{"id":1402,"date":"2026-06-01T10:39:20","date_gmt":"2026-06-01T02:39:20","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1402"},"modified":"2026-06-01T10:39:22","modified_gmt":"2026-06-01T02:39:22","slug":"skillopt-executive-strategy-forself-evolving-agent-skills","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1402","title":{"rendered":"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\"><strong>1. \u8bba\u6587\u6982\u8ff0\u4e0e\u5f15\u8a00<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SKILLOPT\u63d0\u51fa\u4e86\u4e00\u79cd\u5f00\u521b\u6027\u7684\u6587\u672c\u7a7a\u95f4\u4f18\u5316\u6846\u67b6\uff0c\u7528\u4e8e\u5b9e\u73b0\u4ee3\u7406\u6280\u80fd\u7684\u81ea\u8fdb\u5316\u3002\u8be5\u8bba\u6587\u6df1\u523b\u6307\u51fa\uff0c\u5f53\u4eca\u4ee3\u7406\u6280\u80fd\u4e3b\u8981\u4f9d\u8d56\u624b\u5de5 crafting\u3001\u5355\u6b21 LLM \u751f\u6210\u6216\u677e\u6563\u7684\u81ea\u4fee\u8ba2\u5faa\u73af\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u5747\u65e0\u6cd5\u50cf\u6df1\u5ea6\u5b66\u4e60\u4f18\u5316\u5668\u90a3\u6837\u63d0\u4f9b\u7cfb\u7edf\u3001\u53ef\u63a7\u4e14\u53ef\u91cd\u590d\u7684\u6539\u8fdb\u8def\u5f84\u3002\u4f5c\u8005\u4e3b\u5f20\uff0c\u5e94\u5c06\u6280\u80fd\u6587\u6863\u89c6\u4e3a\u51bb\u7ed3\u4ee3\u7406\u6a21\u578b\u7684\u5916\u90e8\u53ef\u8bad\u7ec3\u72b6\u6001\uff0c\u5f15\u5165\u72ec\u7acb\u7684 frontier \u4f18\u5316\u5668\u6a21\u578b\u3002\u8be5\u4f18\u5316\u5668\u901a\u8fc7\u5206\u6790\u5e26\u5206\u6570\u7684 rollout \u8f68\u8ff9\uff0c\u751f\u6210\u53d7\u9650\u7684 add\/delete\/replace \u7f16\u8f91\u64cd\u4f5c\uff0c\u5e76\u4ec5\u5728\u7f16\u8f91\u80fd\u4e25\u683c\u63d0\u5347 held-out \u9a8c\u8bc1\u96c6\u5f97\u5206\u65f6\u624d\u63a5\u53d7\u3002\u8fd9\u79cd\u4e25\u683c\u7684\u9a8c\u8bc1\u95e8\u63a7\u673a\u5236\uff0c\u5b8c\u7f8e\u6a21\u62df\u4e86\u6743\u91cd\u7a7a\u95f4\u4f18\u5316\u7684\u7eaa\u5f8b\u6027\u548c\u7a33\u5b9a\u6027\uff0c\u907f\u514d\u4e86\u4f20\u7edf\u63d0\u793a\u5de5\u7a0b\u4e2d\u5e38\u89c1\u7684\u6f02\u79fb\u548c\u9000\u5316\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u5f3a\u8c03\uff0c\u4ee3\u7406\u6280\u80fd\u4f5c\u4e3a\u9886\u57df\u9002\u5e94\u7684\u6838\u5fc3\u63a5\u53e3\uff0c\u80fd\u591f\u7cfb\u7edf\u6027\u5730\u5c01\u88c5\u7a0b\u5e8f\u6027\u77e5\u8bc6\u3001\u9886\u57df\u7279\u5b9a\u542f\u53d1\u5f0f\u3001\u5de5\u5177\u4f7f\u7528\u7b56\u7565\u3001\u8f93\u51fa\u683c\u5f0f\u7ea6\u675f\u4ee5\u53ca\u5e38\u89c1\u5931\u8d25\u6a21\u5f0f\uff0c\u4ece\u800c\u8ba9\u51bb\u7ed3\u7684\u5e95\u5c42\u6a21\u578b\u901a\u8fc7\u5916\u90e8\u6587\u672c\u5b9e\u73b0\u9ad8\u6548\u9002\u5e94\u3002SKILLOPT \u521b\u65b0\u6027\u5730\u5f15\u5165\u4e86\u6587\u672c\u5f62\u5f0f\u7684 learning-rate \u9884\u7b97\u3001\u62d2\u7edd\u7f16\u8f91\u7f13\u51b2\u533a\u4ee5\u53ca\u9010 epoch \u7684\u6162\/\u5143\u66f4\u65b0\u673a\u5236\uff0c\u786e\u4fdd\u6574\u4e2a\u6280\u80fd\u8bad\u7ec3\u8fc7\u7a0b\u65e2\u7a33\u5b9a\u53c8\u9ad8\u6548\u3002\u5728\u516d\u4e2a\u591a\u6837\u5316\u57fa\u51c6\uff08\u6db5\u76d6 QA\u3001\u7535\u5b50\u8868\u683c\u3001\u6587\u6863\u3001\u591a\u6a21\u6001\u3001\u6570\u5b66\u548c\u5177\u8eab\u51b3\u7b56\uff09\u3001\u4e03\u4e2a\u4e0d\u540c\u89c4\u6a21\u7684\u76ee\u6807\u6a21\u578b\u4ee5\u53ca\u4e09\u79cd\u6267\u884c harness\uff08\u76f4\u63a5\u804a\u5929\u3001Codex\u3001Claude Code\uff09\u4e0a\uff0cSKILLOPT \u5728\u5168\u90e8 52 \u4e2a\u8bc4\u4f30\u5355\u5143\u4e2d\u5747\u53d6\u5f97\u6700\u4f73\u6216\u5e76\u5217\u6700\u4f73\u6210\u7ee9\u3002\u5728 GPT-5.5 \u76f4\u63a5\u804a\u5929\u6a21\u5f0f\u4e0b\uff0c\u5e73\u5747\u51c6\u786e\u7387\u63d0\u5347\u9ad8\u8fbe 23.5 \u4e2a\u767e\u5206\u70b9\uff0c\u5145\u5206\u5c55\u793a\u4e86\u5176\u5728\u5b9e\u9645\u4ee3\u7406\u4efb\u52a1\u4e2d\u7684\u5f3a\u5927\u6f5c\u529b\u3002\u8be5\u65b9\u6cd5\u7684\u6838\u5fc3\u4ef7\u503c\u5728\u4e8e\u5c06\u6280\u80fd\u8fdb\u5316\u8f6c\u5316\u4e3a\u4e00\u4e2a\u53ef\u4e25\u683c\u9a8c\u8bc1\u7684\u4f18\u5316\u5faa\u73af\uff0c\u800c\u975e\u4f20\u7edf\u7684 ad-hoc \u63d0\u793a\u4fee\u6539\uff0c\u8fd9\u4e3a\u672a\u6765\u4ee3\u7406\u7cfb\u7edf\u7684\u6301\u7eed\u6539\u8fdb\u63d0\u4f9b\u4e86\u575a\u5b9e\u7684\u7406\u8bba\u548c\u5b9e\u8df5\u57fa\u7840\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SKILLOPT \u7684\u53e6\u4e00\u91cd\u5927\u521b\u65b0\u5728\u4e8e\u5176\u96f6\u63a8\u7406\u65f6\u5f00\u9500\u8bbe\u8ba1\u3002\u4f18\u5316\u8fc7\u7a0b\u5b8c\u5168\u79bb\u7ebf\u8fdb\u884c\uff0c\u6700\u7ec8\u4ec5\u5bfc\u51fa\u4e00\u4e2a\u7d27\u51d1\u7684 best_skill.md \u6587\u4ef6\uff0c\u90e8\u7f72\u65f6\u65e0\u9700\u989d\u5916\u8c03\u7528\u4f18\u5316\u5668\u6a21\u578b\u3002\u8f6c\u79fb\u5b9e\u9a8c\u8fdb\u4e00\u6b65\u5f3a\u5316\u4e86\u5176\u5b9e\u7528\u6027\uff1a\u4f18\u5316\u540e\u7684\u6280\u80fd\u80fd\u5728\u4e0d\u540c\u6a21\u578b\u89c4\u6a21\u4e4b\u95f4\u3001\u4e0d\u540c\u6267\u884c\u73af\u5883\u4e4b\u95f4\u4ee5\u53ca\u76f8\u8fd1\u57fa\u51c6\u4efb\u52a1\u4e0a\u4fdd\u6301\u663e\u8457\u6b63\u5411\u8fc1\u79fb\u6548\u679c\uff0c\u8fd9\u5927\u5927\u964d\u4f4e\u4e86\u5b9e\u9645\u751f\u4ea7\u73af\u5883\u4e2d\u9886\u57df\u9002\u5e94\u7684\u6210\u672c\uff0c\u4e3a\u4f01\u4e1a\u7ea7\u4ee3\u7406\u90e8\u7f72\u63d0\u4f9b\u4e86\u53ef\u6269\u5c55\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"470\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/1-1024x470.png\"  class=\"wp-image-1403\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/1-1024x470.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/1-300x138.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/1-768x352.