{"id":20932,"date":"2024-07-02T11:35:13","date_gmt":"2024-07-02T03:35:13","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=20932"},"modified":"2024-07-02T11:35:13","modified_gmt":"2024-07-02T03:35:13","slug":"%e7%ad%89%e4%b8%8d%e6%9d%a5openai%e7%9a%84q%ef%bc%8c%e5%8d%8e%e4%b8%ba%e8%af%ba%e4%ba%9a%e6%8e%a2%e7%b4%a2llm%e6%8e%a8%e7%90%86%e7%9a%84%e7%a7%98%e5%af%86%e6%ad%a6%e5%99%a8mindstar%e5%85%88%e6%9d%a5","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2024\/07\/02\/20932\/","title":{"rendered":"\u7b49\u4e0d\u6765OpenAI\u7684Q*\uff0c\u534e\u4e3a\u8bfa\u4e9a\u63a2\u7d22LLM\u63a8\u7406\u7684\u79d8\u5bc6\u6b66\u5668MindStar\u5148\u6765\u4e86"},"content":{"rendered":"
\"\"<\/div>\n
\u672c\u6587\u4ecb\u7ecd\u4e86 MindStar\uff08M*\uff09\uff0c\u4e00\u79cd\u65b0\u9896\u7684\u57fa\u4e8e\u641c\u7d22\u7684\u63a8\u7406\u6846\u67b6\uff0c\u7528\u4e8e\u589e\u5f3a\u9884\u8bad\u7ec3\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002\u901a\u8fc7\u5c06\u63a8\u7406\u4efb\u52a1\u89c6\u4e3a\u641c\u7d22\u95ee\u9898\u5e76\u5229\u7528\u8fc7\u7a0b\u76d1\u7763\u7684\u5956\u52b1\u6a21\u578b\uff0cM* \u5728\u63a8\u7406\u6811\u7a7a\u95f4\u4e2d\u6709\u6548\u5bfc\u822a\uff0c\u8bc6\u522b\u8fd1\u4f3c\u6700\u4f18\u8def\u5f84\u3002<\/div>\n
\n
\n

\u672c\u8bba\u6587\u4f5c\u8005\u6765\u81ea\u534e\u4e3a\u8499\u7279\u5229\u5c14\u8bfa\u4e9a\u65b9\u821f\u5b9e\u9a8c\u5ba4\u7684\u5eb7\u8ba1\u5803\uff0c\u674e\u4fe1\u62e9\uff0c\u9648\u7199\uff0c Amirreza Kazemi\uff0c\u9648\u535a\u5174\u3002<\/strong><\/p>\n

\u4eba\u5de5\u667a\u80fd\uff08AI\uff09\u5728\u8fc7\u53bb\u5341\u5e74\u91cc\u53d6\u5f97\u4e86\u957f\u8db3\u8fdb\u6b65\uff0c\u7279\u522b\u662f\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u3002\u7136\u800c\uff0c\u5982\u4f55\u63d0\u5347 AI \u7684\u8ba4\u77e5\u80fd\u529b\u548c\u63a8\u7406\u80fd\u529b\uff0c\u4ecd\u7136\u662f\u4e00\u4e2a\u5de8\u5927\u7684\u6311\u6218\u3002<\/p>\n

\u8fd1\u671f\uff0c\u4e00\u7bc7\u9898\u4e3a\u300aMindStar: Enhancing Math Reasoning in Pre-trained LLMs at Inference Time\u300b\u7684\u8bba\u6587\u63d0\u51fa\u4e86\u57fa\u4e8e\u6811\u641c\u7d22\u7684\u63a8\u7406\u65f6\u95f4\u80fd\u529b\u63d0\u5347\u65b9\u6cd5 MindStar [1]\uff0c\u8be5\u65b9\u6cd5\u5728\u5f00\u6e90\u6a21\u578b Llama-13-B \u4e0e Mistral-7B \u4e0a\u8fbe\u5230\u4e86\u8fd1\u4f3c\u95ed\u6e90\u5927\u6a21\u578b GPT-3.5 \u4e0e Grok-1 \u5728\u6570\u5b66\u95ee\u9898\u4e0a\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n

