Harness\u5c42\u7684RSI<\/p>\n
\u8c37\u6b4cRSI\u65b0\u6210\u679c\uff0c\u5f00\u6e90\u4e86\u3002<\/p>\n
\u540d\u53ebHarness\u00b2\uff0c\u8ba9\u5927\u6a21\u578b\u5728\u8fd0\u884c\u4e2d\u81ea\u5df1\u7ed9\u81ea\u5df1\u9b54\u6539\u5916\u6302\u4ee3\u7801\u3001\u7cfb\u7edf\u63d0\u793a\u8bcd\u3001\u5de5\u5177\u94fe\u548c\u6280\u80fd\u5e93\u3002<\/p>\n
\u662f\u7684\uff0cHarness\u5c42\u7684RSI\u3002<\/p>\n
\u6548\u679c\u4e5f\u662f\u5f88\u731b\u3002<\/p>\n
\u88c5\u914d\u4e0a\u8fd9\u4ef6RSI\u4e13\u7cbe\u7684GLM-5\uff0c\u5728\u529e\u516c\u57fa\u51c6WorkBuddy-Bench\u4e0a\uff0c\u76f4\u63a5\u98d9\u5347\u4e8611.5\u5206\u2014\u2014<\/p>\n
\u77ac\u95f4\u53cd\u8d85\u539f\u751fGemini 3.5 Flash\u3002<\/p>\n
\u603b\u4e4b\u5c31\u662f\u4e00\u4e2a\u975e\u5e38\u79bb\u8c31\u7684\u5916\u6302\uff0c\u4efb\u4f55\u6a21\u578b\u5957\u4e0a\u6765\uff0c\u8dd1\u5206\u90fd\u76f4\u63a5\u8d77\u98de\u4e86\u3002<\/p>\n
\u76ee\u524d\uff0c\u8be5\u9879\u76ee\u5df2\u5168\u9762\u5f00\u6e90\uff0c\u5df2\u652f\u6301OpenCode\u4e0eCodex\u3002<\/p>\n
\u8fc7\u53bb\u4e1a\u5185\u4e3a\u4e86\u8ba9Agent\u53d8\u806a\u660e\uff0c\u641e\u8fc7\u4e0d\u5c11\u9012\u5f52\u81ea\u6211\u6539\u8fdb\uff08Recursive Self-Improvement\uff0c\u7b80\u79f0RSI\uff09\u3002<\/p>\n
\u601d\u8def\u542c\u8d77\u6765\u5f88\u7f8e\uff1a\u8ba9AI\u81ea\u5df1\u5199\u4ee3\u7801\u3001\u81ea\u5df1\u6539\u63d0\u793a\u8bcd\u3001\u81ea\u5df1\u4f18\u5316\u7cfb\u7edf\u3002\u4f46\u53ea\u8981\u4e00\u653e\u8fdb\u771f\u5b9e\u7684\u5f00\u653e\u4e16\u754c\uff0c\u51e0\u4e4e\u5168\u519b\u8986\u6ca1\u3002<\/p>\n
\u4e3a\u5565\u5462\uff1f<\/p>\n
\u4e3b\u8981\u5c31\u5361\u5728\u4e24\u9053\u6b7b\u7a74\u4e0a\uff1a<\/p>\n
\u6ca1\u6709\u6807\u51c6\u7b54\u6848\u3002<\/p>\n
\u5b9e\u9a8c\u5ba4\u91cc\u5237\u9898\u6709\u5355\u5143\u6d4b\u8bd5\uff0c\u6709Ground Truth\u544a\u8bc9\u4f60\u5bf9\u4e0d\u5bf9\u3002\u4f46\u5230\u4e86\u771f\u5b9e\u6253\u5de5\u573a\u666f\uff0c\u5199\u4e00\u4efd\u5546\u4e1a\u4f01\u5212\u6848\u3001\u5ba1\u4e00\u4efd\u6cd5\u5f8b\u5408\u540c\uff0c\u7ebf\u4e0a\u6839\u672c\u6ca1\u4eba\u7ed9\u4f60\u6253\u5206\u3002<\/p>\n
