{"id":4733,"date":"2026-01-15T23:27:08","date_gmt":"2026-01-15T15:27:08","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4733"},"modified":"2026-01-15T23:27:08","modified_gmt":"2026-01-15T15:27:08","slug":"202601-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-llm-self-correction-with-decrim-decompose-critique-and-refine-for-enhanced-following-of-instructionswith-multiple-constraints","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202601-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-llm-self-correction-with-decrim-decompose-critique-and-refine-for-enhanced-following-of-instructionswith-multiple-constraints\/","title":{"rendered":"202601 \u8bba\u6587\u7814\u8bfb – LLM Self-Correction with DECRIM: DECOMPOSE, CRITIQUE, AND REFINE for Enhanced Following of Instructionswith Multiple Constraints"},"content":{"rendered":"\n

\u4f5c\u8005\uff1aThomas Palmeira Ferraz, Kartik Mehta, Yu-Hsiang Lin\u7b49
\u5355\u4f4d\uff1a\u4e9a\u9a6c\u900a\u901a\u7528\u4eba\u5de5\u667a\u80fd\u57fa\u7840\u7814\u7a76\u56e2\u961f\u3001\u6cd5\u56fd\u5df4\u9ece\u7535\u4fe1\u5b66\u9662\u7b49
\u6765\u6e90\uff1aEMNLP
\u65f6\u95f4\uff1a2024.10<\/p>\n\n\n\n

\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a<\/p>\n\n\n\n

\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u9075\u5faa\u6307\u4ee4\u65b9\u9762\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\uff0c\u4f46\u5f53\u6307\u4ee4\u5305\u542b\u591a\u4e2a\u7ea6\u675f\u6761\u4ef6\u65f6\uff08\u4f8b\u5982\uff0c\u201c\u5199\u4e00\u6761\u63a8\u6587\uff0c\u8bed\u6c14\u8981\u5e7d\u9ed8\uff0c\u4e0d\u80fd\u4f7f\u7528\u6807\u7b7e\uff0c\u5e76\u4e14\u8981\u63a8\u5e7fAI\u201d\uff09\uff0c\u5176\u8868\u73b0\u5f80\u5f80\u4e0d\u5c3d\u5982\u4eba\u610f\u3002\u73b0\u6709\u7814\u7a76\u548c\u8bc4\u4f30\u57fa\u51c6\u5b58\u5728\u4e24\u4e2a\u4e3b\u8981\u95ee\u9898\uff1a
\u6570\u636e\u4e0d\u771f\u5b9e\uff1a\u5927\u591a\u6570\u8bc4\u4f30\u57fa\u51c6\uff08\u5982IFEval\uff09\u4f9d\u8d56\u4e8e\u5408\u6210\u6570\u636e\uff08synthetic data\uff09\u3002\u8fd9\u4e9b\u6570\u636e\u867d\u7136\u4fbf\u4e8e\u81ea\u52a8\u5316\u8bc4\u4f30\uff0c\u4f46\u53ef\u80fd\u65e0\u6cd5\u51c6\u786e\u6355\u6349\u771f\u5b9e\u4e16\u754c\u7528\u6237\u8bf7\u6c42\u7684\u590d\u6742\u6027\u548c\u7ec6\u5fae\u5dee\u522b\uff0c\u751a\u81f3\u53ef\u80fd\u4eba\u4e3a\u5730\u5236\u9020\u51fa\u4e0d\u5207\u5b9e\u9645\u7684\u96be\u9898\u3002
\u8bc4\u4f30\u6210\u672c\u9ad8\uff1a\u8bc4\u4f30\u771f\u5b9e\u4e16\u754c\u7684\u5f00\u653e\u5f0f\u6307\u4ee4\u975e\u5e38\u56f0\u96be\uff0c\u901a\u5e38\u9700\u8981\u6602\u8d35\u4e14\u8017\u65f6\u7684\u4eba\u5de5\u6807\u6ce8\u3002
\u56e0\u6b64\uff0c\u8be5\u9886\u57df\u7f3a\u4e4f\u4e00\u4e2a\u80fd\u6709\u6548\u8861\u91cfLLM\u5728\u5904\u7406\u771f\u5b9e\u3001\u591a\u7ea6\u675f\u7528\u6237\u8bf7\u6c42\u65f6\u80fd\u529b\u7684\u57fa\u51c6\uff0c\u4e5f\u7f3a\u4e4f\u4e00\u79cd\u9ad8\u6548\u3001\u4f4e\u6210\u672c\u7684\u8bc4\u4f30\u65b9\u6cd5\u3002\u540c\u65f6\uff0c\u5f00\u6e90\u6a21\u578b\u4e0e\u9876\u5c16\u95ed\u6e90\u6a21\u578b\uff08\u5982GPT-4\uff09\u4e4b\u95f4\u5728\u8be5\u4efb\u52a1\u4e0a\u5b58\u5728\u660e\u663e\u7684\u6027\u80fd\u5dee\u8ddd\u3002<\/p>\n\n\n\n

