{"id":4104,"date":"2025-07-16T16:23:54","date_gmt":"2025-07-16T08:23:54","guid":{"rendered":"https:\/\/www.ihub1.com\/?p=4104"},"modified":"2025-07-16T16:26:05","modified_gmt":"2025-07-16T08:26:05","slug":"%e5%8f%aa%e5%9b%a0%e4%b8%80%e4%b8%aa%ef%bc%8c%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%85%a8%e5%86%9b%e8%a6%86%e6%b2%a1","status":"publish","type":"post","link":"https:\/\/www.ihub1.com\/index.php\/2025\/07\/16\/%e5%8f%aa%e5%9b%a0%e4%b8%80%e4%b8%aa%ef%bc%8c%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%85%a8%e5%86%9b%e8%a6%86%e6%b2%a1\/","title":{"rendered":"\u53ea\u56e0\u4e00\u4e2a\u201c:\u201d\uff0c\u5927\u6a21\u578b\u5168\u519b\u8986\u6ca1"},"content":{"rendered":"\t\t
\n\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

\u4e00\u4e2a\u5192\u53f7\uff0c\u7adf\u7136\u8ba9\u5927\u6a21\u578b\u96c6\u4f53\u7ffb\u8f66\uff1f<\/p>

<\/div>

\u660e\u660e\u5e94\u8be5\u88ab\u62e6\u4e0b\u6765\u7684\u865a\u5047\u56de\u7b54\uff0c\u7ed3\u679cLLM\u901a\u901a\u5f00\u7eff\u706f\u3002<\/p>

<\/div>

\u8be5\u53d1\u73b0\u6765\u81ea\u4e00\u7bc7\u540d\u53eb\u201c\u4e00\u4e2atoken\u5c31\u80fd\u6b3a\u9a97LLM\u201d\u7684\u8bba\u6587\u3002<\/p>

<\/div>

\u4e0d\u4ec5\u5982\u6b64\uff0c\u9664\u4e86<\/span>\u5192\u53f7<\/span>\u3001<\/span>\u7a7a\u683c<\/span>\u8fd9\u7c7b\u7b26\u53f7\uff0c\u8fd8\u6709\u8bf8\u5982\u6b64\u7c7b\u7684\u63a8\u7406\u5f00\u5934\u8bed\uff1a<\/span>\u201cThought process\uff1a\u201d<\/span>\u3001<\/span>\u201c\u89e3\u201d<\/span>\uff0c\u4e5f\u662f\u8f7b\u677e\u901a\u8fc7\u3002<\/span><\/p>

\u597d\u5bb6\u4f19\uff0c\u539f\u6765\u4e00\u4e2a\u201c\u89e3\u201d\u5b57\uff0c\u6570\u5b66\u8003\u8bd5\u80fd\u5f97\u5206\uff0cLLM\u4e5f\u4f1a\u88ab\u9a97\u5230\u2026\u2026<\/p>

<\/div>

\u800c\u4e14\u8fd9\u4e00\u6ce2\u662f\u51b2\u7740\u6240\u6709\u901a\u7528LLM\u6765\u7684\uff0c<\/span>GPT-4o<\/span>\u3001<\/span>Claude-4<\/span>\u3001<\/span>LLaMA3-70B<\/span>\u901a\u901a\u88ab\u65a9\u4e8e\u9a6c\u4e0b\u3002<\/span><\/p>

\u90a3\u548b\u529e\uff1fbug\u6709\u4e86\uff0c\u6765\u81ea<\/span>\u817e\u8baf<\/span>AI Lab<\/span>\u3001<\/span>\u666e\u6797\u65af\u987f\u5927\u5b66<\/span>\u548c<\/span>\u5f17\u5409\u5c3c\u4e9a\u5927\u5b66<\/span>\u7684\u7814\u7a76\u4eba\u5458\u5c31\u5f00\u59cb\u54fc\u54e7\u54fc\u54e7\u89e3bug\u3002<\/span><\/p>

