{"id":728376,"date":"2026-02-23T14:43:47","date_gmt":"2026-02-23T06:43:47","guid":{"rendered":"https:\/\/www.alphaseek.cn\/?p=728376"},"modified":"2026-02-23T14:43:47","modified_gmt":"2026-02-23T06:43:47","slug":"%e9%87%8d%e7%a3%85%ef%bc%81%e6%99%ba%e8%b0%b1%e6%8a%8aglm-5%e7%9a%84%e6%89%80%e6%9c%89%e6%8a%80%e6%9c%af%e7%bb%86%e8%8a%82%e9%83%bd%e5%85%ac%e5%bc%80%e4%ba%86","status":"publish","type":"post","link":"https:\/\/www.alphaseek.cn\/?p=728376","title":{"rendered":"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86"},"content":{"rendered":"<p>\u6211\u4eec\u63a8\u51fa\u4e86 GLM-5\uff0c\u8fd9\u662f\u4e00\u6b3e\u65e8\u5728\u63a8\u52a8\u7f16\u7a0b\u8303\u5f0f\u4ece\u201cVibe Coding\u201d\uff08\u6c1b\u56f4\u7f16\u7a0b\uff09\u8f6c\u5411\u201cAgentic Engineering\u201d\uff08\u667a\u80fd\u4f53\u5de5\u7a0b\uff09\u7684\u4e0b\u4e00\u4ee3\u57fa\u7840\u6a21\u578b\u3002GLM-5 \u5728\u524d\u4ee3\u6a21\u578b GLM-4.5 \u7684\u667a\u80fd\u4f53\u3001\u63a8\u7406\u4e0e\u7f16\u7a0b\uff08Agentic, Reasoning and Coding, ARC\uff09\u80fd\u529b\u57fa\u7840\u4e0a\uff0c\u91c7\u7528\u7a00\u758f\u6ce8\u610f\u529b\uff08DeepSeek Sparse Attention\uff0cDSA\uff09\u4ee5\u5927\u5e45\u964d\u4f4e\u63a8\u7406\u6210\u672c\uff0c\u540c\u65f6\u4fdd\u6301\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u65e0\u635f\u3002\u4e3a\u4e86\u8ba9\u6a21\u578b\u66f4\u597d\u5730\u4e0e\u5404\u7c7b\u4efb\u52a1\u5bf9\u9f50\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u5957\u65b0\u578b\u5f02\u6b65\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u57fa\u7840\u8bbe\u65bd\uff0c\u901a\u8fc7\u5c06\u751f\u6210\u8fc7\u7a0b\u4e0e\u8bad\u7ec3\u8fc7\u7a0b\u89e3\u8026\uff0c\u4ece\u800c\u5927\u5e45\u63d0\u5347\u4e86\u540e\u8bad\u7ec3\u7684\u8fed\u4ee3\u6548\u7387\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u63d0\u51fa\u4e86\u5168\u65b0\u7684\u5f02\u6b65 Agent \u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u5f3a\u5316\u5b66\u4e60\u7684\u6548\u679c\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u66f4\u6709\u6548\u5730\u4ece\u590d\u6742\u3001\u957f\u7a0b\u4ea4\u4e92\u4e2d\u5b66\u4e60\u3002\u57fa\u4e8e\u4e0a\u8ff0<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e5%88%9b%e6%96%b0\" title=\"View all posts in \u521b\u65b0\" target=\"_blank\" class=\"tag_link\">\u521b\u65b0<\/a>\uff0cGLM-5 \u5728\u4e3b\u6d41\u7684\u5f00\u653e\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5b9e\u73b0\u4e86 SOTA \u6027\u80fd\u3002\u6700\u5173\u952e\u7684\u662f\uff0cGLM-5 \u5728\u771f\u5b9e\u4e16\u754c\u7f16\u7a0b\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u524d\u6240\u672a\u6709\u7684\u80fd\u529b\uff0c\u5728\u5904\u7406\u7aef\u5230\u7aef\u8f6f\u4ef6\u5de5\u7a0b\u6311\u6218\u65b9\u9762\u8d85\u8d8a\u4e86\u6b64\u524d\u6240\u6709\u5f00\u6e90\u57fa\u7ebf\u3002\u4ee3\u7801\u3001\u6a21\u578b\u53ca\u66f4\u591a\u4fe1\u606f\u8bf7\u8bbf\u95ee\uff1ahttps:\/\/z.ai\/blog\/glm-5<\/p>\n<p>1 \u5f15\u8a00<\/p>\n<p>\u8ffd\u6c42 AGI \u5e76\u975e\u5355\u7eaf\u6269\u5f20\u53c2\u6570\u89c4\u6a21\uff0c\u800c\u662f\u8981\u4ece\u5e95\u5c42\u91cd\u6784\u667a\u80fd\u6548\u7387\u4e0e\u81ea\u4e3b\u6f14\u8fdb\u67b6\u6784\u3002\u5728 GLM-4.5 \u4e2d\uff0c\u6211\u4eec\u6210\u529f\u9a8c\u8bc1\u4e86\u5c06\u667a\u80fd\u4f53\u3001\u63a8\u7406\u4e0e\u7f16\u7a0b\uff08ARC\uff09\u80fd\u529b\u878d\u5408\u81f3\u5355\u4e00 MoE \u67b6\u6784\u7684\u53ef\u884c\u6027\uff0c\u5e76\u5728\u591a\u9879\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u53d6\u5f97\u4e0d\u9519\u7684\u6210\u7ee9\u3002\u4e0d\u8fc7\uff0c\u968f\u7740 LLM \u4ece\u9759\u6001\u77e5\u8bc6\u5e93\u5411\u4e3b\u52a8\u95ee\u9898\u6c42\u89e3\u5668\u8f6c\u53d8\uff0c\u8ba1\u7b97\u6210\u672c\u4e0e\u73b0\u5b9e\u9002\u5e94\u6027\u6210\u4e3a\u6838\u5fc3\u74f6\u9888\uff0c\u8fd9\u5728\u590d\u6742\u7684\u8f6f\u4ef6\u5de5\u7a0b\u573a\u666f\u4e2d\u8868\u73b0\u5f97\u5c24\u4e3a\u660e\u663e\u3002<\/p>\n<p>\u4e3a\u6b64\uff0c\u6211\u4eec\u63a8\u51fa\u4e86\u4e0b\u4e00\u4ee3\u65d7\u8230\u6a21\u578b GLM-5 \u4ee5\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\u3002GLM-5 \u5728\u6027\u80fd\u4e0e\u8ba1\u7b97\u6548\u7387\u4e0a\u5b9e\u73b0\u4e86\u8dc3\u5347\uff0c\u4e0d\u4ec5\u5728 ArtificialAnalysis.ai\u3001LMArena \u6587\u672c\u4e0e\u4ee3\u7801\u7b49\u4e3b\u8981\u699c\u5355\u4e2d\u5747\u8fbe\u5230 SOTA \u6c34\u5e73\uff0c\u66f4\u91cd\u5851\u4e86\u771f\u5b9e\u4e16\u754c\u7684\u7f16\u7a0b\u6807\u51c6\u3002\u5b83\u7a81\u7834\u4e86 SWE-bench \u7b49\u4f20\u7edf\u9759\u6001\u6d4b\u8bc4\u7684\u8003\u5bdf\u8fb9\u754c\uff0c\u5728\u5904\u7406\u590d\u6742\u7684\u7aef\u5230\u7aef\u8f6f\u4ef6\u5f00\u53d1\u4efb\u52a1\u65f6\uff0c\u5c55\u73b0\u51fa\u4e86\u524d\u6240\u672a\u6709\u7684\u5f3a\u5927\u80fd\u529b\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u679c\u3002\u56fe 1 \u5c55\u793a\u4e86 GLM-5\u3001GLM-4.7\u3001Claude Opus 4.5\u3001Gemini 3 Pro \u548c GPT-5.2\uff08xhigh\uff09\u5728 8 \u4e2aAgent\u3001\u63a8\u7406\u4e0e\u7f16\u7a0b\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u7ed3\u679c\u3002\u5e73\u5747\u800c\u8a00\uff0cGLM-5 \u76f8\u6bd4\u4e0a\u4e00\u7248\u672c GLM-4.7 \u63d0\u5347\u7ea6 20%\uff0c\u4e0e Claude Opus 4.5 \u548c GPT-5.2\uff08xhigh\uff09\u76f8\u5f53\uff0c\u5e76\u4f18\u4e8e Gemini 3 Pro\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/3BCFFC05D22B821EC66CFA05C1D876820EB7263B_size325_w1080_h632.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe1\uff1aGLM-5\u3001DeepSeek-V3.2\u3001Claude Opus 4.5\u3001Gemini 3 Pro \u548c GPT-5.2\uff08xhigh\uff09\u5728 Humanity's Last Exam\u3001SWE-bench Verified\u3001SWE-bench Multilingual\u3001Terminal-Bench 2.0\u3001BrowseComp\u3001MCP-Atlas\u3001\u03c4&sup2;-Bench \u548c Vending Bench 2 \u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u7ed3\u679c\u3002<\/p>\n<p>GLM-5 \u5728 Artificial Analysis Intelligence Index v4.0 \u4e2d\u5f97\u5230 50 \u5206\uff0c\u6210\u4e3a\u65b0\u7684\u5f00\u6e90 SOTA \u6a21\u578b\uff08\u89c1\u56fe2\uff09\uff0c\u8f83 GLM-4.7 \u63d0\u5347 8 \u5206\uff0c\u8fd9\u4e00\u63d0\u5347\u6e90\u4e8e Agent \u6027\u80fd\u53ca\u77e5\u8bc6\/\u5e7b\u89c9\u65b9\u9762\u7684\u5168\u9762\u6539\u8fdb\u3002\u8fd9\u662f\u5f00\u653e\u6743\u91cd\u6a21\u578b\u9996\u6b21\u5728 Artificial Analysis Intelligence Index v4.0 \u4e2d\u8fbe\u5230 50 \u5206\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/CF40727922F19A1B538EABFB78ACC69FB037BB2C_size220_w1080_h315.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe2\uff1aArtificial Analysis Intelligence Index v4.0 \u6db5\u76d6 10 \u9879\u8bc4\u4f30\uff1aGDPval-AA\u3001\u03c4&sup2;-Bench Telecom\u3001Terminal-Bench Hard\u3001SciCode\u3001AA-LCR\u3001AA-Omniscience\u3001IFBench\u3001Humanity's Last Exam\u3001GPQA Diamond \u548c CritPt\u3002<\/p>\n<p>LMArena \u6700\u521d\u7531\u52a0\u5dde\u5927\u5b66\u4f2f\u514b\u5229\u5206\u6821\u53d1\u8d77\uff0c\u662f\u4e00\u4e2a\u901a\u8fc7\u6570\u767e\u4e07\u771f\u5b9e\u4efb\u52a1\uff08\u5305\u62ec\u5199\u4f5c\u3001\u7f16\u7a0b\u3001\u63a8\u7406\u3001\u8bbe\u8ba1\u3001\u641c\u7d22\u548c\u521b\u4f5c\uff09\u4ee5\u4eba\u7c7b\u8bc4\u5224\u65b9\u5f0f\u900f\u660e\u8bc4\u4f30\u548c\u6bd4\u8f83\u524d\u6cbf AI \u80fd\u529b\u7684\u5e73\u53f0\u3002\u5927\u91cf\u7684\u4eba\u7c7b\u4ea4\u4e92\u4ea7\u751f\u4e86\u771f\u5b9e\u4e16\u754c\u6548\u7528\u4fe1\u53f7\uff0c\u4f7f\u5176\u6709\u522b\u4e8e\u5176\u4ed6\u9759\u6001\u57fa\u51c6\u6d4b\u8bd5\u3002\u5982\u56fe 3 \u663e\u793a\uff0cGLM-5 \u5728\u6587\u672c\u7ade\u6280\u573a\u548c\u4ee3\u7801\u7ade\u6280\u573a\u4e2d\u5747\u4f4d\u5217\u5f00\u653e\u6a21\u578b\u7b2c\u4e00\uff0c\u6574\u4f53\u6c34\u5e73\u4e0e Claude Opus 4.5 \u548c Gemini 3 Pro \u76f8\u5f53\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/063F178DF6F2AB61DC292836EA8D1C54DB09CFEC_size282_w1080_h452.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe3\uff1a\u5728 LMArena \u4e0a\uff0cGLM-5 \u5728\u6587\u672c\u7ade\u6280\u573a\uff08Text Arena\uff09\u548c\u4ee3\u7801\u7ade\u6280\u573a\uff08Code Arena\uff09\u4e2d\u5747\u4f4d\u5c45\u5f00\u6e90\u6a21\u578b\u7b2c\u4e00\u3002<\/p>\n<p>\u667a\u80fd\u4f53\uff08Agent\uff09\u5728\u957f\u5468\u671f\u4efb\u52a1\u4e2d\u7684\u8fde\u8d2f\u6027\u6b63\u53d8\u5f97\u6108\u53d1\u5173\u952e\u3002\u5982\u4eca\uff0c\u7f16\u7a0b\u667a\u80fd\u4f53\uff08Coding Agent\uff09\u5df2\u80fd\u8fde\u7eed\u6570\u5c0f\u65f6\u81ea\u4e3b\u7f16\u5199\u4ee3\u7801\uff0cAI \u6a21\u578b\u7684\u4efb\u52a1\u80fd\u529b\u8fb9\u754c\u4e5f\u5728\u4e0d\u65ad\u62d3\u5c55\u3002\u4e3a\u4e86\u8bc4\u4f30 GLM-5 \u5904\u7406\u957f\u7a0b\u4efb\u52a1\u7684\u80fd\u529b\uff0c\u6211\u4eec\u5f15\u5165\u4e86 Vending-Bench 2 \u548c CC-Bench-V2 \u4e24\u5927\u57fa\u51c6\u6d4b\u8bd5\u3002<\/p>\n<p>Vending-Bench 2 \u65e8\u5728\u8bc4\u4f30 AI \u6a21\u578b\u5728\u957f\u5468\u671f\u5185\u7684\u5546\u4e1a\u8fd0\u8425\u8868\u73b0\u3002\u8be5\u6d4b\u8bd5\u8981\u6c42\u6a21\u578b\u5728\u8bbe\u5b9a\u7684\u4e00\u5e74\u671f\u9650\u5185\u6a21\u62df\u7ecf\u8425\u4e00\u5bb6\u81ea\u52a8\u552e\u8d27\u673a\uff0c\u5e76\u4ee5\u6700\u7ec8\u7684\u8d26\u6237\u4f59\u989d\u4f5c\u4e3a\u8bc4\u5206\u6807\u51c6\u3002\u5982\u56fe 4\uff08\u5de6\uff09\u6240\u793a\uff0cGLM-5 \u5728\u6240\u6709\u5f00\u6e90\u6a21\u578b\u4e2d\u62d4\u5f97\u5934\u7b79\uff0c\u6700\u7ec8\u8d26\u6237\u4f59\u989d\u8fbe\u5230 4432 \u7f8e\u5143\uff0c\u5c55\u73b0\u51fa\u5353\u8d8a\u7684\u957f\u671f\u89c4\u5212\u4e0e<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e8%b5%84%e6%ba%90\" title=\"View all posts in \u8d44\u6e90\" target=\"_blank\" class=\"tag_link\">\u8d44\u6e90<\/a>\u7ba1\u7406\u80fd\u529b\uff0c\u5176\u8868\u73b0\u5df2\u903c\u8fd1 Claude Opus 4.5\u3002<\/p>\n<p>\u56fe 4\uff08\u53f3\uff09\u8fdb\u4e00\u6b65\u5c55\u793a\u4e86\u57fa\u4e8e\u5185\u90e8\u8bc4\u4f30\u5957\u4ef6 CC-Bench-V2 \u7684\u6d4b\u8bd5\u7ed3\u679c\u3002\u6570\u636e\u663e\u793a\uff0cGLM-5 \u5728\u524d\u7aef\u3001\u540e\u7aef\u53ca\u957f\u7a0b\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u5747\u663e\u8457\u8d85\u8d8a\u4e86\u524d\u4ee3 GLM-4.7\uff0c\u8fdb\u4e00\u6b65\u7f29\u5c0f\u4e86\u4e0e Claude Opus 4.5 \u7684\u5dee\u8ddd\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/E74D413F1D03186CF22BE920D72E5A1A40532AC7_size200_w1080_h350.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe4\uff1a\u5728\u591a\u4e2a\u957f\u7a0b\u4efb\u52a1\u4e0a\u7684\u7ed3\u679c\u3002\u5de6\uff1aVending-Bench 2\uff1b\u53f3\uff1aCC-Bench-V2\u3002<\/p>\n<p>\u56fe 5 \u8be6\u7ec6\u5c55\u793a\u4e86 GLM-5 \u7684\u6574\u4f53\u8bad\u7ec3\u6d41\u7a0b\u3002\u5728\u57fa\u7840\u6a21\u578b\u8bad\u7ec3\u9636\u6bb5\uff0c\u6211\u4eec\u4f7f\u7528\u4e86\u89c4\u6a21\u8fbe 27 \u4e07\u4ebf token \u7684\u6d77\u91cf\u8bed\u6599\u5e93\uff0c\u5e76\u5728\u8bad\u7ec3\u521d\u671f\u91cd\u70b9\u5f15\u5165\u4ee3\u7801\u4e0e\u63a8\u7406\u6570\u636e\u3002\u968f\u540e\u8fdb\u5165\u4e2d\u671f\u8bad\u7ec3\uff08Mid-training\uff09\u9636\u6bb5\uff0c\u6211\u4eec\u5c06\u4e0a\u4e0b\u6587\u7a97\u53e3\u4ece 4K \u9010\u6b65\u6269\u5c55\u81f3 200K\uff0c\u5e76\u4e13\u95e8\u9488\u5bf9\u957f\u4e0a\u4e0b\u6587 Agent \u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u4ee5\u4fdd\u969c\u6a21\u578b\u5728\u590d\u6742\u5de5\u4f5c\u6d41\u4e2d\u7684\u6267\u884c\u7a33\u5b9a\u6027\u3002\u5728\u540e\u8bad\u7ec3\uff08Post-training\uff09\u9636\u6bb5\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u5957\u5e8f\u5217\u5316\u7684\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6d41\u7a0b\uff1a\u4f9d\u6b21\u5728\u63a8\u7406\u3001\u667a\u80fd\u4f53\u3001\u901a\u7528\u9886\u57df\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6211\u4eec\u5728\u5168\u6d41\u7a0b\u4e2d\u5e94\u7528\u4e86\u8de8\u9636\u6bb5\u5728\u7ebf\u84b8\u998f\uff08Cross-stage online distillation\uff09\u6280\u672f\uff0c\u6709\u6548\u514b\u670d\u4e86\u707e\u96be\u6027\u9057\u5fd8\u95ee\u9898\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/65362A818D970D06C8AFCD46A8220424239149CE_size256_w1080_h609.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe5\uff1aGLM-5 \u7684\u6574\u4f53\u8bad\u7ec3\u6d41\u7a0b\u3002<\/p>\n<p>\u603b\u4f53\u800c\u8a00\uff0cGLM-5 \u80fd\u591f\u5b9e\u73b0\u6027\u80fd\u7684\u5927\u5e45\u8dc3\u5347\uff0c\u4e3b\u8981\u5f97\u76ca\u4e8e\u4ee5\u4e0b\u56db\u5927\u6280\u672f\u521b\u65b0\uff1a<\/p>\n<p>\u7b2c\u4e00\uff0c\u5f15\u5165 DSA \u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\uff08DeepSeek Sparse Attention, DSA\uff09\u3002\u8fd9\u4e00\u5168\u65b0\u67b6\u6784\u6781\u5927\u964d\u4f4e\u4e86\u8bad\u7ec3\u4e0e\u63a8\u7406\u6210\u672c\u3002\u6b64\u524d\u7684 GLM-4.5 \u4f9d\u8d56\u6807\u51c6 MoE \u67b6\u6784\u63d0\u5347\u6548\u7387\uff0c\u800c DSA \u673a\u5236\u5219\u4f7f GLM-5 \u80fd\u591f\u6839\u636e Token \u7684\u91cd\u8981\u6027\u52a8\u6001\u5206\u914d\u6ce8\u610f\u529b\u8d44\u6e90\u3002\u5728\u4e0d\u6298\u635f\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u548c\u63a8\u7406\u6df1\u5ea6\u7684\u524d\u63d0\u4e0b\uff0c\u7b97\u529b\u5f00\u9500\u5f97\u4ee5\u5927\u5e45\u524a\u51cf\u3002\u5f97\u76ca\u4e8e\u6b64\uff0c\u6211\u4eec\u5c06\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u6210\u529f\u6269\u5c55\u81f3 744B\uff087440 \u4ebf\uff09\uff0c\u540c\u65f6\u5c06\u8bad\u7ec3 Token \u89c4\u6a21\u63d0\u5347\u81f3 28.5T\uff0828.5 \u4e07\u4ebf\uff09\u3002<\/p>\n<p>\u7b2c\u4e8c\uff0c\u6784\u5efa\u5168\u65b0\u7684\u5f02\u6b65 RL \u57fa\u7840\u8bbe\u65bd\u3002\u57fa\u4e8e GLM-4.5 \u65f6\u671f slime \u6846\u67b6\u201c\u8bad\u7ec3\u4e0e\u63a8\u7406\u89e3\u8026\u201d\u7684\u8bbe\u8ba1\uff0c\u6211\u4eec\u7684<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e6%96%b0%e5%9f%ba%e5%bb%ba\" title=\"View all posts in \u65b0\u57fa\u5efa\" target=\"_blank\" class=\"tag_link\">\u65b0\u57fa\u5efa<\/a>\u8fdb\u4e00\u6b65\u5b9e\u73b0\u4e86\u201c\u751f\u6210\u4e0e\u8bad\u7ec3\u201d\u7684\u6df1\u5ea6\u89e3\u8026\uff0c\u5c06 GPU \u5229\u7528\u7387\u63a8\u5411\u6781\u81f4\u3002\u8be5\u7cfb\u7edf\u652f\u6301\u6a21\u578b\u5f00\u5c55\u5927\u89c4\u6a21\u7684\u667a\u80fd\u4f53\uff08Agent\uff09\u8f68\u8ff9\u63a2\u7d22\uff0c\u5927\u5e45\u51cf\u7f13\u4e86\u4ee5\u5f80\u62d6\u6162\u8fed\u4ee3\u901f\u5ea6\u7684\u540c\u6b65\u74f6\u9888\uff0c\u8ba9 RL \u540e\u8bad\u7ec3\u6d41\u7a0b\u7684\u6548\u7387\u5b9e\u73b0\u4e86\u8d28\u7684\u98de\u8dc3\u3002<\/p>\n<p>\u7b2c\u4e09\uff0c\u63d0\u51fa\u5168\u65b0\u7684\u5f02\u6b65 Agent RL \u7b97\u6cd5\u3002\u8be5\u7b97\u6cd5\u65e8\u5728\u5168\u9762\u63d0\u5347\u6a21\u578b\u7684\u81ea\u4e3b\u51b3\u7b56\u8d28\u91cf\u3002GLM-4.5 \u66fe\u4f9d\u9760\u8fed\u4ee3\u81ea\u84b8\u998f\u548c\u7ed3\u679c\u76d1\u7763\u6765\u8bad\u7ec3 Agent\uff1b\u800c\u5728 GLM-5 \u4e2d\uff0c\u6211\u4eec\u7814\u53d1\u7684\u5f02\u6b65\u7b97\u6cd5\u4f7f\u6a21\u578b\u80fd\u591f\u4ece\u591a\u6837\u5316\u7684\u957f\u5468\u671f\u4ea4\u4e92\u4e2d\u6301\u7eed\u5b66\u4e60\u3002\u8fd9\u4e00\u7b97\u6cd5\u9488\u5bf9\u52a8\u6001\u73af\u5883\u4e0b\u7684\u89c4\u5212\u4e0e\u81ea\u6211\u7ea0\u9519\u80fd\u529b\u8fdb\u884c\u4e86\u6df1\u5ea6\u4f18\u5316\uff0c\u8fd9\u4e5f\u6b63\u662f GLM-5 \u80fd\u591f\u5728\u771f\u5b9e\u7f16\u7a0b\u573a\u666f\u4e2d\u8868\u73b0\u5353\u8d8a\u7684\u5e95\u5c42\u903b\u8f91\u3002<\/p>\n<p>\u7b2c\u56db\uff0c\u5168\u9762\u62e5\u62b1\u56fd\u4ea7\u7b97\u529b\u751f\u6001\u3002\u4ece\u6a21\u578b\u53d1\u5e03\u4f0a\u59cb\uff0cGLM-5 \u5c31\u539f\u751f\u9002\u914d\u4e86\u4e2d\u56fd GPU \u751f\u6001\u3002\u6211\u4eec\u5df2\u5b8c\u6210\u4ece\u5e95\u5c42\u5185\u6838\u5230\u4e0a\u5c42\u63a8\u7406\u6846\u67b6\u7684\u6df1\u5ea6\u4f18\u5316\uff0c\u5168\u9762\u517c\u5bb9\u4e03\u5927\u4e3b\u6d41\u56fd\u4ea7<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e8%8a%af%e7%89%87\" title=\"View all posts in \u82af\u7247\" target=\"_blank\" class=\"tag_link\">\u82af\u7247<\/a>\u5e73\u53f0\uff1a<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e5%8d%8e%e4%b8%ba\" title=\"View all posts in \u534e\u4e3a\" target=\"_blank\" class=\"tag_link\">\u534e\u4e3a<\/a>\u6607\u817e\u3001\u6469\u5c14\u7ebf\u7a0b\u3001\u6d77\u5149\u3001\u5bd2\u6b66\u7eaa\u3001\u6606\u4ed1\u82af\u3001\u5929\u6570\u667a\u82af\u4e0e\u71e7\u539f\u3002<\/p>\n<p>\u51ed\u501f\u4e0a\u8ff0\u8fdb\u6b65\uff0cGLM-5 \u4e0d\u4ec5\u662f\u4e00\u4e2a\u66f4\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u66f4\u662f\u4e0b\u4e00\u4ee3 AI Agent \u66f4\u9ad8\u6548\u3001\u66f4\u5b9e\u7528\u7684\u57fa\u7840\u6a21\u578b\u3002\u6211\u4eec\u5411\u793e\u533a\u5f00\u6e90 GLM-5\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63a8\u52a8\u9ad8\u6548\u7684\u3001\u9762\u5411 Agent \u7684\u901a\u7528<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e4%ba%ba%e5%b7%a5%e6%99%ba%e8%83%bd\" title=\"View all posts in \u4eba\u5de5\u667a\u80fd\" target=\"_blank\" class=\"tag_link\">\u4eba\u5de5\u667a\u80fd<\/a>\u7684\u53d1\u5c55\u3002<\/p>\n<p>2 \u9884\u8bad\u7ec3<\/p>\n<p>\u4e0e GLM-4.5 \u7c7b\u4f3c\uff0cGLM-5 \u7684\u57fa\u7840\u6a21\u578b\u7ecf\u5386\u4e24\u4e2a\u9636\u6bb5\uff1a\u7528\u4e8e\u901a\u7528\u8bed\u8a00\u548c\u7f16\u7a0b\u80fd\u529b\u7684\u9884\u8bad\u7ec3\uff0c\u4ee5\u53ca\u7528\u4e8e Agent \u548c\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u7684\u4e2d\u671f\u8bad\u7ec3\u3002\u6211\u4eec\u589e\u52a0\u4e86 GLM-5 \u6240\u6709\u8bad\u7ec3\u9636\u6bb5\u7684\u8bad\u7ec3 token \u9884\u7b97\uff0c\u57fa\u7840\u6a21\u578b\u603b\u8bad\u7ec3\u91cf\u8fbe 28.5 \u4e07\u4ebf token\u3002<\/p>\n<p>2.1 \u67b6\u6784<\/p>\n<p>\u6a21\u578b\u89c4\u6a21 scaling\u3002GLM-5 \u5c06\u4e13\u5bb6\u6570\u91cf\u6269\u5c55\u81f3 256 \u4e2a\uff0c\u5e76\u5c06\u5c42\u6570\u51cf\u5c11\u81f3 80 \u5c42\uff0c\u4ee5\u51cf\u5c11\u4e13\u5bb6\u5e76\u884c\u901a\u4fe1\u5f00\u9500\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u53c2\u6570\u91cf\u8fbe\u5230 744B\uff08\u6fc0\u6d3b\u53c2\u6570 40B\uff09\uff0c\u662f GLM-4.5\uff08\u603b\u53c2\u6570 355B\uff0c\u6fc0\u6d3b\u53c2\u6570 32B\uff09\u7684\u4e24\u500d\u3002<\/p>\n<p>Multi-latent \u6ce8\u610f\u529b\uff08MLA\uff09\u3002\u901a\u8fc7\u91c7\u7528\u538b\u7f29KV\u7f13\u5b58\u7684\u7ef4\u5ea6\uff0cMLA \u5728\u6548\u679c\u4e0a\u4e0e\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff08GQA\uff09\u76f8\u5f53\uff0c\u4f46\u80fd\u63d0\u4f9b\u66f4\u5c11\u7684\u663e\u5b58\u5360\u7528\u548c\u66f4\u5feb\u7684\u957f\u4e0a\u4e0b\u6587\u5e8f\u5217\u5904\u7406\u901f\u5ea6\u3002<\/p>\n<p>\u7136\u800c\uff0c\u5728\u4f7f\u7528 Muon \u4f18\u5316\u5668\u7684\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0\u5177\u6709 576 \u7ef4\u6f5c\u53d8\u91cf KV \u7f13\u5b58\u7684 MLA \u65e0\u6cd5\u5339\u914d\u5177\u6709 8 \u4e2a\u67e5\u8be2\u7ec4\u7684 GQA\uff08GQA-8\uff0c2048 \u7ef4 KV \u7f13\u5b58\uff09\u7684\u6027\u80fd\u3002\u4e3a\u5f25\u8865\u6027\u80fd\u5dee\u8ddd\uff0c\u6211\u4eec\u5bf9 GLM-4.5 \u4e2d Muon \u4f18\u5316\u5668\u7684\u65b9\u6848\u63d0\u51fa\u4e86\u6539\u8fdb\u3002\u5728\u539f\u59cb\u65b9\u6848\u4e2d\uff0c\u6211\u4eec\u5bf9\u591a\u5934\u67e5\u8be2\u3001\u952e\u3001\u503c\u7684\u4e0a\u6295\u5f71\u77e9\u9635 WUQ\u3001WUK\u3001WUV \u5e94\u7528\u77e9\u9635\u6b63\u4ea4\u5316\u3002\u53d6\u800c\u4ee3\u4e4b\u7684\u662f\uff0c\u6211\u4eec\u5c06\u8fd9\u4e9b\u77e9\u9635\u62c6\u5206\u4e3a\u9002\u7528\u4e8e\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u7684\u66f4\u5c0f\u77e9\u9635\uff0c\u5e76\u5bf9\u8fd9\u4e9b\u72ec\u7acb\u77e9\u9635\u5e94\u7528\u77e9\u9635\u6b63\u4ea4\u5316\u3002\u8be5\u65b9\u6cd5\u79f0\u4e3a Muon Split\uff0c\u4f7f\u4e0d\u540c\u6ce8\u610f\u529b\u5934\u7684\u6295\u5f71\u6743\u91cd\u80fd\u591f\u4ee5\u4e0d\u540c\u901f\u7387\u66f4\u65b0\u3002\u5982\u8868 1 \u6240\u793a\uff0c\u8be5\u65b9\u6cd5\u6709\u6548\u63d0\u5347\u4e86 MLA \u7684\u6027\u80fd\uff0c\u4f7f\u5176\u4e0e GQA-8 \u76f8\u5f53\u3002\u5b9e\u8df5\u4e2d\uff0c\u6211\u4eec\u8fd8\u53d1\u73b0\u4f7f\u7528 Muon Split \u540e\uff0cGLM-5 \u6ce8\u610f\u529b\u5206\u6570\u7684\u89c4\u6a21\u5728\u9884\u8bad\u7ec3\u671f\u95f4\u65e0\u9700\u4efb\u4f55\u88c1\u526a<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e7%ad%96%e7%95%a5\" title=\"View all posts in \u7b56\u7565\" target=\"_blank\" class=\"tag_link\">\u7b56\u7565<\/a>\u5373\u53ef\u4fdd\u6301\u7a33\u5b9a\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/F67E84A69577B636F33F00FCCD8CB9C06E79E3B8_size48_w1080_h174.