\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u4e4b\u6240\u4ee5\u6602\u8d35\uff0c\u74f6\u9888\u4e0d\u5728\u7b97\u529b\uff0c\u800c\u5728\u6807\u6ce8\uff1a\u8981\u8ba9\u5b83\u6307\u51fa\u89e3\u7b54\u5728\u54ea\u4e00\u6b65\u5f00\u59cb\u51fa\u9519\uff0c\u5c31\u5f97\u6709\u4eba\u9010\u6b65\u6807\u6ce8\u6bcf\u6761\u63a8\u7406\u8f68\u8ff9\u7684\u6bcf\u4e00\u6b65\u5bf9\u9519\u3002\u8fd9\u4e00\u72b6\u51b5\u6b63\u5728\u88ab\u4e00\u79cd\u53ea\u7528\u7ed3\u679c\u6807\u7b7e\u7684\u8bad\u7ec3\u65b9\u6cd5\u6539\u53d8\u3002\u4e2d\u56fd\u4eba\u6c11\u5927\u5b66\u3001\u6e05\u534e\u5927\u5b66\u3001\u7535\u5b50\u79d1\u6280\u5927\u5b66\u4e0e\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66\u56db\u6821\u56e2\u961f\u5728\u9884\u5370\u672c\u5e73\u53f0 arXiv \u516c\u5f00\u4e86\u4e00\u7bc7\u7814\u7a76\u5982\u4f55\u4ec5\u4ece\u4f9d\u8d56\u6700\u7ec8\u7ed3\u679c\u7684\u5f3a\u5316\u5b66\u4e60\u4e2d\u8bf1\u5bfc\u51fa\u8fc7\u7a0b\u76d1\u7763\u7684\u8bba\u6587\uff08\u7f16\u53f7 arXiv:2609.36641\uff09\uff0c\u63d0\u51faTIPS\uff08Thinking-Induced Process Supervision\uff0c\u601d\u8003\u8bf1\u5bfc\u8fc7\u7a0b\u76d1\u7763\uff09\u6846\u67b6\uff1a\u8bad\u7ec3\u65f6\u53ea\u4f9d\u636e”\u6700\u7ec8\u7b54\u6848\u5bf9\u4e0d\u5bf9”\u8fd9\u4e00\u4e2a\u4fe1\u53f7\uff0c\u6a21\u578b\u5374\u80fd\u5b66\u4f1a\u9010\u6b65\u5b9a\u4f4d\u63a8\u7406\u9519\u8bef\u3002<\/p>\n
\u65b9\u6cd5\u672c\u8eab\u5e76\u4e0d\u590d\u6742\uff0c\u4f46\u8bbe\u8ba1\u524d\u63d0\u503c\u5f97\u62c6\u5f00\u770b\u3002\u7ed9\u5b9a\u4e00\u9053\u9898\u548c\u4e00\u6761\u5f85\u8bc4\u5224\u7684\u89e3\u7b54\uff0c\u6a21\u578b\u5148\u81ea\u7531\u751f\u6210\u4e00\u6bb5\u601d\u7ef4\u94fe\uff0c\u518d\u8f93\u51fa\u4e24\u7c7b\u5224\u65ad\u2014\u2014\u6bcf\u4e00\u6b65\u662f\u5426\u5728\u7ed9\u5b9a\u524d\u5e8f\u6b65\u9aa4\u4e0b\u6709\u6548\uff08\u4e8c\u503c\u6807\u7b7e\uff09\uff0c\u4ee5\u53ca\u6574\u6761\u8f68\u8ff9\u6700\u7ec8\u5bf9\u4e0d\u5bf9\uff08\u4e8c\u503c\u6807\u7b7e\uff09\u3002\u5956\u52b1\u51fd\u6570\u53ea\u8bfb\u53d6\u540e\u8005\uff1a\u9884\u6d4b\u7ed3\u679c\u4e0e\u771f\u5b9e\u7b54\u6848\u4e00\u81f4\u8bb01\u5206\uff0c\u5426\u5219\u8bb00\u5206\uff0c\u8fc7\u7a0b\u6807\u7b7e\u5b8c\u5168\u4e0d\u8fdb\u5165\u5956\u52b1\u8ba1\u7b97\u3002\u8bad\u7ec3\u7528GRPO\uff08\u7fa4\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff09\u8fd9\u4e00\u65e0\u9700critic\u7684PPO\u53d8\u4f53\uff0c\u540c\u4e00\u9898\u76ee\u91c7\u6837\u4e00\u7ec4\u6eda\u52a8\u8f93\u51fa\uff0c\u6309\u7ec4\u5185\u5956\u52b1\u5747\u503