\u4e00\u9053\u6570\u5b66\u9898\u6362\u4e2a\u95ee\u6cd5\u3001\u610f\u601d\u548c\u7b54\u6848\u90fd\u6ca1\u53d8\uff0c\u4e00\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\u5374\u53ef\u80fd\u4ece\u7b54\u5bf9\u53d8\u6210\u7b54\u9519\u2014\u2014\u8fd9\u79cd\u770b\u4f3c\u79bb\u8c31\u7684\u73b0\u8c61\uff0c\u88ab\u4e00\u7bc7\u7531\u6d59\u6c5f\u5927\u5b66\u3001\u897f\u6e56\u5927\u5b66\u4e0e\u4e0a\u6d77\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u56e2\u961f\u63d0\u4ea4\u7684\u8bba\u6587\u7cfb\u7edf\u62c6\u89e3\u3002\u7814\u7a76\u8005\u628a\u6a21\u578b\u5bf9\u9898\u76ee”\u8868\u9762\u63aa\u8f9e”\u7684\u8106\u5f31\u4f9d\u8d56\u79f0\u4e3a\u6f5c\u5728\u7684\u865a\u5047\u4f9d\u8d56\uff0c\u5e76\u636e\u6b64\u63d0\u51fa\u4e00\u79cd\u65b0\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u65b9\u6cd5 SCAPO\uff0c\u5728\u4e0d\u6539\u52a8\u6a21\u578b\u67b6\u6784\u7684\u524d\u63d0\u4e0b\u663e\u8457\u63d0\u5347\u4e86\u6570\u5b66\u63a8\u7406\u51c6\u786e\u7387\u4e0e\u6297\u5e72\u6270\u80fd\u529b\u3002<\/p>\n
\u8bba\u6587\u7684\u6838\u5fc3\u8d77\u70b9\u662f\u5f53\u524d\u4e3b\u6d41\u7684 RLVR\uff08\u53ef\u9a8c\u8bc1\u5956\u52b1\u5f3a\u5316\u5b66\u4e60\uff09\u8bad\u7ec3\u8303\u5f0f\uff0c\u5c24\u5176\u662f\u5176\u4e2d\u6700\u5e38\u7528\u7684 GRPO \u7b97\u6cd5\u3002\u5b83\u7684\u903b\u8f91\u5f88\u76f4\u63a5\uff1a\u8ba9\u6a21\u578b\u751f\u6210\u4e00\u957f\u4e32\u63a8\u7406\u8fc7\u7a0b\uff0c\u53ea\u770b\u6700\u7ec8\u7b54\u6848\u5bf9\u4e0d\u5bf9\uff0c\u5bf9\u4e86\u5c31\u7ed9\u6574\u6bb5\u56de\u7b54\u540c\u4e00\u4e2a\u6b63\u4f18\u52bf\u3001\u9519\u4e86\u5c31\u7ed9\u8d1f\u4f18\u52bf\u3002\u95ee\u9898\u5728\u4e8e\uff0c\u4e00\u65e6\u7b54\u6848\u88ab\u5224\u5bf9\uff0c\u56de\u7b54\u91cc\u7684\u6bcf\u4e00\u4e2a\u8bcd\u2014\u2014\u65e0\u8bba\u662f\u5173\u952e\u7684\u6570\u5b66\u7b26\u53f7\uff0c\u8fd8\u662f”\u4e0d\u8fc7””\u8ba9\u6211\u518d\u60f3\u60f3”\u8fd9\u7c7b\u8fc7\u6e21\u8bcd\u2014\u2014\u90fd\u62ff\u5230\u4e86\u76f8\u540c\u7684\u4fe1\u7528\u3002\u6a21\u578b\u65e0\u4ece\u533a\u5206\u54ea\u4e9b\u8bcd\u771f\u6b63\u63a8\u52a8\u4e86\u6b63\u786e\u89e3\u7b54\uff0c\u54ea\u4e9b\u53ea\u662f\u6070\u597d\u88ab\u91c7\u6837\u5230\u3001\u4e0e\u7b54\u6848\u5e76\u65e0\u56e0\u679c\u5173\u7cfb\u7684\u566a\u58f0\u3002<\/p>\n