\u7814\u7a76\u95ee\u9898<\/p>\n\n\n\n

\u672c\u6587\u65e8\u5728\u89e3\u51b3\u4ee5\u4e0b\u4e09\u4e2a\u6838\u5fc3\u95ee\u9898\uff1a
\u8bc4\u4f30\u95ee\u9898\uff1a\u5982\u4f55\u6784\u5efa\u4e00\u4e2a\u57fa\u4e8e\u771f\u5b9e\u7528\u6237\u8bf7\u6c42\u7684\u3001\u53ef\u9760\u7684\u591a\u7ea6\u675f\u6307\u4ee4\u9075\u5faa\u8bc4\u4f30\u57fa\u51c6\uff1f
\u8bc4\u4f30\u6210\u672c\u95ee\u9898\uff1a\u80fd\u5426\u7528\u6a21\u578b\u81ea\u8eab\uff08LLM-as-a-Judge\uff09\u6765\u66ff\u4ee3\u4eba\u5de5\u8fdb\u884c\u4f4e\u6210\u672c\u3001\u9ad8\u6548\u7387\u7684\u8bc4\u4f30\uff1f
\u6027\u80fd\u63d0\u5347\u95ee\u9898\uff1a\u5982\u4f55\u8bbe\u8ba1\u4e00\u4e2a\u901a\u7528\u7684\u81ea\u4fee\u6b63\u6846\u67b6\uff0c\u4ee5\u63d0\u5347\uff08\u5c24\u5176\u662f\u5f00\u6e90\uff09LLM\u5728\u591a\u7ea6\u675f\u6307\u4ee4\u9075\u5faa\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\uff0c\u5e76\u7f29\u5c0f\u4e0e\u9876\u5c16\u95ed\u6e90\u6a21\u578b\u7684\u5dee\u8ddd\uff1f
\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e09\u9879\u4e3b\u8981\u8d21\u732e\uff1a
REALINSTRUCT \u57fa\u51c6\uff1a\u8fd9\u662f\u7b2c\u4e00\u4e2a\u5229\u7528\u771f\u5b9e\u7528\u6237\u5411AI\u52a9\u624b\u63d0\u95ee\u7684\u6570\u636e\u6784\u5efa\u7684\u591a\u7ea6\u675f\u6307\u4ee4\u9075\u5faa\u8bc4\u4f30\u57fa\u51c6\u3002
LLM-as-a-Judge \u8bc4\u4f30\u534f\u8bae\uff1a\u7cfb\u7edf\u6027\u5730\u7814\u7a76\u4e86\u4f7f\u7528\u5f00\u6e90\u548c\u95ed\u6e90LLM\u4f5c\u4e3a\u8bc4\u4f30\u5668\u6765\u5224\u65ad\u7ea6\u675f\u6ee1\u8db3\u60c5\u51b5\u7684\u6709\u6548\u6027\u3002
DECRIM \u81ea\u4fee\u6b63\u6d41\u6c34\u7ebf\uff1a\u4e00\u4e2a\u65b0\u9896\u7684\u3001\u65e0\u9700\u5bf9\u7ea6\u675f\u7c7b\u578b\u505a\u4efb\u4f55\u5047\u8bbe\u7684\u81ea\u4fee\u6b63\u6846\u67b6\uff0c\u7528\u4e8e\u589e\u5f3aLLM\u9075\u5faa\u591a\u7ea6\u675f\u6307\u4ee4\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n

