{"id":25860,"date":"2025-05-15T10:01:21","date_gmt":"2025-05-15T02:01:21","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=25860"},"modified":"2025-05-15T10:01:21","modified_gmt":"2025-05-15T02:01:21","slug":"dancegrpo%ef%bc%9a%e9%a6%96%e4%b8%aa%e7%bb%9f%e4%b8%80%e8%a7%86%e8%a7%89%e7%94%9f%e6%88%90%e7%9a%84%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e6%a1%86%e6%9e%b6","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/05\/15\/25860\/","title":{"rendered":"DanceGRPO\uff1a\u9996\u4e2a\u7edf\u4e00\u89c6\u89c9\u751f\u6210\u7684\u5f3a\u5316\u5b66\u4e60\u6846\u67b6"},"content":{"rendered":"
\"\"<\/div>\n
<\/div>\n
R1 \u6a2a\u7a7a\u51fa\u4e16\uff0c\u5e26\u706b\u4e86 GRPO \u7b97\u6cd5\uff0cRL \u4e5f\u968f\u4e4b\u6210\u4e3a 2025 \u5e74\u7684\u70ed\u95e8\u6280\u672f\u63a2\u7d22\u65b9\u5411\uff0c\u8fd1\u671f\uff0c\u5b57\u8282 Seed \u56e2\u961f\u5c31\u5728\u56fe\u50cf\u751f\u6210\u65b9\u5411\u8fdb\u884c\u4e86\u76f8\u5173\u63a2\u7d22\u3002<\/div>\n
\n
\n

\u672c\u6587\u7531\u5b57\u8282\u8df3\u52a8 Seed \u548c\u9999\u6e2f\u5927\u5b66\u8054\u5408\u5b8c\u6210\u3002\u7b2c\u4e00\u4f5c\u8005\u859b\u6cfd\u5cb3\u4e3a\u9999\u6e2f\u5927\u5b66 MMLab@HKU \u5728\u8bfb\u535a\u58eb\u751f\uff0c\u5728 CVPR\u3001NeurIPS \u7b49\u56fd\u9645\u9876\u7ea7\u4f1a\u8bae\u4e0a\u53d1\u8868\u591a\u7bc7\u7814\u7a76\u6210\u679c\u3002\u9879\u76ee\u901a\u8baf\u4f5c\u8005\u4e3a\u9ec4\u4f1f\u6797\u535a\u58eb\u548c\u7f57\u5e73\u6559\u6388\u3002<\/p>\n

R1 \u6a2a\u7a7a\u51fa\u4e16\uff0c\u5e26\u706b\u4e86 GRPO \u7b97\u6cd5\uff0cRL \u4e5f\u968f\u4e4b\u6210\u4e3a 2025 \u5e74\u7684\u70ed\u95e8\u6280\u672f\u63a2\u7d22\u65b9\u5411\uff0c\u8fd1\u671f\uff0c\u5b57\u8282 Seed \u56e2\u961f\u5c31\u5728\u56fe\u50cf\u751f\u6210\u65b9\u5411\u8fdb\u884c\u4e86\u76f8\u5173\u63a2\u7d22\u3002<\/p>\n

\u73b0\u5728\uff0c\u6211\u4eec\u63a8\u51fa\u540d\u4e3a\u00a0DanceGRPO\u00a0\u7684\u521b\u65b0\u6846\u67b6\uff0c\u8fd9\u662f\u9996\u4e2a\u65e8\u5728\u7edf\u4e00\u89c6\u89c9\u751f\u6210\u5f3a\u5316\u5b66\u4e60\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u5b9e\u73b0\u4e86\u5355\u4e00\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u5728\u4e24\u5927\u751f\u6210\u8303\u5f0f\uff08diffusion\/rectified flow\uff09\u3001\u4e09\u9879\u4efb\u52a1\uff08\u6587\u672c\u5230\u56fe\u50cf\u3001\u6587\u672c\u5230\u89c6\u9891\u3001\u56fe\u50cf\u5230\u89c6\u9891\uff09\u3001\u56db\u79cd\u57fa\u7840\u6a21\u578b\uff08SD\u3001HunyuanVideo\u3001FLUX\u3001SkyReels-I2V\uff09\u4ee5\u53ca\u4e94\u7c7b\u5956\u52b1\u6a21\u578b\uff08\u56fe\u50cf \/ \u89c6\u9891\u7f8e\u5b66\u3001\u56fe\u6587\u5bf9\u9f50\u3001\u89c6\u9891\u52a8\u6001\u8d28\u91cf\u3001\u4e8c\u5143\u5956\u52b1\uff09\u4e2d\u7684\u5168\u9762\u8986\u76d6\u3002<\/p>\n

