{"id":1747,"date":"2025-09-27T17:15:53","date_gmt":"2025-09-27T09:15:53","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=1747"},"modified":"2025-09-27T17:15:53","modified_gmt":"2025-09-27T09:15:53","slug":"agentic-coding%e8%a1%a8%e7%8e%b0%e5%88%9b%e6%96%b0%e9%ab%98%ef%bc%8c%e5%85%a8%e6%96%b0kat%e7%b3%bb%e5%88%97%e6%a8%a1%e5%9e%8b%e4%b8%8a%e6%a6%9cswe-bench","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/09\/27\/agentic-coding%e8%a1%a8%e7%8e%b0%e5%88%9b%e6%96%b0%e9%ab%98%ef%bc%8c%e5%85%a8%e6%96%b0kat%e7%b3%bb%e5%88%97%e6%a8%a1%e5%9e%8b%e4%b8%8a%e6%a6%9cswe-bench\/","title":{"rendered":"Agentic Coding\u8868\u73b0\u521b\u65b0\u9ad8\uff0c\u5168\u65b0KAT\u7cfb\u5217\u6a21\u578b\u4e0a\u699cSWE-Bench"},"content":{"rendered":"\n

\u56fe 1\uff1a\u5728 SWE-Bench Verified \u4e0a\uff0c\u548c\u5168\u5c3a\u5bf8\u5f00\u6e90\u6a21\u578b\u5bf9\u6bd4\uff0cKAT-Dev \u7528\u6781\u5c0f\u7684\u6a21\u578b\u5c3a\u5bf8\u53d6\u5f97\u4e86\u7b2c\u4e00\u68af\u961f\u7684\u6027\u80fd<\/sup><\/p>\n\n\n\n

\u5feb\u624b Kwaipilot \u56e2\u961f\u63a8\u51fa\u4e86 KAT \u7cfb\u5217\u4e24\u6b3e\u7a81\u7834\u6027 Agentic Coding \u5927\u6a21\u578b<\/strong>\uff1a\u5f00\u6e90 32B \u53c2\u6570\u6a21\u578b KAT-Dev-32B <\/strong>\u4e0e\u95ed\u6e90\u65d7\u8230\u6a21\u578b  KAT-Coder<\/strong>\u3002<\/p>\n\n\n\n

\u8fd9\u4e24\u6b3e\u6a21\u578b\u5728 Code Intelligence \u9886\u57df\u5206\u522b\u4f53\u73b0\u51fa\u8f7b\u91cf\u7ea7\u7684\u8d85\u5f3a\u8868\u73b0\u548c\u6781\u81f4\u6027\u80fd\u3002\u5176\u4e2d\uff0c\u5728 SWE-Bench Verified \u4e0a\uff0cKAT-Dev-32B \u5c55\u73b0\u51fa\u5f3a\u52b2\u6027\u80fd\u5e76\u53d6\u5f97\u4e86 62.4% \u7684\u89e3\u51b3\u7387\uff0c\u5728\u6240\u6709\u4e0d\u540c\u89c4\u6a21\u7684\u5f00\u6e90\u6a21\u578b\u4e2d\u6392\u540d\u7b2c 5\u3002\u4e0e\u6b64\u540c\u65f6\uff0cKAT-Coder \u4ee5 73.4% \u7684\u89e3\u51b3\u7387\u5728 SWE-Bench Verified \u4e0a\u53d6\u5f97\u4e86\u6781\u4f73\u7684\u5355\u6a21\u578b\u8868\u73b0\uff0c\u6bd4\u80a9\u5168\u7403\u9876\u5c16\u95ed\u6e90\u6a21\u578b\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u56fe 2\uff1a\u5728 SWE-Bench Verified \u4e0a\uff0cKAT-Coder \u53d6\u5f97\u6781\u4f73\u7684\u5355\u6a21\u578b\u8868\u73b0\uff0c\u6bd4\u80a9\u5168\u7403\u9876\u5c16\u95ed\u6e90\u6a21\u578b\u6027\u80fd<\/sup><\/p>\n\n\n\n

\u6a21\u578b\u5f00\u6e90\u548c API \u5f00\u653e<\/strong>
KAT-Dev-32B \u5df2\u5728\u5f00\u6e90\u6a21\u578b\u6258\u7ba1\u5e73\u53f0 Hugging Face \u4e0a\u7ebf\uff0c\u53ef\u4f9b\u8fdb\u4e00\u6b65\u7814\u7a76\u548c\u5f00\u53d1\u4f7f\u7528\u3002KAT-Coder \u6a21\u578b\u7684 API \u5bc6\u94a5\u8fd1\u671f\u4e5f\u5728 \u201c\u5feb\u624b\u4e07\u64ce\u201d \u4f01\u4e1a\u7ea7\u5927\u6a21\u578b\u670d\u52a1\u4e0e\u5f00\u53d1\u5e73\u53f0\u4e0a\u5f00\u653e\u7533\u8bf7\uff0c\u7528\u6237\u5c06\u80fd\u591f\u901a\u8fc7 Claude Code \u7b49\u5de5\u5177\u76f4\u63a5\u8bbf\u95ee\u5e76\u8fdb\u884c\u7f16\u7801\u3002