<\/div>

\u7528\u589e\u5f3a\u6570\u636e\u96c6\u8bad\u7ec3\u51fa\u4e00\u4e2a\u9760\u8c31\u7684\u201c\u8bc4\u59d4\u201d\u6a21\u578b<\/span>Master-RM<\/span>\uff0c\u88ab\u9a97\u6982\u7387\u76f4\u63a5\u65e0\u9650\u63a5\u8fd10\uff0c\u6b63\u5e38\u8bc4\u4f30\u80fd\u529b\u8fd8\u80fd\u4e0d\u53d7\u5f71\u54cd\u3002<\/span><\/p>

\u5177\u4f53\u4ec0\u4e48\u60c5\u51b5\uff0c\u54b1\u4e14\u63a5\u7740\u5f80\u4e0b\u770b\u3002<\/p>

\u4e00\u628a\u80fd\u6b3a\u9a97LLM\u7684\u201c\u4e07\u80fd\u94a5\u5319\u201d<\/p>

\u8fd1\u6765\uff0c\u5229\u7528LLM\u5145\u5f53\u8bc4\u5224\u5de5\u5177\uff0c\u5728\u5e26\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60<\/span>\uff08RLVR\uff09<\/span>\u4e2d\u8bc4\u4f30\u7b54\u6848\u8d28\u91cf\u7684\u573a\u666f\u6108\u52a0\u666e\u904d\u3002<\/span><\/p>

LLM\u8bc4\u5224\u6a21\u578b\u901a\u8fc7\u6bd4\u5bf9\u751f\u6210\u7684\u5019\u9009\u7b54\u6848\u4e0e\u53c2\u8003\u7b54\u6848\uff0c\u8f93\u51fa\u4e8c\u5143\u5956\u52b1\u4fe1\u53f7\uff0c\u4ece\u800c\u6307\u5bfc\u7b56\u7565\u6a21\u578b\u66f4\u65b0\u3002<\/p>

\u7136\u800c\u7814\u7a76\u53d1\u73b0\uff0c<\/span>LLM\u201c\u5d29\u6e83\u201d\u4e86\uff1f<\/span><\/p>

\u54cd\u5e94\u957f\u5ea6\u4e0d\u4ec5\u9510\u51cf\u81f330 tokens\u4ee5\u4e0b\uff0c\u4e00\u4e9b\u610f\u4e49\u4e0d\u5927\u7684\u8bed\u53e5\u6216\u6587\u5b57\u7b26\u53f7\uff0c\u5374\u4eceLLM\u5904\u9a97\u5f97\u4e86\u5047\u9633\u6027\u5956\u52b1\uff0c\u4e5f\u5c31\u662f\u6253\u5f00LLM\u540e\u95e8\u7684\u4e00\u628a<\/span>\u201c\u4e07\u80fd\u94a5\u5319\u201d<\/span>\u3002<\/span><\/p>

<\/div>

\u8fd9\u628a\u80fd\u8bf1\u5bfcLLM\u8bc4\u5224\u6a21\u578b\u4ea7\u751f\u5047\u9633\u6027\u5224\u65ad\u7684\u201c\u4e07\u80fd\u94a5\u5319\u201d\u53ef\u5206\u4e3a\u4e24\u7c7b\uff1a<\/p>

  • \u975e\u6587\u5b57\u7b26\u53f7<\/span>\uff1a\u5982\u7a7a\u683c\u3001\u201c.\u201d\u3001\u201c,\u201d\u3001\u201c\uff1a\u201d\u3002<\/span><\/p><\/li>

  • \u63a8\u7406\u5f00\u5934\u8bed<\/span>\uff1a\u5982\u201cThought process:\u201d\u3001\u201cSolution\u201d\u3001\u201cLet\u2019s solve this problem step by step\u201d\u7b49\uff0c\u4ec5\u8868\u793a\u63a8\u7406\u5f00\u59cb\u4f46\u5e76\u6ca1\u6709\u5b9e\u8d28\u5185\u5bb9\u3002<\/span><\/p><\/li><\/ul>