AI\u81ea\u5df1\u6539\u4e86\u4e00\u7248Prompt\uff0c\u5b83\u600e\u4e48\u77e5\u9053\u6539\u5b8c\u662f\u53d8\u806a\u660e\u4e86\u8fd8\u662f\u5728\u778e\u7f16\uff1f<\/p>\n
\u8fc7\u5ea6\u62df\u5408\u5355\u6b21\u4efb\u52a1\u3002<\/p>\n
\u597d\u4e0d\u5bb9\u6613\u4e3a\u4e86\u4fee\u67d0\u4e2a\u4efb\u52a1\u628a\u811a\u624b\u67b6\u9b54\u6539\u4e86\u4e00\u901a\uff0c\u7ed3\u679c\u4e0b\u4e00\u4e2a\u4efb\u52a1\u6362\u6210\u4e86\u5199\u722c\u866b\uff0c\u4e4b\u524d\u6539\u7684\u89c4\u5219\u5168\u6210\u4e86\u7d2f\u8d58\u3002<\/p>\n
\u76f4\u63a5\u628a\u6539\u5b8c\u7684\u811a\u624b\u67b6\u786c\u5957\u7ed9\u4e0b\u4e00\u4e2a\u4efb\u52a1\uff0c\u7cfb\u7edf\u5f88\u5feb\u5c31\u9677\u5165\u566a\u58f0\u87ba\u65cb\uff0c\u8d8a\u6539\u8d8a\u50bb\u3002<\/p>\n
\u8c37\u6b4c\u56e2\u961f\u8fd9\u6b21\u63d0\u51fa\u6765\u7684Harness\u00b2\uff0c\u5c31\u662f\u4e13\u95e8\u51b2\u7740\u8fd9\u4e24\u4e2a\u6b7b\u7a74\u5f00\u5200\u7684\u3002<\/p>\n
\u4ed6\u4eec\u642d\u4e86\u4e00\u5957\u53cc\u5c42\u9012\u5f52\u67b6\u6784\u3002<\/p>\n
\u53cc\u5c42\u9012\u5f52\uff1a\u5de6\u624b\u505a\u5b9e\u9a8c\uff0c\u53f3\u624b\u8bb0\u7b14\u8bb0<\/p>\n
Harness\u00b2 \u540d\u5b57\u91cc\u7684\u8fd9\u4e2a\u201c\u5e73\u65b9\u201d\uff0c\u6307\u7684\u5c31\u662f\u4e24\u5c42\u4e92\u76f8\u5d4c\u5957\u7684\u9012\u5f52\u5faa\u73af\uff1a\u4efb\u52a1\u7ea7\u9012\u5f52\uff08Task-level\uff09\u548c\u6539\u8fdb\u7ea7\u9012\u5f52\uff08Harnessing-level\uff09\u3002<\/p>\n
\u7b2c\u4e00\u5c42\uff0c\u4efb\u52a1\u7ea7\u9012\u5f52\uff1a\u5728\u6ca1\u6709\u6807\u51c6\u7b54\u6848\u7684\u60c5\u51b5\u4e0b\u505a\u5bf9\u7167\u5b9e\u9a8c\u3002<\/p>\n
\u9762\u5bf9\u773c\u524d\u8fd9\u4e2a\u65b0\u4efb\u52a1\uff0c\u7cfb\u7edf\u5148\u7528\u521d\u59cb\u811a\u624b\u67b6\u8dd1\u4e00\u7248\u53c2\u8003\u8f68\u8ff9\u51fa\u6765\u3002\u8fd9\u7248\u7ed3\u679c\u5bf9\u4e0d\u5bf9\u5148\u4e0d\u7ba1\uff0c\u6838\u5fc3\u662f\u770b\u6a21\u578b\u5728\u6267\u884c\u8fc7\u7a0b\u4e2d\u66b4\u9732\u51fa\u4ec0\u4e48\u7f3a\u9677\u3002<\/p>\n
\u7d27\u63a5\u7740\uff0c\u5b83\u542f\u52a8\u4e86\u4e00\u5957\u53eb\u505aPropose\u2013Probe\u2013Compose\uff08\u63d0\u8bae-\u63a2\u6d4b-\u5408\u6210\uff09\u7684\u673a\u5236\uff1a<\/p>\n