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

    \n
  • \u8bba\u6587\u6807\u9898\uff1aMindStar: Enhancing Math Reasoning in Pre-trained LLMs at Inference Time<\/li>\n
  • \u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2405.16265v2<\/li>\n<\/ul>\n

    MindStar \u5728\u6570\u5b66\u95ee\u9898\u4e0a\u7684\u5e94\u7528\u6548\u679c\uff1a<\/p>\n

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

    \u56fe 1 \uff1a\u4e0d\u540c\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u6570\u5b66\u51c6\u786e\u7387\u3002LLaMA-2-13B \u5728\u6570\u5b66\u6027\u80fd\u4e0a\u4e0e GPT-3.5 (4-shot) \u7c7b\u4f3c\uff0c\u4f46\u8282\u7701\u4e86\u5927\u7ea6 200 \u500d\u7684\u8ba1\u7b97\u8d44\u6e90\u3002<\/p>\n

    1. \u5f15\u8a00<\/h4>\n

    \u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u5feb\u901f\u589e\u957f\uff0c\u57fa\u4e8e Transformer \u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u6307\u4ee4\u9075\u5faa [1,2]\u3001\u7f16\u7801\u8f85\u52a9 [3,4] \u548c\u521b\u610f\u5199\u4f5c [5] \u7b49\u9886\u57df\u5c55\u793a\u4e86\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u6210\u679c\u3002\u7136\u800c\uff0c\u89e3\u9501 LLMs \u89e3\u51b3\u590d\u6742\u63a8\u7406\u4efb\u52a1\u7684\u80fd\u529b\u4ecd\u7136\u662f\u4e00\u5927\u6311\u6218\u3002\u6700\u8fd1\u7684\u4e00\u4e9b\u7814\u7a76 [6,7] \u5c1d\u8bd5\u901a\u8fc7\u76d1\u7763\u5fae\u8c03\uff08Supervised Fine-Tuning, SFT\uff09\u6765\u89e3\u51b3\uff0c\u901a\u8fc7\u5c06\u65b0\u7684\u63a8\u7406\u6570\u636e\u6837\u672c\u4e0e\u539f\u59cb\u6570\u636e\u96c6\u6df7\u5408\uff0c\u4f7f LLMs \u5b66\u4e60\u8fd9\u4e9b\u6837\u672c\u7684\u5e95\u5c42\u5206\u5e03\uff0c\u5e76\u5c1d\u8bd5\u6a21\u4eff\u6240\u5b66\u903b\u8f91\u6765\u89e3\u51b3\u672a\u89c1\u8fc7\u7684\u63a8\u7406\u4efb\u52a1\u3002\u5c3d\u7ba1\u8fd9\u79cd\u65b9\u6cd5\u6709\u6027\u80fd\u63d0\u5347\uff0c\u4f46\u5b83\u4e25\u91cd\u4f9d\u8d56\u4e8e\u5927\u91cf\u7684\u8bad\u7ec3\u548c\u989d\u5916\u7684\u6570\u636e\u51c6\u5907 [8,9]\u3002<\/p>\n