<\/p>\n\n\n\n

    \n
  • \u5feb\u624b Kwaipilot \u56e2\u961f\u7684\u5b98\u65b9\u6280\u672f Blog\uff1ahttps:\/\/kwaipilot.github.io\/KAT-Coder\/<\/li>\n\n\n\n
  • KAT-Dev-32B \u6a21\u578b\u5f00\u6e90\u5730\u5740\uff1ahttps:\/\/huggingface.co\/Kwaipilot\/KAT-Dev<\/li>\n\n\n\n
  • KAT-Coder \u5f00\u53d1\u5de5\u5177\u63a5\u5165\u6307\u5357\uff1ahttps:\/\/www.streamlake.com\/document\/WANQING\/me6ymdjrqv8lp4iq0o9<\/li>\n\n\n\n
  • KAT-Coder API Key \u7533\u8bf7\uff1ahttps:\/\/console.streamlake.com\/wanqing\/<\/li>\n<\/ul>\n\n\n\n

    \u6838\u5fc3\u8d21\u732e\u70b9\u6458\u8981<\/strong><\/p>\n\n\n\n

    KAT-Dev-32B \u548c KAT-Coder \u5728\u591a\u4e2a\u8bad\u7ec3\u9636\u6bb5\u8fdb\u884c\u4e86\u521b\u65b0\u548c\u4f18\u5316\uff0c\u5305\u62ec Mid-Training \u9636\u6bb5\u3001\u76d1\u7763\u5fae\u8c03 (SFT) \u9636\u6bb5\u3001\u5f3a\u5316\u5fae\u8c03 (RFT) \u9636\u6bb5\uff0c\u4ee5\u53ca\u5927\u89c4\u6a21\u667a\u80fd\u4f53\u5f3a\u5316\u5b66\u4e60 (RL) \u9636\u6bb5\uff0c\u5177\u4f53\u5982\u4e0b\uff1a<\/p>\n\n\n\n

      \n
    • Mid-Training<\/strong>\uff1aKwaipilot \u56e2\u961f\u53d1\u73b0\uff0c\u5728\u8fd9\u4e00\u9636\u6bb5\u5927\u91cf\u589e\u52a0\u5de5\u5177\u4f7f\u7528\u80fd\u529b\u3001\u591a\u8f6e\u4ea4\u4e92\u548c\u6307\u4ee4\u9075\u5faa\u7684\u8bad\u7ec3\uff0c\u867d\u7136\u5728\u5f53\u524d\u7ed3\u679c\u4e0a\uff08\u4f8b\u5982\u5728 SWE-bench \u7b49\u6392\u884c\u699c\uff09\u53ef\u80fd\u4e0d\u4f1a\u5e26\u6765\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff0c\u4f46\u5bf9\u540e\u7eed\u7684 SFT \u548c RL \u9636\u6bb5\u5177\u6709\u91cd\u5927\u5f71\u54cd\u3002<\/li>\n\n\n\n
    • SFT & RFT<\/strong>\uff1a\u56e2\u961f\u5728 SFT \u9636\u6bb5\u7cbe\u5fc3\u7b56\u5212\u4e86\u516b\u79cd\u4efb\u52a1\u7c7b\u578b\u548c\u516b\u79cd\u7f16\u7a0b\u573a\u666f\uff0c\u4ee5\u786e\u4fdd\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u548c\u7efc\u5408\u80fd\u529b\u3002\u6b64\u5916\uff0c\u5728 RL \u4e4b\u524d\uff0c\u521b\u65b0\u6027\u5730\u5f15\u5165\u4e86 RFT \u9636\u6bb5\uff0c\u4f7f\u7528\u4eba\u7c7b\u5de5\u7a0b\u5e08\u6807\u6ce8\u7684 “\u6559\u5e08\u8f68\u8ff9” \u4f5c\u4e3a\u8bad\u7ec3\u671f\u95f4\u7684\u6307\u5bfc\u3002<\/li>\n\n\n\n
    • \u5927\u89c4\u6a21 Agentic RL<\/strong>\uff1a\u5f53\u524d\uff0c\u6269\u5c55\u667a\u80fd\u4f53 RL \u9762\u4e34\u4e09\u4e2a\u6311\u6218\uff1a\u975e\u7ebf\u6027\u8f68\u8ff9\u5386\u53f2\u7684\u9ad8\u6548\u5b66\u4e60\u3001\u5229\u7528\u5185\u5728\u6a21\u578b\u4fe1\u53f7\u4ee5\u53ca\u6784\u5efa\u53ef\u6269\u5c55\u7684\u9ad8\u541e\u5410\u91cf\u57fa\u7840\u8bbe\u65bd\u3002\u5bf9\u6b64\uff0cKwaipilot \u56e2\u961f\u901a\u8fc7\u5bf9\u6570\u6982\u7387\u8ba1\u7b97\u7684\u524d\u7f00\u7f13\u5b58\uff08Log-Probability Prefix Caching\uff09\u3001\u57fa\u4e8e\u71b5\u7684\u8f68\u8ff9\u526a\u679d\uff08Entropy-based Tree Pruning\uff09\u548c\u81ea\u7814\u7684\u5de5\u4e1a\u7ea7\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u6846\u67b6 SeamlessFlow \u6765\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u3002<\/li>\n<\/ul>\n\n\n\n

      KAT \u7cfb\u5217\u6a21\u578b\u7684\u6838\u5fc3\u6280\u672f\u8def\u7ebf<\/strong><\/p>\n\n\n\n

      \u4e00\u3001Mid-Training<\/strong><\/p>\n\n\n\n

      Kwaipilot \u56e2\u961f\u5bf9\u7ecf\u8fc7\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u8fdb\u884c\u4e86\u4e24\u9636\u6bb5\u8bad\u7ec3\uff0c\u8be5\u9636\u6bb5\u88ab\u79f0\u4e3a Mid-Training\u3002\u5728\u5176\u4e2d\u7684\u7b2c\u4e00\u4e2a\u9636\u6bb5\uff0c\u589e\u5f3a\u4e86\u6a21\u578b\u4e0e \u201cLLM-as-Agent\u201d \u76f8\u5173\u7684\u5168\u65b9\u4f4d\u80fd\u529b\uff0c\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e\u4ee5\u4e0b\u51e0\u79cd\u80fd\u529b\uff1a<\/p>\n\n\n\n