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88681\uff1a\u5728 GQA-8 \u548c MLA \u7b49\u5404\u53d8\u4f53\u4e0a\u7684\u8bc4\u4f30\u7ed3\u679c\u3002<\/p>\n<p>MLA \u7684\u53e6\u4e00\u4e2a\u7f3a\u70b9\u662f\u5176\u5728\u89e3\u7801\u671f\u95f4\u7684\u9ad8\u8ba1\u7b97\u6210\u672c\u3002\u5728\u89e3\u7801\u65f6\uff0cMLA \u6267\u884c 576 \u7ef4\u70b9\u79ef\u8fd0\u7b97\uff0c\u9ad8\u4e8e GQA \u7684 128 \u7ef4\u8ba1\u7b97\u91cf\u3002\u867d\u7136 DeepSeek-V3 \u4e2d\u6ce8\u610f\u529b\u5934\u6570\u91cf\u662f\u6839\u636e H800 \u7684 roofline \u9009\u62e9\u7684\uff0c\u4f46\u8fd9\u5bf9\u5176\u4ed6\u786c\u4ef6\u5e76\u4e0d\u9002\u5408\u3002\u9274\u4e8e MLA \u5728\u8bad\u7ec3\u548c\u9884\u586b\u5145\u671f\u95f4\u7684\u591a\u5934\u6ce8\u610f\u529b\uff08MHA\uff09\u98ce\u683c\uff0c\u6211\u4eec\u5c06\u5934\u7ef4\u5ea6\u4ece 192 \u589e\u5927\u5230 256\uff0c\u5e76\u5c06\u6ce8\u610f\u529b\u5934\u6570\u91cf\u51cf\u5c11 1\/3\u3002\u8fd9\u5728\u4fdd\u6301\u8bad\u7ec3\u8ba1\u7b97\u91cf\u548c\u53c2\u6570\u6570\u91cf\u4e0d\u53d8\u7684\u540c\u65f6\u964d\u4f4e\u4e86\u89e3\u7801\u8ba1\u7b97\u91cf\u3002\u8be5\u53d8\u4f53\u5728\u8868 1 \u4e2d\u4e3a MLA-256\uff0c\u5728 Muon Split \u4e0b\u4e0e MLA \u6027\u80fd\u76f8\u5f53\u3002<\/p>\n<p>\u53c2\u6570\u5171\u4eab\u7684\u591a token \u9884\u6d4b\uff08MTP\uff09\u3002\u591a token \u9884\u6d4b\u53ef\u4ee5\u63d0\u5347\u57fa\u7840\u6a21\u578b\u6027\u80fd\uff0c\u5e76\u4f5c\u4e3a\u63a8\u6d4b\u89e3\u7801\u7684\u8349\u7a3f\u6a21\u578b\u3002\u7136\u800c\uff0c\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u9884\u6d4b\u63a5\u4e0b\u6765\u7684 n \u4e2a token \u9700\u8981 n \u4e2a MTP \u5c42\u3002\u56e0\u6b64\uff0cMTP \u53c2\u6570\u548c KV \u7f13\u5b58\u7684\u663e\u5b58\u5360\u7528\u968f\u63a8\u6d4b\u6b65\u6570\u7ebf\u6027\u589e\u957f\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cDeepSeek-V3 \u4f7f\u7528\u5355\u4e2a MTP \u5c42\u8bad\u7ec3\uff0c\u5e76\u5728\u63a8\u7406\u671f\u95f4\u9884\u6d4b\u63a5\u4e0b\u6765 2 \u4e2a token\u3002\u8bad\u7ec3\u4e0e\u63a8\u7406\u7684\u4e0d\u4e00\u81f4\u6027\u964d\u4f4e\u4e86\u7b2c\u4e8c\u4e2a token \u7684\u63a5\u53d7\u7387\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u63d0\u51fa\u5728\u8bad\u7ec3\u671f\u95f4\u5171\u4eab 3 \u4e2a MTP \u5c42\u7684\u53c2\u6570\u3002\u8fd9\u5728\u4fdd\u6301\u8349\u7a3f\u6a21\u578b\u5185\u5b58\u6210\u672c\u4e0e DeepSeek-V3 \u4e00\u81f4\u7684\u540c\u65f6\u63d0\u5347\u4e86\u63a5\u53d7\u7387\u3002\u8868 2 \u663e\u793a\uff0c\u5728\u76f8\u540c\u63a8\u6d4b\u6b65\u6570\uff084 \u6b65\uff09\u7684\u79c1\u6709\u63d0\u793a\u96c6\u4e0a\uff0cGLM-5 \u7684\u63a5\u53d7\u957f\u5ea6\u6bd4 DeepSeek-V3.2 \u66f4\u957f\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/E4ECEA7B87301AB7F736D9BE6B961D760FCF5525_size21_w748_h248.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88682\uff1aDeepSeek-V3.2 \u4e0e GLM-5 \u63a5\u53d7\u957f\u5ea6\u7684\u5bf9\u6bd4\u3002<\/p>\n<p>2.1.1 \u57fa\u4e8e DSA \u7684\u7ee7\u7eed\u9884\u8bad\u7ec3<\/p>\n<p>\u6211\u4eec\u5728\u8bad\u7ec3\u4e2d\u91c7\u7528\u4e86 DSA\u3002DSA \u7684\u6838\u5fc3\u7406\u5ff5\u662f\u5c06\u4f20\u7edf\u7684\u5bc6\u96c6 O(L&sup2;) \u6ce8\u610f\u529b\uff08\u5728 200K \u4e0a\u4e0b\u6587\u4e2d\u8ba1\u7b97\u91cf\u6781\u4e3a\u6602\u8d35\uff09\u66ff\u6362\u4e3a\u52a8\u6001\u7684\u7ec6\u7c92\u5ea6\u9009\u62e9\u673a\u5236\u3002\u4e0e\u56fa\u5b9a\u6a21\u5f0f\uff08\u5982\u6ed1\u52a8\u7a97\u53e3\uff09\u4e0d\u540c\uff0cDSA\u201c\u5ba1\u89c6\u201d\u5185\u5bb9\u4ee5\u51b3\u5b9a\u54ea\u4e9b token \u91cd\u8981\u3002\u4ece\u7814\u7a76\u8005\u89d2\u5ea6\u6765\u770b\uff0cDSA \u7279\u522b\u6709\u8da3\u4e4b\u5904\u5728\u4e8e\u5b83\u662f\u901a\u8fc7\u4ece\u7a20\u5bc6\u57fa\u7840\u6a21\u578b\u8fdb\u884c\u7ee7\u7eed\u9884\u8bad\u7ec3\u5f15\u5165\u7684\uff0c\u4ece\u800c\u907f\u514d\u4e86\u201c\u5929\u6587\u6570\u5b57\u201d\u7ea7\u522b\u7684\u4ece\u5934\u8bad\u7ec3\u6210\u672c\u3002\u8fc7\u6e21\u9075\u5faa\u201c\u7a20\u5bc6\u9884\u70ed\u201d\u4e0e\u201c\u7a00\u758f\u8bad\u7ec3\u9002\u914d\u201d\u7684\u4e24\u9636\u6bb5\u7b56\u7565\u3002DeepSeek-V3.2-Exp \u4fdd\u6301\u4e86\u4e0e\u5176\u7a20\u5bc6\u524d\u8eab\u76f8\u540c\u7684\u57fa\u51c6\u6027\u80fd\uff0c\u8bc1\u660e\u957f\u4e0a\u4e0b\u6587\u4e2d 90% \u7684\u6ce8\u610f\u529b\u6761\u76ee\u786e\u5b9e\u662f\u5197\u4f59\u7684\u3002DSA \u5c06\u957f\u5e8f\u5217\u7684\u6ce8\u610f\u529b\u8ba1\u7b97\u964d\u4f4e\u7ea6 1.5-2 \u500d\uff0c\u8fd9\u5bf9\u4e8e\u6211\u4eec\u6b63\u5728\u6784\u5efa\u7684\u63a8\u7406\u5bc6\u96c6\u578b Agent \u975e\u5e38\u91cd\u8981\uff0c\u80fd\u591f\u4ee5\u4e00\u534a GPU \u6210\u672c\u5904\u7406 200K \u4e0a\u4e0b\u6587\u3002<\/p>\n<p>DSA \u8bad\u7ec3\u4ece\u4e2d\u8bad\u7ec3\u7ed3\u675f\u65f6\u7684\u57fa\u7840\u6a21\u578b\u5f00\u59cb\u3002\u9884\u70ed\u9636\u6bb5\u7ecf\u5386 1000 \u6b65\uff0c\u6bcf\u6b65\u5728 14 \u4e2a\u5e8f\u5217\uff08\u6bcf\u4e2a\u5e8f\u5217\u6709 202752 \u4e2a token\uff09\u4e0a\u8bad\u7ec3\uff0c\u6700\u5927\u5b66\u4e60\u7387\u4e3a 5e-3\u3002\u7a00\u758f\u9002\u914d\u9636\u6bb5\u9075\u5faa\u4e2d\u671f\u8bad\u7ec3\u7684\u6570\u636e\u548c\u8d85\u53c2\u6570\u8bbe\u7f6e\uff0c\u7ecf\u5386 20B token\u3002\u5c3d\u7ba1\u8bad\u7ec3\u9884\u7b97\u8fdc\u5c0f\u4e8e DeepSeek-V3.2\uff08943.7B token\uff09\uff0c\u6211\u4eec\u53d1\u73b0\u8fd9\u5df2\u8db3\u4ee5\u4f7f DSA \u6a21\u578b\u7684\u6027\u80fd\u4e0e\u539f\u59cb MLA \u6a21\u578b\u76f8\u5f53\u3002\u5982\u8868 3 \u6240\u793a\uff0cDSA \u6a21\u578b\u7684\u957f\u4e0a\u4e0b\u6587\u6027\u80fd\u4e0e MLA \u6a21\u578b\u63a5\u8fd1\u3002\u4e3a\u8fdb\u4e00\u6b65\u9a8c\u8bc1 DSA \u8bad\u7ec3\u7684\u6709\u6548\u6027\uff0c\u6211\u4eec\u5206\u522b\u7528\u76f8\u540c\u7684 SFT \u6570\u636e\u5bf9 DSA \u548c MLA \u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u53d1\u73b0\u4e24\u4e2a\u6a21\u578b\u5728\u8bad\u7ec3\u635f\u5931\u548c\u8bc4\u4f30\u57fa\u51c6\u4e0a\u65d7\u9f13\u76f8\u5f53\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/7223986DDBE1AC004FE9D3E96CA4474958A594FA_size33_w1080_h138.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88683\uff1aMLA \u4e0e DSA \u57fa\u7840\u6a21\u578b\u5728\u957f\u4e0a\u4e0b\u6587\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u5bf9\u6bd4\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/316A98FDBAA4A80ED9775A8F177EBD17F9ABF1CE_size240_w1080_h631.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe6\uff1aMLA \u4e0e DSA \u8bad\u7ec3\u7684 SFT \u635f\u5931\u66f2\u7ebf\u5bf9\u6bd4\u3002\u7ed3\u679c\u91c7\u7528\u7a97\u53e3\u5927\u5c0f\u4e3a 50 \u7684\u6ed1\u52a8\u5e73\u5747\u8fdb\u884c\u5e73\u6ed1\u5904\u7406\u3002<\/p>\n<p>2.1.2 \u9ad8\u6548\u6ce8\u610f\u529b\u53d8\u4f53\u7684\u6d88\u878d\u7814\u7a76<\/p>\n<p>\u9664 DSA \u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u57fa\u4e8e GLM-9B \u63a2\u7d22\u4e86\u591a\u79cd\u66ff\u4ee3\u7684\u9ad8\u6548\u6ce8\u610f\u529b\u673a\u5236\u3002\u57fa\u7ebf\u5728\u6240\u6709 40 \u5c42\u4e2d\u91c7\u7528\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff0c\u5e76\u5df2\u5728 128K token \u4e0a\u4e0b\u6587\u7a97\u53e3\u4e0b\u8fdb\u884c\u5fae\u8c03\u3002\u6211\u4eec\u8bc4\u4f30\u4e86\u4ee5\u4e0b\u65b9\u6cd5\uff1a<\/p>\n<p>\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\uff08SWA\uff09\u4ea4\u9519\uff1a\u5168\u6ce8\u610f\u529b\u5c42\u4e0e\u7a97\u53e3\u6ce8\u610f\u529b\u5c42\u5728\u7f51\u7edc\u4e2d\u5747\u5300\u4ea4\u66ff\u7684\u56fa\u5b9a\u6a21\u5f0f\u3002<\/p>\n<p>\u95e8\u63a7 DeltaNet\uff08GDN\uff09\uff1a\u4e00\u79cd\u7ebf\u6027\u6ce8\u610f\u529b\u53d8\u4f53\uff0c\u5c06\u4e8c\u6b21\u65b9 softmax \u6ce8\u610f\u529b\u8ba1\u7b97\u66ff\u6362\u4e3a\u95e8\u63a7\u7ebf\u6027\u9012\u5f52\uff0c\u5c06\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4ece\u5e8f\u5217\u957f\u5ea6\u7684\u4e8c\u6b21\u65b9\u964d\u81f3\u7ebf\u6027\u3002<\/p>\n<p>\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e24\u9879\u6539\u8fdb\uff1a<\/p>\n<p>\u57fa\u4e8e\u641c\u7d22\u7684 SWA \u6a21\u5f0f\uff1a\u53d7 PostNAS \u542f\u53d1\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u57fa\u4e8e\u641c\u7d22\u7684\u9002\u914d\u65b9\u6cd5\uff0c\u8bc6\u522b\u7528\u4e8e SWA \u8f6c\u6362\u7684\u6700\u4f18\u5c42\u5b50\u96c6\uff0c\u540c\u65f6\u5728\u5176\u4f59\u5c42\u4fdd\u7559\u5168\u6ce8\u610f\u529b\u3002\u6211\u4eec\u91c7\u7528\u675f\u641c\u7d22\u7b56\u7565\u786e\u5b9a\u80fd\u6700\u5927\u5316\u957f\u4e0a\u4e0b\u6587\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\u7684\u914d\u7f6e\u3002\u4e3a\u964d\u4f4e\u8ba1\u7b97\u6210\u672c\uff0c\u6211\u4eec\u4ec5\u5728 16K \u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0b\u8fdb\u884c\u641c\u7d22\uff0c\u5e76\u5c06\u7ed3\u679c\u6a21\u5f0f\u63a8\u5e7f\u81f3\u6240\u6709\u5176\u4ed6\u8f93\u5165\u957f\u5ea6\u3002\u6700\u7ec8\u63a8\u5bfc\u51fa\u7684\u6a21\u5f0f\u4e3a SFSSFFSSSFFFFSSFSFFFFFFSFSFSSFSSFSFSSFSSS\uff0c\u5176\u4e2d S \u548c F \u5206\u522b\u8868\u793a SWA \u548c\u5168\u6ce8\u610f\u529b\u5c42\u3002\u5982\u8868 4 \u6240\u793a\uff0c\u8fd9\u79cd\u57fa\u4e8e\u641c\u7d22\u7684\u914d\u7f6e\u663e\u8457\u4f18\u4e8e\u56fa\u5b9a\u4ea4\u9519\u65b9\u6cd5\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5c3d\u7ba1\u4ec5\u5728 16K \u4e0b\u4f18\u5316\uff0c\u8be5\u6a21\u5f0f\u4ecd\u8868\u73b0\u51fa\u826f\u597d\u7684\u957f\u5ea6\u6cdb\u5316\u80fd\u529b\uff0c\u5728\u6240\u6709\u6d4b\u8bd5\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0b\u5747\u6709\u6548\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/1181CBABE96EF129BC2DA5FFB77134B4BE11F9E7_size55_w1080_h199.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88684\uff1aGLM-9B \u57fa\u7ebf\u6a21\u578b\u53ca\u4e24\u79cd SWA \u53d8\u4f53\u5728 RULER \u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u7ed3\u679c\uff08\u5747\u672a\u7ecf\u8fc7\u4efb\u4f55\u989d\u5916\u8bad\u7ec3\uff09\u3002\u4e24\u79cd SWA \u65b9\u6cd5\u5747\u91c7\u7528\u5168\u6ce8\u610f\u529b\u5c42\u4e0e SWA \u5c42 1:1 \u7684\u6bd4\u4f8b\uff0c\u6ed1\u52a8\u7a97\u53e3\u5927\u5c0f\u4e3a 4096 \u4e2a token\u3002\u57fa\u4e8e\u641c\u7d22\u7684 SWA \u6a21\u5f0f\u5728 16k \u4e0a\u4e0b\u6587\u957f\u5ea6\u65f6\u53d1\u73b0\u4e00\u6b21\uff0c\u5e76\u7edf\u4e00\u5e94\u7528\u4e8e\u6240\u6709\u8f93\u5165\u957f\u5ea6\u3002<\/p>\n<p>SimpleGDN\uff1a\u4e00\u79cd\u6781\u7b80\u7ebf\u6027\u5316\u7b56\u7565\uff0c\u8bbe\u8ba1\u7528\u4e8e\u6700\u5927\u7a0b\u5ea6\u590d\u7528\u9884\u8bad\u7ec3\u6743\u91cd\uff0c\u5728\u6301\u7eed\u8bad\u7ec3\u9002\u914d\u4e0a\u6539\u8fdb\u4e86 GDN\u3002\u6211\u4eec\u5b8c\u5168\u79fb\u9664\u4e86 Conv1d \u548c\u663e\u5f0f\u95e8\u63a7\u6a21\u5757\uff0c\u76f4\u63a5\u5c06\u9884\u8bad\u7ec3\u7684\u67e5\u8be2\u3001\u952e\u3001\u503c\u6295\u5f71\u6743\u91cd\u6620\u5c04\u5230\u7ebf\u6027\u9012\u5f52\u516c\u5f0f\u4e2d\u3002\u8fd9\u79cd\u7b80\u5316\u65e0\u9700\u989d\u5916\u53c2\u6570\uff0c\u540c\u65f6\u4fdd\u7559\u4e86\u7ebf\u6027\u6ce8\u610f\u529b\u7684\u6548\u7387\u4f18\u52bf\u3002<\/p>\n<p>\u8868 5 \u7684\u7ed3\u679c\u63ed\u793a\u4e86\u9ad8\u6548\u6ce8\u610f\u529b\u65b9\u6cd5\u4e4b\u95f4\u6e05\u6670\u7684\u6743\u8861\u5c42\u6b21\u3002\u539f\u751f\u4ea4\u9519\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\uff08SWA\uff09\u5728\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e0a\u5bfc\u81f4\u707e\u96be\u6027\u9000\u5316\uff08\u5982 RULER@128K \u4e0b\u964d 30.35\uff09\uff0c\u800c\u57fa\u4e8e\u641c\u7d22\u7684\u5c42\u9009\u62e9\u901a\u8fc7\u5728\u6700\u5173\u952e\u7684\u5c42\u4fdd\u7559\u5168\u6ce8\u610f\u529b\uff0c\u663e\u8457\u7f29\u5c0f\u4e86\u8fd9\u4e00\u5dee\u8ddd\u3002\u7ebf\u6027\u6ce8\u610f\u529b\u53d8\u4f53\uff08\u5982 GDN\uff09\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u8d28\u91cf\uff0c\u4f46\u4ee3\u4ef7\u662f\u589e\u52a0\u4e86\u989d\u5916\u53c2\u6570\uff1bSimpleGDN \u901a\u8fc7\u6700\u5927\u7a0b\u5ea6\u590d\u7528\u9884\u8bad\u7ec3\u6743\u91cd\u53d6\u5f97\u4e86\u6700\u4f73\u5e73\u8861\u3002\u7136\u800c\uff0c\u6240\u6709\u8fd9\u4e9b\u65b9\u6cd5\u5728\u7ec6\u7c92\u5ea6\u68c0\u7d22\u4efb\u52a1\u4e0a\u90fd\u5b58\u5728\u56fa\u6709\u7684\u7cbe\u5ea6\u5dee\u8ddd\u2014\u2014\u5373\u4f7f\u5728\u4e00\u534a\u5c42\u4fdd\u7559\u5168\u6ce8\u610f\u529b\u7684\u60c5\u51b5\u4e0b\uff0cRULER@128K \u4e0a\u6700\u9ad8\u635f\u5931 5.69 \u5206\uff0cRepoQA@128K \u4e0a\u6700\u9ad8\u635f\u5931 7.33 \u5206\u2014\u2014\u8fd9\u662f\u7ee7\u7eed\u8bad\u7ec3\u9002\u914d\u4e2d\u9ad8\u6548\u6ce8\u610f\u529b\u673a\u5236\u5f15\u5165\u7684\u4e0d\u53ef\u907f\u514d\u7684\u4fe1\u606f\u635f\u5931\u6240\u81f4\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cDSA \u5728\u8bbe\u8ba1\u4e0a\u662f\u65e0\u635f\u7684\uff1a\u5176\u8f7b\u91cf\u7ea7\u7d22\u5f15\u5668\u5728\u4e0d\u4e22\u5f03\u4efb\u4f55\u957f\u7a0b\u4f9d\u8d56\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0 token \u7ea7\u7a00\u758f\u6027\uff0c\u80fd\u591f\u5e94\u7528\u4e8e\u6240\u6709\u5c42\u800c\u4e0d\u964d\u4f4e\u8d28\u91cf\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/F7A9C7F5C0FD2A4C074AF8A806757A6B2E96D324_size64_w1080_h172.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88685\uff1a\u957f\u4e0a\u4e0b\u6587\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c\u3002\u6240\u6709\u9ad8\u6548\u6ce8\u610f\u529b\u53d8\u4f53\u5747\u4ece GLM-9B \u5168\u6ce8\u610f\u529b\u57fa\u7ebf\u6a21\u578b\u8fdb\u884c\u7ee7\u7eed\u8bad\u7ec3\u3002SWA pattern \u8868\u793a\u57fa\u4e8e\u641c\u7d22\u7684\u5c42\u9009\u62e9\u65b9\u5f0f\uff1bSWA interleave \u8868\u793a\u56fa\u5b9a\u4ea4\u66ff\u6a21\u5f0f\u3002\u2206@64K \u548c \u2206@128K \u5206\u522b\u8868\u793a\u5728 64K \u548c 128K \u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0b\u76f8\u5bf9\u4e8e\u5168\u6ce8\u610f\u529b\u57fa\u7ebf\u7684\u5dee\u503c\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/9F8AC6C8B5A5F0521748E8ACCB8DF8B09AC05C64_size48_w1080_h178.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88686\uff1aGLM-4.7-Flash \u642d\u914d DSA \u7684 RULER \u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c\u3002\u4ec5\u9884\u70ed\u53d8\u4f53\uff08warmup-only\uff09\u53ea\u8bad\u7ec3\u7d22\u5f15\u5668\u800c\u4fdd\u6301\u57fa\u7840\u6a21\u578b\u51bb\u7ed3\uff0c\u5b8c\u6574 DSA \u53d8\u4f53\u5219\u5bf9\u4e24\u8005\u8fdb\u884c\u8054\u5408\u8bad\u7ec3\uff0c\u5171\u8bad\u7ec3 150B token\u3002<\/p>\n<p>2.2 \u9884\u8bad\u7ec3\u6570\u636e<\/p>\n<p>\u7f51\u9875\u6570\u636e\u3002\u5728 GLM-4.5 \u6570\u636e\u6d41\u7a0b\u7684\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u5b8c\u5584\u4e86\u6d77\u91cf\u7f51\u9875\u6570\u636e\u96c6\u7684\u7b5b\u9009\u6807\u51c6\u3002\u6211\u4eec\u5f15\u5165\u4e86\u53e6\u4e00\u4e2a\u57fa\u4e8e\u8bed\u53e5\u5d4c\u5165\u7684 DCLM \u5206\u7c7b\u5668\uff0c\u4ee5\u8bc6\u522b\u548c\u6c47\u805a\u6807\u51c6\u5206\u7c7b\u5668\u4e4b\u5916\u7684\u9ad8\u8d28\u91cf\u6570\u636e\u3002\u4e3a\u5e94\u5bf9\u957f\u5c3e\u77e5\u8bc6\u6311\u6218\uff0c\u6211\u4eec\u5229\u7528\u4e16\u754c\u77e5\u8bc6\u5206\u7c7b\u5668\uff0c\u901a\u8fc7\u7ef4\u57fa\u767e\u79d1\u6761\u76ee\u548c LLM \u6807\u6ce8\u6570\u636e\u4f18\u5316\uff0c\u4ece\u539f\u672c\u4e2d\u4f4e\u8d28\u91cf\u6570\u636e\u4e2d\u7b5b\u9009\u6709\u4ef7\u503c\u4fe1\u606f\u3002<\/p>\n<p>\u4ee3\u7801\u6570\u636e\u3002\u6211\u4eec\u901a\u8fc7\u66f4\u65b0\u4e3b\u8981\u4ee3\u7801\u6258\u7ba1\u5e73\u53f0\u7684\u5feb\u7167\u548c\u66f4\u5927\u89c4\u6a21\u7684\u542b\u4ee3\u7801\u7f51\u9875\u96c6\u5408\u6765\u6269\u5145\u4ee3\u7801\u9884\u8bad\u7ec3\u8bed\u6599\u5e93\uff0c\u4f7f\u6a21\u7cca\u53bb\u91cd\u540e\u7684 token \u589e\u52a0 28%\u3002\u4e3a\u63d0\u5347\u8bed\u6599\u5e93\u5b8c\u6574\u6027\u5e76\u51cf\u5c11\u566a\u58f0\uff0c\u6211\u4eec\u4fee\u590d\u4e86 Software Heritage \u4ee3\u7801\u6587\u4ef6\u4e2d\u7684\u5143\u6570\u636e\u5bf9\u9f50\u95ee\u9898\uff0c\u5e76\u91c7\u7528\u4e86\u66f4\u7cbe\u51c6\u7684\u8bed\u8a00\u5206\u7c7b\u6d41\u7a0b\u3002\u6211\u4eec\u9075\u5faa GLM-4.5 \u7684\u8d28\u91cf\u611f\u77e5\u91c7\u6837\u7b56\u7565\u5904\u7406\u6e90\u4ee3\u7801\u548c\u4ee3\u7801\u76f8\u5173\u7f51\u9875\u6587\u6863\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u4e3a\u66f4\u591a\u4f4e\u8d44\u6e90\u7f16\u7a0b\u8bed\u8a00\uff08\u5982 Scala\u3001Swift\u3001Lua \u7b49\uff09\u8bad\u7ec3\u4e86\u4e13\u7528\u5206\u7c7b\u5668\uff0c\u63d0\u5347\u4e86\u8fd9\u4e9b\u8bed\u8a00\u7684\u91c7\u6837\u8d28\u91cf\u3002<\/p>\n<p>\u6570\u5b66\u4e0e\u79d1\u5b66\u6570\u636e\u3002\u6211\u4eec\u4ece\u7f51\u9875\u3001\u4e66\u7c4d\u548c\u8bba\u6587\u4e2d\u6536\u96c6\u9ad8\u8d28\u91cf\u6570\u5b66\u4e0e\u79d1\u5b66\u6570\u636e\uff0c\u4ee5\u8fdb\u4e00\u6b65\u589e\u5f3a\u63a8\u7406\u80fd\u529b\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u5b8c\u5584\u4e86\u7f51\u9875\u5185\u5bb9\u63d0\u53d6\u6d41\u7a0b\u4ee5\u53ca\u4e66\u7c4d\u548c\u8bba\u6587\u7684 PDF \u89e3\u6790\u673a\u5236\uff0c\u4ee5\u63d0\u5347\u6570\u636e\u8d28\u91cf\u3002\u6211\u4eec\u91c7\u7528\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5bf9\u5019\u9009\u6587\u6863\u8fdb\u884c\u8bc4\u5206\uff0c\u4ec5\u4fdd\u7559\u6700\u5177\u6559\u80b2\u4ef7\u503c\u7684\u5185\u5bb9\u3002\u5bf9\u4e8e\u957f\u6587\u6863\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u5206\u5757\u805a\u5408\u8bc4\u5206\u7b97\u6cd5\u4ee5\u63d0\u5347\u8bc4\u5206\u7cbe\u5ea6\u3002\u8fc7\u6ee4\u6d41\u7a0b\u4e25\u683c\u907f\u514d\u4f7f\u7528\u5408\u6210\u3001AI \u751f\u6210\u6216\u6a21\u677f\u7c7b\u6570\u636e\u3002<\/p>\n<p>2.3 \u4e2d\u671f\u8bad\u7ec3<\/p>\n<p>\u5728 GLM-4.5 \u5f15\u5165\u7684\u4e2d\u671f\u8bad\u7ec3\u6846\u67b6\u57fa\u7840\u4e0a\uff0cGLM-5 \u6269\u5927\u4e86\u8bad\u7ec3\u89c4\u6a21\u5e76\u5ef6\u957f\u4e86\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u4ee5\u8fdb\u4e00\u6b65\u589e\u5f3a\u6a21\u578b\u7684\u63a8\u7406\u3001\u957f\u4e0a\u4e0b\u6587\u548c Agent \u80fd\u529b\u3002<\/p>\n<p>\u6269\u5c55\u4e0a\u4e0b\u6587\u4e0e\u8bad\u7ec3\u89c4\u6a21\u3002\u6211\u4eec\u5206\u4e09\u4e2a\u9636\u6bb5\u9010\u6b65\u6269\u5c55\u4e0a\u4e0b\u6587\u7a97\u53e3\uff1a32K\uff081T token\uff09\u3001128K\uff08500B token\uff09\u548c 200K\uff0850B token\uff09\u3002\u4e0e GLM-4.5 \u6700\u5927 128K \u76f8\u6bd4\uff0c\u989d\u5916\u7684 200K \u9636\u6bb5\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u5904\u7406\u8d85\u957f\u6587\u6863\u548c\u590d\u6742\u591a\u6587\u4ef6\u4ee3\u7801\u5e93\u7684\u80fd\u529b\u3002\u957f\u6587\u6863\u548c\u5408\u6210 Agent \u8f68\u8ff9\u5728\u540e\u671f\u9636\u6bb5\u76f8\u5e94\u5730\u88ab\u4e0a\u91c7\u6837\u3002<\/p>\n<p>\u8f6f\u4ef6\u5de5\u7a0b\u6570\u636e\u3002\u6211\u4eec\u4fdd\u7559\u4e86\u5c06\u4ed3\u5e93\u7ea7\u4ee3\u7801\u6587\u4ef6\u3001\u63d0\u4ea4\u5dee\u5f02\u3001GitHub issues\u3001\u62c9\u53d6\u8bf7\u6c42\u548c\u76f8\u5173\u6e90\u6587\u4ef6\u62fc\u63a5\u6210\u7edf\u4e00\u8bad\u7ec3\u5e8f\u5217\u7684\u8303\u5f0f\u3002\u5728 GLM-5 \u4e2d\uff0c\u6211\u4eec\u653e\u5bbd\u4e86\u4ed3\u5e93\u7ea7\u8fc7\u6ee4\u6807\u51c6\u4ee5\u6269\u5927\u5408\u683c\u4ed3\u5e93\u6c60\uff0c\u83b7\u5f97\u7ea6 1000 \u4e07\u4e2a issue-PR \u5bf9\uff0c\u540c\u65f6\u52a0\u5f3a\u4e86\u5355\u4e2a issue \u7ea7\u522b\u7684\u8d28\u91cf\u8fc7\u6ee4\u4ee5\u964d\u4f4e\u566a\u58f0\u3002\u6211\u4eec\u8fd8\u4e3a\u6bcf\u4e2a issue-PR \u5bf9\u68c0\u7d22\u4e86\u66f4\u591a\u76f8\u5173\u6587\u4ef6\uff0c\u4ea7\u751f\u4e86\u66f4\u4e30\u5bcc\u7684\u5f00\u53d1\u4e0a\u4e0b\u6587\uff0c\u5e76\u6269\u5927\u4e86\u771f\u5b9e\u8f6f\u4ef6\u5de5\u7a0b\u573a\u666f\u7684\u8986\u76d6\u8303\u56f4\u3002\u8fc7\u6ee4\u540e\uff0cissue-PR \u90e8\u5206\u7684\u6570\u636e\u96c6\u5305\u542b\u7ea6 160B token\u3002<\/p>\n<p>\u957f\u4e0a\u4e0b\u6587\u6570\u636e\u3002\u6211\u4eec\u7684\u957f\u4e0a\u4e0b\u6587\u8bad\u7ec3\u96c6\u5305\u542b\u81ea\u7136\u6570\u636e\u548c\u5408\u6210\u6570\u636e\u3002\u81ea\u7136\u6570\u636e\u4ece\u4e66\u7c4d\u3001\u5b66\u672f\u8bba\u6587\u548c\u901a\u7528\u9884\u8bad\u7ec3\u8bed\u6599\u5e93\u6587\u6863\u4e2d\u7cbe\u9009\uff0c\u91c7\u7528\u591a\u9636\u6bb5\u8fc7\u6ee4\uff08\u56f0\u60d1\u5ea6\u3001\u53bb\u91cd\u3001\u957f\u5ea6\uff09\u5e76\u5bf9\u77e5\u8bc6\u5bc6\u96c6\u578b\u9886\u57df\u8fdb\u884c\u4e0a\u91c7\u6837\u3002\u5728\u5408\u6210\u6570\u636e\u6784\u5efa\u4e2d\uff0c\u53d7 NextLong \u548c EntropyLong \u542f\u53d1\uff0c\u6211\u4eec\u91c7\u7528\u591a\u79cd\u6280\u672f\u6784\u5efa\u957f\u7a0b\u4f9d\u8d56\u5173\u7cfb\u3002\u9ad8\u5ea6\u76f8\u4f3c\u7684\u6587\u672c\u901a\u8fc7\u4ea4\u9519\u6253\u5305\u805a\u5408\u6210\u5e8f\u5217\uff0c\u65e8\u5728\u7f13\u89e3\u201c\u4e2d\u95f4\u4e22\u5931\u201d\u73b0\u8c61\u5e76\u63d0\u5347\u5404\u7c7b\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u7684\u6027\u80fd\u3002\u5728 200K \u9636\u6bb5\uff0c\u6211\u4eec\u989d\u5916\u52a0\u5165\u4e86\u5c11\u91cf MRCR \u7c7b\u6570\u636e\uff0c\u5e76\u8bbe\u8ba1\u4e86\u591a\u79cd\u53d8\u4f53\u4ee5\u6269\u5c55 <a href=\"https:\/\/www.alphaseek.cn\/?tag=openai\" title=\"View all posts in OpenAI\" target=\"_blank\" class=\"tag_link\">OpenAI<\/a> \u539f\u59cb\u8303\u5f0f\uff0c\u4ee5\u589e\u5f3a\u6269\u5c55\u591a\u8f6e\u5bf9\u8bdd\u4e2d\u7684\u53ec\u56de\u80fd\u529b\u3002<\/p>\n<p>2.4 \u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd<\/p>\n<p>2.4.1 \u5185\u5b58\u6548\u7387<\/p>\n<p>\u7075\u6d3b\u7684 MTP \u5e03\u5c40\u3002\u5728\u4ea4\u9519\u6d41\u6c34\u5e76\u884c\u4e0b\uff0cMTP \u6a21\u5757\uff08\u5305\u542b\u5d4c\u5165\u5c42\u3001Transformer \u5757\u548c\u8f93\u51fa\u5c42\uff09\u663e\u5b58\u5360\u7528\u663e\u8457\u9ad8\u4e8e\u5176\u4ed6\u6a21\u5757\uff0c\u5bb9\u6613\u5bfc\u81f4\u4e0d\u540c\u6d41\u6c34\u5e76\u884c rank \u4e4b\u95f4\u663e\u5b58\u8d1f\u8f7d\u4e0d\u5747\u8861\u3002\u6211\u4eec\u5bf9 MTP \u6a21\u5757\u8fdb\u884c\u62c6\u5206\uff0c\u5c06\u8f93\u51fa\u5c42\u4e0e\u4e3b\u8f93\u51fa\u5c42\u7edf\u4e00\u653e\u7f6e\u5728\u6700\u540e\u9636\u6bb5\u4ee5\u5b9e\u73b0\u53c2\u6570\u5171\u4eab\uff0c\u5176\u4f59\u90e8\u5206\u524d\u79fb\u81f3\u524d\u4e00\u9636\u6bb5\uff0c\u4ece\u800c\u6539\u5584\u5404 rank \u4e4b\u95f4\u7684\u663e\u5b58\u5747\u8861\u3002<\/p>\n<p>\u6d41\u6c34\u7ebf ZeRO2 \u68af\u5ea6\u5206\u7247\u3002\u5728\u6d41\u6c34\u7ebf\u5e76\u884c\u4e2d\uff0c\u6bcf\u4e2a rank \u7ef4\u62a4\u591a\u4e2a stage\uff0c\u5e38\u89c4\u5b9e\u73b0\u4e0b\u6bcf\u4e2a stage \u5747\u9700\u7ef4\u62a4\u5b8c\u6574\u68af\u5ea6 buffer\uff0c\u663e\u5b58\u5f00\u9500\u8f83\u5927\u3002\u53d7 ZeRO2 \u601d\u60f3\u542f\u53d1\uff0c\u6211\u4eec\u5c06\u68af\u5ea6\u6309\u6570\u636e\u5e76\u884c\u7ef4\u5ea6\u5206\u7247\uff0c\u4f7f\u6bcf\u4e2a stage \u4ec5\u6301\u6709\u5176\u5bf9\u5e94\u5206\u7247\u3002\u540c\u65f6\u7ed3\u5408\u53cc\u7f13\u51b2\u7b56\u7565\uff0c\u5728\u4efb\u610f\u65f6\u523b\u4ec5\u4fdd\u7559\u4e24\u4e2a\u5b8c\u6574\u68af\u5ea6\u7d2f\u52a0 buffer\uff0c\u5b9e\u73b0\u68af\u5ea6\u7d2f\u52a0\u4e0e\u901a\u4fe1\u540c\u6b65\u5e76\u884c\u6267\u884c\uff0c\u5e76\u5c06\u5b8c\u6574\u68af\u5ea6\u7684\u5e38\u9a7b\u663e\u5b58\u89c4\u6a21\u63a7\u5236\u5728\u56fa\u5b9a\u8303\u56f4\u5185\u3002\u5728\u4e0d\u5f15\u5165\u989d\u5916\u540c\u6b65\u5f00\u9500\u7684\u60c5\u51b5\u4e0b\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u68af\u5ea6\u76f8\u5173\u663e\u5b58\u5f00\u9500\u3002<\/p>\n<p>Muon \u5206\u5e03\u5f0f\u4f18\u5316\u5668\u96f6\u5197\u4f59\u901a\u4fe1\u3002\u539f\u751f Muon \u5b9e\u73b0\u4f1a\u5728\u6bcf\u4e2a\u6570\u636e\u5e76\u884c rank \u4e0a\u6267\u884c\u5b8c\u6574\u53c2\u6570 all-gather\uff0c\u5e26\u6765\u989d\u5916\u901a\u4fe1\u548c\u663e\u5b58\u5360\u7528\u3002\u6211\u4eec\u5c06 all-gather \u8303\u56f4\u9650\u5236\u4e3a\u672c rank \u8d1f\u8d23\u7684\u53c2\u6570\u5206\u7247\uff0c\u5e76\u4e0e\u672c\u5730\u8ba1\u7b97\u91cd\u53e0\u6267\u884c\uff0c\u4ece\u800c\u6d88\u9664\u5197\u4f59\u901a\u4fe1\u5e76\u964d\u4f4e\u4f18\u5316\u5668\u9636\u6bb5\u7684\u663e\u5b58\u5f00\u9500\u3002<\/p>\n<p>\u6d41\u6c34\u7ebf\u6fc0\u6d3b\u5378\u8f7d\u3002\u5728\u6d41\u6c34\u7ebf\u9884\u70ed\u9636\u6bb5\uff0c\u4e2d\u95f4\u6fc0\u6d3b\u5728\u663e\u5b58\u4e2d\u7684\u9a7b\u7559\u65f6\u95f4\u8f83\u957f\u3002\u6211\u4eec\u6309\u5c42\u7c92\u5ea6\u5c06\u6fc0\u6d3b\u5728\u524d\u5411\u5b8c\u6210\u540e\u5378\u8f7d\u81f3 DRAM\uff0c\u5e76\u5728\u53cd\u5411\u8ba1\u7b97\u524d\u6309\u9700\u52a0\u8f7d\u3002\u5378\u8f7d\u4e0e\u52a0\u8f7d\u8fc7\u7a0b\u4e0e\u8ba1\u7b97\u91cd\u53e0\u6267\u884c\uff0c\u540c\u65f6\u907f\u514d\u4e0e P2P \u901a\u4fe1\u548c MoE token dispatch\/combine \u7ade\u4e89 PCIe \u5e26\u5bbd\u3002\u5728\u51e0\u4e4e\u4e0d\u589e\u52a0\u989d\u5916\u5f00\u9500\u7684\u524d\u63d0\u4e0b\uff0c\u6709\u6548\u538b\u7f29\u4e86\u6fc0\u6d3b\u663e\u5b58\u5360\u7528\uff0c\u4f7f\u663e\u5b58\u4f7f\u7528\u66f4\u52a0\u7a33\u5b9a\u53ef\u63a7\u3002<\/p>\n<p>\u6309\u5e8f\u5217\u5206\u5757\u7684\u8f93\u51fa\u6295\u5f71\u4f18\u5316\u3002\u5728\u957f\u5e8f\u5217\u8bad\u7ec3\u573a\u666f\u4e0b\uff0c\u8f93\u51fa\u6295\u5f71\u4e0e cross entropy loss \u8ba1\u7b97\u4f1a\u5e26\u6765\u8f83\u5927\u7684\u663e\u5b58\u538b\u529b\u3002\u6211\u4eec\u6309\u5e8f\u5217\u7ef4\u5ea6\u5206\u5757\u6267\u884c\u8f93\u51fa\u5c42\u4e0e loss \u8ba1\u7b97\uff0c\u5bf9\u6bcf\u4e2a\u5757\u72ec\u7acb\u5b8c\u6210\u524d\u5411\u4e0e\u53cd\u5411\u4f20\u64ad\u540e\u518d\u7ee7\u7eed\u5904\u7406\u4e0b\u4e00\u5757\uff0c\u4ece\u800c\u63a7\u5236\u8f93\u51fa\u9636\u6bb5\u7684\u663e\u5b58\u5cf0\u503c\u3002\u8be5\u5b9e\u73b0\u4e0e\u5e38\u89c4\u5b9e\u73b0\u76f8\u6bd4\u6027\u80fd\u4fdd\u6301\u63a5\u8fd1\u3002<\/p>\n<p>2.4.2 \u5e76\u884c\u6548\u7387<\/p>\n<p>\u9ad8\u6548\u7684\u5ef6\u8fdf\u6743\u91cd\u68af\u5ea6\u8ba1\u7b97\u3002\u4e3a\u51cf\u5c11\u6d41\u6c34\u7ebf\u201c\u6c14\u6ce1\u201d\uff0c\u6211\u4eec\u63a8\u8fdf\u5173\u952e\u8def\u5f84\u4e0a\u7684\u90e8\u5206\u6743\u91cd\u68af\u5ea6\u8ba1\u7b97\u3002\u901a\u8fc7\u7cbe\u7ec6\u7684\u663e\u5b58\u7ba1\u7406\u7b56\u7565\u548c\u8ba1\u7b97<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e9%80%9a%e8%ae%af\" title=\"View all posts in \u901a\u8baf\" target=\"_blank\" class=\"tag_link\">\u901a\u8baf<\/a>\u91cd\u53e0\uff0c\u5b9e\u73b0\u8bad\u7ec3\u541e\u5410\u63d0\u5347\u5e76\u4e14\u4fdd\u8bc1\u989d\u5916\u663e\u5b58\u5f00\u9500\u53d7\u63a7\uff0c\u4e0d\u81f3\u4e8e OOM\u3002<\/p>\n<p>\u9ad8\u6548\u957f\u5e8f\u5217\u8bad\u7ec3\u3002\u52a8\u6001\u957f\u5ea6\u7684\u957f\u5e8f\u5217\u8bad\u7ec3\u52a0\u5267\u4e86\u6570\u636e\u5e76\u884c\u548c\u6d41\u6c34\u5e76\u884c\u4e4b\u95f4\u7684\u8d1f\u8f7d\u4e0d\u5747\u8861\u3002\u6211\u4eec\u901a\u8fc7\u5de5\u4f5c\u8d1f\u8f7d\u5efa\u6a21\uff0c\u5728\u6570\u636e\u5e76\u884c\u7ec4\u4e4b\u95f4\u8fdb\u884c\u957f\u5e8f\u5217\u6570\u636e\u91cd\u6392\uff0c\u4e00\u5b9a\u7a0b\u5ea6\u7f13\u89e3\u4e86\u8d1f\u8f7d\u4e0d\u5747\u8861\u95ee\u9898\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u8fd8\u5728\u6570\u636e\u5e76\u884c\u7ec4\u95f4\u52a8\u6001\u8c03\u6574\u8ba1\u7b97\u8d1f\u8f7d\uff0c\u8fdb\u4e00\u6b65\u7f13\u89e3\u4e86\u6570\u636e\u5e76\u884c\u7ec4\u95f4\u7684\u6162\u7b49\u5f85\u95ee\u9898\u3002\u6b64\u5916\u6211\u4eec\u8fd8\u901a\u8fc7\u52a8\u6001\u521b\u5efa\u865a\u62df CP \u7ec4\u6765\u7f13\u89e3\u56fa\u5b9a\u5927\u5c0f CP \u7ec4\u5e26\u6765\u7684\u4e00\u4e9b\u8ba1\u7b97\u4f4e\u6548\u95ee\u9898\u3002\u6700\u540e\uff0c\u6211\u4eec\u5229\u7528\u673a\u5185\/\u673a\u95f4\u7684 all2all \u901a\u8baf\u91cd\u53e0\u6765\u51cf\u5c0f\u7aef\u5230\u7aef\u7684\u5ef6\u8fdf\u3002<\/p>\n<p>2.4.3 INT4 \u91cf\u5316\u611f\u77e5\u8bad\u7ec3<\/p>\n<p>\u4e3a\u5728\u4f4e\u7cbe\u5ea6\u4e0b\u63d0\u4f9b\u66f4\u597d\u7684\u51c6\u786e\u6027\uff0c\u6211\u4eec\u5728 SFT \u9636\u6bb5\u5e94\u7528\u4e86 INT4 \u91cf\u5316\u611f\u77e5\u8bad\u7ec3\uff08QAT\uff09\u3002\u6b64\u5916\uff0c\u4e3a\u8fdb\u4e00\u6b65\u964d\u4f4e\u8bad\u7ec3\u65f6\u95f4\u5f00\u9500\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u9002\u7528\u4e8e\u8bad\u7ec3\u548c\u79bb\u7ebf\u6743\u91cd\u91cf\u5316\u7684\u91cf\u5316 kernel \uff0c\u786e\u4fdd\u8bad\u7ec3\u4e0e\u63a8\u7406\u53ef\u4ee5\u505a\u5230\u6bd4\u7279\u7ea7\u522b\u5bf9\u9f50\u3002<\/p>\n<p>3 \u540e\u8bad\u7ec3<\/p>\n<p>GLM-5 \u7684\u540e\u8bad\u7ec3\u9636\u6bb5\u65e8\u5728\u5c06\u57fa\u7840\u6a21\u578b\u8f6c\u5316\u4e3a\u5177\u5907\u5f3a\u5927\u63a8\u7406\u3001\u7f16\u7a0b\u548c Agent \u80fd\u529b\u7684\u9ad8\u6548\u52a9\u624b\u3002\u5982\u4e0a\u56fe 5 \u6240\u793a\uff0c\u6211\u4eec\u7684\u6d41\u7a0b\u9075\u5faa\u6e10\u8fdb\u5f0f\u5bf9\u9f50\u7b56\u7565\uff1a\u4ece\u5f15\u5165\u590d\u6742\u4ea4\u9519\u601d\u8003\u6a21\u5f0f\u7684\u591a\u4efb\u52a1 SFT \u5f00\u59cb\uff0c\u968f\u540e\u662f\u4e13\u95e8\u9488\u5bf9\u63a8\u7406\u548c Agent \u4efb\u52a1\u7684 RL \u9636\u6bb5\uff0c\u6700\u540e\u662f\u9762\u5411\u4eba\u7c7b\u98ce\u683c\u5bf9\u9f50\u7684\u901a\u7528 RL \u9636\u6bb5\u3002\u901a\u8fc7\u8de8\u9636\u6bb5\u5728\u7ebf\u84b8\u998f\u4f5c\u4e3a\u6700\u7ec8\u7cbe\u70bc\u6b65\u9aa4\uff0cGLM-5 \u6709\u6548\u51cf\u8f7b\u4e86\u80fd\u529b\u9000\u5316\uff0c\u540c\u65f6\u5145\u5206\u5229\u7528\u4e86\u6bcf\u4e2a\u8bad\u7ec3\u9636\u6bb5\u7684\u6027\u80fd\u589e\u76ca\u3002<\/p>\n<p>3.1 \u76d1\u7763\u5fae\u8c03\uff08SFT\uff09<\/p>\n<p>\u4e0e GLM-4.5 \u76f8\u6bd4\uff0cGLM-5 \u5728 SFT \u9636\u6bb5\u663e\u8457\u6269\u5927\u4e86 Agent \u548c\u7f16\u7a0b\u6570\u636e\u7684\u89c4\u6a21\u3002GLM-5 \u7684 SFT \u8bed\u6599\u5e93\u6db5\u76d6\u4e09\u5927\u7c7b\u522b\uff1a<\/p>\n<p>\u901a\u7528\u5bf9\u8bdd\uff08\u95ee\u7b54\u3001\u5199\u4f5c\u3001\u89d2\u8272\u626e\u6f14\u3001\u7ffb\u8bd1\u3001\u591a\u8f6e\u5bf9\u8bdd\u548c\u957f\u4e0a\u4e0b\u6587\u4ea4\u4e92\uff09\uff1b<\/p>\n<p>\u63a8\u7406\uff08\u6570\u5b66\u3001\u7f16\u7a0b\u548c\u79d1\u5b66\u63a8\u7406\uff09\uff1b<\/p>\n<p>\u7f16\u7a0b\u4e0e Agent \uff08\u524d\u7aef\u548c\u540e\u7aef\u5de5\u7a0b\u4ee3\u7801\u3001\u5de5\u5177\u8c03\u7528\u3001Coding Agent \u3001\u641c\u7d22 Agent \u548c\u901a\u7528 Agent \uff09\u3002<\/p>\n<p>\u6b64\u5916\uff0cGLM-5 \u5728 SFT \u9636\u6bb5\u5c06\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u81f3 202752 \u4e2a token\u3002\u914d\u5408\u66f4\u65b0\u540e\u7684\u804a\u5929\u6a21\u677f\uff0c\u6a21\u578b\u652f\u6301\u4e09\u79cd\u4e0d\u540c\u7684\u601d\u8003\u7279\u5f81\uff08\u56fe 7\uff09\uff1a<\/p>\n<p>\u4ea4\u9519\u601d\u8003\uff08Interleaved Thinking\uff09\uff1a\u6a21\u578b\u5728\u6bcf\u6b21\u54cd\u5e94\u548c\u5de5\u5177\u8c03\u7528\u524d\u8fdb\u884c\u601d\u8003\uff0c\u63d0\u5347\u6307\u4ee4\u9075\u5faa\u548c\u751f\u6210\u8d28\u91cf\uff1b<\/p>\n<p>\u4fdd\u7559\u601d\u8003\uff08reserved Thinking\uff09\uff1a\u5728 Coding Agent \u573a\u666f\u4e2d\uff0c\u6a21\u578b\u81ea\u52a8\u5728\u591a\u8f6e\u5bf9\u8bdd\u4e2d\u4fdd\u7559\u6240\u6709\u601d\u8003\u5757\uff0c\u590d\u7528\u5df2\u6709\u63a8\u7406\u800c\u975e\u91cd\u65b0\u63a8\u5bfc\uff0c\u4ece\u800c\u51cf\u5c11\u4fe1\u606f\u4e22\u5931\u548c\u4e0d\u4e00\u81f4\u6027\uff0c\u9002\u5408\u957f\u7a0b\u590d\u6742\u4efb\u52a1\uff1b<\/p>\n<p>\u8f6e\u7ea7\u601d\u8003\uff08Turn-level Thinking\uff09\uff1a\u6a21\u578b\u652f\u6301\u5728\u4f1a\u8bdd\u4e2d\u5bf9\u6bcf\u8f6e\u63a8\u7406\u8fdb\u884c\u7cbe\u7ec6\u63a7\u5236\uff0c\u9762\u5bf9\u8f7b\u91cf\u7ea7\u8bf7\u6c42\u65f6\u53ef\u7981\u7528\u601d\u8003\uff0c\u4ece\u800c\u964d\u4f4e\u5ef6\u8fdf\/\u6210\u672c\uff0c\u9762\u5bf9\u590d\u6742\u4efb\u52a1\u65f6\u53ef\u542f\u7528\u601d\u8003\uff0c\u4ece\u800c\u63d0\u5347\u7cbe\u5ea6\u548c\u7a33\u5b9a\u6027\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/5E6577DB93D64FCDEA6F85F143E7AE587F5FEF81_size269_w1080_h576.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe7\uff1a\u201c\u4ea4\u9519\u601d\u8003\u201d\u4e0e\u201c\u4fdd\u7559\u601d\u8003\u201d\u793a\u610f\u56fe\u3002<\/p>\n<p>\u901a\u8fc7\u5728\u52a8\u4f5c\u95f4\u8fdb\u884c\u601d\u8003\u5e76\u5728\u8f6e\u6b21\u95f4\u4fdd\u6301\u4e00\u81f4\u6027\uff0cGLM-5 \u5728\u590d\u6742\u4efb\u52a1\u4e0a\u5b9e\u73b0\u4e86\u66f4\u7a33\u5b9a\u3001\u53ef\u63a7\u7684\u884c\u4e3a\u3002<\/p>\n<p>\u5728\u901a\u7528\u5bf9\u8bdd\u65b9\u9762\uff0c\u76f8\u6bd4 GLM-4.5\uff0c\u6211\u4eec\u4f18\u5316\u4e86\u54cd\u5e94\u98ce\u683c\uff0c\u4f7f GLM-5 \u66f4\u52a0\u903b\u8f91\u6e05\u6670\u3001\u7b80\u6d01\u7cbe\u7ec3\uff1b\u5bf9\u4e8e\u89d2\u8272\u626e\u6f14\u4efb\u52a1\uff0c\u6211\u4eec\u6536\u96c6\u5e76\u6784\u5efa\u4e86\u8986\u76d6\u591a\u79cd\u8bed\u8a00\u548c\u591a\u79cd\u89d2\u8272\u8bbe\u5b9a\u7684\u66f4\u5e7f\u6cdb\u3001\u66f4\u591a\u6837\u7684\u6570\u636e\u96c6\u3002\u7279\u522b\u662f\uff0c\u6211\u4eec\u5b9a\u4e49\u4e86\u82e5\u5e72\u8bc4\u4f30\u7ef4\u5ea6\uff0c\u5305\u62ec\u6307\u4ee4\u9075\u5faa\u3001\u8bed\u8a00\u8868\u73b0\u529b\u3001\u521b\u9020\u529b\u3001\u903b\u8f91\u8fde\u8d2f\u6027\u548c\u957f\u5bf9\u8bdd\u4e00\u81f4\u6027\uff0c\u5e76\u7ed3\u5408\u81ea\u52a8\u5316\u4e0e\u4eba\u5de5\u7b5b\u9009\u5bf9\u6570\u636e\u8fdb\u884c\u6574\u7406\u548c\u7cbe\u70bc\u3002<\/p>\n<p>\u5728\u63a8\u7406\u4efb\u52a1\u65b9\u9762\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86\u6a21\u578b\u63a8\u7406\u6df1\u5ea6\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5bf9\u4e8e\u903b\u8f91\u63a8\u7406\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u53ef\u9a8c\u8bc1\u7684\u95ee\u9898\uff0c\u5e76\u901a\u8fc7\u62d2\u7edd\u91c7\u6837\u5408\u6210\u9ad8\u8d28\u91cf\u6570\u636e\uff1b\u5bf9\u4e8e\u6570\u5b66\u548c\u79d1\u5b66\u95ee\u9898\uff0c\u6211\u4eec\u91c7\u7528\u57fa\u4e8e\u96be\u5ea6\u7684\u8fc7\u6ee4\u6d41\u7a0b\uff0c\u4ec5\u4fdd\u7559\u5bf9 GLM-4.7 \u5177\u6709\u6311\u6218\u6027\u7684\u95ee\u9898\u3002<\/p>\n<p>\u5bf9\u4e8e\u7f16\u7a0b\u548c Agent \u4efb\u52a1\uff0c\u4e0e GLM-4.5 \u76f8\u6bd4\uff0c\u6211\u4eec\u4e3a GLM-5 \u6784\u5efa\u4e86\u5927\u91cf\u6267\u884c\u73af\u5883\uff0c\u4ece\u800c\u83b7\u53d6\u9ad8\u8d28\u91cf\u8f68\u8ff9\uff0c\u5c24\u5176\u6ce8\u91cd\u771f\u5b9e\u4e16\u754c\u573a\u666f\u548c\u957f\u7a0b\u4efb\u52a1\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u5229\u7528\u4e13\u5bb6 RL \u548c\u62d2\u7edd\u91c7\u6837\u6765\u6539\u8fdb SFT \u6570\u636e\u3002\u8f68\u8ff9\u4e2d\u7684\u9519\u8bef\u7247\u6bb5\u88ab\u4fdd\u7559\uff0c\u4f46\u5728\u635f\u5931\u51fd\u6570\u4e2d\u88ab\u63a9\u7801\u5904\u7406\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u7ea0\u9519\u884c\u4e3a\uff0c\u800c\u4e0d\u4f1a\u5f3a\u5316\u9519\u8bef\u52a8\u4f5c\u3002<\/p>\n<p>3.2 Reasoning RL<\/p>\n<p>RL \u7b97\u6cd5\u3002\u6211\u4eec\u7684 RL \u7b97\u6cd5\u4ee5 GRPO \u4e3a\u57fa\u7840\uff0c\u5e76\u5f15\u5165 IcePop \u6280\u672f\u6765\u7f13\u89e3\u8bad\u7ec3-\u63a8\u7406\u4e0d\u5339\u914d\u95ee\u9898\uff0c\u5373 RL \u4f18\u5316\u8fc7\u7a0b\u4e2d\u63a8\u7406\u5206\u5e03\u4e0e\u8bad\u7ec3\u5206\u5e03\u4e4b\u95f4\u7684\u5dee\u5f02\u3002\u6211\u4eec\u660e\u786e\u533a\u5206\u4e86\u7528\u4e8e\u68af\u5ea6\u66f4\u65b0\u7684\u8bad\u7ec3\u6a21\u578b\u548c\u7528\u4e8e\u8f68\u8ff9\u91c7\u6837\u7684 \u63a8\u7406\u6a21\u578b\u3002\u4e0e\u539f\u59cb IcePop \u516c\u5f0f\u76f8\u6bd4\uff0c\u6211\u4eec\u53bb\u9664\u4e86 KL \u6b63\u5219\u9879\uff0c\u6765\u52a0\u901f RL \u7684\u6539\u8fdb\u3002\u8bad\u7ec3\u5b8c\u5168\u4ee5\u540c\u7b56\u7565\u8fdb\u884c\uff0cgroup \u5927\u5c0f\u4e3a 32\uff0cbatch \u5927\u5c0f\u4e3a 32\u3002<\/p>\n<p>DSA RL \u8981\u70b9\u3002\u6211\u4eec\u5728 DSA \u67b6\u6784\u7684\u6a21\u578b\u4e0a\u8fdb\u884c\u4e86\u5927\u89c4\u6a21 RL \u8bad\u7ec3\u3002\u4e0e MLA \u76f8\u6bd4\uff0cDSA \u5f15\u5165\u4e86\u4e00\u4e2a\u989d\u5916\u7684Indexer\uff0c\u7528\u4e8e\u68c0\u7d22\u524d k \u4e2a\u6700\u76f8\u5173\u7684 KV \u6761\u76ee\uff0c\u5e76\u5bf9\u68c0\u7d22\u5230\u7684\u5b50\u96c6\u8fdb\u884c\u7a00\u758f\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u68c0\u7d22\u5230\u7684\u524d k \u4e2a\u7ed3\u679c\u5bf9 RL \u7a33\u5b9a\u6027\u81f3\u5173\u91cd\u8981\u3002\u8fd9\u7c7b\u4f3c\u4e8e MoE \u6a21\u578b\u4f7f\u7528\u8def\u7531\u56de\u653e\u6765\u4fdd\u7559\u6fc0\u6d3b\u7684\u524d