    \u540c\u65f6\u4e3a\u4e86\u8fdb\u4e00\u6b65\u7814\u7a76\u8fd9\u79cd\u201c\u5956\u52b1\u6a21\u578b\u6b3a\u9a97\u201d\u73b0\u8c61\u662f\u5426\u5b58\u5728\u666e\u904d\u6027\uff0c\u7814\u7a76\u4eba\u5458\u5728\u591a\u6570\u636e\u96c6\u3001\u63d0\u793a\u8bcd\u683c\u5f0f\u4e0a\u5bf9\u5404\u79cdLLM\u5747\u8fdb\u884c\u4e86\u7cfb\u7edf\u6027\u8bc4\u4f30\u3002<\/p>

    <\/div>

    \u5b9e\u9a8c\u5206\u522b\u9009\u53d6\u4e24\u7c7b\u6a21\u578b\uff0c\u5206\u522b\u662f<\/span>\u4e13\u7528\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b<\/span>\uff08\u5982Multi-sub RM\u3001Omni-Judge\uff09<\/span>\uff0c\u4ee5\u53ca<\/span>\u901a\u7528LLM<\/span>\uff08\u5982GPT-4o\u3001Claude-4\u3001LLaMA3-70B\u3001Qwen2.5-72B\u7b49\uff09<\/span>\u3002<\/span><\/p>

    \u4e13\u7528\u6a21\u578b\u4f7f\u7528\u9ed8\u8ba4\u63d0\u793a\uff0c\u800c\u901a\u7528LLM\u91c7\u7528\u6807\u51c6\u5316\u63d0\u793a\u6a21\u677f\u3002<\/p>

    \u7136\u540e\u9009\u62e910\u79cd\u53ef\u89e6\u53d1\u5047\u9633\u6027\u7684\u5bf9\u6297\u6027\u54cd\u5e94\uff0c\u5305\u62ec\u975e\u6587\u5b57\u7b26\u53f7<\/span>\uff08\u5982\u7a7a\u683c\u3001\u201c:\u201d\uff09<\/span>\u548c\u591a\u8bed\u8a00\u63a8\u7406\u5f00\u5934\u8bed<\/span>\uff08\u5982\u82f1\u6587\u7684 \u201cThought process:\u201d\u3001\u4e2d\u6587\u7684\u201c\u89e3\u201d\u3001\u65e5\u8bed\u7684\u201c\u304b\u3044\u305b\u3064\u201d\uff09<\/span>\u3002<\/span><\/p>

    \u53e6\u5916\u4e3a\u4e86\u6d4b\u8bd5\u6a21\u578b\u8de8\u9886\u57df\u7684\u7a33\u5065\u6027\uff0c\u5b9e\u9a8c\u6db5\u76d6\u901a\u7528\u63a8\u7406\u548c\u6570\u5b66\u63a8\u7406\u7684\u51715\u4e2a\u63a8\u7406\u57fa\u51c6\u3002<\/p>

    <\/div>

    \u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c<\/span>\u6240\u6709\u6d4b\u8bd5\u6a21\u578b\u65e0\u4e00\u5e78\u514d<\/span>\uff0c\u5168\u90e8\u90fd\u4f1a\u89e6\u53d1\u5047\u9633\u6027\u54cd\u5e94\u3002<\/span><\/p>

    \u4f8b\u5982GPT-4o\u5bf9\u7b26\u53f7 \u201c:\u201d \u7684\u5047\u9633\u6027\u7387<\/span>\uff08FPR\uff09<\/span>\u53ef\u8fbe<\/span>35%<\/span>\uff0cLLaMA3-70B\u5bf9 \u201cThought process:\u201d \u7684FPR\u751a\u81f3\u9ad8\u8fbe<\/span>60%-90%<\/span>\uff0c\u4e13\u6709\u6a21\u578bGeneral-Verifier\u5728MATH\u6570\u636e\u96c6\u4e0a\u5bf9\u7a7a\u683c\u7684FPR\u4e5f\u8fbe<\/span>66.8%<\/span>\u3002<\/span><\/p>