    Llama-3 \u62a5\u544a [10] \u5f3a\u8c03\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u89c2\u5bdf\uff1a\u5f53\u9762\u5bf9\u4e00\u4e2a\u5177\u6709\u6311\u6218\u6027\u7684\u63a8\u7406\u95ee\u9898\u65f6\uff0c\u6a21\u578b\u6709\u65f6\u4f1a\u751f\u6210\u6b63\u786e\u7684\u63a8\u7406\u8f68\u8ff9\u3002\u8fd9\u8868\u660e\u6a21\u578b\u77e5\u9053\u5982\u4f55\u4ea7\u751f\u6b63\u786e\u7b54\u6848\uff0c\u4f46\u5728\u9009\u62e9\u4e0a\u5b58\u5728\u56f0\u96be\u3002\u57fa\u4e8e\u8fd9\u4e00\u53d1\u73b0\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u7b80\u5355\u7684\u95ee\u9898\uff1a\u6211\u4eec\u80fd\u5426\u901a\u8fc7\u5e2e\u52a9 LLMs \u9009\u62e9\u6b63\u786e\u7684\u8f93\u51fa\u6765\u589e\u5f3a\u5b83\u4eec\u7684\u63a8\u7406\u80fd\u529b\uff1f\u4e3a\u63a2\u7d22\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u8fdb\u884c\u4e86\u4e00\u9879\u5b9e\u9a8c\uff0c\u5229\u7528\u4e0d\u540c\u7684\u5956\u52b1\u6a21\u578b\u8fdb\u884c LLMs \u8f93\u51fa\u9009\u62e9\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u6b65\u9aa4\u7ea7\u9009\u62e9\u663e\u8457\u4f18\u4e8e\u4f20\u7edf\u7684 CoT \u65b9\u6cd5\u3002<\/p>\n

    2. MindStar \u65b9\u6cd5<\/h4>\n

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

    \u56fe 2 MindStar \u7684\u7b97\u6cd5\u67b6\u6784\u56fe<\/p>\n

    \u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u7684\u63a8\u7406\u641c\u7d22\u6846\u67b6 \u2014\u2014MindStar\uff08M*\uff09\uff0c\u901a\u8fc7\u5c06\u63a8\u7406\u4efb\u52a1\u89c6\u4e3a\u641c\u7d22\u95ee\u9898\uff0c\u5e76\u5229\u7528\u8fc7\u7a0b\u76d1\u7763\u7684\u5956\u52b1\u6a21\u578b\uff08Process-supervised Reward Model, PRM\uff09\uff0cM * \u5728\u63a8\u7406\u6811\u7a7a\u95f4\u4e2d\u6709\u6548\u5bfc\u822a\uff0c\u8bc6\u522b\u8fd1\u4f3c\u6700\u4f18\u8def\u5f84\u3002\u7ed3\u5408\u675f\u641c\u7d22\uff08Beam Search, BS\uff09\u548c Levin \u6811\u641c\u7d22\uff08Levin Tree Search, LevinTS\uff09\u7684\u601d\u60f3\uff0c\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86\u641c\u7d22\u6548\u7387\uff0c\u5e76\u4fdd\u8bc1\u5728\u6709\u9650\u8ba1\u7b97\u590d\u6742\u5ea6\u5185\u627e\u5230\u6700\u4f73\u63a8\u7406\u8def\u5f84\u3002<\/p>\n

    2.1 \u8fc7\u7a0b\u76d1\u7763\u5956\u52b1\u6a21\u578b<\/strong><\/p>\n

    \u8fc7\u7a0b\u76d1\u7763\u5956\u52b1\u6a21\u578b (PRM) \u7684\u8bbe\u8ba1\u76ee\u7684\u662f\u8bc4\u4f30\u5927\u8bed\u8a00\u6a21\u578b (LLM) \u751f\u6210\u7684\u4e2d\u95f4\u6b65\u9aa4\uff0c\u4ee5\u5e2e\u52a9\u9009\u62e9\u6b63\u786e\u7684\u63a8\u7406\u8def\u5f84\u3002\u8fd9\u79cd\u65b9\u6cd5\u501f\u9274\u4e86\u5176\u4ed6\u5e94\u7528\u4e2d PRM \u7684\u6210\u529f\u7ecf\u9a8c\u3002\u5177\u4f53\u800c\u8a00\uff0cPRM \u4ee5\u5f53\u524d\u63a8\u7406\u8def\u5f84\"\u56fe\u7247\"\u548c\u6f5c\u5728\u7684\u4e0b\u4e00\u6b65\"\u56fe\u7247\"\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u8fd4\u56de\u5956\u52b1\u503c\"\u56fe\u7247\"\u3002<\/p>\n

