{"id":26789,"date":"2025-07-22T10:30:30","date_gmt":"2025-07-22T02:30:30","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=26789"},"modified":"2025-07-22T10:30:30","modified_gmt":"2025-07-22T02:30:30","slug":"%e7%aa%81%e7%a0%b4%e9%ab%98%e5%88%86%e8%be%a8%e7%8e%87%e5%9b%be%e5%83%8f%e6%8e%a8%e7%90%86%e7%93%b6%e9%a2%88%ef%bc%8c%e5%a4%8d%e6%97%a6%e8%81%94%e5%90%88%e5%8d%97%e6%b4%8b%e7%90%86%e5%b7%a5%e6%8f%90","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/07\/22\/26789\/","title":{"rendered":"\u7a81\u7834\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u63a8\u7406\u74f6\u9888\uff0c\u590d\u65e6\u8054\u5408\u5357\u6d0b\u7406\u5de5\u63d0\u51fa\u57fa\u4e8e\u89c6\u89c9Grounding\u7684\u591a\u8f6e\u5f3a\u5316\u5b66\u4e60\u6846\u67b6MGPO"},"content":{"rendered":"
\"\"<\/div>\n
<\/div>\n
MGPO \u901a\u8fc7\u591a\u8f6e\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u6fc0\u6d3b\u89c6\u89c9 Grounding \u80fd\u529b\uff0c\u6709\u6548\u63d0\u5347\u4e86\u591a\u6a21\u6001\u5927\u6a21\u578b\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u7684 \u300c\u89c6\u89c9 Token \u5197\u4f59\u300d \u548c \u300c\u5173\u952e\u4fe1\u606f\u4e22\u5931\u300d \u7b49\u95ee\u9898\u3002<\/div>\n
\n
\n

\u672c\u6587\u7684\u4e3b\u8981\u4f5c\u8005\u6765\u81ea\u590d\u65e6\u5927\u5b66\u548c\u5357\u6d0b\u7406\u5de5\u5927\u5b66 S-Lab\uff0c\u7814\u7a76\u65b9\u5411\u805a\u7126\u4e8e\u89c6\u89c9\u63a8\u7406\u4e0e\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\u3002<\/span><\/p>\n

\u5148\u8fdb\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08Large Multi-Modal Models, LMMs\uff09\u901a\u5e38\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Models, LLMs\uff09\u7ed3\u5408\u539f\u751f\u5206\u8fa8\u7387\u89c6\u89c9 Transformer\uff08NaViT\uff09\u6784\u5efa\u3002\u7136\u800c\uff0c\u8fd9\u7c7b\u6a21\u578b\u5728\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u9762\u4e34\u74f6\u9888\uff1a\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u4f1a\u8f6c\u5316\u4e3a\u6d77\u91cf\u89c6\u89c9 Token\uff0c\u5176\u4e2d\u5927\u90e8\u5206\u4e0e\u4efb\u52a1\u65e0\u5173\uff0c\u65e2\u589e\u52a0\u4e86\u8ba1\u7b97\u8d1f\u62c5\uff0c\u4e5f\u5e72\u6270\u4e86\u6a21\u578b\u5bf9\u5173\u952e\u4fe1\u606f\u7684\u6355\u6349\u3002<\/span><\/p>\n