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

    \n
  • \u8bba\u6587\u6807\u9898\uff1aDanceGRPO: Unleashing GRPO on Visual Generation<\/li>\n
  • arXiv \u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2505.07818<\/li>\n<\/ul>\n

    \u52a8\u673a<\/h4>\n

    \u5728\u751f\u6210\u5f0f AI \u5feb\u901f\u53d1\u5c55\u7684\u8fd9\u4e09\u5e74\uff0cRLHF \u9010\u6e10\u7684\u8d70\u8fdb\u4e86\u5927\u5bb6\u7684\u89c6\u91ce\uff0c\u9996\u5148\u662f GPT-3.5\/o1 \u7b49\u4e00\u7cfb\u5217\u5de5\u4f5c\u8ba9\u5927\u5bb6\u660e\u767d\u4e86 RLHF \u5728 LLM \u4e2d\u7684\u610f\u4e49\uff0cLLM \u9886\u57df\u7684 RLHF \u65b9\u6848\u767e\u82b1\u9f50\u653e\uff0c\u53d1\u5c55\u7684\u76f8\u5bf9\u6210\u719f\u4e00\u4e9b\uff0c\u4eca\u5e74\u66f4\u50ac\u751f\u4e86 R1 \u8fd9\u4e00\u7c7b\u7684\u5927\u7206\u6b3e\u5de5\u4f5c\u3002\u540c\u65f6\uff0c\u5927\u5bb6\u4e5f\u53d1\u73b0\uff0c\u5bf9\u4e8e\u89c6\u89c9\u751f\u6210\u4efb\u52a1\uff0cRLHF \u5bf9\u7f8e\u611f\u7b49\u6307\u6807\u51e0\u4e4e\u6709\u7740\u51b3\u5b9a\u6027\u5f71\u54cd\uff0c\u4f46\u76f8\u8f83 LLM \u7684\u6210\u719f\u5ea6\uff0c\u751f\u6210\u9886\u57df\u7684 RLHF \u65b9\u6848\u663e\u7136\u6ca1\u90a3\u4e48\u6210\u719f\uff0c\u76ee\u524d\u7684\u4e3b\u6d41\u65b9\u6848\u5927\u6982\u5206\u4e3a\u4e24\u7c7b\uff1a<\/p>\n