    PRM \u901a\u8fc7\u8003\u8651\u6574\u4e2a\u5f53\u524d\u63a8\u7406\u8f68\u8ff9\u6765\u8bc4\u4f30\u65b0\u6b65\u9aa4\uff0c\u9f13\u52b1\u4e0e\u6574\u4f53\u8def\u5f84\u7684\u4e00\u81f4\u6027\u548c\u5fe0\u5b9e\u6027\u3002\u9ad8\u5956\u52b1\u503c\u8868\u660e\uff0c\u65b0\u7684\u6b65\u9aa4\"\u56fe\u7247\"\uff09\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u63a8\u7406\u8def\u5f84\"\u56fe\u7247\"\u53ef\u80fd\u662f\u6b63\u786e\u7684\uff0c\u4ece\u800c\u4f7f\u6269\u5c55\u8def\u5f84\u503c\u5f97\u8fdb\u4e00\u6b65\u63a2\u7d22\u3002\u76f8\u53cd\uff0c\u4f4e\u5956\u52b1\u503c\u5219\u8868\u793a\u65b0\u6b65\u9aa4\u53ef\u80fd\u4e0d\u6b63\u786e\uff0c\u8fd9\u610f\u5473\u7740\u9075\u5faa\u6b64\u8def\u5f84\u7684\u89e3\u51b3\u65b9\u6848\u4e5f\u53ef\u80fd\u4e0d\u6b63\u786e\u3002<\/p>\n

    M* \u7b97\u6cd5\u5305\u542b\u4e24\u4e2a\u4e3b\u8981\u6b65\u9aa4\uff0c\u8fed\u4ee3\u76f4\u5230\u627e\u5230\u6b63\u786e\u7684\u89e3\u51b3\u65b9\u6848\uff1a<\/p>\n

    1. \u63a8\u7406\u8def\u5f84\u6269\u5c55\uff1a\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\uff0c\u57fa\u7840 LLM \u751f\u6210\u5f53\u524d\u63a8\u7406\u8def\u5f84\u7684\u4e0b\u4e00\u6b65\u3002<\/p>\n

    2. \u8bc4\u4f30\u548c\u9009\u62e9\uff1a\u4f7f\u7528 PRM \u8bc4\u4f30\u751f\u6210\u7684\u6b65\u9aa4\uff0c\u5e76\u6839\u636e\u8fd9\u4e9b\u8bc4\u4f30\u9009\u62e9\u4e0b\u4e00\u6b21\u8fed\u4ee3\u7684\u63a8\u7406\u8def\u5f84\u3002<\/p>\n

    2.2 \u63a8\u7406\u8def\u5f84\u6269\u5c55<\/strong><\/p>\n

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

    \u5728\u9009\u62e9\u8981\u6269\u5c55\u7684\u63a8\u7406\u8def\u5f84\"\u56fe\u7247\"\u540e\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u63d0\u793a\u6a21\u677f\uff08Example 3.1\uff09\uff0c\u4ee5\u4ece LLM \u4e2d\u6536\u96c6\u4e0b\u4e00\u6b65\u3002\u6b63\u5982\u793a\u4f8b\u6240\u793a\uff0cLLM \u5c06\u539f\u59cb\u95ee\u9898\u4f5c\u4e3a {question}\uff0c\u5c06\u5f53\u524d\u63a8\u7406\u8def\u5f84\u4f5c\u4e3a {answer}\u3002\u6ce8\u610f\uff0c\u5728\u7b97\u6cd5\u7684\u7b2c\u4e00\u6b21\u8fed\u4ee3\u4e2d\uff0c\u6240\u9009\u62e9\u7684\u8282\u70b9\u662f\u4ec5\u5305\u542b\u95ee\u9898\u7684\u6839\u8282\u70b9\uff0c\u56e0\u6b64 {answer} \u4e3a\u7a7a\u3002\u5bf9\u4e8e\u63a8\u7406\u8def\u5f84\"\u56fe\u7247\"\uff0cLLM \u751f\u6210 N \u4e2a\u4e2d\u95f4\u6b65\u9aa4\uff0c\u5e76\u5c06\u5b83\u4eec\u4f5c\u4e3a\u5f53\u524d\u8282\u70b9\u7684\u5b50\u8282\u70b9\u9644\u52a0\u3002\u5728\u7b97\u6cd5\u7684\u4e0b\u4e00\u6b65\u4e2d\uff0c\u5c06\u8bc4\u4f30\u8fd9\u4e9b\u65b0\u751f\u6210\u7684\u5b50\u8282\u70b9\uff0c\u5e76\u9009\u62e9\u4e00\u4e2a\u65b0\u7684\u8282\u70b9\u8fdb\u884c\u8fdb\u4e00\u6b65\u6269\u5c55\u3002\u6211\u4eec\u8fd8\u610f\u8bc6\u5230\uff0c\u751f\u6210\u6b65\u9aa4\u7684\u53e6\u4e00\u79cd\u65b9\u6cd5\u662f\u4f7f\u7528\u6b65\u9aa4\u6807\u8bb0\u5bf9 LLM \u8fdb\u884c\u5fae\u8c03\u3002\u7136\u800c\uff0c\u8fd9\u53ef\u80fd\u4f1a\u964d\u4f4e LLM \u7684\u63a8\u7406\u80fd\u529b\uff0c\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8fd9\u4e0e\u672c\u6587\u7684\u91cd\u70b9 \u2014\u2014 \u5728\u4e0d\u4fee\u6539\u6743\u91cd\u7684\u60c5\u51b5\u4e0b\u589e\u5f3a LLM \u63a8\u7406\u80fd\u529b\u76f8\u6096\u3002<\/p>\n