\u4e3a\u4e86\u91cf\u5316\u8fd9\u79cd\u654f\u611f\u5ea6\uff0c\u7814\u7a76\u8005\u8bbe\u8ba1\u4e86”\u534a\u4e8b\u5b9e\u6270\u52a8”\uff1a\u5728\u4fdd\u6301\u9898\u76ee\u6570\u5b66\u5185\u5bb9\u4e0e\u6807\u51c6\u7b54\u6848\u5b8c\u5168\u4e0d\u53d8\u7684\u524d\u63d0\u4e0b\uff0c\u53ea\u6539\u52a8\u4e0e\u89e3\u9898\u65e0\u5173\u7684\u8868\u9762\u7279\u5f81\u3002\u5177\u4f53\u505a\u6cd5\u5305\u62ec\u628a\u53e5\u5b50\u6539\u5199\u6210\u540c\u4e49\u8868\u8ff0\u3001\u5728\u65e0\u5173\u5355\u8bcd\u91cc\u6545\u610f\u6253\u9519\u4e00\u4e2a\u5b57\u6bcd\u3001\u7ed9\u9898\u76ee\u52a0\u4e0a\u573a\u666f\u5305\u88c5\uff0c\u6216\u5728\u672b\u5c3e\u9644\u4e00\u4e32\u65e0\u610f\u4e49\u7684\u6807\u7b7e\u3002\u4ed6\u4eec\u7528 GPT-5.5 \u751f\u6210\u8fd9\u4e9b\u53d8\u4f53\uff0c\u5e76\u4e25\u683c\u8981\u6c42\u6570\u5b57\u3001\u516c\u5f0f\u3001\u7ea6\u675f\u4e0e\u7b54\u6848\u4e00\u5b57\u4e0d\u6539\u3002\u968f\u540e\uff0c\u8bba\u6587\u7528”\u6982\u7387\u6f02\u79fb”\u6307\u6807\u8861\u91cf\u540c\u4e00\u4e2a\u8bcd\u5728\u539f\u9898\u4e0e\u6270\u52a8\u9898\u4e0b\u88ab\u5206\u914d\u6982\u7387\u7684\u5dee\u8ddd\uff0c\u53d6\u503c\u5f52\u4e00\u5316\u5230 0 \u5230 2 \u4e4b\u95f4\u3002<\/p>\n
\u628a 1000 \u9053\u6570\u5b66\u9898\u8dd1\u5b8c\u8fd9\u5957\u6d41\u7a0b\u540e\uff0c\u7ed3\u679c\u76f8\u5f53\u624e\u5fc3\uff1a33.4% \u7684\u8bcd\u5728\u6270\u52a8\u540e\u6982\u7387\u4e0a\u5347\uff0c42.6% \u4e0b\u964d\uff0c\u53ea\u6709 24% \u7eb9\u4e1d\u4e0d\u52a8\u3002\u66f4\u5173\u952e\u7684\u662f\u654f\u611f\u5ea6\u5728\u4e0d\u540c\u7c7b\u578b\u8bcd\u4e0a\u5dee\u5f02\u5de8\u5927\u2014\u2014\u5e26\u6709”\u53cd\u601d”\u610f\u5473\u7684\u8bcd\uff08\u5982”\u7b49\u4e00\u4e0b””\u518d\u68c0\u67e5\u4e00\u4e0b”\uff09\u654f\u611f\u5ea6\u662f\u5e73\u5747\u6c34\u5e73\u7684 2.74 \u500d\uff0c\u903b\u8f91\u8fde\u63a5\u8bcd\uff08\u5982”\u56e0\u6b64””\u6240\u4ee5”\uff09\u4e3a 2.32 \u500d\uff0c\u800c\u6570\u5b66\u7b26\u53f7\u548c\u6570\u5b57\u53ea\u6709 0.47 \u548c 0.37 \u500d\u3002\u6362\u8a00\u4e4b\uff0c\u6a21\u578b\u5185\u90e8\u6700\u4e0d\u7a33\u7684\uff0c\u6070\u6070\u662f\u90a3\u4e9b\u7ec4\u7ec7\u8bed\u8a00\u3001\u81ea\u6211\u53cd\u601d\u7684”\u5199\u4f5c\u4e60\u60ef”\uff0c\u771f\u6b63\u505a\u8ba1\u7b97\u7684\u90e8\u5206\u53cd\u800c\u5f88\u7a33\uff0c\u8fd9\u8bf4\u660e\u6a21\u578b\u53ef\u80fd\u628a\u4e00\u90e8\u5206\u4e0e\u63a8\u7406\u65e0\u5173\u7684\u8868\u9762\u89c4\u5f8b\u5f53\u6210\u4e86\u89e3\u9898\u4f9d\u636e\u3002<\/p>\n
\u6709\u610f\u601d\u7684\u662f\uff0c\u7814\u7a76\u8005\u5148\u5728\u63a8\u7406\u9636\u6bb5\u505a\u4e86\u4e00\u6b21”\u7eaf\u8fc7\u6ee4”\u9a8c\u8bc1\uff1a\u4e0d\u66f4\u65b0\u4efb\u4f55\u6743\u91cd\uff0c\u53ea\u5728\u67d0\u4e9b\u751f\u6210\u6b65\u9aa4\u5c4f\u853d\u6389\u6982\u7387\u6f02\u79fb\u6700\u9ad8\u7684\u90a3\u4e00\u6279\u5019\u9009\u8bcd\uff0c\u5c31\u8ba9\u672a\u7ecf\u8bad\u7ec3\u7684 