        \n
      • \u5de5\u5177\u8c03\u7528\u80fd\u529b<\/strong>\uff1a\u6784\u5efa\u4e86\u5728\u6c99\u76d2\u73af\u5883\u771f\u5b9e\u6267\u884c\u5de5\u5177\u7684\u8c03\u7528\u65b9\u6cd5\u4ee5\u53ca\u6267\u884c\u7ed3\u679c\u7684\u4ea4\u4e92\u6570\u636e\uff0c\u7528\u4e8e\u63d0\u5347\u6a21\u578b\u7684\u5de5\u5177\u8c03\u7528\u80fd\u529b\uff1b<\/li>\n\n\n\n
      • \u591a\u8f6e\u4ea4\u4e92\u80fd\u529b<\/strong>\uff1a\u6784\u5efa\u4e86\u6700\u957f\u6570\u767e\u8f6e\u7684\u4eba\u7c7b\u3001\u6a21\u578b\u3001\u5de5\u5177\u7684\u4ea4\u4e92\u6570\u636e\uff0c\u7528\u4e8e\u63d0\u5347\u5728\u957f\u6587\u672c\u60c5\u51b5\u4e0b\u6a21\u578b\u7684\u591a\u8f6e\u4ea4\u4e92\u80fd\u529b\uff1b<\/li>\n\n\n\n
      • \u7f16\u7801\u77e5\u8bc6\u6ce8\u5165<\/strong>\uff1a\u52a0\u5165\u4e86\u9ad8\u8d28\u91cf\u7684\u4e0e\u7f16\u7801\u76f8\u5173\u7684\u9886\u57df\u77e5\u8bc6\u6570\u636e\uff0c\u7528\u4e8e\u8fdb\u4e00\u6b65\u589e\u5f3a\u6a21\u578b\u5728\u7f16\u7801\u573a\u666f\u4e0b\u7684\u6027\u80fd\uff1b<\/li>\n\n\n\n
      • Git Commit \u6570\u636e<\/strong>\uff1a\u52a0\u5165\u4e86\u5927\u91cf\u6765\u81ea\u4e8e\u771f\u5b9e Git \u4ed3\u5e93\u7684 PR \u6570\u636e\uff0c\u7528\u4e8e\u8fdb\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u5728\u771f\u5b9e\u7f16\u7a0b\u4efb\u52a1\u4e0b\u7684\u8868\u73b0\uff1b<\/li>\n\n\n\n
      • \u6307\u4ee4\u8ddf\u968f\u6570\u636e<\/strong>\uff1a\u6536\u96c6\u4e86 30 + \u7c7b\u5e38\u89c1\u7684\u7528\u6237\u6307\u4ee4\uff0c\u7528\u4e8e\u589e\u5f3a\u6a21\u578b\u5bf9\u7528\u6237\u6307\u4ee4\u7684\u7406\u89e3\u80fd\u529b\uff1b<\/li>\n\n\n\n
      • \u901a\u7528\u53ca\u601d\u8003\u6570\u636e<\/strong>\uff1a\u6784\u5efa\u4e86\u591a\u7c7b\u901a\u7528\u6570\u636e\uff0c\u7528\u4e8e\u589e\u5f3a\u6a21\u578b\u5728\u901a\u7528\u9886\u57df\u4ee5\u53ca\u5728\u8c03\u7528\u5de5\u5177\u65f6\u8fdb\u884c\u601d\u8003\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n

        \u4e8c\u3001\u76d1\u7763\u5fae\u8c03 (Supervised Fine-Tuning, SFT)<\/strong><\/p>\n\n\n\n

        \u5728\u7b2c\u4e8c\u9636\u6bb5\uff0cKwaipilot \u56e2\u961f\u6536\u96c6\u4e86\u5927\u91cf\u4eba\u7c7b\u5de5\u7a0b\u5e08\u6807\u8bb0\u7684\u771f\u5b9e\u9700\u6c42\u4ea4\u4ed8\u8f68\u8ff9\uff0c\u5e76\u57fa\u4e8e\u6b64\u5408\u6210\u4e86\u5927\u91cf\u7684\u8f68\u8ff9\u6570\u636e\uff0c\u8fdb\u4e00\u6b65\u5bf9\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\uff0c\u4ee5\u589e\u5f3a\u5176\u7aef\u5230\u7aef\u9700\u6c42\u4ea4\u4ed8\u7684\u80fd\u529b\u3002\u5176\u4e2d\u8986\u76d6\u4e86\u591a\u79cd\u4efb\u52a1\u7c7b\u578b\uff1a<\/p>\n\n\n\n

        \u516b\u5927\u7528\u6237\u4efb\u52a1\u7c7b\u578b\uff1a<\/strong><\/p>\n\n\n\n

          \n
        • Feature Implementation\uff08\u529f\u80fd\u5b9e\u73b0\uff09<\/li>\n\n\n\n
        • Feature Enhancement\uff08\u529f\u80fd\u589e\u5f3a\uff09<\/li>\n\n\n\n
        • Bug Fixing\uff08\u7f3a\u9677\u4fee\u590d\uff09<\/li>\n\n\n\n
        • Refactoring\uff08\u7ed3\u6784\u4f18\u5316\uff09<\/li>\n\n\n\n
        • Performance Optimization\uff08\u6027\u80fd\u4f18\u5316\uff09<\/li>\n\n\n\n
        • Test Case Generation\uff08\u6d4b\u8bd5\u7528\u4f8b\u751f\u6210\uff09<\/li>\n\n\n\n
        • Code Understanding\uff08\u4ee3\u7801\u7406\u89e3\uff09<\/li>\n\n\n\n
        • Configuration & Deployment\uff08\u914d\u7f6e\u4e0e\u90e8\u7f72\uff09<\/li>\n<\/ul>\n\n\n\n

