{"id":1659,"date":"2026-09-20T11:59:23","date_gmt":"2026-09-20T03:59:23","guid":{"rendered":"https:\/\/lingbo.online\/index.php\/learning_note\/aacr-bench\/"},"modified":"2026-09-20T11:59:23","modified_gmt":"2026-09-20T03:59:23","slug":"aacr-bench","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/learning_note\/aacr-bench\/","title":{"rendered":"AACR-Bench: Evaluating Automatic Code Review with Holistic Repository-Level Context"},"content":{"rendered":"<h3>Meta Data<\/h3>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4\uff1a2026-01-30<\/li>\n<li>\u4f5c\u8005\uff1aLei Zhang, Yongda Yu, Minghui Yu, Xinxin Guo, Zhengqi Zhuang, Guoping Rong, Dong Shao, Haifeng Shen, Hongyu Kuang, Zhengfeng Li, Boge Wang, Guoan Zhang, Bangyu Xiang, Xiaobin Xu\uff08\u5357\u4eac\u5927\u5b66\u3001\u5357\u5341\u5b57\u661f\u5927\u5b66\u3001\u963f\u91cc\u5df4\u5df4\uff09<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2601.19494.pdf\" target=\"_blank\" rel=\"noopener\" rel=\"nofollow\" >https:\/\/arxiv.org\/pdf\/2601.19494.pdf<\/a><\/li>\n<li>arXiv \u9875\u9762\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2601.19494\" target=\"_blank\" rel=\"noopener\" rel=\"nofollow\" >https:\/\/arxiv.org\/abs\/2601.19494<\/a><\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/alibaba\/aacr-bench\" target=\"_blank\" rel=\"noopener\" rel=\"nofollow\" >https:\/\/github.com\/alibaba\/aacr-bench<\/a><\/li>\n<\/ul>\n<h1>AACR-Bench\uff1a\u57fa\u4e8e\u5b8c\u6574\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u7684\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\u8bc4\u6d4b<\/h1>\n<blockquote>\n<p>\u539f\u6587\u6807\u9898\uff1aAACR-Bench: Evaluating Automatic Code Review with Holistic Repository-Level Context<\/p>\n<\/blockquote>\n<p><strong>Lei Zhang*<\/strong>\u00b9, <strong>Yongda Yu*<\/strong>\u00b9, Minghui Yu\u00b9, Xinxin Guo\u00b9, Zhengqi Zhuang\u00b9, Guoping Rong\u00b9, Dong Shao\u00b9, Haifeng Shen\u00b2, Hongyu Kuang\u00b9, Zhengfeng Li\u00b3, Boge Wang\u00b3, Guoan Zhang\u00b3, Bangyu Xiang\u00b3, Xiaobin Xu\u00b3<\/p>\n<p>\u00b9 \u5357\u4eac\u5927\u5b66\u8f6f\u4ef6\u5b66\u9662\uff0c\u5357\u4eac\uff0c\u4e2d\u56fd<br \/>\n\u00b2 \u5357\u5341\u5b57\u661f\u5927\u5b66\uff08Southern Cross University\uff09\uff0c\u9ec4\u91d1\u6d77\u5cb8\uff0c\u6fb3\u5927\u5229\u4e9a<br \/>\n\u00b3 \u963f\u91cc\u5df4\u5df4 TRE\uff0c\u676d\u5dde\uff0c\u4e2d\u56fd  <\/p>\n<p>\u901a\u8baf\u4f5c\u8005\uff1aGuoping Rong\uff08ronggp@nju.edu.cn\uff09\uff0cZhengfeng Li\uff08lizhengfeng.lzf@alibaba-inc.com\uff09<br \/>\n\u9884\u5370\u672c\u30022026 \u5e74 2 \u6708 2 \u65e5\u3002<br \/>\n* \u540c\u7b49\u8d21\u732e\u3002<\/p>\n<h2>\u6458\u8981<\/h2>\n<p>\u9ad8\u8d28\u91cf\u8bc4\u6d4b\u57fa\u51c6\u5bf9\u4e8e\u5c06\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u90e8\u7f72\u5230\u7279\u5b9a\u5e94\u7528\u9886\u57df\u81f3\u5173\u91cd\u8981\u3002\u7136\u800c\uff0c\u73b0\u6709\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\uff08Automated Code Review, ACR\uff09\u57fa\u51c6\u5b58\u5728\u4e24\u9879\u5173\u952e\u5c40\u9650\uff1a\u5176\u4e00\uff0c\u4f9d\u8d56\u4ece\u539f\u59cb Pull Request\uff08PR\uff09\u8bc4\u8bba\u4e2d\u63d0\u53d6\u7684\u566a\u58f0\u3001\u4e0d\u5b8c\u6574\u7684 Ground Truth\uff0c\u4ece\u800c\u9650\u5236\u4e86\u95ee\u9898\u68c0\u6d4b\u7684\u8986\u76d6\u8303\u56f4\uff1b\u5176\u4e8c\uff0c\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u7f3a\u4e4f\u591a\u8bed\u8a00\u652f\u6301\uff0c\u4ece\u800c\u9650\u5236\u4e86\u8bc4\u6d4b\u7ed3\u679c\u7684\u53ef\u63a8\u5e7f\u6027\u3002\u4e3a\u5e94\u5bf9\u8fd9\u4e9b\u6311\u6218\uff0c\u6211\u4eec\u63d0\u51fa AACR-Bench\uff0c\u8fd9\u662f\u4e00\u4e2a\u5728\u591a\u79cd\u7f16\u7a0b\u8bed\u8a00\u4e0a\u63d0\u4f9b\u5b8c\u6574\u8de8\u6587\u4ef6\u4e0a\u4e0b\u6587\u7684\u7efc\u5408\u6027\u57fa\u51c6\u3002\u4e0e\u4f20\u7edf\u6570\u636e\u96c6\u4e0d\u540c\uff0cAACR-Bench \u91c7\u7528\u300cAI \u8f85\u52a9\u3001\u4eba\u7c7b\u4e13\u5bb6\u6838\u9a8c\u300d\u7684\u6807\u6ce8\u6d41\u6c34\u7ebf\uff0c\u4ee5\u53d1\u73b0\u539f\u59cb PR \u4e2d\u5e38\u88ab\u5ffd\u7565\u7684\u6f5c\u5728\u7f3a\u9677\uff0c\u4f7f\u95ee\u9898\u8986\u76d6\u7387\u63d0\u5347\u4e86 285%\u3002\u5728 AACR-Bench \u4e0a\u5bf9\u4e3b\u6d41 LLM \u7684\u5e7f\u6cdb\u8bc4\u6d4b\u8868\u660e\uff0c\u7531\u4e8e\u6570\u636e\u9650\u5236\uff0c\u4ee5\u5f80\u8bc4\u4f30\u53ef\u80fd\u8bef\u5224\u6216\u4ec5\u90e8\u5206\u523b\u753b\u4e86\u6a21\u578b\u80fd\u529b\u3002\u6211\u4eec\u7684\u5de5\u4f5c\u4e3a ACR \u8bc4\u6d4b\u5efa\u7acb\u4e86\u66f4\u4e25\u683c\u7684\u6807\u51c6\uff0c\u5e76\u5c31\u57fa\u4e8e LLM \u7684 ACR \u7ed9\u51fa\u65b0\u7684\u6d1e\u89c1\uff1a\u4e0a\u4e0b\u6587\u7684\u7c92\u5ea6\/\u5c42\u7ea7\u4ee5\u53ca\u68c0\u7d22\u65b9\u6cd5\u7684\u9009\u62e9\u4f1a\u663e\u8457\u5f71\u54cd ACR \u6027\u80fd\uff0c\u4e14\u8fd9\u79cd\u5f71\u54cd\u968f LLM\u3001\u7f16\u7a0b\u8bed\u8a00\u4ee5\u53ca LLM \u4f7f\u7528\u8303\u5f0f\uff08\u4f8b\u5982\u662f\u5426\u91c7\u7528 Agent \u67b6\u6784\uff09\u800c\u53d8\u5316\u3002\u8bc4\u6d4b\u96c6\u7684\u4ee3\u7801\u3001\u6570\u636e\u53ca\u5176\u4ed6\u4ea7\u7269\u89c1 Github<sup id=\"fnref:1\"><a class=\"footnote-ref\" href=\"1\" target=\"_blank\"  rel=\"nofollow\" >1<\/a><\/sup>\u3002<\/p>\n<h2>1 \u5f15\u8a00<\/h2>\n<p>\u5982\u4eca\uff0c\u5728\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u8d4b\u80fd\u4e0b\uff0c\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\uff08ACR\uff09\u6280\u672f\u5df2\u88ab\u5e7f\u6cdb\u7814\u7a76\u4e0e\u91c7\u7528 (Hou et al. 2024; Zhang et al. 2023)\u3002\u501f\u52a9 LLM \u51fa\u8272\u7684\u7406\u89e3\u80fd\u529b\uff0c\u7814\u7a76\u8005\u63a2\u7d22\u4e86\u591a\u79cd\u7aef\u5230\u7aef\u751f\u6210\u8bc4\u5ba1\u8bc4\u8bba\u7684\u65b9\u6cd5\uff0c\u4f8b\u5982 Tufano \u7b49\u4eba\u63d0\u51fa\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u65b9\u6848\u3001\u517c\u987e\u6210\u672c\u4e0e\u6027\u80fd\u7684\u5fae\u8c03\u7b56\u7565 (Lu et al. 2023)\uff0c\u4ee5\u53ca\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u4f18\u5316\u673a\u5236 (Le et al. 2022)\u3002\u6700\u8fd1\uff0c\u57fa\u4e8e Agent \u7684\u4ee3\u7801\u8bc4\u5ba1\u7cfb\u7edf\u4e5f\u5f00\u59cb\u51fa\u73b0 (Guo et al. 2025b; Ren et al. 2025)\u3002<\/p>\n<p>ACR \u7814\u7a76\u4e0e\u843d\u5730\u7684\u7e41\u8363\uff0c\u52a0\u5267\u4e86\u5bf9\u80fd\u591f\u5168\u9762\u8bc4\u4f30\u5176\u6027\u80fd\u7684\u53ef\u4fe1\u57fa\u51c6\u7684\u9700\u6c42\u3002\u7136\u800c\uff0c\u5f53\u524d\u8bc4\u6d4b\u6846\u67b6\u4ecd\u5b58\u5728\u5173\u952e\u5c40\u9650\uff0c\u53ef\u6982\u62ec\u4e3a\u4e24\u4e2a\u4e3b\u8981\u95ee\u9898\u3002<\/p>\n<ul>\n<li><strong>\u95ee\u9898\u6807\u6ce8\u4e0d\u5b8c\u6574\u3002<\/strong> \u603b\u4f53\u800c\u8a00\uff0c\u73b0\u6709\u57fa\u51c6\u76f4\u63a5\u5c06\u771f\u5b9e\u5386\u53f2 PR \u4e2d\u7684\u539f\u59cb\u8bc4\u5ba1\u8bc4\u8bba\u4f5c\u4e3a Ground Truth\uff0c\u5bfc\u81f4\u6570\u636e\u96c6\u5728\u95ee\u9898\u8986\u76d6\u4e0a\u5b58\u5728\u56fa\u6709\u5c40\u9650 (Liu et al. 2025; Khoshnoud et al. 2022)\u3002\u8fd9\u4f7f\u5f97\u65e0\u6cd5\u771f\u5b9e\u3001\u5fe0\u5b9e\u5730\u523b\u753b\u6a21\u578b\u5728\u771f\u5b9e\u4ee3\u7801\u8bc4\u5ba1\u573a\u666f\u4e2d\u53d1\u73b0\u6f5c\u5728\u95ee\u9898\u7684\u80fd\u529b\u3002<\/li>\n<li><strong>\u4e0a\u4e0b\u6587\u8303\u56f4\u53d7\u9650\u3002<\/strong> \u8bb8\u591a\u4ee3\u7801\u7f3a\u9677\u672c\u8d28\u4e0a\u662f\u8de8\u6587\u4ef6\u7684\uff0c\u8bc4\u5ba1\u7cfb\u7edf\u9700\u8981\u8bbf\u95ee\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u624d\u80fd\u51c6\u786e\u68c0\u6d4b\u3002\u5c3d\u7ba1\u5df2\u6709\u90e8\u5206\u57fa\u51c6\u63d0\u4f9b\u8de8\u6587\u4ef6\u4e0a\u4e0b\u6587\u611f\u77e5 (Zeng et al. 2025; Guo et al. 2025a)\uff0c\u4f46\u5b83\u4eec\u5927\u591a\u5c40\u9650\u4e8e\u5355\u4e00\u7f16\u7a0b\u8bed\u8a00\uff08\u5982 Python\uff09\u3002\u8fd9\u79cd\u8bed\u8a00\u7279\u5b9a\u7684\u5173\u6ce8\u4e0d\u4ec5\u9650\u5236\u4e86\u8bc4\u6d4b\u53d1\u73b0\u7684\u53ef\u63a8\u5e7f\u6027\uff0c\u8fd8\u53ef\u80fd\u5f15\u5165\u4e0e\u8be5\u8bed\u8a00\u8bed\u8a00\u5b66\u7279\u5f81\u7ed1\u5b9a\u7684\u7ed3\u6784\u6027\u504f\u5dee\u3002\u56e0\u6b64\uff0c\u8fd9\u7c7b\u57fa\u51c6\u65e0\u6cd5\u5145\u5206\u4ee3\u8868\u5f53\u4ee3\u8f6f\u4ef6\u5f00\u53d1\u7684\u591a\u8bed\u8a00\u73b0\u5b9e\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u4e0d\u8db3\u5171\u540c\u524a\u5f31\u4e86\u5f53\u524d ACR \u8bc4\u6d4b\u7684\u53ef\u4fe1\u5ea6\u4e0e\u8986\u76d6\u8303\u56f4\uff0c\u51f8\u663e\u4e86\u6784\u5efa\u66f4\u5168\u9762\u3001\u66f4\u53ef\u4fe1\u57fa\u51c6\u7684\u5fc5\u8981\u6027\u3002<\/p>\n<p>\u4e3a\u5e94\u5bf9\u4e0a\u8ff0\u6311\u6218\uff0c\u6211\u4eec\u63d0\u51fa AACR-Bench\u2014\u2014\u4e00\u4e2a\u652f\u6301\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u611f\u77e5\u7684\u591a\u8bed\u8a00 ACR \u57fa\u51c6\u3002\u9996\u5148\uff0c\u9488\u5bf9\u73b0\u6709\u6570\u636e\u96c6\u95ee\u9898\u6807\u6ce8\u6709\u9650\u7684\u95ee\u9898\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u9ad8\u8d28\u91cf\u6df7\u5408\u8bed\u6599\u3002\u9664\u4ece GitHub PR \u4e2d\u6536\u96c6 391 \u6761\u771f\u5b9e\u8bc4\u5ba1\u8bc4\u8bba\u5916\uff0c\u6211\u4eec\u8fd8\u7eb3\u5165\u4e86\u5927\u91cf\u4eba\u5de5\u95ee\u9898\u6807\u6ce8\uff1a80 \u540d\u9ad8\u7ea7\u8f6f\u4ef6\u5de5\u7a0b\u5e08\uff08\u6bcf\u4eba\u5177\u5907 2 \u5e74\u4ee5\u4e0a\u5de5\u4e1a\u7ecf\u9a8c\uff09\u4ed4\u7ec6\u5ba1\u9605\u4e86\u7531\u4e24\u5957 ACR \u7cfb\u7edf\u5728\u516d\u4e2a LLM \u4e0a\u751f\u6210\u7684 2,145 \u6761\u8bc4\u8bba\u3002\u8be5\u6807\u6ce8\u7b56\u7565\u663e\u8457\u63d0\u5347\u4e86\u95ee\u9898\u8986\u76d6\u7387\uff0c\u4f7f AACR-Bench \u76f8\u6bd4\u5e38\u89c4\u6570\u636e\u96c6\u80fd\u591f\u66f4\u51c6\u786e\u3001\u66f4\u5168\u9762\u5730\u8bc4\u4f30\u6a21\u578b\u53d1\u73b0\u6f5c\u5728\u95ee\u9898\u7684\u80fd\u529b\u3002\u5176\u6b21\uff0c\u9488\u5bf9\u4e0a\u4e0b\u6587\u4e0e\u8bed\u8a00\u591a\u6837\u6027\u95ee\u9898\uff0cAACR-Bench \u63d0\u4f9b\u5b8c\u6574\u7684\u4ed3\u5e93\u7ea7\u4f9d\u8d56\u4fe1\u606f\uff0c\u5e76\u5e7f\u6cdb\u8986\u76d6 10 \u79cd\u4e3b\u6d41\u7f16\u7a0b\u8bed\u8a00\u3002\u5b8c\u6574\u7684 AACR-Bench \u6570\u636e\u96c6\u53ef\u5728\u672c\u6587\u8865\u5145\u6750\u6599\u4e2d\u83b7\u53d6\uff0c\u5e76\u5c06\u968f\u540e\u5f00\u6e90\u3002<\/p>\n<p>\u6211\u4eec\u7684\u4e3b\u8981\u8d21\u732e\u6982\u62ec\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\u6211\u4eec\u5c06\u591a\u6a21\u578b\u751f\u6210\u4e0e\u5927\u89c4\u6a21\u4eba\u5de5\u6807\u6ce8\u76f8\u7ed3\u5408\uff0c\u4ee5\u63d0\u5347\u95ee\u9898\u8986\u76d6\u7387\uff0c\u4ece\u800c\u5efa\u7acb\u66f4\u597d\u7684 Ground Truth \u6570\u636e\u96c6\u3002<\/li>\n<li>\u6211\u4eec\u63d0\u51fa AACR-Bench\uff0c\u8fd9\u662f\u9762\u5411 LLM \u8d4b\u80fd ACR \u4efb\u52a1\u7684\u9996\u4e2a\u591a\u8bed\u8a00\u3001\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u611f\u77e5\u57fa\u51c6\u3002<\/li>\n<li>\u6211\u4eec\u57fa\u4e8e AACR-Bench \u5bf9\u4e3b\u6d41 LLM \u8fdb\u884c\u4e86\u5168\u9762\u5b9e\u8bc1\u8bc4\u6d4b\uff0c\u5e76\u63ed\u793a\u4e86\u5173\u4e8e\u4e3b\u6d41 LLM \u7684 ACR \u80fd\u529b\u7684\u65b0\u6d1e\u89c1\u3002<\/li>\n<\/ul>\n<h2>2 \u76f8\u5173\u5de5\u4f5c<\/h2>\n<p>\u6211\u4eec\u4ece\u4e24\u4e2a\u5173\u952e\u7ef4\u5ea6\u5b9a\u4f4d\u672c\u7814\u7a76\uff1a\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684 ACR \u73b0\u72b6\uff0c\u4ee5\u53ca\u76f8\u5e94\u4e13\u7528\u57fa\u51c6\u65b9\u6cd5\u7684\u6f14\u8fdb\u683c\u5c40\u3002<\/p>\n<h3>2.1 \u57fa\u4e8e LLM \u7684\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1<\/h3>\n<p>LLM \u7684\u8fd1\u671f\u8fdb\u5c55\u63a8\u52a8\u4e86\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\uff08ACR\uff09\u7814\u7a76\u7684\u589e\u957f\uff0c\u65b9\u6cd5\u6db5\u76d6\u82e5\u5e72\u7c7b\u522b\u3002\u65e9\u671f\u5de5\u4f5c\u5305\u62ec\u57fa\u4e8e T5 \u7684 ACR (Tufano et al. 2022) \u4ee5\u53ca CodeReviewer (Li et al. 2022)\uff0c\u540e\u8005\u5728\u5927\u89c4\u6a21 diff\u2013\u8bc4\u8bba\u5bf9\u4e0a\u9884\u8bad\u7ec3\u3002\u4e3a\u63d0\u5347\u6548\u7387\uff0cLlama-Reviewer (Lu et al. 2023) \u91c7\u7528\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff08Parameter-Efficient Fine-Tuning\uff09\uff0c\u800c Sun \u7b49\u4eba\u5f15\u5165\u4e86\u901a\u8fc7\u8fed\u4ee3\u53cd\u9988\u5b9e\u73b0\u6a21\u578b\u6301\u7eed\u6f14\u5316\u7684\u300c\u6570\u636e\u98de\u8f6e\u300d\uff08Data Flywheel\uff09(Sun et al. 2025; Yu et al. 2024a)\u3002\u8fd1\u671f\u5de5\u4f5c\u5c06\u6a21\u578b\u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\uff1aYu \u7b49\u4eba (Yu et al. 2024b) \u4f7f\u7528 Kahneman\u2013Tversky Optimization \u4ee5\u589e\u5f3a\u8bc4\u5ba1\u6548\u7528\uff0cKapadnis \u7b49\u4eba (Kapadnis et al. 2025) \u5c06\u9759\u6001\u5206\u6790\u4e0e Direct Preference Optimization \u7ed3\u5408\uff0c\u4ee5\u66f4\u597d\u5730\u68c0\u6d4b\u6f5c\u5728\u7f3a\u9677\u3002\u4e3a\u5e94\u5bf9\u6709\u9650\u4e0a\u4e0b\u6587\uff0cZhang \u7b49\u4eba (Zhang et al. 2025) \u5e94\u7528\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval-Augmented Generation\uff09\u4ee5\u7eb3\u5165\u9879\u76ee\u7ea7\u8bed\u4e49\u3002\u4e3a\u8fdb\u4e00\u6b65\u6269\u5c55\u63a8\u7406\u6df1\u5ea6\uff0c\u57fa\u4e8e Agent \u7684\u6846\u67b6\u91c7\u7528\u591a\u667a\u80fd\u4f53\u534f\u4f5c (Ren et al. 2025; Li et al. 2025; Sharanarthi &amp; Polineni 2025) \u4ee5\u53ca\u8fa9\u8bba\u7b49\u8fa9\u8bc1\u4ea4\u4e92 (Tang et al. 2024)\uff0c\u4ee5\u4ea7\u751f\u66f4\u5168\u9762\u3001\u66f4\u5ba2\u89c2\u7684\u8bc4\u5ba1\u3002<\/p>\n<h3>2.2 \u4ee3\u7801\u8bc4\u5ba1\u57fa\u51c6<\/h3>\n<p>\u57fa\u51c6\u5bf9\u4e8e\u754c\u5b9a LLM \u7684\u80fd\u529b\u8fb9\u754c\u5e76\u5f15\u5bfc\u7b97\u6cd5\u6f14\u8fdb\u81f3\u5173\u91cd\u8981 (Jimenez et al. 2023)\u3002\u5728 ACR \u4e2d\uff0c\u591a\u6837\u5316\u65b9\u6cd5\u7684\u8fc5\u901f\u6d8c\u73b0\u51f8\u663e\u4e86\u5bf9\u6807\u51c6\u5316\u57fa\u51c6\u7684\u9700\u6c42\uff0c\u4ee5\u7cfb\u7edf\u8bc4\u4f30\u6027\u80fd\u5e76\u6307\u5bfc\u540e\u7eed\u4f18\u5316\u3002\u76ee\u524d ACR \u57fa\u51c6\u4ecd\u5904\u4e8e\u840c\u82bd\u9636\u6bb5\uff0c\u76f8\u5173\u5de5\u4f5c\u53ef\u5206\u4e3a\u4e24\u7c7b\uff1a\u5355\u9879\u7814\u7a76\u9644\u5e26\u7684\u6570\u636e\u96c6\uff0c\u4ee5\u53ca\u4e13\u7528\u8bc4\u6d4b\u6570\u636e\u96c6\u3002\u4f7f\u7528\u6700\u5e7f\u7684\u9644\u5e26\u6570\u636e\u96c6\u662f CodeReviewer (Li et al. 2022)\uff0c\u88ab\u8bb8\u591a\u8fd1\u671f\u5de5\u4f5c\u91c7\u7528 (Lu et al. 2023; Yu et al. 2024a; Yu et al. 2024b; Kapadnis et al. 2025; Ren et al. 2025; Li et al. 2025)\u3002\u7136\u800c\uff0c\u5b83\u4ec5\u63d0\u4f9b diff \u7ea7\u4ee3\u7801\u7247\u6bb5\uff0c\u7f3a\u5c11\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\uff0c\u76f8\u5bf9\u751f\u4ea7\u73af\u5883\u7684\u771f\u5b9e\u6027\u6709\u9650\u3002\u5c11\u6570\u7814\u7a76 (Sun et al. 2025; Sharanarthi &amp; Polineni 2025) \u5b8c\u5168\u653e\u5f03\u9759\u6001\u6570\u636e\u96c6\uff0c\u4ec5\u4f9d\u8d56\u7ebf\u4e0a\u7528\u6237\u53cd\u9988\u3002\u4e13\u7528\u57fa\u51c6\u5305\u62ec SWR-Bench (Zeng et al. 2025)\uff08\u6e90\u81ea SWE-Bench \u7684 12 \u4e2a Python \u9879\u76ee\uff09\u4ee5\u53ca CodeFuse-CR-Bench (Guo et al. 2025a)\uff0870 \u4e2a Python \u9879\u76ee\uff09\u3002\u4e8c\u8005\u867d\u63d0\u4f9b\u5b8c\u6574\u4ed3\u5e93\u4e0a\u4e0b\u6587\uff0c\u4f46\u4ec5\u805a\u7126 Python\uff0c\u53ef\u63a8\u5e7f\u6027\u53d7\u9650\u3002ContextCRBench (Hu et al. 2025) \u63d0\u5347\u4e86\u8bed\u8a00\u591a\u6837\u6027\uff0890 \u4e2a\u4ed3\u5e93\u30019 \u79cd\u8bed\u8a00\uff09\uff0c\u4f46\u5c06\u4e0a\u4e0b\u6587\u9650\u5236\u5728\u6587\u4ef6\u7ea7\uff0c\u65e0\u6cd5\u8bc4\u4f30\u590d\u6742\u7f3a\u9677\u6240\u9700\u7684\u8de8\u6587\u4ef6\uff08\u5373\u4ed3\u5e93\u7ea7\uff09\u63a8\u7406\u3002<\/p>\n<p>\u53d7\u4e0a\u8ff0\u683c\u5c40\u542f\u53d1\uff0c\u672c\u6587\u5f15\u5165\u9996\u4e2a\u517c\u5177\u591a\u8bed\u8a00\u4e0e\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u7684 ACR \u8bc4\u6d4b\u6570\u636e\u96c6\uff0c\u5373 AACR-Bench\uff0c\u65e8\u5728\u5efa\u7acb\u66f4\u8d34\u8fd1\u771f\u5b9e\u751f\u4ea7\u73af\u5883\u7684\u57fa\u51c6\u3002<\/p>\n<h2>3 AACR-Bench<\/h2>\n<h3>3.1 AACR-Bench \u6982\u89c8<\/h3>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 1\uff1aAACR-Bench \u6982\u89c8\u3002AACR-Bench \u5305\u542b\u4ece 50 \u4e2a\u70ed\u95e8\u4ed3\u5e93\u63d0\u53d6\u5e76\u6574\u7406\u7684 200 \u4e2a PR \u4e0e 1,505 \u6761\u7ec6\u7c92\u5ea6\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u8986\u76d6 