    1. Diffusion\/Flow-DPO\uff1a\u8fd9\u4e00\u7c7b\u65b9\u6cd5\u662f\u4e5f\u662f\u76f4\u63a5\u6765\u81ea\u4e8e LLM \u7684 DPO \u7684\u5f71\u54cd\uff0c\u5728\u5b9e\u9645\u64cd\u4f5c\u4e2d\uff0c\u5927\u4f53\u5206\u4e3a\u4e24\u79cd\u65b9\u6848\uff0c\u7b2c\u4e00\u79cd\u662f\u79bb\u7ebf DPO\uff0c\u5373\u8ba9\u6a21\u578b\u53bb\u751f\u6210\u4e00\u6279\u6570\u636e\uff0c\u7136\u540e\u4eba\u5de5\u6807\u6ce8\uff0c\u7136\u540e\u8ba9\u597d\u548c\u574f\u7684\u6570\u636e\u7ec4\u6210 pairs \u9001\u56de\u6a21\u578b\u53bb\u4f18\u5316\uff0c\u7b2c\u4e8c\u79cd\u662f\u5728\u7ebf DPO\uff0c\u5373\u5728\u7ebf\u751f\u6210\u6570\u636e\uff0c\u7136\u540e\u8ba9 reward model \u5b9e\u65f6\u6253\u5206\uff0c\u7136\u540e\u62ff\u56de\u53bb\u8bad\u7ec3\uff0c\u8fd9\u4e24\u79cd\u601d\u8def\u5927\u540c\u5c0f\u5f02\u3002\u5728\u5b9e\u9645\u64cd\u4f5c\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0 DPO \u5bf9\u6a21\u578b\u7684\u6548\u679c\u6bd4\u8f83\u5fae\u5f31\uff0c\u6bd4\u5982 DPO \u524d\u540e\u7684\u751f\u6210\u7ed3\u679c\u533a\u522b\u5e76\u4e0d\u662f\u5f88\u5927\uff0c\u539f\u56e0\u4e5f\u5f88\u597d\u7406\u89e3\uff0cDPO \u5e76\u6ca1\u6709\u62ff reward \u4f5c\u4e3a\u4fe1\u53f7\u53bb reinforce \u6a21\u578b\u7684\u5b66\u4e60\uff0c\u800c\u662f\u7528\u4e00\u79cd\u7c7b\u4f3c SFT \u7684\u65b9\u6848\u53bb\u505a\uff0c\u8fd9\u79cd\u60c5\u51b5\u4e0b\u5bf9\u6a21\u578b\u7684\u4f5c\u7528\u5e76\u4e0d\u662f\u5f88\u5927\u3002<\/p>\n

    2. ReFL\uff1a\u8fd9\u4e00\u7c7b\u65b9\u6848\u662f diffusion\/rectified flow \u6240\u4e13\u6709\u7684\uff0c\u601d\u8def\u975e\u5e38\u7b80\u5355\u76f4\u63a5\uff0c\u5c31\u662f\u76f4\u63a5\u5728 z_t \u6b65\u4e0b\u76f4\u63a5\u9884\u6d4b z_0 \u7684\u7ed3\u679c\uff0c\u7136\u540e vae \u53bb decode \u76f4\u63a5\u9001\u8fdb reward model \u53bb\u76f4\u63a5\u53cd\u5411\u4f20\u64ad\u8fdb\u884c\u4f18\u5316\u3002\u8fd9\u4e00\u7c7b\u65b9\u6848\u5728\u56fe\u50cf\u4e0a\u6548\u679c\u5f88\u660e\u663e\uff0c\u4f46\u662f\u56e0\u4e3a\u8981\u5bf9 reward model \u548c decode \u540e\u7684 features \u8fdb\u884c\u53cd\u5411\u4f20\u64ad\uff0c\u5728\u9762\u5bf9 100 \u5e27\u4ee5\u4e0a\u7684\u89c6\u9891\u751f\u6210\u7684\u65f6\u5019\u663e\u5b58\u538b\u529b\u5f88\u5927\u3002\u800c\u4e14\uff0c\u76ee\u524d LLM \u548c diffusion \u8054\u5408\u8bad\u7ec3\u5df2\u6210\u4e3a\u5927\u52bf\u6240\u9a71\uff0cReFL \u8fd9\u79cd\u901a\u8fc7\u5efa\u6a21 z_t \u548c z_0 \u7684\u5173\u7cfb\uff0c\u5e76\u4e14\u76f4\u63a5\u53cd\u5411\u4f20\u64ad\u7684\u65b9\u5f0f\uff0c\u4f3c\u4e4e\u548c\u8fd9\u4e00\u7c7b\u6a21\u578b\u7684\u5efa\u6a21\u7b56\u7565\u6709\u7740\u5f88\u5927\u7684 gap\u3002<\/p>\n

    \u4e8e\u662f\uff0c\u5229\u7528\u5f3a\u5316\u5b66\u4e60\u5bf9\u6a21\u578b\u8fdb\u884c\u4f18\u5316\u7684\u601d\u8def\u4e5f\u5c31\u547c\u4e4b\u6b32\u51fa\u4e86\uff0c\u4e4b\u524d\u793e\u533a\u4e5f\u5bf9\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\u751f\u6210\u6a21\u578b\u6709\u8fc7\u4e00\u4e9b\u63a2\u7d22\uff0c\u4f8b\u5982 DDPO \u548c DPOK\uff0c\u4f46\u662f\u4ed6\u4eec\u90fd\u6709\u5f88\u5f3a\u7684\u5c40\u9650\u6027\uff1a<\/p>\n