          \u516b\u5927\u7528\u6237\u7f16\u7a0b\u573a\u666f\uff1a<\/strong><\/p>\n\n\n\n

            \n
          • Application Development\uff08\u5e94\u7528\u5f00\u53d1\uff09<\/li>\n\n\n\n
          • UI\/UX Engineering\uff08\u754c\u9762\u4e0e\u7528\u6237\u4f53\u9a8c\u5de5\u7a0b\uff09<\/li>\n\n\n\n
          • Data Science & Engineering\uff08\u6570\u636e\u79d1\u5b66\u4e0e\u5de5\u7a0b\uff09<\/li>\n\n\n\n
          • Machine Learning & AI\uff08\u673a\u5668\u5b66\u4e60\u4e0e\u4eba\u5de5\u667a\u80fd\uff09<\/li>\n\n\n\n
          • Database Systems\uff08\u6570\u636e\u5e93\u7cfb\u7edf\uff09<\/li>\n\n\n\n
          • Infrastructure Development\uff08\u57fa\u7840\u8bbe\u65bd\u5f00\u53d1\uff09<\/li>\n\n\n\n
          • Specialized Programming Domains\uff08\u4e13\u4e1a\u7f16\u7a0b\u9886\u57df\uff09<\/li>\n\n\n\n
          • Security Engineering\uff08\u5b89\u5168\u5de5\u7a0b\uff09<\/li>\n<\/ul>\n\n\n\n

            \u4e09\u3001\u5f3a\u5316\u5fae\u8c03\uff08Reinforcement Finetune\uff0cRFT\uff09<\/strong><\/p>\n\n\n\n

            \u5728\u8fd9\u4e00\u9636\u6bb5\uff0cKwaipilot \u56e2\u961f\u5728\u5f3a\u5316\u5b66\u4e60\u6d41\u7a0b\u7684\u57fa\u7840\u4e0a\uff0c\u989d\u5916\u5f15\u5165\u4e86\u591a\u4e2a ground truth \u7528\u4e8e\u8f68\u8ff9\u63a2\u7d22\u7684\u6307\u5bfc\uff0c\u63d0\u5347 rollout \u6548\u7387\uff0c\u4ece\u7edd\u5bf9 reward \u5230\u8861\u91cf\u4e0e ground truth \u7684\u5dee\u5f02\uff0c\u63d0\u5347\u4e86\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\u7684\u6548\u7387\u548c\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n

            \u4ece\u76f4\u63a5\u7ed9\u5b9a\u7edd\u5bf9 reward \u66f4\u65b0\u4e3a\u8861\u91cf rollout \u6837\u672c\u548c ground truth \u4e4b\u95f4\u7684\u76f8\u5bf9\u5dee\u5f02\u7ed9\u4e86\u5f3a\u5316\u5b66\u4e60\u66f4\u7a33\u5b9a\u548c\u66f4\u51c6\u786e\u7684\u5956\u52b1\u4fe1\u53f7\uff0c\u540c\u65f6\u4e5f\u4f1a\u5728 rollout \u9636\u6bb5\u5b9e\u65f6\u76d1\u7763\u6837\u672c\u7684\u6b63\u786e\u6027\uff0c\u5e76\u53ca\u65f6\u7ec8\u6b62\u4e0e ground truth \u6709\u660e\u663e\u504f\u79bb\u7684\u6837\u672c\u751f\u6210\uff0c\u8fd9\u4e5f\u7ed9\u5f3a\u5316\u5b66\u4e60\u5e26\u6765\u4e86\u66f4\u9ad8\u7684\u6837\u672c\u6548\u7387\u3002<\/p>\n\n\n\n

            \"\"<\/figure>\n\n\n\n

                  \u56fe 3\uff1a\u5728\u5f3a\u5316\u5fae\u8c03\uff08RFT\uff09\u6d41\u7a0b\u4e2d\uff0c\u5f15\u5165\u6559\u5e08\u8f68\u8ff9\u4f5c\u4e3a\u6307\u5bfc<\/sup><\/p>\n\n\n\n

            \u7ecf\u8fc7\u4e09\u9636\u6bb5\u7684\u8bad\u7ec3\uff0c\u56e2\u961f\u83b7\u5f97\u4e86\u4e3a RL \u9636\u6bb5\u51c6\u5907\u7684\u51b7\u542f\u52a8\u6a21\u578b\uff0cRFT \u7684\u52a0\u5165\u4e5f\u4e3a SFT \u548c RL \u4e4b\u95f4\u6784\u5efa\u4e86\u6865\u6881\u3002<\/p>\n\n\n\n