c\u7b97\u51fa\u76f8\u5bf9\u4f18\u52bf\uff0c\u518d\u628a\u8fd9\u4e2a\u4f18\u52bf\u65bd\u52a0\u5230\u751f\u6210\u7684\u5168\u90e8token\u4e0a\uff0c\u5305\u62ec\u601d\u7ef4\u94fe\u548c\u8fc7\u7a0b\u6807\u7b7e\u672c\u8eab\u3002\u7531\u6b64\u5f62\u6210\u4e00\u6761\u95f4\u63a5\u94fe\u8def\uff1a\u51c6\u786e\u5224\u65ad\u4e2d\u95f4\u6b65\u9aa4\u6709\u52a9\u4e8e\u5224\u5bf9\u6700\u7ec8\u7ed3\u679c\uff0c\u5224\u5bf9\u7ed3\u679c\u62ff\u5230\u66f4\u9ad8\u4f18\u52bf\uff0c\u4f18\u52bf\u53c8\u53cd\u8fc7\u6765\u5f3a\u5316\u90a3\u6bb5”\u6311\u9519”\u884c\u4e3a\u3002<\/p>\n
\u8bba\u6587\u7528\u4fe1\u606f\u8bba\u7ed9\u51fa\u4e86\u8fd9\u6761\u94fe\u8def\u6210\u7acb\u7684\u6761\u4ef6\u3002\u8bbeR\u4e3a\u6a21\u578b\u63a8\u7406\u524d\u7684\u8868\u793a\u3001Z\u4e3a\u751f\u6210\u7684\u601d\u7ef4\u94fe\u3001D\u4e3a\u771f\u5b9e\u7ed3\u679c\u6807\u7b7e\uff0c\u5b9a\u7406\u4e00\u7ed9\u51fa\u7ed3\u8bba\u4e92\u4fe1\u606f\u4e0b\u754c I(Z;D|R) \u2265 H(D|R) \u2212 H_b(P_e)\uff0c\u5176\u4e2dH_b\u4e3a\u4e8c\u5143\u71b5\uff1b\u5b9a\u7406\u4e8c\u8fdb\u4e00\u6b65\u8bf4\u660e I(Z;Y|R) \u2265 I(Z;D|R) \u2212 H(D|Y,R)\u3002\u5408\u5e76\u540e\u5f97\u5230\u7684\u4e0b\u754c\u7531\u4e09\u9879\u6784\u6210\uff1a\u9a8c\u8bc1\u96be\u5ea6H(D|R)\u3001\u7ed3\u679c\u9519\u8bef\u7387H_b(P_e)\u3001\u4ee5\u53ca\u6b65\u9aa4\u4e0e\u7ed3\u679c\u7684\u8026\u5408\u6b8b\u4f59H(D|Y,R)\u3002\u8fd9\u610f\u5473\u7740\u65b9\u6cd5\u751f\u6548\u9700\u8981\u4e09\u4e2a\u6761\u4ef6\u540c\u65f6\u6210\u7acb\uff1a\u7ed3\u679c\u65e0\u6cd5\u4ece\u8f93\u5165\u76f4\u63a5\u63a8\u51fa\u3001\u7ed3\u679c\u9884\u6d4b\u51c6\u786e\u7387\u8db3\u591f\u9ad8\u3001\u4ee5\u53ca\u6700\u7ec8\u7ed3\u679c\u4e0e\u6b65\u9aa4\u5bf9\u9519\u9ad8\u5ea6\u8026\u5408\u2014\u2014\u6570\u5b66\u9898\u6070\u597d\u6ee1\u8db3\u6700\u540e\u4e00\u6761\uff0c\u4e00\u65e6\u4e2d\u95f4\u6b65\u9aa4\u7684\u6709\u6548\u6027\u5df2\u77e5\uff0c\u7b54\u6848\u5bf9\u9519\u57fa\u672c\u786e\u5b9a\u3002\u8bba\u6587\u540c\u65f6\u5217\u51fa\u65b9\u6cd5\u5931\u6548\u7684\u51e0\u79cd\u60c5\u5f62\uff1a\u8f93\u5165\u5c42\u9762\u5b58\u5728\u6377\u5f84\u3001\u7ed3\u679c\u9884\u6d4b\u51c6\u786e\u7387\u4e0a\u4e0d\u53bb\u3001\u6216\u4efb\u52a1\u672c\u8eab\u7ed3\u679c\u4e0e\u8fc7\u7a0b\u5173\u8054\u677e\u6563\u3002<\/p>\n
\u6570\u636e\u89c4\u6a21\u662f\u8fd9\u9879\u5de5\u4f5c\u6700\u7a81\u51fa\u7684\u90e8\u5206\u3002\u6570\u5b66\u4efb\u52a1\u53ea\u4ece\u516c\u5f00\u7684SCAN-Pro\u6570\u636e\u96c6\u91c7\u68373200\u6761\u8f68\u8ff9\uff0c\u667a\u80fd\u4f53\u4efb\u52a1\u53e6\u6784\u5efa2905\u6761\u6765\u81eaHotPotQA\u3001GAIA\u3001\u03c4-Bench\u4e0eBFCL\u7684\u8bed\u6599\uff0c\u8bad\u7ec3\u57288\u575780GB\u663e\u5b58\u7684A800\u4e0a\u5b8c\u6210\u3002<\/p>\n