Propose\uff08\u63d0\u8bae\uff09\uff1a<\/p>\n
\u7531\u4e13\u95e8\u7684\u7f16\u8f91\u89d2\u8272\u9488\u5bf9\u521a\u624d\u7684\u8f68\u8ff9\u63d0\u51fa\u4e24\u5957\u4fee\u6539\u65b9\u6848\u3002<\/p>\n
\u4e00\u5957\u662f\u201c\u901a\u7528\u7f3a\u9677\u4fee\u8865\u201d\uff08Invariant\uff09\uff0c\u4e13\u95e8\u6293\u8de8\u9886\u57df\u901a\u7528\u7684\u6bdb\u75c5\uff0c\u6bd4\u5982\u89c4\u5212\u6df7\u4e71\u3001\u5de5\u5177\u8c03\u7528\u6b7b\u5faa\u73af\u3001\u4e0a\u4e0b\u6587\u7206\u70b8\uff1b<\/p>\n
\u53e6\u4e00\u5957\u662f\u201c\u9886\u57df\u4e13\u5c5e\u4fee\u8865\u201d\uff08Domain\uff09\uff0c\u4e13\u95e8\u9488\u5bf9\u5f53\u524d\u4e1a\u52a1\u7684\u7279\u5b9a\u77e5\u8bc6\u548c\u7ea6\u675f\u3002<\/p>\n
Probe\uff08\u63a2\u6d4b\uff09\uff1a<\/p>\n
\u628a\u8fd9\u4e24\u5957\u4fee\u6539\u540e\u7684\u5019\u9009\u811a\u624b\u67b6\u653e\u56de\u540c\u4e00\u4e2a\u4efb\u52a1\u91cc\u5206\u522b\u8dd1\u4e00\u904d\u3002<\/p>\n
\u6ca1\u6709\u88c1\u5224\u6ca1\u5173\u7cfb\uff0c\u7cfb\u7edf\u6293\u53d6\u5b83\u4eec\u7684\u6267\u884c\u8f68\u8ff9\u548c\u53c2\u8003\u8f68\u8ff9\u505a\u884c\u4e3a\u5bf9\u6bd4\uff08Behavioral Contrast\uff09\u3002\u770b\u6539\u52a8\u5230\u5e95\u6709\u6ca1\u6709\u89e3\u51b3\u521a\u624d\u5361\u4f4f\u7684\u6b65\u9aa4\uff0c\u6709\u6ca1\u6709\u5f15\u53d1\u65b0\u7684\u903b\u8f91\u9000\u5316\u3002<\/p>\n
Compose\uff08\u5408\u6210\uff09\uff1a<\/p>\n
\u5408\u6210\u5668\u628a\u6240\u6709\u6267\u884c\u8bc1\u636e\u6446\u5728\u4e00\u8d77\uff0c\u6311\u51fa\u90a3\u4e9b\u771f\u6b63\u88ab\u884c\u4e3a\u8bc1\u636e\u8bc1\u660e\u6709\u6548\u7684\u6539\u52a8\uff0c\u628a\u5b83\u4eec\u5408\u5e76\u8fdb\u6700\u7ec8\u7684\u6267\u884c\u811a\u624b\u67b6\u91cc\u3002<\/p>\n
\u8fd9\u8fd8\u4e0d\u7b97\u5b8c\u3002\u5b83\u8fd8\u652f\u6301\u5e76\u884c\u6269\u5c55\uff08Parallel Recursion\uff09\uff0c\u4e00\u6b21\u6027\u94fa\u5f00 $k$ \u4e2a\u5019\u9009\u65b9\u6848\u5e76\u53d1\u63a2\u6d4b\uff0c\u5728\u66f4\u5bbd\u7684\u63a2\u7d22\u7a7a\u95f4\u91cc\u62e9\u4f18\u5408\u6210\u3002<\/p>\n
\u795e\u4e86\u3002<\/p>\n
\u7b2c\u4e8c\u5c42\uff0c\u6539\u8fdb\u7ea7\u9012\u5f52\uff1a\u628a\u6539\u811a\u624b\u67b6\u7684\u7ecf\u9a8c\u6c89\u6dc0\u6210\u5143\u80fd\u529b\u3002<\/p>\n