              \n
            • Mid-Training<\/strong>\uff1a\u9996\u5148\uff0c\u56e2\u961f\u6559\u4f1a\u5927\u6a21\u578b\u5404\u79cd\u57fa\u672c\u6280\u80fd\uff0c\u5305\u62ec\u5982\u4f55\u4f7f\u7528\u5de5\u5177\u3001\u5982\u4f55\u7406\u89e3\u7528\u6237\u610f\u56fe\u7b49\uff1b<\/li>\n\n\n\n
            • SFT<\/strong>\uff1a\u5176\u6b21\uff0c\u7528\u9ad8\u8d28\u91cf\u7684\u8f68\u8ff9\u6570\u636e\uff0c\u8ba9\u6a21\u578b\u5b66\u4e60\u5982\u4f55\u6267\u884c\u771f\u5b9e\u7684\u4e0b\u6e38\u4efb\u52a1\uff1b<\/li>\n\n\n\n
            • RFT<\/strong>\uff1a\u6700\u540e\uff0c\u5728\u6a21\u578b\u51c6\u5907 \u201c\u81ea\u7531\u63a2\u7d22\u201d \u4e4b\u524d\uff0c\u5148\u7531\u6559\u5e08\u8f68\u8ff9\u624b\u628a\u624b\u6559\u4f1a\u6a21\u578b\u5982\u4f55\u63a2\u7d22\uff0c\u4fdd\u969c\u4e86\u6a21\u578b\u540e\u7eed\u5728 RL \u9636\u6bb5\u7684\u7a33\u5b9a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n

              \u56db\u3001\u5927\u89c4\u6a21 Agentic RL<\/strong><\/p>\n\n\n\n

              1\u3001\u57fa\u4e8e\u71b5\u7684\u6811\u526a\u679d\uff08Entropy Based Tree Pruning\uff09<\/p>\n\n\n\n

              Kwaipilot \u56e2\u961f\u53d1\u73b0\uff0c\u5373\u4fbf\u4f7f\u7528\u4e0a\u8ff0\u6280\u672f\uff0c\u5bf9\u5b8c\u6574\u6811\u4e2d\u7684\u6240\u6709 token \u8fdb\u884c\u8bad\u7ec3\u7684\u6210\u672c\u4ecd\u7136\u8fc7\u9ad8\uff0c\u56e0\u6b64\u4e9f\u9700\u8bbe\u8ba1\u4e00\u79cd\u80fd\u591f\u4f18\u5148\u805a\u7126\u4e8e\u643a\u5e26\u6700\u5f3a\u8bad\u7ec3\u4fe1\u53f7\u8282\u70b9\u7684\u673a\u5236\u3002<\/p>\n\n\n\n

              \u4e3a\u6b64\uff0c\u56e2\u961f\u5c06\u8f68\u8ff9\u538b\u7f29\u6210\u4e00\u4e2a\u524d\u7f00\u6811\uff0c\u5176\u4e2d\u6bcf\u4e2a\u8282\u70b9\u8868\u793a\u4e00\u4e2a\u5171\u4eab\u524d\u7f00\uff0c\u6bcf\u6761\u8fb9\u5bf9\u5e94\u4e00\u6bb5 token\u3002\u5728\u56fa\u5b9a\u7684\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0c\u76ee\u6807\u662f\u53ea\u4fdd\u7559\u6700\u6709\u4ef7\u503c\u7684\u8282\u70b9\u8fdb\u884c\u8bad\u7ec3\u3002\u56e2\u961f\u57fa\u4e8e\u6811\u4e2d\u805a\u5408\u7684\u71b5\u4fe1\u53f7\u548c\u8282\u70b9\u88ab\u5230\u8fbe\u7684\u53ef\u80fd\u6027\u6765\u4f30\u8ba1\u8282\u70b9\u7684\u4fe1\u606f\u91cf\uff0c\u5e76\u6309\u7167\u91cd\u8981\u6027\u987a\u5e8f\u6269\u5c55\u8282\u70b9\u6765\u526a\u679d\u6811\uff0c\u76f4\u5230\u9884\u7b97\u8017\u5c3d\u3002\u989d\u5916\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\u786e\u4fdd\u4fdd\u7559\u7ed3\u6784\u4e0a\u7684\u91cd\u8981\u533a\u57df\uff08\u4f8b\u5982\uff0c\u5de5\u5177\u6216\u5185\u5b58\u4e8b\u4ef6\uff09\uff0c\u5e76\u7ef4\u62a4\u5c40\u90e8\u4e0a\u4e0b\u6587\u4ee5\u7a33\u5b9a\u8bad\u7ec3\u3002\u8fd9\u79cd\u57fa\u4e8e\u71b5\u7684\u526a\u679d\u5927\u5e45\u51cf\u5c11\u5197\u4f59\u8ba1\u7b97\uff0c\u540c\u65f6\u4fdd\u7559\u5927\u90e8\u5206\u6709\u6548\u7684\u8bad\u7ec3\u4fe1\u53f7\uff0c\u4ece\u800c\u5b9e\u73b0\u663e\u8457\u7684\u541e\u5410\u91cf\u63d0\u5347\u548c\u66f4\u4f4e\u7684\u603b\u4f53\u6210\u672c\u3002<\/p>\n\n\n\n

              2\u3001RL infra\uff1a\u81ea\u7814 SeamlessFlow \u6846\u67b6<\/p>\n\n\n\n

              \"\"<\/figure>\n\n\n\n

                    \u56fe 4\uff1aKwaipilot \u56e2\u961f\u81ea\u7814\u7684 RL \u8bad\u7ec3\u6846\u67b6 SeamlessFlow \u67b6\u6784<\/sup><\/p>\n\n\n\n