\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u590d\u65e6\u5927\u5b66\u3001\u5357\u6d0b\u7406\u5de5\u5927\u5b66\u7684\u7814\u7a76\u8005\u63d0\u51fa\u4e00\u79cd\u57fa\u4e8e\u89c6\u89c9 Grounding \u7684\u591a\u8f6e\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5 MGPO\uff0c\u4f7f LMM \u80fd\u5728\u591a\u8f6e\u4ea4\u4e92\u4e2d\u6839\u636e\u95ee\u9898\uff0c\u81ea\u52a8\u9884\u6d4b\u5173\u952e\u533a\u57df\u5750\u6807\uff0c\u88c1\u526a\u5b50\u56fe\u50cf\u5e76\u6574\u5408\u5386\u53f2\u4e0a\u4e0b\u6587\uff0c\u6700\u7ec8\u5b9e\u73b0\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684\u7cbe\u51c6\u63a8\u7406\u3002\u76f8\u6bd4\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u9700\u8981\u6602\u8d35\u7684 Grounding \u6807\u6ce8\u4f5c\u4e3a\u76d1\u7763\uff0cMGPO \u8bc1\u660e\u4e86\u5728\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8303\u5f0f\u4e2d\uff0c\u5373\u4f7f\u6ca1\u6709 Grounding \u6807\u6ce8\uff0c\u6a21\u578b\u4e5f\u80fd\u4ece \u300c\u6700\u7ec8\u7b54\u6848\u662f\u5426\u6b63\u786e\u300d\u7684\u53cd\u9988\u4e2d\uff0c\u6d8c\u73b0\u51fa\u9c81\u68d2\u7684\u89c6\u89c9 Grounding \u80fd\u529b\u3002<\/span><\/p>\n

MGPO \u7684\u6838\u5fc3\u521b\u65b0\u70b9\u5305\u62ec\uff1a 1\uff09\u81ea\u4e0a\u800c\u4e0b\u7684\u53ef\u89e3\u91ca\u89c6\u89c9\u63a8\u7406\uff1a\u8d4b\u4e88\u4e86 LMMs \u9488\u5bf9\u9ad8\u5206\u8fa8\u7387\u573a\u666f\u7684 \u300c\u81ea\u4e0a\u800c\u4e0b\u3001\u95ee\u9898\u9a71\u52a8\u300d \u89c6\u89c9\u641c\u7d22\u673a\u5236\uff0c\u63d0\u4f9b\u53ef\u89e3\u91ca\u7684\u89c6\u89c9 Grounding \u8f93\u51fa\uff1b 2\uff09\u7a81\u7834\u6700\u5927\u50cf\u7d20\u9650\u5236\uff1a\u5373\u4f7f\u56e0\u89c6\u89c9 Token \u6570\u53d7\u9650\u5bfc\u81f4\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7f29\u653e\u540e\u6a21\u7cca\uff0c\u6a21\u578b\u4ecd\u80fd\u51c6\u786e\u8bc6\u522b\u76f8\u5173\u533a\u57df\u5750\u6807\uff0c\u4ece\u539f\u59cb\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u4e2d\u88c1\u526a\u51fa\u6e05\u6670\u5b50\u56fe\u50cf\u7528\u4e8e\u540e\u7eed\u5206\u6790\uff1b 3\uff09\u65e0\u9700\u989d\u5916 Grounding \u6807\u6ce8\uff1a\u53ef\u76f4\u63a5\u5728\u6807\u51c6 VQA \u6570\u636e\u96c6\u4e0a\u8fdb\u884c RL \u8bad\u7ec3\uff0c\u4ec5\u57fa\u4e8e\u7b54\u6848\u76d1\u7763\u5c31\u80fd\u8ba9\u6a21\u578b\u6d8c\u73b0\u51fa\u9c81\u68d2\u7684\u89c6\u89c9 Grounding \u80fd\u529b\u3002<\/span><\/p>\n

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

    \n
  • \u6807\u9898\uff1aHigh-Resolution Visual Reasoning via Multi-Turn Grounding-Based Reinforcement Learning<\/span><\/li>\n
  • \u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2507.05920<\/span><\/li>\n
  • \u4ee3\u7801\u5730\u5740\uff1ahttps:\/\/github.com\/EvolvingLMMs-Lab\/MGPO<\/span><\/li>\n<\/ul>\n

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

    \u56fe 1\uff1a\u57fa\u4e8e MGPO \u8bad\u7ec3\u7684\u6a21\u578b\u6027\u80fd\u5c55\u793a\uff0c\u5728\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\uff0c\u6a21\u578b\u4f1a\u6839\u636e\u95ee\u9898\u8f93\u51fa\u5173\u952e\u533a\u57df\u5750\u6807\uff0c\u7136\u540e\u81ea\u52a8\u89e6\u53d1\u56fe\u50cf\u88c1\u526a\u51fd\u6570\uff0c\u8fd4\u56de\u6e05\u6670\u7684\u5b50\u56fe\u5e2e\u52a9\u6a21\u578b\u56de\u7b54\u95ee\u9898\u3002<\/span><\/p>\n