    \u53e6\u5916\uff0c\u4e0d\u540c\u8bed\u8a00\u4e5f\u4e0d\u4f1a\u5f71\u54cd\u8fd9\u79cd\u6b3a\u9a97\u73b0\u8c61\u7684\u51fa\u73b0\uff0c\u65e0\u8bba\u662f\u4e2d\u6587\u8fd8\u662f\u65e5\u8bed\uff0c\u90fd\u540c\u6837\u80fd\u591f\u8bf1\u53d1\u9ad8FPR\uff0c\u8be5\u6f0f\u6d1e\u5177\u6709<\/span>\u8de8\u8bed\u8a00<\/span>\u7684\u666e\u904d\u6027\u3002<\/span><\/p>

    <\/div>

    \u7814\u7a76\u4eba\u5458\u8fd8\u5206\u6790\u4e86<\/span>0.5B\u81f372B<\/span>\u7684<\/span>Qwen2.5-Instruct<\/span>\u7cfb\u5217\u6a21\u578b\uff0c\u53d1\u73b0\uff1a<\/span><\/p>

    • 0.5B\u6a21\u578b<\/span>\uff1a\u4f9d\u8d56\u5b57\u9762\u5339\u914d\uff0cFPR\u4f4e\u4f46\u4e0eGPT-4o\u4e00\u81f4\u6027\u5dee\uff1b<\/span><\/p><\/li>

    • 1.5B-3B\u6a21\u578b<\/span>\uff1a\u80fd\u68c0\u6d4b\u8bed\u4e49\u76f8\u4f3c\u6027\u4f46\u7f3a\u4e4f\u7cbe\u7ec6\u9a8c\u8bc1\uff0cFPR\u9aa4\u5347\uff1b<\/span><\/p><\/li>

    • 7B-14B\u6a21\u578b<\/span>\uff1a\u5e73\u8861\u9a8c\u8bc1\u80fd\u529b\u4e0e\u8c28\u614e\u6027\uff0cFPR\u6700\u4f4e\u4e14\u4e00\u81f4\u6027\u9ad8\uff1b<\/span><\/p><\/li>

    • 32B-72B\u6a21\u578b<\/span>\uff1a\u56e0\u4e3a\u66f4\u503e\u5411\u4e8e\u81ea\u5df1\u89e3\u9898\u800c\u975e\u5bf9\u6bd4\u54cd\u5e94\u4e0e\u53c2\u8003\u7b54\u6848\uff0cFPR\u518d\u6b21\u4e0a\u5347\u3002<\/span><\/p><\/li><\/ul>

      \u6240\u4ee5\u6a21\u578b\u7684\u5927\u5c0f\u4e0eFPR\u4e4b\u95f4\u5e76\u975e\u5b8c\u5168\u7684\u5355\u8c03\u5173\u7cfb\uff0c<\/span>\u4e0d\u662f\u6a21\u578b\u8d8a\u5927\u5c31\u8d8a\u4e0d\u5bb9\u6613\u88ab\u9a97<\/span>\u3002<\/span><\/p>

      \u5982\u679c\u60f3\u901a\u8fc7\u4e00\u4e9b\u63a8\u7406\u65f6\u7684\u6280\u5de7\u6765\u51cf\u5c11\u8fd9\u79cd\u6f0f\u6d1e\uff0c\u6548\u679c\u4e5f\u4e0d\u592a\u7a33\u5b9a\uff0c\u8fd8\u5f97\u770b\u5177\u4f53\u6a21\u578b\u548c\u5e94\u7528\u573a\u666f\u3002<\/p>

      \u6b64\u5916\uff0c\u7814\u7a76\u4eba\u5458\u8fd8\u53d1\u73b0\uff0c<\/span>\u8fd9\u79cdbug\u8fd8\u80fd\u65e0\u9650\u7e41\u6b96\u2026\u2026<\/span><\/p>

      \u53ea\u9700\u8981\u57fa\u4e8eall-MiniLM-L6-v2\u7f16\u7801\u5668\u8fdb\u884c\u5d4c\u5165\u76f8\u4f3c\u5ea6\u641c\u7d22\uff0c\u4ece\u5927\u89c4\u6a21\u8bed\u6599\u4e2d\u81ea\u52a8\u751f\u6210\u4e0e\u5df2\u77e5 \u201c\u4e07\u80fd\u94a5\u5319\u201d \u76f8\u4f3c\u7684\u65b0\u5bf9\u6297\u6027\u54cd\u5e94\uff0c\u65b0\u7684\u201c\u4e07\u80fd\u94a5\u5319\u201d\u5c31\u80fd\u540c\u6837\u4ea7\u751f\u51fa\u9ad8\u6c34\u5e73FPR\u3002<\/p>