              \u4e3a\u6269\u5c55 RL\uff0c\u5fc5\u987b\u5c06 RL \u8bad\u7ec3\u4e0e\u667a\u80fd\u4f53\u7684\u591a\u6837\u5316\u5185\u90e8\u903b\u8f91\u5b8c\u5168\u89e3\u8026\uff0c\u540c\u65f6\u6700\u5927\u5316\u5f02\u6784\u8ba1\u7b97\u67b6\u6784\u7684\u5229\u7528\u7387\u3002\u9075\u5faa SeamlessFlow \u7684\u8bbe\u8ba1\uff0cKwaipilot \u56e2\u961f\u5728\u667a\u80fd\u4f53\u548c RL \u8bad\u7ec3\u4e4b\u95f4\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u4e13\u95e8\u7528\u4e8e\u8f68\u8ff9\u6811\u7ba1\u7406\u7684\u4e2d\u95f4\u5c42\uff0c\u786e\u4fdd\u4e24\u8005\u4e4b\u95f4\u7684\u4e25\u683c\u5206\u79bb\u3002\u6b64\u5916\uff0c\u91c7\u7528\u63d0\u51fa\u7684\u6807\u7b7e\u9a71\u52a8\u8c03\u5ea6\u673a\u5236\u6765\u534f\u8c03\u5f02\u6784\u96c6\u7fa4\u4e2d\u7684\u4efb\u52a1\u5206\u914d\uff0c\u4ece\u800c\u6700\u5c0f\u5316\u7ba1\u9053\u6c14\u6ce1\u5e76\u7ef4\u6301\u9ad8\u541e\u5410\u91cf\u8bad\u7ec3\u3002<\/p>\n\n\n\n

              3\u3001\u7edf\u4e00\u73af\u5883\u63a5\u53e3\u548c\u4f01\u4e1a\u7ea7 RL \u6570\u636e\u6784\u5efa<\/p>\n\n\n\n

              Kwaipilot \u56e2\u961f\u8fd8\u901a\u8fc7\u7edf\u4e00\u4e0d\u540c RL \u6267\u884c\u73af\u5883\u7684\u90e8\u7f72\u548c\u8bc4\u4f30\u63a5\u53e3\uff0c\u4f7f\u4efb\u4f55\u65b0\u6dfb\u52a0\u7684\u73af\u5883\u90fd\u80fd\u4ee5\u4f4e\u6210\u672c\u65e0\u7f1d\u96c6\u6210\u3002\u8fd9\u79cd\u7edf\u4e00\u8bbe\u8ba1\u4e3a\u8de8\u5f02\u6784\u6570\u636e\u6e90\u548c\u4efb\u52a1\u7c7b\u578b\u6269\u5c55 RL \u8bad\u7ec3\u5960\u5b9a\u4e86\u575a\u5b9e\u57fa\u7840\u3002\u5177\u4f53\u5230\u8f6f\u4ef6\u5f00\u53d1\u573a\u666f\uff0c\u56e2\u961f\u805a\u96c6\u4e8e\u4e09\u4e2a\u57fa\u672c\u7ec4\u4ef6\uff1a\u4e0e\u76f8\u5e94\u5206\u652f\u4ee3\u7801\u914d\u5bf9\u7684\u95ee\u9898\u63cf\u8ff0\u3001\u53ef\u6267\u884c\u73af\u5883\u548c\u53ef\u9a8c\u8bc1\u7684\u6d4b\u8bd5\u7528\u4f8b\u3002<\/p>\n\n\n\n

              Kwaipilot \u56e2\u961f\u4ece\u5f00\u6e90\u4ed3\u5e93\u6536\u96c6\u62c9\u53d6\u8bf7\u6c42\u548c\u76f8\u5173\u95ee\u9898\uff0c\u5e76\u6839\u636e\u8fd9\u4e9b\u4ed3\u5e93\u7684\u661f\u6807\u3001PR \u6d3b\u52a8\u548c\u95ee\u9898\u5185\u5bb9\u8fc7\u6ee4\u4f4e\u8d28\u91cf\u6570\u636e\uff0c\u968f\u540e\u7cfb\u7edf\u5730\u4e3a\u6bcf\u4e2a\u6536\u96c6\u7684\u5b9e\u4f8b\u6784\u5efa\u53ef\u6267\u884c\u73af\u5883\u955c\u50cf\u5e76\u751f\u6210\u5355\u5143\u6d4b\u8bd5\u7528\u4f8b\u3002\u9664\u4e86\u8f6f\u4ef6\u5de5\u7a0b\u6570\u636e\uff0c\u56e2\u961f\u8fd8\u7eb3\u5165\u4e86\u5176\u4ed6\u53ef\u9a8c\u8bc1\u9886\u57df\uff0c\u5982\u6570\u5b66\u548c\u63a8\u7406\u4efb\u52a1\uff0c\u8fdb\u4e00\u6b65\u4e30\u5bcc\u4e86 RL \u4fe1\u53f7\u7684\u591a\u6837\u6027\u3002<\/p>\n\n\n\n