10 \u79cd\u4e3b\u6d41\u7f16\u7a0b\u8bed\u8a00\u3002\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig01.png\" \/><\/p>\n<p><strong>\u56fe 1\uff1a<\/strong> AACR-Bench \u6982\u89c8\u3002AACR-Bench \u5305\u542b\u4ece 50 \u4e2a\u70ed\u95e8\u4ed3\u5e93\u63d0\u53d6\u5e76\u6574\u7406\u7684 200 \u4e2a PR \u4e0e 1,505 \u6761\u7ec6\u7c92\u5ea6\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u8986\u76d6 10 \u79cd\u4e3b\u6d41\u7f16\u7a0b\u8bed\u8a00\u3002<\/p>\n<p>AACR-Bench \u662f\u4e00\u4e2a\u9762\u5411\u8bc4\u6d4b ACR \u65b9\u6cd5\/\u7cfb\u7edf\u7aef\u5230\u7aef\u6027\u80fd\u7684\u4ed3\u5e93\u7ea7\u57fa\u51c6\u3002\u56fe 1 \u5c55\u793a\u4e86 AACR-Bench \u7684\u6982\u89c8\u3002\u603b\u4f53\u800c\u8a00\uff0cAACR-Bench \u5305\u542b\u4ece 50 \u4e2a\u70ed\u95e8\u4ed3\u5e93\u63d0\u53d6\u5e76\u6574\u7406\u7684 200 \u4e2a PR \u4e0e 1,505 \u6761\u7ec6\u7c92\u5ea6\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u8986\u76d6 10 \u79cd\u4e3b\u6d41\u7f16\u7a0b\u8bed\u8a00\u3002\u6b64\u5916\uff0cAACR-Bench \u878d\u5408\u4e86\u6a21\u578b\u589e\u5f3a\u7684\u4eba\u5de5\u8bc4\u5ba1\u4e0e\u5b8c\u5168\u7531\u6a21\u578b\u751f\u6210\u7684\u8bc4\u5ba1\uff0c\u4e8c\u8005\u5747\u7ecf\u8fc7\u4eba\u7c7b\u4e13\u5bb6\u6807\u6ce8\u7684\u4e25\u683c\u6838\u9a8c\uff0c\u4ee5\u786e\u4fdd\u57fa\u51c6\u53ef\u4fe1\u5ea6\u3002<\/p>\n<p><strong>\u8868 1\uff1a<\/strong> \u8bc4\u5ba1\u8bc4\u8bba\u7684\u4e0a\u4e0b\u6587\u8303\u56f4\u7c7b\u522b<\/p>\n<table>\n<thead>\n<tr>\n<th>\u4e0a\u4e0b\u6587\u5c42\u7ea7<\/th>\n<th>\u63cf\u8ff0<\/th>\n<th style=\"text-align: right\">\u8bc4\u8bba\u6570<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Diff<\/td>\n<td>\u4ec5\u67e5\u770b\u5f53\u524d diff hunk \u5373\u53ef\u7ed9\u51fa\u7684\u8bc4\u5ba1\u8bc4\u8bba<\/td>\n<td style=\"text-align: right\">754<\/td>\n<\/tr>\n<tr>\n<td>File<\/td>\n<td>\u9700\u8981\u5305\u542b\u8be5 diff hunk \u7684\u6574\u4e2a\u6587\u4ef6\u4e0a\u4e0b\u6587\u7684\u8bc4\u5ba1\u8bc4\u8bba<\/td>\n<td style=\"text-align: right\">518<\/td>\n<\/tr>\n<tr>\n<td>Repo<\/td>\n<td>\u9700\u8981\u4ed3\u5e93\u8303\u56f4\u4e0a\u4e0b\u6587\u7684\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u5305\u62ec PR \u5143\u6570\u636e\u4ee5\u53ca\u4ee3\u7801\u5e93\u4e2d\u5176\u4ed6\u6587\u4ef6\u7684\u5185\u5bb9<\/td>\n<td style=\"text-align: right\">233<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4e3a\u4e0e\u73b0\u4ee3\u4ee3\u7801\u8bc4\u5ba1\u4ee5 diff \u4e3a\u5bfc\u5411\u7684\u8303\u5f0f\u5bf9\u9f50\uff0cAACR-Bench \u5c06\u6bcf\u4e2a PR \u6307\u5b9a\u4e3a\u4e00\u4e2a\u8bc4\u6d4b\u5355\u5143\u3002\u6267\u884c\u8fc7\u7a0b\u4e2d\uff0c\u88ab\u8bc4\u6d4b\u7684 ACR \u65b9\u6cd5\u904d\u5386\u4e00\u4e2a PR \u5185\u7684\u5168\u90e8 diff Hunk \u4ee5\u751f\u6210\u8bc4\u5ba1\u8bc4\u8bba\u3002\u6027\u80fd\u901a\u8fc7\u5c06\u8fd9\u4e9b\u751f\u6210\u8bc4\u8bba\u4e0e Ground Truth\uff08\u5171 1,505 \u6761\uff09\u5339\u914d\uff0c\u5e76\u8ba1\u7b97\u7279\u5b9a\u51c6\u786e\u7387\u6307\u6807\uff08\u4f8b\u5982 Precision\u3001Recall \u4e0e F1-score\uff09\u6765\u8bc4\u4f30\u3002<\/p>\n<h3>3.2 \u6570\u636e\u96c6\u6784\u5efa<\/h3>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 2\uff1aAACR-Bench \u4e2d\u8bc4\u5ba1\u8bc4\u8bba\u7684\u5206\u5e03\u3002TS\u3001JS \u4e0e Py \u5206\u522b\u4ee3\u8868 TypeScript\u3001JavaScript \u4e0e Python\u3002\u201cAug\u201d \u8868\u793a\u7531\u539f\u59cb PR \u8bc4\u5ba1\u589e\u5f3a\u5f97\u5230\u7684\u8bc4\u8bba\uff0c\u201cGen\u201d \u8868\u793a\u7531 6 \u4e2a LLM \u751f\u6210\u7684\u8bc4\u8bba\u3002\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig02.png\" \/><\/p>\n<p><strong>\u56fe 2\uff1a<\/strong> AACR-Bench \u4e2d\u8bc4\u5ba1\u8bc4\u8bba\u7684\u5206\u5e03\u3002TS\u3001JS \u4e0e Py \u5206\u522b\u4ee3\u8868 TypeScript\u3001JavaScript \u4e0e Python\u3002\u201cAug\u201d \u8868\u793a\u7531\u539f\u59cb PR \u8bc4\u5ba1\u589e\u5f3a\u5f97\u5230\u7684\u8bc4\u8bba\uff0c\u201cGen\u201d \u8868\u793a\u7531 6 \u4e2a LLM \u751f\u6210\u7684\u8bc4\u8bba\u3002<\/p>\n<p>\u5982\u7b2c 1 \u8282\u6240\u8ff0\uff0c\u73b0\u6709\u57fa\u51c6\u5b58\u5728\u5355\u4e00\u8bed\u8a00\u5c40\u9650\uff0c\u4ee5\u53ca\u4ec5\u4ece\u5386\u53f2 PR \u6570\u636e\u5f97\u5230\u7684\u4e0d\u5b8c\u6574 Ground Truth\u3002\u4e3a\u5f25\u5408\u8fd9\u4e9b\u5dee\u8ddd\uff0c\u6211\u4eec\u6784\u5efa\u4e86 AACR-Bench\u2014\u2014\u4e00\u4e2a\u591a\u8bed\u8a00\u3001\u4ed3\u5e93\u7ea7\u57fa\u51c6\u3002\u6838\u5fc3\u8003\u8651\u4e0e\u5904\u7406\u5982\u4e0b\u3002<\/p>\n<h5>\u7f16\u7a0b\u8bed\u8a00\u4e0e\u4ed3\u5e93\u7684\u9009\u62e9<\/h5>\n<p>\u4e3a\u786e\u4fdd\u8bc4\u6d4b\u57fa\u51c6\u7684\u65f6\u6548\u6027\u4e0e\u591a\u6837\u6027\uff0c\u6211\u4eec\u4f9d\u636e StackOverflow Developer Survey 2025<sup id=\"fnref:2\"><a class=\"footnote-ref\" href=\"2\" target=\"_blank\"  rel=\"nofollow\" >2<\/a><\/sup> \u9009\u53d6\u4e86\u6392\u540d\u524d 10 \u7684\u7f16\u7a0b\u8bed\u8a00\uff0c\u5373\uff1aJavaScript\u3001Python\u3001TypeScript\u3001Java\u3001C#\u3001C++\u3001C\u3001PHP\u3001Go \u4e0e Rust\u3002\u5bf9\u6bcf\u79cd\u8bed\u8a00\uff0c\u6211\u4eec\u5728 GitHub \u4e0a\u8bc6\u522b\u5728 2024 \u5e74 12 \u6708 1 \u65e5\u81f3 2025 \u5e74 12 \u6708 1 \u65e5\u671f\u95f4\u65b0\u589e star \u6570\u4e0e\u5df2\u5173\u95ed PR \u6570\u5747\u8fdb\u5165\u524d 2,000 \u7684\u5019\u9009\u4ed3\u5e93\u3002\u4ece\u8be5\u5019\u9009\u6c60\u4e2d\uff0c\u6211\u4eec\u9009\u53d6\u65b0\u589e star \u6570\u6700\u9ad8\u7684\u524d 5 \u4e2a\u4ed3\u5e93\u3002\u56e0\u6b64\uff0c\u6700\u7ec8\u6570\u636e\u96c6\u5305\u542b\u8fd9 10 \u79cd\u4e0d\u540c\u8bed\u8a00\u4e0a\u7684 50 \u4e2a\u4ed3\u5e93\uff08\u6bcf\u79cd\u8bed\u8a00 5 \u4e2a\u4ed3\u5e93\uff09\uff0c\u4f5c\u4e3a\u63d0\u53d6 PR \u4e0e\u8bc4\u5ba1\u8bc4\u8bba\u7684\u539f\u59cb\u6765\u6e90\u3002<\/p>\n<h5>PR \u8fc7\u6ee4\u4e0e\u8bc4\u8bba\u589e\u5f3a<\/h5>\n<p>\u5728\u4ece 50 \u4e2a\u9009\u5b9a\u4ed3\u5e93\u63d0\u53d6\u7684 PR \u4e2d\uff0c\u6211\u4eec\u4e22\u5f03\u672a\u8bc6\u522b\u51fa\u660e\u786e\u95ee\u9898\u7684\u8bc4\u8bba\u3002\u4e3a\u4fdd\u8bc1\u6570\u636e\u8d28\u91cf\uff0c\u6211\u4eec\u5e94\u7528\u4e86\u4e94\u6761\u8fc7\u6ee4\u6807\u51c6\uff1a(1) PR \u6807\u9898\u4e0e\u63cf\u8ff0\u5fc5\u987b\u4e3a\u82f1\u6587\uff1b(2) \u53d8\u66f4\u4ee3\u7801\u884c\u6570\u5fc5\u987b $leq 1{,}000$\uff08\u4e0e Google \u5173\u4e8e\u6709\u6548\u8bc4\u5ba1\u7684\u4ee3\u7801\u8bc4\u5ba1\u5b9e\u8df5\u5bf9\u9f50 (Solmaz 2025)\uff09\uff1b(3) \u88ab\u4fee\u6539\u6587\u4ef6\u7684\u4e3b\u8981\u7f16\u7a0b\u8bed\u8a00\u5fc5\u987b\u4e0e\u4ed3\u5e93\u4e3b\u8bed\u8a00\u5339\u914d\uff1b(4) \u4e00\u4e2a PR \u5fc5\u987b\u5305\u542b $&gt;2$ \u6761\u884c\u5185\u8bc4\u8bba\uff0c\u5176\u4e2d\u5305\u62ec\u81f3\u5c11\u4e00\u6761\u5bfc\u81f4\u4ee3\u7801\u4fee\u6539\u7684\u88ab\u91c7\u7eb3\u8bc4\u8bba\uff1b(5) \u8131\u79bb\u9879\u76ee\u4e1a\u52a1\u4e0a\u4e0b\u6587\u6216\u7f3a\u4e4f\u8bed\u4e49\u542b\u4e49\u7684\u53d8\u66f4\u88ab\u6392\u9664\u3002\u6700\u540e\uff0c\u4e3a\u786e\u4fdd\u57fa\u51c6\u7684\u4ee3\u8868\u6027\u4e0e\u591a\u6837\u6027\uff0c\u6211\u4eec\u57fa\u4e8e\u4ed3\u5e93\u3001PR \u95ee\u9898\u57df (Jimenez et al. 2023; Guo et al. 2025a) \u4ee5\u53ca\u53d8\u66f4\u89c4\u6a21\uff0c\u5bf9\u8fc7\u6ee4\u540e\u7684\u5019\u9009\u8fdb\u884c\u5206\u5c42\u62bd\u6837\uff0c\u6784\u5efa\u4e86\u5305\u542b 200 \u4e2a PR \u7684\u6838\u5fc3\u6570\u636e\u96c6\u3002\u7531\u4e8e\u4ee3\u7801\u8bc4\u5ba1\u5e38\u6d89\u53ca\u591a\u8f6e\u5bf9\u8bdd\uff0c\u76f4\u63a5\u4f7f\u7528\u539f\u59cb\u8bc4\u8bba\u53ef\u80fd\u4e22\u5931\u4e0a\u4e0b\u6587\u6216\u5f15\u5165\u566a\u58f0\u3002\u6211\u4eec\u805a\u7126\u6bcf\u4e2a PR \u4e2d\u884c\u5185\u8bc4\u8bba\u6700\u591a\u7684\u4fee\u8ba2\u7248\u672c\uff0c\u4f7f\u7528 LLM \u5bf9\u8bc4\u5ba1\u7ebf\u7a0b\u8fdb\u884c\u6df1\u5ea6\u8bed\u4e49\u5206\u6790\u3002\u8fd9\u4f7f\u5f97\u6211\u4eec\u80fd\u591f\u4ece\u591a\u8f6e\u4ea4\u4e92\u4e2d\u63d0\u53d6\u5df2\u786e\u8ba4\u7684\u4ee3\u7801\u7f3a\u9677\uff0c\u5e76\u5c06\u5176\u91cd\u7ec4\u4e3a\u300c\u589e\u5f3a\u8bc4\u5ba1\u8bc4\u8bba\u300d\uff08Augmented Review Comments\uff09\u3002<\/p>\n<h5>\u8bc4\u5ba1\u8865\u5168\u4e0e\u4e13\u5bb6\u6807\u6ce8<\/h5>\n<p>\u4e3a\u5e94\u5bf9 GitHub PR \u5e38\u88ab\u8bc4\u5ba1\u4e0d\u8db3\u6240\u5bfc\u81f4\u7684\u6807\u7b7e\u4e0d\u5b8c\u6574\uff08Label Incompleteness\uff09(Bacchelli &amp; Bird 2013)\uff0c\u6211\u4eec\u5229\u7528 LLM \u5168\u9762\u8865\u5145\u6bcf\u4e2a PR \u7684\u8bc4\u5ba1\u8bc4\u8bba\u3002\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u7531 6 \u4e2a\u4e3b\u6d41\u5f00\u6e90\u4e0e\u95ed\u6e90\u6a21\u578b\u7ec4\u6210\u7684\u751f\u6210\u77e9\u9635\uff08Claude-4.5-Sonnet\u3001Qwen3-Coder-480B-A35B-Instruct (Team 2025)\u3001GPT-5.2\u3001Deepseek-V3.2 (DeepSeek-AI 2025)\u3001GLM-4.7 (Team et al. 2025)\u3001Gemini-3-Pro\uff09\uff0c\u4ee5\u51cf\u8f7b\u5355\u6a21\u578b\u504f\u5dee\u5e76\u786e\u4fdd\u8f93\u51fa\u591a\u6837\u6027\u3002\u8bc4\u5ba1\u8bc4\u8bba\u901a\u8fc7\u4e24\u4e2a\u5f02\u6784\u6846\u67b6\u5e76\u884c\u751f\u6210\uff1a\u5185\u90e8\u8bc4\u5ba1\u7cfb\u7edf\u4e0e\u5f00\u6e90 Agent Claude Code\u3002\u7ecf\u8fc7\u8bed\u4e49\u53bb\u91cd\u540e\uff0c\u751f\u6210\u7684\u8bc4\u8bba\u4e0e\u589e\u5f3a\u540e\u7684\u4eba\u5de5\u8bc4\u5ba1\u5408\u5e76\uff0c\u5f62\u6210\u540e\u7eed\u4eba\u5de5\u6838\u9a8c\u7684\u5019\u9009\u96c6\u3002\u8be5\u96c6\u5408\u7531 80 \u540d\u9ad8\u7ea7\u8f6f\u4ef6\u5de5\u7a0b\u5e08\u8fdb\u884c\u4e25\u683c\u4eba\u5de5\u6807\u6ce8\uff0c\u6bcf\u4eba\u5177\u5907\u4e24\u5e74\u4ee5\u4e0a\u4e13\u4e1a\u7ecf\u9a8c\u3002\u6bcf\u6761\u8bc4\u8bba\u81f3\u5c11\u7531\u4e24\u540d\u6807\u6ce8\u8005\u72ec\u7acb\u6807\u6ce8\uff0c\u4efb\u4f55\u5206\u6b67\u7531\u516d\u4eba\u6838\u5fc3\u56e2\u961f\u8ba8\u8bba\u89e3\u51b3\u3002\u6807\u6ce8\u8005\u6838\u9a8c\u8bc4\u8bba\u6b63\u786e\u6027\u5e76\u5bf9\u95ee\u9898\u7c7b\u578b\u8fdb\u884c\u5206\u7c7b (Sun et al. 2025)\u3002\u4e0e\u73b0\u6709\u57fa\u51c6\u4e0d\u540c\uff0c\u6211\u4eec\u521b\u65b0\u6027\u5730\u6807\u6ce8\u4e86\u5f62\u6210\u6bcf\u6761\u8bc4\u8bba\u6240\u9700\u7684\u4e0a\u4e0b\u6587\u5c42\u7ea7\uff0c\u4ece\u800c\u80fd\u591f\u8bc4\u4f30\u4e0d\u540c\u4e0a\u4e0b\u6587\u4f9d\u8d56\u4e0b\u7684\u68c0\u6d4b\u96be\u5ea6\u3002\u8be5\u6b65\u9aa4\u5171\u5f97\u5230 1,505 \u6761\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u5176\u4e2d 391 \u6761\u7531\u539f\u59cb\u8bc4\u5ba1\u589e\u5f3a\u800c\u6765\uff0c1,114 \u6761\u7531 LLM \u4e0e\u4eba\u7c7b\u4e13\u5bb6\u589e\u5f3a\u800c\u6765\uff0c\u95ee\u9898\u8986\u76d6\u7387\u63d0\u5347\u4e86 285%\u3002\u56fe 2 \u5c55\u793a\u4e86\u8bed\u8a00\u5206\u5e03\u4ee5\u53ca\u589e\u5f3a\u7684\u4eba\u5de5\u8bc4\u8bba\u4e0e\u6a21\u578b\u751f\u6210\u8bc4\u8bba\u7684\u6bd4\u4f8b\u3002<\/p>\n<h3>3.3 \u4e0e\u73b0\u6709\u6570\u636e\u96c6\u7684\u6bd4\u8f83<\/h3>\n<p><strong>\u8868 2\uff1a<\/strong> \u73b0\u6709\u57fa\u51c6\u4e0e\u672c\u6587\u57fa\u51c6\u7684\u6bd4\u8f83<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6570\u636e\u96c6<\/th>\n<th>\u591a\u8bed\u8a00\u652f\u6301<\/th>\n<th>\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u652f\u6301<\/th>\n<th>\u8bc4\u5ba1\u8bc4\u8bba\u6765\u6e90<\/th>\n<th>\u4e0a\u4e0b\u6587\u8303\u56f4\u6807\u6ce8<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>CodeFuse-CR-Bench<\/td>\n<td>\u5426<\/td>\n<td>\u662f<\/td>\n<td>\u539f\u59cb\u6570\u636e<\/td>\n<td>\u5426<\/td>\n<\/tr>\n<tr>\n<td>SWR-Bench<\/td>\n<td>\u5426<\/td>\n<td>\u662f<\/td>\n<td>\u539f\u59cb\u6570\u636e<\/td>\n<td>\u5426<\/td>\n<\/tr>\n<tr>\n<td>ContextCRBench<\/td>\n<td>\u662f<\/td>\n<td>\u5426<\/td>\n<td>\u539f\u59cb\u6570\u636e<\/td>\n<td>\u5426<\/td>\n<\/tr>\n<tr>\n<td>AACR-Bench\uff08\u672c\u6587\uff09<\/td>\n<td>\u662f<\/td>\n<td>\u662f<\/td>\n<td>LLM \u4e0e\u4e13\u5bb6\u589e\u5f3a<\/td>\n<td>\u662f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u73b0\u6709 ACR \u57fa\u51c6\u4e0e\u672c\u6587\u57fa\u51c6\u7684\u7b80\u8981\u6bd4\u8f83\u89c1\u8868 2\u3002\u6211\u4eec\u7684\u57fa\u51c6\u4f5c\u4e3a\u9996\u4e2a\u5728\u591a\u8bed\u8a00\u73af\u5883\u4e2d\u63d0\u4f9b\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u7684\u57fa\u51c6\u800c\u4e0e\u4f17\u4e0d\u540c\uff0c\u7efc\u5408\u4e86\u73b0\u6709\u5de5\u4f5c\u7684\u4f18\u52bf\u3002\u5c3d\u7ba1 CodeFuse-CR-Bench (Guo et al. 2025a) \u4e0e SWR-Bench (Zeng et al. 2025) \u63d0\u4f9b\u4ed3\u5e93\u4e0a\u4e0b\u6587\uff0c\u4f46\u5b83\u4eec\u4ec5\u9650\u4e8e Python\uff1b\u76f8\u53cd\uff0cContextCRBench (Hu et al. 2025) \u4e0e CodeReview (Li et al. 2022) \u7b49\u591a\u8bed\u8a00\u6570\u636e\u96c6\u5219\u5c40\u9650\u4e8e\u6587\u4ef6\u7ea7\u3001\u51fd\u6570\u7ea7\u751a\u81f3 diff \u7ea7\u4e0a\u4e0b\u6587\u3002\u6211\u4eec\u901a\u8fc7\u5728 10 \u79cd\u6d41\u884c\u7f16\u7a0b\u8bed\u8a00\u4e0a\u8bc4\u6d4b\u5de5\u5177\u3001\u5e76\u63d0\u4f9b\u5305\u62ec PR \u5143\u6570\u636e\u4e0e\u5b8c\u6574\u4ee3\u7801\u4ed3\u5e93\u5728\u5185\u7684\u5168\u9762\u4e0a\u4e0b\u6587\u6765\u5f25\u5408\u8fd9\u4e00\u5dee\u8ddd\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u5e94\u5bf9\u5148\u524d\u57fa\u51c6\u6240\u4f7f\u7528\u7684\u539f\u59cb GitHub \u6570\u636e\u4e2d\u56fa\u6709\u7684\u566a\u58f0\u4e0e\u4e0d\u5b8c\u6574\u6027 (Guo et al. 2025a; Zeng et al. 2025; Hu et al. 2025; Li et al. 2022)\u3002\u6211\u4eec\u4e0d\u4f9d\u8d56\u539f\u59cb\u62bd\u53d6\uff0c\u800c\u662f\u91c7\u7528\u4e25\u683c\u7684\u6784\u5efa\u65b9\u6cd5\uff1a\u4f7f\u7528 LLM \u4ece\u4eba\u5de5\u8f93\u5165\u4e2d\u589e\u5f3a\u6280\u672f\u95ee\u9898\uff0c\u5e76\u7528\u516d\u4e2a\u5148\u8fdb\u6a21\u578b\uff08\u5305\u62ec Qwen3-Coder-480B\u3001GPT-5.2 \u4e0e Gemini-3-Pro\uff09\u6269\u5c55\u7f3a\u9677\u8986\u76d6\u3002\u4e3a\u786e\u4fdd\u53ef\u9760\u6027\uff0c\u6240\u6709\u8bc4\u8bba\u2014\u2014\u65e0\u8bba\u662f\u589e\u5f3a\u8fd8\u662f\u751f\u6210\u2014\u2014\u5747\u7531\u4e00\u4e2a\u7531 80 \u540d\u4e13\u4e1a\u8f6f\u4ef6\u5de5\u7a0b\u5e08\u7ec4\u6210\u7684\u5927\u89c4\u6a21\u56e2\u961f\u4ed4\u7ec6\u6838\u9a8c\u3002<\/p>\n<p>\u6700\u540e\uff0c\u9664\u6807\u51c6 PR \u4e0e\u8bc4\u8bba\u6807\u7b7e\u5916 (Guo et al. 2025a; Zeng et al. 2025; Hu et al. 2025)\uff0c\u6211\u4eec\u521b\u65b0\u6027\u5730\u6807\u6ce8\u6bcf\u6761\u8bc4\u5ba1\u8bc4\u8bba\u6240\u9700\u7684\u4e0a\u4e0b\u6587\u8303\u56f4\u3002\u8fd9\u4e00\u65b0\u7ef4\u5ea6\u4f7f\u5f97\u80fd\u591f\u7ec6\u7c92\u5ea6\u5206\u6790 ACR \u65b9\u6cd5\u5728\u4e0d\u540c\u4e0a\u4e0b\u6587\u6df1\u5ea6\u4e0e\u8303\u56f4\u4f9d\u8d56\u4e0b\u68c0\u6d4b\u95ee\u9898\u7684\u80fd\u529b\u3002<\/p>\n<h2>4 \u5b9e\u9a8c<\/h2>\n<p>\u6211\u4eec\u7684\u5b9e\u9a8c\u65e8\u5728\u9a8c\u8bc1\u6240\u63d0\u51fa\u6570\u636e\u96c6\uff08AACR-Bench\uff09\u6838\u5fc3\u65b0\u7279\u5f81\u6240\u5e26\u6765\u7684\u5f71\u54cd\uff0c\u5373\u591a\u8bed\u8a00\u652f\u6301\u3001\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\uff0c\u4ee5\u53ca\u66f4\u5168\u9762\u7684\u7f3a\u9677\u66b4\u9732\u6c34\u5e73\u3002<\/p>\n<h3>4.1 \u8bc4\u6d4b\u8bbe\u7f6e<\/h3>\n<p><strong>\u6a21\u578b\u3002<\/strong> \u6211\u4eec\u9009\u53d6\u4e86\u4e3b\u6d41\u5f00\u6e90\u63d0\u4f9b\u5546\u7684\u6700\u65b0\u6a21\u578b\u4ee5\u53ca\u4e3b\u8981\u5546\u4e1a\u5927\u6a21\u578b\u7684\u6700\u65b0\u7248\u672c\u3002\u56e0\u6b64\uff0c\u5b9e\u9a8c\u5305\u542b\u4e09\u4e2a\u5f00\u6e90\u6a21\u578b\uff08Qwen3-Coder-480B-A35B-Instruct (Team 2025)\u3001DeepSeek-V3.2 (DeepSeek-AI 2025)\u3001GLM-4.7 (Team et al. 2025)\uff09\u4ee5\u53ca\u4e24\u4e2a\u5546\u4e1a\u6a21\u578b\uff08GPT-5.2\u3001Claude-4.5-Sonnet\uff09\u3002\u6211\u4eec\u4f7f\u7528\u5b8c\u6574\u6570\u636e\u96c6\u8bc4\u6d4b\u4e86\u6240\u6709\u9009\u5b9a\u6a21\u578b\u3002\u4e3a\u7b80\u4fbf\u8d77\u89c1\uff0c\u540e\u6587\u5c06 Qwen3-Coder-480B-A35B-Instruct \u79f0\u4e3a Qwen-480B-Coder\u3002<\/p>\n<p><strong>\u4e0a\u4e0b\u6587\u68c0\u7d22\u65b9\u6cd5\u3002<\/strong> \u6211\u4eec\u6570\u636e\u96c6\u7684\u4e00\u4e2a\u663e\u8457\u7279\u5f81\u662f\u63d0\u4f9b\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u4f9d\u8d56\u3002\u7136\u800c\uff0c\u4e0d\u540c\u68c0\u7d22\u65b9\u6cd5\u4e5f\u4f1a\u663e\u8457\u5f71\u54cd\u6a21\u578b\u7684\u63a8\u7406\u6027\u80fd (Zhang et al.; Liu et al. 2023b)\u3002\u672c\u5b9e\u9a8c\u91c7\u7528\u4ee5\u4e0b\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li><strong>\u65e0\u4e0a\u4e0b\u6587\uff08No context\uff09\uff1a<\/strong> \u4f5c\u4e3a\u5bf9\u7167\u57fa\u7ebf\u3002<\/li>\n<li><strong>BM25\uff1a<\/strong> \u7ecf\u5178\u6587\u672c\u76f8\u4f3c\u5ea6\u65b9\u6cd5 (Robertson et al. 2009)\uff0c\u4e5f\u66fe\u7528\u4e8e\u5148\u524d\u7c7b\u4f3c\u7814\u7a76 (Guo et al. 2025a)\u3002<\/li>\n<li><strong>Embedding\uff1a<\/strong> \u91c7\u7528\u5f53\u524d\u6700\u5148\u8fdb\uff08SOTA\uff09\u6a21\u578b\u4e4b\u4e00 Qwen3-Embedding-8B \u8fdb\u884c\u5411\u91cf\u76f8\u4f3c\u5ea6\u68c0\u7d22\u3002<\/li>\n<li><strong>\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\uff1a<\/strong> \u6211\u4eec\u9009\u62e9\u4e86\u5e7f\u6cdb\u4f7f\u7528\u3001\u652f\u6301\u4ee3\u7801\u8bc4\u5ba1\u7684 Agent \u6846\u67b6 Claude Code\u3002<\/li>\n<\/ul>\n<p>\u6b64\u5916\u9700\u8981\u6307\u51fa\uff0c\u5bf9\u4e8e\u57fa\u4e8e\u76f8\u4f3c\u5ea6\u7684\u68c0\u7d22\u65b9\u6cd5\uff0c\u68c0\u7d22\u5230\u7684\u4ee3\u7801\u4e0a\u4e0b\u6587\u6570\u91cf\u7edf\u4e00\u8bbe\u4e3a 3\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cAgent \u65b9\u6cd5\u5141\u8bb8 Claude Code \u6846\u67b6\u81ea\u4e3b\u51b3\u5b9a\u68c0\u7d22\u7684\u4e0a\u4e0b\u6587\u6570\u91cf\u3002\u5e76\u4e14\u5728\u6240\u6709\u60c5\u51b5\u4e0b\uff08\u5373\u4f7f\u662f\u65e0\u4e0a\u4e0b\u6587\u65b9\u6cd5\uff09\uff0c\u90fd\u63d0\u4f9b\u5b8c\u5168\u76f8\u540c\u7684 PR \u6807\u9898\u4e0e\u63cf\u8ff0\u3002<\/p>\n<h5>\u6307\u6807<\/h5>\n<p>\u6211\u4eec\u4f7f\u7528\u5e38\u89c1\u7684\u4e09\u9879\u6307\u6807\u8bc4\u6d4b\u5404\u7c7b ACR \u65b9\u6cd5\u5728 AACR-Bench \u4e0a\u7684\u8868\u73b0\uff0c\u5373 Precision\u3001Recall \u4e0e F1-score\u3002<\/p>\n<p><strong>\u8868 3\uff1a<\/strong> \u4e0d\u540c ACR \u65b9\u6cd5\u4e0b\u5404\u6a21\u578b\u7684\u6027\u80fd\u6bd4\u8f83\uff08%\uff09<\/p>\n<table>\n<thead>\n<tr>\n<th>\u65b9\u6cd5<\/th>\n<th>\u6a21\u578b<\/th>\n<th style=\"text-align: right\">\u5e73\u5747\u8bc4\u8bba\u6570<\/th>\n<th style=\"text-align: right\">Recall (%)<\/th>\n<th style=\"text-align: right\">Precision (%)<\/th>\n<th style=\"text-align: right\">F1 (%)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Agent<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">0.0890<\/td>\n<td style=\"text-align: right\">10.10<\/td>\n<td style=\"text-align: right\">39.90<\/td>\n<td style=\"text-align: right\">16.12<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">0.1525<\/td>\n<td style=\"text-align: right\">4.78<\/td>\n<td style=\"text-align: right\">11.00<\/td>\n<td style=\"text-align: right\">6.67<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">0.1448<\/td>\n<td style=\"text-align: right\">4.72<\/td>\n<td style=\"text-align: right\">11.50<\/td>\n<td style=\"text-align: right\">6.69<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">0.1063<\/td>\n<td style=\"text-align: right\">2.99<\/td>\n<td style=\"text-align: right\">9.90<\/td>\n<td style=\"text-align: right\">4.59<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">0.1004<\/td>\n<td style=\"text-align: right\">4.39<\/td>\n<td style=\"text-align: right\">15.30<\/td>\n<td style=\"text-align: right\">6.82<\/td>\n<\/tr>\n<tr>\n<td>Embedding<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">1.8902<\/td>\n<td style=\"text-align: right\">42.86<\/td>\n<td style=\"text-align: right\">8.00<\/td>\n<td style=\"text-align: right\">13.48<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">2.5230<\/td>\n<td style=\"text-align: right\">36.35<\/td>\n<td style=\"text-align: right\">5.10<\/td>\n<td style=\"text-align: right\">8.94<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">0.8657<\/td>\n<td style=\"text-align: right\">26.98<\/td>\n<td style=\"text-align: right\">11.00<\/td>\n<td style=\"text-align: right\">15.63<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">2.4709<\/td>\n<td style=\"text-align: right\">47.24<\/td>\n<td style=\"text-align: right\">6.70<\/td>\n<td style=\"text-align: right\">11.74<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">0.8335<\/td>\n<td style=\"text-align: right\">24.25<\/td>\n<td style=\"text-align: right\">10.20<\/td>\n<td style=\"text-align: right\">14.36<\/td>\n<\/tr>\n<tr>\n<td>No context<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">1.7220<\/td>\n<td style=\"text-align: right\">42.86<\/td>\n<td style=\"text-align: right\">8.70<\/td>\n<td style=\"text-align: right\">14.46<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">2.2875<\/td>\n<td style=\"text-align: right\">36.54<\/td>\n<td style=\"text-align: right\">5.60<\/td>\n<td style=\"text-align: right\">9.71<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">0.8573<\/td>\n<td style=\"text-align: right\">27.57<\/td>\n<td style=\"text-align: right\">11.30<\/td>\n<td style=\"text-align: right\">16.03<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">2.3537<\/td>\n<td style=\"text-align: right\">47.11<\/td>\n<td style=\"text-align: right\">7.00<\/td>\n<td style=\"text-align: right\">12.19<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">1.0217<\/td>\n<td style=\"text-align: right\">27.44<\/td>\n<td style=\"text-align: right\">9.40<\/td>\n<td style=\"text-align: right\">14.00<\/td>\n<\/tr>\n<tr>\n<td>BM25<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">2.1698<\/td>\n<td style=\"text-align: right\">35.75<\/td>\n<td style=\"text-align: right\">5.80<\/td>\n<td style=\"text-align: right\">9.98<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">0.8851<\/td>\n<td style=\"text-align: right\">27.38<\/td>\n<td style=\"text-align: right\">10.90<\/td>\n<td style=\"text-align: right\">15.59<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">0.9070<\/td>\n<td style=\"text-align: right\">26.25<\/td>\n<td style=\"text-align: right\">10.20<\/td>\n<td style=\"text-align: right\">14.69<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">1.7461<\/td>\n<td style=\"text-align: right\">43.59<\/td>\n<td style=\"text-align: right\">8.80<\/td>\n<td style=\"text-align: right\">14.64<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">2.3906<\/td>\n<td style=\"text-align: right\">45.85<\/td>\n<td style=\"text-align: right\">6.70<\/td>\n<td style=\"text-align: right\">11.69<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>4.2 \u4e3b\u8981\u7ed3\u679c<\/h3>\n<p>\u672c\u8282\u7ed9\u51fa\u591a\u6837\u5316 ACR \u65b9\u6cd5\u4e0e\u6a21\u578b\u5728\u6211\u4eec\u57fa\u51c6\u4e0a\u7684\u5168\u9762\u8bc4\u6d4b\u3002\u6211\u4eec\u7279\u522b\u5206\u6790\u4e0a\u4e0b\u6587\u68c0\u7d22\u3001\u6a21\u578b\u67b6\u6784\u4e0e Agent \u6846\u67b6\u5bf9\u6240\u751f\u6210\u8bc4\u5ba1\u603b\u4f53\u8d28\u91cf\u7684\u5f71\u54cd\u3002\u9700\u8981\u6307\u51fa\u7684\u662f\uff0c\u5f53\u524d ACR \u7684\u5b9e\u9645\u5e94\u7528 (Sadowski et al. 2015; Distefano et al. 2019) \u4ee5\u53ca\u5148\u524d\u7814\u7a76 (Tao et al. 2012) \u5747\u4e00\u81f4\u5f3a\u8c03\u4e0a\u4e0b\u6587\u4f9d\u8d56\u4fe1\u606f\u5728 ACR \u4e2d\u7684\u5173\u952e\u4f5c\u7528\u3002\u56e0\u6b64\uff0c\u672c\u6587\u6240\u6709\u5b9e\u9a8c\u5747\u57fa\u4e8e\u8fd9\u4e00\u5047\u8bbe\uff1a\u8fd9\u4ee3\u8868 ACR \u6240\u786e\u7acb\u7684\u771f\u6b63\u8303\u5f0f\u3002\u7b80\u8a00\u4e4b\uff0c\u9664\u975e\u4f5c\u4e3a\u5bf9\u7167\u57fa\u7ebf\uff0c\u6211\u4eec\u4e0d\u8bc4\u6d4b LLM \u5728\u65e0\u4e0a\u4e0b\u6587\u4fe1\u606f\u65f6\u7684 ACR \u4efb\u52a1\u8868\u73b0\u3002<\/p>\n<h4>4.2.1 \u5728\u66f4\u5145\u5206\u7684\u7f3a\u9677\u66b4\u9732\u4e0b\u8bc4\u6d4b ACR \u6027\u80fd<\/h4>\n<p>\u5982\u8868 3 \u7684\u5b9e\u9a8c\u6570\u636e\u6240\u793a\uff0c\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\u5448\u73b0\u51fa\u4e0e\u975e Agent \u57fa\u7ebf\u660e\u663e\u4e0d\u540c\u7684\u8bc4\u5ba1\u8868\u73b0\u3002\u540c\u65f6\uff0c\u5f15\u5165\u4e0a\u4e0b\u6587\u4fe1\u606f\u5e76\u4e0d\u603b\u80fd\u5728 ACR \u4efb\u52a1\u6027\u80fd\u4e0a\u5e26\u6765\u6b63\u5411\u589e\u76ca\u3002<\/p>\n<h5>\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\u4e0e\u4f20\u7edf\u65b9\u6cd5\u7684\u6bd4\u8f83<\/h5>\n<p>\u4e0e\u4f20\u7edf\u65b9\u6cd5\u76f8\u6bd4\uff0c\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\uff08\u4f8b\u5982 Claude Code\uff09\u6bcf\u4e2a patch \u6301\u7eed\u751f\u6210\u663e\u8457\u66f4\u5c11\u7684\u8bc4\u5ba1\u8bc4\u8bba\uff08$0.08sim 0.15$\uff09\u3002\u4f8b\u5982\uff0cClaude-4.5-Sonnet \u5728 Agent \u6a21\u5f0f\u4e0b\u8fbe\u5230 $39.90%$ \u7684 Precision\uff0c\u8fdc\u8d85\u5176\u5728\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u6a21\u5f0f\u4e0b\u7684 $8.70%$\uff0c\u4f46\u4f34\u968f\u660e\u663e\u66f4\u4f4e\u7684 Recall\uff08$10.10%$\uff09\u3002\u8fd9\u8868\u660e\uff0c\u5c3d\u7ba1 Agent \u5728\u5bf9\u7cbe\u5ea6\u8981\u6c42\u82db\u523b\u7684\u573a\u666f\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u5b83\u4eec\u53ef\u80fd\u5ffd\u7565\u8bb8\u591a\u6f5c\u5728\u7f3a\u9677\u2014\u2014\u8fd9\u4e00\u503e\u5411\u53ef\u5f52\u56e0\u4e8e\u5176\u805a\u7126\u68c0\u7d22\u673a\u5236\u6240\u8bf1\u53d1\u7684\u300c\u4e0a\u4e0b\u6587\u96a7\u9053\u89c6\u91ce\u300d\uff08contextual tunnel vision\uff09\u6548\u5e94 (Liu et al. 2023a)\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u4f20\u7edf\u65b9\u6cd5\u5f80\u5f80\u4ea7\u751f\u8fc7\u91cf\u8bc4\u8bba\uff08\u4f8b\u5982 GPT-5.2 \u5e73\u5747\u6bcf\u4e2a patch $2.47$ \u6761\uff09\uff0c\u53ef\u80fd\u5728\u566a\u58f0\u4e2d\u63a9\u76d6\u6709\u610f\u4e49\u7684\u6d1e\u89c1\u3002\u6b64\u5916\uff0c\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\u7684\u6709\u6548\u6027\u8868\u73b0\u51fa\u5f3a\u70c8\u7684\u6a21\u578b\u7279\u5f02\u6027\u4f9d\u8d56\uff0c\u8fd9\u4e00\u654f\u611f\u6027\u5728\u5176\u4ed6\u68c0\u7d22\u7b56\u7565\u4e2d\u5e76\u672a\u89c2\u5bdf\u5230\u3002\u5c3d\u7ba1 Claude-4.5-Sonnet \u5728 Agent \u6a21\u5f0f\u4e0b\u53d6\u5f97\u51fa\u8272\u7cbe\u5ea6\uff0c\u5176\u4ed6\u6a21\u578b\u5e76\u672a\u5c55\u73b0\u7c7b\u4f3c\u589e\u76ca\uff0c\u751a\u81f3\u53ef\u80fd\u51fa\u73b0\u6027\u80fd\u4e0b\u964d\u3002\u4f8b\u5982\uff0cGPT-5.2 \u5c3d\u7ba1\u5728\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u6a21\u5f0f\u4e0b\u57fa\u7ebf\u8868\u73b0\u5f3a\u52b2\uff0c\u4f46\u5728 Agent \u6a21\u5f0f\u4e0b\u6025\u5267\u4e0b\u964d\uff0cPrecision \u964d\u81f3 $9.90%$\uff0cRecall \u964d\u81f3 $2.99%$\u3002\u8fd9\u4e00\u5bf9\u6bd4\u8868\u660e\uff0c\u901a\u7528\u6a21\u578b\u80fd\u529b\u5e76\u4e0d\u80fd\u76f4\u63a5\u8f6c\u5316\u4e3a\u57fa\u4e8e Agent \u7684 ACR \u4efb\u52a1\u4e0a\u7684\u719f\u7ec3\u5ea6 (Liu et al. 2023c)\u3002<\/p>\n<h5>\u4e0a\u4e0b\u6587\u68c0\u7d22\u65b9\u6cd5\u7684\u5f71\u54cd<\/h5>\n<p>\u8868 3 \u8fdb\u4e00\u6b65\u8868\u660e\uff0c\u4e0a\u4e0b\u6587\u68c0\u7d22\u5e76\u975e\u666e\u904d\u6709\u76ca\u3002\u5bf9\u4e8e\u56fa\u6709\u63a8\u7406\u80fd\u529b\u5f3a\u7684\u6a21\u578b\uff0c\u6734\u7d20 RAG \u53ef\u80fd\u5f15\u5165\u6709\u5bb3\u566a\u58f0\uff0c\u800c\u5404\u7c7b\u5f00\u6e90\u6a21\u578b\u5bf9\u7279\u5b9a\u68c0\u7d22\u7b56\u7565\u8868\u73b0\u51fa\u4e0d\u540c\u504f\u597d\u3002\u4f8b\u5982\uff0cClaude-4.5-Sonnet \u5728\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u6a21\u5f0f\u4e0b\u8868\u73b0\u7a33\u5065\uff08F1=$14.46$\uff09\u3002\u7136\u800c\uff0c\u901a\u8fc7 BM25 \u63d0\u4f9b top-3 \u4e0a\u4e0b\u6587\u4f1a\u5bfc\u81f4\u6027\u80fd\u6025\u5267\u4e0b\u964d\uff08F1=$9.98$\uff0c\u4e0b\u964d $31%$\uff1bPrecision \u4ece $8.70%$ \u964d\u81f3 $5.80%$\uff09\u3002\u57fa\u4e8e Embedding \u7684\u68c0\u7d22\u540c\u6837\u4f1a\u4f7f\u5176\u7ed3\u679c\u53d8\u5dee\u3002\u4e0d\u540c\u6a21\u578b\u504f\u597d\u4e0d\u540c\u7b56\u7565\u3002DeepSeek-V3.2 \u5728 BM25 \u4e0b\u53d6\u5f97\u6700\u4f73\u8868\u73b0\uff1a\u6bcf\u4e2a patch \u7684\u5e73\u5747\u8bc4\u8bba\u6570\u4ece\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u7684 2.29 \u964d\u81f3 $0.89$\uff0c\u540c\u65f6 Precision \u4ece $5.10%$ \u8dc3\u5347\u81f3 $10.90%$\uff0cF1 \u8fbe\u5230 $15.59$\u3002\u76f8\u53cd\uff0cQwen-480B-Coder \u5728\u57fa\u4e8e Embedding \u7684\u68c0\u7d22\u4e0b\u8fbe\u5230\u5cf0\u503c\uff08F1=$14.36$\uff09\uff0c\u663e\u8457\u4f18\u4e8e\u5176 BM25 \u5206\u6570\uff08$11.69$\uff09\u3002\u56e0\u6b64\uff0c\u6ca1\u6709\u4e00\u79cd\u68c0\u7d22\u65b9\u6cd5\u5bf9\u6240\u6709\u6a21\u578b\u90fd\u6700\u4f18\uff1b\u6027\u80fd\u9ad8\u5ea6\u4f9d\u8d56\u4e8e\u6a21\u578b\u4e0e\u68c0\u7d22\u6a21\u5f0f\u7684\u7279\u5b9a\u7ec4\u5408\u3002<\/p>\n<h5>\u5173\u952e\u89c2\u5bdf<\/h5>\n<p>\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\u6301\u7eed\u4ea7\u751f\u8fdc\u5c11\u4e8e\u5176\u4ed6\u65b9\u6cd5\u7684\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u4e14\u5176\u6709\u6548\u6027\u5f3a\u70c8\u4f9d\u8d56\u4e8e\u6a21\u578b\u3002\u8be5\u8303\u5f0f\u4e0b\u4e0d\u540c LLM \u7684\u8868\u73b0\u5dee\u5f02\u663e\u8457\u3002\u6b64\u5916\uff0c\u4e0d\u540c\u6a21\u578b\u5bf9\u4e0a\u4e0b\u6587\u68c0\u7d22\u65b9\u6cd5\u7684\u54cd\u5e94\u4e0d\u540c\uff0c\u4f8b\u5982\u5c06 Claude-4.5-Sonnet \u4e0e Agent \u6846\u67b6\u914d\u5bf9\u3001DeepSeek \u4e0e BM25 \u914d\u5bf9\u3001Qwen \u4e0e Embedding \u914d\u5bf9\uff0c\u6301\u7eed\u5448\u73b0\u6700\u4f18\u8868\u73b0\u3002<\/p>\n<p><strong>\u8868 4\uff1a<\/strong> \u4e0d\u540c ACR \u65b9\u6cd5\u5728\u4e0d\u540c\u4e0a\u4e0b\u6587\u8303\u56f4\u4e0b\u53d1\u73b0\u95ee\u9898\u7684\u8868\u73b0\uff08\u4ec5 Recall\uff09<\/p>\n<table>\n<thead>\n<tr>\n<th>\u65b9\u6cd5<\/th>\n<th>\u6a21\u578b<\/th>\n<th style=\"text-align: right\">Diff<\/th>\n<th style=\"text-align: right\">File<\/th>\n<th style=\"text-align: right\">Repo<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Agent<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">11.95<\/td>\n<td style=\"text-align: right\">16.84<\/td>\n<td style=\"text-align: right\">13.77<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">4.28<\/td>\n<td style=\"text-align: right\">4.64<\/td>\n<td style=\"text-align: right\">8.00<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">6.50<\/td>\n<td style=\"text-align: right\">5.95<\/td>\n<td style=\"text-align: right\">4.40<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">3.21<\/td>\n<td style=\"text-align: right\">2.52<\/td>\n<td style=\"text-align: right\">3.45<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">4.49<\/td>\n<td style=\"text-align: right\">5.24<\/td>\n<td style=\"text-align: right\">5.94<\/td>\n<\/tr>\n<tr>\n<td>No context<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">45.76<\/td>\n<td style=\"text-align: right\">40.54<\/td>\n<td style=\"text-align: right\">38.63<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">39.26<\/td>\n<td style=\"text-align: right\">32.43<\/td>\n<td style=\"text-align: right\">36.91<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">31.43<\/td>\n<td style=\"text-align: right\">24.90<\/td>\n<td style=\"text-align: right\">21.03<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">50.66<\/td>\n<td style=\"text-align: right\">43.82<\/td>\n<td style=\"text-align: right\">42.92<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">33.82<\/td>\n<td style=\"text-align: right\">22.59<\/td>\n<td style=\"text-align: right\">17.60<\/td>\n<\/tr>\n<tr>\n<td>BM25<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">46.68<\/td>\n<td style=\"text-align: right\">41.31<\/td>\n<td style=\"text-align: right\">38.63<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">38.20<\/td>\n<td style=\"text-align: right\">32.82<\/td>\n<td style=\"text-align: right\">34.33<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">29.71<\/td>\n<td style=\"text-align: right\">27.61<\/td>\n<td style=\"text-align: right\">19.31<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">48.94<\/td>\n<td style=\"text-align: right\">43.82<\/td>\n<td style=\"text-align: right\">40.34<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">29.97<\/td>\n<td style=\"text-align: right\">23.94<\/td>\n<td style=\"text-align: right\">19.31<\/td>\n<\/tr>\n<tr>\n<td>Embedding<\/td>\n<td>Claude-4.5-Sonnet<\/td>\n<td style=\"text-align: right\">46.82<\/td>\n<td style=\"text-align: right\">38.03<\/td>\n<td style=\"text-align: right\">40.77<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Deepseek-V3.2<\/td>\n<td style=\"text-align: right\">39.39<\/td>\n<td style=\"text-align: right\">33.40<\/td>\n<td style=\"text-align: right\">33.05<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GLM-4.7<\/td>\n<td style=\"text-align: right\">30.11<\/td>\n<td style=\"text-align: right\">26.64<\/td>\n<td style=\"text-align: right\">17.60<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>GPT-5.2<\/td>\n<td style=\"text-align: right\">52.39<\/td>\n<td