\u5728ProcessBench\u8fd9\u4e00”\u627e\u51fa\u89e3\u7b54\u4e2d\u7b2c\u4e00\u4e2a\u9519\u8bef\u6b65\u9aa4”\u7684\u57fa\u51c6\u4e0a\uff0cTIPS-Qwen3-4B-Thinking-2507\u62ff\u523085.2\u7684F1\u5206\u6570\uff0c\u8d85\u8fc7\u7528\u8499\u7279\u5361\u6d1b\u4f30\u8ba1\u52a0\u84b8\u998f\u8bad\u7ec3\u7684Qwen2.5-Math-PRM-72B\uff0878.3\uff09\u3001GenPRM-32B\uff0877.6\uff09\uff0c\u4e5f\u8d85\u8fc7\u76f4\u63a5\u4ee5\u63d0\u793a\u65b9\u5f0f\u4f5c\u7b54\u7684GPT-5.4-Instruct\uff0879.4\uff09\u548cClaude-4.7-Opus\uff0879.3\uff09\uff0c\u4f46\u4ecd\u4f4e\u4e8eo1-mini\u768487.9\u3002\u540c\u4e00\u57fa\u5ea7\u7684\u6307\u4ee4\u578b\u7248\u672cTIPS-Qwen3-4B-Instruct-2507\u4e3a77.7\uff0c\u800c\u672a\u7ecf\u8bad\u7ec3\u7684Qwen3-4B-Thinking-2507\u53ea\u670960.3\u3001\u6307\u4ee4\u578b\u4e3a63.8\u2014\u2014\u540c\u4e00\u6a21\u578b\u5bb6\u65cf\u5728\u8bad\u7ec3\u524d\u540e\u51fa\u73b0\u8fd125\u5206\u7684\u843d\u5dee\u3002\u5bf9\u7167\u5b9e\u9a8c\u540c\u6837\u8bf4\u660e\u95ee\u9898\u4e0d\u5728\u6570\u636e\u91cf\uff1a\u4e24\u79cd\u540c\u6837\u53ea\u7528\u7ed3\u679c\u6807\u7b7e\u7684\u65e2\u6709\u65b9\u6cd5AutoPSV\u4e0eImplicitPRM\uff0c\u5728197K\u6570\u636e\u4e0b\u5206\u522b\u53ea\u670945.3\u548c41.5\uff0c\u75283.2K\u6570\u636e\u7684TIPS\u53cd\u800c\u9ad8\u51fa\u4e00\u5927\u622a\u3002<\/p>\n
\u51e0\u7ec4\u6d88\u878d\u5b9e\u9a8c\u628a\u65b9\u6cd5\u7684\u9002\u7528\u8fb9\u754c\u5212\u5f97\u66f4\u6e05\u695a\u3002\u53bb\u6389\u601d\u7ef4\u94fe\u3001\u6539\u4e3a\u76f4\u63a5\u8f93\u51fa\u6807\u7b7e\u7684\u7248\u672c\uff0cProcessBench\u57fa\u7840\u5206\u6570\u53ea\u670947.5\uff0c\u8fc7\u7a0b\u5224\u65ad\u80fd\u529b\u51e0\u4e4e\u96f6\u63d0\u5347\uff080.7\u5206\uff09\uff0c\u800c\u4fdd\u7559\u601d\u7ef4\u94fe\u7684\u7248\u672c\u63d0\u534713.9\u5206\u2014\u2014\u8bf4\u660e\u627f\u8f7d\u4fe1\u53f7\u7684\u662f\u90a3\u6bb5\u53ef\u88ab\u5f3a\u5316\u7684\u63a8\u7406\u8fc7\u7a0b\uff0c\u800c\u4e0d\u662f\u6807\u7b7e\u672c\u8eab\u3002\u4ee5Qwen2.5-7B-Instruct\u4e3a\u9aa8\u5e72\u65f6\u51fa\u73b0\u4e86\u503c\u5f97\u8b66\u60d5\u7684\u73b0\u8c61\uff1a\u7ed3\u679c\u5224\u65ad\u51c6\u786e\u7387\u5feb\u901f\u63d0\u5347\uff0c\u4f46\u8fc7\u7a0b\u5224\u65adF1\u5148\u5347\u540e\u964d\uff0c\u8bad\u7ec3\u540e\u671f\u56de\u843d\uff1b\u76f8\u6bd4\u4e4b\u4e0b\u7ecfDeepSeek-R1\u84b8\u998f\