    1. \u5c1d\u8bd5\u7684\u6570\u636e\u96c6\u975e\u5e38\u5c0f\uff0c\u4e00\u822c\u5c0f\u4e8e 100 \u4e2a prompts\uff0c<\/p>\n

    2. \u53ea\u9488\u5bf9\u6587\u751f\u56fe diffusion model \u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u5e76\u6ca1\u6709\u6d89\u53ca\u5230\u76ee\u524d\u6d41\u884c\u7684 rectified flow \u548c\u89c6\u9891\u751f\u6210\u6a21\u578b<\/p>\n

    \u5b9e\u73b0\u76ee\u6807<\/h4>\n

    \u4e8e\u662f\uff0c\u6211\u4eec\u7684\u76ee\u6807\u4e5f\u547c\u4e4b\u6b32\u51fa\uff0c\u6211\u4eec\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u9700\u8981\u6ee1\u8db3\u4ee5\u4e0b\u7279\u5f81\uff1a<\/p>\n

    1. \u63d0\u5347\u660e\u663e\uff0creward \u4e0a\u6da8\u7684\u8db3\u591f\u591a<\/p>\n

    2. \u5728 video \u4e0a\u5b9e\u73b0\u7684\u65f6\u5019\u663e\u5b58\u538b\u529b\u4e0d\u80fd\u5927\uff0c\u5373\uff0c\u4e0d\u80fd\u76f4\u63a5\u53cd\u5411\u4f20\u64ad reward model \u548c vae<\/p>\n

    3. \u80fd\u5728\u5927\u578b prompt \u6570\u636e\u96c6\u4e0a\u8bad\u7ec3<\/p>\n

    4. \u80fd\u8fc1\u79fb\u5230 rectified flow \u548c\u89c6\u9891\u751f\u6210\u6a21\u578b<\/p>\n

    DanceGRPO<\/h4>\n

    \u6838\u5fc3\u8d21\u732e<\/strong><\/p>\n

    \u6211\u4eec\u662f\u7b2c\u4e00\u4e2a\u89c6\u89c9\u751f\u6210 + \u5f3a\u5316\u5b66\u4e60\u7684\u5927\u4e00\u7edf\u6846\u67b6\uff0c\u4e00\u4e2a DanceGRPO\uff0c\u53ef\u4ee5\u5e94\u7528\u5728 diffusion \u548c rectified flow\uff0c\u53ef\u4ee5\u5e94\u7528\u5728\u6587\u751f\u56fe\uff0c\u6587\u751f\u89c6\u9891\uff0c\u56fe\u751f\u89c6\u9891\u4e09\u7c7b\u4efb\u52a1\uff0c\u6211\u4eec\u5728\u56db\u4e2a foundation model \u4e0a\u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u5305\u62ec\u4e86 Stable Diffusion\uff0cFLUX\uff0cHunyuanVideo\uff0cSkyReel-I2V\uff0c\u6211\u4eec\u8fd8\u5728\u4e94\u4e2a reward model \u4e0a\u8fdb\u884c\u4e86\u6d4b\u8bd5\u3002<\/p>\n

    \u65b9\u6848\u8bbe\u8ba1<\/strong><\/p>\n

    \u53d7\u6700\u8fd1\u5728 LLM \u5927\u706b\u7684 GRPO \u5f71\u54cd\uff0c\u6211\u4eec\u5f00\u59cb\u5c1d\u8bd5 GRPO \u8fd9\u4e00\u7c7b\u65b9\u6848\u662f\u5426\u80fd\u5728 diffusion model \u6216\u8005 rectified flow \u4e0a\u7a33\u5b9a\u8bad\u7ec3\uff0c\u9996\u5148\u6211\u4eec\u8981 claim \u7684\u70b9\u662f\uff0cdiffusion model \u548c rectified flow \u7684\u51fa\u53d1\u70b9\u867d\u7136\u4e0d\u540c\uff0c\u5373 diffusion \u548c rectified flow \u90fd\u53ef\u4ee5\u8868\u793a\u6210 stochastic interpolant \u7684\u7279\u6b8a\u60c5\u51b5\uff0c\u5728\u8fd9\u79cd case \u4e0b\uff0c\u4ed6\u4eec\u7684\u91c7\u6837\u65b9\u7a0b\u90fd\u53ef\u4ee5\u7528 SDE \u5b9e\u73b0\uff0c\u5177\u4f53\u7684\u516c\u5f0f\u63a8\u5bfc\u53c2\u89c1\u539f\u6587\u5373\u53ef\u3002<\/p>\n