    \u4ecb\u7ecd<\/span><\/h4>\n

    \u5f53\u524d\uff0c\u4ee5 Qwen2.5-VL \u4e3a\u4ee3\u8868\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08LMMs\uff09\u901a\u5e38\u57fa\u4e8e\u5f3a\u5927\u7684\u8bed\u8a00\u6a21\u578b\uff08\u5982 Qwen2.5\uff09\u7ed3\u5408\u5916\u90e8\u539f\u751f\u5206\u8fa8\u7387\u89c6\u89c9 Transformer\uff08NaViT\uff09\u6784\u5efa\u3002\u7136\u800c\uff0c\u8fd9\u7c7b\u6a21\u578b\u5728\u5904\u7406\u9ad8\u5206\u8fa8\u56fe\u50cf\u4efb\u52a1\u65f6\u9762\u4e34\u6311\u6218\uff1a\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u4f1a\u8f6c\u6362\u6210\u6d77\u91cf\u89c6\u89c9 Token\uff0c\u5176\u4e2d\u5927\u90e8\u5206\u4e0e\u4efb\u52a1\u65e0\u5173\uff0c\u65e2\u589e\u52a0\u4e86\u8ba1\u7b97\u8d1f\u62c5\uff0c\u4e5f\u5e72\u6270\u4e86\u6a21\u578b\u5bf9\u5173\u952e\u4fe1\u606f\u7684\u6355\u6349\u3002<\/span><\/p>\n

    \u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5728\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u771f\u5b9e\u573a\u666f\u65f6\uff0c\u4eba\u7c7b\u89c6\u89c9\u7cfb\u7edf\u4f1a\u91c7\u7528\u4efb\u52a1\u9a71\u52a8\u7684\u89c6\u89c9\u641c\u7d22\u7b56\u7565\uff0c\u9996\u5148\u5b9a\u4f4d\uff0c\u518d\u4ed4\u7ec6\u5ba1\u89c6\u5173\u952e\u5174\u8da3\u533a\u57df\u3002\u53d7\u8fd9\u4e00\u751f\u7269\u673a\u5236\u542f\u53d1\uff0c\u6211\u4eec\u5c1d\u8bd5\u901a\u8fc7\u89c6\u89c9 Grounding \u4e3a LMMs \u8d4b\u4e88\u7c7b\u4f3c\u7684\u89c6\u89c9\u641c\u7d22\u80fd\u529b\uff0c\u4f7f\u5176\u805a\u7126\u4e8e\u56fe\u50cf\u4e2d\u7684\u5173\u952e\u533a\u57df\u3002<\/span><\/p>\n

    \u4f46\u4f20\u7edf\u89c6\u89c9 Grounding \u6a21\u578b\u9700\u4f9d\u8d56\u5927\u91cf Grounding \u6807\u6ce8\u8fdb\u884c\u8bad\u7ec3\uff0c\u800c\u6b64\u7c7b\u6807\u6ce8\u6210\u672c\u8f83\u9ad8\u3002\u6709\u6ca1\u6709\u53ef\u80fd\u4e0d\u9700\u8981\u989d\u5916 Grounding \u6807\u6ce8\uff0c\u4ec5\u901a\u8fc7\u6700\u7ec8\u7b54\u6848\u7684\u6b63\u786e\u6027\u5bf9\u6a21\u578b\u8fdb\u884c\u5956\u52b1\uff0c\u5c31\u8ba9\u6a21\u578b\u81ea\u52a8\u5b66\u4f1a \u300c\u627e\u91cd\u70b9\u300d\uff1f<\/span><\/p>\n