u7684\u540c\u57fa\u5ea7\u6a21\u578b\u5219\u7a33\u6b65\u4e0a\u5347\u3002\u8bba\u6587\u7531\u6b64\u63d0\u51fa\u4e00\u6761\u65e0\u9700\u8fc7\u7a0b\u6807\u6ce8\u7684\u65e9\u505c\u542f\u53d1\u5f0f\uff1a\u5f53\u7ed3\u679c\u5224\u65ad\u51c6\u786e\u7387\u8d8b\u4e8e\u5e73\u53f0\u3001\u540c\u65f6\u751f\u6210\u957f\u5ea6\u5f00\u59cb\u4e0b\u964d\u65f6\u505c\u6b62\u8bad\u7ec3\u3002<\/p>\n
\u53e6\u4e00\u7ec4\u5b9e\u9a8c\u6392\u9664\u4e86”\u53ea\u662f\u628a\u89e3\u9898\u80fd\u529b\u7ec3\u5f3a\u4e86”\u7684\u89e3\u91ca\uff1a\u53ea\u7ec3\u89e3\u9898\u7684\u5bf9\u7167\u7248\u672c\u5728AIME 2026\u548cHMMT 2026\u4e0a\u7684\u89e3\u9898\u51c6\u786e\u7387\u66f4\u9ad8\uff0855.8%\u5bf953.3%\uff09\uff0c\u4f46ProcessBench\u4e0a\u7684\u627e\u9519\u80fd\u529b\u53ea\u670962.8\u5206\uff0c\u6bd4TIPS\u768477.7\u4f4e\u8fd115\u5206\uff1b\u53c2\u6570\u91cf\u5927\u5f97\u591a\u7684Qwen3-30B-A3B-Thinking-2507\u89e3\u9898\u5206\u6570\u66f4\u9ad8\uff0cProcessBench\u5374\u6bd44B\u7684TIPS\u6a21\u578b\u4f4e\u7ea65\u5206\u3002<\/p>\n
\u667a\u80fd\u4f53\u4efb\u52a1\u7684\u6536\u76ca\u660e\u663e\u66f4\u5c0f\u3002AgentProcessBench\u4e0a\uff0cQwen3-4B-Thinking-2507\u7684\u9996\u9519\u5b9a\u4f4d\u51c6\u786e\u7387\u753143.7%\u5347\u81f349.1%\uff0c\u53e6\u4e24\u4e2a\u9aa8\u5e72\u5206\u522b\u63d0\u53472.5\u548c4.5\u4e2a\u767e\u5206\u70b9\uff0c\u5176\u4e2dQwen3-8B-Thinking\u5728HotPotQA\u4e00\u9879\u4e0a\u8fd8\u4e0b\u964d\u4e863.2\u4e2a\u70b9\u3002\u8bba\u6587\u628a\u539f\u56e0\u5f52\u5230\u7ed3\u679c\u4e0e\u8fc7\u7a0b\u7684\u8026\u5408\u66f4\u677e\uff1a\u667a\u80fd\u4f53\u5931\u8d25\u53ef\u80fd\u6765\u81ea\u6f0f\u505a\u52a8\u4f5c\u3001\u8bc1\u636e\u4e0d\u8db3\u6216\u63a2\u7d22\u65e0\u6548\uff0c\u672a\u5fc5\u5bf9\u5e94\u67d0\u4e2a\u660e\u786e\u7684\u9519\u8bef\u6b65\u9aa4\u3002\u4eba\u5de5\u590d\u6838\u7684\u90e8\u5206\u540c\u6837\u7559\u6709\u8fb9\u754c\uff1a\u4e24\u4f4d\u535a\u58eb\u4eceProcessBench\u968f\u673a\u62bd\u53d660\u6761\u8f68\u8ff9\u68c0\u67e5\uff0c\u6a21\u578b\u6b63\u786e\u5b9a\u4f4d\u9996\u4e2a\u9519\u8bef\u536078.3%\uff0c\u5176\u4e2d\u89e3\u91ca\u4e5f\u51c6\u786e\u7684\u6bd4\u4f8b\u4e3a80.9%\uff0c\u4e24\u8005\u517c\u5177\u7684\u536063.3%\u3002<\/p>\n