style=\"text-align: right\">42.28<\/td>\n<td style=\"text-align: right\">41.63<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>Qwen-480B-Coder<\/td>\n<td style=\"text-align: right\">27.06<\/td>\n<td style=\"text-align: right\">22.78<\/td>\n<td style=\"text-align: right\">18.45<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>4.2.2 \u4e0a\u4e0b\u6587\u5c42\u7ea7\u5bf9 ACR \u6027\u80fd\u7684\u5f71\u54cd<\/h4>\n<p>\u6211\u4eec\u8bc4\u4f30\u5404\u7c7b\u65b9\u6cd5\u5728\u8bc6\u522b\u9700\u8981\u4e0d\u540c\u4e0a\u4e0b\u6587\u5c42\u7ea7\uff08\u5b9a\u4e49\u89c1\u8868 1\uff09\u7684\u95ee\u9898\u4e0a\u7684\u6548\u529b\u3002\u7531\u4e8e\u65e0\u6cd5\u786e\u77e5\u6bcf\u6b21\u63a8\u7406\u4e2d\u9690\u5f0f\u4f7f\u7528\u7684\u7cbe\u786e\u4e0a\u4e0b\u6587\u5c42\u7ea7\uff0c\u6211\u4eec\u4ec5\u62a5\u544a Recall \u6307\u6807\u3002<\/p>\n<h5>\u6765\u81ea\u4e0a\u4e0b\u6587\u5c42\u7ea7\u7684\u5f71\u54cd<\/h5>\n<p>\u5982\u8868 4 \u6240\u793a\uff0c\u9664 Agent \u6846\u67b6\u8fd9\u4e00\u663e\u8457\u4f8b\u5916\uff0c\u6240\u6709\u4e0a\u4e0b\u6587\u68c0\u7d22\u65b9\u6cd5\uff08\u65e0\u4e0a\u4e0b\u6587\u3001BM25\u3001Embedding\uff09\u5728\u4e0d\u540c\u4e0a\u4e0b\u6587\u5c42\u7ea7\u4e0a\u90fd\u5448\u73b0\u51fa\u660e\u663e\u7684\u6027\u80fd\u8870\u51cf\u8d8b\u52bf\uff0c\u4e00\u81f4\u9075\u5faa $text{Diff} &gt; text{File} &gt; text{Repo}$ \u7684\u5c42\u7ea7\u3002\u4ee5\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u65b9\u6cd5\u4e3a\u4f8b\uff1aQwen-480B-Coder \u7684\u8868\u73b0\u4ece Diff \u7ea7\u7684 $33.82%$ \u4e0b\u964d\u5230 File \u7ea7\u7684 $22.59%$\uff0c\u5e76\u8fdb\u4e00\u6b65\u4e0b\u964d\u5230 Repo \u7ea7\u7684 $17.60%$\u3002\u7c7b\u4f3c\u5730\uff0cGPT-5.2 \u4e5f\u5448\u73b0\u5e73\u884c\u7684\u4e0b\u884c\u8f68\u8ff9\u3002\u5373\u4fbf\u5f15\u5165\u68c0\u7d22\u589e\u5f3a\uff08BM25 \u6216 Embedding\uff09\uff0c\u4e5f\u65e0\u6cd5\u9006\u8f6c\u8fd9\u79cd\u4e0e\u4e0a\u4e0b\u6587\u8303\u56f4\u6269\u5927\u76f8\u5173\u7684\u6027\u80fd\u8870\u51cf\u73b0\u8c61\u3002\u5f62\u6210\u9c9c\u660e\u5bf9\u6bd4\u7684\u662f\uff0cAgent \u6846\u67b6\uff08\u901a\u8fc7 Claude Code \u5b9e\u73b0\uff09\u603b\u4f53\u4e0a\u5448\u73b0\u76f8\u53cd\u8d8b\u52bf\uff0c\u5728\u590d\u6742 Repo \u7ea7\u573a\u666f\u4e2d\u5e38\u5e38\u4f18\u4e8e\u5b64\u7acb\u7684 diff \u573a\u666f\u3002\u4f8b\u5982\uff0c\u5728 Agent \u6846\u67b6\u4e0b\uff0cDeepSeek-V3.2 \u7684\u8868\u73b0\u4ece diff \u7ea7\u7684 $4.28%$ \u63d0\u5347\u5230 Repo \u7ea7\u7684 $8.00%$\uff1b\u7c7b\u4f3c\u5730\uff0cQwen-480B-Coder \u4ece $4.49%$ \u5347\u81f3 $5.94%$\u3002\u5c3d\u7ba1\u8fd9\u8868\u660e Agent \u53ef\u4ee5\u5229\u7528\u591a\u8f6e\u4ea4\u4e92\u6709\u6548\u68c0\u7d22\u590d\u6742\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u4f46 diff \u7ea7\u4e0a\u6781\u4f4e\u7684\u5206\u6570\uff08\u4f8b\u5982 DeepSeek \u7684 $4.28%$ \u5bf9\u6bd4\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u7684 $39.26%$\uff09\u53ef\u80fd\u610f\u5473\u7740 Agent \u4f1a\u8fc7\u5206\u5173\u6ce8\u5916\u90e8\u4f9d\u8d56\uff0c\u4ece\u800c\u5ffd\u7565 diff \u672c\u8eab\u56fa\u6709\u7684\u663e\u773c\u5c40\u90e8\u95ee\u9898\u3002<\/p>\n<h5>\u5173\u952e\u89c2\u5bdf<\/h5>\n<p>\u9664\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\u5916\uff0cACR \u65b9\u6cd5\u68c0\u6d4b\u95ee\u9898\u7684\u80fd\u529b\u968f\u6240\u9700\u4e0a\u4e0b\u6587\u5c42\u7ea7\/\u8303\u56f4\u589e\u52a0\u800c\u4e0b\u964d\uff0c\u800c Agent \u65b9\u6cd5\u5448\u73b0\u76f8\u53cd\u8d8b\u52bf\u3002<\/p>\n<h4>4.2.3 \u8bed\u8a00\u7ef4\u5ea6\u5bf9 ACR \u6027\u80fd\u7684\u5f71\u54cd<\/h4>\n<p>\u56fe 3 \u663e\u793a\uff0c\u7f16\u7a0b\u8bed\u8a00\u4e5f\u53ef\u80fd\u5f71\u54cd\u5404\u7c7b ACR \u65b9\u6cd5\u7684\u8868\u73b0\uff0c\u8fd9\u5728\u76f8\u5f53\u7a0b\u5ea6\u4e0a\u8bc1\u5b9e\u4e86\u5c06\u57fa\u51c6\u6269\u5c55\u5230\u591a\u79cd\u7f16\u7a0b\u8bed\u8a00\u7684\u91cd\u8981\u6027\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 3\uff1a\u5206\u8bed\u8a00\u7684\u4ee3\u7801\u8bc4\u5ba1\u8868\u73b0\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig03.png\" \/><\/p>\n<p><strong>\u56fe 3\uff1a<\/strong> \u5206\u8bed\u8a00\u7684\u4ee3\u7801\u8bc4\u5ba1\u8868\u73b0<\/p>\n<h5>\u6a21\u578b\u8868\u73b0\u4e2d\u7684\u8bed\u8a00\u7279\u5f02\u6027\u504f\u5dee<\/h5>\n<p>\u4e0d\u540c ACR \u65b9\u6cd5\u5728\u5404\u7f16\u7a0b\u8bed\u8a00\u4e0a\u7684\u6709\u6548\u6027\u5dee\u5f02\u76f8\u5f53\u5927\uff0c\u63ed\u793a\u51fa\u5f3a\u70c8\u7684\u8bed\u8a00\u7279\u5f02\u6027\u504f\u5dee\u3002\u4f8b\u5982\uff0cClaude-4.5-Sonnet\u2014\u2014Agent \u4efb\u52a1\u4e2d\u8868\u73b0\u6700\u597d\u7684\u6a21\u578b\u2014\u2014\u5448\u73b0\u51fa\u6e05\u6670\u7684\u6027\u80fd\u5c42\u7ea7\uff1a\u5b83\u5728 Python\uff080.247\uff09\u3001Java\uff080.241\uff09\u3001Go\uff080.218\uff09\u4e0e C\uff080.189\uff09\u4e0a\u53d6\u5f97\u660e\u663e\u66f4\u9ad8\u7684 F1 \u5206\u6570\uff0c\u6784\u6210\u4e00\u4e2a\u9c9c\u660e\u7684\u7b2c\u4e00\u68af\u961f\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5176\u5728 TypeScript\uff080.081\uff09\u3001PHP\uff080.082\uff09\u4e0e Rust\uff080.091\uff09\u4e0a\u7684\u8868\u73b0\u6025\u5267\u4e0b\u964d\uff0cPython \u4e0e TypeScript \u4e4b\u95f4\u7684\u5dee\u8ddd\u53ef\u8fbe 3 \u500d\u3002\u6211\u4eec\u5c06\u8fd9\u4e00\u5dee\u5f02\u4e3b\u8981\u5f52\u56e0\u4e8e\u6a21\u578b\u8bad\u7ec3\u6570\u636e\u7684\u4e0d\u5747\u8861\u5206\u5e03\u3002Python \u4e0e Java \u7b49\u8bed\u8a00\u62e5\u6709\u5e7f\u6cdb\u7684\u751f\u6001\u7cfb\u7edf\u4e0e\u5927\u91cf\u9ad8\u8d28\u91cf\u5f00\u6e90\u4ed3\u5e93\uff0c\u5f88\u53ef\u80fd\u63d0\u4f9b\u4e86\u5145\u8db3\u4e14\u7ecf\u8fc7\u826f\u597d\u6574\u7406\u7684\u8bad\u7ec3\u8bed\u6599\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5bf9 Rust \u8fd9\u7c7b\u8bed\u8a00\uff0c\u76f8\u5bf9\u7a00\u7f3a\u7684\u8bed\u6599\u5f80\u5f80\u963b\u788d LLM\/Agent \u83b7\u5f97\u6267\u884c\u51c6\u786e\u89c4\u5212\u6b65\u9aa4\u6240\u9700\u7684\u77e5\u8bc6\u3002<\/p>\n<p>\u540c\u65f6\uff0c\u5bf9\u6240\u6709\u4e3b\u6d41\u6846\u67b6\uff08\u65e0\u4e0a\u4e0b\u6587\u3001BM25\u3001Embedding\u3001Agent\uff09\u7684\u4ea4\u53c9\u6bd4\u8f83\u4e5f\u663e\u793a\uff0cC# \u8868\u73b0\u51fa\u5f02\u5e38\u7684\u53ef\u89e3\u91ca\u6027\u4e0e\u6a21\u578b\u53cb\u597d\u6027\uff0c\u800c C \u8bed\u8a00\u5219\u6301\u7eed\u5904\u4e8e\u6027\u80fd\u8c31\u7cfb\u7684\u5e95\u90e8\u3002\u5728\u65e0\u4e0a\u4e0b\u6587\u7684\u300cNo context\u300d\u6a21\u5f0f\u4e0b\uff0cGPT-5.2 \u5728 C# \u4e0a\u53d6\u5f97\u5168\u9762\u6700\u9ad8\u7684 F1 \u5206\u6570\uff080.309\uff09\uff0c\u800c\u540c\u4e00\u6a21\u578b\u5728 C \u4e0a\u4ec5\u5f97 0.085\u3002\u8fd9\u4e00\u8d8b\u52bf\u8fdb\u4e00\u6b65\u88ab Qwen-480B-Coder \u8bc1\u5b9e\uff08C# $0.268$ vs. C $0.104$\uff09\u3002\u9274\u4e8e\u4e24\u79cd\u8bed\u8a00\u7684\u5e94\u7528\u73b0\u72b6\uff0c\u6211\u4eec\u53ef\u5408\u7406\u5047\u8bbe C# \u4e0e C \u4e3a\u8bad\u7ec3\u8bed\u6599\u63d0\u4f9b\u4e86\u53ef\u6bd4\u7684\u4e30\u5bcc\u7a0b\u5ea6\u3002\u5728\u6b64\u610f\u4e49\u4e0a\uff0c\u8fd9\u4e00\u73b0\u8c61\u6df1\u523b\u53cd\u6620\u4e86\u7f16\u7a0b\u8bed\u8a00\u5185\u5728\u7279\u5f81\u5bf9 LLM \u6027\u80fd\u7684\u5f71\u54cd\u3002\u4f8b\u5982\uff0c\u4f5c\u4e3a\u5f3a\u7c7b\u578b\u3001\u9762\u5411\u5bf9\u8c61\u7684\u8bed\u8a00\uff0cC# \u62e5\u6709\u4e25\u683c\u7684\u547d\u540d\u7a7a\u95f4\u7ba1\u7406\u4e0e\u663e\u5f0f\u7c7b\u578b\u5b9a\u4e49\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cC \u4e25\u91cd\u4f9d\u8d56\u6307\u9488\u64cd\u4f5c\u3001\u5b8f\u5b9a\u4e49\u4e0e\u9690\u5f0f\u5185\u5b58\u7ba1\u7406\u3002\u5176\u4f9d\u8d56\u4fe1\u606f\u5f80\u5f80\u9690\u542b\u5728\u975e\u7ed3\u6784\u5316\u5934\u6587\u4ef6\u6216\u94fe\u63a5\u903b\u8f91\u4e2d\u3002\u6211\u4eec\u5047\u8bbe\u8fd9\u7c7b\u7ed3\u6784\u7279\u5f81\u5bf9 LLM \u7684 ACR \u6027\u80fd\u5177\u6709\u5b9e\u8d28\u6027\u5f71\u54cd\u3002<\/p>\n<h5>\u4e0a\u4e0b\u6587\u8de8\u8bed\u8a00\u7684\u6027\u80fd\u5f71\u54cd<\/h5>\n<p>\u5c3d\u7ba1\u901a\u5e38\u9884\u671f\u4e0a\u4e0b\u6587\u4f1a\u6539\u5584\u6a21\u578b\u8868\u73b0\uff0c\u6211\u4eec\u7684\u5b9e\u9a8c\u63ed\u793a\u4e86\u4e00\u4e2a\u53cd\u76f4\u89c9\u8d8b\u52bf\uff1a\u5f15\u5165\u4e0a\u4e0b\u6587\u5728\u5927\u591a\u6570\u8bed\u8a00\u4e0a\u5bfc\u81f4\u663e\u8457\u7684\u300c\u4e0a\u4e0b\u6587\u9000\u6b65\u300d\uff08Contextual Backwardness\uff09\uff0c\u4ec5\u5c11\u6570\u8bed\u8a00\u8868\u73b0\u51fa\u7a33\u5065\u6027\u3002\u6bd4\u8f83\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u6a21\u5f0f\u4e0e\u300cAgent\/\u68c0\u7d22\u300d\u6a21\u5f0f\uff0c\u63ed\u793a\u51fa\u4e24\u79cd\u4e0d\u540c\u7684\u884c\u4e3a\u6a21\u5f0f\u3002\u5bf9 C#\u3001C++\u3001JavaScript\u3001PHP\u3001Python\u3001Rust \u4e0e TypeScript\uff0c\u5f15\u5165\u590d\u6742\u4e0a\u4e0b\u6587\u6846\u67b6\u53cd\u800c\u5f15\u5165\u4e86\u566a\u58f0\u3002\u4ee5 GPT-5.2 \u4e3a\u4f8b\uff0c\u5176\u5728 C# \u4e0a\u7684\u8868\u73b0\u4ece\u300c\u65e0\u4e0a\u4e0b\u6587\u300d\u6a21\u5f0f\u7684 0.309 \u9aa4\u964d\u81f3 Agent \u6a21\u5f0f\u7684 0.095\uff1b\u7c7b\u4f3c\u5730\uff0cPython \u4ece 0.165 \u964d\u81f3 0.071\u3002\u8fd9\u8868\u660e\u5bf9\u8fd9\u4e9b\u8bed\u8a00\uff0c\u5916\u90e8\u68c0\u7d22\u7684\u4e0a\u4e0b\u6587\u6216 Agent \u751f\u6210\u7684\u5197\u4f59\u89c4\u5212\u6b65\u9aa4\u4e25\u91cd\u5e72\u6270\u4e86\u6a21\u578b\u7684\u5185\u5728\u5224\u65ad\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u4ec5 C\u3001Go \u4e0e Java \u5728\u590d\u6742\u4e0a\u4e0b\u6587\u4e2d\u4fdd\u6301\u7a33\u5b9a\u751a\u81f3\u53d6\u5f97\u63d0\u5347\u3002\u6700\u7a81\u51fa\u7684\u4f8b\u5b50\u662f Claude-4.5-Sonnet\uff0c\u5b83\u901a\u8fc7 Agent \u6846\u67b6\u5728 Go\uff08$0.120rightarrow 0.218$\uff09\u3001Java\uff08$0.142rightarrow 0.241$\uff09\u4e0e C\uff08$0.106rightarrow 0.189$\uff09\u4e0a\u5b9e\u73b0\u4e86\u5b9e\u8d28\u6027\u6027\u80fd\u8dc3\u5347\u3002\u8fd9\u8868\u660e\uff0c\u5f53\u8003\u8651\u662f\u5426\u5411\u6a21\u578b\u63d0\u4f9b\u4e0a\u4e0b\u6587\u4fe1\u606f\u4ee5\u6539\u5584\u5176 ACR \u8868\u73b0\u65f6\uff0c\u7b54\u6848\u5b8c\u5168\u53d6\u51b3\u4e8e\u7f16\u7a0b\u8bed\u8a00\u3002<\/p>\n<h5>\u5173\u952e\u89c2\u5bdf<\/h5>\n<p>\u4e0d\u540c ACR \u65b9\u6cd5\u8868\u73b0\u51fa\u663e\u8457\u7684\u8bed\u8a00\u7279\u5f02\u6027\u504f\u5dee\u3002\u5bf9\u67d0\u4e9b\u8bed\u8a00\uff0c\u5f15\u5165\u4e0a\u4e0b\u6587\u68c0\u7d22\u65b9\u6cd5\u751a\u81f3\u53ef\u80fd\u964d\u4f4e\u6027\u80fd\u3002<\/p>\n<h2>5 \u9519\u8bef\u5206\u6790<\/h2>\n<p>\u6a21\u578b\u751f\u6210\u7684\u9519\u8bef\u8bc4\u5ba1\u8bc4\u8bba\u7684\u8be6\u7ec6\u6848\u4f8b\u7814\u7a76\u89c1\u9644\u5f55 D\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u5f53\u524d\u6a21\u578b\u5728\u4ee3\u7801\u8bc4\u5ba1\u4e2d\u4ecd\u6301\u7eed\u906d\u53d7\u77e5\u8bc6\u9519\u8bef\u3002\u6b64\u5916\uff0c\u4e0a\u4e0b\u6587\u68c0\u7d22\u5f15\u5165\u7684\u566a\u58f0\u6570\u636e\u88ab\u8bc6\u522b\u4e3a\u5bfc\u81f4\u751f\u6210\u9519\u8bef\u8bc4\u5ba1\u8bc4\u8bba\u7684\u91cd\u8981\u56e0\u7d20\u3002<\/p>\n<h2>6 \u7ed3\u8bba<\/h2>\n<p>\u6211\u4eec\u63d0\u51fa\u4e86 AACR-Bench\uff0c\u4e00\u4e2a\u65e8\u5728\u4ee5\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u8bc4\u6d4b ACR \u7cfb\u7edf\u7684\u591a\u8bed\u8a00\u57fa\u51c6\u3002AACR-Bench \u901a\u8fc7\u66f4\u597d\u5730\u8861\u91cf\u5404\u7c7b\u65b9\u6cd5\u5229\u7528\u590d\u6742\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u6267\u884c ACR \u7684\u80fd\u529b\uff0c\u586b\u8865\u4e86\u8be5\u9886\u57df\u7684\u4e00\u9879\u5173\u952e\u7a7a\u767d\u3002\u6211\u4eec\u5728 AACR-Bench \u4e0a\u7684\u5e7f\u6cdb\u5b9e\u8bc1\u8bc4\u6d4b\u8868\u660e\uff1a<\/p>\n<p>\u4e0a\u4e0b\u6587\u7684\u7c92\u5ea6\/\u5c42\u7ea7\u4ee5\u53ca\u68c0\u7d22\u65b9\u6cd5\u7684\u9009\u62e9\u4f1a\u663e\u8457\u5f71\u54cd ACR \u6027\u80fd\uff0c\u4e14\u8fd9\u79cd\u5f71\u54cd\u968f LLM\u3001\u7f16\u7a0b\u8bed\u8a00\u4ee5\u53ca LLM \u4f7f\u7528\u8303\u5f0f\uff08\u4f8b\u5982\u662f\u5426\u91c7\u7528 Agent \u67b6\u6784\uff09\u800c\u53d8\u5316\u3002<\/p>\n<p>\u5c3d\u7ba1\u4e0a\u8ff0\u53d1\u73b0\u6709\u529b\u5f3a\u8c03\u4e86\u6784\u5efa AACR-Bench \u8fd9\u7c7b\u6570\u636e\u96c6\u7684\u5fc5\u8981\u6027\uff0c\u6211\u4eec\u7684\u5de5\u4f5c\u4e5f\u5b58\u5728\u5c40\u9650\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5c3d\u7ba1\u6211\u4eec\u5229\u7528 LLM \u751f\u6210\u7684\u8bc4\u5ba1\u6765\u589e\u5f3a\u6570\u636e\u96c6\uff0c\u4f46\u7531\u4e8e\u771f\u5b9e\u8f6f\u4ef6\u7cfb\u7edf\u56fa\u6709\u7684\u590d\u6742\u6027\u4e0e\u4e3b\u89c2\u6027\uff0c\u6784\u5efa\u5b8c\u5168\u5168\u9762\u7684 Ground Truth \u4ecd\u662f\u4e00\u9879\u8270\u5de8\u6311\u6218\u3002\u672a\u6765\u5de5\u4f5c\u5c06\u805a\u7126\u4e8e\u8fdb\u4e00\u6b65\u6269\u5927\u6570\u636e\u96c6\u89c4\u6a21\uff0c\u5e76\u63a2\u7d22\u66f4\u5148\u8fdb\u7684\u534a\u81ea\u52a8\u65b9\u6cd5\u4ee5\u7cbe\u70bc Ground Truth \u8d28\u91cf\u3002<\/p>\n<h2>\u5f71\u54cd\u58f0\u660e<\/h2>\n<p>\u672c\u5de5\u4f5c\u5f15\u5165 AACR-Bench\uff0c\u8fd9\u662f\u9762\u5411\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\u7684\u9996\u4e2a\u591a\u8bed\u8a00\u3001\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u611f\u77e5\u57fa\u51c6\u3002\u901a\u8fc7\u5168\u9762\u523b\u753b\u5f53\u524d\u5927\u8bed\u8a00\u6a21\u578b\u7684\u80fd\u529b\u8fb9\u754c\uff0c\u6211\u4eec\u7684\u7814\u7a76\u8bc6\u522b\u51fa\u5173\u952e\u6311\u6218\uff0c\u5e76\u4e3a\u4e0b\u4e00\u4ee3 ACR \u7cfb\u7edf\u6307\u660e\u65b9\u5411\u3002<\/p>\n<h5>\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\u4e2d\u7684\u8303\u5f0f\u8f6c\u53d8<\/h5>\n<p>\u6211\u4eec\u7684\u7814\u7a76\u5efa\u7acb\u4e86\u4e25\u683c\u7684\u8bc4\u6d4b\u6807\u51c6\uff0c\u5e76\u63ed\u793a\u4e86\u672a\u6765 ACR \u65b9\u6cd5\u4e3a\u5b9e\u73b0\u5b9e\u7528\u6548\u7528\u5e94\u5e94\u5bf9\u7684\u4e09\u9879\u6839\u672c\u6311\u6218\uff1a<\/p>\n<ul>\n<li><strong>\u9a7e\u9a6d Precision\u2013Recall \u6743\u8861\uff1a<\/strong> \u6211\u4eec\u8bc6\u522b\u51fa\u4e00\u79cd\u9c9c\u660e\u7684\u4e8c\u5206\uff1a\u4f20\u7edf\u751f\u6210\u65b9\u6cd5\u6700\u5927\u5316 Recall \u4f46\u906d\u53d7\u9891\u7e41\u5e7b\u89c9\uff0c\u800c\u57fa\u4e8e Agent \u7684\u65b9\u6cd5\u5728\u7f3a\u9677\u5b9a\u4f4d\u4e0a\u8fbe\u5230\u9ad8 Precision \u4f46\u7f3a\u4e4f\u5168\u9762\u8986\u76d6\u3002\u8fd9\u4e00\u53d1\u73b0\u6566\u4fc3\u793e\u533a\u8d85\u8d8a\u5355\u4e00\u6307\u6807\u4f18\u5316\uff0c\u805a\u7126\u4e8e\u5728\u4fdd\u6301\u9ad8\u7cbe\u5ea6\u7684\u540c\u65f6\u6269\u5927\u7f3a\u9677\u68c0\u6d4b\u8986\u76d6\u7684\u6df7\u5408\u67b6\u6784\u3002<\/li>\n<li><strong>\u8fc8\u5411\u81ea\u9002\u5e94\u4e0a\u4e0b\u6587\u611f\u77e5\uff1a<\/strong> \u6211\u4eec\u6311\u6218\u300c\u66f4\u591a\u4e0a\u4e0b\u6587\u603b\u662f\u66f4\u597d\u300d\u7684\u6d41\u884c\u5047\u8bbe\u3002\u5b9e\u8bc1\u7ed3\u679c\u8868\u660e\uff0c\u68c0\u7d22\u4e0a\u4e0b\u6587\u7684\u5f71\u54cd\u5728\u4e0d\u540c\u8bed\u8a00\u95f4\u5dee\u5f02\u663e\u8457\uff08\u4f8b\u5982 C# vs. Python\uff09\uff0c\u4e0d\u76f8\u5173\u4e0a\u4e0b\u6587\u5f80\u5f80\u5145\u5f53\u566a\u58f0\u3002\u8fd9\u9700\u8981\u8f6c\u5411\u81ea\u9002\u5e94\u4e0a\u4e0b\u6587\u611f\u77e5\uff08Adaptive Context Awareness\uff09\uff0c\u8981\u6c42\u7cfb\u7edf\u5177\u5907\u5143\u8ba4\u77e5\u80fd\u529b\uff0c\u4ee5\u6839\u636e\u4ee3\u7801\u7279\u5f81\u4e0e\u4efb\u52a1\u7c7b\u578b\u52a8\u6001\u786e\u5b9a\u68c0\u7d22\u7684\u5fc5\u8981\u6027\u4e0e\u7c92\u5ea6\u3002<\/li>\n<li><strong>\u7edf\u4e00\u5c40\u90e8\u4e0e\u5168\u5c40\u89c6\u89d2\uff1a<\/strong> \u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u5c3d\u7ba1\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u6709\u52a9\u4e8e\u5168\u5c40\u4f9d\u8d56\u7406\u89e3\uff0c\u5176\u6240\u8bf1\u53d1\u7684\u566a\u58f0\u4f1a\u635f\u5bb3 Diff \u5185\u5c40\u90e8\u7f3a\u9677\u7684\u68c0\u6d4b\u3002\u76f8\u53cd\uff0c\u6613\u51fa\u73b0\u300c\u4e0a\u4e0b\u6587\u96a7\u9053\u300d\u7684 Agent \u5e38\u5e38\u5ffd\u7565\u660e\u663e\u7684\u5c40\u90e8\u9519\u8bef\u3002\u672a\u6765\u8fed\u4ee3\u5fc5\u987b\u53d1\u5c55\u52a8\u6001\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5c06\u5fae\u89c2\u8bed\u6cd5\u6838\u9a8c\u4e0e\u5b8f\u89c2\u8de8\u6587\u4ef6\u98ce\u9669\u8bc4\u4f30\u6709\u673a\u6574\u5408\u3002<\/li>\n<\/ul>\n<h5>\u4ece\u88ab\u52a8\u6444\u5165\u5230\u4e3b\u52a8\u5ba1\u8ba1<\/h5>\n<p>\u672c\u5de5\u4f5c\u91cd\u65b0\u6846\u5b9a\u4e86 ACR \u4e2d\u300c\u4e0a\u4e0b\u6587\u6709\u6548\u6027\u300d\u7684\u8bc4\u6d4b\u3002\u6211\u4eec\u8ba4\u4e3a\uff0c\u9ad8\u8d28\u91cf\u4ee3\u7801\u8bc4\u5ba1\u8f83\u5c11\u4f9d\u8d56\u4e8e\u4fe1\u606f\u68c0\u7d22\uff0c\u800c\u66f4\u591a\u4f9d\u8d56\u4e8e\u4fe1\u606f\u5229\u7528\u4e0e\u63a8\u7406\u7a33\u5065\u6027\u3002\u6211\u4eec\u7684\u53d1\u73b0\u8868\u660e\uff0c\u82e5\u6ca1\u6709\u8db3\u591f\u7684\u566a\u58f0\u5bb9\u5fcd\u80fd\u529b\uff0c\u5373\u4f7f\u901a\u8fc7 BM25 \u6216 Embedding \u68c0\u7d22\u5230\u76f8\u5173\u4ee3\u7801\u4e5f\u53ef\u80fd\u964d\u4f4e\u6027\u80fd\u3002Agent \u5728\u590d\u6742\u573a\u666f\u4e2d\u66f4\u4f18\u7684\u63a8\u7406\u8868\u660e\uff0c\u4ece\u300c\u88ab\u52a8\u4ee3\u7801\u6444\u5165\u300d\uff08\u63a5\u6536\u9884\u5148\u68c0\u7d22\u7684\u7247\u6bb5\uff09\u8f6c\u5411\u300c\u4e3b\u52a8\u4ee3\u7801\u5ba1\u8ba1\u300d\u7684\u8303\u5f0f\u8f6c\u53d8\u3002\u4f7f\u6a21\u578b\u80fd\u591f\u4e3b\u52a8\u63a2\u7d22\u4e0a\u4e0b\u6587\u3001\u901a\u8fc7\u591a\u8f6e\u4ea4\u4e92\u9a8c\u8bc1\u5047\u8bbe\u5e76\u8fc7\u6ee4\u566a\u58f0\u2014\u2014\u6a21\u4eff\u4eba\u7c7b\u4e13\u5bb6\u884c\u4e3a\u2014\u2014\u4ee3\u8868\u4e86\u5b9e\u73b0\u53ef\u9760\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\u6700\u6709\u524d\u666f\u7684\u65b9\u5411\u3002<\/p>\n<h2>\u53c2\u8003\u6587\u732e<\/h2>\n<ul>\n<li>Bacchelli &amp; Bird (2013) Bacchelli, A. and Bird, C. Expectations, outcomes, and challenges of modern code review. In <em>2013 35th International Conference on Software Engineering (ICSE)<\/em>, pp. 712\u2013721. IEEE, 2013.