{"id":40242,"date":"2025-04-20T22:02:55","date_gmt":"2025-04-20T14:02:55","guid":{"rendered":"https:\/\/17aitech.com\/?p=40242"},"modified":"2025-04-20T22:02:55","modified_gmt":"2025-04-20T14:02:55","slug":"%e6%89%a9%e6%95%a3llm%e6%8e%a8%e7%90%86%e7%94%a8%e4%b8%8a%e7%b1%bbgrpo%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%ef%bc%81%e4%bc%98%e4%ba%8e%e5%8d%95%e7%8b%acsft%ef%bc%8cucla%e3%80%81meta%e6%96%b0%e6%a1%86","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=40242","title":{"rendered":"\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-20-7\" target=\"_blank\">\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90<\/a><\/p>\n<blockquote>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4e0d\u518d\u662f AR\uff08\u81ea\u56de\u5f52\uff09\u7684\u4e13\u5c5e\u3002\u6269\u6563\u6a21\u578b\u73b0\u5728\u4e5f\u80fd\u300c\u52a8\u8111\u5b50\u300d\uff0c\u65b0\u6846\u67b6 d1 \u8ba9\u5b83\u4eec\u5b66\u4f1a\u4e86\u89e3\u6570\u5b66\u3001\u61c2\u903b\u8f91\u3001\u4f1a\u601d\u8003\u3002<\/p>\n<\/blockquote>\n<p>\u5f53\u524d\uff0c\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u65b9\u6cd5\u5728\u6700\u8fd1\u6a21\u578b\u7684\u63a8\u7406\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6539\u8fdb\uff0c\u6bd4\u5982 DeepSeek-R1\u3001Kimi K1.5\uff0c\u663e\u793a\u4e86\u5c06 RL \u76f4\u63a5\u7528\u4e8e\u57fa\u7840\u6a21\u578b\u53ef\u4ee5\u53d6\u5f97\u5ab2\u7f8e OpenAI o1 \u7684\u6027\u80fd\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0c\u57fa\u4e8e RL \u7684\u540e\u8bad\u7ec3\u8fdb\u5c55\u4e3b\u8981\u53d7\u9650\u4e8e\u81ea\u56de\u5f52\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\uff0c\u5b83\u4eec\u901a\u8fc7\u4ece\u5de6\u5230\u53f3\u7684\u5e8f\u5217\u63a8\u7406\u6765\u8fd0\u884c\u3002<\/p>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c\u79bb\u6563\u6269\u6563\u5927\u8bed\u8a00\u6a21\u578b\uff08dLLM\uff09\u6210\u4e3a\u6709\u6f5c\u529b\u7684\u8bed\u8a00\u5efa\u6a21\u7684\u975e\u81ea\u56de\u5f52\u66ff\u4ee3\u3002\u4e0d\u50cf\u4ee5\u56e0\u679c\u65b9\u5f0f\u9010 token \u751f\u6210\u6587\u672c\u7684\u81ea\u56de\u5f52\u6a21\u578b\u90a3\u6837\uff0cdLLM \u901a\u8fc7\u8fed\u4ee3\u53bb\u566a\u8fc7\u7a0b\u751f\u6210\u6587\u672c\uff0c\u5728\u591a\u6b65\u9aa4\u64cd\u4f5c\u4e2d\u4f18\u5316\u5e8f\u5217\u7684\u540c\u65f6\u5e76\u901a\u8fc7\u53cc\u5411\u6ce8\u610f\u529b\u5229\u7528\u8fc7\u53bb\u548c\u672a\u6765\u7684\u4e0a\u4e0b\u6587\u3002\u5176\u4e2d\uff0cLLaDA \u7b49\u5f00\u653e\u7684\u63a9\u7801 dLLM \u5b9e\u73b0\u4e86\u5ab2\u7f8e\u540c\u5c3a\u5bf8\u81ea\u56de\u5f52\u6a21\u578b\u7684\u6027\u80fd\uff0c\u800c Mercury \u7b49\u95ed\u6e90 dLLM \u8fdb\u4e00\u6b65\u5c55\u73b0\u4e86\u51fa\u8272\u7684\u63a8\u7406\u5ef6\u8fdf\u3002<\/p>\n<p>\u7136\u800c\uff0c\u9876\u7ea7\u7684\u5f00\u6e90 dLLM \u5e76\u6ca1\u6709\u4f7f\u7528 RL \u540e\u8bad\u7ec3\uff0c\u4f7f\u5f97\u8fd9\u4e00\u6709\u6f5c\u529b\u7684\u7814\u7a76\u65b9\u5411\u8fd8\u6709\u5f88\u5927\u7684\u6316\u6398\u7a7a\u95f4\u3002\u8fd9\u4e00\u8303\u5f0f\u8f6c\u53d8\u5f15\u51fa\u4e86\u91cd\u8981\u7684\u95ee\u9898\uff1aRL \u540e\u8bad\u7ec3\u5982\u4f55\u5728\u975e\u81ea\u56de\u5f52\u4e0a\u4e0b\u6587\u4e2d\u9ad8\u6548\u5730\u5b9e\u73b0\uff1f<\/p>\n<p>RL \u7b97\u6cd5\u9002\u5e94\u63a9\u7801 dLLM \u9762\u4e34\u4e00\u4e9b\u72ec\u7279\u7684\u6311\u6218\uff0c\u539f\u56e0\u5728\u4e8e\u81ea\u56de\u5f52\u6a21\u578b\u91c7\u7528\u7684\u5df2\u6709\u65b9\u6cd5\uff08\u5982 PPO\u3001GRPO\uff09\u901a\u8fc7\u8ba1\u7b97\u751f\u6210\u5e8f\u5217\u7684\u5bf9\u6570\u6982\u7387\u6765\u4f30\u8ba1\u548c\u4f18\u5316\u7b56\u7565\u5206\u5e03\uff0c\u5bfc\u81f4\u65e0\u6cd5\u76f4\u63a5\u5e94\u7528\u4e8e dLLM\u3002\u867d\u7136\u8fd9\u79cd\u8ba1\u7b97\u5728\u81ea\u56de\u5f52\u6a21\u578b\u4e2d\u901a\u8fc7\u5e8f\u5217\u56e0\u5f0f\u5206\u89e3\u5f88\u5bb9\u6613\u5b9e\u73b0\uff0c\u4f46 dLLM \u7531\u4e8e\u5b83\u4eec\u7684\u8fed\u4ee3\u3001\u975e\u5e8f\u5217\u751f\u6210\u8fc7\u7a0b\u800c\u7f3a\u4e4f\u8fd9\u79cd\u81ea\u7136\u5206\u89e3\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c<strong>\u6765\u81ea UCLA \u548c Meta AI \u7684\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e00\u4e2a\u4e24\u9636\u6bb5\u540e\u8bad\u7ec3\u6846\u67b6 d1\uff0c\u4ece\u800c\u53ef\u4ee5\u5728\u63a9\u7801 dLLM \u4e2d\u8fdb\u884c\u63a8\u7406<\/strong>\u3002\u5728\u7b2c\u4e00\u9636\u6bb5\uff0c\u6a21\u578b\u5728\u9ad8\u8d28\u91cf\u63a8\u7406\u8f68\u8ff9\u4e2d\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff1b\u5728\u7b2c\u4e8c\u5373 RL \u9636\u6bb5\uff0c\u7814\u7a76\u8005\u5f15\u5165\u4e86\u7528\u4e8e\u63a9\u7801 dLLM \u7684\u65b0\u9896\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5 diffu-GRPO\uff0c\u5b83\u5229\u7528\u63d0\u51fa\u7684\u9ad8\u6548\u4e00\u6b65\uff08one-step\uff09\u5bf9\u6570\u6982\u7387\u4f30\u8ba1\u5728 GRPO \u7684\u57fa\u7840\u4e0a\u521b\u5efa\u3002<\/p>\n<p>\u7814\u7a76\u8005\u8868\u793a\uff0c\u4ed6\u4eec\u7684\u4f30\u8ba1\u5668\u5229\u7528\u4e86\u968f\u673a\u63d0\u793a\u8bcd\u63a9\u7801\uff0c\u4f5c\u4e3a\u7b56\u7565\u4f18\u5316\u7684\u4e00\u79cd\u6b63\u5219\u5316\uff0c\u4f7f\u5f97\u53ef\u4ee5\u6269\u5c55 per batch \u7684\u68af\u5ea6\u66f4\u65b0\u6570\u91cf\u5e76\u51cf\u5c11 RL \u8bad\u7ec3\u6240\u9700\u7684\u5728\u7ebf\u751f\u6210\u6570\u91cf\u3002\u8fd9\u5c06\u6781\u5927\u5730\u964d\u4f4e\u8ba1\u7b97\u65f6\u95f4\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png\"><\/a><\/section>\n<ul>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1ad1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2504.12216<\/p>\n<\/li>\n<li>\n<p>\u9879\u76ee\u4e3b\u9875\uff1ahttps:\/\/dllm-reasoning.github.io\/<\/p>\n<\/li>\n<li>\n<p>GitHub \u5730\u5740\uff1ahttps:\/\/github.com\/dllm-reasoning\/d1<\/p>\n<\/li>\n<\/ul>\n<p>\u5728\u5b9e\u9a8c\u90e8\u5206\uff0c\u7814\u7a76\u8005\u4f7f\u7528 LLaDA-8B-Instruct \u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\u5b9e\u4f8b\u5316 d1\u3002\u4ed6\u4eec\u5c06 d1-LLaDA \u7684\u6027\u80fd\u4e0e\u57fa\u7840 LLaDA \u6a21\u578b\u4ee5\u53ca\u4ec5\u4f7f\u7528 SFT \u548c\u4ec5\u4f7f\u7528 diffu-GRPO \u8bad\u7ec3\u7684 LLaDA \u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\u3002\u7ed3\u679c\u8868\u660e\uff0cd1 \u5728\u56db\u4e2a\u6570\u5b66\u548c\u903b\u8f91\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u59cb\u7ec8\u4f18\u4e8e\u57fa\u7840\u6a21\u578b\uff0c\u5982\u4e0b\u56fe 1 \u6240\u793a\u3002d1-LLaDA \u540c\u6837\u4f18\u4e8e\u4ec5\u4f7f\u7528 SFT \u65b9\u6cd5\u548c\u4ec5\u4f7f\u7528 diffu-GRPO \u65b9\u6cd5\u7684\u6a21\u578b\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-93b93e5c168e8ac061ea2a28a9fa7e82.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-93b93e5c168e8ac061ea2a28a9fa7e82.png\"><\/a><\/section>\n<p><strong>\u65b9\u6cd5\u6982\u89c8<\/strong><\/p>\n<p>d1 \u662f\u4e00\u4e2a\u4e24\u9636\u6bb5\u6846\u67b6\uff0c\u901a\u8fc7\u4f9d\u6b21\u7ed3\u5408\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u5728\u7ebf\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6765\u589e\u5f3a\u9884\u8bad\u7ec3\u63a9\u7801 dLLMs \u7684\u63a8\u7406\u6027\u80fd\u3002<\/p>\n<p>\u5176\u4e2d\uff0c\u5728\u7ebf\u5f3a\u5316\u5b66\u4e60\uff08\u7279\u522b\u662f GRPO \u7b97\u6cd5\uff09\u5df2\u88ab\u8bc1\u660e\u80fd\u6709\u6548\u63d0\u5347\u79bb\u7ebf\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u6027\u80fd\u3002\u7136\u800c\uff0cGRPO \u7684\u5b66\u4e60\u7b56\u7565\u5e76\u4e0d\u80fd\u76f4\u63a5\u6cdb\u5316\u5230 dLLMs\u3002<\/p>\n<p>GRPO \u7684\u76ee\u6807\u51fd\u6570\uff08\u5982\u516c\u5f0f 3 \u6240\u793a\uff09\u9700\u8981\u540c\u65f6\u8ba1\u7b97\u5f53\u524d\u7b56\u7565 \u03c0_\u03b8 \u548c\u65e7\u7b56\u7565 \u03c0_\u03b8old \u5728\u4ee5\u4e0b\u4e24\u4e2a\u5c42\u9762\u7684\uff08\u5bf9\u6570\uff09\u4f3c\u7136\u6bd4\uff1a<\/p>\n<ol>\n<li>\n<p>token \u5c42\u9762\uff08\u7528\u4e8e\u4f18\u52bf\u6743\u91cd\u8ba1\u7b97\uff09\uff1b<\/p>\n<\/li>\n<li>\n<p>\u5e8f\u5217\u5c42\u9762\uff08\u7528\u4e8e\u53cd\u5411 KL \u6563\u5ea6\u9879\uff09\u3002<\/p>\n<\/li>\n<\/ol>\n<p>\u6838\u5fc3\u95ee\u9898\u5728\u4e8e\uff1a\u7814\u7a76\u8005\u9700\u8981\u9ad8\u6548\u8ba1\u7b97 dLLMs \u751f\u6210\u5185\u5bb9\u7684\u9010 token \u5bf9\u6570\u6982\u7387\u548c\u5e8f\u5217\u5bf9\u6570\u6982\u7387\u3002<\/p>\n<p>\u81ea\u56de\u5f52\uff08AR\uff09\u6a21\u578b\uff0c\u5982 Transformer\uff0c\u76f4\u63a5\u5bf9\u6bcf\u4e2a token \u7684\u5bf9\u6570\u6982\u7387\u8fdb\u884c\u5efa\u6a21\uff0c\u5e76\u4e14\u53ef\u4ee5\u901a\u8fc7\u94fe\u5f0f\u6cd5\u5219\u4f7f\u7528\u4e00\u6b21\u524d\u5411\u4f20\u9012\u8f7b\u677e\u8ba1\u7b97\u51fa\u5e8f\u5217\u7ea7\u522b\u7684\u5bf9\u6570\u6982\u7387<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-34f4427334d73eab47e9af59dddf2808.