              \u66f4\u91cd\u8981\u7684\u662f\uff0c\u9664\u4e86\u5f00\u6e90\u6570\u636e\uff0c\u56e2\u961f\u8fd8\u8fdb\u4e00\u6b65\u6536\u96c6\u5e76\u5229\u7528\u6765\u81ea\u771f\u5b9e\u4e16\u754c\u5de5\u4e1a\u7cfb\u7edf\u7684\u533f\u540d\u4f01\u4e1a\u7ea7\u4ee3\u7801\u5e93\u8fdb\u884c RL \u8bad\u7ec3\u3002\u4e0e\u4ec5\u5728\u516c\u5171\u4ed3\u5e93\uff08\u5982 GitHub \u4e0a\u7684\u4ed3\u5e93\uff09\u4e0a\u8bad\u7ec3\u4e0d\u540c\uff0c\u8fd9\u4e9b\u4ed3\u5e93\u901a\u5e38\u5305\u542b\u8f83\u7b80\u5355\u7684\u9879\u76ee\uff0c\u800c\u8fd9\u4e9b\u5927\u89c4\u6a21\u3001\u590d\u6742\u7684\u4ee3\u7801\u5e93 \u2014\u2014 \u8de8\u8d8a\u591a\u79cd\u7f16\u7a0b\u8bed\u8a00\u5e76\u4ee3\u8868\u771f\u5b9e\u7684\u4e1a\u52a1\u903b\u8f91 \u2014\u2014 \u8ba9\u6a21\u578b\u63a5\u89e6\u5230\u66f4\u5177\u6311\u6218\u6027\u7684\u5f00\u53d1\u573a\u666f\uff0c\u4e3a RL \u63d0\u4f9b\u4e86\u9ad8\u4ef7\u503c\u7684\u8d44\u4ea7\u3002\u8bad\u7ec3\u667a\u80fd\u4f53\u89e3\u51b3\u8fd9\u4e9b\u771f\u5b9e\u4e16\u754c\u7684\u5de5\u4e1a\u95ee\u9898\u4e0d\u4ec5\u589e\u5f3a\u4e86\u5b66\u4e60\u7684\u9c81\u68d2\u6027\uff0c\u8fd8\u5c06\u6240\u5f97\u6a21\u578b\u7684\u7f16\u7a0b\u80fd\u529b\u5efa\u7acb\u5728\u73b0\u5b9e\u7684\u751f\u4ea7\u7ea7\u73af\u5883\u4e2d\u3002<\/p>\n\n\n\n

              \"\"<\/figure>\n\n\n\n

                    \u56fe 5\uff1a\u5728 SWE-Bench Verified \u4e0a\uff0c\u5404\u9636\u6bb5\u8bad\u7ec3\u5bf9\u6a21\u578b\u7684\u6027\u80fd\u5f71\u54cd<\/sup><\/p>\n\n\n\n

              \u6a21\u578b\u6548\u679c\u5c55\u793a<\/strong><\/p>\n\n\n\n

              KAT-Coder \u6a21\u578b\u5177\u5907\u5f3a\u5927\u7684\u4ee3\u7801\u751f\u6210\u80fd\u529b\uff0c\u53ef\u72ec\u7acb\u5b8c\u6210\u5b8c\u6574\u7684\u9879\u76ee\u5f00\u53d1\uff0c\u901a\u8fc7\u8c03\u7528\u7f16\u7a0b\u5de5\u5177\u53ef\u5b9e\u73b0\u4ece\u4ea4\u4e92\u5f0f\u6e38\u620f\u5230\u4ee3\u7801\u91cd\u6784\u7b49\u591a\u6837\u5316\u7f16\u7a0b\u4efb\u52a1\u3002\u7528\u6237\u4ec5\u9700\u63cf\u8ff0\u9700\u6c42\uff0c\u6a21\u578b\u5373\u53ef\u4ea4\u4ed8\u5b8c\u6574\u7684\u4ee3\u7801\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n

              1\u3001\u661f\u7a7a\u6548\u679c<\/p>\n\n\n\n

              \"\"<\/figure>\n\n\n\n

              2\u3001\u6c34\u679c\u5fcd\u8005\ud83e\udd77<\/p>\n\n\n\n

              \"\"<\/figure>\n\n\n\n

              3\u3001\u4ee3\u7801\u91cd\u6784<\/p>\n\n\n\n

              \"\"<\/figure>\n\n\n\n

              \u5927\u89c4\u6a21 Agentic RL \u540e\u7684\u6d8c\u73b0\u80fd\u529b<\/strong><\/p>\n\n\n\n

              \u5bf9\u7ecf\u8fc7\u5927\u89c4\u6a21 Agentic RL \u8bad\u7ec3\u540e\u7684\u6a21\u578b\u8fdb\u884c\u5206\u6790\uff0cKwaipilot \u56e2\u961f\u89c2\u6d4b\u5230\u4e86\u4e24\u4e2a\u663e\u8457\u7684\u6d8c\u73b0\u73b0\u8c61\uff1a<\/p>\n\n\n\n

                \n
              1. \u5bf9\u8bdd\u8f6e\u6b21\u663e\u8457\u964d\u4f4e<\/strong>\uff1a\u6a21\u578b\u503e\u5411\u4e8e\u7528\u66f4\u5c11\u7684\u4ea4\u4e92\u8f6e\u6b21\u5b8c\u6210\u4efb\u52a1\uff0c\u76f8\u8f83\u4e8e SFT \u6a21\u578b\uff0c\u5e73\u5747\u5bf9\u8bdd\u8f6e\u6b21\u4e0b\u964d\u4e86 32%\uff1b<\/li>\n\n\n\n
              2. \u591a\u5de5\u5177\u5e76\u884c\u8c03\u7528<\/strong>\uff1a\u6a21\u578b\u5c55\u73b0\u51fa\u540c\u65f6\u8c03\u7528\u591a\u4e2a\u5de5\u5177\u7684\u80fd\u529b\uff0c\u800c\u975e\u4f20\u7edf\u7684\u4e32\u884c\u8c03\u7528\u3002<\/li>\n<\/ol>\n\n\n\n

                \u56e2\u961f\u63a8\u6d4b\uff0c\u8fd9\u6e90\u4e8e\u8f68\u8ff9\u6811\u7ed3\u6784\u5e26\u6765\u7684\u9690\u5f0f\u4f18\u5316\u538b\u529b\uff0c\u4f7f\u6a21\u578b\u81ea\u7136\u5f62\u6210\u6548\u7387\u504f\u597d\u4e0e\u5e76\u884c\u8c03\u7528\u80fd\u529b\u3002<\/p>\n\n\n\n