Qwen3-4B-Base \u5728\u4e00\u9879\u5343\u9898\u8bca\u65ad\u96c6\u4e0a\u7684\u51c6\u786e\u7387\u4ece 15.8% \u8df3\u5230 30.0%\uff0c\u51e0\u4e4e\u7ffb\u500d\u3002\u8fd9\u8bc1\u660e”\u534a\u4e8b\u5b9e\u654f\u611f\u5ea6”\u4e0d\u662f\u7eb8\u4e0a\u6307\u6807\uff0c\u800c\u662f\u771f\u5b9e\u53ef\u7528\u7684\u4fe1\u53f7\u2014\u2014\u6a21\u578b\u786e\u5b9e\u5728\u4f9d\u8d56\u4e0d\u8be5\u4f9d\u8d56\u7684\u4e1c\u897f\uff0c\u628a\u8fd9\u4e9b\u4e1c\u897f\u6458\u6389\u53cd\u800c\u505a\u5f97\u66f4\u597d\u3002<\/p>\n
\u987a\u7740\u8fd9\u6761\u7ebf\u7d22\uff0c\u8bba\u6587\u63d0\u51fa SCAPO\uff08Semifactual Credit-Augmented Policy Optimization\uff0c\u534a\u4e8b\u5b9e\u4fe1\u7528\u589e\u5f3a\u7b56\u7565\u4f18\u5316\uff09\u3002\u5b83\u6cbf\u7528 GRPO \u7684\u6846\u67b6\uff0c\u4f46\u989d\u5916\u628a”\u534a\u4e8b\u5b9e\u7a33\u5b9a\u6027”\u5f15\u5165\u8bcd\u5143\u7ea7\u4fe1\u7528\u5206\u914d\uff1a\u7528\u6559\u5e08\u5f3a\u5236\u628a\u5df2\u91c7\u6837\u597d\u7684\u56de\u7b54\u5582\u7ed9\u5404\u6270\u52a8\u7248\u672c\u9898\u76ee\uff0c\u8bb0\u5f55\u6bcf\u4e2a\u8bcd\u7684\u6982\u7387\u6f02\u79fb\uff1b\u5728\u540c\u4e00\u7ec4\u56de\u7b54\u5185\u505a\u6807\u51c6\u5316\uff0c\u5f97\u5230\u6bcf\u4e2a\u8bcd\u7684\u76f8\u5bf9\u7a33\u5b9a\u6027\u5206\u6570\uff1b\u518d\u628a\u8fd9\u4e2a\u5206\u6570\u91cc\u504f\u8d1f\u7684\u90e8\u5206\uff08\u5373\u6bd4\u7ec4\u5185\u5e73\u5747\u66f4\u4e0d\u7a33\u5b9a\u7684\u8bcd\uff09\u53e0\u52a0\u5230 GRPO \u539f\u672c\u7684\u4f18\u52bf\u503c\u4e0a\u505a\u8d1f\u5411\u4fee\u6b63\uff0c\u800c\u5bf9\u7a33\u5b9a\u7684\u8bcd\u4e0d\u989d\u5916\u52a0\u5206\u3002\u8bba\u6587\u7279\u522b\u89e3\u91ca\uff0c\u7a33\u5b9a\u6027\u4e0d\u7b49\u4e8e\u6b63\u786e\u6027\uff0c\u4e00\u4e2a\u65e0\u8bba\u600e\u4e48\u6270\u52a8\u90fd\u51fa\u73b0\u7684\u8fde\u63a5\u8bcd\u672a\u5fc5\u6709\u7528\uff0c\u56e0\u6b64\u5956\u52b1\u673a\u5236\u5fc5\u987b\u662f”\u53ea\u7f5a\u4e0d\u5956”\u7684\u975e\u5bf9\u79f0\u8bbe\u8ba1\u3002\u8fd9\u4e2a\u4fee\u6b63\u53ea\u5728\u8bad\u7ec3\u65e9\u671f\u7a97\u53e3\u542f\u7528\uff084B \u6a21\u578b\u524d 120 \u6b65\u30011.7B \u6a21\u578b\u524d 200 \u6b65\uff09\uff0c\u4e4b\u540e\u5207\u56de\u6807\u51c6 GRPO\u3002<\/p>\n