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-34f4427334d73eab47e9af59dddf2808.png\"><\/a>\u3002\u540c\u6837\uff0cKL \u9879\u53ef\u4ee5\u5206\u89e3\u4e3a<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a90e8c067a6d2d93596c8a6cb27d97ec.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a90e8c067a6d2d93596c8a6cb27d97ec.png\"><\/a>\u3002<\/p>\n<p>\u4e0e AR \u6a21\u578b\u4e0d\u540c\uff0cdLLMs \u4e0d\u9075\u5faa\u5e8f\u5217\u5bf9\u6570\u6982\u7387\u7684\u987a\u5e8f\u5206\u89e3\u3002\u540c\u65f6\uff0c\u6bcf\u4e2a token \u7684\u5bf9\u6570\u6982\u7387\u8ba1\u7b97\u6210\u672c\u4e5f\u5f88\u9ad8\uff0c\u56e0\u4e3a\u89e3\u7801\u8fc7\u7a0b\u4e2d\u9700\u8981\u591a\u6b21\u8c03\u7528\u63a9\u7801\u9884\u6d4b\u5668 f_\u03b8\u3002\u57fa\u4e8e\u6b64\uff0c\u8be5\u7814\u7a76\u63d0\u51fa\u4e86\u4e00\u4e2a\u9ad8\u6548\u7684\u5bf9\u6570\u6982\u7387\u4f30\u8ba1\u5668\u3002<\/p>\n<p>\u5bf9\u4e8e\u5e8f\u5217\u5bf9\u6570\u6982\u7387\uff0c\u8be5\u7814\u7a76\u4f7f\u7528\u5747\u573a\u8fd1\u4f3c\u65b9\u6cd5\uff0c\u5c06\u5176\u5206\u89e3\u4e3a\u72ec\u7acb\u7684\u6bcf\u4e2a token \u5bf9\u6570\u6982\u7387\u7684\u4e58\u79ef\u3002<\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e2a token \u7684\u5bf9\u6570\u6982\u7387\uff0c\u8be5\u7814\u7a76\u5f15\u5165\u4e86\u4e00\u79cd\u4f30\u8ba1\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u4ec5\u8c03\u7528\u4e00\u6b21 f_\u03b8\u3002<\/p>\n<p>\u57fa\u4e8e\u65b0\u5f15\u5165\u7684\u5bf9\u6570\u6982\u7387\u4f30\u8ba1\u5668\uff0c\u8be5\u7814\u7a76\u5c06 GRPO \u6269\u5c55\u5230\u63a9\u7801 dLLMs\uff0c\u63a8\u5bfc\u51fa diffu-GRPO \u7684\u635f\u5931\u51fd\u6570\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-9539c9159b5ea97edc50f212c39653c8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-9539c9159b5ea97edc50f212c39653c8.png\"><\/a><\/section>\n<p>\u7b97\u6cd5\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-8c6a787a96608c8e04efdfda1ede9c19.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-8c6a787a96608c8e04efdfda1ede9c19.png\"><\/a><\/section>\n<p><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/p>\n<p>\u8868 1 \u62a5\u544a\u4e86\u57fa\u7ebf\u6a21\u578b LLaDA-8B-Instruct \u4e0e\u91c7\u7528\u4e0d\u540c\u540e\u8bad\u7ec3\u4f18\u5316\u65b9\u6848\u7684\u6a21\u578b\uff0c\u5728\u56db\u9879\u4efb\u52a1\u4e0a\u7684\u96f6\u6837\u672c\u6027\u80fd\u5bf9\u6bd4\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b76978bee6e48f84020f00f3564c42b4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b76978bee6e48f84020f00f3564c42b4.png\"><\/a><\/section>\n<p>\u56fe 3 \u7ed8\u5236\u4e86\u6709\u6548 token \u7684\u5e73\u5747\u6570\u91cf\uff1a<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-e538e23b76ee0bb2c2da75f243ec8b11.