    2.3 \u63a8\u7406\u8def\u5f84\u9009\u62e9<\/strong><\/p>\n

    \u5728\u6269\u5c55\u63a8\u7406\u6811\u540e\uff0c\u6211\u4eec\u4f7f\u7528\u9884\u8bad\u7ec3\u7684\u8fc7\u7a0b\u76d1\u7763\u5956\u52b1\u6a21\u578b\uff08PRM\uff09\u6765\u8bc4\u4f30\u6bcf\u4e2a\u65b0\u751f\u6210\u7684\u6b65\u9aa4\u3002\u6b63\u5982\u524d\u9762\u63d0\u5230\u7684\uff0cPRM \u91c7\u7528\u8def\u5f84\u548c\u6b65\u9aa4 \uff0c\u5e76\u8fd4\u56de\u76f8\u5e94\u7684\u5956\u52b1\u503c\u3002\u5728\u8bc4\u4f30\u4e4b\u540e\uff0c\u6211\u4eec\u9700\u8981\u4e00\u79cd\u6811\u641c\u7d22\u7b97\u6cd5\u6765\u9009\u62e9\u4e0b\u4e00\u4e2a\u8981\u6269\u5c55\u7684\u8282\u70b9\u3002\u6211\u4eec\u7684\u6846\u67b6\u4e0d\u4f9d\u8d56\u4e8e\u7279\u5b9a\u7684\u641c\u7d22\u7b97\u6cd5\uff0c\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5b9e\u4f8b\u5316\u4e86\u4e24\u79cd\u6700\u4f73\u4f18\u5148\u641c\u7d22\u65b9\u6cd5\uff0c\u5373 Beam Search \u548c Levin Tree Search\u3002<\/p>\n

    3. \u7ed3\u679c\u4e0e\u8ba8\u8bba<\/h4>\n

    \u5728 GSM8K \u548c MATH \u6570\u636e\u96c6\u4e0a\u7684\u5e7f\u6cdb\u8bc4\u4f30\u663e\u793a\uff0cM * \u663e\u8457\u63d0\u5347\u4e86\u5f00\u6e90\u6a21\u578b\uff08\u5982 LLaMA-2\uff09\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5176\u8868\u73b0\u53ef\u4e0e\u66f4\u5927\u89c4\u6a21\u7684\u95ed\u6e90\u6a21\u578b\uff08\u5982 GPT-3.5 \u548c Grok-1\uff09\u5ab2\u7f8e\uff0c\u540c\u65f6\u5927\u5e45\u51cf\u5c11\u4e86\u6a21\u578b\u89c4\u6a21\u548c\u8ba1\u7b97\u6210\u672c\u3002\u8fd9\u4e9b\u53d1\u73b0\u7a81\u663e\u4e86\u5c06\u8ba1\u7b97\u8d44\u6e90\u4ece\u5fae\u8c03\u8f6c\u79fb\u5230\u63a8\u7406\u65f6\u95f4\u641c\u7d22\u7684\u6f5c\u529b\uff0c\u4e3a\u672a\u6765\u9ad8\u6548\u63a8\u7406\u589e\u5f3a\u6280\u672f\u7684\u7814\u7a76\u5f00\u8f9f\u4e86\u65b0\u9014\u5f84\u3002<\/p>\n