\u5f53\u524d\u4efb\u52a1\u505a\u5b8c\u4e86\uff0c\u4ea4\u4ed8\u7269\u4ea4\u51fa\u53bb\u4e86\uff0c\u5f88\u591a\u7cfb\u7edf\u7684\u751f\u547d\u5468\u671f\u5c31\u5230\u6b64\u4e3a\u6b62\u4e86\u3002<\/p>\n
\u4f46Harness\u00b2 \u771f\u6b63\u5389\u5bb3\u7684\u5728\u540e\u5934\u2014\u2014<\/p>\n
\u628a\u6574\u4e2a\u8c03\u8bd5\u8bb0\u5f55\uff08\u53c2\u8003\u8f68\u8ff9\u3001\u5019\u9009\u4fee\u6539\u3001\u5bf9\u6bd4\u8bc1\u636e\u3001\u5408\u6210\u51b3\u7b56\uff09\u4ea4\u7ed9\u53cd\u601d\u6a21\u5757\uff08Experience Reflector\uff09\u3002<\/p>\n
\u53cd\u601d\u6a21\u5757\u590d\u76d8\u4e4b\u540e\uff0c\u4f1a\u5199\u4e00\u4efdEXPERIENCE.md\uff0c\u628a\u8fd9\u6b21\u4e3a\u4ec0\u4e48\u8fd9\u4e48\u6539\u3001\u5728\u4ec0\u4e48\u6761\u4ef6\u4e0b\u8d77\u6548\u3001\u5e26\u6765\u4e86\u4ec0\u4e48\u540e\u679c\uff0c\u6e05\u6e05\u695a\u695a\u5b58\u8fdb\u6301\u7eed\u6539\u8fdb\u7ecf\u9a8c\u5e93\uff08Harness Improvement Library\uff09\u3002<\/p>\n
\u7b49\u5230\u4e0b\u4e00\u4e2a\u65b0\u4efb\u52a1\u8fdb\u6765\u65f6\uff1a<\/p>\n
\u6267\u884c\u7aef\u62ff\u5230\u7684\uff0c\u4f9d\u7136\u662f\u4e00\u4e2a\u5e72\u5e72\u51c0\u51c0\u7684\u57fa\u7840\u811a\u624b\u67b6\u3002<\/p>\n
\u4f46\u662f\u8d1f\u8d23\u6539\u811a\u624b\u67b6\u7684\u90a3\u4e2a\u201c\u6539\u8fdb\u5927\u8111\u201d\uff0c\u624b\u91cc\u5df2\u7ecf\u6525\u7740\u539a\u539a\u7684\u5386\u53f2\u7ecf\u9a8c\u624b\u518c\u3002\u5728\u63d0\u8bae\u548c\u5408\u6210\u9636\u6bb5\uff0c\u5b83\u80fd\u76f4\u63a5\u68c0\u7d22\u8fc7\u5f80\u76f8\u4f3c\u573a\u666f\u4e0b\u7684\u6709\u6548\u7ec4\u4ef6\u548c\u907f\u5751\u6307\u5357\u3002<\/p>\n
\u6539\u8fdb\u7684\u4e0d\u53ea\u662f\u773c\u524d\u7684\u811a\u624b\u67b6\uff0c\u66f4\u662f\u672a\u6765\u6539\u8fdb\u811a\u624b\u67b6\u7684\u80fd\u529b\u3002<\/p>\n
\u8fd9\u5c31\u628a\u5355\u6b21\u4efb\u52a1\u7684\u4e34\u65f6\u4fee\u8865\uff0c\u53d8\u6210\u4e86\u7ec8\u8eab\u5b66\u4e60\u7684\u8ba4\u77e5\u8d44\u4ea7\u3002<\/p>\n
\u673a\u5236\u8bb2\u5f97\u518d\u6f02\u4eae\uff0c\u7ec8\u7a76\u5f97\u62ff\u771f\u5b9e\u6218\u573a\u7684\u6570\u636e\u8bf4\u8bdd\u3002<\/p>\n