      <\/div>

      \u5b9e\u9a8c\u6700\u7ec8\u8bf4\u660e\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\u5176\u5b9e\u5b58\u5728\u4e00\u4e2a\u76f8\u5f53\u5173\u952e\u7684\u6838\u5fc3\u673a\u5236\u6f0f\u6d1e\uff1a<\/span>\u539f\u672c\u7528\u4e8e\u8fc7\u6ee4\u65e0\u6548\u6216\u9519\u8bef\u7b54\u6848\u7684\u9a8c\u8bc1\u5668\uff0c\u5bb9\u6613\u88ab\u65e0\u5173\u7d27\u8981\u7684\u8868\u9762\u5185\u5bb9\u64cd\u7eb5\uff0c\u4ece\u800c\u4ea7\u751f\u5047\u9633\u6027\u7ed3\u679c\u3002<\/span><\/p>

      \u8fd9\u5bf9\u4efb\u4f55\u4f9d\u8d56\u9a8c\u8bc1\u5668\u63d0\u4f9b\u53cd\u9988\u7684RLVR\u6d41\u7a0b\u90fd\u63d0\u51fa\u4e86\u7834\u574f\u6027\u7684\u6311\u6218\u3002<\/p>

      \u4e00\u4e2a\u4e0d\u4f1a\u88ab\u9a97\u7684\u201c\u8bc4\u59d4\u201d\u6a21\u578b<\/p>

      \u4e3a\u4e86\u7f13\u89e3\u201c\u4e07\u80fd\u94a5\u5319\u201d\u7684\u5f71\u54cd\uff0c\u7814\u7a76\u4eba\u5458\u4e13\u95e8\u6784\u5efa\u4e86\u65b0\u7684\u201c\u8bc4\u59d4\u201d\u6a21\u578b<\/span>Master-RM<\/span>\uff08Master Reward Model\uff09<\/span>\u3002<\/span><\/p>

      \u9996\u5148\u4ece\u539f\u59cb\u768416\u4e07\u6761\u8bad\u7ec3\u6570\u636e\u4e2d\u968f\u673a\u91c7\u68372\u4e07\u6761\uff0c\u7528GPT-4o-mini\u751f\u6210\u5e26\u63a8\u7406\u5f00\u5934\u8bed\u53e5\u7684\u54cd\u5e94\uff0c\u4f46\u4ec5\u4fdd\u7559\u65e0\u5b9e\u8d28\u5185\u5bb9\u7684\u7b2c\u4e00\u53e5\u8bdd\uff0c\u5e76\u6807\u8bb0\u4e3a\u201c\u9519\u8bef\u201d\u3002<\/p>

      \u5c06\u8fd92\u4e07\u6761\u5bf9\u6297\u6837\u672c\u4e0e\u539f\u59cb\u6570\u636e\u7ed3\u5408\uff0c\u6784\u6210<\/span>\u589e\u5f3a\u8bad\u7ec3\u6570\u636e\u96c6<\/span>\u3002<\/span><\/p>

      \u7136\u540e\u57fa\u4e8e<\/span>Qwen2.5-7B-Instruct<\/span>\u8fdb\u884c\u6709\u76d1\u7763\u5fae\u8c03<\/span>\uff08SFT\uff09<\/span>\uff0c\u4fdd\u8bc1\u6700\u5c0f\u5316\u4ea4\u53c9\u71b5\u635f\u5931\uff0c\u8ba9\u6a21\u578b\u5b66\u4e60\u5982\u4f55\u533a\u5206\u6709\u6548\u54cd\u5e94\u4e0e\u8868\u9762\u6b3a\u9a97\u6027\u54cd\u5e94\u3002<\/span><\/p>