k \u4e2a\u4e13\u5bb6\uff0c\u4ee5\u786e\u4fdd\u8bad\u7ec3\u4e0e\u63a8\u7406\u7684\u4e00\u81f4\u6027\u3002\u7136\u800c\uff0c\u5728\u6bcf\u4e2a token \u4f4d\u7f6e\u5b58\u50a8\u7d22\u5f15\u5668\u7684\u524d k \u4e2a\u7d22\u5f15\u663e\u7136\u662f\u4e0d\u5207\u5b9e\u9645\u7684\uff0c\u56e0\u4e3a\u7d22\u5f15\u5668\u4f7f\u7528\u7684 k = 2048 \u8fdc\u5927\u4e8e MoE \u4e2d\u901a\u5e38\u4f7f\u7528\u7684 k \u503c\uff0c\u5b58\u50a8\u6240\u6709\u8fd9\u4e9b\u7d22\u5f15\u5c06\u5e26\u6765\u5de8\u5927\u7684\u5b58\u50a8\u5f00\u9500\uff0c\u4ee5\u53ca\u8bad\u7ec3\u5f15\u64ce\u4e0e\u63a8\u7406\u5f15\u64ce\u4e4b\u95f4\u663e\u8457\u7684\u901a\u4fe1\u5f00\u9500\u3002<\/p>\n<p>\u6211\u4eec\u53d1\u73b0\uff0c\u91c7\u7528\u786e\u5b9a\u6027 top-k \u7b97\u5b50\u80fd\u6709\u6548\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u4e0e SGLang \u7684 DSA Indexer \u4e2d\u4f7f\u7528\u7684\u57fa\u4e8e CUDA \u7684\u975e\u786e\u5b9a\u6027 top-k \u5b9e\u73b0\u76f8\u6bd4\uff0c\u76f4\u63a5\u4f7f\u7528\u539f\u751f\u7684 torch.topk \u901f\u5ea6\u7a0d\u6162\uff0c\u4f46\u5177\u6709\u786e\u5b9a\u6027\u3002\u5b83\u80fd\u4ea7\u751f\u66f4\u4e00\u81f4\u7684\u8f93\u51fa\uff0c\u5e76\u5e26\u6765\u663e\u8457\u7684 RL \u589e\u76ca\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5176\u4ed6\u975e\u786e\u5b9a\u6027 top-k \u7b97\u5b50\uff08\u5982 CUDA \u6216 TileLang \u5b9e\u73b0\uff09\u5728 RL \u4ec5\u51e0\u6b65\u540e\u5c31\u5bfc\u81f4\u6027\u80fd\u6025\u5267\u4e0b\u964d\uff0c\u5e76\u4f34\u968f\u71b5\u503c\u9aa4\u964d\u3002\u56e0\u6b64\uff0c\u5728\u6211\u4eec\u6240\u6709\u7684 RL \u9636\u6bb5\u4e2d\uff0c\u6211\u4eec\u5728\u8bad\u7ec3\u5f15\u64ce\u7684 DSA Indexer \u4e2d\u9ed8\u8ba4\u4f7f\u7528 torch.topk \u4f5c\u4e3a top-k \u7b97\u5b50\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5728 RL \u8fc7\u7a0b\u4e2d\u9ed8\u8ba4\u51bb\u7ed3\u7d22\u5f15\u5668\u53c2\u6570\uff0c\u4ece\u800c\u52a0\u901f\u8bad\u7ec3\u5e76\u9632\u6b62\u7d22\u5f15\u5668\u51fa\u73b0\u4e0d\u7a33\u5b9a\u7684\u5b66\u4e60\u884c\u4e3a\u3002<\/p>\n<p>\u591a\u9886\u57df\u6df7\u5408 RL\u3002\u5728\u63a8\u7406 RL \u9636\u6bb5\uff0c\u6211\u4eec\u5728\u56db\u4e2a\u9886\u57df\u4e0a\u8fdb\u884c\u6df7\u5408 RL \u8bad\u7ec3\uff1a\u6570\u5b66\u3001\u79d1\u5b66\u3001\u4ee3\u7801\u548c\u5de5\u5177\u96c6\u6210\u63a8\u7406\uff08tool-integrated reasoning\uff0cTIR\uff09\u3002<\/p>\n<p>\u5bf9\u4e8e\u6570\u5b66\u548c\u79d1\u5b66\uff0c\u6211\u4eec\u4ece\u5f00\u6e90\u6570\u636e\u96c6\u548c\u4e0e\u5916\u90e8\u6807\u6ce8\u4f9b\u5e94\u5546\u5171\u540c\u6784\u5efa\u7684\u6570\u636e\u96c6\u4e2d\u6574\u7406\u6570\u636e\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u5e94\u7528\u96be\u5ea6\u8fc7\u6ee4\uff0c\u5c06\u8bad\u7ec3\u91cd\u70b9\u96c6\u4e2d\u5728 GLM-4.7 \u6781\u5c11\u7b54\u5bf9\u6216\u59cb\u7ec8\u65e0\u6cd5\u89e3\u51b3\u3001\u4f46\u8f83\u5f3a\u7684\u6a21\u578b\uff08\u5982 GPT-5.2 xhigh \u548c Gemini 3 Pro Preview\uff09\u53ef\u4ee5\u89e3\u51b3\u7684\u95ee\u9898\u4e0a\u3002<\/p>\n<p>\u5bf9\u4e8e\u4ee3\u7801\uff0c\u6211\u4eec\u540c\u65f6\u6db5\u76d6\u7ade\u6280\u7f16\u7a0b\u98ce\u683c\u4efb\u52a1\u548c\u79d1\u5b66\u7f16\u7a0b\u4efb\u52a1\u3002\u524d\u8005\u4e3b\u8981\u6765\u6e90\u4e8e Codeforces \u4ee5\u53ca TACO\u3001SYNTHETIC-2-RL \u7b49\u4ee3\u8868\u6027\u6570\u636e\u96c6\uff0c\u540e\u8005\u5219\u901a\u8fc7\u5c06\u95ee\u9898\u5206\u89e3\u4e3a\u6b63\u786e\u89e3\u7b54\u6240\u9700\u7684\u6700\u5c0f\u4ee3\u7801\u5b9e\u73b0\uff0c\u4ece\u5185\u90e8\u9898\u5e93\u4e2d\u6784\u5efa\u800c\u6210\u3002<\/p>\n<p>\u5bf9\u4e8e TIR\uff0c\u6211\u4eec\u590d\u7528\u6570\u5b66\u548c\u79d1\u5b66\u5f3a\u5316\u5b66\u4e60\u6570\u636e\u4e2d\u96be\u5ea6\u8f83\u9ad8\u7684\u5b50\u96c6\uff0c\u5e76\u4e0e\u6807\u6ce8\u4f9b\u5e94\u5546\u8054\u5408\u6784\u5efa\u660e\u786e\u8bbe\u8ba1\u4e3a\u9700\u501f\u52a9\u5916\u90e8\u5de5\u5177\u4f5c\u7b54\u7684 STEM \u95ee\u9898\u3002<\/p>\n<p>\u5728 RL \u8bad\u7ec3\u671f\u95f4\uff0c\u6211\u4eec\u4e3a\u5404\u9886\u57df\u548c\u7279\u5b9a\u6765\u6e90\u5206\u914d\u4e13\u5c5e\u88c1\u5224\u6a21\u578b\u6216\u8bc4\u4f30\u7cfb\u7edf\uff0c\u7528\u6765\u751f\u6210\u4e8c\u5143\u7ed3\u679c\u5956\u52b1\u3002\u6211\u4eec\u4fdd\u6301\u56db\u4e2a\u9886\u57df\u7684\u6574\u4f53\u6df7\u5408\u5927\u81f4\u4fdd\u6301\u5e73\u8861\uff0c\u5e76\u5728\u6df7\u5408 RL \u8bbe\u7f6e\u4e0b\u6301\u7eed\u89c2\u5bdf\u5230\u6bcf\u4e2a\u9886\u57df\u7684\u7a33\u5b9a\u4e14\u663e\u8457\u7684\u589e\u76ca\u3002<\/p>\n<p>3.3 Agentic RL<\/p>\n<p>\u4e3a\u63d0\u5347 GLM-5 \u7684 Agent \u6027\u80fd\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u5957\u5b8c\u5168\u5f02\u6b65\u4e14\u89e3\u8026\u7684 RL \u6846\u67b6\uff0c\u5e76\u5728\u7f16\u7a0b\u548c\u641c\u7d22 Agent \u4efb\u52a1\u4e0a\u4f18\u5316\u4e86 GLM-5\u3002\u539f\u751f\u7684\u540c\u6b65 RL \u5728\u957f\u7a0b Agent rollout \u65f6\u4f1a\u5bfc\u81f4\u5927\u91cf\u7684 GPU \u65f6\u95f4\u6d6a\u8d39\u3002\u901a\u8fc7 Multi-Task Rollout Orchestrator \u5c06\u63a8\u7406\u5f15\u64ce\u548c\u8bad\u7ec3\u5f15\u64ce\u89e3\u8026\uff0c\u6211\u4eec\u5728\u591a\u4e2a\u7684 Agent \u4efb\u52a1\u4e2d\u5b9e\u73b0\u4e86\u9ad8\u541e\u5410\u91cf\u7684\u8054\u5408\u8bad\u7ec3\u3002<\/p>\n<p>\u4e3a\u5728\u5f02\u6b65\u79bb\u7b56\u7565\u6761\u4ef6\u4e0b\u4fdd\u6301\u8bad\u7ec3\u7a33\u5b9a\u6027\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e24\u9879\u5173\u952e\u7b97\u6cd5\u7b56\u7565\u3002\u9996\u5148\uff0cToken-in-Token-out\uff08TITO\uff09\u4ee3\u66ffText-in-Text-out\uff0c\u901a\u8fc7\u4fdd\u7559\u7cbe\u786e\u7684token\u7ea7\u522b\u5bf9\u5e94\u5173\u7cfb\uff0c\u6d88\u9664\u4e86 re-tokenization \u5e26\u6765\u7684\u4e0d\u5339\u914d\u95ee\u9898\u3002\u5176\u6b21\uff0c\u6211\u4eec\u91c7\u7528\u76f4\u63a5\u53cc\u4fa7\u91cd\u8981\u6027\u91c7\u6837\u65b9\u6cd5\uff0c\u5bf9 rollout \u6982\u7387\u5e94\u7528 token \u7ea7\u88c1\u526a\u673a\u5236\uff0c\u5728\u65e0\u9700\u4fdd\u7559\u591a\u4e2a\u5386\u53f2\u7b56\u7565\u6a21\u578b\u6743\u91cd\u7684\u60c5\u51b5\u4e0b\u6709\u6548\u63a7\u5236\u5f53\u524d\u8bad\u7ec3\u6a21\u578b\u4e0e\u91c7\u6837\u7b56\u7565\u504f\u5dee\u3002\u6211\u4eec\u8fd8\u91c7\u7528\u4e86 DP-aware \u8def\u7531\uff0c\u4ee5\u5728\u5927\u89c4\u6a21 MoE \u6a21\u578b\u7684\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u8fc7\u7a0b\u4e2d\u6700\u5927\u5316 KV \u7f13\u5b58\u590d\u7528\uff0c\u4ece\u800c\u52a0\u901f\u63a8\u7406\u3002\u4e3a\u4e86\u6269\u5c55 Agent \u73af\u5883\uff0c\u6211\u4eec\u5728\u4e09\u4e2a\u9886\u57df\u4e2d\u6269\u5c55\u4e86\u53ef\u9a8c\u8bc1\u7684\u8bad\u7ec3\u73af\u5883\uff1a\u8d85\u8fc7\u4e00\u4e07\u4e2a\u771f\u5b9e\u4e16\u754c\u8f6f\u4ef6\u5de5\u7a0b\uff08SWE\uff09\u4efb\u52a1\u3001\u7ec8\u7aef\u4efb\u52a1\uff0c\u4ee5\u53ca\u9ad8\u96be\u5ea6\u591a\u8df3\u641c\u7d22\u4efb\u52a1\u3002\u66f4\u591a Agent RL \u7ec6\u8282\u8bf7\u53c2\u89c1\u7b2c4\u8282\u3002<\/p>\n<p>3.4 General RL<\/p>\n<p>\u591a\u7ef4\u4f18\u5316\u76ee\u6807\u3002\u6211\u4eec\u5c06\u901a\u7528\u5f3a\u5316\u5b66\u4e60\uff08General RL\uff09\u7684\u4f18\u5316\u76ee\u6807\u5206\u89e3\u4e3a\u4e09\u4e2a\u7ef4\u5ea6\uff1a\u6b63\u786e\u6027\u3001\u60c5\u5546\u3001\u7279\u5b9a\u4efb\u52a1\u80fd\u529b\u3002<\/p>\n<p>\u6b63\u786e\u6027\u7ef4\u5ea6\u662f\u56de\u590d\u8d28\u91cf\u7684\u57fa\u77f3\u3002\u5b83\u9488\u5bf9\u5e7f\u6cdb\u5b58\u5728\u7684\u5404\u79cd\u9519\u8bef\u7c7b\u578b\uff0c\u4f8b\u5982\u6307\u4ee4\u4e0d\u9075\u5faa\u3001\u903b\u8f91\u77db\u76fe\u3001\u4e8b\u5b9e\u4e0d\u51c6\u786e\u3001\u77e5\u8bc6\u5e7b\u89c9\u3001\u8bed\u8a00\u4e0d\u6d41\u7545\u7b49\u9519\u8bef\u7c7b\u578b\uff0c\u8fd9\u4e9b\u9519\u8bef\u4f1a\u7834\u574f\u56de\u590d\u7684\u53ef\u7528\u6027\u3002\u6b63\u786e\u6027\u7ef4\u5ea6\u7684\u4f18\u5316\u76ee\u6807\u662f\u964d\u4f4e\u9519\u8bef\u7387\uff0c\u4f7f\u56de\u590d\u8d28\u91cf\u8fbe\u5230\u53ef\u7528\u7684\u57fa\u51c6\u7ebf\u3002\u6211\u4eec\u5c06\u5176\u89c6\u4e3a\u5176\u4ed6\u4f18\u5316\u7684\u5148\u51b3\u6761\u4ef6\uff1a\u5982\u679c\u56de\u590d\u5305\u542b\u9519\u8bef\uff08\u4f8b\u5982\u4e8b\u5b9e\u9519\u8bef\u6216\u8bef\u89e3\u4e86\u7528\u6237\u7684\u610f\u56fe\uff09\uff0c\u65e0\u8bba\u5b83\u8868\u9762\u4e0a\u770b\u8d77\u6765\u591a\u4e48\u5b8c\u7f8e\uff0c\u90fd\u53ef\u80fd\u5bf9\u7528\u6237\u9020\u6210\u5b9e\u8d28\u6027\u7684\u8bef\u5bfc\u3002<\/p>\n<p>\u60c5\u5546\u7ef4\u5ea6\u5728\u6b63\u786e\u6027\u4e4b\u4e0a\u8fdb\u4e00\u6b65\u4f18\u5316\u7528\u6237\u4f53\u9a8c\u3002\u5b83\u65e8\u5728\u751f\u6210\u5177\u6709\u540c\u7406\u5fc3\u3001\u5bcc\u6709\u6d1e\u5bdf\u529b\u4e14\u5728\u98ce\u683c\u4e0a\u63a5\u8fd1\u81ea\u7136\u4eba\u7c7b\u4ea4\u6d41\u7684\u56de\u590d\uff0c\u4ece\u800c\u4f7f\u4e0e\u6a21\u578b\u7684\u4ea4\u4e92\u66f4\u52a0\u81ea\u7136\u4e14\u5f15\u4eba\u5165\u80dc\u3002<\/p>\n<p>\u7279\u5b9a\u4efb\u52a1\u80fd\u529b\u7ef4\u5ea6\u9488\u5bf9\u5404\u79cd\u5177\u4f53\u4efb\u52a1\u8fdb\u884c\u7ec6\u7c92\u5ea6\u7684\u4f18\u5316\u3002\u5728\u6b63\u786e\u6027\u7684\u57fa\u7840\u4e0a\uff0c\u5b83\u65e8\u5728\u5c06\u5404\u4efb\u52a1\u7c7b\u522b\u4e2d\u7684\u56de\u590d\u4ece\u201c\u4ec5\u4ec5\u662f\u6b63\u786e\u7684\u201d\u63d0\u5347\u5230\u201c\u771f\u6b63\u9ad8\u8d28\u91cf\u7684\u201d\u3002\u8be5\u7ef4\u5ea6\u6db5\u76d6\u4e86\u5e7f\u6cdb\u7684\u4efb\u52a1\u9886\u57df\uff0c\u5305\u62ec\u5199\u4f5c\u3001\u6587\u672c\u5904\u7406\u3001\u4e3b\u89c2\u4e0e\u5ba2\u89c2\u95ee\u7b54\u3001\u89d2\u8272\u626e\u6f14\u3001\u7ffb\u8bd1\u7b49\u5404\u79cd\u9886\u57df\u3002\u6bcf\u4e2a\u4efb\u52a1\u9886\u57df\u90fd\u9700\u8981\u4e0d\u540c\u7684\u5956\u52b1\u4fe1\u53f7\uff0c\u56e0\u6b64\u6211\u4eec\u91c7\u7528\u4e86\u6df7\u5408\u5956\u52b1\u7cfb\u7edf\u3002<\/p>\n<p>\u6df7\u5408\u5956\u52b1\u7cfb\u7edf\u3002\u4e3a\u4e86\u7ed9\u4e0a\u8ff0\u591a\u7ef4\u4f18\u5316\u76ee\u6807\u63d0\u4f9b\u76d1\u7763\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u96c6\u6210\u4e09\u79cd\u4e92\u8865\u5956\u52b1\u4fe1\u53f7\u7684\u6df7\u5408\u5956\u52b1\u7cfb\u7edf\uff1a\u57fa\u4e8e\u89c4\u5219\u7684\u5956\u52b1\u51fd\u6570\u3001\u5224\u522b\u5f0f\u5956\u52b1\u6a21\u578b\uff08ORM\uff09\u548c\u751f\u6210\u5f0f\u5956\u52b1\u6a21\u578b\uff08GRM\uff09\u3002\u89c4\u5219\u5956\u52b1\u51fd\u6570\u53ef\u4ee5\u63d0\u4f9b\u51c6\u786e\u7684\u3001\u53ef\u89e3\u91ca\u7684\u4fe1\u53f7\uff0c\u4f46\u9002\u7528\u8303\u56f4\u53d7\u9650\uff1bORM \u63d0\u4f9b\u4f4e\u65b9\u5dee\u5956\u52b1\u4fe1\u53f7\u548c\u9ad8\u8bad\u7ec3\u6548\u7387\uff0c\u4f46\u53ef\u80fd\u9677\u5165\u5956\u52b1\u6b3a\u9a97\uff08reward hacking\uff09\uff0c\u5373\u7b56\u7565\u6a21\u578b\uff08policy model\uff09\u5229\u7528\u865a\u5047\u7684\u9ad8\u5206\u7279\u5f81\u6765\u9a97\u53d6\u5956\u52b1\uff0c\u800c\u975e\u771f\u6b63\u63d0\u5347\u6838\u5fc3\u80fd\u529b\uff1bGRM \u5229\u7528\u8bed\u8a00\u6a21\u578b\u751f\u6210\u8bc4\u4ef7\u4e0e\u5206\u6570\uff0c\u5bf9\u6b64\u7c7b\u653b\u51fb\u66f4\u5177\u9c81\u68d2\u6027\uff0c\u4f46\u65b9\u5dee\u8f83\u5927\u3002\u4e09\u8005\u7ec4\u5408\u5f97\u5230\u7684\u5956\u52b1\u7cfb\u7edf\u5e73\u8861\u4e86\u51c6\u786e\u6027\u3001\u8bad\u7ec3\u6548\u7387\u548c\u9c81\u68d2\u6027\uff0c\u7f13\u89e3\u4e86\u5355\u4e00\u7c7b\u578b\u5956\u52b1\u4fe1\u53f7\u7684\u7f3a\u70b9\u3002<\/p>\n<p>\u4eba\u673a\u534f\u4f5c\u98ce\u683c\u5bf9\u9f50\u3002\u5728\u901a\u7528\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\uff0c\u6211\u4eec\u6ca1\u6709\u4ec5\u4ec5\u4f9d\u8d56\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\uff0c\u800c\u662f\u5f15\u5165\u4e86\u7531\u4eba\u7c7b\u64b0\u5199\u7684\u9ad8\u8d28\u91cf\u56de\u590d\uff0c\u5c06\u5176\u4f5c\u4e3a\u98ce\u683c\u548c\u8d28\u91cf\u7684\u5bf9\u9f50\u951a\u70b9\u3002\u8fd9\u79cd\u505a\u6cd5\u7684\u52a8\u673a\u6e90\u4e8e\u6211\u4eec\u7684\u4e00\u79cd\u89c2\u5bdf\uff1a\u7eaf\u7cb9\u4f9d\u8d56\u6a21\u578b\u56de\u590d\u8fdb\u884c\u7684\u5f3a\u5316\u5b66\u4e60\uff0c\u5f80\u5f80\u4f1a\u6536\u655b\u4e8e\u90a3\u4e9b\u6781\u5177\u8fa8\u8bc6\u5ea6\u7684\u201c\u673a\u5668\u611f\u201d\u6a21\u5f0f\u2014\u2014\u8fd9\u4e9b\u6a21\u5f0f\u901a\u5e38\u663e\u5f97\u5197\u957f\u3001\u516c\u5f0f\u5316\uff0c\u6216\u8005\u7f3a\u4e4f\u4eba\u7c7b\u5199\u4f5c\u4e2d\u6240\u5e26\u6709\u7684\u7ec6\u817b\u611f\u3002\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8ba9\u6a21\u578b\u5b66\u4e60\u4eba\u7c7b\u7684\u56de\u590d\u8303\u5f0f\uff0c\u6211\u4eec\u4fc3\u4f7f\u6a21\u578b\u91c7\u7528\u66f4\u52a0\u81ea\u7136\u3001\u66f4\u7b26\u5408\u4eba\u7c7b\u4e60\u60ef\u7684\u56de\u590d\u6a21\u5f0f\u3002<\/p>\n<p>3.5 \u8de8\u9636\u6bb5\u5728\u7ebf\u84b8\u998f<\/p>\n<p>\u5728\u591a\u9636\u6bb5 RL \u6d41\u7a0b\u4e2d\uff0c\u5bf9\u4e0d\u540c\u76ee\u6807\u7684\u987a\u5e8f\u4f18\u5316\u53ef\u80fd\u5bfc\u81f4\u5148\u524d\u83b7\u5f97\u80fd\u529b\u7684\u7d2f\u79ef\u9000\u5316\u3002\u4e3a\u7f13\u89e3\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u5c06\u8de8\u9636\u6bb5\u5728\u7ebf\u84b8\u998f\u4f5c\u4e3a\u6700\u7ec8\u9636\u6bb5\uff0c\u91c7\u7528\u5728\u7ebf\u84b8\u998f\u7b97\u6cd5\u5feb\u901f\u6062\u590d\u65e9\u671f SFT \u548c RL \u9636\u6bb5\uff08\u63a8\u7406 RL \u548c\u901a\u7528 RL\uff09\u83b7\u5f97\u7684\u80fd\u529b\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6765\u81ea\u524d\u5e8f\u8bad\u7ec3\u9636\u6bb5\u7684\u6700\u7ec8\u68c0\u67e5\u70b9\u4f5c\u4e3a\u6559\u5e08\u6a21\u578b\uff0c\u8bad\u7ec3 Prompts \u4ece\u76f8\u5e94\u6559\u5e08\u7684 RL \u8bad\u7ec3\u96c6\u4e2d\u91c7\u6837\u5e76\u4ee5\u9002\u5f53\u6bd4\u4f8b\u6df7\u5408\u3002\u76ee\u524d\uff0c\u6211\u4eec\u5229\u7528\u63a8\u7406\u5f15\u64ce\u83b7\u53d6\u6559\u5e08\u7684 logits\uff0c\u6279\u5927\u5c0f\u8bbe\u7f6e\u4e3a 1024\uff0c\u7ec4\u5927\u5c0f\u914d\u7f6e\u4e3a 1 \u4ee5\u63d0\u9ad8\u6570\u636e\u541e\u5410\u91cf\u2014\u2014\u5728\u6b64\u9636\u6bb5\u4e0d\u518d\u9700\u8981\u6bcf\u4e2a\u63d0\u793a\u7ef4\u6301\u5927\u91cf\u6837\u672c\u6765\u4f30\u8ba1 Advantage\uff0cAdvantage \u76f4\u63a5\u4ece\u4e0e\u6559\u5e08\u6a21\u578b\u7684\u5dee\u8ddd\u4e2d\u8ba1\u7b97\u5f97\u51fa\u3002<\/p>\n<p>3.6 RL \u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\uff1aslime \u6846\u67b6<\/p>\n<p>\u6211\u4eec\u7ee7\u7eed\u4f7f\u7528 slime \u4f5c\u4e3a GLM-5 \u7edf\u4e00\u7684\u540e\u8bad\u7ec3\u8bad\u7ec3\u6846\u67b6\uff0c\u5b9e\u73b0\u5927\u89c4\u6a21\u7aef\u5230\u7aef RL\u3002GLM-5 \u5145\u5206\u5229\u7528\u4e86 slime \u7684\u80fd\u529b\uff1a(1) \u901a\u8fc7\u7075\u6d3b\u7684 rollout \u5b9a\u5236\u5316\u63a5\u53e3\u548c\u670d\u52a1\u5316\uff08server-based\uff09\u7684\u63a8\u7406\u65b9\u6848\u6269\u5927\u8bad\u7ec3\u4efb\u52a1\u7684\u8986\u76d6\u9762\uff1b(2) \u901a\u8fc7\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3 \/ rollout \u4ee5\u53ca\u591a token \u9884\u6d4b\uff08MTP\uff09\u548c\u9884\u586b\u5145-\u89e3\u7801\uff08PD\uff09\u5206\u79bb\u5927\u5e45\u964d\u4f4e\u5ef6\u8fdf\uff0c\u7279\u522b\u662f\u5728\u591a\u8f6e RL \u5de5\u4f5c\u8d1f\u8f7d\u4e0b\u7684\u7aef\u5230\u7aef\u5ef6\u8fdf\uff1b(3) \u4f7f\u7528\u57fa\u4e8e\u5fc3\u8df3\u7684\u5bb9\u707e\u548c\u751f\u547d\u5468\u671f\u7ba1\u7406\u63d0\u5347\u7a33\u5b9a\u6027\u3002<\/p>\n<p>3.6.1 \u6a2a\u5411\u6269\u5c55\uff1a\u901a\u8fc7\u9ad8\u5ea6\u53ef\u5b9a\u5236 rollout \u5b9e\u73b0\u7075\u6d3b\u8bad\u7ec3<\/p>\n<p>GLM-5 \u540e\u8bad\u7ec3\u7684\u8bad\u7ec3\u76ee\u6807\u591a\u6837\u3002\u4e3a\u4e86\u590d\u7528\u540c\u4e00\u5957\u7cfb\u7edf\u4ee5\u652f\u6301\u8fd9\u4e9b\u9700\u6c42\uff0cGLM-5 \u5229\u7528\u4e86 slime \u9ad8\u5ea6\u53ef\u5b9a\u5236\u7684 rollout \u63a5\u53e3\u548c\u670d\u52a1\u5316\u7684\u63a8\u7406\u65b9\u6848\u3002slime \u7684\u7075\u6d3b\u63a5\u53e3\u8ba9\u6211\u4eec\u65e0\u9700\u4fee\u6539\u5c31\u80fd\u652f\u6301\u591a\u8f6e\u4ea4\u4e92\u3001\u5de5\u5177\u8c03\u7528\u3001\u590d\u6742\u5916\u90e8\u73af\u5883\u63a5\u5165\u7b49\u8bad\u7ec3\u9700\u6c42\u3002\u53e6\u5916\uff0cslime \u901a\u8fc7\u6807\u51c6 HTTP API \u66b4\u9732\u63a8\u7406\u670d\u52a1\u548c\u8def\u7531\uff0c\u5141\u8bb8\u7528\u6237\u50cf\u4f7f\u7528\u5e38\u89c4\u63a8\u7406\u5f15\u64ce\u4e00\u6837\u4e0e slime \u4ea4\u4e92\uff0c\u5b9e\u73b0\u63a8\u7406\u7684\u670d\u52a1\u5316\uff0c\u5c06\u63a8\u7406\u903b\u8f91\u4e0e\u8bad\u7ec3\u903b\u8f91\u89e3\u8026\u3002<\/p>\n<p>3.6.2 \u7eb5\u5411\u6269\u5c55\uff1aRL rollout \u7684\u957f\u5c3e\u5ef6\u8fdf\u4f18\u5316<\/p>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cRL \u63a8\u7406\u7684\u4f18\u5316\u76ee\u6807\u4e0d\u662f\u541e\u5410\u91cf\uff0c\u800c\u662f\u7531\u6bcf\u6b65\u4e2d\u6700\u6162\u7684\uff08\u957f\u5c3e\uff09\u6837\u672c\u4e3b\u5bfc\u7684\u7aef\u5230\u7aef\u5ef6\u8fdf\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u5355\u4e2a rollout \u8f68\u8ff9\u53ef\u80fd\u4f1a\u5ef6\u540e\u6574\u4f53\u8fdb\u7a0b\uff0c\u4ece\u800c\u76f4\u63a5\u62d6\u6162\u8bad\u7ec3\u65f6\u95f4\u3002\u56e0\u6b64\uff0cGLM-5 \u5728 slime \u4e2d\u8fdb\u884c\u4e86\u5ef6\u8fdf\u5bfc\u5411\u7684\u6027\u80fd\u4f18\u5316\u548c\u8c03\u5ea6\uff1a\u91c7\u7528\u591a\u8282\u70b9\u63a8\u7406\u90e8\u7f72\uff08\u5982\u8de8 8 \u8282\u70b9\u7684 EP64 \u914d\u5408 DP64\uff09\u4ee5\u63d0\u4f9b\u5145\u8db3\u7684 KV \u7f13\u5b58\uff1b\u4f7f\u7528 FP8 \u7cbe\u5ea6\u8fdb\u884c\u63a8\u7406\u4ee5\u964d\u4f4e\u6bcf\u4e2a token \u7684\u89e3\u7801\u5ef6\u8fdf\uff1b\u5229\u7528 MTP \u6765\u5728 RL \u4e2d\u5178\u578b\u7684\u5c0f\u6279\u6b21\u89e3\u7801\u573a\u666f\u4e0b\u83b7\u5f97\u663e\u8457\u6027\u80fd\u6536\u76ca\uff1b\u901a\u8fc7 PD \u5206\u79bb\u786e\u4fdd\u5728\u591a\u8f6e\u4ea4\u4e92\u7684\u8bad\u7ec3\u4efb\u52a1\u4e2d\u89e3\u7801\u901f\u5ea6\u7684\u7a33\u5b9a\uff0c\u4ece\u800c\u663e\u8457\u6539\u5584\u957f\u7a0b Agent RL \u8bad\u7ec3\u4e2d\u7684\u957f\u5c3e\u95ee\u9898\uff0c\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n<p>3.6.3 Rollout \u7a33\u5b9a\u6027\uff1a\u57fa\u4e8e\u5fc3\u8df3\u7684\u5bb9\u707e<\/p>\n<p>\u5728\u5927\u89c4\u6a21\u573a\u666f\u4e0b\uff0c\u77ac\u6001\u6545\u969c\uff08\u5982\u5355\u4e2a\u670d\u52a1\u5668\u5d29\u6e83\u3001\u7f51\u7edc\u95ee\u9898\u6216\u6027\u80fd\u4e0b\u964d\uff09\u4e0d\u53ef\u907f\u514d\u3002GLM-5 \u5229\u7528 slime \u7684\u5bb9\u707e\u673a\u5236\u6765\u786e\u4fdd\u5728\u6b64\u7c7b\u4e8b\u4ef6\u4e0b\u8bad\u7ec3\u7684\u8fde\u7eed\u6027\uff1aslime \u4f1a\u5b9a\u671f\u53d1\u51fa\u5fc3\u8df3\u4ee5\u53ca\u65f6\u8bc6\u522b\u5e76\u7ec8\u6b62\u4e0d\u5065\u5eb7\u7684\u63a8\u7406\u670d\u52a1\uff0c\u5e76\u4ece\u8def\u7531\u4e2d\u6ce8\u9500\u5bf9\u5e94\u63a5\u53e3\uff0c\u4ece\u800c\u5728\u8bf7\u6c42\u91cd\u8bd5\u65f6\u81ea\u52a8\u8def\u7531\u5230\u5065\u5eb7\u7684\u63a8\u7406\u670d\u52a1\uff0c\u9632\u6b62\u4e2a\u522b\u6545\u969c\u5f71\u54cd\u7aef\u5230\u7aef RL \u8bad\u7ec3\uff0c\u63d0\u5347\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/p>\n<p>4 \u667a\u80fd\u4f53\u5de5\u7a0b Agentic Engineering<\/p>\n<p>\u5728\u8fd9\u4e00\u8282\u4e2d\uff0c\u6211\u4eec\u63cf\u8ff0\u4ece\u6c1b\u56f4\u7f16\u7a0b\uff08vibe coding\uff09\u5230\u667a\u80fd\u4f53\u5de5\u7a0b\uff08agentic engineering\uff09\u7684\u8f6c\u53d8\u3002\u5728\u6c1b\u56f4\u7f16\u7a0b\u4e2d\uff0c\u4eba\u7c7b\u63d0\u793a AI \u6a21\u578b\u7f16\u5199\u4ee3\u7801\uff1b\u800c\u5728\u667a\u80fd\u4f53\u5de5\u7a0b\u4e2d\uff0cAI Agent \u81ea\u884c\u7f16\u5199\u4ee3\u7801\u2014\u2014\u5b83\u4eec\u8fdb\u884c\u89c4\u5212\u3001\u5b9e\u73b0\u548c\u8fed\u4ee3\u3002\u4e3a\u652f\u6301\u8fd9\u4e9b\u957f\u7a0b\u4efb\u52a1\uff0cGLM-5 \u5229\u7528\u5b8c\u5168\u5f02\u6b65\u4e14\u89e3\u8026\u7684 RL \u6846\u67b6\uff0c\u901a\u8fc7\u51cf\u5c11 Agent rollout \u671f\u95f4\u7684\u7a7a\u95f2\u65f6\u95f4\u663e\u8457\u63d0\u5347 GPU \u5229\u7528\u7387\u3002\u4e3a\u6269\u5c55 Agent \u73af\u5883\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u73af\u5883\u6784\u5efa\u6d41\u7a0b\uff1a\u5bf9\u4e8e\u7f16\u7a0b Agent\uff0c\u901a\u8fc7\u521b\u5efa\u8d85\u8fc7 \u4e0a\u4e07\u4e2a\u53ef\u9a8c\u8bc1\u8bad\u7ec3\u573a\u666f\u6765\u642d\u5efa\u771f\u5b9e\u4e16\u754c\u8f6f\u4ef6\u5de5\u7a0b\u95ee\u9898\u548c\u7ec8\u7aef\u4efb\u52a1\uff1b\u5bf9\u4e8e\u641c\u7d22 Agent \uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u81ea\u52a8\u5316\u4e14\u53ef\u6269\u5c55\u7684\u590d\u6742\u591a\u6b65\u63a8\u7406\u6570\u636e\u5408\u6210\u6d41\u7a0b\u3002<\/p>\n<p>4.1 \u5f02\u6b65 Agent RL<\/p>\n<p>\u4e3a\u5bf9 Agent \u4efb\u52a1\u8fdb\u884c RL\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u5b8c\u5168\u5f02\u6b65\u4e14\u89e3\u8026\u7684 RL \u57fa\u7840\u8bad\u7ec3\u6846\u67b6\uff0c\u80fd\u591f\u9ad8\u6548\u5904\u7406\u957f\u7a0b Agent rollout\uff0c\u5e76\u652f\u6301\u8de8\u591a\u6837\u5316 Agent \u6846\u67b6\u7684\u7075\u6d3b\u591a\u4efb\u52a1 RL \u8bad\u7ec3\u3002\u6211\u4eec\u91c7\u7528\u5206\u7ec4\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\u8fdb\u884c RL \u8bad\u7ec3\uff0c\u4ec5\u5bf9\u6a21\u578b\u751f\u6210\u7684 token \u8fdb\u884c\u4f18\u5316\uff0c\u5ffd\u7565\u73af\u5883\u53cd\u9988\u90e8\u5206\u7684\u635f\u5931\u8ba1\u7b97\u3002<\/p>\n<p>4.1.1 \u5f02\u6b65 Agent RL \u8bad\u7ec3\u8bbe\u8ba1<\/p>\n<p>\u7531\u4e8e rollout \u8fc7\u7a0b\u7684\u957f\u5c3e\u7279\u6027\uff0c\u539f\u751f\u540c\u6b65 RL \u8bad\u7ec3\u5728 rollout \u9636\u6bb5\u4f1a\u5f15\u5165\u5927\u91cf\u201c\u6c14\u6ce1\u201d\uff0c\u56e0\u4e3a Agent \u4efb\u52a1\u7684\u751f\u6210\u4e25\u91cd\u4e0d\u5747\u8861\uff0c\u5bfc\u81f4\u5927\u91cf GPU \u7a7a\u95f2\u65f6\u95f4\u3002\u4e3a\u63d0\u5347\u8bad\u7ec3\u541e\u5410\u91cf\uff0c\u6211\u4eec\u5bf9 Agent RL \u91c7\u7528\u5b8c\u5168\u5f02\u6b65\u8bad\u7ec3\u8303\u5f0f\u4ee5\u63d0\u5347 GPU \u5229\u7528\u7387\u548c\u8bad\u7ec3\u6548\u7387\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u5c06\u8bad\u7ec3\u5f15\u64ce\u548c\u63a8\u7406\u5f15\u64ce\u89e3\u8026\u5230\u4e0d\u540c GPU \u8bbe\u5907\u4e0a\u3002\u63a8\u7406\u5f15\u64ce\u6301\u7eed\u751f\u6210\u8f68\u8ff9\uff0c\u4e00\u65e6\u751f\u6210\u8f68\u8ff9\u6570\u91cf\u8fbe\u5230\u9884\u5b9a\u9608\u503c\uff0c\u8be5\u6279\u6b21\u5c31\u88ab\u53d1\u9001\u5230\u8bad\u7ec3\u5f15\u64ce\u7528\u4e8e\u6a21\u578b\u8bad\u7ec3\u3002\u4e3a\u51cf\u5c11\u7b56\u7565\u6ede\u540e\u5e76\u4fdd\u6301\u8bad\u7ec3\u7684\u8fd1\u4f3c\u540c\u7b56\u7565\u6027\uff0c\u63a8\u7406\u6a21\u578b\u4f7f\u7528\u7684\u6a21\u578b\u6743\u91cd\u4f1a\u5b9a\u671f\u4e0e\u8bad\u7ec3\u4fa7\u540c\u6b65\u3002<\/p>\n<p>\u57fa\u4e8e\u670d\u52a1\u7684\u591a\u4efb\u52a1\u8bad\u7ec3\u8bbe\u8ba1\u3002\u4e3a\u89e3\u51b3\u591a\u4efb\u52a1 RL \u4e2d\u8f68\u8ff9\u751f\u6210\u7684\u5f02\u6784\u6027\uff08\u4e0d\u540c\u4efb\u52a1\u901a\u5e38\u4f9d\u8d56\u4e0d\u540c\u7684\u5de5\u5177\u96c6\u548c\u4efb\u52a1\u7279\u5b9a\u7684 rollout \u811a\u624b\u67b6\uff09\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u57fa\u4e8e\u670d\u52a1\u5668\u7684 Multi-Task Rollout Orchestrator\u3002\u8be5\u7ec4\u4ef6\u88ab\u8bbe\u8ba1\u901a\u8fc7\u5177\u6709\u591a\u4e2a\u6ce8\u518c\u4efb\u52a1\u670d\u52a1\u7684\u4e2d\u592e\u7f16\u6392\u5668\uff0c\u786e\u4fdd slime RL \u8bad\u7ec3\u6846\u67b6\u4e0e\u591a\u6837\u5316\u4e0b\u6e38\u4efb\u52a1\u4e4b\u95f4\u7684\u65e0\u7f1d\u517c\u5bb9\u6027\u3002\u6bcf\u4e2a\u4efb\u52a1\u4ee5\u72ec\u7acb\u5fae\u670d\u52a1\u5f62\u5f0f\u5b9e\u73b0\u5176 rollout \u548c\u5956\u52b1\u903b\u8f91\uff0c\u6ce8\u518c\u5230\u4e2d\u592e\u7f16\u6392\u5668\u8fdb\u884c\u7ba1\u7406\u548c\u8c03\u5ea6\u3002\u8be5\u7f16\u6392\u5668\u652f\u6301\u8d85\u8fc7 \u4e0a\u5343\u5e76\u53d1 rollout\uff0c\u5b9e\u73b0\u4e86\u4efb\u52a1\u91c7\u6837\u6bd4\u4f8b\u7684\u81ea\u52a8\u5316\u52a8\u6001\u8c03\u6574\u548c\u4efb\u52a1\u8fdb\u5ea6\u7684\u7ec6\u7c92\u5ea6\u76d1\u63a7\u3002<\/p>\n<p>4.1.2 \u4f18\u5316\u5f02\u6b65\u8bad\u7ec3\u7a33\u5b9a\u6027<\/p>\n<p>Token-in-Token-out \u5bf9\u6bd4 Text-in-Text-out\u3002\u5728 RL rollout \u8bbe\u7f6e\u4e2d\uff0cToken-in-Token-out\uff08TITO\uff09\u610f\u5473\u7740\u8bad\u7ec3\u6d41\u7a0b\u76f4\u63a5\u6d88\u8d39\u63a8\u7406\u5f15\u64ce\u751f\u6210\u7684\u7cbe\u786e tokenization \u548c\u89e3\u7801 token \u6d41\u6765\u6784\u5efa\u5b66\u4e60\u8f68\u8ff9\uff1b\u76f8\u6bd4\u4e4b\u4e0b\uff0cText-in-Text-out \u5c06 rollout \u5f15\u64ce\u89c6\u4e3a\u8fd4\u56de\u6700\u7ec8\u6587\u672c\u7684\u9ed1\u7bb1\uff0c\u8bad\u7ec3\u5668\u901a\u8fc7 re-tokenizing \u91cd\u5efa\u8f68\u8ff9\u3002\u8fd9\u770b\u4f3c\u5fae\u5c0f\u7684\u9009\u62e9\u5b9e\u9645\u4e0a\u5f71\u54cd\u5f88\u5927\uff1are-tokenizing \u53ef\u80fd\u5728 token \u8fb9\u754c\u3001\u7a7a\u767d\/\u5f52\u4e00\u5316\u5904\u7406\u3001\u622a\u65ad\u6216\u7279\u6b8a token \u653e\u7f6e\u4e0a\u5f15\u5165\u7ec6\u5fae\u4e0d\u5339\u914d\uff0c\u4ece\u800c\u5f71\u54cd\u5bf9\u5355\u4e2a token \u91c7\u6837\u6982\u7387\u7684\u4f30\u8ba1\u3002\u6211\u4eec\u5b9e\u73b0\u4e86\u4e00\u4e2a TITO \u4e2d\u95f4\u4ef6\uff0c\u62e6\u622a rollout \u4efb\u52a1\u7684\u6240\u6709\u751f\u6210\u8bf7\u6c42\u5e76\u8bb0\u5f55\u6bcf\u4e2a\u8f68\u8ff9\u7684 token ID \u548c\u5143\u6570\u636e\uff0c\u5c06\u7e41\u7410\u7684 token ID \u5904\u7406\u4ece\u4e0b\u6e38 Agent rollout \u903b\u8f91\u4e2d\u9694\u79bb\u51fa\u6765\u3002<\/p>\n<p>\u7528\u4e8e token \u88c1\u526a\u7684\u76f4\u63a5\u53cc\u4fa7\u91cd\u8981\u6027\u91c7\u6837\u3002\u4e0e\u7b2c3\u8282\u7684\u540c\u6b65 RL \u8bad\u7ec3\u8bbe\u7f6e\u4e0d\u540c\uff0c\u5728\u5f02\u6b65\u8bbe\u7f6e\u4e2d\uff0crollout \u5f15\u64ce\u53ef\u80fd\u5728\u5355\u4e2a\u8f68\u8ff9\u751f\u6210\u8fc7\u7a0b\u4e2d\u7ecf\u5386\u591a\u6b21\u66f4\u65b0\uff0c\u8fd9\u4f7f\u5f97\u8ffd\u8e2a\u5386\u53f2\u8bad\u7ec3\u4fa7\u6a21\u578b\u7684\u7cbe\u786e\u884c\u4e3a\u6982\u7387\u5728\u8ba1\u7b97\u4e0a\u4e0d\u53ef\u884c\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u9996\u5148\u91c7\u7528\u7b80\u5316\u7684 token \u7ea7\u91cd\u8981\u6027\u91c7\u6837\u673a\u5236\uff0c\u5c06 rollout \u671f\u95f4\u751f\u6210\u7684\u5bf9\u6570\u6982\u7387\u4f5c\u4e3a\u76f4\u63a5\u884c\u4e3a\u4ee3\u7406\uff0c\u901a\u8fc7\u8ba1\u7b97$ r_t(\u03b8) = \u03c0_\u03b8\/\u03c0_{rollout} $\u5e76\u4e22\u5f03\u4f20\u7edf\u7684 $\u03c0_{\u03b8_{old}}$ \u6765\u6d88\u9664\u5355\u72ec\u65e7\u7b56\u7565\u63a8\u7406\u7684\u8ba1\u7b97\u5f00\u9500\u3002\u5176\u6b21\uff0c\u6211\u4eec\u91c7\u7528\u53cc\u4fa7\u6821\u51c6 token \u7ea7\u63a9\u7801\u7b56\u7565\uff0c\u5c06\u4fe1\u4efb\u57df\u9650\u5236\u5728$ [1-\u03b5_l, 1+\u03b5_h]$\uff0c\u5bf9\u843d\u5728\u6b64\u533a\u95f4\u4e4b\u5916\u7684 token \u5b8c\u5168\u5c4f\u853d\u68af\u5ea6\u8ba1\u7b97\u3002\u5b9e\u9a8c\u4e2d\u6211\u4eec\u53d1\u73b0\uff0c\u590d\u7528\u63a8\u51fa\u5bf9\u6570\u6982\u7387\u63a5\u53d7\u53d7\u63a7\u7a0b\u5ea6\u7684\u79bb\u7b56\u7565\u504f\u5dee\uff0c\u4ee5\u7ed5\u8fc7\u5386\u53f2\u7b56\u7565\u8ffd\u8e2a\u9700\u6c42\u540c\u65f6\u63d0\u5347\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u4e22\u5f03\u79bb\u7ebf\u7a0b\u5ea6\u9ad8\u7684\u6837\u672c\u548c\u566a\u58f0\u6837\u672c\u3002\u5728\u5f02\u6b65 RL \u4e2d\uff0c\u8fc7\u957f\u7684\u8f68\u8ff9\u53ef\u80fd\u53d8\u5f97\u9ad8\u5ea6\u79bb\u7b56\u7565\uff0c\u4ece\u800c\u7834\u574f\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002\u6211\u4eec\u8bb0\u5f55 rollout \u5f15\u64ce\u5728\u751f\u6210\u65f6\u4f7f\u7528\u7684\u7b56\u7565\u6743\u91cd\u7248\u672c\uff0c\u4e22\u5f03\u6700\u65e7 rollout \u7248\u672c\u8fc7\u65f6\u7684\u6837\u672c\u3002\u6b64\u5916\uff0cCoding Agent \u6c99\u7bb1\u53ef\u80fd\u56e0\u6a21\u578b\u65e0\u5173\u7684\u539f\u56e0\uff08\u5982\u73af\u5883\u5d29\u6e83\uff09\u800c\u4e0d\u7a33\u5b9a\uff0c\u5f15\u5165\u566a\u58f0\u8bad\u7ec3\u4fe1\u53f7\u3002\u6211\u4eec\u8bb0\u5f55\u6bcf\u4e2a\u6837\u672c\u7684\u5931\u8d25\u539f\u56e0\uff0c\u6392\u9664\u56e0\u73af\u5883\u5d29\u6e83\u800c\u5931\u8d25\u7684\u6837\u672c\uff0c\u5e76\u5728\u6709\u6548\u6837\u672c\u8d85\u8fc7\u7ec4\u5927\u5c0f\u4e00\u534a\u65f6\u901a\u8fc7\u91cd\u590d\u6709\u6548\u6837\u672c\u6765\u8865\u9f50\u4e0d\u5b8c\u6574\u7ec4\uff0c\u5426\u5219\u4e22\u5f03\u6574\u4e2a\u7ec4\u3002<\/p>\n<p>\u7528\u4e8e\u52a0\u901f\u7684 DP \u611f\u77e5\u8def\u7531\u3002\u6211\u4eec\u63d0\u51fa\u4e00\u79cd DP \u611f\u77e5\u8def\u7531\u673a\u5236\uff0c\u5728\u5927\u89c4\u6a21 MoE \u63a8\u7406\u7684\u6570\u636e\u5e76\u884c\u4e0b\u4fdd\u7559 KV \u7f13\u5b58\u5c40\u90e8\u6027\u3002\u5728\u591a\u8f6e Agent \u5de5\u4f5c\u8d1f\u8f7d\u4e2d\uff0c\u6765\u81ea\u76f8\u540c rollout \u7684\u987a\u5e8f\u8bf7\u6c42\u5171\u4eab\u76f8\u540c\u524d\u7f00\uff0c\u6211\u4eec\u901a\u8fc7\u4e00\u81f4\u6027\u54c8\u5e0c\u5c06\u6bcf\u4e2a rollout ID \u6620\u5c04\u5230\u56fa\u5b9a DP rank\uff0c\u5e76\u7ed3\u5408\u54c8\u5e0c\u7a7a\u95f4\u4e0a\u7684\u8f7b\u91cf\u7ea7\u52a8\u6001\u8d1f\u8f7d\u91cd\u65b0\u5e73\u8861\u6765\u9632\u6b62\u957f\u671f\u4e0d\u5747\u8861\u3002\u8fd9\u907f\u514d\u4e86\u5197\u4f59\u9884\u586b\u5145\u8ba1\u7b97\uff0c\u65e0\u9700\u8de8 DP rank \u7684 KV \u540c\u6b65\uff0c\u968f\u7740 rollout \u957f\u5ea6\u589e\u52a0\uff0c\u9884\u586b\u5145\u6210\u672c\u4ecd\u4e0e\u589e\u91cf token \u6210\u6b63\u6bd4\u3002<\/p>\n<p>4.2 Agent \u73af\u5883\u6269\u5c55<\/p>\n<p>\u4e3a\u652f\u6301\u8de8\u591a\u6837\u5316\u4efb\u52a1\u7684 Agent RL\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u53ef\u9a8c\u8bc1\u7684\u53ef\u6267\u884c\u73af\u5883\uff0c\u4e3a\u4ee3\u7801\u4e2d\u5fc3\u548c\u5185\u5bb9\u751f\u6210\u5de5\u4f5c\u6d41\u63d0\u4f9b\u6709\u6839\u636e\u7684\u53cd\u9988\u3002\u5bf9\u4e8e Agent \u7f16\u7a0b\u4efb\u52a1\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e24\u79cd\u6784\u5efa\u53ef\u9a8c\u8bc1\u53ef\u6267\u884c\u73af\u5883\u7684\u6d41\u7a0b\uff1a\u57fa\u4e8e\u771f\u5b9e\u4e16\u754c\u8f6f\u4ef6\u5de5\u7a0b\u95ee\u9898\u7684\u73af\u5883\u8bbe\u7f6e\u6d41\u7a0b\uff0c\u4ee5\u53ca\u7528\u4e8e\u7ec8\u7aef Agent \u73af\u5883\u7684\u5408\u6210\u6d41\u7a0b\u3002\u8d85\u8d8a\u7f16\u7a0b\u8303\u7574\uff0c\u6211\u4eec\u8fd8\u5f15\u5165\u4e86\u5e7b\u706f\u7247\u751f\u6210\u73af\u5883\uff0c\u5176\u4e2d Agent \u5728\u7ed3\u6784\u5316 HTML \u4e0a\u64cd\u4f5c\uff0c\u914d\u6709\u53ef\u6267\u884c\u6e32\u67d3\u548c\u57fa\u4e8e\u5e03\u5c40\u7684\u9a8c\u8bc1\u3002<\/p>\n<p>4.2.1 \u8f6f\u4ef6\u5de5\u7a0b\uff08SWE\uff09\u73af\u5883<\/p>\n<p>\u5728\u6784\u5efa\u53ef\u6267\u884c\u73af\u5883\u4e4b\u524d\uff0c\u6211\u4eec\u6536\u96c6\u4e86\u5927\u91cf\u771f\u5b9e\u4e16\u754c\u7684 Issue-PR \u5bf9\uff0c\u5e76\u5e94\u7528\u4e25\u683c\u7684\u57fa\u4e8e\u89c4\u5219\u548c LLM \u7684\u8fc7\u6ee4\uff0c\u786e\u4fdd\u83b7\u53d6\u771f\u5b9e\u9ad8\u8d28\u91cf\u7684 issue \u63cf\u8ff0\u3002\u6211\u4eec\u5c06\u8fd9\u4e9b\u5b9e\u4f8b\u5206\u7c7b\u4e3a\u4e0d\u540c\u4efb\u52a1\u7c7b\u578b\uff08\u9519\u8bef\u4fee\u590d\u3001\u529f\u80fd\u5b9e\u73b0\u3001\u91cd\u6784\u7b49\uff09\u5e76\u5305\u542b\u5fc5\u8981\u7684\u4efb\u52a1\u9700\u6c42\u3002\u6211\u4eec\u91c7\u7528\u57fa\u4e8e RepoLaunch \u6846\u67b6\u7684\u73af\u5883\u8bbe\u7f6e\u6d41\u7a0b\uff0c\u81ea\u52a8\u5206\u6790\u4ed3\u5e93\u7684\u5b89\u88c5\u548c\u4f9d\u8d56\u8bbe\u7f6e\u4ee5\u6784\u5efa\u53ef\u6267\u884c\u73af\u5883\u5e76\u751f\u6210\u6d4b\u8bd5\u547d\u4ee4\uff0c\u518d\u5229\u7528 LLM \u4ece\u6d4b\u8bd5\u8f93\u51fa\u751f\u6210\u8bed\u8a00\u611f\u77e5\u7684\u65e5\u5fd7\u89e3\u6790\u51fd\u6570\uff0c\u5b9e\u73b0\u5931\u8d25\u8f6c\u901a\u8fc7\uff08F2P\uff09\u548c\u901a\u8fc7\u8f6c\u901a\u8fc7\uff08P2P\uff09\u6d4b\u8bd5\u7528\u4f8b\u7684\u63d0\u53d6\u3002\u4f7f\u7528\u8be5\u6d41\u7a0b\uff0c\u6211\u4eec\u8de8\u6570\u5343\u4e2a\u4ed3\u5e93\u3001\u6db5\u76d6 Python\u3001Java\u3001Go\u3001C\u3001C++\u3001JavaScript\u3001TypeScript\u3001PHP \u548c Ruby 9 \u79cd\u7f16\u7a0b\u8bed\u8a00\uff0c\u6784\u5efa\u4e86\u8d85\u8fc7 10000 \u4e2a\u53ef\u9a8c\u8bc1\u73af\u5883\u3002<\/p>\n<p>4.2.2 \u7ec8\u7aef\u73af\u5883<\/p>\n<p>\u4ece\u79cd\u5b50\u6570\u636e\u5408\u6210\u3002\u4e3a\u5927\u89c4\u6a21\u6784\u5efa\u53ef\u9a8c\u8bc1\u7684\u7ec8\u7aef Agent \u73af\u5883\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5305\u542b\u4e09\u4e2a\u9636\u6bb5\u7684 Agent \u6570\u636e\u5408\u6210\u6d41\u7a0b\uff1a\u4efb\u52a1\u8349\u7a3f\u751f\u6210\u3001\u5177\u4f53\u4efb\u52a1\u5b9e\u73b0\u548c\u8fed\u4ee3\u4efb\u52a1\u4f18\u5316\u3002\u4ece\u771f\u5b9e\u4e16\u754c\u8f6f\u4ef6\u5de5\u7a0b\u548c\u57fa\u4e8e\u7ec8\u7aef\u7684 computer-use \u573a\u666f\u6536\u96c6\u7684\u79cd\u5b50\u4efb\u52a1\u51fa\u53d1\uff0c\u5229\u7528 LLM \u5934\u8111\u98ce\u66b4\u5e76\u751f\u6210\u5927\u91cf\u53ef\u9a8c\u8bc1\u7684\u7ec8\u7aef\u4efb\u52a1\u8349\u7a3f\u3002\u8fd9\u4e9b\u8349\u7a3f\u7531\u6784\u5efa Agent \u5728 Harbor \u683c\u5f0f\u4e2d\u5b9e\u4f8b\u5316\u4e3a\u5177\u4f53\u4efb\u52a1\uff0c\u5305\u62ec\u7ed3\u6784\u5316\u4efb\u52a1\u63cf\u8ff0\u3001Docker \u5316\u6267\u884c\u73af\u5883\u548c\u76f8\u5e94\u7684\u6d4b\u8bd5\u811a\u672c\u3002\u968f\u540e\uff0c\u7cbe\u70bc Agent \u6839\u636e\u624b\u52a8\u5b9a\u4e49\u7684\u8bc4\u5206\u6807\u51c6\u68c0\u67e5\u5e76\u8fed\u4ee3\u4f18\u5316\u751f\u6210\u7684\u4efb\u52a1\u3002\u6574\u4f53\u6d41\u7a0b\u4ea7\u751f\u6570\u5343\u4e2a\u591a\u6837\u5316\u4e14\u53ef\u9a8c\u8bc1\u7684\u7ec8\u7aef Agent \u73af\u5883\uff0cDocker \u6784\u5efa\u7cbe\u5ea6\u8d85\u8fc7 90%\u3002<\/p>\n<p>\u4ece\u7f51\u9875\u8bed\u6599\u5e93\u5408\u6210\u3002\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u53ef\u6269\u5c55\u7684\u81ea\u52a8\u5316\u6d41\u7a0b\uff0c\u57fa\u4e8e\u7f51\u9875\u8bed\u6599\u5e93\u6784\u5efa\u4e86 LLM \u9a8c\u8bc1\u7684\u3001\u57fa\u4e8e\u7ec8\u7aef\u7684\u7f16\u7a0b\u4efb\u52a1\uff0c\u91c7\u7528\u95ed\u73af\u8bbe\u8ba1\uff0c\u5176\u4e2d\u6784\u5efa Agent \u540c\u65f6\u5145\u5f53\u5176\u7b2c\u4e00\u8f6e\u8bc4\u4f30\u8005\u3002\u6211\u4eec\u6536\u96c6\u5927\u89c4\u6a21\u4ee3\u7801\u76f8\u5173\u7f51\u9875\u5e76\u5e94\u7528\u6570\u636e\u8d28\u91cf\u5206\u7c7b\u5668\u8fdb\u884c\u8fc7\u6ee4\uff0c\u7136\u540e\u5bf9\u6613\u4e8e\u7ec8\u7aef\u5f0f\u4efb\u52a1\u8868\u8ff0\u7684\u7f51\u9875\u8fdb\u884c\u5206\u5c42\u91c7\u6837\u3002\u63a5\u7740\uff0c\u6211\u4eec\u63d0\u793a Coding Agent \u5408\u6210\u5b8c\u6574\u7684\u7ec8\u7aef\u4efb\u52a1\uff0c\u5e76\u6267\u884c Harbor \u9a8c\u8bc1\u811a\u672c\u5bf9\u5176\u8f93\u51fa\u8fdb\u884c\u9a8c\u8bc1\uff0c\u5931\u8d25\u65f6\u8fed\u4ee3\u8bca\u65ad\u548c\u4fee\u8ba2\u4efb\u52a1\u76f4\u81f3\u901a\u8fc7\u6240\u6709\u81ea\u52a8\u5316\u68c0\u67e5\u3002\u53ea\u6709\u6210\u529f\u901a\u8fc7\u6b64\u81ea\u9a8c\u8bc1\u5faa\u73af\u7684\u4efb\u52a1\u624d\u88ab\u7eb3\u5165\u6700\u7ec8\u6570\u636e\u96c6\u3002<\/p>\n<p>4.2.3 \u641c\u7d22\u4efb\u52a1<\/p>\n<p>\u5bf9\u4e8e\u6df1\u5ea6\u641c\u7d22\u4fe1\u606f\u68c0\u7d22\u4efb\u52a1\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u6570\u636e\u5408\u6210\u6d41\u7a0b\uff0c\u751f\u6210\u5177\u6709\u6311\u6218\u6027\u7684\u591a\u8df3\u95ee\u7b54\u5bf9\u3002\u6bcf\u4e2a\u95ee\u9898\u9700\u8981\u57fa\u4e8e\u4ece\u591a\u4e2a\u7f51\u7edc\u6765\u6e90\u6c47\u805a\u7684\u8bc1\u636e\u8fdb\u884c\u591a\u6b65\u63a8\u7406\u3002\u4ece\u65e9\u671f\u641c\u7d22 Agent \u7684\u8f68\u8ff9\u51fa\u53d1\uff0c\u6211\u4eec\u6536\u96c6\u5e76\u53bb\u91cd\u6240\u6709\u8bbf\u95ee\u7684 URL\uff0c\u4fdd\u7559\u8986\u76d6\u591a\u6837\u9886\u57df\u7684\u8d85\u8fc7 200 \u4e07\u4e2a\u9ad8\u4fe1\u606f\u91cf\u7f51\u9875\uff0c\u901a\u8fc7 LLM \u5bf9\u5b9e\u4f53\u8fdb\u884c\u8bc6\u522b\u3001\u566a\u58f0\u8fc7\u6ee4\u548c\u7ed3\u6784\u5316\u4fe1\u606f\u63d0\u53d6\uff0c\u6301\u7eed\u66f4\u65b0 Web \u77e5\u8bc6\u56fe\u8c31\uff08WKG\uff09\u3002\u57fa\u4e8e WKG\uff0c\u6211\u4eec\u4ee5\u4f4e\u81f3\u4e2d\u9891\u5b9e\u4f53\u4e3a\u79cd\u5b50\u8282\u70b9\uff0c\u6269\u5c55\u5176\u591a\u8df3\u90bb\u57df\u5f62\u6210\u5b8c\u6574\u5b50\u56fe\uff0c\u7136\u540e\u5c06\u6bcf\u4e2a\u5b50\u56fe\u8f6c\u5316\u4e3a\u9690\u5f0f\u7f16\u7801\u591a\u5b9e\u4f53\u5173\u7cfb\u94fe\u7684\u95ee\u9898\u3002\u6211\u4eec\u5e94\u7528\u4e09\u9636\u6bb5\u6d41\u7a0b\u5e73\u8861\u96be\u5ea6\u4e0e\u6b63\u786e\u6027\uff1a\uff081\uff09\u5220\u9664\u4e0d\u4f7f\u7528\u5de5\u5177\u63a8\u7406\u6a21\u578b\u5728 8 \u6b21\u72ec\u7acb\u5c1d\u8bd5\u4e2d\u81f3\u5c11\u89e3\u5bf9\u4e00\u6b21\u7684\u95ee\u9898\uff1b\uff082\uff09\u8fc7\u6ee4\u53ef\u88ab\u65e9\u671f Agent \u7528\u57fa\u672c\u641c\u7d22\u5728\u51e0\u6b65\u5185\u89e3\u51b3\u7684\u95ee\u9898\uff1b\uff083\uff09\u8bbe\u8ba1 Verification Agent \u8fdb\u884c\u53cc\u5411\u9a8c\u8bc1\uff0c\u62d2\u7edd\u5177\u6709\u975e\u552f\u4e00\u7b54\u6848\u3001\u4e0d\u4e00\u81f4\u8bc1\u636e\u6216\u9519\u8bef\u6807\u7b7e\u7684\u6837\u672c\uff0c\u6700\u7ec8\u83b7\u5f97\u9ad8\u8d28\u91cf\u3001\u9ad8\u96be\u5ea6\u3001\u53ef\u9760\u7684\u591a\u8df3\u95ee\u7b54\u5bf9\u3002<\/p>\n<p>4.2.4 \u641c\u7d22 Agent \u7684\u4e0a\u4e0b\u6587\u7ba1\u7406\u63a8\u7406<\/p>\n<p>\u6211\u4eec\u53d1\u73b0 BrowseComp \u4e0a\u7684\u6027\u80fd\u5bf9\u88c1\u5224\u63d0\u793a\u548c\u88c1\u5224\u6a21\u578b\u5747\u654f\u611f\uff0c\u5f00\u6e90\u6a21\u578b\u4f5c\u4e3a\u88c1\u5224\u53ef\u80fd\u5f15\u5165\u7cfb\u7edf\u6027\u504f\u5dee\u3002\u4e3a\u786e\u4fdd\u4e00\u81f4\u6027\u548c\u53ef\u91cd\u73b0\u6027\uff0c\u6211\u4eec\u5224\u5b9a\u7b54\u6848\u6b63\u786e\u6027\u91c7\u7528\u5b98\u65b9 OpenAI \u8bc4\u4f30\u63d0\u793a\u548c\u95ed\u6e90\u6a21\u578b o3-mini \u4f5c\u4e3a\u7ec4\u4ef6\u3002\u5148\u524d\u5de5\u4f5c\u5f15\u5165\u4e86\u4e0a\u4e0b\u6587\u7ba1\u7406\uff0c\u5176\u4e2d Discard-all \u7b56\u7565\u901a\u8fc7\u5220\u9664\u6574\u4e2a\u5de5\u5177\u8c03\u7528\u5386\u53f2\u6765\u91cd\u7f6e\u4e0a\u4e0b\u6587\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u89c2\u5bdf\u5230\uff0c\u5728\u6781\u957f\u4e0a\u4e0b\u6587\uff08\u5982\u8d85\u8fc7 100K token\uff09\u4e0b\uff0c\u6a21\u578b\u6027\u80fd\u4f1a\u663e\u8457\u4e0b\u964d\u3002\u53d7\u6b64\u542f\u53d1\uff0c\u6211\u4eec\u91c7\u7528\u7b80\u5355\u7684\u201cKeep-recent-k\u201d\u7b56\u7565\uff1a\u5f53\u4ea4\u4e92\u5386\u53f2\u8d85\u8fc7\u9608\u503c k \u65f6\uff0c\u6700\u8fd1 k \u8f6e\u4e4b\u524d\u7684\u5de5\u5177\u5185\u5bb9\u5c06\u88ab\u6298\u53e0\u4ee5\u63a7\u5236\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u6211\u4eec\u8bbe\u5b9a k=5\uff0c\u4ece\u800c\u5c06 GLM-5 \u4ece 55.3%\uff08\u4e0d\u542b\u4fdd\u7559\u6700\u8fd1 k \u8f6e\uff09\u7a33\u5b9a\u63d0\u5347\u81f3 62.0%\uff08\u542b\u4fdd\u7559\u6700\u8fd1 k \u8f6e\uff09\u3002<\/p>\n<p>\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u5c06 Keep-recent-k \u4e0e Discard-all \u7ed3\u5408\uff0c\u5f62\u6210\u6df7\u5408\u5c42\u6b21\u4e0a\u4e0b\u6587\u7ba1\u7406\u7b56\u7565\uff1a\u5728\u4f7f\u7528 Keep-recent-k \u8fdb\u884c\u63a8\u7406\u65f6\uff0c\u5982\u679c\u603b\u4e0a\u4e0b\u6587\u957f\u5ea6\u8d85\u8fc7\u9608\u503c T\uff0c\u5219\u4e22\u5f03\u6574\u4e2a\u5de5\u5177\u8c03\u7528\u5386\u53f2\u5e76\u4ee5\u65b0\u4e0a\u4e0b\u6587\u91cd\u65b0\u5f00\u59cb\uff0c\u540c\u65f6\u7ee7\u7eed\u5e94\u7528 Keep-recent-k \u7b56\u7565\uff0c\u901a\u8fc7\u53c2\u6570\u641c\u7d22\u9009\u53d6 T=32K\u3002\u5982\u56fe 8 \u6240\u793a\uff0c\u5728\u4e0d\u540c\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0c\u8be5\u7b56\u7565\u6709\u6548\u91ca\u653e\u4e0a\u4e0b\u6587\u7a7a\u95f4\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u6267\u884c\u66f4\u591a\u6b65\u9aa4\u5e76\u6301\u7eed\u6539\u8fdb\u6027\u80fd\uff0c\u6700\u7ec8\u5f97\u5206\u8fbe\u5230 75.9\uff0c\u8d85\u8d8a\u6240\u6709\u914d\u5907\u4e0a\u4e0b\u6587\u7ba1\u7406\u7684\u5f00\u6e90\u6a21\u578b\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/54902B2FED8D20751BC9AC993A9E2FD78C2E4D33_size242_w1080_h670.