\u56e2\u961f\u76f4\u63a5\u4e0a\u4e86\u4e09\u4e2a\u6781\u5177\u5206\u91cf\u7684\u786c\u6838\u804c\u4e1a\u57fa\u51c6\uff1a<\/p>\n
WorkBuddy-Bench\uff1a\u6db5\u76d6\u4ee3\u7801\uff08Code\uff09\u3001\u7f51\u9875\uff08Web\uff09\u3001\u529e\u516c\u5957\u4ef6\uff08Office\uff09\u4e09\u5927\u9ad8\u96be\u5ea6\u529e\u516c\u5de5\u4f5c\u6d41\uff1b<\/p>\n
JobBench\uff1a\u957f\u7a0b\u7aef\u5230\u7aef\u590d\u6742\u4e13\u4e1a\u5c97\u4f4d\u4efb\u52a1\uff1b<\/p>\n
LAB\uff1aHarvey\u5f00\u6e90\u7684\u9876\u7ea7\u6cd5\u5f8b\u667a\u80fd\u4f53\u57fa\u51c6\u3002<\/p>\n
\u5e95\u5ea7\u6a21\u578b\u9009\u4e86Gemini 3.5 Flash\u548cGLM-5\uff0c\u6267\u884c\u811a\u624b\u67b6\u6311\u4e86\u5f00\u6e90\u4e3b\u6d41\u7684OpenCode\u548cCodex\u3002<\/p>\n
\u7ed3\u679c\u6781\u5176\u5938\u5f20\u3002<\/p>\n
\u5728WorkBuddy-Bench\u6d4b\u8bd5\u91cc\uff0c\u4ec5\u4ec5\u8fdb\u884c\u5355\u6b65Harness\u00b2 \u6539\u8fdb\uff08$k=1$\uff09\uff1a<\/p>\n
GLM-5\u5728OpenCode\u4e0b\u7efc\u5408\u5f97\u5206\u4ece64.2\u72c2\u98d9\u523075.7\uff0c\u66b4\u6da811.5\u5206\uff1b\u5728Codex\u4e0b\u4e5f\u4ece64.4\u98d9\u5230\u4e8675.3\u3002<\/p>\n
Gemini 3.5 Flash\u5728OpenCode\u4e0b\u4ece72.4\u63d0\u5347\u523077.5\uff0c\u5728Codex\u4e0b\u4ece74.5\u63d0\u5347\u523077.5\u3002<\/p>\n
\u4f60\u770b\u8fd9\u4e2a\u6570\u636e\uff1a<\/p>\n
\u88ab\u4f18\u5316\u8fc7\u811a\u624b\u67b6\u7684GLM-5\uff0875.7\u5206 \/ 75.3\u5206\uff09\uff0c\u76f4\u63a5\u628a\u539f\u7248\u6ca1\u6539\u811a\u624b\u67b6\u7684Gemini 3.5 Flash\uff0872.4\u5206 \/ 74.5\u5206\uff09\u7ed9\u5e72\u7ffb\u4e86\u3002<\/p>\n
\u5728\u7aef\u5230\u7aef\u5c97\u4f4d\u4efb\u52a1\u7684JobBench\u4e0a\uff0c\u5355\u6b65\u6539\u8fdb\u540c\u6837\u5e26\u6765\u5168\u9762\u63d0\u5347\uff0cGemini 3.5 Flash\u6da8\u4e864.5\u5206\uff0cGLM-5\u6da8\u4e863.6\u5206\uff1b\u5728\u6cd5\u5f8b\u57fa\u51c6LAB\u4e0a\uff0cGLM-5\u7684\u90e8\u5206\u901a\u8fc7\u7387\u76f4\u63a5\u62d4\u9ad8\u4e8611.8\u4e2a\u767e\u5206\u70b9\u3002<\/p>\n