      \u5c06Master-RM\u653e\u5165\u76f8\u540c\u6761\u4ef6\u4e0b\u5b9e\u9a8c\u518d\u6b21\u9a8c\u8bc1\uff0c\u53d1\u73b0\u6b64\u65f6\u5728\u8de8\u6570\u636e\u96c6\u6d4b\u8bd5\u4e2d\uff0c\u6a21\u578b\u5bf9\u6240\u6709 \u201c\u4e07\u80fd\u94a5\u5319\u201d \u7684\u5047\u9633\u6027\u7387\u63a5\u8fd1<\/span>0%<\/span>\uff08\u751a\u81f3\u5b8c\u5168\u4e3a\u96f6\uff09<\/span>\uff0c\u4e14\u9c81\u68d2\u6027\u53ef\u6cdb\u5316\u5230\u672a\u89c1\u8fc7\u7684\u6570\u636e\u96c6\u548c\u6b3a\u9a97\u653b\u51fb\u4e2d\u3002<\/span><\/p>

      <\/div>

      \u540c\u65f6\u6a21\u578b\u4fdd\u6301\u4e0eGPT-4o\u7684\u8bc4\u4f30\u4e00\u81f4\u6027\u53ef\u8fbe<\/span>0.96<\/span>\uff0c\u9a8c\u8bc1\u4e86\u5176\u4f5c\u4e3a\u901a\u7528\u9886\u57df\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\u7684\u6709\u6548\u6027\u3002<\/span><\/p>

      \u6240\u4ee5LLM\u4f5c\u4e3a\u201c\u8bc4\u59d4\u201d\u6a21\u578b\u5176\u5b9e\u76f8\u5f53\u8106\u5f31\uff0c<\/span>\u5c0f\u5c0f\u4e00\u4e2a\u5192\u53f7<\/span>\u5c31\u53ef\u80fd\u8ba9\u5b83\u51fa\u9519\u3002<\/span><\/p>

      \u56e0\u6b64\u6709\u7f51\u53cb\u8868\u793a\uff0c\u8be5\u53d1\u73b0\u63ed\u793a\u4e86\u6a21\u578b\u7a33\u5065\u7684\u91cd\u8981\u6027\uff0c\u800cRLHF\u4e5f\u9700\u8981\u4e25\u683c\u5bf9\u6297\u8bc4\u4f30\uff0c\u6784\u5efa\u66f4\u4e3a\u53ef\u9760\u7684LLM\u5de5\u4f5c\u6d41\u7a0b\u3002<\/p>

      <\/div>

      \u4f5c\u8005\u672c\u4eba\u4e5f\u73b0\u8eab\u8bc4\u8bba\u533a\uff0c\u4ed6\u8ba4\u4e3a\uff0c\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\u5bb9\u6613\u53d7\u5230\u865a\u5047\u5956\u52b1\u653b\u51fb\uff0c\u5982\u4f55\u66f4\u597d\u5730\u907f\u514d\u7c7b\u4f3c\u60c5\u51b5\u53d1\u751f\uff0c\u5c06\u662f\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u3002<\/p>

      <\/div>

      \u5168\u534e\u4eba\u56e2\u961f<\/p>

      \u6700\u540e\u6765\u770b\u4e0b\u7814\u7a76\u56e2\u961f\uff0c\u5206\u522b\u6765\u81ea\u817e\u8bafAI Lab\u3001\u666e\u6797\u65af\u987f\u5927\u5b66\u548c\u5f17\u5409\u5c3c\u4e9a\u5927\u5b66\u3002<\/p>

      \u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u5176\u4e2d\u770b\u5230\u4e86\u5927\u4f6c<\/span>\u4fde\u680b<\/span>\u7684\u8eab\u5f71\u3002<\/span><\/p>

      \u4ed6\u88ab\u79f0\u4e3a\u817e\u8bafAI Lab\u4e09\u5251\u5ba2\u4e4b\u4e00\uff0c\u76ee\u524d\u662f\u817e\u8baf\u4e91\u4eba\u5de5\u667a\u80fd\u9996\u5e2d\u79d1\u5b66\u5bb6\u517c\u526f\u603b\u7ecf\u7406\uff0c\u4e4b\u524d\u66fe\u662f\u5fae\u8f6f\u9996\u5e2d\u7814\u7a76\u5458\uff0c2017\u5e74\u52a0\u5165\u817e\u8baf\u3002<\/p>