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe8\uff1a\u4ece GLM-4.7\uff08\u7070\u8272\uff09\u5230 GLM-5\uff08\u5f69\u8272\uff09\u5728\u4e0d\u540c\u4e0a\u4e0b\u6587\u7ba1\u7406\u7b56\u7565\u4e0b BrowseComp \u7684\u51c6\u786e\u7387\u3002<\/p>\n<p>4.2.5 PPT\u751f\u6210<\/p>\n<p>\u6211\u4eec\u91c7\u7528\u81ea\u6539\u8fdb\u6d41\u7a0b\uff0c\u901a\u8fc7 RL \u548c\u62d2\u7edd\u91c7\u6837\u5fae\u8c03\u6765\u7cfb\u7edf\u6027\u63d0\u5347PPT\u751f\u6210\u6027\u80fd\u3002\u9996\u5148\u901a\u8fc7 SFT \u521d\u59cb\u5316\u6a21\u578b\u4ee5\u63d0\u4f9b\u57fa\u672c\u7684PPT\u751f\u6210\u80fd\u529b\uff0c\u7136\u540e\u4f7f\u7528\u57fa\u4e8e\u5e38\u89c1\u7f8e\u5b66\u548c\u7ed3\u6784\u5c5e\u6027\u7684\u591a\u7ea7\u5956\u52b1\u516c\u5f0f\u8fdb\u884c RL\uff0c\u8be5\u9636\u6bb5\u5e26\u6765\u663e\u8457\u7684\u751f\u6210\u8d28\u91cf\u63d0\u5347\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u8fdb\u884c\u62d2\u7edd\u91c7\u6837\u5fae\u8c03\u548c\u63a9\u7801\u5fae\u8c03\uff0c\u5c06 RL \u83b7\u5f97\u7684\u77e5\u8bc6\u6ce8\u56de\u8bad\u7ec3\u8bed\u6599\u5e93\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/DDAFF9993F648B7C8854DA8B70E29AC68C50E479_size95_w1080_h632.jpg\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe9\uff1a\u9488\u5bf9PPT\u751f\u6210\u7684 RL \u8bad\u7ec3\u4e2d\u7684\u5956\u52b1\u6b3a\u9a97\uff08reward hacking\uff09\u793a\u4f8b\u3002\u6211\u4eec\u7684\u8fd0\u884c\u65f6\u6e32\u67d3\u53ef\u83b7\u53d6\u66f4\u8d34\u8fd1\u5c55\u793a\u65f6\u7684\u5c5e\u6027\u503c\uff0c\u4ece\u800c\u4f7f\u8bc4\u4f30\u5bf9\u6b64\u7c7b\u6b3a\u9a97\u884c\u4e3a\u5177\u6709\u9c81\u68d2\u6027\u3002<\/p>\n<p>\u6211\u4eec\u63d0\u51fa\u4e00\u79cd\u591a\u5c42\u7ea7\u5956\u52b1\u673a\u5236\uff0c\u5c06\u57fa\u4e8e HTML \u7684PPT\u751f\u6210\u8fc7\u7a0b\u4e2d\u7684\u5956\u52b1\u4fe1\u53f7\u5212\u5206\u4e3a\u4e09\u4e2a\u5c42\u7ea7\uff1a<\/p>\n<p>\u7b2c1\u7ea7\uff08\u9759\u6001\u6807\u8bb0\u5c5e\u6027\uff09\uff1a\u5173\u6ce8\u751f\u6210 HTML \u4e2d\u7684\u58f0\u660e\u5f0f\u5c5e\u6027\uff0c\u5305\u62ec\u5b9a\u4f4d\u3001\u95f4\u8ddd\u3001\u989c\u8272\u3001\u5b57\u4f53\u6392\u7248\u3001\u9971\u548c\u5ea6\u7b49\u6837\u5f0f\u5c5e\u6027\uff0c\u57fa\u4e8e\u4e13\u4e1a\u8bbe\u8ba1\u539f\u5219\u8bbe\u8ba1\u89c4\u5219\u96c6\uff0c\u540c\u65f6\u5f15\u5165\u5e7b\u89c9\u56fe\u50cf\u548c\u91cd\u590d\u56fe\u50cf\u68c0\u6d4b\u673a\u5236\u3002<\/p>\n<p>\u7b2c2\u7ea7\uff08\u8fd0\u884c\u65f6\u6e32\u67d3\u5c5e\u6027\uff09\uff1a\u8bc4\u4f30\u6e32\u67d3\u671f\u95f4 DOM \u8282\u70b9\u7684\u8fd0\u884c\u65f6\u5c5e\u6027\uff0c\u5982\u5143\u7d20\u5bbd\u9ad8\u3001\u8fb9\u754c\u6846\u548c\u5176\u4ed6\u51e0\u4f55\u5e03\u5c40\u6307\u6807\uff0c\u7ea6\u675f\u8fd9\u4e9b\u5c5e\u6027\u4ee5\u9f13\u52b1\u751f\u6210\u7684PPT\u66f4\u8d34\u8fd1\u4eba\u7c7b\u5ba1\u7f8e\u504f\u597d\u7684\u7a7a\u95f4\u7ec4\u7ec7\u3002<\/p>\n<p>\u7b2c3\u7ea7\uff08\u89c6\u89c9\u611f\u77e5\u7279\u5f81\uff09\uff1a\u5728\u8fd0\u884c\u65f6\u6e32\u67d3\u7ea6\u675f\u4e4b\u5916\uff0c\u5bf9\u6e32\u67d3\u5e7b\u706f\u7247\u8fdb\u884c\u611f\u77e5\u5c42\u9762\u8bc4\u4f30\uff0c\u5982\u68c0\u6d4b\u5f02\u5e38\u7a7a\u767d\u6a21\u5f0f\u4f5c\u4e3a\u8f85\u52a9\u4fe1\u53f7\uff0c\u4ee5\u8fdb\u4e00\u6b65\u6539\u5584\u6574\u4f53\u6784\u56fe\u5e73\u8861\u548c\u89c6\u89c9\u7f8e\u5b66\u3002<\/p>\n<p>\u8bad\u7ec3\u7b56\u7565\u3002\u8fd9\u4e9b\u4fe1\u53f7\u5728RL\u8fc7\u7a0b\u4e2d\u7efc\u5408\u5229\u7528\uff0c\u4ee5\u63d0\u9ad8\u751f\u6210\u7684HTML\u7ed3\u6784\u7684\u6709\u6548\u6027\uff0c\u589e\u5f3a\u5e03\u5c40\u7ec4\u7ec7\uff0c\u5e76\u63d0\u5347\u6574\u4f53\u89c6\u89c9\u7f8e\u5b66\u8d28\u91cf\u3002\u9664\u4e86\u5956\u52b1\u8bbe\u8ba1\u5916\uff0c\u6211\u4eec\u8fd8\u901a\u8fc7\u52a8\u6001\u91c7\u6837\u91cd\u65b0\u8c03\u6574\u8bad\u7ec3\u5206\u5e03\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u4e00\u90e8\u5206\u7ed3\u6784\u7b80\u5355\u7684\u6837\u672c\u4f1a\u6309\u6982\u7387\u88ab\u4e22\u5f03\uff0c\u4ece\u800c\u4f7f\u4f18\u5316\u8fc7\u7a0b\u80fd\u591f\u96c6\u4e2d\u4e8e\u66f4\u5177\u6311\u6218\u6027\u7684\u9875\u9762\uff0c\u5e76\u5728\u590d\u6742\u7ec4\u5408\u573a\u666f\u4e0b\u63d0\u9ad8\u9c81\u68d2\u6027\u3002\u6211\u4eec\u8fd8\u91c7\u7528\u4e86\u57fa\u4e8e\u4ee4\u724c\u7ea7\u7684\u7b56\u7565\u68af\u5ea6\u635f\u5931\u6765\u7a33\u5b9a\u4f18\u5316\u8fc7\u7a0b\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u5e73\u8861\u7b56\u7565\uff0c\u5c06\u540c\u4e00\u6837\u672c\u7684\u4e0d\u540c\u7ed3\u679c\u5206\u5e03\u5230\u591a\u4e2a\u8bad\u7ec3\u6279\u6b21\u4e2d\uff0c\u4ece\u800c\u51cf\u5c11\u4f18\u5316\u504f\u5dee\u5e76\u63d0\u9ad8\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u62d2\u7edd\u91c7\u6837\u3002\u5728\u62d2\u7edd\u91c7\u6837\u9636\u6bb5\uff0c\u5f3a\u5316\u5b66\u4e60\u4e2d\u4f7f\u7528\u7684\u5956\u52b1\u51fd\u6570\u88ab\u8f6c\u6362\u4e3a\u6570\u636e\u8fc7\u6ee4\u6d41\u6c34\u7ebf\uff0c\u7528\u4e8e\u6784\u5efa\u9ad8\u8d28\u91cf\u7684\u8bad\u7ec3\u5b50\u96c6\u3002\u5728\u9875\u9762\u7ea7\u522b\uff0c\u8fc7\u6ee4\u6807\u51c6\u5305\u62ec\u4ee3\u7801\u7684\u6709\u6548\u6027\u548c\u53ef\u7f16\u8bd1\u6027\u3002\u5728\u8f68\u8ff9\u7ea7\u522b\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u5f3a\u5316\u5de5\u5177\u6267\u884c\u7684\u6b63\u786e\u6027\u4ee5\u53ca\u5168\u5c40\u5185\u5bb9\u591a\u6837\u6027\u7ea6\u675f\uff0c\u4ee5\u786e\u4fdd\u7ed3\u6784\u4e00\u81f4\u6027\u3002\u6211\u4eec\u91c7\u7528\u4e86\u4ece\u591a\u4e2a\u72ec\u7acb\u751f\u6210\u7684\u5019\u9009\u4e2d\u4fdd\u7559\u8d28\u91cf\u6700\u9ad8\u7684\u6837\u672c\u7684\u65b9\u5f0f\uff0c\u8fd9\u4e00\u673a\u5236\u6709\u6548\u5730\u91cd\u65b0\u52a0\u6743\u5206\u5e03\uff0c\u5411\u66f4\u9ad8\u8d28\u91cf\u7684\u5b9e\u4f8b\u503e\u659c\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u6837\u672c\u6548\u7387\u5e76\u589e\u5f3a\u4e86\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u63a9\u7801\u4fee\u6b63\u3002\u5c3d\u7ba1\u62d2\u7edd\u91c7\u6837\u53bb\u9664\u4e86\u5927\u90e8\u5206\u4f4e\u8d28\u91cf\u8f93\u51fa\uff0c\u4f46\u4ecd\u6709\u4e00\u4e9b\u8f68\u8ff9\u4ec5\u5728\u5c11\u6570\u9875\u9762\u4e2d\u5b58\u5728\u7f3a\u9677\u3002\u4e22\u5f03\u8fd9\u4e9b\u6837\u672c\u4f1a\u964d\u4f4e\u6709\u6548\u6570\u636e\u5229\u7528\u7387\u5e76\u589e\u52a0\u751f\u6210\u6210\u672c\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u57fa\u4e8e\u63a9\u7801\u7684\u4fee\u6b63\u673a\u5236\uff0c\u80fd\u591f\u81ea\u52a8\u8bc6\u522b\u6709\u7f3a\u9677\u7684\u9875\u9762\u5e76\u5e94\u7528\u63a9\u7801\uff0c\u540c\u65f6\u4fdd\u7559\u540c\u4e00\u8f68\u8ff9\u4e2d\u7684\u9ad8\u8d28\u91cf\u5185\u5bb9\u3002\u5982\u6b64\u4fdd\u7559\u4e86\u6709\u4ef7\u503c\u7684\u76d1\u7763\u4fe1\u53f7\uff0c\u63d0\u9ad8\u4e86\u6709\u6548\u6570\u636e\u6548\u7387\uff0c\u5e76\u51cf\u5c11\u4e86\u5197\u4f59\u7684\u91cd\u65b0\u751f\u6210\u5f00\u9500\uff0c\u4ece\u800c\u63d0\u5347\u4e86\u6574\u4f53\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n<p>\u6548\u679c\u63d0\u5347\u3002\u5728\u5185\u90e8\u8bc4\u6d4b\u96c6\u4e2d\uff0c\u751f\u6210\u7684\u9875\u9762\u4e2d\u4e25\u683c\u7b26\u5408 16:9 \u5bbd\u9ad8\u6bd4\u7684\u6bd4\u4f8b\u4ece 40% \u63d0\u5347\u81f3 92%\uff0c\u9875\u9762\u6ea2\u51fa\u60c5\u51b5\u663e\u8457\u51cf\u5c11\u3002\u4eba\u5de5\u8bc4\u4f30\u8fdb\u4e00\u6b65\u663e\u793a\uff0c\u4e0e GLM-4.5 \u76f8\u6bd4\uff0cGLM-5 \u5728\u5185\u5bb9\u8d28\u91cf\u4e0a\u7684\u80dc\u7387\u4e3a 60%\uff0c\u5e03\u5c40\u5408\u7406\u6027\u4e3a 57.5%\uff0c\u89c6\u89c9\u7f8e\u5b66\u4e3a 65%\uff0c\u7efc\u5408\u80dc\u7387\u8fbe\u5230 67.5%\u3002<\/p>\n<p>5 \u9002\u914d\u4e2d\u56fd\u82af\u7247\u57fa\u7840\u8bbe\u65bd<\/p>\n<p>\u786c\u4ef6\u751f\u6001\u7684\u5f02\u6784\u6027\u5f80\u5f80\u4f1a\u8ba9\u9ad8\u6027\u80fd\u90e8\u7f72\u53d8\u5f97\u5f02\u5e38\u590d\u6742\uff0c\u56e0\u6b64\uff0c\u5c06 GLM-5 \u9002\u914d\u81f3\u591a\u6837\u5316\u7684\u56fd\u4ea7\u82af\u7247\u57fa\u7840\u8bbe\u65bd\u662f\u4e00\u9879\u5de8\u5927\u6311\u6218\u3002\u5c3d\u7ba1\u56f0\u96be\u91cd\u91cd\uff0c\u6211\u4eec\u901a\u8fc7\u4e0e\u534e\u4e3a\u6607\u817e\u3001\u6469\u5c14\u7ebf\u7a0b\u3001\u6d77\u5149\u3001\u5bd2\u6b66\u7eaa\u3001\u6606\u4ed1\u82af\u3001\u6c90\u66e6\u4ee5\u53ca\u71e7\u539f\u7b49\u56fd\u5185\u4e3b\u6d41\u82af\u7247\u5e73\u53f0\u7684\u6df1\u5ea6\u5408\u4f5c\uff0c\u6210\u529f\u5b8c\u6210\u4e86 GLM-5 \u7684\u5168\u6808\u9002\u914d\u3002<\/p>\n<p>\u672c\u8282\u5c06\u4ee5\u534e\u4e3a\u6607\u817e Atlas \u7cfb\u5217\u4e3a\u4f8b\uff0c\u8be6\u7ec6\u89e3\u6790\u6211\u4eec\u7684\u9002\u914d\u65b9\u6848\uff0c\u91cd\u70b9\u805a\u7126\u4e09\u5927\u6838\u5fc3\u6280\u672f\u7ef4\u5ea6\uff1a\u6781\u81f4\u91cf\u5316\u3001\u9ad8\u6027\u80fd\u7b97\u5b50\u878d\u5408\u4ee5\u53ca\u5148\u8fdb\u7684\u63a8\u7406\u5f15\u64ce\u8c03\u5ea6\u3002<\/p>\n<p>W4A8 \u6df7\u5408\u7cbe\u5ea6\u91cf\u5316\uff1a\u4e3a\u4e86\u8ba9\u5355\u53f0 Atlas 800T A3 \u670d\u52a1\u5668\u80fd\u591f\u5bb9\u7eb3\u9ad8\u8fbe 750B \u53c2\u6570\u7684 GLM-5\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u7cbe\u5bc6\u7684\u6df7\u5408\u7cbe\u5ea6\u91cf\u5316\u7b56\u7565\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6807\u51c6\u7684 Attention \u548c MLP \u6a21\u5757\u91c7\u7528 W8A8\uff08INT8\uff09\u91cf\u5316\uff0c\u800c MoE \u4e13\u5bb6\u6a21\u5757\u5219\u88ab\u8fdb\u4e00\u6b65\u538b\u7f29\u81f3 W4A8\uff08INT4\uff09\u3002\u8fd9\u4e00\u8bbe\u8ba1\u5728\u4e0d\u663e\u8457\u727a\u7272\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\uff0c\u5927\u5e45\u524a\u51cf\u4e86\u663e\u5b58\u5360\u7528\u3002\u540c\u65f6\uff0c\u6211\u4eec\u5f15\u5165\u4e86 QuaRot \u5f02\u5e38\u503c\u6291\u5236\u4e0e Flex_AWQ_SSZ \u7f29\u653e\u6821\u51c6\u7b49\u5148\u8fdb\u7b97\u6cd5\uff0c\u4e3a\u4f4e\u7cbe\u5ea6\u90e8\u7f72\u7684\u7a33\u5b9a\u6027\u4fdd\u9a7e\u62a4\u822a\u3002<\/p>\n<p>\u9ad8\u6027\u80fd\u5b9a\u5236\u878d\u5408\u7b97\u5b50\uff1a\u9488\u5bf9\u6607\u817e NPU \u5728\u5904\u7406\u7a00\u758f\u6ce8\u610f\u529b\u65f6\u7684\u8ba1\u7b97\u74f6\u9888\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u5957\u4e13\u5c5e\u7684\u878d\u5408\u7b97\u5b50\u3002\u5176\u4e2d\uff0cLightning Indexer \u5c06\u5206\u6570\u8ba1\u7b97\u3001ReLU \u6fc0\u6d3b\u4e0e TopK \u805a\u5408\u4e3a\u5355\u4e00\u7b97\u5b50\uff0c\u5b9e\u73b0\u4e86\u8ba1\u7b97\u4e0e\u8bbf\u5b58\u7684\u6d41\u6c34\u7ebf\u91cd\u53e0\uff1bSparse Flash Attention \u9488\u5bf9 GLM-5 \u7684\u7a00\u758f\u6a21\u5f0f\u8fdb\u884c\u4e86\u4e13\u9879\u8c03\u4f18\uff0c\u5b9e\u73b0\u4e86 KV Cache TopK \u68c0\u7d22\u4e0e\u7a00\u758f\u6ce8\u610f\u529b\u8ba1\u7b97\u7684\u5e76\u884c\u6267\u884c\uff1b\u6b64\u5916\uff0cMLAPO\uff08\u591a\u5934\u6f5c\u53d8\u91cf\u6ce8\u610f\u529b\u9884\u5904\u7406\u4f18\u5316\uff09 \u521b\u9020\u6027\u5730\u5c06 13 \u4e2a\u788e\u7247\u5316\u7684\u9884\u5904\u7406\u7b97\u5b50\u878d\u5408\u6210\u4e00\u4e2a\u201c\u8d85\u7ea7\u7b97\u5b50\u201d\uff0c\u901a\u8fc7\u5145\u5206\u8c03\u7528\u5411\u91cf\u4e0e\u77e9\u9635\u5355\u5143\u7684\u5e76\u884c\u7b97\u529b\uff0c\u6781\u5927\u63d0\u5347\u4e86\u7aef\u5230\u7aef\u6548\u7387\u3002<\/p>\n<p>\u4e13\u9879\u63a8\u7406\u5f15\u64ce\u4f18\u5316\uff1a\u6211\u4eec\u6df1\u5ea6\u9002\u914d\u4e86 vLLM-Ascend \u4e0e SGLang \u4e24\u5927\u4e3b\u6d41\u63a8\u7406\u5f15\u64ce\u3002\u5728 vLLM \u4e2d\uff0c\u6211\u4eec\u5f15\u5165\u5f02\u6b65\u8c03\u5ea6\u673a\u5236\uff0c\u5c06 Device-to-Host (D2H) \u7684\u91c7\u6837\u6570\u636e\u56de\u4f20\u4e0e\u4e0b\u4e00\u6b65\u89e3\u7801\u7684\u51c6\u5907\u5de5\u4f5c\u5b8c\u7f8e\u5e76\u884c\uff0c\u6709\u6548\u6d88\u9664\u4e86\u8c03\u5ea6\u201c\u6c14\u6ce1\u201d\uff1b\u501f\u52a9 RadixCache\uff08\u524d\u7f00\u5171\u4eab\uff09\u548c\u524d\u7f00\u7f13\u5b58\uff08\u5c06 KV \u5b58\u50a8\u5378\u8f7d\u81f3\u7cfb\u7edf\u5185\u5b58\uff09\u6280\u672f\uff0c\u5b9e\u73b0\u4e86 KV Cache \u7684\u9ad8\u6548\u590d\u7528\uff0c\u8fd9\u6210\u4e3a\u7a81\u7834\u957f\u4e0a\u4e0b\u6587\u6027\u80fd\u7684\u5173\u952e\uff1b\u6211\u4eec\u8fd8\u91c7\u7528\u4e86\u6ce8\u610f\u529b\u6570\u636e\u5e76\u884c\uff08DP\uff09\u4e0e MoE \u4e13\u5bb6\u5e76\u884c\uff08EP\uff09\u76f8\u878d\u5408\u7684\u6df7\u5408\u5e76\u884c\u7b56\u7565\uff0c\u5e76\u642d\u914d FlashComm \u5207\u5206 AllReduce \u901a\u4fe1\uff1b\u6700\u540e\uff0c\u901a\u8fc7\u652f\u6301\u5355\u6b65\u751f\u6210\u591a\u4e2a token \u7684 MTP \u6280\u672f\uff0c\u8fdb\u4e00\u6b65\u62c9\u9ad8\u4e86 NPU \u7684\u8ba1\u7b97\u5bc6\u5ea6\u5e76\u7f29\u77ed\u4e86\u603b\u751f\u6210\u8017\u65f6\u3002<\/p>\n<p>\u5f97\u76ca\u4e8e\u4e0a\u8ff0\u8f6f\u786c\u534f\u540c\u7684\u6781\u81f4\u4f18\u5316\uff0cGLM-5 \u5728\u5355\u53f0\u56fd\u4ea7\u7b97\u529b\u8282\u70b9\u4e0a\u7684\u6027\u80fd\u8868\u73b0\uff0c\u5df2\u8db3\u53ef\u5ab2\u7f8e\u7531\u4e24\u53f0\u56fd\u9645\u4e3b\u6d41 GPU \u7ec4\u6210\u7684\u8ba1\u7b97\u96c6\u7fa4\uff1b\u4e0d\u4ec5\u5982\u6b64\uff0c\u5728\u957f\u5e8f\u5217\u5904\u7406\u573a\u666f\u4e0b\uff0c\u5176\u90e8\u7f72\u6210\u672c\u66f4\u662f\u5927\u5e45\u964d\u4f4e\u4e86 50%\u3002<\/p>\n<p>6 \u8bc4\u4f30<\/p>\n<p>\u5982\u4e0a\u6240\u8ff0\uff0cGLM-5 \u6807\u5fd7\u7740\u4ece Vibe Coding \u5230 Agentic Engineering \u65b0\u65f6\u4ee3\u7684\u8f6c\u53d8\u3002\u6211\u4eec\u9996\u5148\u5728\u667a\u80fd\u4f53\u3001\u63a8\u7406\u548c\u7f16\u7a0b\uff08ARC\uff09\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u8bc4\u4f30 GLM-5 \u4e0e\u524d\u6cbf\u6a21\u578b\u7684\u8868\u73b0\u3002\u4e3a\u5168\u9762\u8bc4\u4f30 GLM-5 \u5728\u771f\u5b9e\u4e16\u754c\u7684 Agentic Engineering \u573a\u666f\u4e2d\u7684\u6027\u80fd\uff0c\u6211\u4eec\u63d0\u51fa\u65b0\u7684\u5185\u90e8\u8bc4\u4f30\u57fa\u51c6 CC-Bench-V2\uff0c\u6db5\u76d6\u524d\u7aef\u3001\u540e\u7aef\u548c\u957f\u7a0b\u4efb\u52a1\u3002\u6700\u540e\uff0c\u6211\u4eec\u5728\u4e94\u4e2a\u5e38\u89c1\u7684\u771f\u5b9e\u573a\u666f\u4e2d\u8bc4\u4f30\u4e86 GLM-5 \u7684\u901a\u7528\u80fd\u529b\u3002<\/p>\n<p>6.1 ARC \u57fa\u51c6\u6d4b\u8bd5\u8bc4\u4f30<\/p>\n<p>\u6211\u4eec\u5728\u8868 7 \u4e2d\u62a5\u544a\u4e86 ARC \u57fa\u51c6\u6d4b\u8bd5\u7684\u4e3b\u8981\u7ed3\u679c\uff0c\u5c06 GLM-5 \u4e0e GLM-4.7\u3001DeepSeek-V3.2\u3001Kimi-K2.5\u3001Claude Opus 4.5\u3001Gemini 3 Pro \u548c GPT-5.2\uff08xhigh\uff09\u8fdb\u884c\u6bd4\u8f83\u3002\u603b\u4f53\u800c\u8a00\uff0cGLM-5 \u76f8\u5bf9 GLM-4.7 \u5b9e\u73b0\u4e86\u663e\u8457\u63d0\u5347\uff0c\u5728\u5f00\u6e90\u6a21\u578b\u4e2d\u8fbe\u5230\u4e1a\u754c\u6700\u4f18\u6027\u80fd\uff0c\u7f29\u5c0f\u4e86\u4e0e Claude Opus 4.5 \u7b49\u95ed\u6e90\u6a21\u578b\u7684\u5dee\u8ddd\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/95539575E5B2C06360D56CC7CAA9589AE06A4ABC_size280_w1080_h1049.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88687\uff1aGLM-5 \u4e0e\u5f00\u6e90\/\u95ed\u6e90\u6a21\u578b\u7684\u5bf9\u6bd4\u3002\u6807\u6709 * \u7684\u7ed3\u679c\u6765\u81ea\u5b8c\u6574\u7248 HLE \u6d4b\u8bd5\u96c6\u3002\u6807\u6709 \u2020 \u7684\u7ed3\u679c\u5728\u9a8c\u8bc1\u7248\u672c\u7684 Terminal-Bench 2.0 \u4e0a\u8bc4\u4f30\uff0c\u4fee\u6b63\u4e86\u90e8\u5206\u6a21\u7cca\u6307\u4ee4\u3002GDPval-AA Elo \u5206\u6570\u8bb0\u5f55\u4e8e 2026 \u5e74 2 \u6708 15 \u65e5\u3002\u6bcf\u9879\u57fa\u51c6\u6d4b\u8bd5\u7684\u6700\u9ad8\u5206\u52a0\u7c97\u663e\u793a\uff0c\u6b21\u9ad8\u5206\u52a0\u4e0b\u5212\u7ebf\u6807\u6ce8\u3002<\/p>\n<p>6.1.1 \u63a8\u7406\u4e0e\u901a\u7528\u57fa\u51c6\u6d4b\u8bd5\u8bc4\u4f30<\/p>\n<p>\u5728\u63a8\u7406\u548c\u901a\u7528\u57fa\u51c6\u6d4b\u8bd5\u65b9\u9762\uff0c\u6211\u4eec\u8bc4\u4f30\u4e86 Humanity's Last Exam\uff08HLE\uff09\u3001AIME 2026\u3001HMMT 2025\u3001IMO-AnswerBench\u3001GPQA-Diamond \u548c LongBench v2\u3002GLM-5 \u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u4e0e\u5f3a\u5927\u7684\u5f00\u6e90\u57fa\u7ebf Kimi-K2.5 \u76f8\u5f53\u3002\u4e0e\u95ed\u6e90\u6a21\u578b\u76f8\u6bd4\uff0cGLM-5 \u5728 HLE\uff08\u542b\u5de5\u5177\uff09\u4e0a\u4f18\u4e8e Claude Opus 4.5 \u548c Gemini 3 Pro\u3002GLM-5 \u5728 HMMT Feb.