\u5b9e\u9a8c\u5728 Qwen3-4B-Base \u4e0e Qwen3-1.7B-Base \u4e0a\u5b8c\u6210\uff0c\u8bad\u7ec3\u6570\u636e\u4e3a\u7ea6 1.7 \u4e07\u9053\u7684\u7ade\u8d5b\u7ea7\u6570\u5b66\u9898\u96c6 DAPO-Math-17K\u3002\u5728 AIME 2024 \u81f3 2026 \u7684\u7efc\u5408\u51c6\u786e\u7387\u4e0a\uff0c4B \u6a21\u578b\u4ece GRPO \u7684 22.08% \u63d0\u5347\u5230 27.71%\uff0c1.7B \u6a21\u578b\u4ece 7.71% \u63d0\u5347\u5230 11.88%\uff1b\u800c\u4e14 SCAPO \u4ec5\u7528\u4e0d\u5230 GRPO \u4e00\u534a\u7684\u8bad\u7ec3\u6b65\u6570\u5c31\u8ffd\u5e73\u4e86\u540e\u8005\u7684\u6700\u7ec8\u6210\u7ee9\u3002\u5728\u8986\u76d6\u516b\u9879\u6570\u5b66\u57fa\u51c6\u7684\u5e73\u5747\u5206\u4e0a\uff0cSCAPO \u540c\u6837\u5c45\u9996\uff084B \u4e3a 39.65%\u30011.7B \u4e3a 23.08%\uff0c\u800c GRPO \u5206\u522b\u4e3a 34.93% \u4e0e 17.86%\uff09\uff0c\u5e76\u5728\u5168\u90e8\u516b\u4e2a\u57fa\u51c6\u4e0a\u81f3\u5c11\u6709\u4e00\u9879\u767b\u9876\u3002<\/p>\n
\u66f4\u503c\u5f97\u505c\u7559\u7684\u662f\u5b83\u7684\u8de8\u9886\u57df\u6cdb\u5316\u3002\u7814\u7a76\u8005\u7528\u7814\u7a76\u751f\u6c34\u5e73\u7684\u79d1\u5b66\u95ee\u7b54\u57fa\u51c6 GPQA-Diamond \u6d4b\u8bd5\uff0cSCAPO \u5728 4B \u6a21\u578b\u4e0a\u62ff\u5230 42.45%\uff0c\u6bd4 GRPO \u7684 36.26% \u9ad8\u51fa 6.19 \u4e2a\u767e\u5206\u70b9\u3002\u7531\u4e8e\u8bad\u7ec3\u53ea\u6d89\u53ca\u6570\u5b66\uff0c\u8fd9\u79cd”\u5b66\u4f1a\u66f4\u7a33\u5065\u7684\u63a8\u7406\u4e60\u60ef\u800c\u975e\u67d0\u7c7b\u9898\u76ee\u7684\u5957\u8def”\u7684\u8fc1\u79fb\uff0c\u6697\u793a\u53bb\u9664\u865a\u5047\u4f9d\u8d56\u5e26\u6765\u7684\u6536\u76ca\u53ef\u80fd\u6bd4\u5237\u65b0\u5355\u4e00\u699c\u5355\u66f4\u672c\u8d28\u3002\u5728\u63d2\u5165\u65e0\u5173\u4fe1\u606f\u7684 NoOp \u7248\u672c\u4e0e\u4e13\u95e8\u6d4b\u8f93\u5165\u6270\u52a8\u7684 ThinkBench \u7248\u672c\u4e0a\uff0cSCAPO \u4e5f\u8868\u73b0\u6700\u5f3a\uff0c\u4e0e\u8bba\u6587\u6700\u521d\u53d1\u73b0\u7684\u95ee\u9898\u5f62\u6210\u95ed\u73af\u3002<\/p>\n
\u628a\u89c6\u89d2\u62c9\u5230\u8bad\u7ec3\u4fe1\u53f7\u8bbe\u8ba1\u7684\u5c42\u9762\u770b\uff0c\u8fd9\u7bc7\u8bba\u6587\u7684\u4ef7\u503c\u4e0d\u5728\u4e8e\u53c8\u4e00\u6b21\u5237\u5206\uff0c\u800c\u5728\u4e8e\u6362\u4e86\u4e00\u4e2a\u63d0\u95ee\u7ef4\u5ea6\uff1a\u8fc7\u53bb\u6539\u8fdb RLVR \u5927\u591a\u56f4\u7ed5”\u4f18\u52bf\u503c\u600e\u4e48\u7b97\u66f4\u7a33\u3001\u88c1\u526a\u600e\u4e48\u66f4\u5408\u7406”\uff0c\u800c SCAPO \u8ffd\u95ee\u7684\u662f”\u5956\u52b1\u4fe1\u53f7\u4f20\u64ad\u5230\u6bcf\u4e00\u4e2a\u8bcd\u65f6\uff0c\u662f\u4e0d\u662f\u4f20\u5f97\u592a\u7c97”\u3002\u5b83\u501f\u7528\u4e86\u56e0\u679c\u63a8\u65ad\u7684\u601d\u8def\u2014\u2014\u4e00\u4e2a\u53d8\u91cf\u82e5\u5728\u591a\u79cd\u5e72\u9884\u4e0b\u4fdd\u6301\u7a33\u5b9a\uff0c\u624d\u66f4\u53ef\u80fd\u662f\u771f\u6b63\u7684\u539f\u56e0\uff1b\u82e5\u5bf9\u65e0\u5173\u6270\u52a8\u5f02\u5e38\u654f\u611f\uff0c\u90a3\u5b83\u66f4\u53ef\u80fd\u53ea\u662f\u4e0e\u7ed3\u679c\u76f8\u5173\u3001\u800c\u975e\u56e0\u679c\u3002\u628a\u8fd9\u5957\u56e0\u679c\u601d\u7ef4\u5ac1\u63a5\u5230\u5927\u6a21\u578b\u5f3a\u5316\u5b66\u4e60\u91cc\uff0c\u672c\u8eab\u5177\u6709\u65b9\u6cd5\u8bba\u4e0a\u7684\u542f\u53d1\u6027\u3002<\/p>\n