    \u6211\u4eec\u7684\u7b54\u6848\u662f\uff1a\u53ef\u4ee5\u3002\u672c\u6587\u63d0\u51fa\u57fa\u4e8e\u89c6\u89c9 Grounding \u7684\u591a\u8f6e\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5 MGPO\uff08Multi-turn Grounding-based Policy Optimization\uff09\uff0c\u4f7f LMMs \u80fd\u5728\u591a\u8f6e\u4ea4\u4e92\u4e2d\u81ea\u52a8\u9884\u6d4b\u5173\u952e\u533a\u57df\u5750\u6807\u3001\u88c1\u526a\u5b50\u56fe\u50cf\u5e76\u6574\u5408\u5386\u53f2\u4e0a\u4e0b\u6587\uff0c\u6700\u7ec8\u5b9e\u73b0\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684\u7cbe\u51c6\u63a8\u7406\u3002\u6211\u4eec\u7684\u5b9e\u9a8c\u8bc1\u660e\uff0c\u5373\u4f7f\u6ca1\u6709\u4efb\u4f55 Grounding \u6807\u6ce8\uff0c\u6a21\u578b\u4e5f\u80fd\u4ece \u300c\u6700\u7ec8\u7b54\u6848\u662f\u5426\u6b63\u786e\u300d \u7684\u5956\u52b1\u53cd\u9988\u4e2d\uff0c\u6d8c\u73b0\u51fa\u9c81\u68d2\u7684\u89c6\u89c9\u5b9a\u4f4d\u80fd\u529b\u3002<\/span><\/p>\n

    \u65b9\u6cd5\u6982\u89c8<\/span><\/h4>\n

    MGPO \u7684\u6838\u5fc3\u601d\u60f3\u662f\u6a21\u62df\u4eba\u7c7b\u7684\u591a\u6b65\u89c6\u89c9\u63a8\u7406\u8fc7\u7a0b\uff1a\u7ed9\u5b9a\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u548c\u95ee\u9898\uff0c\u6a21\u578b\u5148\u9884\u6d4b\u5173\u952e\u533a\u57df\u7684\u5750\u6807\uff0c\u88c1\u526a\u51fa\u5b50\u56fe\u50cf\uff1b\u518d\u7ed3\u5408\u539f\u59cb\u56fe\u50cf\u548c\u5b50\u56fe\u50cf\u7684\u4e0a\u4e0b\u6587\uff0c\u8fdb\u884c\u4e0b\u4e00\u6b65\u63a8\u7406\u3002<\/span><\/p>\n

    \u4e0b\u56fe\u6bd4\u8f83\u4e86 MGPO \u4e0e SFT\u3001GRPO \u7684\u533a\u522b\uff0cMGPO \u53ef\u4ee5\u4ec5\u9760\u6b63\u786e\u7b54\u6848\u7684\u76d1\u7763\u4fe1\u606f\uff0c\u6d8c\u73b0\u9c81\u68d2\u7684\u89c6\u89c9 Grounding \u80fd\u529b\u3002<\/span><\/p>\n

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

    \u89e3\u51b3 \u300c\u51b7\u542f\u52a8\u300d\uff1a\u56fa\u5b9a\u4e24\u56de\u5408\u5bf9\u8bdd\u6a21\u677f<\/strong><\/span><\/p>\n

    \u5728\u5b9e\u9645\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0 LLMs \u5728 Rollout \u8fc7\u7a0b\u4e2d\uff0c\u96be\u4ee5\u81ea\u4e3b\u5728\u4e2d\u95f4\u8fc7\u7a0b\u8c03\u7528 Grounding \u80fd\u529b\uff0c\u4f7f\u5f97 RL \u8bad\u7ec3\u8fc7\u7a0b\u7f13\u6162\u3002\u4e3a\u4e86\u89e3\u51b3\u6a21\u578b\u7684\u51b7\u542f\u52a8\u95ee\u9898\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u56fa\u5b9a\u4e24\u8f6e\u5bf9\u8bdd\u6a21\u677f\uff08\u5982\u4e0b\u56fe\u6240\u793a\uff09\uff0c\u5728\u7b2c\u4e00\u8f6e\u5bf9\u8bdd\u4e2d\u660e\u786e\u8981\u6c42\u6a21\u578b\u53ea\u8f93\u51fa\u4e0e\u95ee\u9898\u76f8\u5173\u7684\u533a\u57df\u5750\u6807\uff0c\u5728\u7b2c\u4e8c\u8f6e\u5bf9\u8bdd\u4e2d\u518d\u8981\u6c42\u6a21\u578b\u56de\u7b54\u95ee\u9898\u3002\u00a0<\/span><\/p>\n