<\/li>\n<li>DeepSeek-AI (2025) DeepSeek-AI. Deepseek-v3.2: Pushing the frontier of open large language models, 2025.<\/li>\n<li>Distefano et al. (2019) Distefano, D., F\u00e4hndrich, M., Logozzo, F., and O\u2019Hearn, P. W. Scaling static analyses at facebook. <em>Communications of the ACM<\/em>, 62(8):62\u201370, 2019.<\/li>\n<li>Guo et al. (2025a) Guo, H., Zheng, X., Liao, Z., Yu, H., Di, P., Zhang, Z., and Dai, H.-N. Codefuse-cr-bench: A comprehensiveness-aware benchmark for end-to-end code review evaluation in python projects. <em>arXiv preprint arXiv:2509.14856<\/em>, 2025a.<\/li>\n<li>Guo et al. (2025b) Guo, J., Wang, C., Xu, X., Su, Z., and Zhang, X. Repoaudit: An autonomous llm-agent for repository-level code auditing. <em>arXiv preprint arXiv:2501.18160<\/em>, 2025b.<\/li>\n<li>Hou et al. (2024) Hou, X., Zhao, Y., Liu, Y., Yang, Z., Wang, K., Li, L., Luo, X., Lo, D., Grundy, J., and Wang, H. Large language models for software engineering: A systematic literature review. <em>ACM Transactions on Software Engineering and Methodology<\/em>, 33(8):1\u201379, 2024.<\/li>\n<li>Hu et al. (2025) Hu, R., Wang, X., Wen, X.-C., Zhang, Z., Jiang, B., Gao, P., Peng, C., and Gao, C. Benchmarking llms for fine-grained code review with enriched context in practice. <em>arXiv preprint arXiv:2511.07017<\/em>, 2025.<\/li>\n<li>Jimenez et al. (2023) Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., and Narasimhan, K. Swe-bench: Can language models resolve real-world github issues? <em>arXiv preprint arXiv:2310.06770<\/em>, 2023.<\/li>\n<li>Kapadnis et al. (2025) Kapadnis, M. N., Naik, A., and Rose, C. Crscore++: Reinforcement learning with verifiable tool and ai feedback for code review. <em>arXiv preprint arXiv:2506.00296<\/em>, 2025.<\/li>\n<li>Khoshnoud et al. (2022) Khoshnoud, F., Nasab, A. R., Toudeji, Z., and Sami, A. Which bugs are missed in code reviews: An empirical study on smartshark dataset. In <em>Proceedings of the 19th International Conference on Mining Software Repositories<\/em>, pp. 137\u2013141, 2022.<\/li>\n<li>Le et al. (2022) Le, H., Wang, Y., Gotmare, A. D., Savarese, S., and Hoi, S. C. H. Coderl: Mastering code generation through pretrained models and deep reinforcement learning. <em>Advances in Neural Information Processing Systems<\/em>, 35:21314\u201321328, 2022.<\/li>\n<li>Li et al. (2025) Li, S., Wang, D., Thongtanunam, P., Wang, Z., Yu, J., and Chen, J. Issue-oriented agent-based framework for automated review comment generation. <em>arXiv preprint arXiv:2511.00517<\/em>, 2025.<\/li>\n<li>Li et al. (2022) Li, Z., Lu, S., Guo, D., Duan, N., Jannu, S., Jenks, G., Majumder, D., Green, J., Svyatkovskiy, A., Fu, S., et al. Automating code review activities by large-scale pre-training. In <em>Proceedings of the 30th ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering<\/em>, pp. 1035\u20131047, 2022.<\/li>\n<li>Liu et al. (2025) Liu, C., Lin, H. Y., and Thongtanunam, P. Too noisy to learn: Enhancing data quality for code review comment generation. In <em>2025 IEEE\/ACM 22nd International Conference on Mining Software Repositories (MSR)<\/em>, pp. 236\u2013248. IEEE, 2025.<\/li>\n<li>Liu et al. (2023a) Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P. Lost in the middle: How language models use long contexts, 2023. URL https:\/\/arxiv.org\/abs\/2307.03172, 2023a.<\/li>\n<li>Liu et al. (2023b) Liu, T., Xu, C., and McAuley, J. Repobench: Benchmarking repository-level code auto-completion systems. <em>arXiv preprint arXiv:2306.03091<\/em>, 2023b.<\/li>\n<li>Liu et al. (2023c) Liu, X., Yu, H., Zhang, H., Xu, Y., Lei, X., Lai, H., Gu, Y., Ding, H., Men, K., Yang, K., et al. Agentbench: Evaluating llms as agents. <em>arXiv preprint arXiv:2308.03688<\/em>, 2023c.<\/li>\n<li>Lu et al. (2023) Lu, J., Yu, L., Li, X., Yang, L., and Zuo, C. Llama-reviewer: Advancing code review automation with large language models through parameter-efficient fine-tuning. In <em>2023 IEEE 34th International Symposium on Software Reliability Engineering (ISSRE)<\/em>, pp. 647\u2013658. IEEE, 2023.<\/li>\n<li>Ren et al. (2025) Ren, X., Dai, C., Huang, Q., Wang, Y., Liu, C., and Jiang, B. Hydra-reviewer: A holistic multi-agent system for automatic code review comment generation. <em>IEEE Transactions on Software Engineering<\/em>, 2025.<\/li>\n<li>Robertson et al. (2009) Robertson, S., Zaragoza, H., et al. The probabilistic relevance framework: Bm25 and beyond. <em>Foundations and trends\u00ae in information retrieval<\/em>, 3(4):333\u2013389, 2009.<\/li>\n<li>Sadowski et al. (2015) Sadowski, C., Van Gogh, J., Jaspan, C., Soderberg, E., and Winter, C. Tricorder: Building a program analysis ecosystem. In <em>2015 IEEE\/ACM 37th IEEE International Conference on Software Engineering<\/em>, volume 1, pp. 598\u2013608. IEEE, 2015.<\/li>\n<li>Sharanarthi &amp; Polineni (2025) Sharanarthi, T. and Polineni, S. Multi-agent llm collaboration for adaptive code review, debugging, and security analysis. In <em>2025 International Conference on Mechatronics, Robotics, and Artificial Intelligence (MRAI)<\/em>, pp. 541\u2013546. IEEE, 2025.<\/li>\n<li>Solmaz (2025) Solmaz, O. Google\u2019s Code Review Guidelines (GitHub Adaptation), September 2025. URL https:\/\/solmaz.io\/google-eng-practices-github. Accessed: 2026-01-28.<\/li>\n<li>Sun et al. (2025) Sun, T., Xu, J., Li, Y., Yan, Z., Zhang, G., Xie, L., Geng, L., Wang, Z., Chen, Y., Lin, Q., et al. Bitsai-cr: Automated code review via llm in practice. In <em>Proceedings of the 33rd ACM International Conference on the Foundations of Software Engineering<\/em>, pp. 274\u2013285, 2025.<\/li>\n<li>Tang et al. (2024) Tang, X., Kim, K., Song, Y., Lothritz, C., Li, B., Ezzini, S., Tian, H., Klein, J., and Bissyand\u00e9, T. F. Codeagent: Autonomous communicative agents for code review. <em>arXiv preprint arXiv:2402.02172<\/em>, 2024.<\/li>\n<li>Tao et al. (2012) Tao, Y., Dang, Y., Xie, T., Zhang, D., and Kim, S. How do software engineers understand code changes? an exploratory study in industry. In <em>Proceedings of the ACM SIGSOFT 20th International symposium on the foundations of software engineering<\/em>, pp. 1\u201311, 2012.<\/li>\n<li>Team et al. (2025) Team, G., Zeng, A., Lv, X., Zheng, Q., Hou, Z., Chen, B., Xie, C., Wang, C., Yin, D., Zeng, H., Zhang, J., Wang, K., Zhong, L., Liu, M., Lu, R., Cao, S., Zhang, X., Huang, X., Wei, Y., Cheng, Y., An, Y., Niu, Y., Wen, Y., Bai, Y., Du, Z., Wang, Z., Zhu, Z., Zhang, B., Wen, B., Wu, B., Xu, B., Huang, C., Zhao, C., Cai, C., Yu, C., Li, C., Ge, C., Huang, C., Zhang, C., Xu, C., Zhu, C., Li, C., Yin, C., Lin, D., Yang, D., Jiang, D., Ai, D., Zhu, E., Wang, F., Pan, G., Wang, G., Sun, H., Li, H., Li, H., Hu, H., Zhang, H., Peng, H., Tai, H., Zhang, H., Wang, H., Yang, H., Liu, H., Zhao, H., Liu, H., Yan, H., Liu, H., Chen, H., Li, J., Zhao, J., Ren, J., Jiao, J., Zhao, J., Yan, J., Wang, J., Gui, J., Zhao, J., Liu, J., Li, J., Li, J., Lu, J., Wang, J., Yuan, J., Li, J., Du, J., Du, J., Liu, J., Zhi, J., Gao, J., Wang, K., Yang, L., Xu, L., Fan, L., Wu, L., Ding, L., Wang, L., Zhang, M., Li, M., Xu, M., Zhao, M., Zhai, M., Du, P., Dong, Q., Lei, S., Tu, S., Yang, S., Lu, S., Li, S., Li, S., Shuang-Li, Yang, S., Yi, S., Yu, T., Tian, W., Wang, W., Yu, W., Tam, W. L., Liang, W., Liu, W., Wang, X., Jia, X., Gu, X., Ling, X., Wang, X., Fan, X., Pan, X., Zhang, X., Zhang, X., Fu, X., Zhang, X., Xu, Y., Wu, Y., Lu, Y., Wang, Y., Zhou, Y., Pan, Y., Zhang, Y., Wang, Y., Li, Y., Su, Y., Geng, Y., Zhu, Y., Yang, Y., Li, Y., Wu, Y., Li, Y., Liu, Y., Wang, Y., Li, Y., Zhang, Y., Liu, Z., Yang, Z., Zhou, Z., Qiao, Z., Feng, Z., Liu, Z., Zhang, Z., Wang, Z., Yao, Z., Wang, Z., Liu, Z., Chai, Z., Li, Z., Zhao, Z., Chen, W., Zhai, J., Xu, B., Huang, M., Wang, H., Li, J., Dong, Y., and Tang, J. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models, 2025. URL https:\/\/arxiv.org\/abs\/2508.06471.<\/li>\n<li>Team (2025) Team, Q. Qwen3 technical report, 2025. URL https:\/\/arxiv.org\/abs\/2505.09388.<\/li>\n<li>Tufano et al. (2022) Tufano, R., Masiero, S., Mastropaolo, A., Pascarella, L., Poshyvanyk, D., and Bavota, G. Using pre-trained models to boost code review automation. In <em>Proceedings of the 44th international conference on software engineering<\/em>, pp. 2291\u20132302, 2022.<\/li>\n<li>Yu et al. (2024a) Yu, Y., Rong, G., Shen, H., Zhang, H., Shao, D., Wang, M., Wei, Z., Xu, Y., and Wang, J. Fine-tuning large language models to improve accuracy and comprehensibility of automated code review. <em>ACM transactions on software engineering and methodology<\/em>, 34(1):1\u201326, 2024a.<\/li>\n<li>Yu et al. (2024b) Yu, Y., Zhang, L., Rong, G., Shen, H., Zhang, J., Yan, H., Shi, G., Shao, D., Pan, R., Li, Y., et al. Distilling desired comments for enhanced code review with large language models. <em>arXiv preprint arXiv:2412.20340<\/em>, 2024b.<\/li>\n<li>Zeng et al. (2025) Zeng, Z., Shi, R., Han, K., Li, Y., Sun, K., Wang, Y., Yu, Z., Xie, R., Ye, W., and Zhang, S. Benchmarking and studying the llm-based code review. <em>arXiv preprint arXiv:2509.01494<\/em>, 2025.<\/li>\n<li>Zhang et al. Zhang, F., Chen, B., Zhang, Y., Keung, J., Liu, J., Zan, D., Mao, Y., Lou, J.-G., and Chen, W. Repocoder: Repository-level code completion through iterative retrieval and generation. In <em>The 2023 Conference on Empirical Methods in Natural Language Processing<\/em>.<\/li>\n<li>Zhang et al. (2023) Zhang, Q., Fang, C., Xie, Y., Zhang, Y., Yang, Y., Sun, W., Yu, S., and Chen, Z. A survey on large language models for software engineering. <em>arXiv preprint arXiv:2312.15223<\/em>, 2023.<\/li>\n<li>Zhang et al. (2025) Zhang, Y., Zhang, Y., Sun, Z., Jiang, Y., and Liu, H. Laura: Enhancing code review generation with context-enriched retrieval-augmented llm. <em>arXiv preprint arXiv:2512.01356<\/em>, 2025.<\/li>\n<\/ul>\n<h2>\u9644\u5f55 A \u9644\u5f55\u6982\u89c8<\/h2>\n<p>\u9644\u5f55\u7ec4\u7ec7\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\u7b2c B \u8282\u63d0\u4f9b AACR-Bench \u7684\u8be6\u7ec6\u4fe1\u606f\uff0c\u5305\u62ec\u6570\u636e\u6784\u5efa\u8fc7\u7a0b\u3001\u4e0e\u73b0\u6709\u57fa\u51c6\u7684\u6bd4\u8f83\uff0c\u4ee5\u53ca\u4ee3\u8868\u6027\u6570\u636e\u6837\u672c\u3002<\/li>\n<li>\u7b2c C \u8282\u63cf\u8ff0 AACR-Bench \u7684\u5b9e\u9a8c\u8bbe\u7f6e\uff0c\u6db5\u76d6\u5b8c\u6574\u8bc4\u6d4b\u6d41\u6c34\u7ebf\u3001\u6240\u6d4b\u6a21\u578b\u3001\u8d85\u53c2\u6570\u8bbe\u7f6e\uff0c\u4ee5\u53ca\u7ec6\u7c92\u5ea6\u7ed3\u679c\u5206\u6790\u3002<\/li>\n<li>\u7b2c D \u8282\u63d0\u4f9b\u57fa\u4e8e AACR-Bench \u8bc4\u6d4b\u7ed3\u679c\u7684\u6848\u4f8b\u7814\u7a76\u3002<\/li>\n<\/ul>\n<h2>\u9644\u5f55 B \u6570\u636e\u96c6\u7ec6\u8282<\/h2>\n<h3>B.1 \u8be6\u7ec6\u6570\u636e\u96c6\u6784\u5efa\u8fc7\u7a0b<\/h3>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 4\uff1aAACR-Bench \u7684\u6784\u5efa\u8fc7\u7a0b\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig04.png\" \/><\/p>\n<p><strong>\u56fe 4\uff1a<\/strong> AACR-Bench \u7684\u6784\u5efa\u8fc7\u7a0b<\/p>\n<p>AACR-Bench \u7684\u6784\u5efa\u8fc7\u7a0b\u5982\u56fe 4 \u6240\u793a\u3002<\/p>\n<h5>\u7f16\u7a0b\u8bed\u8a00\u4e0e\u4ed3\u5e93\u7684\u9009\u62e9<\/h5>\n<p>\u4e3a\u786e\u4fdd\u8bc4\u6d4b\u6570\u636e\u7684\u65f6\u6548\u6027\u4e0e\u7f16\u7a0b\u8bed\u8a00\u7684\u591a\u6837\u6027\uff0c\u6211\u4eec\u4f9d\u636e StackOverflow Developer Survey 2025 \u9009\u53d6\u4e86\u6392\u540d\u524d\u5341\u7684\u7f16\u7a0b\u8bed\u8a00\uff0c\u5177\u4f53\u4e3a\uff1aJavaScript\u3001Python\u3001TypeScript\u3001Java\u3001C#\u3001C++\u3001C\u3001PHP\u3001Go \u4e0e Rust\u3002\u5bf9\u6bcf\u79cd\u8bed\u8a00\uff0c\u6211\u4eec\u9009\u53d6\u8be5\u8bed\u8a00\u4f5c\u4e3a\u4e3b\u8981\u7f16\u7a0b\u8bed\u8a00\u7684\u4ed3\u5e93\uff0c\u4f5c\u4e3a PR \u4fe1\u606f\u63d0\u53d6\u7684\u6570\u636e\u6e90\uff0c\u5e76\u9075\u5faa\u4ee5\u4e0b\u6807\u51c6\uff1a\u9996\u5148\uff0c\u6211\u4eec\u5b9a\u4e49\u4e00\u4e2a\u6d3b\u8dc3 GitHub \u4ed3\u5e93\u5019\u9009\u96c6\uff0c\u8fd9\u4e9b\u4ed3\u5e93\u5728 2024 \u5e74 12 \u6708 1 \u65e5\u81f3 2025 \u5e74 12 \u6708 1 \u65e5\u671f\u95f4\u65b0\u589e star \u6570\u4e0e\u5df2\u5173\u95ed PR \u6570\u5747\u8fdb\u5165\u524d 2,000\u3002\u968f\u540e\uff0c\u6211\u4eec\u6309\u65b0\u589e star \u6570\u5bf9\u5019\u9009\u4ed3\u5e93\u6392\u5e8f\uff0c\u5e76\u4e3a\u6bcf\u79cd\u8bed\u8a00\u9009\u53d6\u524d\u4e94\u540d\u3002\u6700\u7ec8\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u8de8\u8d8a 10 \u79cd\u7f16\u7a0b\u8bed\u8a00\u7684 50 \u4e2a\u9ad8\u6d3b\u8dc3\u4ed3\u5e93\u96c6\u5408\uff08\u6bcf\u79cd\u8bed\u8a00\u4e94\u4e2a\u4ed3\u5e93\uff09\uff0c\u4f5c\u4e3a\u63d0\u53d6 PR \u4e0e\u8bc4\u5ba1\u8bc4\u8bba\u7684\u6765\u6e90\u3002<\/p>\n<h5>PR \u8fc7\u6ee4\u4e0e\u8bc4\u5ba1\u8bc4\u8bba\u589e\u5f3a<\/h5>\n<p>\u805a\u7126\u4e8e\u9009\u5b9a\u7684 50 \u4e2a\u4ed3\u5e93\uff0c\u6211\u4eec\u5229\u7528 GitHub API \u6536\u96c6 2024 \u5e74 12 \u6708 1 \u65e5\u81f3 2025 \u5e74 12 \u6708 1 \u65e5\u671f\u95f4\u521b\u5efa\u7684\u5168\u90e8 Pull Request\uff08PR\uff09\uff0c\u5171\u5f97\u5230 $12{,}715$ \u6761\u8bb0\u5f55\u3002\u5bf9\u6bcf\u4e2a PR\uff0c\u6211\u4eec\u63d0\u53d6\u4ee5\u4e0b\u5173\u952e\u4fe1\u606f\uff1a<\/p>\n<ul>\n<li>PR \u7684\u6807\u9898\u4e0e\u63cf\u8ff0\uff1b<\/li>\n<li>\u53d8\u66f4\u7684\u4ee3\u7801\u884c\u6570\uff1b<\/li>\n<li>PR \u7684 base commit\uff1b<\/li>\n<li>\u5168\u9762\u7684\u8bc4\u5ba1\u8bc4\u8bba\u6570\u636e\uff0c\u5305\u62ec\u76ee\u6807\u4fee\u8ba2\u7248\u672c\u3001\u76f8\u5173\u6587\u4ef6\u8def\u5f84\u3001\u884c\u8303\u56f4\uff0c\u4ee5\u53ca\u8bc4\u5ba1\u7684\u5177\u4f53\u5185\u5bb9\u3002<\/li>\n<\/ul>\n<p>\u57fa\u4e8e\u4e0a\u8ff0\u539f\u59cb\u6570\u636e\uff0c\u6211\u4eec\u5b9e\u65bd\u4e86\u4ee5\u4e0b\u9884\u5904\u7406\u6b65\u9aa4\uff1a<\/p>\n<ul>\n<li><strong>\u9886\u57df\u5206\u7c7b\uff1a<\/strong> \u4f7f\u7528\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5206\u6790 PR \u7684\u95ee\u9898\u57df\uff0c\u9075\u5faa SWE-Bench \u4e2d\u5b9a\u4e49\u7684\u5206\u7c7b\u4f53\u7cfb\uff0c\u5982\u8868 5 \u6240\u793a\uff1b<\/li>\n<li><strong>\u8bed\u8a00\u8bc6\u522b\uff1a<\/strong> \u5206\u6790 PR \u6807\u9898\u4e0e\u63cf\u8ff0\u6240\u4f7f\u7528\u7684\u81ea\u7136\u8bed\u8a00\uff1b<\/li>\n<li><strong>\u4fee\u8ba2\u7248\u672c\u9009\u62e9\uff1a<\/strong> \u8bc6\u522b\u5e76\u63d0\u53d6\u5305\u542b\u8bc4\u5ba1\u8bc4\u8bba\u6570\u91cf\u6700\u591a\u7684\u4fee\u8ba2\u7248\u672c\uff0c\u4ee5\u53ca\u5168\u90e8\u5bf9\u5e94\u8bc4\u5ba1\u6570\u636e\uff1b<\/li>\n<li><strong>\u89c4\u6a21\u5206\u7c7b\uff1a<\/strong> \u57fa\u4e8e\u53d8\u66f4\u884c\u6570\u5bf9 PR \u89c4\u6a21\u8fdb\u884c\u5206\u7c7b\uff0c\u9075\u5faa T-Shirt Size \u5206\u7c7b\u65b9\u6cd5\uff0c\u5982\u8868 6 \u6240\u793a\u3002<\/li>\n<li><strong>\u88ab\u8bc4\u5ba1\u7f16\u7a0b\u8bed\u8a00\u8bc6\u522b\uff1a<\/strong> \u57fa\u4e8e\u8bc4\u5ba1\u8bc4\u8bba\u6240\u9488\u5bf9\u6587\u4ef6\u7684\u6269\u5c55\u540d\uff0c\u8bc6\u522b\u88ab\u8bc4\u5ba1\u4ee3\u7801\u7684\u4e3b\u8981\u7f16\u7a0b\u8bed\u8a00\u3002<\/li>\n<\/ul>\n<p><strong>\u8868 