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-e538e23b76ee0bb2c2da75f243ec8b11.png\"><\/a><\/section>\n<p>\u57fa\u4e8e\u5b9e\u9a8c\uff0c\u8be5\u7814\u7a76\u5f97\u51fa\u4ee5\u4e0b\u4e3b\u8981\u53d1\u73b0\uff1a<\/p>\n<p>diffu-GRPO \u5728\u6240\u6709 12 \u79cd\u8bbe\u7f6e\u4e2d\u90fd\u4e00\u81f4\u4f18\u4e8e\u57fa\u7840\u7684 LLaDA \u548c SFT\uff08\u76d1\u7763\u5f0f\u5fae\u8c03\uff09\u3002diffu-GRPO \u548c SFT \u90fd\u76f8\u8f83\u4e8e LLaDA-8B-Instruct \u57fa\u7ebf\u6709\u6240\u63d0\u5347\uff0c\u4f46 diffu-GRPO \u663e\u793a\u51fa\u66f4\u6301\u7eed\u4e14\u5e45\u5ea6\u66f4\u5927\u7684\u589e\u76ca\u3002\u5177\u4f53\u6765\u8bf4\uff0cdiffu-GRPO \u5728\u6240\u6709 12 \u79cd\u8bbe\u7f6e\u4e2d\u90fd\u4f18\u4e8e LLaDA-8B-Instruct \u548c SFT\uff0c\u800c SFT \u4ec5\u5728\u5176\u4e2d\u7684 7 \u79cd\u8bbe\u7f6e\u4e2d\u4f18\u4e8e LLaDA-8B-Instruct\uff0c\u8fd9\u8868\u660e<strong> diffu-GRPO \u76f8\u6bd4\u4e8e\u5355\u72ec\u7684 SFT \u5b9e\u73b0\u4e86\u66f4\u5f3a\u7684\u6574\u4f53\u6027\u80fd\u63d0\u5347\u3002<\/strong><\/p>\n<p>LLaDA+diffu-GRPO \u5728\u6240\u6709\u8bbe\u7f6e\u4e2d\u90fd\u4f18\u4e8e\u57fa\u7840\u7684 LLaDA-8B-Instruct \u6a21\u578b\uff0c\u800c d1-LLaDA \u5728\u6bcf\u79cd\u60c5\u51b5\u4e0b\u90fd\u8d85\u8fc7\u4e86 LLaDA+SFT\u3002\u8fd9\u8868\u660e\uff0c\u65e0\u8bba\u521d\u59cb\u5316\u662f\u6765\u81ea\u9884\u8bad\u7ec3\u6a21\u578b\u8fd8\u662f\u7ecf\u8fc7 SFT \u8c03\u6574\u7684\u68c0\u67e5\u70b9\uff0c<strong>diffu-GRPO \u90fd\u80fd\u63d0\u4f9b\u53ef\u9760\u7684\u6027\u80fd\u63d0\u5347<\/strong>\u3002<\/p>\n<p><strong>d1 \u8bad\u7ec3\u65b9\u6848\u5b9e\u73b0\u4e86\u6700\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347<\/strong>\u3002\u901a\u8fc7\u5148\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u3001\u518d\u7ed3\u5408 diffu-GRPO \u8bad\u7ec3\u6240\u5f62\u6210\u7684 d1-LLaDA \u6a21\u578b\uff0c\u4ea7\u751f\u4e86\u8d85\u8d8a\u5355\u4e00\u65b9\u6cd5\u7684\u53e0\u52a0\u589e\u76ca\u3002\u8fd9\u79cd\u7ec4\u5408\u5f0f\u65b9\u6cd5\u5728 12 \u4e2a\u5b9e\u9a8c\u8bbe\u7f6e\u4e2d\u6709 11 \u9879\u4f18\u4e8e\u7eaf diffu-GRPO \u65b9\u6848\uff0c\u8868\u660e\u4e24\u4e2a\u8bad\u7ec3\u9636\u6bb5\u5b58\u5728\u534f\u540c\u6548\u5e94\u3002<\/p>\n<p>\u5b9a\u6027\u7ed3\u679c\u8868\u660e\uff0c\u5728 SFT \u548c d1-LLaDA \u751f\u6210\u4e2d\u51fa\u73b0\u4e86\u987f\u609f\u65f6\u523b\u3002\u5c3d\u7ba1\u4e0e LLaDA-8B-Instruct \u76f8\u6bd4\uff0c\u751f\u6210\u5e8f\u5217\u957f\u5ea6\u4e3a 128 \u548c 256 \u7684\u6027\u80fd\u968f\u7740 SFT\u3001diffu-GRPO \u548c