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

    \u5904\u7406\u9ad8\u5206\u8fa8\u7387\uff1a\u5750\u6807\u5f52\u4e00\u5316\u4e0e\u5b50\u56fe\u50cf\u88c1\u526a<\/span><\/p>\n

    \u53d7\u9650\u4e8e\u6a21\u578b\u80fd\u591f\u5904\u7406\u7684\u89c6\u89c9 Token \u6570\u91cf\uff0c\u9ad8\u5206\u8fa8\u7387\u56fe\u5f80\u5f80\u4f1a\u88ab\u7f29\u653e\u6210\u6a21\u7cca\u56fe\u50cf\uff0c\u5bfc\u81f4\u7ec6\u8282\u4e22\u5931\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5f53\u5904\u7406\u7f29\u653e\u56fe\u50cf\u65f6\uff0cMGPO \u4f1a\u5148\u5b9a\u4f4d\u5230\u4e0e\u95ee\u9898\u76f8\u5173\u7684\u533a\u57df\uff0c\u518d\u4ece\u539f\u59cb\u56fe\u50cf\u4e2d\u88c1\u526a\u51fa\u6e05\u6670\u7684\u5b50\u56fe\uff0c\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u6b63\u786e\u56de\u7b54\u76f8\u5173\u95ee\u9898\u3002<\/span><\/p>\n

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

    \u5b9e\u9a8c\u7ed3\u679c<\/span><\/p>\n

    1.\u4e0d\u540c\u8303\u5f0f\u5bf9\u6bd4<\/strong><\/span><\/p>\n

    \u57fa\u4e8e\u76f8\u540c\u8bad\u7ec3\u6570\u636e\u4e0b\uff0c\u6211\u4eec\u5bf9\u6bd4\u4e86 SFT\u3001GRPO\u3001MGPO \u5728\u4e24\u4e2a\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf Benchmark \u7684\u8868\u73b0\uff1aMME-Realworld\uff08In-Distribution\uff09\u548c V* Bench (Out of Distribution)\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0cGRPO \u76f8\u8f83\u4e8e SFT \u5e76\u672a\u5e26\u6765\u663e\u8457\u6027\u80fd\u63d0\u5347\uff0c\u8fd9\u4e0e\u4e4b\u524d\u591a\u6a21\u6001\u6570\u5b66\u4efb\u52a1\u7684\u7814\u7a76\u7ed3\u8bba\u76f8\u53cd\u3002\u6211\u4eec\u63a8\u6d4b\uff0c\u5bf9\u4e8e\u9ad8\u5206\u8fa8\u7387\u89c6\u89c9\u4e2d\u5fc3\u4efb\u52a1\uff0c\u6838\u5fc3\u6311\u6218\u5728\u4e8e\u8ba9\u6a21\u578b\u611f\u77e5\u7ec6\u7c92\u5ea6\u56fe\u50cf\u7ec6\u8282\uff0c\u800c\u975e\u8fdb\u884c\u590d\u6742\u7684\u957f\u94fe\u63a8\u7406\u3002<\/span><\/p>\n