      \u4ed6\u5728\u6df1\u5ea6\u5b66\u4e60\u7684\u81ea\u52a8\u8bed\u97f3\u8bc6\u522b\u548c\u5904\u7406\u9886\u57df\u6df1\u8015\u591a\u5e74\uff0c\u5171\u53d1\u8868\u4e24\u672c\u4e13\u8457\u548c400\u591a\u7bc7\u8bba\u6587\uff0c\u5e76\u83b7\u5f974\u5e74\u7684IEEE\u4fe1\u53f7\u5904\u7406\u5b66\u4f1a\u6700\u4f73\u8bba\u6587\u5956\u30012021\u5e74NAACL\u6700\u4f73\u957f\u8bba\u6587\u5956\u7b49\u3002<\/p>

      <\/div>

      \u6b64\u5916\uff0c\u8bba\u6587\u4e00\u4f5c<\/span>Yulai Zhao<\/span>\u672c\u79d1\u6bd5\u4e1a\u4e8e\u6e05\u534e\u5927\u5b66\uff0c\u76ee\u524d\u5728\u666e\u6797\u65af\u987f\u5927\u5b66\u653b\u8bfb\u673a\u5668\u5b66\u4e60\u7684\u535a\u58eb\u5b66\u4f4d\uff0c\u53e6\u5916\u4ed6\u8fd8\u5728\u817e\u8bafAI Lab\u62c5\u4efb\u7814\u7a76\u4eba\u5458\u3002<\/span><\/p>

      <\/div>

      \u7814\u7a76\u65b9\u5411\u4e3b\u8981\u662f\u901a\u8fc7\u6570\u636e\u9a71\u52a8\u7684\u65b9\u6cd5\u63a2\u7d22\u73b0\u4ee3\u5f3a\u5316\u5b66\u4e60\u548c\u6269\u6563\u6a21\u578b\uff0c\u53e6\u5916\u4ed6\u7684\u4e00\u7bc7\u6709\u5173\u8ba9\u6269\u6563\u6a21\u578b\u5728\u751f\u6210\u86cb\u767d\u8d28\u548cDNA\u5e8f\u5217\u8868\u73b0\u66f4\u4f18\u7684\u8bba\u6587\uff0c\u6700\u8fd1\u521a\u521a\u88abICML 2025\u5f55\u7528\u3002<\/p>

      \u5171\u540c\u4e00\u4f5c<\/span>Haolin Liu<\/span>\u5219\u662f\u5f17\u5409\u5c3c\u4e9a\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u7cfb\u7684\u535a\u58eb\u751f\uff0c\u5e08\u4eceChen-Yu Wei\u6559\u6388\u3002<\/span><\/p>

      \u4e4b\u524d\u5728\u4e0a\u6d77\u79d1\u6280\u5927\u5b66\u5c31\u8bfb\u672c\u79d1\uff0c\u5148\u5b66\u4e60\u4e861.5\u5e74\u5316\u5b66\uff0c\u540e\u624d\u8f6c\u4e3a\u8ba1\u7b97\u673a\u79d1\u5b66\uff0c\u4e3b\u8981\u7814\u7a76\u4ee5\u5f3a\u5316\u5b66\u4e60\u4e3a\u4e2d\u5fc3\uff0c\u81f4\u529b\u4e8e\u63a8\u8fdbRL\u5728LLM\u540e\u8bad\u7ec3\u4e2d\u7684\u5e94\u7528\u3002<\/p>

      Dian Yu<\/span>\u76ee\u524d\u662f\u817e\u8bafAI Lab\u7684\u4e00\u540dNLP\u7814\u7a76\u5458\uff0c\u66fe\u5728\u4f26\u65af\u52d2\u7406\u5de5\u5b66\u9662\u53d6\u5f97\u535a\u58eb\u5b66\u4f4d\uff0c\u76ee\u524d\u7814\u7a76\u65b9\u5411\u5305\u542b\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u4fe1\u606f\u62bd\u53d6\u3001\u673a\u5668\u9605\u8bfb\u7406\u89e3\u548c\u5bf9\u8bdd\u7406\u89e3\u3002<\/span><\/p>