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

    \u8868 1 \u5c55\u793a\u4e86\u5404\u79cd\u65b9\u6848\u5728 GSM8K \u548c MATH \u63a8\u7406\u57fa\u51c6\u4e0a\u7684\u5bf9\u6bd4\u7ed3\u679c\u3002\u6bcf\u4e2a\u6761\u76ee\u7684\u6570\u5b57\u8868\u793a\u95ee\u9898\u89e3\u51b3\u7684\u767e\u5206\u6bd4\u3002\u7b26\u53f7 SC@32 \u8868\u793a\u5728 32 \u4e2a\u5019\u9009\u7ed3\u679c\u4e2d\u7684\u81ea\u4e00\u81f4\u6027\uff0c\u800c n-shot \u8868\u793a\u5c11\u6837\u672c\u4f8b\u5b50\u7684\u7ed3\u679c\u3002CoT-SC@16 \u6307\u7684\u662f\u5728 16 \u4e2a\u601d\u7ef4\u94fe\uff08CoT\uff09\u5019\u9009\u7ed3\u679c\u4e2d\u7684\u81ea\u4e00\u81f4\u6027\u3002BS@16 \u4ee3\u8868\u675f\u641c\u7d22\u65b9\u6cd5\uff0c\u5373\u5728\u6bcf\u4e2a\u6b65\u9aa4\u7ea7\u522b\u6d89\u53ca 16 \u4e2a\u5019\u9009\u7ed3\u679c\uff0c\u800c LevinTS@16 \u8be6\u7ec6\u8bf4\u660e\u4e86\u4f7f\u7528\u76f8\u540c\u6570\u91cf\u5019\u9009\u7ed3\u679c\u7684 Levin \u6811\u641c\u7d22\u65b9\u6cd5\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cMATH \u6570\u636e\u96c6\u4e0a GPT-4 \u7684\u6700\u65b0\u7ed3\u679c\u4e3a GPT-4-turbo-0409\uff0c\u6211\u4eec\u7279\u522b\u5f3a\u8c03\u8fd9\u4e00\u70b9\uff0c\u56e0\u4e3a\u5b83\u4ee3\u8868\u4e86 GPT-4 \u5bb6\u65cf\u4e2d\u7684\u6700\u4f73\u6027\u80fd\u3002<\/p>\n

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

    \u56fe 3 \u6211\u4eec\u7814\u7a76\u4e86 M * \u6027\u80fd\u5982\u4f55\u968f\u7740\u6b65\u9aa4\u7ea7\u522b\u5019\u9009\u6570\u91cf\u7684\u53d8\u5316\u800c\u53d8\u5316\u3002\u6211\u4eec\u9009\u62e9 Llama-2-13B \u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\uff0c\u5e76\u5206\u522b\u9009\u62e9\u675f\u641c\u7d22\uff08BS\uff09\u4f5c\u4e3a\u641c\u7d22\u7b97\u6cd5\u3002<\/p>\n

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

    \u56fe 4 Llama-2 \u548c Llama-3 \u6a21\u578b\u5bb6\u65cf\u5728 MATH \u6570\u636e\u96c6\u4e0a\u7684\u5c3a\u5ea6\u5b9a\u5f8b\u3002\u6240\u6709\u7ed3\u679c\u5747\u6765\u81ea\u5b83\u4eec\u7684\u539f\u59cb\u8d44\u6e90\u3002\u6211\u4eec\u4f7f\u7528 Scipy \u5de5\u5177\u548c\u5bf9\u6570\u51fd\u6570\u6765\u8ba1\u7b97\u62df\u5408\u66f2\u7ebf\u3002<\/p>\n