    \u76f8\u6bd4\u4e4b\u4e0b\uff0cMGPO \u53d6\u5f97\u4e86\u663e\u8457\u63d0\u5347\uff0c\u76f8\u6bd4 GRPO \u5728 MME-Realworld\u3001V* Bench \u5206\u522b\u63d0\u5347 5.4%\u30015.2%\u3002\u6211\u4eec\u8fd8\u5c06\u7ed3\u679c\u4e0e OpenAI \u7684 o1\u3001GPT-4o \u5728 V* Bench \u4e0a\u8fdb\u884c\u4e86\u5bf9\u6bd4\uff0c\u5c3d\u7ba1\u6211\u4eec\u7684\u6a21\u578b\u4ec5\u57fa\u4e8e 7B \u6a21\u578b\u3001\u7528 2.1 \u4e07\u6837\u672c\u8bad\u7ec3\uff0c\u7ecf\u8fc7 MGPO \u8bad\u7ec3\u7684\u6a21\u578b\u4ecd\u8d85\u8fc7\u4e86\u8fd9\u4e24\u4e2a\u5546\u4e1a\u5927\u6a21\u578b\u3002<\/span><\/p>\n

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

    2.RL \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u89c6\u89c9 Grounding \u80fd\u529b\u7684\u6d8c\u73b0<\/strong><\/span><\/p>\n

    \u6211\u4eec\u7edf\u8ba1\u4e86 GRPO \u4e0e MGPO \u4e24\u79cd RL \u6846\u67b6\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u751f\u6210\u7684\u6709\u6548 Grounding \u5750\u6807\u6bd4\u4f8b\u3002\u7ed3\u679c\u663e\u793a\uff0cMGPO \u7684\u6709\u6548\u6bd4\u4f8b\u968f\u8bad\u7ec3\u8fed\u4ee3\u5448\u73b0\u663e\u8457\u4e0a\u5347\u8d8b\u52bf\uff0c\u8bc1\u660e\u4e86 MGPO \u4ec5\u9700\u5229\u7528\u6807\u51c6 VQA \u6570\u636e\uff08\u65e0\u9700\u989d\u5916 Grounding \u6807\u6ce8\uff09\uff0c\u5c31\u80fd\u5728 RL \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u81ea\u4e3b\u6d8c\u73b0\u51fa\u7a33\u5b9a\u3001\u7cbe\u51c6\u7684\u89c6\u89c9 Grounding \u80fd\u529b\u3002<\/span><\/p>\n

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

    \u603b\u7ed3<\/span><\/h4>\n

    MGPO \u901a\u8fc7\u591a\u8f6e\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u6fc0\u6d3b\u89c6\u89c9 Grounding \u80fd\u529b\uff0c\u6709\u6548\u63d0\u5347\u4e86\u591a\u6a21\u6001\u5927\u6a21\u578b\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u7684 \u300c\u89c6\u89c9 Token \u5197\u4f59\u300d \u548c \u300c\u5173\u952e\u4fe1\u606f\u4e22\u5931\u300d \u7b49\u95ee\u9898\u3002\u540c\u65f6\uff0c\u5b9e\u9a8c\u8bc1\u660e\u4e86\uff0c\u76f8\u6bd4 SFT \u9700\u8981\u6602\u8d35\u7684 Grounding \u6807\u6ce8\uff0cRL \u7b97\u6cd5\u53ef\u4ee5\u4ec5\u901a\u8fc7\u6700\u7ec8\u7b54\u6848\u7684\u5956\u52b1\u53cd\u9988\uff0c\u4f7f\u5f97\u6a21\u578b\u81ea\u4e3b\u6d8c\u73b0\u51fa\u9c81\u68d2\u7684 Grounding \u80fd\u529b\uff0c\u907f\u514d\u4e86\u5bf9\u6602\u8d35 Grounding \u6807\u6ce8\u7684\u4f9d\u8d56\u3002<\/span><\/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":"

    MGPO \u901a\u8fc7\u591a\u8f6e\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u6fc0\u6d3b\u89c6\u89c9 Grounding \u80fd\u529b\uff0c\u6709\u6548\u63d0\u5347\u4e86\u591a\u6a21\u6001\u5927\u6a21\u578b\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u65f6\u7684 […]<\/p>\n

    <\/div>\n

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

    <\/div>\n","protected":false},"author":56,"featured_media":26791,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,20,80],"tags":[1821],"class_list":["post-26789","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-20","category-80","tag-1821"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26789","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=26789"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26789\/revisions"}],"predecessor-version":[{"id":26792,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26789\/revisions\/26792"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/26791"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=26789"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=26789"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=26789"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}