\u5f53\u7136\uff0c\u8fd9\u79cd\u589e\u76ca\u76ee\u524d\u4ecd\u53d7\u9650\u4e8e\u9a8c\u8bc1\u8303\u56f4\u3002\u8bba\u6587\u53ea\u5728 1.7B \u4e0e 4B \u4e24\u4e2a\u76f8\u5bf9\u8f83\u5c0f\u7684\u6a21\u578b\u3001\u4e14\u53ea\u5728\u6570\u5b66\u9886\u57df\u505a\u4e86\u5b8c\u6574\u5b9e\u9a8c\uff1b\u66f4\u5927\u89c4\u6a21\u6a21\u578b\u3001\u6df7\u5408\u4e13\u5bb6\u67b6\u6784\uff0c\u4ee5\u53ca\u4ee3\u7801\u3001\u5e38\u8bc6\u63a8\u7406\u7b49\u9886\u57df\u80fd\u5426\u590d\u73b0\u540c\u6837\u7684\u6548\u679c\uff0c\u4f5c\u8005\u5e76\u672a\u7ed9\u51fa\u7b54\u6848\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u8fd9\u9879\u53d1\u73b0\u6307\u51fa\u4e86\u4e00\u4e2a\u6e05\u6670\u53ef\u6539\u8fdb\u7684\u65b9\u5411\uff0c\u4f46\u8ddd\u79bb\u8bc1\u660e\u5b83\u662f\u4e00\u6761\u666e\u9002\u7684\u8bad\u7ec3\u539f\u5219\uff0c\u8fd8\u5dee\u66f4\u591a\u8de8\u89c4\u6a21\u3001\u8de8\u4efb\u52a1\u7684\u8bc1\u636e\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4e00\u9053\u6570\u5b66\u9898\u6362\u4e2a\u95ee\u6cd5\u3001\u610f\u601d\u548c\u7b54\u6848\u90fd\u6ca1\u53d8\uff0c\u4e00<\/p>\n","protected":false},"author":9,"featured_media":20798,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"colormag_page_container_layout":"default_layout","colormag_page_sidebar_layout":"default_layout","footnotes":""},"categories":[1],"tags":[],"class_list":["post-20799","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"acf":[],"_links":{"self":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/20799","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/users\/9"}],"replies":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/comments?post=20799"}],"version-history":[{"count":1,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/20799\/revisions"}],"predecessor-version":[{"id":20872,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/20799\/revisions\/20872"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/media\/20798"}],"wp:attachment":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/media?parent=20799"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/categories?post=20799"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/tags?post=20799"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}