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/1-1536x704.png 1536w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/1.png 1701w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe\" alt=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>2. \u65b9\u6cd5\u8bba\u6838\u5fc3\u8bbe\u8ba1<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SKILLOPT \u7684\u4f18\u5316\u5faa\u73af\u7531\u591a\u4e2a\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u9636\u6bb5\u6784\u6210\uff0c\u5305\u62ec\u524d\u5411 rollout \u8bc1\u636e\u6536\u96c6\u3001\u540e\u5411\u5c0f\u6279\u91cf\u53cd\u601d\u3001\u6709\u754c\u6587\u672c\u66f4\u65b0\u3001\u9a8c\u8bc1\u95e8\u63a7\u4ee5\u53ca\u9010\u7eaa\u5143\u6162\/\u5143\u66f4\u65b0\u3002\u5728\u524d\u5411\u9636\u6bb5\uff0c\u76ee\u6807\u6a21\u578b\u4f7f\u7528\u5f53\u524d\u7248\u672c\u7684\u6280\u80fd\u5728\u8bad\u7ec3\u96c6\u4e0a\u6267\u884c\u6279\u91cf\u4efb\u52a1\uff0charness \u8be6\u7ec6\u8bb0\u5f55\u8f68\u8ff9\u4e2d\u7684\u5143\u6570\u636e\u3001\u5de5\u5177\u8c03\u7528\u3001\u89c2\u5bdf\u7ed3\u679c\u3001\u6700\u7ec8\u7b54\u6848\u548c\u9a8c\u8bc1\u5668\u53cd\u9988\uff0c\u4e3a\u540e\u7eed\u4f18\u5316\u63d0\u4f9b\u4e30\u5bcc\u4e14\u7ed3\u6784\u5316\u7684\u8bc1\u636e\u3002\u540e\u5411\u53cd\u601d\u9636\u6bb5\uff0c\u4f18\u5316\u5668\u6a21\u578b\u9996\u5148\u5c06\u6210\u529f\u548c\u5931\u8d25\u8f68\u8ff9\u5206\u79bb\uff0c\u5e76\u8fdb\u4e00\u6b65\u5212\u5206\u4e3a\u5c0f\u6279\u91cf\u8fdb\u884c\u5206\u6790\u3002\u8fd9\u79cd\u5c0f\u6279\u91cf\u5904\u7406\u65b9\u5f0f\u80fd\u6709\u6548\u53d1\u73b0\u7cfb\u7edf\u6027\u7684\u7a0b\u5e8f\u6027\u9519\u8bef\uff0c\u800c\u975e\u4ec5\u9488\u5bf9\u5355\u4e2a\u5b64\u7acb\u6848\u4f8b\u63d0\u51fa\u4e34\u65f6\u4fee\u590d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"560\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/2-1024x560.png\"  class=\"wp-image-1404\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/2-1024x560.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/2-300x164.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/2-768x420.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/2-1536x839.png 1536w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/2.png 1614w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe1\" alt=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u6709\u754c\u6587\u672c\u66f4\u65b0\u662f SKILLOPT \u6700\u6838\u5fc3\u7684\u521b\u65b0\u4e4b\u4e00\u3002\u901a\u8fc7\u5f15\u5165\u6587\u672c\u5b66\u4e60\u7387 Lt\uff08\u6bcf\u6b21\u6b65\u9aa4\u5141\u8bb8\u7684\u6700\u5927\u7f16\u8f91\u6570\u91cf\uff09\uff0c\u7cfb\u7edf\u6709\u6548\u9632\u6b62\u4e86\u65e0\u754c\u91cd\u5199\u53ef\u80fd\u5e26\u6765\u7684\u6709\u7528\u89c4\u5219\u64e6\u9664\u6216\u4e0d\u517c\u5bb9\u6307\u4ee4\u5f15\u5165\u7684\u95ee\u9898\u3002\u4f18\u5316\u5668\u4f1a\u5bf9\u5408\u5e76\u540e\u7684\u7f16\u8f91\u6c60\u8fdb\u884c\u6392\u540d\uff0c\u5e76\u4e25\u683c\u9650\u5236\u5728\u5f53\u524d\u9884\u7b97\u5185\u3002\u5019\u9009\u6280\u80fd\u751f\u6210\u540e\uff0c\u5fc5\u987b\u901a\u8fc7 