u90e8\u7f72\u7684\u6548\u7387\u4f18\u5316\u5177\u6709\u91cd\u8981\u610f\u4e49\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s9.51cto.com\/oss\/202504\/21\/06272171920e3a4a13626497a0f7776996328b.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>[15] \u51b7\u9759\u770b\u5f85\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7684\u8fdb\u5c55<\/p>\n<p>9 Apr, A Sober Look at Progress in Language Model Reasoning: Pitfalls and Paths to Reproducibility, https:\/\/arxiv.org\/abs\/2504.07086<\/p>\n<p>\u672c\u8bba\u6587\u5bf9\u8fd1\u671f\u5173\u4e8e\u5f3a\u5316\u5b66\u4e60\u53ef\u63d0\u5347\u84b8\u998f\u8bed\u8a00\u6a21\u578b\uff08distilled language models\uff09\u6027\u80fd\u7684\u4e3b\u5f20\u8fdb\u884c\u4e86\u66f4\u4e3a\u5ba1\u614e\u7684\u5ba1\u89c6\uff0c\u5c24\u5176\u805a\u7126\u4e8e\u57fa\u4e8e DeepSeek-R1 \u6a21\u578b\u7684\u7814\u7a76\u3002<\/p>\n<p>\u4f8b\u5982\uff0c\u6211\u66fe\u8ba8\u8bba\u8fc7 2024 \u5e74 3 \u6708 20 \u65e5\u53d1\u5e03\u7684\u300aReinforcement Learning for Reasoning in Small LLMs: What Works and What Doesn&#8217;t\u300b\u4e00\u6587\uff0c\u5176\u53d1\u73b0 RL \u5bf9\u84b8\u998f\u6a21\u578b\u5177\u6709\u826f\u597d\u6548\u679c\u3002\u6b64\u5916\uff0c\u5728 DeepSeek-R1 \u7684\u76f8\u5173\u8bba\u6587\u4e2d\u4e5f\u6307\u51fa\uff0c\u91c7\u7528 RL \u65b9\u6cd5\u53ef\u4ee5\u8fdb\u4e00\u6b65\u663e\u8457\u63d0\u5347\u8fd9\u4e9b\u84b8\u998f\u6a21\u578b\u7684\u6027\u80fd\u3002\u56e0\u6b64\uff0c\u8be5\u7814\u7a76\u503e\u5411\u4e8e\u8fdb\u4e00\u6b65\u63a2\u7d22\u76f8\u5173\u673a\u5236\uff0c\u5e76\u5728\u672c\u6587\u4e2d\u4ec5\u5c55\u793a\u901a\u8fc7\u7b80\u5355\u7684\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u8bad\u7ec3\u5f97\u5230\u7684\u84b8\u998f\u6a21\u578b\u7ed3\u679c\u3002<\/p>\n<p>\u7136\u800c\uff0c\u4e0e\u65e9\u671f\u7814\u7a76\u62a5\u544a\u7684 RL \u6548\u679c\u5927\u5e45\u63d0\u5347\u7ed3\u8bba\u4e0d\u540c\uff0c\u672c\u8bba\u6587\u6307\u51fa\uff0c\u90e8\u5206\u6027\u80fd\u63d0\u5347\u53ef\u80fd\u53ea\u662f\u968f\u673a\u6ce2\u52a8\u5e26\u6765\u7684\u5047\u8c61\u3002\u4f5c\u8005\u901a\u8fc7\u5b9e\u9a8c\u8bc1\u660e\uff0c\u5728\u5c0f\u578b\u57fa\u51c6\u6d4b\u8bd5\u4efb\u52a1\uff08\u5982 AIME24\uff09\u4e2d\uff0c\u4ec5\u968f\u673a\u79cd\u5b50\u7684\u6539\u53d8\u5c31\u53ef\u80fd\u5bfc\u81f4\u6027\u80fd\u5206\u6570\u51fa\u73b0\u51e0\u4e2a\u767e\u5206\u70b9\u7684\u6ce2\u52a8\uff0c\u8bf4\u660e\u7ed3\u679c\u5b58\u5728\u6781\u9ad8\u7684\u4e0d\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u5f53\u4f7f\u7528\u66f4\u4e3a\u53d7\u63a7\u4e0e\u6807\u51c6\u5316\u7684\u5b9e\u9a8c\u8bbe\u7f6e\u6765\u8bc4\u4f30 RL \u6a21\u578b\u65f6\uff0c\u5176\u6027\u80fd\u63d0\u5347\u901a\u5e38\u663e\u8457\u5c0f\u4e8e\u539f\u5148\u62a5\u9053\uff0c\u4e14\u591a\u6570\u60c5\u51b5\u4e0b\u7f3a\u4e4f\u7edf\u8ba1\u5b66\u663e\u8457\u6027\u3002\u5c3d\u7ba1\u67d0\u4e9b\u91c7\u7528 RL \u8bad\u7ec3\u7684\u6a21\u578b\u786e\u5b9e\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u8868\u73b0\u51fa\u9002\u5ea6\u63d0\u5347\uff0c\u4f46\u8fd9\u4e9b\u63d0\u5347\u901a\u5e38\u4e0d\u5982\u76d1\u7763\u5fae\u8c03\u5f3a\u6548\uff0c\u4e14\u5f80\u5f80\u7f3a\u4e4f\u8de8\u57fa\u51c6\u8fc1\u79fb\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p>\u7efc\u4e0a\u6240\u8ff0\uff0c\u5c3d\u7ba1 