\u5982\u679c\u628a\u5e76\u884c\u9884\u7b97\u62c9\u5230 $k=5$\uff08\u5e76\u53d15\u8def\u5019\u9009\uff09\uff0c\u6027\u80fd\u8fd8\u4f1a\u8fdb\u4e00\u6b65\u4e0a\u626c\uff1aWorkBuddy-Bench\u6700\u9ad8\u518d\u6da83.9\u5206\uff0cJobBench\u518d\u6da84.2\u5206\uff0cLAB\u901a\u8fc7\u7387\u518d\u6da86.7\u4e2a\u767e\u5206\u70b9\u3002<\/p>\n
\u66f4\u7edd\u7684\u662f\u7ecf\u9a8c\u5e93\u7684\u8de8\u9886\u57df\u8fc1\u79fb\u6d4b\u8bd5\u3002<\/p>\n
\u56e2\u961f\u628a\u5728JobBench\u4e0a\u8dd1\u4efb\u52a1\u79ef\u7d2f\u4e0b\u6765\u7684\u7ecf\u9a8c\u5e93\uff0c\u76f4\u63a5\u62d4\u4e0b\u6765\u63d2\u5230WorkBuddy-Bench\u4e0a\u53bb\u7528\uff0c\u5728\u5b8c\u5168\u4e0d\u505a\u989d\u5916\u63a2\u6d4b\u7684\u60c5\u51b5\u4e0b\uff0c\u76f4\u63a5\u5e26\u6765\u4e86\u4ee3\u7801 +0.6\u3001\u7f51\u9875 +0.5\u3001\u529e\u516c +2.9\u7684\u65e0\u75db\u6da8\u5206\u3002<\/p>\n
\u5728\u8fc7\u53bb\u4e24\u5e74\u91cc\uff0c\u5927\u5bb6\u5bf9AI\u7684\u671f\u5f85\u51e0\u4e4e\u5168\u90e8\u62bc\u6ce8\u5728\u201c\u5e95\u5ea7\u6a21\u578bScaling\u201d\u4e0a\uff0c\u603b\u89c9\u5f97\u53ea\u8981\u7b49\u4e0b\u4e00\u4ee3GPT\u3001\u4e0b\u4e00\u4ee3Gemini\u51fa\u6765\uff0c\u624b\u5934\u7684\u5e94\u7528\u95ee\u9898\u5c31\u8fce\u5203\u800c\u751f\u4e86\u3002<\/p>\n
\u4f46\u73b0\u5b9e\u7ed9\u6240\u6709\u4ece\u4e1a\u8005\u6cfc\u4e86\u4e00\u76c6\u51b7\u6c34\uff1a\u57fa\u5ea7\u6a21\u578b\u8d8a\u6765\u8d8a\u8d35\uff0c\u8fb9\u9645\u6548\u76ca\u5728\u653e\u7f13\uff0c\u800c\u843d\u5730\u573a\u666f\u91cc\u7684\u957f\u5c3ebug\u4f9d\u7136\u65e0\u7a77\u65e0\u5c3d\u3002<\/p>\n
\u4f60\u7ed9\u6a21\u578b\u518d\u5927\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u5b83\u5728\u5904\u7406\u8de8\u6587\u4ef6\u5bf9\u8d26\u65f6\u4f9d\u7136\u4f1a\u5077\u61d2\uff1b\u4f60\u7ed9\u5b83\u518d\u9ad8\u7684\u8dd1\u5206\uff0c\u5b83\u5199\u5b8c\u62a5\u544a\u4f9d\u7136\u4e0d\u4f1a\u4e3b\u52a8\u6253\u5f00\u6587\u4ef6\u53bb\u6821\u9a8c\u683c\u5f0f\u3002<\/p>\n
\u8fd9\u4e9b\u4e1c\u897f\uff0c\u9760\u9884\u8bad\u7ec3\u53c2\u6570\u662f\u6c38\u8fdc\u5582\u4e0d\u9971\u7684\u3002<\/p>\n
\u5b83\u9700\u8981\u7684\u662f\u4e00\u5957\u80fd\u5728\u8fd0\u884c\u4e2d\u81ea\u6211\u611f\u77e5\u3001\u81ea\u6211\u4fee\u8865\u3001\u6301\u7eed\u8fdb\u5316\u7684Runtime\u8ba4\u77e5\u5de5\u7a0b\u3002<\/p>\n