\/Nov. 2025 \u57fa\u51c6\u6d4b\u8bd5\u4e0a\u4f18\u4e8e Claude Opus 4.5 \u548c Gemini 3 Pro\u3002GLM-5 \u8fd8\u5728\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u6b65\uff0c\u5728\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5 LongBench v2 \u4e0a\u5f97\u5206\u6700\u9ad8\uff0c\u4ec5\u6b21\u4e8e Gemini 3 Pro\u3002<\/p>\n<p>6.1.2 \u7f16\u7a0b\u57fa\u51c6\u6d4b\u8bd5\u8bc4\u4f30<\/p>\n<p>\u5728\u7f16\u7a0b\u57fa\u51c6\u6d4b\u8bd5\u65b9\u9762\uff0c\u6211\u4eec\u5728 SWE-bench Verified\u3001SWE-bench Multilingual\u3001Terminal-Bench 2.0 \u548c CyberGym \u4e0a\u8fdb\u884c\u8bc4\u4f30\u3002GLM-5 \u5728\u5f00\u6e90 LLM \u4e2d\u8fbe\u5230\u7f16\u7a0b\u57fa\u51c6\u6d4b\u8bd5\u7684\u6700\u4f18\u6027\u80fd\uff0c\u4e0e\u95ed\u6e90 LLM \u76f8\u6bd4\uff0c\u5728 SWE-bench Verified \u4e0a\u4f18\u4e8e Gemini 3 Pro\uff0c\u5728 SWE-bench Multilingual \u4e0a\u51fb\u8d25 Gemini 3 Pro \u548c GPT-5.2\uff08xhigh\uff09\u3002\u5728 Terminal-Bench 2.0 \u4e0a\uff0cGLM-5 \u4e0e Claude Opus 4.5 \u76f8\u5f53\uff0c\u4fee\u6b63\u6a21\u7cca\u6307\u4ee4\u540e\u751a\u81f3\u8d85\u8d8a\uff0c\u5e76\u4e14\u5728 Claude Code \u548c Terminus-2 \u8fd9\u4e24\u79cd Agent \u6846\u67b6\u4e0b\u7684\u8868\u73b0\u4e00\u81f4\u3002GLM-5 \u5728\u7f51\u7edc\u5b89\u5168\u7f16\u7a0b\u57fa\u51c6\u6d4b\u8bd5 CyberGym \u4e0a\u76f8\u8f83 GLM-4.7 \u6709\u663e\u8457\u63d0\u5347\uff0c\u4ec5\u6b21\u4e8e Claude Opus 4.5\u3002<\/p>\n<p>6.1.3 Agent \u80fd\u529b\u8bc4\u4f30<\/p>\n<p>\u5728 Agent \u57fa\u51c6\u6d4b\u8bd5\u65b9\u9762\uff0c\u6211\u4eec\u5728 BrowseComp\u3001BrowseComp-ZH\u3001\u03c4&sup2;-Bench\u3001MCP-Atlas\u3001Tool-Decathlon\u3001Vending-Bench 2 \u548c GDPval-AA \u4e0a\u8fdb\u884c\u8bc4\u4f30\u3002GLM-5 \u5728\u6240\u6709\u524d\u6cbf LLM \u4e2d\uff0c\u5728\u6709\u65e0\u4e0a\u4e0b\u6587\u7ba1\u7406\u7684 BrowseComp \u4e0a\u5747\u8fbe\u5230\u6700\u4f18\u6027\u80fd\u3002\u5728 BrowseComp-ZH \u4e0a\u4e5f\u4f18\u4e8e Claude Opus 4.5 \u548c Gemini 3 Pro\u3002\u5728\u4e09\u9879\u5de5\u5177\u4f7f\u7528 Agent \u4efb\u52a1\uff08\u03c4&sup2;-Bench\u3001MCP-Atlas \u548c Tool-Decathlon\uff09\u4e0a\uff0cGLM-5 \u4e0e Claude Opus 4.5 \u76f8\u5f53\uff0c\u5c55\u793a\u4e86\u5f3a\u5927\u7684\u5de5\u5177\u4f7f\u7528\u80fd\u529b\u3002Vending-Bench 2 \u4e0a\u7684\u8bc4\u4f30\u7ed3\u679c\uff08$4432 \u4f59\u989d\uff09\u8fdb\u4e00\u6b65\u5c55\u793a\u4e86\u5176\u957f\u7a0b\u4efb\u52a1\u80fd\u529b\u3002\u5728 GDPval-AA \u6240\u4ee3\u8868\u7684\u5177\u5907\u7ecf\u6d4e\u4ef7\u503c\u7684\u573a\u666f\u4e2d\uff0cGLM-5 \u4f18\u4e8e Claude Opus 4.5\uff0c\u4ec5\u6b21\u4e8e GPT-5.2\uff08xhigh\uff09\u3002<\/p>\n<p>6.2 \u771f\u5b9e\u4e16\u754c Agentic Engineering \u4f53\u9a8c\u8bc4\u4f30<\/p>\n<p>\u771f\u5b9e\u4e16\u754c\u4f53\u9a8c\u6bd4\u6392\u884c\u699c\u66f4\u91cd\u8981\u3002\u6211\u4eec\u5c06\u5185\u90e8\u7684 CC-Bench \u5347\u7ea7\u4e3a CC-Bench-V2\uff0c\u7528\u4e8e\u8bc4\u4f30\u6a21\u578b\u5728\u771f\u5b9e\u7684 Agentic Engineering \u73af\u5883\u4e2d\uff0c\u80fd\u5426\u5728\u524d\u7aef\u3001\u540e\u7aef\u548c\u957f\u7a0b\u4efb\u52a1\u4e0a\u6b63\u786e\u5b8c\u6210\u7aef\u5230\u7aef\u4efb\u52a1\u3002CC-Bench-V2 \u5b8c\u5168\u53bb\u9664\u4e86\u4eba\u5de5\u6807\u6ce8\uff0c\u501f\u52a9 Claude Code \u548c\u5176\u4ed6 Agent \u6846\u67b6\uff0c\u7ed3\u5408\u5355\u5143\u6d4b\u8bd5\u4e0e Agent-as-a-Judge \u6280\u672f\uff0c\u5b9e\u73b0\u4e86\u5168\u81ea\u52a8\u8bc4\u4f30\u3002<\/p>\n<p>\u524d\u7aef\uff1a\u6211\u4eec\u91c7\u7528\u4ee5\u4e0b\u6d41\u7a0b\uff1a\u5148\u6784\u5efa Agent \u751f\u6210\u7684\u524d\u7aef\u9879\u76ee\uff0c\u5e76\u68c0\u67e5\u8bed\u6cd5\u3001\u4f9d\u8d56\u4e0e\u517c\u5bb9\u6027\u9519\u8bef\uff0c\u7136\u540e\u4f7f\u7528 Agent-as-a-Judge\uff0c\u901a\u8fc7\u914d\u5907 Playwright \u548c bash \u5de5\u5177\u7684 GUI Agent \u6a21\u62df\u7528\u6237\u4ea4\u4e92\uff0c\u9a8c\u8bc1\u7aef\u5230\u7aef\u6b63\u786e\u6027\u3002<\/p>\n<p>\u540e\u7aef\uff1a\u4efb\u52a1\u6765\u81ea C++\u3001Rust\u3001Go\u3001Java\u3001TypeScript \u548c Python \u7684\u771f\u5b9e\u5f00\u6e90\u9879\u76ee\uff0c\u6db5\u76d6\u529f\u80fd\u5b9e\u73b0\u3001\u9519\u8bef\u4fee\u590d\u3001\u56de\u5f52\u4fee\u590d\u548c\u6027\u80fd\u4f18\u5316\u3002\u6240\u6709\u6539\u52a8\u90fd\u5fc5\u987b\u5728\u771f\u5b9e\u5de5\u7a0b\u7ea6\u675f\u4e0b\u901a\u8fc7\u5b8c\u6574\u7684\u5355\u5143\u6d4b\u8bd5\u3002<\/p>\n<p>\u957f\u7a0b\u4efb\u52a1\uff1a\u6211\u4eec\u5148\u8bc4\u4f30\u6a21\u578b\u5728\u5927\u89c4\u6a21\u4ee3\u7801\u5e93\u4e2d\u7684\u4fe1\u606f\u68c0\u7d22\u80fd\u529b\uff0c\u8fd9\u662f\u50cf\u4eba\u7c7b\u5f00\u53d1\u4eba\u5458\u4e00\u6837\u5b9a\u4f4d\u6b63\u786e\u6587\u4ef6\u5e76\u7406\u89e3\u9879\u76ee\u4e0a\u4e0b\u6587\u7684\u524d\u63d0\u3002\u7136\u540e\uff0c\u6211\u4eec\u901a\u8fc7\u591a\u6b65\u94fe\u5f0f\u4efb\u52a1\u8bc4\u4f30\u7aef\u5230\u7aef\u6b63\u786e\u6027\u3002\u8fd9\u4e9b\u4efb\u52a1\u901a\u8fc7\u6316\u6398\u5177\u6709\u4e30\u5bcc commit \u5386\u53f2\u7684\u5df2\u5408\u5e76\u8fc7\u7684 Pull Request\uff0c\u5e76\u5c06\u8fd9\u4e9b commit \u805a\u7c7b\u6765\u6784\u5efa\u8fde\u8d2f\u7684\u4efb\u52a1\u94fe\u3002Agent \u6309\u987a\u5e8f\u6267\u884c\u8fd9\u4e9b\u4efb\u52a1\u94fe\uff0c\u4ee5\u6d4b\u8bd5\u5176\u7ef4\u6301\u4e0a\u4e0b\u6587\u5e76\u5904\u7406\u5404\u9636\u6bb5\u95f4\u4f9d\u8d56\u5173\u7cfb\u7684\u80fd\u529b\u3002\u8bc4\u4f30\u8fc7\u7a0b\u7ed3\u5408\u5355\u5143\u6d4b\u8bd5\u548c Agent-as-a-Judge \u6765\u9a8c\u8bc1\u529f\u80fd\u6b63\u786e\u6027\u548c\u4ee3\u7801\u8bed\u4e49\u5408\u7406\u6027\u3002<\/p>\n<p>6.2.1 \u524d\u7aef\u8bc4\u4f30\u2014\u2014Agent-as-a-Judge<\/p>\n<p>\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u9762\u5411\u524d\u7aef\u5f00\u53d1\u573a\u666f\u7684\u7efc\u5408\u81ea\u52a8\u5316\u8bc4\u6d4b\u57fa\u51c6\u3002\u8be5\u57fa\u51c6\u8986\u76d6\u5f00\u53d1\u8005\u5e38\u89c1\u5e94\u7528\u7c7b\u578b\uff1a\u843d\u5730\u9875\u3001\u7ba1\u7406\u4eea\u8868\u76d8\u3001\u6570\u636e\u53ef\u89c6\u5316\u3001\u56fe\u5f62\u4e0e\u52a8\u753b\u3001\u5728\u7ebf\u751f\u4ea7\u529b\u5de5\u5177\u3001\u4ea4\u4e92\u5f0f<a href=\"https:\/\/www.alphaseek.cn\/?tag=%e6%b8%b8%e6%88%8f\" title=\"View all posts in \u6e38\u620f\" target=\"_blank\" class=\"tag_link\">\u6e38\u620f<\/a>\u3001\u8868\u5355\u9a71\u52a8\u5de5\u4f5c\u6d41\uff0c\u6db5\u76d6 HTML\u3001React\u3001Vue\u3001Svelte \u548c Next.js \u4e3b\u6d41\u6280\u672f\u6808\u3002<\/p>\n<p>\u6bcf\u4e2a\u6d4b\u8bd5\u6837\u4f8b\u7531\u4e00\u4e2a\u5177\u4f53\u4efb\u52a1\u9700\u6c42\u548c\u4e00\u4efd\u68c0\u67e5\u6e05\u5355\u7ec4\u6210\uff0c\u5176\u4e2d\u6bcf\u4e2a\u68c0\u67e5\u9879\u90fd\u548c\u9700\u6c42\u70b9\u5bf9\u5e94\u3002\u8bc4\u4f30\u6d41\u7a0b\u6709\u4e24\u4e2a\u9636\u6bb5\uff1a1) \u9759\u6001\u9a8c\u8bc1\uff1a\u68c0\u67e5\u751f\u6210\u4ee3\u7801\u80fd\u5426\u6210\u529f\u6784\u5efa\u5e76\u8fd0\u884c\u30022) Agent-as-a-Judge\uff1a\u5bf9\u53ef\u6b63\u786e\u8fd0\u884c\u7684\u4ee3\u7801\uff0c\u4f7f\u7528 GUI Agent \u6a21\u62df\u4eba\u5de5\u6d4b\u8bd5\uff0c\u4ea4\u4e92\u5f0f\u9a8c\u8bc1\u6bcf\u4e2a\u68c0\u67e5\u9879\uff0c\u5e76\u6309\u9700\u6c42\u6ee1\u8db3\u60c5\u51b5\u6253\u5206\u3002\u6211\u4eec\u5b9a\u4e49\u4e86\u4ee5\u4e0b\u6307\u6807\uff1a\u6784\u5efa\u6210\u529f\u7387 (Build Sucess Rate, BSR)\uff0c\u6210\u529f\u521d\u59cb\u5316\u5e76\u8fd0\u884c\u7684\u9879\u76ee\u5360\u6bd4\uff1b\u5b9e\u4f8b\u6210\u529f\u7387 (Instance Success Rate\uff0cISR) \uff0c\u901a\u8fc7\u6240\u6709\u68c0\u67e5\u9879\u7684\u9879\u76ee\u5360\u6bd4\uff1b\u68c0\u67e5\u9879\u6210\u529f\u7387 (Check-item Success Rate\uff0cCSR) \uff0c\u6240\u6709\u68c0\u67e5\u9879\u7684\u7ec6\u7c92\u5ea6\u5b8c\u6210\u7387\u3002<\/p>\n<p>Agent-as-a-Judge\uff1a\u524d\u7aef\u5929\u7136\u5177\u6709\u53ef\u89c6\u5316\u4e0e\u4ea4\u4e92\u6027\uff0c\u5f88\u591a\u7f3a\u9677\u5f80\u5f80\u53ea\u6709\u5728\u7528\u6237\u70b9\u51fb\u6309\u94ae\u6216\u8c03\u6574\u7a97\u53e3\u5927\u5c0f\u7b49\u64cd\u4f5c\u65f6\u624d\u4f1a\u66b4\u9732\uff0c\u9759\u6001\u5206\u6790\u4e0e\u56fa\u5b9a\u5355\u5143\u6d4b\u8bd5\u4e0d\u8db3\u4ee5\u8986\u76d6\u3002\u56e0\u6b64\u6211\u4eec\u5f15\u5165\u4e86 Agent-as-a-Judge\uff08\u56fe 10\uff09\uff1a\u6bcf\u4e2a\u751f\u6210\u597d\u7684\u9879\u76ee\u5148\u90e8\u7f72\u5230 Docker \u5bb9\u5668\u4e2d\u5e76\u6784\u5efa\uff0c\u4ee5\u9a8c\u8bc1\u9759\u6001\u6b63\u786e\u6027\uff1b\u6784\u5efa\u6210\u529f\u7684\u5b9e\u4f8b\u968f\u540e\u79fb\u4ea4\u7ed9 Judge Agent\uff08Claude Code + Claude Sonnet 4.5\uff0c\u914d\u5907 Playwright MCP \u5de5\u5177\uff09\u6267\u884c\u95ed\u73af\u8bc4\u6d4b\u5faa\u73af\uff1a\u5bf9\u6bcf\u4e2a\u68c0\u67e5\u9879\uff0cAgent \u4f1a\u9605\u8bfb\u6e90\u7801\u3001\u8fdb\u884c\u5b9e\u65f6 UI \u4ea4\u4e92\uff08\u70b9\u51fb\u3001\u6309\u952e\u3001\u622a\u56fe\uff09\u3001\u68c0\u67e5\u7ec8\u7aef\u8f93\u51fa\u7b49\uff0c\u5e76\u7ed9\u51fa\u901a\u8fc7\u6216\u5931\u8d25\u7684\u5224\u65ad\u7ed3\u679c\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/E07A35FC55C5D7E9CE17073EBE1D080670A010F9_size249_w1080_h396.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u56fe10\uff1aAgent-as-a-Judge \u8bc4\u4f30\u6d41\u7a0b\u3002\u6bcf\u4e2a\u751f\u6210\u7684\u524d\u7aef\u9879\u76ee\u9996\u5148\u7ecf\u8fc7\u6784\u5efa\u4ee5\u9a8c\u8bc1\u9759\u6001\u6b63\u786e\u6027\uff0c\u6784\u5efa\u6210\u529f\u7684\u5b9e\u4f8b\u968f\u540e\u7531 Judge Agent \u8fdb\u884c\u4ea4\u4e92\u5f0f\u6d4b\u8bd5\uff0c\u4ee5\u5224\u65ad\u6bcf\u4e2a\u68c0\u67e5\u9879\u7684\u529f\u80fd\u6b63\u786e\u6027\u3002<\/p>\n<p>\u4e3a\u9a8c\u8bc1\u53ef\u9760\u6027\uff0c\u6211\u4eec\u4ece\u4e24\u4e2a\u7ef4\u5ea6\u5c06 Agent-as-a-Judge \u8bc4\u5206\u4e0e\u72ec\u7acb\u4eba\u7c7b\u4e13\u5bb6\u5224\u65ad\u8fdb\u884c\u5bf9\u6bd4\uff1a\u5728\u9010\u70b9\u4e00\u81f4\u6027\uff08point-wise consistency\uff09\u65b9\u9762\uff0c\u62bd\u6837 130 \u4e2a\u68c0\u67e5\u9879\uff0c\u7531\u4eba\u7c7b\u4e13\u5bb6\u72ec\u7acb\u8bc4\u5206\uff0c\u4e0e Judge Agent \u7ed3\u679c\u6709 94% \u4e00\u81f4\u6027\uff0c\u5206\u6b67\u96c6\u4e2d\u5728\u4e3b\u89c2\u89c6\u89c9\u8d28\u91cf\u6807\u51c6\u800c\u975e\u529f\u80fd\u6b63\u786e\u6027\uff1b\u5728\u6392\u540d\u4e00\u81f4\u6027\uff08ranking consistency\uff09\u65b9\u9762\uff0c\u5bf9 8 \u4e2a\u524d\u6cbf\u6a21\u578b\uff08\u5982 Claude Sonnet 4.5\u3001Claude Opus 4.5\u3001Gemini 3 Pro\u3001GLM-4.7\u3001DeepSeek-V3.2 \u7b49\uff09\u540c\u65f6\u91c7\u7528\u81ea\u52a8\u6846\u67b6\u8bc4\u4f30\u548c\u4eba\u7c7b\u4e13\u5bb6\u8bc4\u4f30\uff0c\u5f97\u5230\u7684\u6a21\u578b\u6392\u540d\u7684 Spearman \u76f8\u5173\u7cfb\u6570\u4e3a 85.7%\uff0c\u8868\u660e\u5f3a\u6b63\u76f8\u5173\u3002\u5982\u8868 8 \u6240\u793a\uff0cGLM-5 \u7684\u5e73\u5747 BSR \u8fbe\u5230 98.0% \uff0c\u5728 CSR \u4e0a\u4e0e Claude Opus 4.5 \u76f8\u6bd4\u5177\u5907\u7ade\u4e89\u529b\uff0c\u4f46\u5728\u4e09\u79cd\u6280\u672f\u6808\u4e2d\u5747\u5b58\u5728\u660e\u663e\u7684 ISR \u5dee\u8ddd\uff0c\u8bf4\u660e GLM-5 \u5728\u7aef\u5230\u7aef\u4efb\u52a1\u5b8c\u6210\u5ea6\u4e0a\u548c Claude Opus 4.5 \u76f8\u6bd4\u8fd8\u6709\u63d0\u5347\u7a7a\u95f4\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/B3BBADA2883056D0D5CC6172CF9D488D66F1F571_size117_w1080_h521.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88688\uff1aCC-Bench-V2 \u5728\u524d\u7aef\u3001\u540e\u7aef\u53ca\u957f\u7a0b\u4efb\u52a1\u4e0a\u7684\u8bc4\u4f30\u7ed3\u679c\u3002BSR\uff1a\u6784\u5efa\u6210\u529f\u7387\uff1bISR\uff1a\u5b9e\u4f8b\u6210\u529f\u7387\uff1bCSR\uff1a\u68c0\u67e5\u9879\u6210\u529f\u7387\u3002<\/p>\n<p>6.2.2 \u540e\u7aef\u8bc4\u4f30<\/p>\n<p>\u540e\u7aef\u8bc4\u4f30\u65e8\u5728\u8861\u91cf Agent \u5728\u771f\u5b9e\u7684\u5de5\u7a0b\u7ea6\u675f\u4e0b\uff0c\u80fd\u5426\u5bf9\u771f\u5b9e\u7684\u670d\u52a1\u5668\u7aef\u4ee3\u7801\u5e93\u8fdb\u884c\u6b63\u786e\u4e14\u80fd\u901a\u8fc7\u6d4b\u8bd5\u7684\u4fee\u6539\u3002\u6211\u4eec\u63d0\u51fa\u4e86 85 \u4e2a\u4efb\u52a1\uff0c\u6db5\u76d6 Python\u3001Go\u3001C++\u3001Rust\u3001Java \u548c TypeScript \u516d\u79cd\u8bed\u8a00\uff0c\u6d89\u53ca\u641c\u7d22\u5f15\u64ce\u3001\u6570\u636e\u5e93\u5f15\u64ce\u3001Web \u6846\u67b6\u3001AI \u63a8\u7406\u670d\u52a1\u3001\u77e5\u8bc6\u7ba1\u7406\u7cfb\u7edf\u4ee5\u53ca\u72ec\u7acb\u7b97\u6cd5\u548c\u7cfb\u7edf\u7f16\u7a0b\u7b49\u9886\u57df\u3002\u4efb\u52a1\u7c7b\u578b\u5305\u62ec\u529f\u80fd\u5b9e\u73b0\u3001\u7f3a\u9677\u4fee\u590d\u3001\u56de\u5f52\u4fee\u590d\u3001\u6027\u80fd\u4f18\u5316\u7b49\uff0c\u53cd\u6620\u65e5\u5e38\u540e\u7aef\u5f00\u53d1\u591a\u6837\u6027\u3002<\/p>\n<p>\u4e3a\u5b9e\u73b0\u5b8c\u5168\u81ea\u52a8\u5316\u7684\u8bc4\u4f30\uff0c\u6bcf\u4e2a\u4efb\u52a1\u90fd\u914d\u5907\u4e86\u4eba\u5de5\u7f16\u5199\u7684\u5355\u5143\u6d4b\u8bd5\uff08\u6bcf\u4e2a\u4efb\u52a1 5 - 10 \u4e2a\uff09\uff0c\u7528\u4e8e\u9a8c\u8bc1\u529f\u80fd\u6b63\u786e\u6027\u548c\u8fb9\u754c\u60c5\u51b5\u5904\u7406\u3002\u4efb\u52a1\u683c\u5f0f\u53c2\u8003 terminal-bench \u98ce\u683c\uff1a\u5728\u57fa\u4e8e\u9879\u76ee\u771f\u5b9e\u73af\u5883\u521d\u59cb\u5316\u7684 Docker \u5bb9\u5668\u5185\u8fd0\u884c\uff0c\u5411 Agent \u63d0\u4f9b\u5177\u4f53\u4efb\u52a1\u63cf\u8ff0\u4ee5\u89e3\u51b3\u4efb\u52a1\u3002\u6211\u4eec\u62a5\u544a Pass@1 \u6307\u6807\uff0c\u5373\u53ea\u6709\u5f53\u4efb\u52a1\u5173\u8054\u7684\u6240\u6709\u5355\u5143\u6d4b\u8bd5\u5168\u90e8\u901a\u8fc7\u65f6\uff0c\u624d\u89c6\u4e3a\u89e3\u51b3\u3002\u8fd9\u79cd\u4e25\u82db\u7684\u201c\u5168\u6709\u6216\u5168\u65e0\u201d\u6807\u51c6\u4f7f\u5f97\u8be5\u57fa\u51c6\u6d4b\u8bd5\u6781\u5177\u6311\u6218\u6027\uff1aGLM-5 \u548c Claude Opus 4.5 \u7684\u8868\u73b0\u76f8\u5f53\uff08\u89c1\u8868 8\uff09\uff0c\u4e24\u8005\u5747\u663e\u8457\u9886\u5148\u4e8e GLM-4.7\u3002<\/p>\n<p>6.2.3 \u957f\u7a0b\u4efb\u52a1\u8bc4\u4f30<\/p>\n<p>\u957f\u7a0b\u8bc4\u4f30\u805a\u7126\u4e8e\u8bc4\u4f30\u5c06\u751f\u4ea7\u7ea7 Agentic Engineering \u4e0e\u5355\u8f6e vibe coding \u533a\u5206\u5f00\u7684\u80fd\u529b\uff1a\u5728\u8d85\u5927\u4ee3\u7801\u5e93\u4e2d\u63a2\u7d22\uff0c\u5e76\u6267\u884c\u591a\u6b65\u5f00\u53d1\uff08\u6bcf\u4e00\u6b65\u90fd\u4f1a\u6539\u53d8\u540e\u7eed\u6b65\u9aa4\u7684\u4e0a\u4e0b\u6587\uff09\u3002\u6211\u4eec\u5c06\u6b64\u5206\u89e3\u4e3a\u4e24\u4e2a\u4e92\u8865\u4efb\u52a1\u3002<\/p>\n<p>\u5927\u89c4\u6a21\u4ee3\u7801\u5e93\u63a2\u7d22\uff08Large Repo Exploration\uff09\uff1a\u8fd9\u662f\u4efb\u4f55\u590d\u6742\u7f16\u7801\u4efb\u52a1\u7684\u524d\u63d0\uff0c\u662f\u5728\u5927\u578b\u964c\u751f\u4ed3\u5e93\u4e2d\u5b9a\u4f4d\u6b63\u786e\u6e90\u6587\u4ef6\u3002\u6211\u4eec\u9488\u5bf9\u5305\u542b\u6570\u4e07\u4e2a\u6587\u4ef6\u7684\u771f\u5b9e GitHub \u9ad8\u661f\u4ed3\u5e93\u6784\u5efa\u4e86\u81ea\u52a8\u5316\u57fa\u51c6\u3002\u6bcf\u4e2a\u95ee\u9898\u5747\u4ee5\u81ea\u7136\u3001\u9762\u5411\u4e1a\u52a1\u7684\u8bed\u4e49\u8868\u8ff0\uff0c\u4e25\u683c\u4e0d\u63d0\u6587\u4ef6\u540d\u3001\u7c7b\u540d\u3001\u51fd\u6570\u540d\u3002\u5e76\u4e14\u95ee\u9898\u901a\u5e38\u8981\u6c42\u591a\u8df3\u903b\u8f91\u63a8\u7406\uff0c\u624d\u80fd\u4ece\u7528\u6237\u7684\u63cf\u8ff0\u6620\u5c04\u5230\u5b9e\u9645\u7684\u4ee3\u7801\u5b9e\u73b0\u3002\u4f8b\u5982\uff0c\u201c\u5173\u4e8e\u751f\u6210\u89c6\u9891\u4e2d\u53e3\u578b\u540c\u6b65\u4e0d\u51c6\u201d\u8fd9\u4e00\u95ee\u9898\uff0c\u5b9e\u9645\u5bf9\u5e94\u5230\u89c6\u9891\u751f\u6210\u540e\u7aef\u4e2d\u7684\u53c2\u6570\u8c03\u6574\u6a21\u5757\u3002\u76ee\u6807\u6587\u4ef6\u7684\u9009\u62e9\u65e8\u5728\u6700\u5927\u5316\u63a2\u7d22\u96be\u5ea6\uff1a\u5b83\u4eec\u81f3\u5c11\u4f4d\u4e8e\u4e09\u7ea7\u76ee\u5f55\u6df1\u5904\uff1b\u547d\u540d\u6666\u6da9\uff0c\u96be\u4ee5\u5173\u952e\u8bcd\u68c0\u7d22\uff1b\u529f\u80fd\u5728\u4ed3\u5e93\u5185\u5177\u6709\u552f\u4e00\u6027\uff08\u65e0\u91cd\u590d\u5b9e\u73b0\uff09\uff1b\u4e14\u5728\u4e3b\u529f\u80fd\u8def\u5f84\u4e4b\u5916\u3002<\/p>\n<p>\u6211\u4eec\u62a5\u544a\u4e09\u6b21\u8fd0\u884c\u7684\u5e73\u5747 Pass@1\uff0c\u5982\u679c Agent \u5728\u63a2\u7d22\u8fc7\u7a0b\u4e2d\u6210\u529f\u8bfb\u53d6\u4e86\u76ee\u6807\u6587\u4ef6\uff0c\u5219\u89c6\u4e3a\u95ee\u9898\u89e3\u51b3\u3002\u5728\u6b64\u4efb\u52a1\u4e2d\uff0cGLM-5 \u7684\u8868\u73b0\u4f18\u4e8e Claude Opus 4.5\uff08\u89c1\u8868 8\uff09\uff0c\u4e14\u4e24\u8005\u5747\u8fdc\u8d85 GLM-4.7\u3002\u7ed3\u679c\u8868\u660e\uff0c\u9ad8\u6548\u4ed3\u5e93\u63a2\u7d22\u5bf9\u4ee3\u7801\u751f\u6210\u80fd\u529b\u7684\u4f9d\u8d56\u5c0f\u4e8e\u5bf9\u7b56\u7565\u6027\u641c\u7d22\u80fd\u529b\u7684\u4f9d\u8d56\uff0c\u9700\u8981\u6a21\u578b\u901a\u8fc7\u63a8\u7406\u6765\u6536\u7f29\u6587\u4ef6\u7a7a\u95f4\uff0cGLM-5 \u5728 Agent \u5de5\u5177\u4f7f\u7528\u8f68\u8ff9\u4e0a\u7684\u8bad\u7ec3\u4f7f\u5176\u5728\u8fd9\u65b9\u9762\u5177\u6709\u660e\u663e\u4f18\u52bf\u3002<\/p>\n<p>\u591a\u6b65\u94fe\u5f0f\u4efb\u52a1\uff08Multi-step Chained Tasks\uff09\uff1aSWE-bench \u7b49\u4e3b\u6d41\u7f16\u7a0b\u57fa\u51c6\u5c06\u8bc4\u4f30\u7b80\u5316\u4e3a\u201c\u5355\u6b21\u63d0\u4ea4\u3001\u9694\u79bb\u7f16\u8f91\u201d\uff0c\u56e0\u6b64\u65e0\u6cd5\u8bc4\u4f30 Agent \u8fdb\u884c\u589e\u91cf\u5f00\u53d1\u7684\u80fd\u529b\uff0c\u5373\u6bcf\u4e00\u6b65\u90fd\u4f1a\u6539\u53d8\u4ee3\u7801\u5e93\u72b6\u6001\u5e76\u5f71\u54cd\u540e\u7eed\u6b65\u9aa4\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u901a\u8fc7\u6316\u6398\u9ad8\u8d28\u91cf\u4ed3\u5e93\u4e2d\u5df2\u5408\u5e76\u7684PR\uff08Pull Request\uff09\uff0c\u5e76\u6309\u4ee5\u4e0b\u6d41\u7a0b\u6784\u5efa\u4e86\u957f\u7a0b\u4efb\u52a1\u8bc4\u4f30\u57fa\u51c6\uff1a<\/p>\n<p>PR \u7b5b\u9009\uff1a\u4ec5\u4fdd\u7559\u5305\u542b\u6d4b\u8bd5\u3001\u5305\u542b 3 \u5230 15 \u4e2a\u63d0\u4ea4\u4e14\u9075\u5faa\u7ebf\u6027\uff08\u975e\u5408\u5e76\uff09\u5386\u53f2\u7684\u5df2\u5408\u5e76 PR\u3002<\/p>\n<p>\u8bed\u4e49\u5206\u7ec4\uff1a\u4f7f\u7528 LLM \u5bf9\u76f8\u90bb\u63d0\u4ea4\u4e4b\u95f4\u7684\u8bed\u4e49\u76f8\u5173\u6027\u8fdb\u884c\u4e24\u4e24\u8bc4\u5206\uff1b\u52a8\u6001\u89c4\u5212\u6c42\u89e3\u6700\u4f18\u5207\u5206\uff0c\u5728\u4fdd\u6301\u63d0\u4ea4\u987a\u5e8f\u524d\u63d0\u4e0b\u6700\u5927\u5316\u7ec4\u5185\u8bed\u4e49\u4e00\u81f4\u6027\u3002<\/p>\n<p>\u8865\u4e01\u5206\u7c7b\uff1a\u6bcf\u4e2a\u4efb\u52a1\u7684\u7d2f\u8ba1 diff \u5212\u5206\u4e3a\u4e09\u7c7b\uff1agolden patch\uff08Agent \u5fc5\u987b\u4ea7\u51fa\u7684\u6838\u5fc3\u4ee3\u7801\uff09\u3001test patch\uff08\u7528\u4e8e\u9a8c\u8bc1\u7684\u6d4b\u8bd5\u8865\u4e01\uff09\u3001auto-apply patch\uff08\u81ea\u52a8\u5e94\u7528\u7684\u914d\u7f6e\u8865\u4e01\uff09\u3002<\/p>\n<p>\u95ee\u9898\u9648\u8ff0\u751f\u6210\uff1aLLM \u57fa\u4e8e\u8865\u4e01\u4e0e\u63d0\u4ea4\u4fe1\u606f\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u751f\u6210\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u9648\u8ff0\u3002<\/p>\n<p>\u4efb\u52a1\u5206\u7c7b\uff1a\u81ea\u52a8\u5206\u7c7b\u4e3a feature \/ bug-fix \/ refactor \/ test \/ config\uff0c\u5e76\u4ece\u9519\u8bef\u4fee\u590d\u3001\u5173\u952e\u8def\u5f84\u51c6\u786e\u6027\u3001\u6d4b\u8bd5\u901a\u8fc7\u7387\u4e09\u4e2a\u7ef4\u5ea6\u8fdb\u884c\u8bc4\u4f30\u3002<\/p>\n<p>\u73af\u5883\u9a8c\u8bc1\uff1a\u6784\u5efa Docker \u73af\u5883\uff0c\u5e94\u7528 golden patch\uff0c\u9a8c\u8bc1\u6574\u4e2a\u4efb\u52a1\u94fe\u96f6\u56de\u5f52\u3002<\/p>\n<p>\u7ed9\u5b9a\u4e00\u4e2a\u5305\u542b K \u4e2a\u4efb\u52a1\u7684\u94fe\uff0cAgent \u4ece\u57fa\u7840\u63d0\u4ea4\u5f00\u59cb\u987a\u5e8f\u5de5\u4f5c\uff1a\u5b8c\u6210\u4efb\u52a1 k \u540e\uff0c\u5176\u66f4\u6539\u4f1a\u88ab\u63d0\u4ea4\uff0c\u5e76\u5e94\u7528\u4efb\u52a1 k+1 \u7684 auto-apply