d1 \u6709\u6240\u63d0\u9ad8\uff0c\u4f46\u4ece\u8d28\u7684\u65b9\u9762\u770b\uff0c\u5728\u751f\u6210\u7684\u63a8\u7406\u8f68\u8ff9\u4e2d\u5e76\u672a\u89c2\u5bdf\u5230\u663e\u8457\u5dee\u5f02\u3002\u7136\u800c\u5f53\u5e8f\u5217\u957f\u5ea6\u8fbe\u5230 512 \u65f6\uff0c\u8be5\u7814\u7a76\u5f00\u59cb\u89c2\u5bdf\u5230 SFT \u548c d1-LLaDA \u6a21\u578b\u5c55\u73b0\u51fa\u4e24\u79cd\u5173\u952e\u80fd\u529b\uff1a\u81ea\u6211\u4fee\u6b63\u673a\u5236\u548c\u56de\u6eaf\u884c\u4e3a\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-20-7\" target=\"_blank\">\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bG [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[],"class_list":["post-40242","post","type-post","status-publish","format-standard","hentry","category-news"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=40242\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=40242\",\"url\":\"https:\/\/17aitech.com\/?p=40242\",\"name\":\"\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=40242#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=40242#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png\",\"datePublished\":\"2025-04-20T14:02:55+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=40242#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=40242\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=40242#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=40242#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=40242","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=40242","url":"https:\/\/17aitech.com\/?p=40242","name":"\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=40242#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=40242#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png","datePublished":"2025-04-20T14:02:55+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=40242#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=40242"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=40242#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d58910f2ae435cdf11eb99b510ba7128.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=40242#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u6269\u6563LLM\u63a8\u7406\u7528\u4e0a\u7c7bGRPO\u5f3a\u5316\u5b66\u4e60\uff01\u4f18\u4e8e\u5355\u72ecSFT\uff0cUCLA\u3001Meta\u65b0\u6846\u67b6d1\u5f00\u6e90"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/40242","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=40242"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/40242\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=40242"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=40242"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=40242"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}