    \u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c31\u5f00\u59cb\u5b9e\u73b0\u5566\uff0c\u6838\u5fc3\u7684\u601d\u8def\u8fd8\u662f follow \u4e86 DeepSeek \u7684 GRPO \u7b56\u7565\uff0c\u5373\uff0c\u7528\u4e00\u4e2a prompt\uff0c\u751f\u6210\u4e00\u6279\u6570\u636e\uff0c\u7136\u540e\u7528 GRPO \u7684\u76ee\u6807\u51fd\u6570\u8fdb\u884c\u4f18\u5316\uff0c\u4f46\u6211\u4eec\u5e76\u6ca1\u6709\u52a0\u5165 KL \u6563\u5ea6\u7684\u6b63\u5219\u9879\uff0c\u56e0\u4e3a\u53d1\u73b0\u8fd9\u4e00\u9879\u5b9e\u9645\u4e0a\u4f5c\u7528\u4e0d\u5927\uff0c\u4ee5\u4e0b\u662f\u6211\u4eec\u5b9e\u9a8c\u8fc7\u7a0b\u4e2d\u7684\u4e00\u4e9b\u6838\u5fc3\u53d1\u73b0\uff1a<\/p>\n

    1. \u540c\u4e00\u4e2a prompt \u7684\u521d\u59cb\u5316\u566a\u58f0\u6700\u597d\u76f8\u540c\uff0c\u4e0d\u7136\u5bb9\u6613 reward hacking<\/p>\n

    2. \u6211\u4eec\u53ef\u4ee5\u91c7\u6837\u4e00\u4e2a\u5b50\u96c6\u7684 timesteps \u6765\u52a0\u901f\u8bad\u7ec3\uff0c\u540c\u65f6\u8ba9\u6a21\u578b\u89c1\u8fc7\u66f4\u591a\u7684 prompts<\/p>\n

    3. \u5b9e\u73b0\u4e2d\u53ef\u4ee5\u4f7f\u7528\u591a\u4e2a reward model \u53e0\u52a0\uff0c\u800c\u4e14\u5b9e\u73b0\u7684\u65f6\u5019\u6700\u597d\u662f\u591a\u4e2a advantage \u6765\u53e0\u52a0<\/p>\n

    4. DanceGRPO \u53ef\u4ee5\u5b66\u4f1a best-of-n inference scaling \u7684\u8f68\u8ff9<\/p>\n

    5. \u5f3a\u5316\u5b66\u4e60\u4f1a\u524a\u5f31\u751f\u6210\u7684\u591a\u6837\u6027<\/p>\n

    6. \u8bad\u7ec3\u5c3d\u91cf\u4e0d\u8981\u6253\u5f00 cfg\uff0c\u5982\u679c\u975e\u8981\u6253\u5f00\u7684\u8bdd\uff0c\u4e00\u6279 prompt \u53ea\u80fd\u66f4\u65b0\u4e00\u6b21\u68af\u5ea6<\/p>\n

    \u7b97\u6cd5\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n

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

    \u63a5\u4e0b\u6765\u662f\u5173\u4e8e reward model \u7684\u8ba8\u8bba\uff0c\u6211\u4eec\u4e00\u5171\u4f7f\u7528\u4e86\u4e94\u7c7b reward model\uff1a<\/p>\n