5\uff1a<\/strong> PR \u95ee\u9898\u57df\u5206\u7c7b<\/p>\n<table>\n<thead>\n<tr>\n<th>\u7c7b\u522b\uff08\u7f29\u5199\uff09<\/th>\n<th>\u63cf\u8ff0<\/th>\n<th style=\"text-align: right\">PR \u6570\u91cf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Bug Fixes (BF)<\/td>\n<td>\u89e3\u51b3\u529f\u80fd\u6027\u9519\u8bef\u3001\u5d29\u6e83\u3001\u4e0d\u6b63\u786e\u8f93\u51fa<\/td>\n<td style=\"text-align: right\">53<\/td>\n<\/tr>\n<tr>\n<td>New Feature Additions (NFA)<\/td>\n<td>\u5411\u5e94\u7528\u6dfb\u52a0\u65b0\u529f\u80fd\u6216\u7279\u6027<\/td>\n<td style=\"text-align: right\">44<\/td>\n<\/tr>\n<tr>\n<td>Code Refactoring (CA)<\/td>\n<td>\u5728\u4e0d\u6539\u53d8\u5916\u90e8\u884c\u4e3a\u7684\u524d\u63d0\u4e0b\u6539\u5584\u4ee3\u7801\u7ed3\u6784\u3001\u53ef\u8bfb\u6027\u3001\u53ef\u7ef4\u62a4\u6027<\/td>\n<td style=\"text-align: right\">28<\/td>\n<\/tr>\n<tr>\n<td>Documentation Update (DU)<\/td>\n<td>\u4e0e\u4ee3\u7801\u6ce8\u91ca\u6216\u5916\u90e8\u6587\u6863\u76f8\u5173\u7684\u53d8\u66f4<\/td>\n<td style=\"text-align: right\">12<\/td>\n<\/tr>\n<tr>\n<td>Test Suite \/ CI Enhancements (TC)<\/td>\n<td>\u6539\u5584\u6d4b\u8bd5\u8986\u76d6\u3001\u6d4b\u8bd5\u8d28\u91cf\u6216\u6301\u7eed\u96c6\u6210\u6d41\u7a0b<\/td>\n<td style=\"text-align: right\">20<\/td>\n<\/tr>\n<tr>\n<td>Performance Optimizations (PO)<\/td>\n<td>\u6539\u5584\u5e94\u7528\u901f\u5ea6\u3001\u54cd\u5e94\u65f6\u95f4\u6216\u8d44\u6e90\u4f7f\u7528\u6548\u7387<\/td>\n<td style=\"text-align: right\">19<\/td>\n<\/tr>\n<tr>\n<td>Security Patches (SV)<\/td>\n<td>\u4fee\u590d\u53ef\u80fd\u5bfc\u81f4\u5b89\u5168\u95ee\u9898\u7684\u4ee3\u7801\u7f3a\u9677<\/td>\n<td style=\"text-align: right\">13<\/td>\n<\/tr>\n<tr>\n<td>Dependency Updates (DE)<\/td>\n<td>\u66f4\u65b0\u7b2c\u4e09\u65b9\u5e93\u4f9d\u8d56\u6216\u786e\u4fdd\u8de8\u73af\u5883\u517c\u5bb9\u6027<\/td>\n<td style=\"text-align: right\">3<\/td>\n<\/tr>\n<tr>\n<td>Code Style, Linting (CLF)<\/td>\n<td>\u786e\u4fdd\u4ee3\u7801\u7b26\u5408\u56e2\u961f\u7f16\u7801\u6807\u51c6\u4e0e\u4e00\u81f4\u6027<\/td>\n<td style=\"text-align: right\">8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u8868 6\uff1a<\/strong> PR \u89c4\u6a21\u7684 T-Shirt Size \u5206\u7c7b<\/p>\n<table>\n<thead>\n<tr>\n<th>\u89c4\u6a21<\/th>\n<th>\u53d8\u66f4\u884c\u6570<\/th>\n<th style=\"text-align: right\">PR \u6570\u91cf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>XS<\/td>\n<td>0 - 9<\/td>\n<td style=\"text-align: right\">8<\/td>\n<\/tr>\n<tr>\n<td>S<\/td>\n<td>10 - 29<\/td>\n<td style=\"text-align: right\">22<\/td>\n<\/tr>\n<tr>\n<td>M<\/td>\n<td>30 - 99<\/td>\n<td style=\"text-align: right\">47<\/td>\n<\/tr>\n<tr>\n<td>L<\/td>\n<td>100 - 499<\/td>\n<td style=\"text-align: right\">89<\/td>\n<\/tr>\n<tr>\n<td>XL<\/td>\n<td>500 - 999<\/td>\n<td style=\"text-align: right\">34<\/td>\n<\/tr>\n<tr>\n<td>XXL<\/td>\n<td>1000+<\/td>\n<td style=\"text-align: right\">0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5bf9\u9886\u57df\u5206\u7c7b\u4efb\u52a1\uff0c\u6211\u4eec\u91c7\u7528 Qwen3-235B-A22B-Thinking-2507 \u6a21\u578b\uff1b\u6240\u7528\u5177\u4f53 prompt \u89c1\u56fe 5\u3002\u4e3a\u8bc4\u4f30\u5206\u7c7b\u53ef\u9760\u6027\uff0c\u6211\u4eec\u4ece\u7ed3\u679c\u4e2d\u968f\u673a\u62bd\u6837 350 \u4e2a\u5b9e\u4f8b\u8fdb\u884c\u4eba\u5de5\u6838\u9a8c\u3002\u5206\u6790\u663e\u793a\u51c6\u786e\u7387\u4e3a $92.36%$\u3002\u5728 $95%$ \u7f6e\u4fe1\u6c34\u5e73\u4e0b\uff0c\u8be5\u7ed3\u679c\u7684\u8bef\u5dee\u8303\u56f4\u4e3a $5.06%$\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 5\uff1a\u7528\u4e8e PR \u95ee\u9898\u57df\u5206\u7c7b\u7684 Prompt\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig05.png\" \/><\/p>\n<p><strong>\u56fe 5\uff1a<\/strong> \u7528\u4e8e PR \u95ee\u9898\u57df\u5206\u7c7b\u7684 Prompt<\/p>\n<p>\u5728\u9884\u5904\u7406\u6570\u636e\u7684\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u5efa\u7acb\u4e86\u4e94\u6761\u8fc7\u6ee4\u6807\u51c6\u4ee5\u81ea\u52a8\u7b5b\u9009\u9ad8\u8d28\u91cf\u6837\u672c\uff0c\u5f97\u5230 3,328 \u4e2a PR \u6761\u76ee\uff1a<\/p>\n<ul>\n<li><strong>\u8bed\u8a00\u6807\u51c6\uff1a<\/strong> PR \u7684\u6807\u9898\u4e0e\u63cf\u8ff0\u5fc5\u987b\u4ee5\u82f1\u6587\u64b0\u5199\u3002<\/li>\n<li><strong>\u89c4\u6a21\u7ea6\u675f\uff1a<\/strong> \u53d8\u66f4\u4ee3\u7801\u884c\u6570\u5fc5\u987b\u9650\u5236\u4e3a 1,000\u3002\u6839\u636e Google \u7684\u4ee3\u7801\u8bc4\u5ba1\u5b9e\u8df5\uff0c\u8d85\u8fc7\u8be5\u9608\u503c\u7684\u53d8\u66f4\u88ab\u8ba4\u4e3a\u8fc7\u5927\u800c\u65e0\u6cd5\u6709\u6548\u8bc4\u5ba1\u3002<\/li>\n<li><strong>\u8bed\u8a00\u4e00\u81f4\u6027\uff1a<\/strong> \u4e3b\u8981\u4fee\u6539\u6587\u4ef6\u7684\u7f16\u7a0b\u8bed\u8a00\u5fc5\u987b\u4e0e\u4ed3\u5e93\u4e3b\u8bed\u8a00\u5339\u914d\u3002<\/li>\n<li><strong>\u8bc4\u5ba1\u4e30\u5bcc\u5ea6\uff1a<\/strong> PR \u5fc5\u987b\u5305\u542b\u81f3\u5c11\u4e24\u6761\u884c\u5185\u8bc4\u8bba\uff0c\u5176\u4e2d\u5305\u62ec\u81f3\u5c11\u4e00\u6761\u88ab\u63a5\u53d7\uff08\u5373\u5bfc\u81f4\u4ee3\u7801\u4fee\u6539\uff09\u7684\u5efa\u8bbe\u6027\u8bc4\u8bba\u3002<\/li>\n<\/ul>\n<p>\u6211\u4eec\u5bf9\u521d\u6b65\u8fc7\u6ee4\u540e\u7684 PR \u8fdb\u884c\u4e86\u7b2c\u4e8c\u8f6e\u4eba\u5de5\u6838\u9a8c\u3002\u805a\u7126\u8bed\u4e49\u6709\u6548\u6027\uff0c\u6211\u4eec\u6392\u9664\u4e86\u8131\u79bb\u9879\u76ee\u4e1a\u52a1\u4e0a\u4e0b\u6587\u6216\u7f3a\u4e4f\u5b9e\u9645\u8bed\u4e49\u542b\u4e49\u7684\u7410\u788e\u53d8\u66f4\u3002\u8be5\u8fc7\u7a0b\u5f97\u5230\u6700\u7ec8\u7684 573 \u4e2a PR \u6570\u636e\u96c6\u3002<\/p>\n<p>\u6700\u540e\uff0c\u4e3a\u786e\u4fdd\u57fa\u51c6\u7684\u4ee3\u8868\u6027\u4e0e\u591a\u6837\u6027\uff0c\u6211\u4eec\u57fa\u4e8e\u6765\u6e90\u4ed3\u5e93\u3001\u95ee\u9898\u57df\u4e0e PR \u89c4\u6a21\u5bf9\u5019\u9009\u6c60\u8fdb\u884c\u5206\u5c42\u62bd\u6837\u3002\u8be5\u8fc7\u7a0b\u5f97\u5230\u7531 200 \u4e2a PR \u7ec4\u6210\u7684\u6700\u7ec8\u6570\u636e\u96c6\u3002\u8fd9\u4e9b PR \u5728\u95ee\u9898\u57df\u4e0e\u89c4\u6a21\u4e0a\u7684\u5206\u5e03\u5206\u522b\u89c1\u56fe 6 \u4e0e\u56fe 7\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 6\uff1aPR \u5728\u5404\u95ee\u9898\u57df\u4e0a\u7684\u5206\u5e03\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig06.png\" \/><\/p>\n<p><strong>\u56fe 6\uff1a<\/strong> PR \u5728\u5404\u95ee\u9898\u57df\u4e0a\u7684\u5206\u5e03<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 7\uff1aPR \u5728\u4e0d\u540c\u89c4\u6a21\u4e0a\u7684\u5206\u5e03\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig07.png\" \/><\/p>\n<p><strong>\u56fe 7\uff1a<\/strong> PR \u5728\u4e0d\u540c\u89c4\u6a21\u4e0a\u7684\u5206\u5e03<\/p>\n<p>\u9274\u4e8e\u4ee3\u7801\u8bc4\u5ba1\u5e38\u6d89\u53ca\u8bc4\u5ba1\u8005\u4e0e\u5f00\u53d1\u8005\u4e4b\u95f4\u7684\u591a\u8f6e\u5bf9\u8bdd\u4ee5\u6838\u9a8c\u95ee\u9898\u6709\u6548\u6027\uff0c\u539f\u59cb\u8bc4\u8bba\u53ef\u80fd\u906d\u53d7\u4e0a\u4e0b\u6587\u7f3a\u5931\u4e0e\u566a\u58f0\u3002\u4e3a\u5e94\u5bf9\u6b64\u95ee\u9898\uff0c\u6211\u4eec\u4f7f\u7528 LLM \u5bf9\u9009\u5b9a\u4fee\u8ba2\u7248\u672c\u7684\u8bc4\u5ba1\u7ebf\u7a0b\u8fdb\u884c\u6df1\u5ea6\u8bed\u4e49\u5206\u6790\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u4ece\u8fd9\u4e9b\u591a\u8f6e\u4ea4\u4e92\u4e2d\u63d0\u53d6\u5df2\u786e\u8ba4\u7684\u4ee3\u7801\u7f3a\u9677\uff0c\u5e76\u5c06\u5176\u5408\u6210\u4e3a\u300c\u589e\u5f3a\u8bc4\u5ba1\u8bc4\u8bba\u300d\uff0c\u540c\u65f6\u4e22\u5f03\u672a\u6307\u51fa\u5b9e\u8d28\u6027\u7684\u8bc4\u8bba\u3002\u8be5\u8fc7\u7a0b\u6240\u7528 prompt \u89c1\u56fe 8\u3002\u6211\u4eec\u5c06\u6b64\u589e\u5f3a\u8fc7\u7a0b\u5e94\u7528\u4e8e $1{,}119$ \u4e2a\u5bf9\u8bdd\u7ebf\u7a0b\uff0c\u5e76\u62bd\u6837 $300$ \u4e2a\u7ed3\u679c\u8fdb\u884c\u6b63\u786e\u6027\u7684\u4eba\u5de5\u8bc4\u4f30\u3002\u82e5\u6a21\u578b\u51c6\u786e\u63d0\u4f9b\u8bc6\u522b\u7ebf\u7a0b\u4e2d\u5df2\u786e\u8ba4\u4ee3\u7801\u95ee\u9898\u7684\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u6216\u6b63\u786e\u5224\u5b9a\u8be5\u7ebf\u7a0b\u4e0d\u5305\u542b\u4ee3\u7801\u95ee\u9898\uff0c\u5219\u589e\u5f3a\u7ed3\u679c\u88ab\u5b9a\u4e49\u4e3a\u300c\u6b63\u786e\u300d\u3002\u8bc4\u4f30\u7ed3\u679c\u8868\u660e\u51c6\u786e\u7387\u4e3a 95%\uff0c\u7f6e\u4fe1\u6c34\u5e73\u4e3a 95%\uff0c\u8bef\u5dee\u8303\u56f4\u4e3a 4.74%\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 8\uff1a\u7528\u4e8e\u8bc4\u5ba1\u8bc4\u8bba\u589e\u5f3a\u7684 Prompt\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig08.png\" \/><\/p>\n<p><strong>\u56fe 8\uff1a<\/strong> \u7528\u4e8e\u8bc4\u5ba1\u8bc4\u8bba\u589e\u5f3a\u7684 Prompt<\/p>\n<p><strong>\u8868 7\uff1a<\/strong> \u8bc4\u5ba1\u8bc4\u8bba\u4e2d\u63ed\u793a\u7684\u95ee\u9898\u7c7b\u522b<\/p>\n<table>\n<thead>\n<tr>\n<th>\u95ee\u9898\u7c7b\u522b<\/th>\n<th>\u63cf\u8ff0<\/th>\n<th style=\"text-align: right\">\u8bc4\u8bba\u6570\u91cf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Security Vulnerability<\/td>\n<td>\u4ee3\u7801\u4e2d\u53ef\u80fd\u5bfc\u81f4\u6570\u636e\u6cc4\u9732\u3001\u672a\u6388\u6743\u8bbf\u95ee\u6216\u6613\u53d7\u5176\u4ed6\u6076\u610f\u653b\u51fb\u7684\u5b89\u5168\u5f31\u70b9<\/td>\n<td style=\"text-align: right\">53<\/td>\n<\/tr>\n<tr>\n<td>Code Defect<\/td>\n<td>\u53ef\u80fd\u5bfc\u81f4\u8fd0\u884c\u65f6\u5d29\u6e83\u3001\u4ea7\u751f\u4e0d\u6b63\u786e\u7ed3\u679c\u6216\u5bfc\u81f4\u610f\u5916\u7cfb\u7edf\u884c\u4e3a\u7684\u903b\u8f91\u9519\u8bef\u6216\u5b9e\u73b0\u7f3a\u9677<\/td>\n<td style=\"text-align: right\">709<\/td>\n<\/tr>\n<tr>\n<td>Maintainability &amp; Readability<\/td>\n<td>\u4e0e\u7f16\u7801\u98ce\u683c\u6216\u7ed3\u6784\u8bbe\u8ba1\u76f8\u5173\u3001\u964d\u4f4e\u4ee3\u7801\u53ef\u8bfb\u6027\u5e76\u963b\u788d\u672a\u6765\u7406\u89e3\u4e0e\u7ef4\u62a4\u7684\u95ee\u9898<\/td>\n<td style=\"text-align: right\">626<\/td>\n<\/tr>\n<tr>\n<td>Performance Issue<\/td>\n<td>\u7b97\u6cd5\u4f4e\u6548\u6216\u8d44\u6e90\u7ba1\u7406\u4e0d\u5f53\u5bfc\u81f4\u7684\u975e\u529f\u80fd\u6027\u74f6\u9888\uff0c\u4f8b\u5982\u9ad8\u5ef6\u8fdf\u3001\u541e\u5410\u91cf\u4e0d\u8db3\u6216\u8d44\u6e90\u6d88\u8017\u8fc7\u9ad8<\/td>\n<td style=\"text-align: right\">117<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h5>\u8bc4\u5ba1\u8865\u5168\u4e0e\u4e13\u5bb6\u6807\u6ce8<\/h5>\n<p>\u53d7\u9650\u4e8e\u4eba\u7c7b\u8bc4\u5ba1\u8005\u7684\u8ba4\u77e5\u8fb9\u754c\u4e0e\u80fd\u529b\u9650\u5236\uff0c\u4ec5\u4f9d\u8d56\u4eba\u5de5\u52aa\u529b\u5f80\u5f80\u65e0\u6cd5\u53d1\u73b0\u5168\u90e8\u6f5c\u5728\u4ee3\u7801\u7f3a\u9677\u3002\u4e3a\u5e94\u5bf9\u7531\u6b64\u4ea7\u751f\u7684\u300c\u95ee\u9898\u6807\u6ce8\u4e0d\u5b8c\u6574\u300d\u95ee\u9898\uff0c\u6211\u4eec\u91c7\u7528 LLM \u751f\u6210\u6280\u672f\u5168\u9762\u589e\u5f3a\u8bc4\u5ba1\u8bc4\u8bba\u3002\u4e3a\u51cf\u8f7b\u5355\u6a21\u578b\u504f\u5dee\u5e76\u786e\u4fdd\u591a\u6837\u6027\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u7531\u516d\u4e2a\u4e3b\u6d41\u5f00\u6e90\u4e0e\u4e13\u6709\u6a21\u578b\u7ec4\u6210\u7684\u751f\u6210\u77e9\u9635\u3002\u8fd9\u4e9b\u6a21\u578b\u901a\u8fc7\u4e24\u4e2a\u5f02\u6784\u6846\u67b6\u5e76\u884c\u751f\u6210\u8bc4\u5ba1\u8bc4\u8bba\uff1a\u5185\u90e8\u8bc4\u5ba1\u7cfb\u7edf\u4e0e\u5f00\u6e90 Agent \u7cfb\u7edf\uff08Claude Code\uff09\u3002\u6240\u6709\u751f\u6210\u8bc4\u8bba\u5728\u4e0e\u5148\u524d\u589e\u5f3a\u7684\u4eba\u5de5\u8bc4\u5ba1\u8bc4\u8bba\u5408\u5e76\u524d\u90fd\u7ecf\u8fc7\u8bed\u4e49\u53bb\u91cd\uff0c\u5f62\u6210\u6838\u9a8c\u5019\u9009\u96c6\u3002\u8be5\u8bed\u4e49\u53bb\u91cd\u4f7f\u7528 Qwen3-235B-A22B-Thinking-2507 \u6267\u884c\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6240\u6709\u8bc4\u5ba1\u8bc4\u8bba\u9996\u5148\u57fa\u4e8e\u5176\u4ed3\u5e93\u3001Pull Request\uff08PR\uff09\u3001\u6587\u4ef6\u8def\u5f84\u4ee5\u53ca\u6240\u9488\u5bf9\u7684\u7279\u5b9a Diff Hunk \u5206\u7ec4\u3002\u5728\u6bcf\u7ec4\u5185\uff0c\u8bc4\u8bba\u901a\u8fc7 LLM \u8fdb\u884c\u4e24\u4e24\u6bd4\u8f83\uff0c\u5e76\u6839\u636e\u6bd4\u8f83\u7ed3\u679c\u53bb\u9664\u91cd\u590d\u9879\u3002\u8be5\u6bd4\u8f83\u6240\u7528 prompt \u89c1\u56fe 9\u3002\u6211\u4eec\u91c7\u7528\u8fd0\u884c 5 \u6b21\u5224\u65ad\u7684\u7ed3\u679c\u9009\u4e3e\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 9\uff1a\u7528\u4e8e\u68c0\u6d4b\u91cd\u590d\u8bc4\u8bba\u7684 Prompt\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig09.png\" \/><\/p>\n<p><strong>\u56fe 9\uff1a<\/strong> \u7528\u4e8e\u68c0\u6d4b\u91cd\u590d\u8bc4\u8bba\u7684 Prompt<\/p>\n<p>\u968f\u540e\uff0c\u6211\u4eec\u5c06\u8be5\u96c6\u5408\u63d0\u4ea4\u7ed9 80 \u4f59\u540d\u5177\u5907\u4e24\u5e74\u4ee5\u4e0a\u7ecf\u9a8c\u7684\u9ad8\u7ea7\u8f6f\u4ef6\u5de5\u7a0b\u5e08\u8fdb\u884c\u4e25\u683c\u7684\u4eba\u5de5\u6807\u6ce8\u3002\u6807\u6ce8\u8fc7\u7a0b\u8986\u76d6\u4e09\u4e2a\u6838\u5fc3\u7ef4\u5ea6\uff1a\u6838\u9a8c\u8bc4\u5ba1\u8bc4\u8bba\u7684\u6b63\u786e\u6027\u3001\u6309\u8868 7 \u5bf9\u95ee\u9898\u7c7b\u578b\u5206\u7c7b\uff0c\u4ee5\u53ca\u5b9a\u4e49\u4e0a\u4e0b\u6587\u4f9d\u8d56\u8303\u56f4\u3002\u6807\u6ce8\u4eba\u529b\u7531 6 \u4eba\u6838\u5fc3\u4e13\u5bb6\u56e2\u961f\u4e0e\u7531\u5176\u4f59\u53c2\u4e0e\u8005\u7ec4\u6210\u7684\u4e00\u822c\u6807\u6ce8\u6c60\u6784\u6210\u3002\u6807\u6ce8\u8fc7\u7a0b\u5206\u4e3a\u4e09\u8f6e\uff1a\u524d\u4e24\u8f6e\u7531\u4e00\u822c\u6807\u6ce8\u6c60\u4ee5\u53cc\u76f2\u673a\u5236\u8fdb\u884c\uff0c\u6bcf\u6761\u8bc4\u8bba\u7531\u4e24\u540d\u4e0d\u540c\u4eba\u5458\u72ec\u7acb\u6807\u6ce8\uff0c\u4efb\u52a1\u5206\u914d\u4e25\u683c\u5339\u914d\u6807\u6ce8\u8005\u7684\u7f16\u7a0b\u8bed\u8a00\u4e13\u957f\u3002\u7b2c\u4e09\u8f6e\u7531\u6838\u5fc3\u4e13\u5bb6\u56e2\u961f\u8fdb\u884c\uff0c\u8d1f\u8d23\u8ba8\u8bba\u5e76\u88c1\u51b3\u524d\u4e24\u8f6e\u7684\u51b2\u7a81\u7ed3\u679c\u5e76\u786e\u5b9a\u6700\u7ec8\u6807\u6ce8\u3002\u901a\u8fc7\u8fd9\u4e00\u300c\u4eba\u673a\u534f\u540c\u300d\u7684\u591a\u8f6e\u6807\u6ce8\u5de5\u4f5c\u6d41\uff0c\u6211\u4eec\u786e\u4fdd\u6700\u7ec8\u8bc4\u6d4b\u57fa\u51c6\u540c\u65f6\u5177\u5907\u9ad8\u8986\u76d6\u4e0e\u9ad8\u51c6\u786e\u3002<\/p>\n<p>\u4e3a\u8bc4\u4f30\u6240\u9009 LLM \u7684\u4e92\u8865\u6027\uff0c\u6211\u4eec\u5206\u6790\u4e86\u751f\u6210\u8bc4\u5ba1\u8bc4\u8bba\u4e0e\u589e\u5f3a\u8bc4\u5ba1\u8bc4\u8bba\u7684\u91cd\u53e0\u3002\u8868 8 \u5c55\u793a\u4e86\u57fa\u4e8e\u5728\u5df2\u6807\u6ce8 Ground Truth \u4e2d\u6210\u529f\u68c0\u6d4b\u5230\u5b83\u4eec\u7684\u6a21\u578b\u6570\u91cf\u7684\u8bc4\u8bba\u5206\u5e03\u3002<\/p>\n<p><strong>\u8868 8\uff1a<\/strong> \u6309\u6a21\u578b\u68c0\u6d4b\u9891\u6b21\u7684\u8bc4\u5ba1\u8bc4\u8bba\u5206\u5e03<\/p>\n<table>\n<thead>\n<tr>\n<th>\u68c0\u6d4b\u5230\u7684\u6a21\u578b\u6570\u91cf<\/th>\n<th style=\"text-align: right\">\u6570\u91cf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>0\uff08\u4ec5\u589e\u5f3a\uff09<\/td>\n<td style=\"text-align: right\">360<\/td>\n<\/tr>\n<tr>\n<td>1<\/td>\n<td style=\"text-align: right\">1027<\/td>\n<\/tr>\n<tr>\n<td>2<\/td>\n<td style=\"text-align: right\">107<\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td style=\"text-align: right\">11<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6709\u8da3\u7684\u662f\uff0c\u5c3d\u7ba1\u4ec5\u6709 $11$ \u6761\u8bc4\u8bba\u5728 3 \u4e2a\u6a21\u578b\u4e2d\u8fbe\u6210\u5171\u8bc6\uff08\u88ab\u4e09\u4e2a\u6a21\u578b\u68c0\u6d4b\u5230\uff09\uff0c\u76f8\u5f53\u4e00\u90e8\u5206\u4ec5\u88ab\u5355\u4e2a\u6a21\u578b\u8bc6\u522b\u3002\u8fd9\u4e00\u4f4e\u91cd\u53e0\u8868\u660e\uff0c\u6709\u5fc5\u8981\u5728\u8bc4\u5ba1\u8bc4\u8bba\u751f\u6210\u8fc7\u7a0b\u4e2d\u7eb3\u5165\u591a\u6a21\u578b\u7ed3\u679c\uff0c\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u95ee\u9898\u8986\u76d6\u3002<\/p>\n<h3>B.2 Ground Truth \u793a\u4f8b<\/h3>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u7ecf\u6700\u7ec8\u6807\u6ce8\u8fc7\u7a0b\u6838\u9a8c\u4e3a\u6709\u6548\u3001\u4ece\u800c\u4f5c\u4e3a AACR-Bench Ground Truth \u7684\u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b\u3002\u5bf9\u6bcf\u4e2a\u793a\u4f8b\uff0c\u6211\u4eec\u63d0\u4f9b\u4ed3\u5e93\u540d\u79f0\u3001Pull Request\uff08PR\uff09ID\u3001\u76ee\u6807\u6587\u4ef6\u3001\u6240\u9488\u5bf9\u7684\u7279\u5b9a Diff Hunk\u3001\u8bc4\u5ba1\u8bc4\u8bba\u5185\u5bb9\uff0c\u4ee5\u53ca\u76f8\u5e94\u5206\u6790\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 