held-out \u9a8c\u8bc1\u96c6\u7684\u4e25\u683c\u8bc4\u4f30\uff08\u5f97\u5206\u9700\u4e25\u683c\u9ad8\u4e8e\u5f53\u524d\u6700\u4f73\uff09\uff0c\u624d\u80fd\u88ab\u6b63\u5f0f\u63a5\u53d7\uff1b\u5426\u5219\u8fdb\u5165\u62d2\u7edd\u7f13\u51b2\u533a\uff0c\u4f5c\u4e3a\u540e\u7eed\u4f18\u5316\u7684\u8d1f\u53cd\u9988\u4fe1\u53f7\u3002\u9010\u7eaa\u5143\u6162\/\u5143\u66f4\u65b0\u5219\u8d1f\u8d23\u6355\u6349\u8de8 epoch \u7684\u957f\u671f\u6a21\u5f0f\uff0c\u901a\u8fc7\u5bf9\u6bd4\u524d\u540e epoch \u7684\u8868\u73b0\uff0c\u751f\u6210\u4fdd\u62a4\u5b57\u6bb5\u4e2d\u7684\u7eb5\u5411\u6307\u5bfc\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4f18\u5316\u7684\u957f\u671f\u7a33\u5b9a\u6027\u548c\u6cdb\u5316\u80fd\u529b\u3002\u8fd9\u79cd\u591a\u5c42\u6b21\u63a7\u5236\u673a\u5236\uff0c\u4f7f\u5f97\u6280\u80fd\u4f18\u5316\u8fc7\u7a0b\u771f\u6b63\u5177\u5907\u4e86\u7c7b\u4f3c\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u7684\u7cfb\u7edf\u6027\u3001\u53ef\u89e3\u91ca\u6027\u548c\u9c81\u68d2\u6027\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"765\" height=\"996\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/3.png\"  class=\"wp-image-1405\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/3.png 765w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/3-230x300.png 230w\" sizes=\"auto, (max-width: 765px) 100vw, 765px\" title=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe2\" alt=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u5916\uff0c\u65b9\u6cd5\u7279\u522b\u6ce8\u91cd harness \u65e0\u5173\u6027\u548c\u53ef\u79fb\u690d\u6027\u3002\u901a\u8fc7\u8f7b\u91cf\u7ea7\u9002\u914d\u5668\u63a5\u53e3\uff0c\u540c\u4e00\u4f18\u5316\u6846\u67b6\u53ef\u65e0\u7f1d\u652f\u6301\u76f4\u63a5\u804a\u5929\u3001Codex \u98ce\u683c\u4ee3\u7801\u6267\u884c\u548c Claude Code \u73af\u5883\u3002\u8fd9\u79cd\u8bbe\u8ba1\u5f7b\u5e95\u5c06\u6280\u80fd\u5b9a\u4f4d\u4e3a\u53ef\u590d\u7528\u7684\u9886\u57df\u9002\u5e94\u5c42\uff1a\u5f3a\u5927\u7684\u4f18\u5316\u5668\u4ec5\u5728\u8bad\u7ec3\u9636\u6bb5\u53d1\u6325\u4f5c\u7528\uff0c\u90e8\u7f72\u65f6\u4ec5\u9700\u52a0\u8f7d\u7cbe\u70bc\u540e\u7684\u6280\u80fd\u6587\u6863\u5373\u53ef\u663e\u8457\u63d0\u5347\u4efb\u610f\u76ee\u6807\u6a21\u578b\u7684\u8868\u73b0\uff0c\u6781\u5927\u964d\u4f4e\u4e86\u5b9e\u9645\u5e94\u7528\u95e8\u69db\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>3. \u5b9e\u9a8c\u7ed3\u679c\u4e0e\u6027\u80fd\u8868\u73b0<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u8bbe\u8ba1\u5168\u9762\u4e14\u4e25\u8c28\uff0c\u6db5\u76d6 SearchQA\u3001SpreadsheetBench\u3001OfficeQA\u3001DocVQA\u3001LiveMathematicianBench \u548c ALFWorld \u516d\u4e2a\u4ee3\u8868\u6027\u57fa\u51c6\uff0c\u6a2a\u8de8 GPT \u7cfb\u5217\u524d\u6cbf\u6a21\u578b\u5230 Qwen \u5c0f\u6a21\u578b\uff0c\u5e76\u5728\u76f4\u63a5\u804a\u5929\u3001Codex \u4ee5\u53ca Claude Code \u4e09\u79cd\u6267\u884c\u6a21\u5f0f\u4e0b\u8fdb\u884c\u8bc4\u4f30\u3002SKILLOPT \u5728\u5168\u90e8 52 \u4e2a\uff08\u6a21\u578b\u3001\u57fa\u51c6\u3001harness\uff09\u8bc4\u4f30\u5355\u5143\u4e2d\u5747\u53d6\u5f97\u6700\u4f73\u6216\u5e76\u5217\u6700\u4f73\u6210\u7ee9\uff0c\u5168\u9762\u8d85\u8d8a\u4eba\u7c7b\u624b\u5de5\u6280\u80fd\u3001\u5355\u6b21 LLM \u751f\u6210\u6280\u80fd\u3001Trace2Skill\u3001TextGrad\u3001GEPA \u4ee5\u53ca EvoSkill \u7b49\u4e3b\u6d41\u57fa\u7ebf\uff0c\u5c55\u73b0\u51fa\u538b\u5012\u6027\u4f18\u52bf\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"783\" height=\"561\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/4.png\"  class=\"wp-image-1406\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/4.png 783w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/4-300x215.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/4-768x550.png 768w\" sizes=\"auto, (max-width: 783px) 100vw, 783px\" title=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe3\" alt=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe3\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7279\u522b\u662f\u5728 GPT-5.5 \u76f4\u63a5\u804a\u5929\u6a21\u5f0f\u4e0b\uff0cSKILLOPT \u5c06\u5e73\u5747\u6027\u80fd\u4ece\u65e0\u6280\u80fd\u7684 58.8 \u63d0\u5347\u81f3 82.3\uff0c\u51c0\u589e 23.5 \u4e2a\u767e\u5206\u70b9\u3002\u5176\u4e2d\u7a0b\u5e8f\u5316\u4efb\u52a1\u53d7\u76ca\u6700\u4e3a\u663e\u8457\uff1aSpreadsheetBench \u4ece 41.8 \u8dc3\u5347\u81f3 80.7\uff0cOfficeQA \u4ece 33.1 \u63d0\u5347\u81f3 72.1\uff0cLiveMathematicianBench \u4ece 37.6 \u5347\u81f3 66.9\u3002\u8fd9\u4e9b\u5927\u5e45\u63d0\u5347\u5145\u5206\u8bf4\u660e\uff0cSKILLOPT \u80fd\u6709\u6548\u6ce8\u5165\u524d\u6cbf\u6a21\u578b\u6b20\u7f3a\u7684\u7a0b\u5e8f\u6027\u7eaa\u5f8b\u548c\u9886\u57df\u77e5\u8bc6\u3002\u5728 Codex \u548c Claude Code \u7b49\u5de5\u5177\u589e\u5f3a\u6267\u884c\u73af\u5883\u4e2d\uff0c\u540c\u6837\u5206\u522b\u53d6\u5f97 +24.8 \u548c +19.1 \u7684\u5e73\u5747\u63d0\u5347\uff0c\u8bc1\u660e\u8be5\u65b9\u6cd5\u5bf9\u590d\u6742\u591a\u8f6e\u4ea4\u4e92\u573a\u666f\u540c\u6837\u9002\u7528\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"765\" height=\"420\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/5.png\"  class=\"wp-image-1407\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/5.png 765w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/5-300x165.png 300w\" sizes=\"auto, (max-width: 765px) 100vw, 