RL \u5728\u67d0\u4e9b\u60c5\u5f62\u4e0b\u53ef\u80fd\u5bf9\u5c0f\u578b\u84b8\u998f\u8bed\u8a00\u6a21\u578b\u6709\u6240\u52a9\u76ca\uff0c\u672c\u6587\u4e3b\u5f20\u5176\u5b9e\u9645\u76ca\u5904\u88ab\u8fc7\u5ea6\u6e32\u67d3\uff0c\u7814\u7a76\u754c\u4e9f\u9700\u66f4\u4e3a\u4e25\u8c28\u7684\u8bc4\u4f30\u6807\u51c6\uff0c\u4ee5\u5398\u6e05\u771f\u6b63\u6709\u6548\u7684\u624b\u6bb5\u4e0e\u6539\u8fdb\u8def\u5f84\u3002<\/p>\n<\/div>\n<p><img data-dominant-color=\"f1f3f5\" data-has-transparency=\"false\" style=\"--dominant-color: #f1f3f5;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-medium wp-image-25538\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/d93fcd624c6f786ab49529c1b7241acd9a9859-300x207.webp\" alt=\"\" width=\"300\" height=\"207\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/d93fcd624c6f786ab49529c1b7241acd9a9859-300x207.webp 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/d93fcd624c6f786ab49529c1b7241acd9a9859-1024x706.webp 1024w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/d93fcd624c6f786ab49529c1b7241acd9a9859-150x103.webp 150w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/d93fcd624c6f786ab49529c1b7241acd9a9859-768x530.webp 768w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/04\/d93fcd624c6f786ab49529c1b7241acd9a9859.webp 1080w\" sizes=\"auto, (max-width: 300px) 100vw, 300px\" \/><\/p>\n<p>\u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_25536\" class=\"pvc_stats total_only  \" data-element-id=\"25536\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u53ea\u9760\u6a21\u578b\u5c3a\u5bf8\u53d8\u5927\u5df2\u7ecf\u4e0d\u884c\u4e86\uff1f\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u63a8\u7406\u9700\u8981\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6765\u300c\u52a0 buff\u300d\u3002 \u8457\u540d AI \u7814\u7a76 [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_25536\" class=\"pvc_stats total_only  \" data-element-id=\"25536\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":56,"featured_media":25539,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,17,20,80],"tags":[124,887,820],"class_list":["post-25536","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-17","category-20","category-80","tag-ai","tag-887","tag-820"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25536","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=25536"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25536\/revisions"}],"predecessor-version":[{"id":25540,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25536\/revisions\/25540"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/25539"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=25536"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=25536"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=25536"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}