    (1) \u56fe\u50cf\u7f8e\u611f<\/p>\n

    (2) \u89c6\u9891\u7f8e\u611f<\/p>\n

    (3) \u56fe\u6587\u5339\u914d<\/p>\n

    (4) \u89c6\u9891\u52a8\u6001\u8d28\u91cf<\/p>\n

    (5) \u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684 reward model\uff0c\u5373\u628a\u7f8e\u611f & \u56fe\u6587\u5339\u914d reward model \u7684\u7ed3\u679c\u7ed9\u4e8c\u503c\u5316\uff0c\u610f\u601d\u662f\u5927\u4e8e\u67d0\u4e2a\u9608\u503c\u5c31\u662f 1\uff0c\u5c0f\u4e8e\u8fd9\u4e2a\u9608\u503c\u5c31\u662f 0<\/p>\n

    \u6211\u4eec\u5728\u6587\u751f\u56fe\uff0c\u6587\u751f\u89c6\u9891\u548c\u56fe\u751f\u89c6\u9891\u4e0a\u8fdb\u884c\u4e86\u6d4b\u8bd5\u3002<\/p>\n

    \u5b9e\u9a8c\u7ed3\u679c<\/h4>\n

    \u6211\u4eec\u4f7f\u7528\u4e86 HPS-v2.1 \u548c Clip score \u53bb\u4f18\u5316\u6a21\u578b\uff0c\u7ed3\u679c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n

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

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

    \u6211\u4eec\u5229\u7528 VideoAlign \u5728 HunyuanVideo \u4e0a\u8fdb\u884c\u8bad\u7ec3\uff1a<\/p>\n

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

    \u4ee5\u4e0b\u662f\u4e00\u4e9b reward \u66f2\u7ebf\uff1a<\/p>\n

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

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

    ps\uff1ai2v \u4efb\u52a1\u6709\u4e00\u4e9b\u533a\u522b\uff0c\u6211\u4eec\u62ff\u89c6\u89c9\u7f8e\u611f reward model \u8bad\u7ec3\u7684\u65f6\u5019\u5f88\u5bb9\u6613\u53d1\u6563\uff0c\u4e2a\u4eba\u8ba4\u4e3a\u662f\u56e0\u4e3a i2v \u7684\u7f8e\u611f\u66f4\u591a\u53d6\u51b3\u4e8e\u9996\u5e27\uff0c\u8fd9\u4e2a\u4efb\u52a1\u672c\u8eab\u53ea\u6709 motion quality \u53ef\u8a00\uff0c\u6240\u4ee5\u6211\u4eec\u53ea\u9009\u62e9\u4e86\u4f7f\u7528 motion quality reward \u53bb\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n

    \u8fd9\u662f\u4e00\u4e2a\u968f\u7740 FLUX \u8bad\u7ec3\u8fed\u4ee3\u6b21\u6570\u7684\u589e\u52a0\uff0c\u53ef\u89c6\u5316\u7684\u53d8\u5316\uff0c\u6211\u4eec\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5f15\u5165\u968f\u673a\u6027\uff0c\u4f46\u662f\u53ef\u89c6\u5316\u662f\u4e0d\u4f1a\u5f15\u5165\u7684\uff0c\u5c31\u662f\u6b63\u5e38\u7684 ODE solver\u3002<\/p>\n

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

    \u66f4\u591a\u7684\u7ed3\u679c\u6b22\u8fce\u770b\u539f\u6587\uff5e<\/p>\n

    \u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<\/div>\n

    <\/div>\n

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

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

    R1 \u6a2a\u7a7a\u51fa\u4e16\uff0c\u5e26\u706b\u4e86 GRPO \u7b97\u6cd5\uff0cRL \u4e5f\u968f\u4e4b\u6210\u4e3a 2025 \u5e74\u7684\u70ed\u95e8\u6280\u672f\u63a2\u7d22\u65b9\u5411\uff0c\u8fd1\u671f\uff0c\u5b57\u8282 Seed […]<\/p>\n

    <\/div>\n

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

    <\/div>\n","protected":false},"author":56,"featured_media":25862,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,20,80],"tags":[124,1407,1688],"class_list":["post-25860","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-20","category-80","tag-ai","tag-1407","tag-1688"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25860","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=25860"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25860\/revisions"}],"predecessor-version":[{"id":25863,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25860\/revisions\/25863"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/25862"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=25860"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=25860"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=25860"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}