                  \n
                1. \u6548\u7387\u504f\u597d\u7684\u5f62\u6210<\/strong>\uff1a\u5728\u8f68\u8ff9\u6811\u7ed3\u6784\u4e2d\uff0c\u8f83\u77ed\u7684\u8def\u5f84\uff08\u66f4\u5c11\u7684\u5bf9\u8bdd\u8f6e\u6b21\uff09\u4f1a\u88ab\u66f4\u591a\u7684\u8bad\u7ec3\u6837\u672c\u5171\u4eab\u3002\u8fd9\u521b\u9020\u4e86\u4e00\u4e2a\u9690\u5f0f\u7684\u4f18\u5316\u538b\u529b\uff1a\u6a21\u578b\u503e\u5411\u4e8e\u5b66\u4e60\u66f4\u9ad8\u6548\u7684\u89e3\u51b3\u65b9\u6848\uff1b<\/li>\n\n\n\n
                2. \u5e76\u884c\u5316\u7684\u81ea\u7136\u9009\u62e9<\/strong>\uff1a\u5728\u6811\u7ed3\u6784\u4e2d\uff0c\u591a\u5de5\u5177\u5e76\u884c\u8c03\u7528\u521b\u9020\u4e86\u66f4\u591a\u7684\u5206\u652f\u53ef\u80fd\u6027\uff0c\u8fd9\u4e9b\u5206\u652f\u5728\u8bad\u7ec3\u65f6\u88ab\u72ec\u7acb\u5904\u7406\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u540c\u65f6\u63a2\u7d22\u591a\u4e2a\u5de5\u5177\u7ec4\u5408\u3002\u540c\u65f6\u71b5\u526a\u679d\u673a\u5236\uff08Long-term Entropy Pruning\uff09\u4fdd\u7559\u4e86\u4fe1\u606f\u91cf\u8f83\u5927\u7684\u8282\u70b9\uff0c\u800c\u591a\u5de5\u5177\u8c03\u7528\u8282\u70b9\u5f80\u5f80\u5177\u6709\u66f4\u9ad8\u7684\u71b5\u503c\uff0c\u4f7f\u6a21\u578b\u9010\u6e10\u5b66\u4f1a\u4e86 “\u6279\u5904\u7406” \u601d\u7ef4\u3002<\/li>\n<\/ol>\n\n\n\n

                  \u672a\u6765\u5c55\u671b<\/strong><\/p>\n\n\n\n

                  Kwaipilot \u56e2\u961f\u5c06\u6301\u7eed\u63a2\u7d22\u4ee3\u7801\u667a\u80fd\u7684\u524d\u6cbf\u9886\u57df\uff0c\u5f00\u62d3\u521b\u65b0\u53ef\u80fd\uff1a<\/p>\n\n\n\n

                    \n
                  1. \u589e\u5f3a\u5de5\u5177\u96c6\u6210<\/strong>\uff1a\u4e0e\u6d41\u884c\u7684 IDE\u3001\u7248\u672c\u63a7\u5236\u7cfb\u7edf\u548c\u5f00\u53d1\u5de5\u4f5c\u6d41\u6df1\u5ea6\u96c6\u6210\uff0c\u521b\u5efa\u65e0\u7f1d\u7684\u7f16\u7801\u4f53\u9a8c\u3002<\/li>\n\n\n\n
                  2. \u591a\u8bed\u8a00\u6269\u5c55<\/strong>\uff1a\u6269\u5c55 KAT \u6a21\u578b\u80fd\u529b\u4ee5\u8986\u76d6\u65b0\u5174\u7684\u7f16\u7a0b\u8bed\u8a00\u548c\u6846\u67b6\uff0c\u786e\u4fdd\u5168\u9762\u7684\u8bed\u8a00\u652f\u6301\u3002<\/li>\n\n\n\n
                  3. \u534f\u4f5c\u7f16\u7801<\/strong>\uff1a\u63a2\u7d22\u591a\u667a\u80fd\u4f53\u7cfb\u7edf\uff0c\u8ba9 KAT \u6a21\u578b\u80fd\u591f\u5728\u590d\u6742\u7684\u8f6f\u4ef6\u9879\u76ee\u4e0a\u534f\u540c\u5de5\u4f5c\uff0c\u5b9e\u73b0\u524d\u6240\u672a\u6709\u7684\u534f\u4f5c\u3002<\/li>\n\n\n\n
                  4. \u591a\u6a21\u6001\u4ee3\u7801\u667a\u80fd<\/strong>\uff1a\u96c6\u6210\u89c6\u89c9\u7406\u89e3\u80fd\u529b\uff0c\u5904\u7406\u67b6\u6784\u56fe\u3001UI \u8bbe\u8ba1\u3001\u8c03\u8bd5\u622a\u56fe\u548c\u6587\u6863\u56fe\u50cf\u4ee5\u53ca\u4ee3\u7801\uff0c\u4f7f\u5f00\u53d1\u8fc7\u7a0b\u66f4\u52a0\u76f4\u89c2\u548c\u9ad8\u6548\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"

                    \u56fe 1\uff1a\u5728 SWE-Bench Verified \u4e0a\uff0c\u548c\u5168\u5c3a\u5bf8\u5f00\u6e90\u6a21\u578b\u5bf9\u6bd4\uff0cKAT-Dev \u7528\u6781\u5c0f\u7684\u6a21\u578b\u5c3a\u5bf8\u53d6 […]<\/p>\n","protected":false},"author":1,"featured_media":1756,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[],"class_list":["post-1747","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-10"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1747","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=1747"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1747\/revisions"}],"predecessor-version":[{"id":1757,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1747\/revisions\/1757"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/1756"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=1747"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=1747"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=1747"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}