765px\" title=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe4\" alt=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe4\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5c0f\u89c4\u6a21\u548c\u8f83\u5f31\u6a21\u578b\u4ece\u4f18\u5316\u6280\u80fd\u4e2d\u83b7\u76ca\u5c24\u4e3a\u660e\u663e\u3002\u4f8b\u5982 GPT-5.4-nano \u5728\u591a\u4e2a\u57fa\u51c6\u4e0a\u5b9e\u73b0\u4e86\u63a5\u8fd1\u7ffb\u500d\u7684\u6027\u80fd\u8dc3\u5347\u3002\u8fd9\u8868\u660e\u7d27\u51d1\u7684\u6280\u80fd\u6587\u6863\u80fd\u6709\u6548\u8865\u5145\u5c0f\u6a21\u578b\u6743\u91cd\u4e2d\u7f3a\u5931\u7684\u7a0b\u5e8f\u6027\u77e5\u8bc6\u3002\u8de8\u6a21\u578b\u3001\u8de8 harness \u548c\u8de8\u57fa\u51c6\u7684\u8f6c\u79fb\u5b9e\u9a8c\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86\u6280\u80fd\u7684\u9c81\u68d2\u6027\u548c\u590d\u7528\u4ef7\u503c\uff0c\u4e3a\u5b9e\u9645\u90e8\u7f72\u63d0\u4f9b\u4e86\u5f3a\u6709\u529b\u7684\u652f\u6491\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>4. \u6d88\u878d\u5206\u6790\u4e0e\u6280\u80fd\u7279\u6027<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5168\u9762\u7684\u6d88\u878d\u5b9e\u9a8c\u6709\u529b\u8bc1\u660e\u4e86 SKILLOPT \u5404\u7ec4\u4ef6\u7684\u5173\u952e\u4f5c\u7528\u3002\u6709\u754c\u5b66\u4e60\u7387\u3001\u62d2\u7edd\u7f16\u8f91\u7f13\u51b2\u3001\u4e25\u683c\u9a8c\u8bc1\u95e8\u63a7\u4ee5\u53ca\u9010\u7eaa\u5143\u6162\/\u5143\u66f4\u65b0\u5171\u540c\u6784\u6210\u4e86\u6027\u80fd\u63d0\u5347\u7684\u6838\u5fc3\u3002\u5176\u4e2d\uff0c\u79fb\u9664\u6162\/\u5143\u66f4\u65b0\u5728 SpreadsheetBench \u4e0a\u5bfc\u81f4\u4e86\u9ad8\u8fbe 22.5 \u4e2a\u767e\u5206\u70b9\u7684\u5267\u70c8\u6027\u80fd\u4e0b\u964d\uff0c\u51f8\u663e\u957f\u671f\u8de8 epoch \u6307\u5bfc\u5bf9\u4e8e\u6355\u6349\u6301\u4e45\u9886\u57df\u6a21\u5f0f\u7684\u91cd\u8981\u6027\u3002\u5176\u4ed6\u7ec4\u4ef6\u7684\u79fb\u9664\u540c\u6837\u4f1a\u5e26\u6765\u663e\u8457\u9000\u5316\uff0c\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86\u6574\u4e2a\u4f18\u5316\u6846\u67b6\u8bbe\u8ba1\u7684\u7cfb\u7edf\u6027\u548c\u5fc5\u8981\u6027\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"717\" height=\"291\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/6.png\"  class=\"wp-image-1408\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/6.png 717w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/6-300x122.png 300w\" sizes=\"auto, (max-width: 717px) 100vw, 717px\" title=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe5\" alt=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe5\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u7ec8\u5b66\u5230\u7684\u6280\u80fd\u6587\u6863\u8868\u73b0\u51fa\u6781\u4f73\u7684\u7d27\u51d1\u6027\u548c\u7f16\u8f91\u7ecf\u6d4e\u6027\uff1a\u957f\u5ea6\u7a33\u5b9a\u63a7\u5236\u5728 300\u20132000 token \u4e4b\u95f4\uff0c\u901a\u5e38\u4ec5\u9700 1\u20134 \u6b21\u63a5\u53d7\u7684\u7f16\u8f91\u5373\u53ef\u5b9e\u73b0\u5927\u5e45\u6027\u80fd\u63d0\u5347\u3002\u8fd9\u79cd\u9ad8\u6548\u7387\u6e90\u4e8e\u9a8c\u8bc1\u95e8\u63a7\u5bf9\u4f4e\u8d28\u91cf\u63d0\u8bae\u7684\u4e25\u683c\u8fc7\u6ee4\uff0c\u5927\u91cf\u6f5c\u5728\u7f16\u8f91\u88ab\u62d2\u7edd\u5e76\u8f6c\u5316\u4e3a\u8d1f\u53cd\u9988\uff0c\u4fdd\u969c\u4e86\u90e8\u7f72\u6280\u80fd\u7684\u9ad8\u8d28\u91cf\u3002\u5b66\u5230\u7684\u89c4\u5219\u9ad8\u5ea6\u7a0b\u5e8f\u5316\uff0c\u805a\u7126\u4e8e\u901a\u7528\u7b56\u7565\u800c\u975e\u5177\u4f53\u5b9e\u4f8b\uff0c\u4f8b\u5982\u8981\u6c42\u7cfb\u7edf\u68c0\u67e5\u5de5\u4f5c\u7c3f\u7ed3\u6784\u3001\u4e25\u683c\u7ed1\u5b9a\u89c6\u89c9\u8bc1\u636e\u3001\u7ef4\u62a4\u641c\u7d22\u524d\u6cbf\u8bb0\u5fc6\u7b49\u3002\u8fd9\u4e9b\u89c4\u5219\u4e0d\u4ec5\u53ef\u8bfb\u6027\u5f3a\u3001\u4fbf\u4e8e\u4eba\u5de5\u5ba1\u8ba1\uff0c\u8fd8\u4f53\u73b0\u4e86\u63a5\u8fd1\u4eba\u7c7b\u4e13\u5bb6\u6c34\u5e73\u7684\u6d1e\u5bdf\u529b\u3002\u8bad\u7ec3\u6210\u672c\u5206\u6790\u663e\u793a\uff0c\u867d\u7136\u4e0d\u540c\u57fa\u51c6\u7684 token \u6d88\u8017\u6709\u6240\u5dee\u5f02\uff0c\u4f46\u6574\u4f53\u53ef\u63a7\uff0c\u4e14\u4e00\u6b21\u6027\u8bad\u7ec3\u540e\u5373\u53ef\u957f\u671f\u590d\u7528\uff0c\u6027\u4ef7\u6bd4\u6781\u9ad8\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"705\" height=\"461\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/7.png\"  class=\"wp-image-1409\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/7.png 705w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/7-300x196.png 300w\" sizes=\"auto, (max-width: 705px) 100vw, 705px\" title=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe6\" alt=\"SkILLOPT: Executive Strategy forSelf-Evolving Agent Skills\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>5. \u7ed3\u8bba\u4e0e\u5c55\u671b<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SKILLOPT \u4e3a LLM \u4ee3\u7406\u6280\u80fd\u4f18\u5316\u63d0\u4f9b\u4e86\u9996\u4e2a\u7cfb\u7edf\u6027\u3001\u53ef\u63a7\u7684\u6587\u672c\u7a7a\u95f4\u4f18\u5316\u6846\u67b6\u3002\u5b83\u6210\u529f\u8bc1\u660e\uff0c\u901a\u8fc7\u5c06\u6280\u80fd\u6587\u6863\u4f5c\u4e3a\u5916\u90e8\u53ef\u8bad\u7ec3\u72b6\u6001\uff0c\u5e76\u5e94\u7528\u7c7b\u4f3c\u6df1\u5ea6\u5b66\u4e60\u7684\u63a7\u5236\u673a\u5236\uff08\u5b66\u4e60\u7387\u3001\u9a8c\u8bc1\u95e8\u63a7\u3001\u6162\u66f4\u65b0\u7b49\uff09\uff0c\u80fd\u591f\u5728\u4e0d\u4fee\u6539\u6a21\u578b\u6743\u91cd\u7684\u524d\u63d0\u4e0b\u5b9e\u73b0\u663e\u8457\u4e14\u7a33\u5b9a\u7684\u6027\u80fd\u63d0\u5347\u3002\u8be5\u65b9\u6cd5\u5728\u5e7f\u6cdb\u5b9e\u9a8c\u4e2d\u5c55\u73b0\u7684\u4f18\u8d8a\u6027\u3001\u8f6c\u79fb\u80fd\u529b\u548c\u53ef\u89e3\u91ca\u6027\uff0c\u4e3a\u4ee3\u7406\u7cfb\u7edf\u7684\u5b9e\u9645\u90e8\u7f72\u5f00\