patch\uff0c\u4f7f\u4ee3\u7801\u5e93\u72b6\u6001\u9010\u6b65\u7d2f\u79ef\u6f14\u5316\u3002\u8bc4\u4f30\u4f1a\u8f6e\u6d41\u68c0\u67e5\u6bcf\u4e2a\u63d0\u4ea4\uff0c\u5e76\u5728\u8fd0\u884c\u5b8c\u6574\u6d4b\u8bd5\u524d\uff0c\u7d2f\u79ef\u5e94\u7528\u4ece\u4efb\u52a1 1 \u5230 k \u7684\u6240\u6709 test patches\uff0c\u540c\u65f6\u6d4b\u8bd5\u5f53\u524d\u4efb\u52a1\u7684\u6b63\u786e\u6027\u5e76\u5bf9\u5148\u524d\u4efb\u52a1\u8fdb\u884c\u56de\u5f52\u6d4b\u8bd5\u3002\u6211\u4eec\u62a5\u544a\u5355\u4e2a\u4efb\u52a1\u7684 Pass@1\u3002<\/p>\n<p>\u8fd9\u79cd\u94fe\u5f0f+\u72b6\u6001\u9012\u5f52\u7684\u8bbe\u8ba1\uff0c\u76f4\u63a5\u8bc4\u4f30\u4e86\u5355\u6b21\u63d0\u4ea4\u57fa\u51c6\u6d4b\u8bd5\u65e0\u6cd5\u6db5\u76d6\u7684\u957f\u7a0b\u4e0a\u4e0b\u6587\u8ddf\u8e2a\u3001\u89c4\u5212\u548c\u589e\u91cf\u5f00\u53d1\u80fd\u529b\u3002\u5982\u8868 8 \u6240\u793a\uff0cGLM-5 \u8f83 GLM-4.7 \u6709\u663e\u8457\u63d0\u5347\uff0c\u4f46\u4e0e Claude Opus 4.5 \u76f8\u6bd4\u4ecd\u6709\u5f88\u5927\u5dee\u8ddd\u3002\u539f\u56e0\u5728\u4e8e\u94fe\u5f0f\u4efb\u52a1\u4e2d\u9519\u8bef\u4f1a\u7d2f\u79ef\u653e\u5927\uff1a\u67d0\u4e00\u6b65\u7684\u6b21\u4f18\u4fee\u6539\u53ef\u80fd\u6084\u7136\u7834\u574f\u540e\u7eed\u6b65\u9aa4\u6d4b\u8bd5\u3002\u7f29\u5c0f\u8fd9\u4e00\u5dee\u8ddd\uff0c\u9700\u8981\u5728\u957f\u4e0a\u4e0b\u6587\u4e00\u81f4\u6027\u548c\u957f\u7a0b\u81ea\u7ea0\u9519\u65b9\u9762\u7ee7\u7eed\u7a81\u7834\uff0c\u8fd9\u4e5f\u662f\u5f53\u524d\u6211\u4eec\u7684\u7814\u7a76\u91cd\u70b9\u3002<\/p>\n<p>6.2.4 \u52a8\u6001\u6f14\u5316\u7684 SWE \u4efb\u52a1\u8bc4\u4f30<\/p>\n<p>\u6211\u4eec\u5728 SWE-rebench \u4e0a\u8fdb\u884c\u8bc4\u4f30\uff0c\u56e0\u4e3a SWE-bench Verified \u662f\u4e00\u4e2a\u9759\u6001\u3001\u516c\u5f00\u3001\u4eba\u5de5\u9a8c\u8bc1\u7684\u6d4b\u8bd5\u96c6\uff0c\u53d1\u5e03\u5df2\u8d85\u8fc7 2 \u5e74\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cSWE-rebench \u57fa\u4e8e\u81ea\u52a8\u5316\u6d41\u7a0b\uff0c\u6301\u7eed\u6316\u6398\u65b0\u9c9c\u7684\u3001\u771f\u5b9e\u7684 GitHub issue \u4fee\u590d\u4efb\u52a1\uff0c\u5b9e\u73b0\u53bb\u6c61\u67d3\uff08decontaminated\uff09\u3001\u5bf9\u65f6\u95f4\u9c81\u68d2\uff08time-robust\uff09\u7684\u8bc4\u4f30\uff0c\u66f4\u597d\u5730\u8861\u91cf\u5bf9\u65b0\u8f6f\u4ef6\u5de5\u7a0b\u95ee\u9898\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u800c\u975e\u5bf9\u9759\u6001\u57fa\u51c6\u7684\u8bb0\u5fc6\u8868\u73b0\u3002\u8868 9 \u5c55\u793a\u4e86 GLM-5 \u5728 SWE-rebench \u4e0a\u7684\u5b98\u65b9\u6027\u80fd\uff0c\u8868\u660e GLM-5 \u80fd\u591f\u6709\u6548\u6cdb\u5316\u5230\u65b0\u7684 SWE \u95ee\u9898\u3002<\/p>\n<p><img data-lazyload=\"https:\/\/x0.ifengimg.com\/res\/2026\/50797FCA80A23EF07AA495B20470F0ECF22C29F0_size87_w1080_h334.png\" src=\"https:\/\/tech.ifeng.com\/c\/data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQMAAAAl21bKAAAAA1BMVEXy8vJkA4prAAAACklEQVQI12NgAAAAAgAB4iG8MwAAAABJRU5ErkJggg==\" style=\"display:block; margin-left:auto; margin-right:auto\" alt=\"\u91cd\u78c5\uff01\u667a\u8c31\u628aGLM-5\u7684\u6240\u6709\u6280\u672f\u7ec6\u8282\u90fd\u516c\u5f00\u4e86\"  \/><\/p>\n<p>\u88689\uff1aSWE-rebench \u4e0a\u7684\u6027\u80fd\u8868\u73b0\uff0c2026 \u5e74 1 \u6708\u3002<\/p>\n<p>6.3 \u771f\u5b9e\u4e16\u754c\u901a\u7528\u80fd\u529b\u8bc4\u4f30<\/p>\n<p>\u5c3d\u7ba1\u6807\u51c6\u5316\u5b66\u672f\u57fa\u51c6\u80fd\u63d0\u4f9b\u6709\u4ef7\u503c\u7684\u4fe1\u53f7\uff0c\u4f46\u5b83\u4eec\u5e76\u4e0d\u80fd\u5b8c\u6574\u523b\u753b\u6a21\u578b\u5728\u5b9e\u8df5\u4e2d\u7684\u4f7f\u7528\u65b9\u5f0f\u3002\u4e3a\u5f25\u8865\u8fd9\u4e00\u5dee\u8ddd\uff0c\u6211\u4eec\u57fa\u4e8e\u771f\u5b9e\u90e8\u7f72\u73af\u5883\u4e2d\u7684\u9ad8\u9891\u7528\u6237\u4ea4\u4e92\u6a21\u5f0f\uff0c\u8bc4\u4f30 GLM-5 \u4e00\u7cfb\u5217\u5728\u771f\u5b9e\u4efb\u52a1\u4e0a\u7684\u901a\u7528\u80fd\u529b\u3002<\/p>\n<p>\u8fd9\u4e9b\u80fd\u529b\u5305\u62ec\uff1a\u673a\u5668\u7ffb\u8bd1\u3001\u591a\u8bed\u8a00\u5bf9\u8bdd\u3001\u6307\u4ee4\u9075\u5faa\u3001\u4e16\u754c\u77e5\u8bc6\u3001\u5de5\u5177\u8c03\u7528\u3002\u4e0e\u4f20\u7edf\u201c\u4ee5\u57fa\u51c6\u4e3a\u4e2d\u5fc3\u201d\u7684\u8bc4\u4f30\u4e0d\u540c\uff0c\u6211\u4eec\u7684\u76ee\u6807\u662f\u8861\u91cf\u90a3\u4e9b\u80fd\u591f\u76f4\u63a5\u8f6c\u5316\u4e3a\u7528\u6237\u53ef\u611f\u77e5\u8d28\u91cf\u63d0\u5347\u7684\u6539\u8fdb\u3002\u5bf9\u6bcf\u9879\u80fd\u529b\uff0c\u6211\u4eec\u7efc\u5408\u91c7\u7528\uff1a\u5185\u90e8\u4eba\u5de5\u8bc4\u6d4b\u3001\u5185\u90e8\u81ea\u52a8\u5316\u8bc4\u6d4b\u3001\u5916\u90e8\u4eba\u5de5\u8bc4\u6d4b\u3001\u5916\u90e8\u81ea\u52a8\u5316\u57fa\u51c6\uff0c\u4ece\u800c\u540c\u65f6\u4fdd\u8bc1\u8bca\u65ad\u7c92\u5ea6\u4e0e\u8de8\u6a21\u578b\u53ef\u6bd4\u6027\u3002\u4f7f\u7528\u5916\u90e8\u57fa\u51c6\u65f6\uff0c\u6211\u4eec\u4f18\u5148\u9009\u62e9\u80fd\u591f\u53cd\u6620\u771f\u5b9e\u4ea4\u4e92\u6a21\u5f0f\u7684\u6570\u636e\u96c6\uff0c\u800c\u4e0d\u662f\u5206\u5e03\u6784\u9020\u8fc7\u7a84\u7684\u6d4b\u8bd5\u96c6\u3002<\/p>\n<p>\u56fe 11 \u5c55\u793a\u4e86 GLM-5 \u4e0e GLM-4.7 \u5728\u4e0a\u8ff0\u4e94\u7c7b\u771f\u5b9e\u4e16\u754c\u80fd\u529b\u4e0a\u7684\u5bf9\u6bd4\u7ed3\u679c\u3002GLM-5 \u5728\u6240\u6709\u8bc4\u4f30\u7ef4\u5ea6\u4e0a\u5747\u53d6\u5f97\u4e86\u4e00\u81f4\u63d0\u5347\uff0c\u5305\u62ec\u673a\u5668\u7ffb\u8bd1\u3001\u591a\u8bed\u8a00\u5bf9\u8bdd\u3001\u6307\u4ee4\u9075\u5faa\u3001\u4e16\u754c\u77e5\u8bc6\u4e0e\u5de5\u5177\u8c03\u7528\u3002\u5404\u80fd\u529b\u7684\u8be6\u7ec6\u8bc4\u4f30\u65b9\u5f0f\u4e0e\u6570\u636e\u96c6\u8bf4\u660e\u5982\u4e0b\uff1a<\/p>\n<\/p>\n<p>\u56fe 11\uff1aGLM-4.7 \u4e0e GLM-5 \u5728\u4e94\u4e2a\u771f\u5b9e\u901a\u7528\u80fd\u529b\u9886\u57df\u7684\u6027\u80fd\u5bf9\u6bd4\u3002<\/p>\n<p>6.3.1 \u673a\u5668\u7ffb\u8bd1<\/p>\n<p>ZMultiTransBench\uff1a\u8fd9\u662f\u4e00\u4e2a\u5185\u90e8\u6570\u636e\u96c6\uff0c\u5305\u542b 1220 \u4e2a\u6837\u672c\uff0c\u6e90\u81ea\u81ea\u4e3b\u641c\u96c6\u7684\u9ad8\u9891\u7ffb\u8bd1\u573a\u666f\uff0c\u6db5\u76d6\u4e03\u4e2a\u8bed\u79cd\u5bf9\uff1a\u4e2d\u8bd1\u897f (300)\u3001\u4e2d\u8bd1\u4fc4 (250)\u3001\u4e2d\u8bd1\u6cd5 (220)\u3001\u4e2d\u8bd1\u97e9 (200)\u3001\u4e2d\u8bd1\u65e5 (150)\u3001\u4e2d\u8bd1\u963f (50) \u4ee5\u53ca\u4e2d\u8bd1\u5fb7 (50)\u3002\u6240\u6709\u6837\u672c\u5747\u7531\u7ecf\u8fc7\u6b63\u89c4\u7ffb\u8bd1\u5b66\u8bad\u7ec3\u7684\u7814\u7a76\u751f\u5b8c\u6210\u7b5b\u9009\u3001\u7ffb\u8bd1\uff0c\u5e76\u8fdb\u884c\u72ec\u7acb\u6838\u9a8c\u3002\u8be5\u6570\u636e\u96c6\u5f3a\u8c03\u81ea\u7136\u53d1\u751f\u7684\u4f7f\u7528\u573a\u666f\uff0c\u800c\u975e\u4eba\u5de5\u6784\u9020\u7684\u6d4b\u8bd5\u7528\u4f8b\u3002\u8bc4\u4f30\u91c7\u7528\u4e0e\u56fa\u5b9a\u57fa\u51c6\u56de\u590d\u8fdb\u884c\u6210\u5bf9\u6bd4\u8f83\u7684\u65b9\u5f0f\u3002\u8bc4\u5224\u7531\u57fa\u4e8e GPT-4.1 \u7684\u81ea\u52a8\u8bc4\u4f30\u5668\u7ed9\u51fa\uff0c\u8bc4\u4f30\u7ef4\u5ea6\u5305\u62ec\u8bed\u4e49\u5fe0\u5b9e\u5ea6\u3001\u6d41\u7545\u6027\u548c\u6574\u4f53\u7ffb\u8bd1\u8d28\u91cf\u3002<\/p>\n<p>MENT-SNS\uff1a\u4e3a\u8fdb\u4e00\u6b65\u8bc4\u4f30\u6a21\u578b\u5728\u8bed\u8a00\u5b66\u4e0a\u66f4\u5177\u6311\u6218\u6027\u7684\u8bed\u5883\u4e2d\u7684\u9c81\u68d2\u6027\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u6765\u81ea MENT \u7684\u6e90\u53e5\uff0c\u5305\u542b 753 \u4e2a\u82f1\u6c49\u53e5\u5bf9\uff0c\u6db5\u76d6\u56db\u4e2a\u9886\u57df\uff1a\u793e\u4ea4\u7f51\u7edc\u670d\u52a1 (SNS)\u3001\u8de8\u6587\u5316\u3001\u8bd7\u6b4c\u548c\u6587\u5b66\u3002\u9009\u62e9\u8fd9\u4e9b\u9886\u57df\u662f\u4e3a\u4e86\u5728\u590d\u6742\u8bed\u8a00\u73b0\u8c61\u4e0b\u8fdb\u884c\u538b\u529b\u6d4b\u8bd5\uff0c\u5305\u62ec\u4fda\u8bed\u3001\u8c10\u97f3\u53cc\u5173\u3001\u60ef\u7528\u8868\u8fbe\u3001\u5386\u53f2\u5178\u6545\u548c\u9690\u55bb\u6027\u8bed\u8a00\u3002\u4e0e ZMultiTransBench \u7c7b\u4f3c\uff0c\u6240\u6709\u6837\u672c\u5747\u7531\u53d7\u8fc7\u4e13\u4e1a\u8bad\u7ec3\u7684\u7814\u7a76\u751f\u5b8c\u6210\u7b5b\u9009\u4e0e\u6838\u9a8c\u3002\u8bc4\u4f30\u6cbf\u7528\u4e0e\u57fa\u7ebf\u56de\u590d\u8fdb\u884c\u4e24\u4e24\u6bd4\u8f83\u7684\u65b9\u5f0f\uff0c\u5e76\u7531 GPT-4.1 \u62c5\u4efb\u81ea\u52a8\u8bc4\u5ba1\u6a21\u578b\u3002<\/p>\n<p>6.3.2 \u591a\u8bed\u8a00\u5bf9\u8bdd<\/p>\n<p>LMArena\uff1a\u6211\u4eec\u62a5\u544a\u4e86\u6765\u81ea LMArena \u7684 Elo \u8bc4\u5206\uff0c\u8fd9\u4e9b\u8bc4\u5206\u6e90\u81ea\u5927\u89c4\u6a21\u793e\u533a\u63d0\u4ea4\u7684\u6210\u5bf9\u6bd4\u8f83\u3002\u8be5\u6307\u6807\u53cd\u6620\u4e86\u5f00\u653e\u5f0f\u5bf9\u8bdd\u573a\u666f\u4e2d\u7684\u76f8\u5bf9\u6a21\u578b\u504f\u597d\uff0c\u53ef\u4f5c\u4e3a\u5bf9\u8bdd\u80fd\u529b\u7684\u5916\u90e8\u4fe1\u53f7\u3002<\/p>\n<p>ZMultiDialBench\uff1a\u9664\u516c\u5171\u699c\u5355\u5916\uff0c\u6211\u4eec\u8fd8\u5728\u5185\u90e8\u591a\u8bed\u8a00\u5bf9\u8bdd\u57fa\u51c6 ZMultiDialBench \u4e0a\u5f00\u5c55\u4eba\u5de5\u8bc4\u6d4b\u3002\u8be5\u6570\u636e\u96c6\u5305\u542b 141 \u6761\u7ecf\u7b5b\u9009\u6837\u4f8b\uff0c\u8986\u76d6\u591a\u7c7b\u5bf9\u8bdd\u573a\u666f\u3002\u6837\u672c\u6765\u6e90\u5305\u62ec\uff1a\u7531\u591a\u56fd\u6bcd\u8bed\u6807\u6ce8\u5458\u63d0\u4f9b\u7684\u9ad8\u8d28\u91cf\u5bf9\u8bdd\u6570\u636e\uff0c\u4ee5\u53ca\u7ebf\u4e0a\u7528\u6237\u53cd\u9988\u7684\u9ad8\u96be\u5931\u8d25\u6848\u4f8b\u3002\u8bc4\u6d4b\u96c6\u7ecf\u8fc7\u4e86\u7c7b\u522b\u5206\u7c7b\u548c\u6807\u51c6\u5316\uff0c\u4eba\u5de5\u8bc4\u6d4b\u5bf9\u533f\u540d\u5316\u7684\u6a21\u578b\u56de\u590d\u8fdb\u884c 1 \u81f3 10 \u5206\u7684\u6253\u5206\u3002<\/p>\n<p>6.3.3 \u6307\u4ee4\u9075\u5faa<\/p>\n<p>IF-Badcase\uff1aIF-Badcase \u662f\u4e00\u4e2a\u6839\u636e\u771f\u5b9e\u7528\u6237\u5728\u751f\u4ea7\u73af\u5883\u4e2d\u62a5\u544a\u7684\u6307\u4ee4\u9075\u5faa\u5931\u8d25\u6848\u4f8b\u6784\u5efa\u7684\u5185\u90e8\u57fa\u51c6\u3002\u8be5\u6570\u636e\u96c6\u7528\u4e8e\u8bc4\u4f30\u6a21\u578b\u5bf9\u201c\u73b0\u5b9e\u3001\u591a\u91cd\u7ea6\u675f\u6307\u4ee4\u201d\u7684\u4e25\u683c\u9075\u5faa\u80fd\u529b\uff0c\u91cd\u70b9\u5173\u6ce8\u6d41\u7a0b\u51c6\u786e\u6027\u3001\u903b\u8f91\u4e00\u81f4\u6027\u4e0e\u4e25\u683c\u683c\u5f0f\u8981\u6c42\u3002\u8bc4\u4f30\u91c7\u7528\u7ec6\u7c92\u5ea6\u7684\u68c0\u67e5\u9879\u8bc4\u4f30\u65b9\u5f0f\uff0c\u68c0\u67e5\u6a21\u578b\u7ed3\u679c\u662f\u5426\u7b26\u5408\u663e\u6027\u7ea6\u675f\uff0c\u5305\u62ec\u6b65\u9aa4\u987a\u5e8f\u3001\u57fa\u4e8e\u89c4\u5219\u7684\u6761\u4ef6\u548c\u7ed3\u6784\u89c4\u8303\u3002\u6240\u6709\u6837\u672c\u5747\u7ecf\u4eba\u5de5\u4e13\u5bb6\u6807\u6ce8\u3001\u590d\u6838\u5e76\u8fed\u4ee3\u7b5b\u9009\uff0c\u6700\u7ec8\u5f62\u6210 450 \u6761\u7cbe\u9009\u6d4b\u8bd5\u6837\u672c\u3002<\/p>\n<p>IF-Bench\uff1aIF-Bench \u8bc4\u4f30 LLM \u9075\u5faa\u590d\u6742\u3001\u5ba2\u89c2\u7ea6\u675f\u7684\u80fd\u529b\uff0c\u4f8b\u5982\u7279\u5b9a\u7684\u683c\u5f0f\u89c4\u5219\u3001\u957f\u5ea6\u9650\u5236\u548c\u5185\u5bb9\u9650\u5236\u3002\u5b83\u63d0\u4f9b\u4e86\u5bf9\u7cbe\u786e\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u7684\u5b9a\u91cf\u8861\u91cf\uff0c\u4fa7\u91cd\u4e8e\u53ef\u9a8c\u8bc1\u7684\u5408\u89c4\u6027\u800c\u975e\u5f00\u653e\u5f0f\u751f\u6210\u8d28\u91cf\u3002<\/p>\n<p>MultiChallenge\uff1aMultiChallenge \u8003\u5bdf LLM \u5728\u771f\u5b9e\u7684\u3001\u591a\u8f6e\u5bf9\u8bdd\u573a\u666f\u4e2d\u7684\u80fd\u529b\uff0c\u76ee\u6807\u662f\u8bc4\u4f30\u5728\u590d\u6742\u4ea4\u4e92\u4e2d\u5bf9\u6307\u4ee4\u7684\u51c6\u786e\u9075\u5faa\u3001\u4e0a\u4e0b\u6587\u5206\u914d\uff08context allocation\uff09\u548c\u4e0a\u4e0b\u6587\u63a8\u7406\uff08in-context reasoning\uff09\u80fd\u529b\u3002<\/p>\n<p>6.3.4 \u4e16\u754c\u77e5\u8bc6<\/p>\n<p>SimpleQA\uff1aSimpleQA \u901a\u8fc7\u5177\u6709\u552f\u4e00\u3001\u65e0\u4e89\u8bae\u7b54\u6848\u7684\u9ad8\u96be\u95ee\u9898\u8861\u91cf\u4e8b\u5b9e\u6027\u3002\u5b83\u901a\u8fc7\u5c06\u56de\u590d\u5206\u7c7b\u4e3a\u6b63\u786e\u3001\u9519\u8bef\u6216\u672a\u4f5c\u7b54\u6765\u8bc4\u4f30\u6a21\u578b\u7684\u6821\u51c6\u80fd\u529b\uff0c\u4f18\u5148\u5173\u6ce8\u51c6\u786e\u6027\u800c\u975e\u751f\u6210\u957f\u5ea6\u3002<\/p>\n<p>Chinese SimpleQA\uff1a\u8be5\u57fa\u51c6\u5c06 SimpleQA \u65b9\u6cd5\u9002\u914d\u5230\u4e2d\u6587\u8bed\u5883\uff0c\u8986\u76d6 6 \u4e2a\u4e3b\u8981\u9886\u57df\u4e0e 99 \u4e2a\u5b50\u4e3b\u9898\u3002\u5176\u4f7f\u7528\u9ad8\u8d28\u91cf\u3001\u9759\u6001\u3001\u77ed\u7b54\u6848\u95ee\u9898\uff0c\u5e76\u652f\u6301\u53ef\u9760\u81ea\u52a8\u8bc4\u5206\uff0c\u7528\u4e8e\u8bc4\u4f30 LLM \u7684\u77e5\u8bc6\u51c6\u786e\u6027\u3002<\/p>\n<p>6.3.5 \u5de5\u5177\u8c03\u7528<\/p>\n<p>ToolCall-Badcase\uff1aToolCall-Badcase \u662f\u4e00\u4e2a\u6e90\u81ea\u751f\u4ea7\u73af\u5883\u4e2d\u7528\u6237\u62a5\u544a\u7684\u5de5\u5177\u8c03\u7528\u573a\u666f\u5931\u8d25\u6848\u4f8b\u7684\u5185\u90e8\u57fa\u51c6\u3002\u6bcf\u4e2a\u5b9e\u4f8b\u90fd\u5173\u8054\u4e86\u4e00\u4e2a\u53ef\u9a8c\u8bc1\u7684\u6807\u51c6\u7b54\u6848\u5de5\u5177\u8c03\u7528\uff0c\u4ece\u800c\u80fd\u591f\u5bf9\u5de5\u5177\u9009\u62e9\u548c\u53c2\u6570\u6b63\u786e\u6027\u8fdb\u884c\u5ba2\u89c2\u8bc4\u4f30\u3002\u8bc4\u4f30\u8fc7\u7a0b\u8003\u6838\u6a21\u578b (1) \u662f\u5426\u8c03\u7528\u4e86\u6b63\u786e\u7684\u5de5\u5177\uff0c\u4ee5\u53ca (2) \u662f\u5426\u63d0\u4f9b\u4e86\u7ed3\u6784\u6b63\u786e\u4e14\u8bed\u4e49\u51c6\u786e\u7684\u53c2\u6570\u3002\u6240\u6709\u6837\u672c\u90fd\u7ecf\u8fc7\u591a\u8f6e\u5ba1\u6838\u3001\u6539\u5199\u548c\u9a8c\u8bc1\uff0c\u4ee5\u6d88\u9664\u6b67\u4e49\u5e76\u786e\u4fdd\u53ef\u8bc4\u4f30\u6027\u3002\u6700\u7ec8\u751f\u6210\u7684\u6570\u636e\u96c6\u5305\u542b 200 \u4e2a\u7cbe\u9009\u6d4b\u8bd5\u6848\u4f8b\uff0c\u53cd\u6620\u4e86\u771f\u5b9e\u7684\u5de5\u5177\u8c03\u7528\u80fd\u529b\u3002<\/p>\n<p>7 \u7ed3\u8bba<\/p>\n<p>\u672c\u6587\u4ecb\u7ecd\u4e86 GLM-5 \u8fd9\u4e00\u4e0b\u4e00\u4ee3\u57fa\u5ea7\u6a21\u578b\u7684\u7814\u53d1\u5de5\u4f5c\u3002\u5728\u5e94\u7528\u5c42\uff0cGLM-5 \u6210\u529f\u5b9e\u73b0\u4e86\u4ece Vibe Coding \u8303\u5f0f\u5411\u771f\u6b63\u7684 Agentic Engineering \u7684\u6f14\u8fdb\uff0c\u8fd9\u8868\u660e\u5728\u5904\u7406\u9ad8\u5ea6\u590d\u6742\u7684\u771f\u5b9e\u5de5\u4f5c\u6d41\u65f6\uff0c\u5f00\u6e90\u6a21\u578b\u5982\u4eca\u5df2\u80fd\u63d0\u4f9b\u4e0e\u9876\u5c16\u95ed\u6e90\u7cfb\u7edf\u76f8\u5ab2\u7f8e\u7684\u80fd\u529b\u3002\u603b\u4f53\u800c\u8a00\uff0cGLM-5 \u5e26\u6765\u4e86 AI \u843d\u5730\u6548\u80fd\u7684\u8303\u5f0f\u8f6c\u53d8\u3002\u901a\u8fc7\u5f00\u6e90 GLM-5\uff0c\u6211\u4eec\u81f4\u529b\u4e8e\u52a9\u529b\u5f00\u6e90\u793e\u533a\u7a81\u7834\u4f20\u7edf\u9759\u6001\u8bc4\u6d4b\u7684\u8fb9\u754c\uff0c\u63a2\u7d22\u9ad8\u80fd\u6548\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff08AGI\uff09\u7684\u524d\u6cbf\u3002\u6211\u4eec\u671f\u671b\u4ee5\u6b64\u4e3a\u5951\u673a\uff0c\u5171\u540c\u5f00\u542f\u7531 AI \u667a\u80fd\u4f53\u81ea\u4e3b\u89c4\u5212\u3001\u6267\u884c\u5e76\u4f18\u5316\u590d\u6742\u4efb\u52a1\u7684\u65b0\u65f6\u4ee3\u3002<\/p>\n<p>8 \u5f69\u86cb\u2014\u2014Pony Alpha<\/p>\n<p>\u5bf9\u6211\u4eec\u800c\u8a00\uff0c\u4ee3\u53f7\u4e3a\u201cPony Alpha\u201d\u7684\u76f2\u6d4b\u5b9e\u9a8c\u662f\u4e00\u4e2a\u91cc\u7a0b\u7891\u65f6\u523b\u3002\u5728 OpenRouter \u5e73\u53f0\u4e0a\u533f\u540d\u53d1\u5e03 GLM-5 \u662f\u4e00\u6b65\u9669\u68cb\uff0c\u4f46\u6700\u7ec8\u7684\u53cd\u9988\u4e0d\u4ec5\u5370\u8bc1\u4e86\u6211\u4eec\u7684\u6280\u672f\u5224\u65ad\uff0c\u66f4\u5e26\u6765\u4e86\u5de8\u5927\u7684\u60ca\u559c\u3002\u901a\u8fc7\u9690\u53bb\u6240\u6709\u54c1\u724c\u4fe1\u606f\uff0c\u6211\u4eec\u8ba9\u6a21\u578b\u5b8c\u6210\u4e86\u6700\u4e25\u82db\u7684\u201c\u81ea\u6211\u8bc1\u660e\u201d\uff0c\u786e\u4fdd\u6240\u6709\u7684\u8bc4\u4ef7\u90fd\u7eaf\u7cb9\u4e14\u5ba2\u89c2\u3002\u6838\u5fc3\u53cd\u9988\u5982\u4e0b\uff1a<\/p>\n<p>\u4e0a\u7ebf\u4ec5\u51e0\u5929\uff0cPony Alpha \u5c31\u5728 OpenRouter \u793e\u533a\u5f15\u53d1\u4e86\u8f70\u52a8\u3002\u5f00\u53d1\u8005\u4eec\u654f\u9510\u5730\u5bdf\u89c9\u5230\u4e86\u5b83\u5728\u5904\u7406\u590d\u6742\u4ee3\u7801\u3001Agent \u4efb\u52a1\u94fe\u8def\u53ca\u89d2\u8272\u626e\u6f14\u65f6\u7684\u5353\u8d8a\u80fd\u529b\u3002\u8fd9\u79cd\u201c\u795e\u79d8\u7684\u5f3a\u5927\u201d\u5f15\u53d1\u4e86\u5e7f\u6cdb\u731c\u6d4b\uff1a\u521d\u6b65\u7edf\u8ba1\u663e\u793a\uff0c25% \u7684\u7528\u6237\u63a8\u6d4b\u5b83\u662f Anthropic \u7684 Claude Sonnet 5\uff0c20% \u8ba4\u4e3a\u662f Grok \u7684\u65b0\u7248\u672c\uff0c10% \u731c\u662f DeepSeek V4\uff0c\u800c\u5176\u4f59\u7528\u6237\u5219\u6210\u529f\u731c\u4e2d\u4e86 GLM-5\u3002<\/p>\n<p>\u6700\u7ec8\u786e\u8ba4 Pony Alpha \u771f\u8eab\u5373\u662f GLM-5\uff0c\u8fd9\u5bf9\u6211\u4eec\u56e2\u961f\u662f\u4e00\u6b21\u5de8\u5927\u7684\u9f13\u821e\uff0c\u4e5f\u6709\u529b\u56de\u51fb\u4e86\u957f\u671f\u4ee5\u6765\u5916\u754c\u5bf9\u4e2d\u56fd\u672c\u571f\u6a21\u578b\u6280\u672f\u6c34\u51c6\u7684\u8d28\u7591\u3002<\/p>\n<p>Pony Alpha\uff08GLM-5\uff09\u4e0d\u4ec5\u8d62\u5728\u4e86\u57fa\u51c6\u6d4b\u8bd5\u7684\u699c\u5355\u4e0a\uff0c\u66f4\u8d62\u5728\u4e86\u771f\u5b9e\u573a\u666f\u91cc\u3002\u8fd9\u6807\u5fd7\u7740\u6211\u4eec\u7684\u7814\u53d1\u91cd\u5fc3\u5df2\u6df1\u5ea6\u8f6c\u5411\u201c\u5de5\u7a0b\u7ea7\u53ef\u9760\u6027\u201d\u3002\u8fd9\u6b21\u533f\u540d\u6d4b\u8bd5\u6253\u7834\u4e86\u5148\u5165\u4e3a\u4e3b\u7684\u5730\u7f18\u653f\u6cbb\u504f\u89c1\uff0c\u8ba9\u793e\u533a\u7684\u8ba4\u53ef\u56de\u5f52\u5230\u4e86\u201c\u597d\u7528\u4e0e\u5426\u201d\u8fd9\u4e00\u6700\u7eaf\u7cb9\u7684\u6280\u672f\u672c\u8d28\u3002\u5e86\u795d\u4e4b\u4f59\uff0c\u6211\u4eec\u6df1\u77e5\u4efb\u91cd\u9053\u8fdc\u3002\u5f00\u6e90\u8ffd\u8d76\u95ed\u6e90\u7684\u6218\u5f79\u4ecd\u5728\u7ee7\u7eed\uff0c\u6211\u4eec\u5c06\u575a\u5b9a\u4e0d\u79fb\u5730\u63a2\u7d22\u6280\u672f\u524d\u6cbf\uff0c\u6784\u5efa\u66f4\u9ad8\u6548\u3001\u66f4\u667a\u80fd\u7684\u5e95\u5c42\u7cfb\u7edf\u3002<\/p>\n<p>(\u6765\u6e90\uff1a\u65b0\u6d6a\u79d1\u6280)<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6211\u4eec\u63a8\u51fa\u4e86 GLM-5\uff0c\u8fd9\u662f\u4e00\u6b3e\u65e8\u5728\u63a8\u52a8\u7f16\u7a0b\u8303\u5f0f\u4ece\u201cVibe Coding\u201d\uff08\u6c1b\u56f4\u7f16\u7a0b\uff09\u8f6c\u5411\u201cAgentic  [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[6],"tags":[172,378,71,148,252,23,82,328,321,88,339,238,24],"views":21,"_links":{"self":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/728376"}],"collection":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=728376"}],"version-history":[{"count":0,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/728376\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=728376"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=728376"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=728376"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}