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

    \u8868 2 \u4e0d\u540c\u65b9\u6cd5\u5728\u56de\u7b54\u95ee\u9898\u65f6\u7684\u5e73\u5747 token \u751f\u4ea7\u6570\u91cf<\/p>\n

    4. \u7ed3\u8bba<\/h4>\n

    \u672c\u6587\u4ecb\u7ecd\u4e86 MindStar\uff08M*\uff09\uff0c\u4e00\u79cd\u65b0\u9896\u7684\u57fa\u4e8e\u641c\u7d22\u7684\u63a8\u7406\u6846\u67b6\uff0c\u7528\u4e8e\u589e\u5f3a\u9884\u8bad\u7ec3\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002\u901a\u8fc7\u5c06\u63a8\u7406\u4efb\u52a1\u89c6\u4e3a\u641c\u7d22\u95ee\u9898\u5e76\u5229\u7528\u8fc7\u7a0b\u76d1\u7763\u7684\u5956\u52b1\u6a21\u578b\uff0cM*\u00a0\u5728\u63a8\u7406\u6811\u7a7a\u95f4\u4e2d\u6709\u6548\u5bfc\u822a\uff0c\u8bc6\u522b\u8fd1\u4f3c\u6700\u4f18\u8def\u5f84\u3002\u7ed3\u5408\u675f\u641c\u7d22\u548c Levin \u6811\u641c\u7d22\u7684\u601d\u60f3\uff0c\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86\u641c\u7d22\u6548\u7387\uff0c\u5e76\u4fdd\u8bc1\u5728\u6709\u9650\u8ba1\u7b97\u590d\u6742\u5ea6\u5185\u627e\u5230\u6700\u4f73\u63a8\u7406\u8def\u5f84\u3002\u5e7f\u6cdb\u7684\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cM*\u00a0\u663e\u8457\u63d0\u5347\u4e86\u5f00\u6e90\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5176\u8868\u73b0\u53ef\u4e0e\u66f4\u5927\u89c4\u6a21\u7684\u95ed\u6e90\u6a21\u578b\u5ab2\u7f8e\uff0c\u540c\u65f6\u5927\u5e45\u51cf\u5c11\u4e86\u6a21\u578b\u89c4\u6a21\u548c\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n

    \u8fd9\u4e9b\u7814\u7a76\u6210\u679c\u8868\u660e\uff0c\u5c06\u8ba1\u7b97\u8d44\u6e90\u4ece\u5fae\u8c03\u8f6c\u79fb\u5230\u63a8\u7406\u65f6\u95f4\u641c\u7d22\u5177\u6709\u5de8\u5927\u7684\u6f5c\u529b\uff0c\u4e3a\u672a\u6765\u9ad8\u6548\u63a8\u7406\u589e\u5f3a\u6280\u672f\u7684\u7814\u7a76\u5f00\u8f9f\u4e86\u65b0\u9014\u5f84\u3002<\/p>\n<\/div>\n<\/div>\n

    <\/div>\n

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

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

    \u672c\u6587\u4ecb\u7ecd\u4e86 MindStar\uff08M*\uff09\uff0c\u4e00\u79cd\u65b0\u9896\u7684\u57fa\u4e8e\u641c\u7d22\u7684\u63a8\u7406\u6846\u67b6\uff0c\u7528\u4e8e\u589e\u5f3a\u9884\u8bad\u7ec3\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002\u901a\u8fc7 […]<\/p>\n

    <\/div>\n

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

    <\/div>\n","protected":false},"author":56,"featured_media":20934,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12,23,20,80],"tags":[1177],"class_list":["post-20932","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-5g","category-23","category-20","category-80","tag--ai"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/20932","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=20932"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/20932\/revisions"}],"predecessor-version":[{"id":20935,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/20932\/revisions\/20935"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/20934"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=20932"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=20932"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=20932"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}