u8f9f\u4e86\u5168\u65b0\u8def\u5f84\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c55\u671b\u672a\u6765\uff0c\u53ef\u8fdb\u4e00\u6b65\u6269\u5c55\u81f3\u591a\u6280\u80fd\u5e93\u534f\u540c\u8fdb\u5316\u3001\u504f\u597d\u9a71\u52a8\u6216\u65e0\u5956\u52b1\u9a8c\u8bc1\u673a\u5236\uff0c\u4ee5\u53ca\u5c06\u4f18\u5316\u6280\u80fd\u81ea\u84b8\u998f\u56de\u76ee\u6807\u6a21\u578b\u4ee5\u5b9e\u73b0\u6df7\u5408\u9002\u5e94\u7b49\u65b9\u5411\u3002\u603b\u4f53\u800c\u8a00\uff0cSKILLOPT \u662f\u4e00\u7bc7\u5177\u6709\u91cd\u8981\u7406\u8bba\u521b\u65b0\u548c\u5f3a\u5927\u5b9e\u8df5\u4ef7\u503c\u7684\u8bba\u6587\uff0c\u4e3a\u6784\u5efa\u66f4\u667a\u80fd\u3001\u66f4\u53ef\u9760\u7684 LLM \u4ee3\u7406\u7cfb\u7edf\u8d21\u732e\u4e86\u5173\u952e\u65b9\u6cd5\u8bba\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. \u8bba\u6587\u6982\u8ff0\u4e0e\u5f15\u8a00 SKILLOPT\u63d0\u51fa\u4e86\u4e00\u79cd\u5f00\u521b\u6027\u7684\u6587\u672c\u7a7a\u95f4\u4f18\u5316\u6846\u67b6\uff0c\u7528\u4e8e\u5b9e\u73b0\u4ee3\u7406\u6280\u80fd\u7684\u81ea\u8fdb\u5316\u3002\u8be5\u8bba\u6587\u6df1\u523b\u6307\u51fa\uff0c\u5f53\u4eca\u4ee3\u7406\u6280\u80fd\u4e3b\u8981\u4f9d\u8d56\u624b\u5de5 crafting\u3001\u5355\u6b21 LLM \u751f\u6210\u6216\u677e\u6563\u7684\u81ea\u4fee\u8ba2\u5faa\u73af\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u5747\u65e0\u6cd5\u50cf\u6df1\u5ea6\u5b66\u4e60\u4f18\u5316\u5668\u90a3\u6837\u63d0\u4f9b\u7cfb\u7edf\u3001\u53ef\u63a7\u4e14\u53ef\u91cd\u590d\u7684\u6539\u8fdb\u8def\u5f84\u3002\u4f5c\u8005\u4e3b\u5f20\uff0c\u5e94\u5c06\u6280\u80fd\u6587\u6863\u89c6\u4e3a\u51bb\u7ed3\u4ee3\u7406\u6a21\u578b\u7684\u5916\u90e8\u53ef\u8bad\u7ec3\u72b6\u6001\uff0c\u5f15\u5165\u72ec\u7acb\u7684 frontier \u4f18\u5316\u5668\u6a21\u578b\u3002\u8be5\u4f18\u5316\u5668\u901a\u8fc7\u5206\u6790\u5e26\u5206\u6570\u7684 rollout \u8f68\u8ff9\uff0c\u751f\u6210\u53d7\u9650\u7684 add\/delete\/replace \u7f16\u8f91\u64cd\u4f5c\uff0c\u5e76\u4ec5\u5728\u7f16 &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1402\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":5,"featured_media":1410,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5,1],"tags":[],"class_list":["post-1402","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1402","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1402"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1402\/revisions"}],"predecessor-version":[{"id":1411,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1402\/revisions\/1411"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1410"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1402"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1402"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1402"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}