\u4ee5\u524d\u6211\u4eec\u642d\u811a\u624b\u67b6\uff0c\u662f\u5de5\u7a0b\u5e08\u5750\u5728\u7535\u8111\u524d\u4e00\u6761\u4e00\u6761\u624b\u5199Prompt\u3001\u914d\u5de5\u5177\u3001\u4fee\u7ba1\u9053\uff0c\u7d2f\u5f97\u534a\u6b7b\u8fd8\u5230\u5904\u6f0f\u98ce\u3002<\/p>\n
\u73b0\u5728Harness\u00b2 \u8bc1\u660e\u4e86\u4e00\u6761\u65b0\u8def\uff1a<\/p>\n
\u628a\u89c4\u5219\u5236\u5b9a\u7684\u6743\u529b\uff0c\u8fd8\u7ed9Agent\u81ea\u5df1\u3002<\/p>\n
\u6700\u540e\uff0c\u7167\u4f8b\u6765\u8bf4\u8bf4\u8bba\u6587\u4f5c\u8005\u3002<\/p>\n
\u4e00\u51719\u4f4d\uff0c\u800c\u4e14\u534e\u4eba\u6d53\u5ea6\u6781\u9ad8\uff0c\u9ad8\u8fbe78%\u3002\u3002\u3002<\/p>\n
\u7531\u4e8e\u6587\u7ae0\u7bc7\u5e45\u6709\u9650\uff0c\u8fd9\u91cc\u5c31\u6309\u987a\u5e8f\u4ecb\u7ecd\u4e09\u4f4d\u3002<\/p>\n
Ruiyao Xu\uff0c\u8bba\u6587\u4e00\u4f5c\u517c\u901a\u8baf\u4f5c\u8005\u3002<\/p>\n
\u7f8e\u56fd\u897f\u5317\u5927\u5b66\u7edf\u8ba1\u4e0e\u6570\u636e\u79d1\u5b66\u7cfb\u535a\u58eb\u751f\uff0c\u5bfc\u5e08\u4e3aKaize Ding\u3002<\/p>\n
\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u5305\u62ec\u5927\u8bed\u8a00\u6a21\u578b\u3001\u56fe\u673a\u5668\u5b66\u4e60\u3001\u5f02\u5e38\u68c0\u6d4b\u53caAI Agent\u7cfb\u7edf\u3002<\/p>\n
\u8fd9\u7bc7\u8bba\u6587\uff0c\u4fbf\u662f\u4ed6\u5728\u8c37\u6b4c\u5b9e\u4e60\u671f\u95f4\u7684\u5de5\u4f5c\u3002<\/p>\n
Yanfei Chen\uff0cGoogle Cloud AI Research\u8d44\u6df1\u7814\u7a76\u5de5\u7a0b\u5e08\u3002<\/p>\n
\u672c\u79d1\u6bd5\u4e1a\u4e8e\u5317\u4eac\u5927\u5b66\uff0c\u4e3b\u4fee\u7535\u6c14\u5de5\u7a0b\u4e0e\u7ecf\u6d4e\u5b66\uff0c2018\u5e74\u83b7\u5f97\u7f8e\u56fd\u5339\u5179\u5821\u5927\u5b66\u5de5\u4e1a\u5de5\u7a0b\u535a\u58eb\u5b66\u4f4d\u3002<\/p>\n
\u6b64\u540e\u66fe\u5728\u673a\u5668\u5b66\u4e60\u516c\u53f8Qeexo\u62c5\u4efb\u7814\u7a76\u5de5\u7a0b\u5e08\uff0c2020\u5e74\u52a0\u5165Google\uff0c\u53c2\u4e0e\u8fc7Vertex AI Search\u3001Gemini Enterprise Deep Research\u7b49\u9879\u76ee\uff0c\u4e3b\u8981\u7814\u7a76\u5927\u8bed\u8a00\u6a21\u578b\u667a\u80fd\u4f53\u3001\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u53ca\u5de5\u5177\u5b66\u4e60\u3002<\/p>\n