10\uff1aGround Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig10.png\" \/><\/p>\n<p><strong>\u56fe 10\uff1a<\/strong> Ground Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 11\uff1aGround Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig11.png\" \/><\/p>\n<p><strong>\u56fe 11\uff1a<\/strong> Ground Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 12\uff1aGround Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig12.png\" \/><\/p>\n<p><strong>\u56fe 12\uff1a<\/strong> Ground Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 13\uff1aGround Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig13.png\" \/><\/p>\n<p><strong>\u56fe 13\uff1a<\/strong> Ground Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 14\uff1aGround Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig14.png\" \/><\/p>\n<p><strong>\u56fe 14\uff1a<\/strong> Ground Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 15\uff1aGround Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig15.png\" \/><\/p>\n<p><strong>\u56fe 15\uff1a<\/strong> Ground Truth \u8bc4\u5ba1\u8bc4\u8bba\u793a\u4f8b<\/p>\n<h2>\u9644\u5f55 C \u5b9e\u9a8c\u7ec6\u8282<\/h2>\n<h3>C.1 \u8be6\u7ec6\u57fa\u51c6\u6d41\u7a0b<\/h3>\n<p>AACR-Bench \u7684\u8bc4\u6d4b\u8fc7\u7a0b\u4e0e\u73b0\u4ee3\u4ee3\u7801\u8bc4\u5ba1\u4ee5 Diff \u4e3a\u5bfc\u5411\u7684\u8303\u5f0f\u5bf9\u9f50\u3002\u5728\u57fa\u51c6\u4e2d\uff0c\u6bcf\u4e2a Pull Request\uff08PR\uff09\u4f5c\u4e3a\u4e00\u4e2a\u8bc4\u6d4b\u5b9e\u4f8b\uff0c\u4e0e\u8be5 PR \u5173\u8054\u7684\u8bc4\u5ba1\u8bc4\u8bba\u96c6\u5408\u6784\u6210 Ground Truth\u3002\u5bf9\u6bcf\u4e2a PR\uff0c\u88ab\u8bc4\u6d4b\u7684\u81ea\u52a8\u5316\u4ee3\u7801\u8bc4\u5ba1\uff08ACR\uff09\u65b9\u6cd5\u9700\u8981\u904d\u5386\u4ee3\u7801\u53d8\u66f4\u4e2d\u7684 Diff Hunk\uff0c\u5e76\u4e3a\u6bcf\u4e2a hunk \u751f\u6210\u8bc4\u5ba1\u8bc4\u8bba\u3002\u968f\u540e\u901a\u8fc7\u5c06\u751f\u6210\u8bc4\u8bba\u4e0e Ground Truth \u6bd4\u8f83\u4ee5\u8ba1\u7b97 Precision\u3001Recall \u4e0e F1-score\uff0c\u6765\u8861\u91cf ACR \u65b9\u6cd5\u7684\u4ee3\u7801\u8bc4\u5ba1\u80fd\u529b\u3002\u8bc4\u5ba1\u8bc4\u8bba\u5339\u914d\u6240\u7528 prompt \u4e0e\u56fe 9 \u76f8\u540c\uff0c\u6240\u7528\u6a21\u578b\u4e3a Qwen3-235B-A22B-Instruct-2507\u3002\u82e5\u4e00\u6761\u751f\u6210\u8bc4\u5ba1\u8bc4\u8bba\u7684\u884c\u8303\u56f4\u4e0e Ground Truth \u4e2d\u67d0\u6761\u8bc4\u5ba1\u8bc4\u8bba\u91cd\u53e0\uff0c\u5219\u89c6\u4e3a\u884c\u6b63\u786e\uff08Line Correct\uff09\u3002\u82e5\u5b83\u5728\u884c\u53f7\u4e0a\u4e0e Ground Truth \u91cd\u53e0\u4e14\u5185\u5bb9\u76f8\u540c\uff0c\u5219\u89c6\u4e3a\u8bed\u4e49\u6b63\u786e\uff08Semantically Correct\uff09\u3002\u6240\u6709\u6a21\u578b\u4ee5 $mathrm{Temperature}=0.7$\u3001$mathrm{Top}<em>{p}=0.95$\u3001$mathrm{Top}<\/em>=20$ \u8fd0\u884c\u3002<\/p>\n<h5>\u975e Agent \u65b9\u6cd5\u7684\u8bc4\u6d4b<\/h5>\n<p>\u975e Agent \u65b9\u6cd5\u7684\u8bc4\u6d4b\u9075\u5faa\u6807\u51c6\u5316\u5de5\u4f5c\u6d41\u3002\u5bf9\u57fa\u51c6\u4e2d\u7684\u6bcf\u4e2a PR\uff0c\u6211\u4eec\u9996\u5148\u5728\u672c\u5730\u514b\u9686\u4ed3\u5e93\uff0c\u5e76\u786e\u4fdd Base \u7248\u672c\u4e0e Target \u7248\u672c\uff08\u5373\u88ab\u8bc4\u5ba1\u7248\u672c\uff09\u5747\u5df2\u540c\u6b65\u3002\u7136\u540e\u6211\u4eec\u4f7f\u7528 GitPython \u6838\u9a8c\u5e76\u63d0\u53d6\u4e24\u4e2a\u7248\u672c\u4e4b\u95f4\u7684\u5168\u90e8 diff hunk\u3002\u88ab\u8bc4\u6d4b\u7684 ACR \u65b9\u6cd5\u4f5c\u4e3a\u626b\u63cf\u5668\uff0c\u904d\u5386\u6bcf\u4e2a diff hunk \u4ee5\u751f\u6210\u8bc4\u5ba1\u8bc4\u8bba\u3002\u5bf9\u7eb3\u5165\u4e0a\u4e0b\u6587\u68c0\u7d22\u7684\u65b9\u6cd5\uff0c\u6211\u4eec\u57fa\u4e8e Target \u7248\u672c\u4e2d\u7684\u4ee3\u7801\u6784\u5efa\u7d22\u5f15\u4ee5\u4fc3\u8fdb\u4e0a\u4e0b\u6587\u53ec\u56de\u3002\u5728\u8bc4\u5ba1\u6bcf\u4e2a Diff Hunk \u671f\u95f4\uff0c\u6211\u4eec\u59cb\u7ec8\u63d0\u4f9b PR Title \u4e0e PR Description \u4f5c\u4e3a\u4ed3\u5e93\u7ea7\u4e0a\u4e0b\u6587\u3002\u6b64\u5916\uff0c\u4ed3\u5e93\u4ee3\u7801\u4e0a\u4e0b\u6587\u6309\u5177\u4f53\u5b9e\u9a8c\u8bbe\u8ba1\u53ef\u9009\u63d0\u4f9b\u3002\u8bc4\u6d4b\u6240\u7528 prompt \u89c1\u56fe 16 \u4e0e\u56fe 17\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5bf9\u6bcf\u6761\u8bc4\u5ba1\u8bc4\u8bba\uff0c\u6a21\u578b\u9700\u8981\u8f93\u51fa\u4e09\u4e2a\u7ec4\u6210\u90e8\u5206\uff1a<\/p>\n<ul>\n<li><strong>\u6709\u95ee\u9898\u7684 diff \u7247\u6bb5\uff1a<\/strong> \u7528\u4e8e\u8bc4\u8bba\u7684\u7cbe\u786e\u5b9a\u4f4d\uff1b<\/li>\n<li><strong>\u8bc4\u8bba\u4fa7\uff1a<\/strong> \u6307\u660e\u8bc4\u8bba\u9002\u7528\u4e8e\u300cleft\u300d\uff08\u65e7\u7248\u672c\u4e2d\u7684\u5220\u9664\u884c\uff09\u8fd8\u662f\u300cright\u300d\uff08\u65b0\u7248\u672c\u4e2d\u7684\u65b0\u589e\u884c\uff09\uff1b<\/li>\n<li><strong>\u8bc4\u5ba1\u8bc4\u8bba\u7684\u5185\u5bb9\u3002<\/strong><\/li>\n<\/ul>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 16\uff1a\u65e0\u4ee3\u7801\u4e0a\u4e0b\u6587\u65f6\u975e Agent \u65b9\u6cd5 ACR \u4efb\u52a1\u7684 Prompt\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig16.png\" \/><\/p>\n<p><strong>\u56fe 16\uff1a<\/strong> \u65e0\u4ee3\u7801\u4e0a\u4e0b\u6587\u65f6\u975e Agent \u65b9\u6cd5 ACR \u4efb\u52a1\u7684 Prompt<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 17\uff1a\u6709\u4ee3\u7801\u4e0a\u4e0b\u6587\u65f6\u975e Agent \u65b9\u6cd5 ACR \u4efb\u52a1\u7684 Prompt\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig17.png\" \/><\/p>\n<p><strong>\u56fe 17\uff1a<\/strong> \u6709\u4ee3\u7801\u4e0a\u4e0b\u6587\u65f6\u975e Agent \u65b9\u6cd5 ACR \u4efb\u52a1\u7684 Prompt<\/p>\n<h5>Agent \u65b9\u6cd5\u7684\u8bc4\u6d4b<\/h5>\n<p>\u6211\u4eec\u9009\u62e9 Claude Code \u4f5c\u4e3a\u8bc4\u6d4b\u57fa\u4e8e Agent \u65b9\u6cd5\u7684\u6846\u67b6\u3002\u501f\u52a9\u5176\u81ea\u4e3b\u5de5\u5177\u4f7f\u7528\u80fd\u529b\uff0cClaude Code \u53ef\u4ee5\u72ec\u7acb\u8c03\u7528 Git \u547d\u4ee4\u4ee5\u63d0\u53d6 Base \u4e0e Target \u7248\u672c\u4e4b\u95f4\u7684 Diff Hunk\uff0c\u5e76\u6267\u884c\u5b8c\u5168\u81ea\u4e3b\u7684\u4e0a\u4e0b\u6587\u68c0\u7d22\u3002\u56e0\u6b64\uff0c\u8bc4\u6d4b\u8bbe\u7f6e\u88ab\u7b80\u5316\uff1a\u6211\u4eec\u901a\u5e38\u5728\u672c\u5730\u514b\u9686\u4ed3\u5e93\u5e76 checkout \u5bf9\u5e94\u7684 PR \u7248\u672c\uff0c\u4ece\u800c\u65e0\u9700\u9884\u5148\u6784\u5efa\u4ee3\u7801\u7d22\u5f15\u3002\u672c\u8bc4\u6d4b\u6240\u7528\u4ee3\u7801\u8bc4\u5ba1 Agent \u7684\u5b9a\u4e49\u89c1\u56fe 18\uff0c\u89e6\u53d1\u8bc4\u5ba1\u8fc7\u7a0b\u6240\u7528\u7684\u5177\u4f53 prompt \u89c1\u56fe 19\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 18\uff1aClaude Code \u7684 Agent \u5b9a\u4e49\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig18.png\" \/><\/p>\n<p><strong>\u56fe 18\uff1a<\/strong> Claude Code \u7684 Agent \u5b9a\u4e49<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 19\uff1aClaude Code \u7684 Agent Prompt\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig19.png\" \/><\/p>\n<p><strong>\u56fe 19\uff1a<\/strong> Claude Code \u7684 Agent Prompt<\/p>\n<p>\u4e3a\u8bc4\u4f30\u6a21\u578b\u68c0\u6d4b\u4e0d\u540c\u7c7b\u522b\u95ee\u9898\u7684\u51c6\u786e\u7387\uff0c\u6211\u4eec\u5bf9 ACR \u65b9\u6cd5\u751f\u6210\u7684\u8bc4\u5ba1\u8bc4\u8bba\u8fdb\u884c\u4e86\u5206\u7c7b\u3002\u8be5\u5206\u7c7b\u4f7f\u7528 Qwen3-235B-A22B-Instruct-2507 \u6267\u884c\uff0c\u6240\u7528 prompt \u89c1\u56fe 20\u3002\u6211\u4eec\u5728\u57fa\u51c6\u7684 Ground Truth \u6570\u636e\u96c6\u4e0a\u6838\u9a8c\u4e86\u8be5 prompt \u7684\u6548\u529b\uff0c\u8fbe\u5230 $97%$ \u7684\u51c6\u786e\u7387\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 20\uff1a\u5bf9 ACR \u65b9\u6cd5\u6240\u53d1\u73b0\u95ee\u9898\u8fdb\u884c\u5206\u7c7b\u7684 Prompt\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig20.png\" \/><\/p>\n<p><strong>\u56fe 20\uff1a<\/strong> \u5bf9 ACR \u65b9\u6cd5\u6240\u53d1\u73b0\u95ee\u9898\u8fdb\u884c\u5206\u7c7b\u7684 Prompt<\/p>\n<h3>C.2 \u9644\u52a0\u7edf\u8ba1<\/h3>\n<p><strong>\u8868 9\uff1a<\/strong> \u4e0d\u540c\u7c7b\u578b\u4e0b\u5404\u6a21\u578b\u7684\u8be6\u7ec6\u6027\u80fd\u6bd4\u8f83\uff08%\uff09<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>\u7c7b\u578b<\/th>\n<th style=\"text-align: right\">\u5e73\u5747\u8bc4\u8bba\u6570\uff08\u6bcf Patch\uff09<\/th>\n<th style=\"text-align: right\">Line Recall<\/th>\n<th style=\"text-align: right\">Semantic Recall<\/th>\n<th style=\"text-align: right\">Line Precision<\/th>\n<th style=\"text-align: right\">Semantic Precision<\/th>\n<th style=\"text-align: right\">Line F1<\/th>\n<th style=\"text-align: right\">Semantic F1<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">0.08<\/td>\n<td style=\"text-align: right\">13.16<\/td>\n<td style=\"text-align: right\">10.10<\/td>\n<td style=\"text-align: right\">52.00<\/td>\n<td style=\"text-align: right\">39.90<\/td>\n<td style=\"text-align: right\">21.00<\/td>\n<td style=\"text-align: right\">16.12<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">0.15<\/td>\n<td style=\"text-align: right\">14.15<\/td>\n<td style=\"text-align: right\">4.78<\/td>\n<td style=\"text-align: right\">32.60<\/td>\n<td style=\"text-align: right\">11.00<\/td>\n<td style=\"text-align: right\">19.74<\/td>\n<td style=\"text-align: right\">6.67<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">0.14<\/td>\n<td style=\"text-align: right\">9.63<\/td>\n<td style=\"text-align: right\">4.72<\/td>\n<td style=\"text-align: right\">23.40<\/td>\n<td style=\"text-align: right\">11.50<\/td>\n<td style=\"text-align: right\">13.65<\/td>\n<td style=\"text-align: right\">6.69<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">0.11<\/td>\n<td style=\"text-align: right\">5.51<\/td>\n<td style=\"text-align: right\">2.99<\/td>\n<td style=\"text-align: right\">18.20<\/td>\n<td style=\"text-align: right\">9.90<\/td>\n<td style=\"text-align: right\">8.46<\/td>\n<td style=\"text-align: right\">4.59<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">0.10<\/td>\n<td style=\"text-align: right\">10.63<\/td>\n<td style=\"text-align: right\">4.39<\/td>\n<td style=\"text-align: right\">37.20<\/td>\n<td style=\"text-align: right\">15.30<\/td>\n<td style=\"text-align: right\">16.54<\/td>\n<td style=\"text-align: right\">6.82<\/td>\n<\/tr>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">1.72<\/td>\n<td style=\"text-align: right\">73.29<\/td>\n<td style=\"text-align: right\">42.86<\/td>\n<td style=\"text-align: right\">15.00<\/td>\n<td style=\"text-align: right\">8.70<\/td>\n<td style=\"text-align: right\">24.90<\/td>\n<td style=\"text-align: right\">14.46<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">2.29<\/td>\n<td style=\"text-align: right\">72.76<\/td>\n<td style=\"text-align: right\">36.54<\/td>\n<td style=\"text-align: right\">11.20<\/td>\n<td style=\"text-align: right\">5.60<\/td>\n<td style=\"text-align: right\">19.41<\/td>\n<td style=\"text-align: right\">9.71<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">0.85<\/td>\n<td style=\"text-align: right\">50.23<\/td>\n<td style=\"text-align: right\">27.57<\/td>\n<td style=\"text-align: right\">20.60<\/td>\n<td style=\"text-align: right\">11.30<\/td>\n<td style=\"text-align: right\">29.22<\/td>\n<td style=\"text-align: right\">16.03<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">2.35<\/td>\n<td style=\"text-align: right\">73.89<\/td>\n<td style=\"text-align: right\">47.11<\/td>\n<td style=\"text-align: right\">11.00<\/td>\n<td style=\"text-align: right\">7.00<\/td>\n<td style=\"text-align: right\">19.15<\/td>\n<td style=\"text-align: right\">12.19<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">1.02<\/td>\n<td style=\"text-align: right\">58.34<\/td>\n<td style=\"text-align: right\">27.44<\/td>\n<td style=\"text-align: right\">20.10<\/td>\n<td style=\"text-align: right\">9.40<\/td>\n<td style=\"text-align: right\">29.90<\/td>\n<td style=\"text-align: right\">14.00<\/td>\n<\/tr>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">2.17<\/td>\n<td style=\"text-align: right\">72.56<\/td>\n<td style=\"text-align: right\">35.75<\/td>\n<td style=\"text-align: right\">11.80<\/td>\n<td style=\"text-align: right\">5.80<\/td>\n<td style=\"text-align: right\">20.30<\/td>\n<td style=\"text-align: right\">9.98<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">0.89<\/td>\n<td style=\"text-align: right\">51.69<\/td>\n<td style=\"text-align: right\">27.38<\/td>\n<td style=\"text-align: right\">20.50<\/td>\n<td style=\"text-align: right\">10.90<\/td>\n<td style=\"text-align: right\">29.36<\/td>\n<td style=\"text-align: right\">15.59<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">0.91<\/td>\n<td style=\"text-align: right\">53.62<\/td>\n<td style=\"text-align: right\">26.25<\/td>\n<td style=\"text-align: right\">20.80<\/td>\n<td style=\"text-align: right\">10.20<\/td>\n<td style=\"text-align: right\">29.97<\/td>\n<td style=\"text-align: right\">14.69<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">1.75<\/td>\n<td style=\"text-align: right\">74.55<\/td>\n<td style=\"text-align: right\">43.59<\/td>\n<td style=\"text-align: right\">15.00<\/td>\n<td style=\"text-align: right\">8.80<\/td>\n<td style=\"text-align: right\">24.97<\/td>\n<td style=\"text-align: right\">14.64<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">2.39<\/td>\n<td style=\"text-align: right\">72.49<\/td>\n<td style=\"text-align: right\">45.85<\/td>\n<td style=\"text-align: right\">10.70<\/td>\n<td style=\"text-align: right\">6.70<\/td>\n<td style=\"text-align: right\">18.65<\/td>\n<td style=\"text-align: right\">11.69<\/td>\n<\/tr>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">1.89<\/td>\n<td style=\"text-align: right\">73.16<\/td>\n<td style=\"text-align: right\">42.86<\/td>\n<td style=\"text-align: right\">13.60<\/td>\n<td style=\"text-align: right\">8.00<\/td>\n<td style=\"text-align: right\">22.94<\/td>\n<td style=\"text-align: right\">13.48<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">2.52<\/td>\n<td style=\"text-align: right\">72.56<\/td>\n<td style=\"text-align: right\">36.35<\/td>\n<td style=\"text-align: right\">10.10<\/td>\n<td style=\"text-align: right\">5.10<\/td>\n<td style=\"text-align: right\">17.73<\/td>\n<td style=\"text-align: right\">8.94<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">0.87<\/td>\n<td style=\"text-align: right\">49.63<\/td>\n<td style=\"text-align: right\">26.98<\/td>\n<td style=\"text-align: right\">20.20<\/td>\n<td style=\"text-align: right\">11.00<\/td>\n<td style=\"text-align: right\">28.71<\/td>\n<td style=\"text-align: right\">15.63<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">2.47<\/td>\n<td style=\"text-align: right\">72.49<\/td>\n<td style=\"text-align: right\">47.24<\/td>\n<td style=\"text-align: right\">10.30<\/td>\n<td style=\"text-align: right\">6.70<\/td>\n<td style=\"text-align: right\">18.04<\/td>\n<td style=\"text-align: right\">11.74<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">0.83<\/td>\n<td style=\"text-align: right\">49.50<\/td>\n<td style=\"text-align: right\">24.25<\/td>\n<td style=\"text-align: right\">20.90<\/td>\n<td style=\"text-align: right\">10.20<\/td>\n<td style=\"text-align: right\">29.39<\/td>\n<td style=\"text-align: right\">14.36<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u8868 10\uff1a<\/strong> \u4e0d\u540c\u6a21\u578b\u4e0e\u65b9\u6cd5\u5728\u56db\u7c7b\u95ee\u9898\u4e0a\u7684\u6027\u80fd\u6bd4\u8f83\uff08%\uff09<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>\u65b9\u6cd5<\/th>\n<th style=\"text-align: right\">Security Vuln. Rec<\/th>\n<th style=\"text-align: right\">Prec<\/th>\n<th style=\"text-align: right\">F1<\/th>\n<th style=\"text-align: right\">Code Defect Rec<\/th>\n<th style=\"text-align: right\">Prec<\/th>\n<th style=\"text-align: right\">F1<\/th>\n<th style=\"text-align: right\">Performance Rec<\/th>\n<th style=\"text-align: right\">Prec<\/th>\n<th style=\"text-align: right\">F1<\/th>\n<th style=\"text-align: right\">Maint. &amp; Read. Rec<\/th>\n<th style=\"text-align: right\">Prec<\/th>\n<th style=\"text-align: right\">F1<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">11.76<\/td>\n<td style=\"text-align: right\">17.39<\/td>\n<td style=\"text-align: right\">14.04<\/td>\n<td style=\"text-align: right\">17.07<\/td>\n<td style=\"text-align: right\">44.02<\/td>\n<td style=\"text-align: right\">24.60<\/td>\n<td style=\"text-align: right\">9.28<\/td>\n<td style=\"text-align: right\">36.00<\/td>\n<td style=\"text-align: right\">14.75<\/td>\n<td style=\"text-align: right\">11.30<\/td>\n<td style=\"text-align: right\">37.60<\/td>\n<td style=\"text-align: right\">17.38<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">2.08<\/td>\n<td style=\"text-align: right\">2.17<\/td>\n<td style=\"text-align: right\">2.13<\/td>\n<td style=\"text-align: right\">6.71<\/td>\n<td style=\"text-align: right\">13.65<\/td>\n<td style=\"text-align: right\">8.99<\/td>\n<td style=\"text-align: right\">2.65<\/td>\n<td style=\"text-align: right\">5.88<\/td>\n<td style=\"text-align: right\">3.66<\/td>\n<td style=\"text-align: right\">3.68<\/td>\n<td style=\"text-align: right\">9.69<\/td>\n<td style=\"text-align: right\">5.33<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">4.88<\/td>\n<td style=\"text-align: right\">4.08<\/td>\n<td style=\"text-align: right\">4.44<\/td>\n<td style=\"text-align: right\">6.72<\/td>\n<td style=\"text-align: right\">14.29<\/td>\n<td style=\"text-align: right\">9.14<\/td>\n<td style=\"text-align: right\">8.42<\/td>\n<td style=\"text-align: right\">20.00<\/td>\n<td style=\"text-align: right\">11.85<\/td>\n<td style=\"text-align: right\">4.67<\/td>\n<td style=\"text-align: right\">8.43<\/td>\n<td style=\"text-align: right\">6.01<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">4.00<\/td>\n<td style=\"text-align: right\">7.14<\/td>\n<td style=\"text-align: right\">5.13<\/td>\n<td style=\"text-align: right\">3.40<\/td>\n<td style=\"text-align: right\">8.36<\/td>\n<td style=\"text-align: right\">4.84<\/td>\n<td style=\"text-align: right\">2.59<\/td>\n<td style=\"text-align: right\">17.65<\/td>\n<td style=\"text-align: right\">4.51<\/td>\n<td style=\"text-align: right\">2.56<\/td>\n<td style=\"text-align: right\">12.31<\/td>\n<td style=\"text-align: right\">4.24<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>Agent<\/td>\n<td style=\"text-align: right\">4.17<\/td>\n<td style=\"text-align: right\">13.33<\/td>\n<td style=\"text-align: right\">6.35<\/td>\n<td style=\"text-align: right\">6.46<\/td>\n<td style=\"text-align: right\">18.22<\/td>\n<td style=\"text-align: right\">9.53<\/td>\n<td style=\"text-align: right\">4.00<\/td>\n<td style=\"text-align: right\">16.67<\/td>\n<td style=\"text-align: right\">6.45<\/td>\n<td style=\"text-align: right\">3.49<\/td>\n<td style=\"text-align: right\">11.31<\/td>\n<td style=\"text-align: right\">5.33<\/td>\n<\/tr>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">45.28<\/td>\n<td style=\"text-align: right\">4.81<\/td>\n<td style=\"text-align: right\">8.70<\/td>\n<td style=\"text-align: right\">46.40<\/td>\n<td style=\"text-align: right\">15.63<\/td>\n<td style=\"text-align: right\">23.38<\/td>\n<td style=\"text-align: right\">53.85<\/td>\n<td style=\"text-align: right\">14.13<\/td>\n<td style=\"text-align: right\">22.38<\/td>\n<td style=\"text-align: right\">36.58<\/td>\n<td style=\"text-align: right\">5.30<\/td>\n<td style=\"text-align: right\">9.26<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">28.30<\/td>\n<td style=\"text-align: right\">3.45<\/td>\n<td style=\"text-align: right\">6.15<\/td>\n<td style=\"text-align: right\">39.63<\/td>\n<td style=\"text-align: right\">10.35<\/td>\n<td style=\"text-align: right\">16.42<\/td>\n<td style=\"text-align: right\">43.59<\/td>\n<td style=\"text-align: right\">11.02<\/td>\n<td style=\"text-align: right\">17.59<\/td>\n<td style=\"text-align: right\">32.43<\/td>\n<td style=\"text-align: right\">3.28<\/td>\n<td style=\"text-align: right\">5.96<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">37.74<\/td>\n<td style=\"text-align: right\">9.57<\/td>\n<td style=\"text-align: right\">15.27<\/td>\n<td style=\"text-align: right\">28.77<\/td>\n<td style=\"text-align: right\">17.82<\/td>\n<td style=\"text-align: right\">22.01<\/td>\n<td style=\"text-align: right\">35.04<\/td>\n<td style=\"text-align: right\">20.71<\/td>\n<td style=\"text-align: right\">26.03<\/td>\n<td style=\"text-align: right\">23.96<\/td>\n<td style=\"text-align: right\">7.08<\/td>\n<td style=\"text-align: right\">10.93<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">56.60<\/td>\n<td style=\"text-align: right\">2.92<\/td>\n<td style=\"text-align: right\">5.56<\/td>\n<td style=\"text-align: right\">50.78<\/td>\n<td style=\"text-align: right\">13.92<\/td>\n<td style=\"text-align: right\">21.85<\/td>\n<td style=\"text-align: right\">47.86<\/td>\n<td style=\"text-align: right\">9.82<\/td>\n<td style=\"text-align: right\">16.30<\/td>\n<td style=\"text-align: right\">42.01<\/td>\n<td style=\"text-align: right\">4.46<\/td>\n<td style=\"text-align: right\">8.07<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>No context<\/td>\n<td style=\"text-align: right\">26.42<\/td>\n<td style=\"text-align: right\">7.87<\/td>\n<td style=\"text-align: right\">12.12<\/td>\n<td style=\"text-align: right\">29.90<\/td>\n<td style=\"text-align: right\">16.99<\/td>\n<td style=\"text-align: right\">21.67<\/td>\n<td style=\"text-align: right\">30.77<\/td>\n<td style=\"text-align: right\">14.17<\/td>\n<td style=\"text-align: right\">19.41<\/td>\n<td style=\"text-align: right\">24.12<\/td>\n<td style=\"text-align: right\">5.61<\/td>\n<td style=\"text-align: right\">9.10<\/td>\n<\/tr>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">47.17<\/td>\n<td style=\"text-align: right\">4.84<\/td>\n<td style=\"text-align: right\">8.77<\/td>\n<td style=\"text-align: right\">48.38<\/td>\n<td style=\"text-align: right\">15.92<\/td>\n<td style=\"text-align: right\">23.95<\/td>\n<td style=\"text-align: right\">48.72<\/td>\n<td style=\"text-align: right\">12.28<\/td>\n<td style=\"text-align: right\">19.62<\/td>\n<td style=\"text-align: right\">36.90<\/td>\n<td style=\"text-align: right\">5.32<\/td>\n<td style=\"text-align: right\">9.31<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">33.96<\/td>\n<td style=\"text-align: right\">3.97<\/td>\n<td style=\"text-align: right\">7.11<\/td>\n<td style=\"text-align: right\">39.49<\/td>\n<td style=\"text-align: right\">10.80<\/td>\n<td style=\"text-align: right\">16.96<\/td>\n<td style=\"text-align: right\">36.75<\/td>\n<td style=\"text-align: right\">9.39<\/td>\n<td style=\"text-align: right\">14.96<\/td>\n<td style=\"text-align: right\">31.47<\/td>\n<td style=\"text-align: right\">3.41<\/td>\n<td style=\"text-align: right\">6.15<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">30.19<\/td>\n<td style=\"text-align: right\">7.08<\/td>\n<td style=\"text-align: right\">11.47<\/td>\n<td style=\"text-align: right\">29.20<\/td>\n<td style=\"text-align: right\">18.00<\/td>\n<td style=\"text-align: right\">22.27<\/td>\n<td style=\"text-align: right\">30.77<\/td>\n<td style=\"text-align: right\">18.95<\/td>\n<td style=\"text-align: right\">23.45<\/td>\n<td style=\"text-align: right\">24.44<\/td>\n<td style=\"text-align: right\">6.88<\/td>\n<td style=\"text-align: right\">10.74<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">56.60<\/td>\n<td style=\"text-align: right\">2.97<\/td>\n<td style=\"text-align: right\">5.64<\/td>\n<td style=\"text-align: right\">49.08<\/td>\n<td style=\"text-align: right\">13.27<\/td>\n<td style=\"text-align: right\">20.89<\/td>\n<td style=\"text-align: right\">47.01<\/td>\n<td style=\"text-align: right\">9.79<\/td>\n<td style=\"text-align: right\">16.20<\/td>\n<td style=\"text-align: right\">41.05<\/td>\n<td style=\"text-align: right\">4.26<\/td>\n<td style=\"text-align: right\">7.71<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>BM25<\/td>\n<td style=\"text-align: right\">20.75<\/td>\n<td style=\"text-align: right\">6.92<\/td>\n<td style=\"text-align: right\">10.38<\/td>\n<td style=\"text-align: right\">28.21<\/td>\n<td style=\"text-align: right\">18.94<\/td>\n<td style=\"text-align: right\">22.66<\/td>\n<td style=\"text-align: right\">24.79<\/td>\n<td style=\"text-align: right\">13.62<\/td>\n<td style=\"text-align: right\">17.58<\/td>\n<td style=\"text-align: right\">24.76<\/td>\n<td style=\"text-align: right\">6.31<\/td>\n<td style=\"text-align: right\">10.06<\/td>\n<\/tr>\n<tr>\n<td>Claude-4.5-Sonnet<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">45.28<\/td>\n<td style=\"text-align: right\">4.45<\/td>\n<td style=\"text-align: right\">8.11<\/td>\n<td style=\"text-align: right\">46.83<\/td>\n<td style=\"text-align: right\">14.36<\/td>\n<td style=\"text-align: right\">21.98<\/td>\n<td style=\"text-align: right\">48.72<\/td>\n<td style=\"text-align: right\">10.65<\/td>\n<td style=\"text-align: right\">17.48<\/td>\n<td style=\"text-align: right\">37.06<\/td>\n<td style=\"text-align: right\">4.93<\/td>\n<td style=\"text-align: right\">8.70<\/td>\n<\/tr>\n<tr>\n<td>Deepseek-V3.2<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">32.08<\/td>\n<td style=\"text-align: right\">3.62<\/td>\n<td style=\"text-align: right\">6.50<\/td>\n<td style=\"text-align: right\">40.48<\/td>\n<td style=\"text-align: right\">9.99<\/td>\n<td style=\"text-align: right\">16.02<\/td>\n<td style=\"text-align: right\">42.74<\/td>\n<td style=\"text-align: right\">8.40<\/td>\n<td style=\"text-align: right\">14.04<\/td>\n<td style=\"text-align: right\">30.83<\/td>\n<td style=\"text-align: right\">2.81<\/td>\n<td style=\"text-align: right\">5.15<\/td>\n<\/tr>\n<tr>\n<td>GLM-4.7<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">30.19<\/td>\n<td style=\"text-align: right\">7.51<\/td>\n<td style=\"text-align: right\">12.03<\/td>\n<td style=\"text-align: right\">29.06<\/td>\n<td style=\"text-align: right\">17.04<\/td>\n<td style=\"text-align: right\">21.48<\/td>\n<td style=\"text-align: right\">33.33<\/td>\n<td style=\"text-align: right\">17.73<\/td>\n<td style=\"text-align: right\">23.15<\/td>\n<td style=\"text-align: right\">23.16<\/td>\n<td style=\"text-align: right\">7.03<\/td>\n<td style=\"text-align: right\">10.78<\/td>\n<\/tr>\n<tr>\n<td>GPT-5.2<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">58.49<\/td>\n<td style=\"text-align: right\">2.93<\/td>\n<td style=\"text-align: right\">5.58<\/td>\n<td style=\"text-align: right\">51.62<\/td>\n<td style=\"text-align: right\">13.43<\/td>\n<td style=\"text-align: right\">21.31<\/td>\n<td style=\"text-align: right\">48.72<\/td>\n<td style=\"text-align: right\">8.85<\/td>\n<td style=\"text-align: right\">14.98<\/td>\n<td style=\"text-align: right\">41.05<\/td>\n<td style=\"text-align: right\">4.18<\/td>\n<td style=\"text-align: right\">7.59<\/td>\n<\/tr>\n<tr>\n<td>Qwen-480B-Coder<\/td>\n<td>Embedding<\/td>\n<td style=\"text-align: right\">22.64<\/td>\n<td style=\"text-align: right\">10.26<\/td>\n<td style=\"text-align: right\">14.12<\/td>\n<td style=\"text-align: right\">26.52<\/td>\n<td style=\"text-align: right\">18.23<\/td>\n<td style=\"text-align: right\">21.61<\/td>\n<td style=\"text-align: right\">27.35<\/td>\n<td style=\"text-align: right\">16.75<\/td>\n<td style=\"text-align: right\">20.78<\/td>\n<td style=\"text-align: right\">21.25<\/td>\n<td style=\"text-align: right\">5.97<\/td>\n<td style=\"text-align: right\">9.32<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8868 9 \u4e0e\u8868 10 \u7ed9\u51fa\u5b9e\u9a8c\u4e2d\u7684\u9644\u52a0\u7edf\u8ba1\u3002<\/p>\n<h2>\u9644\u5f55 D \u6848\u4f8b\u7814\u7a76<\/h2>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u6a21\u578b\u5728\u57fa\u4e8e Agent \u4e0e\u57fa\u4e8e\u76f8\u4f3c\u5ea6\u68c0\u7d22\u8bbe\u7f6e\u4e0b\u751f\u6210\u7684\u6b63\u786e\u4e0e\u9519\u8bef\u8bc4\u5ba1\u8bc4\u8bba\u7684\u6848\u4f8b\u7814\u7a76\u3002\u5bf9\u6bcf\u4e2a\u6848\u4f8b\uff0c\u6211\u4eec\u63d0\u4f9b Repo\u3001PR ID\u3001File Path\uff0c\u4ee5\u53ca\u88ab\u8bc4\u5ba1\u6587\u4ef6\u4e2d\u7684 Diff Hunk\u3001\u751f\u6210\u7684\u8bc4\u5ba1\u8bc4\u8bba\u3001\u68c0\u7d22\u5230\u7684\u4ee3\u7801\u4e0a\u4e0b\u6587\uff08\u5982\u9002\u7528\uff09\uff0c\u4ee5\u53ca\u5bf9\u9519\u8bef\u8bc4\u8bba\u7684\u5206\u6790\u3002\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u5f53\u524d\u6a21\u578b\u5728\u4ee3\u7801\u8bc4\u5ba1\u4e2d\u4ecd\u6301\u7eed\u906d\u53d7\u77e5\u8bc6\u9519\u8bef\u3002\u6b64\u5916\uff0c\u4e0a\u4e0b\u6587\u68c0\u7d22\u5f15\u5165\u7684\u566a\u58f0\u6570\u636e\u88ab\u8bc6\u522b\u4e3a\u5bfc\u81f4\u751f\u6210\u9519\u8bef\u8bc4\u5ba1\u8bc4\u8bba\u7684\u91cd\u8981\u56e0\u7d20\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 21\uff1a\u4f7f\u7528 Agent \u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig21.png\" \/><\/p>\n<p><strong>\u56fe 21\uff1a<\/strong> \u4f7f\u7528 Agent \u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 22\uff1a\u4f7f\u7528 Agent \u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig22.png\" \/><\/p>\n<p><strong>\u56fe 22\uff1a<\/strong> \u4f7f\u7528 Agent \u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 23\uff1a\u4f7f\u7528 Agent \u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig23.png\" \/><\/p>\n<p><strong>\u56fe 23\uff1a<\/strong> \u4f7f\u7528 Agent \u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 24\uff1a\u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig24.png\" \/><\/p>\n<p><strong>\u56fe 24\uff1a<\/strong> \u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u6b63\u786e\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 25\uff1a\u4f7f\u7528 Agent \u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig25.png\" \/><\/p>\n<p><strong>\u56fe 25\uff1a<\/strong> \u4f7f\u7528 Agent \u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 26\uff1a\u4f7f\u7528 Agent \u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig26.png\" \/><\/p>\n<p><strong>\u56fe 26\uff1a<\/strong> \u4f7f\u7528 Agent \u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 27\uff1a\u4f7f\u7528 Agent \u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig27.png\" \/><\/p>\n<p><strong>\u56fe 27\uff1a<\/strong> \u4f7f\u7528 Agent \u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 28\uff1a\u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig28.png\" \/><\/p>\n<p><strong>\u56fe 28\uff1a<\/strong> \u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 29\uff1a\u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig29.png\" \/><\/p>\n<p><strong>\u56fe 29\uff1a<\/strong> \u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 30\uff1a\u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig30.png\" \/><\/p>\n<p><strong>\u56fe 30\uff1a<\/strong> \u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" style=\"max-width:100%;height:auto\" alt=\"\u56fe 31\uff1a\u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c\" src=\"https:\/\/youpaiyun.lingbo.online\/paper-read\/2601.19494\/fig31.png\" \/><\/p>\n<p><strong>\u56fe 31\uff1a<\/strong> \u4f7f\u7528 BM25\/Embedding \u68c0\u7d22\u7684\u9519\u8bef\u6848\u4f8b\u6837\u672c<\/p>\n<div class=\"footnote\">\n<hr \/>\n<ol>\n<li id=\"fn:1\">\n<p>https:\/\/github.com\/alibaba\/aacr-bench&#160;<a class=\"footnote-backref\" href=\"1\" title=\"Jump back to footnote 1 in the text\" target=\"_blank\"  rel=\"nofollow\" >&#8617;<\/a><\/p>\n<\/li>\n<li id=\"fn:2\">\n<p>https:\/\/survey.stackoverflow.co\/2025\/technology&#160;<a class=\"footnote-backref\" href=\"2\" title=\"Jump back to footnote 2 in the text\" target=\"_blank\"  rel=\"nofollow\" >&#8617;<\/a><\/p>\n<\/li>\n<\/ol>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>Meta Data \u53d1\u8868\u65f6\u95f4\uff1a2026-01-30 \u4f5c\u8005\uff1aLei Zhang, Yongda Yu, Minghui Yu, Xi &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[60,41],"tags":[],"class_list":["post-1659","post","type-post","status-publish","format-standard","hentry","category-paper_read","category-learning_note"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1659","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1659"}],"version-history":[{"count":0,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1659\/revisions"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1659"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1659"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1659"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}