      \u53e6\u5916\uff0c\u4f5c\u8005\u91cc\u8fd8\u6709\u666e\u6797\u65af\u987f\u5927\u5b66\u7535\u5b50\u5de5\u7a0b\u4e13\u4e1a\u7684<\/span>\u8d21\u4e09\u5143<\/span>\u6559\u6388\uff0c\u5176\u7814\u7a76\u9886\u57df\u5305\u62ec\u673a\u5668\u5b66\u4e60\u3001\u7cfb\u7edf\u5efa\u6a21\u4e0e\u8bc6\u522b\u3001\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u7b49\uff0c\u53d1\u8868\u8fc7400\u4f59\u7bc7\u8bba\u6587\u53ca\u4e13\u8457\u3002<\/span><\/p>

      <\/div>

      \u56e0\u5176\u5bf9VLSI\u4fe1\u53f7\u5904\u7406\u548c\u795e\u7ecf\u7f51\u7edc\u7684\u8d21\u732e\uff0c\u88ab\u8868\u5f70\u4e3aIEEE\u7ec8\u751f\u4f1a\u58eb\uff0c\u8fd8\u83b7\u5f97\u8fc7IEEE\u4fe1\u53f7\u5904\u7406\u5b66\u4f1a\u6700\u4f73\u8bba\u6587\u5956\u3001IEEE\u4fe1\u53f7\u5904\u7406\u5b66\u4f1a\u6280\u672f\u6210\u5c31\u5956\u7b49\u3002<\/p>

      Haitao Mi<\/span>\u535a\u58eb\u6bd5\u4e1a\u4e8e\u4e2d\u56fd\u79d1\u5b66\u9662\u8ba1\u7b97\u6280\u672f\u7814\u7a76\u6240\uff0c\u66fe\u5728\u652f\u4ed8\u5b9d\u4e2d\u62c5\u4efb\u9996\u5e2d\u7814\u7a76\u5458\u3002<\/span><\/p>

      <\/div>

      \u76ee\u524d\u662f\u817e\u8bafAI Lab\u7684\u9996\u5e2d\u7814\u7a76\u5458,\u96b6\u5c5e\u4e8e\u65d7\u4e0b\u7684\u8bed\u8a00\u667a\u80fd\u7814\u7a76\u5c0f\u7ec4\uff0c\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u662f\u6269\u5c55\u5927\u578b\u57fa\u7840\u6a21\u578b\u548c\u4e0b\u4e00\u4ee3\u667a\u80fd\u4f53\u7cfb\u7edf\u3002<\/p>

      \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2507.08794<\/p>

      \u6570\u636e\u96c6\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/datasets\/sarosavo\/Master-RM<\/p>

      \u6a21\u578b\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/sarosavo\/Master-RM<\/p>

      \u53c2\u8003\u94fe\u63a5\uff1a<\/p>

      [1]https:\/\/x.com\/omarsar0\/status\/1944778174493343771<\/p>

      [2]https:\/\/yulaizhao.com\/<\/p>

      [3]https:\/\/liuhl2000.github.io\/<\/p>

      [4]https:\/\/ece.princeton.edu\/people\/sun-yuan-kung<\/p>

      [5]https:\/\/openreview.net\/profile?id=~Haitao_Mi1<\/p>

      [6]https:\/\/sites.google.com\/view\/dongyu888\/<\/p>

      [7]https:\/\/openreview.net\/profile?id=~Dian_Yu3<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"

      \u9e6d\u7fbd \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n","protected":false},"author":1,"featured_media":4106,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4104","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/4104","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/comments?post=4104"}],"version-history":[{"count":7,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/4104\/revisions"}],"predecessor-version":[{"id":4112,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/4104\/revisions\/4112"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media\/4106"}],"wp:attachment":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media?parent=4104"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/categories?post=4104"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/tags?post=4104"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}