Zhongying CuiZhu\uff0c\u4e5f\u53ebZoey CuiZhu\uff0cGoogle Cloud AI Research\u7814\u7a76\u5de5\u7a0b\u5e08\u3002<\/p>\n
2019\u5e74\u83b7\u5f97\u65af\u5766\u798f\u5927\u5b66\u7535\u6c14\u5de5\u7a0b\u535a\u58eb\u5b66\u4f4d\uff0c\u4e3b\u8981\u7814\u7a76\u901a\u7528\u667a\u80fd\u4f53\u3001\u53ef\u9a8c\u8bc1\u667a\u80fd\u4f53\u53ca\u667a\u80fd\u4f53\u8d28\u91cf\u8bc4\u4f30\u3002<\/p>\n
\u53c2\u8003\u94fe\u63a5\uff1a[1]https:\/\/arxiv.org\/html\/2610.14224v1<\/p>\n
\n\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n
\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n
\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n
\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"
Harness\u5c42\u7684RSI<\/p>\n
\u8c37\u6b4cRSI\u65b0\u6210\u679c\uff0c\u5f00\u6e90\u4e86\u3002<\/p>\n
\u540d\u53ebHarness\u00b2\uff0c\u8ba9\u5927\u6a21\u578b\u5728\u8fd0\u884c\u4e2d\u81ea\u5df1\u7ed9\u81ea\u5df1\u9b54\u6539\u5916\u6302\u4ee3\u7801\u3001\u7cfb\u7edf\u63d0\u793a\u8bcd\u3001\u5de5\u5177\u94fe\u548c\u6280\u80fd\u5e93\u3002<\/p>\n
\u662f\u7684\uff0cHarness\u5c42\u7684RSI\u3002<\/p>\n
\u6548\u679c\u4e5f\u662f\u5f88\u731b\u3002<\/p>\n
\u88c5\u914d\u4e0a\u8fd9\u4ef6RSI\u4e13\u7cbe\u7684GLM-5\uff0c\u5728\u529e\u516c\u57fa\u51c6WorkBuddy-Bench\u4e0a\uff0c\u76f4\u63a5\u98d9\u5347\u4e8611.5\u5206\u2014\u2014<\/p>\n
\u77ac\u95f4\u53cd\u8d85\u539f\u751fGemini 3.5 Flash\u3002<\/p>\n
\u603b\u4e4b\u5c31\u662f\u4e00\u4e2a\u975e\u5e38\u79bb\u8c31\u7684\u5916\u6302\uff0c\u4efb\u4f55\u6a21\u578b\u5957<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[17],"class_list":["post-9099","post","type-post","status-publish","format-standard","hentry","category-xwzx","tag-17"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/9099","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=9099"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/9099\/revisions"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=9099"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=9099"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=9099"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}