\u8fd9\u9879\u5de5\u4f5c\u628a\u8fc7\u7a0b\u76d1\u7763\u7684\u6807\u6ce8\u74f6\u9888\u63a8\u5230\u4e86\u4e00\u4e2a\u65b0\u7684\u4f4d\u7f6e\uff1a\u8fc7\u7a0b\u5224\u65ad\u53ef\u4ee5\u88ab”\u8bf1\u5bfc”\u51fa\u6765\uff0c\u4f46\u524d\u63d0\u662f\u7ed3\u679c\u4e0e\u8fc7\u7a0b\u4e4b\u95f4\u786e\u5b9e\u5b58\u5728\u6e05\u6670\u56e0\u679c\u3002\u8bba\u6587\u81ea\u8eab\u5728\u5c40\u9650\u6027\u4e0e\u66f4\u5e7f\u5f71\u54cd\u4e2d\u5199\u660e\uff0c\u65b9\u6cd5\u53ea\u8986\u76d6\u7eaf\u6587\u672c\u63a8\u7406\u4e0e\u667a\u80fd\u4f53\u8f68\u8ff9\uff0c\u672a\u8bc4\u4f30\u591a\u6a21\u6001\u573a\u666f\uff1b\u7531\u4e8e\u8fc7\u7a0b\u6807\u7b7e\u662f\u7531\u7ed3\u679c\u5956\u52b1\u95f4\u63a5\u8bf1\u5bfc\u4ea7\u751f\u7684\uff0c\u5728\u7ed3\u679c\u4e0e\u5c40\u90e8\u6b65\u9aa4\u6b63\u786e\u6027\u5173\u8054\u8f83\u5f31\u7684\u9886\u57df\u53ef\u80fd\u5931\u51c6\u6216\u4e0d\u53ef\u9760\uff0c\u4e0d\u5e94\u5728\u9ad8\u98ce\u9669\u573a\u666f\u4e2d\u88ab\u5f53\u4f5c\u771f\u503c\u4f7f\u7528\u3002\u540c\u4e00\u8109\u7edc\u4e0a\u7684\u53e6\u4e00\u6761\u8def\u5f84\u5219\u9009\u62e9\u7528\u4fe1\u606f\u8bba\u76f4\u63a5\u4f30\u8ba1\u6bcf\u4e00\u6b65\u5bf9\u6b63\u786e\u7b54\u6848\u4f3c\u7136\u7684\u5f71\u54cd\uff0c\u628a\u8ba1\u7b97\u590d\u6742\u5ea6\u4eceO(N log N)\u964d\u5230O(N)\uff0c\u5728\u6570\u5b66\u3001Python\u3001SQL\u4e0e\u79d1\u5b66\u95ee\u7b54\u7b49\u591a\u7c7b\u57fa\u51c6\u4e0a\u9a8c\u8bc1\u4e86\u94fe\u5f0f\u601d\u7ef4\u9009\u62e9\u7684\u53ef\u884c\u6027\u3002\u4e24\u6761\u8def\u7ebf\u6307\u5411\u540c\u4e00\u5904\u672a\u51b3\u95ee\u9898\uff1a\u5f53”\u7ed3\u679c”\u672c\u8eab\u4e0d\u8db3\u4ee5\u7ea6\u675f”\u8fc7\u7a0b”\u65f6\uff0c\u8fc7\u7a0b\u76d1\u7763\u7684\u4fe1\u53f7\u7a76\u7adf\u8fd8\u80fd\u4ece\u54ea\u91cc\u6765\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u4e4b\u6240\u4ee5\u6602\u8d35\uff0c\u74f6\u9888\u4e0d\u5728\u7b97\u529b\uff0c\u800c<\/p>\n","protected":false},"author":10,"featured_media":22416,"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-22417","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\/22417","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\/10"}],"replies":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/comments?post=22417"}],"version-history":[{"count":2,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/22417\/revisions"}],"predecessor-version":[{"id":22512,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/22417\/revisions\/22512"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/media\/22416"}],"wp:attachment":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/media?parent=22417"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/categories?post=22417"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/tags?post=22417"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}