\u89e3\u51b3\u529e\u6cd5-DECRIM \u65b9\u6cd5<\/p>\n\n\n\n

DECRIM \u662f\u4e00\u4e2a\u8fed\u4ee3\u5f0f\u7684\u81ea\u4fee\u6b63\u6d41\u6c34\u7ebf\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u590d\u6742\u7684\u591a\u7ea6\u675f\u6307\u4ee4\u5206\u89e3\u3001\u6279\u5224\u5e76\u7cbe\u70bc\u3002\u5b83\u5305\u542b\u56db\u4e2a\u5173\u952e\u6b65\u9aa4\uff0c\u5faa\u73af\u6267\u884c\u76f4\u81f3\u6240\u6709\u7ea6\u675f\u88ab\u6ee1\u8db3\u6216\u8fbe\u5230\u6700\u5927\u8fed\u4ee3\u6b21\u6570\u3002
\u6b65\u9aa4\u4e00\uff1a\u521d\u59cb\u54cd\u5e94 (Initial Response)
LLM\u76f4\u63a5\u6839\u636e\u7528\u6237\u7684\u539f\u59cb\u6307\u4ee4\u751f\u6210\u4e00\u4e2a\u521d\u6b65\u7684\u56de\u7b54\u3002
\u6b65\u9aa4\u4e8c\uff1a\u5206\u89e3 (Decompose)
\u4f7f\u7528\u4e00\u4e2a\u201c\u5206\u89e3\u5668\u201d\uff08Decomposer\uff09\u6a21\u578b\uff08\u53ef\u4ee5\u662fLLM\u672c\u8eab\u6216\u5176\u4ed6\u6a21\u578b\uff09\u5c06\u539f\u59cb\u7684\u3001\u590d\u6742\u7684\u7528\u6237\u6307\u4ee4\u5206\u89e3\u6210\u4e00\u4e2a\u6e05\u6670\u3001\u7ec6\u7c92\u5ea6\u7684\u7ea6\u675f\u5217\u8868\u3002<\/p>\n\n\n\n

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

\u6b65\u9aa4\u4e09\uff1a\u6279\u5224 (Critique)
\u4f7f\u7528\u4e00\u4e2a\u201c\u6279\u8bc4\u5bb6\u201d\uff08Critic\uff09\u6a21\u578b\uff0c\u6839\u636e\u4e0a\u4e00\u6b65\u5f97\u5230\u7684\u7ea6\u675f\u5217\u8868\uff0c\u9010\u6761\u68c0\u67e5\u521d\u59cb\u54cd\u5e94\u662f\u5426\u6ee1\u8db3\u4e86\u6240\u6709\u7ea6\u675f\u3002
\u5982\u679c\u6240\u6709\u7ea6\u675f\u90fd\u6ee1\u8db3\uff0c\u5219\u5c06\u5f53\u524d\u54cd\u5e94\u4f5c\u4e3a\u6700\u7ec8\u8f93\u51fa\uff0c\u6d41\u7a0b\u7ed3\u675f\u3002
\u5982\u679c\u6709\u672a\u6ee1\u8db3\u7684\u7ea6\u675f\uff0c\u6279\u8bc4\u5bb6\u4f1a\u4ee5\u81ea\u7136\u8bed\u8a00\u7684\u5f62\u5f0f\u63d0\u4f9b\u5177\u4f53\u7684\u53cd\u9988\uff0c\u660e\u786e\u6307\u51fa\u54ea\u4e9b\u7ea6\u675f\u672a\u88ab\u6ee1\u8db3\u3002
\u6b65\u9aa4\u56db\uff1a\u7cbe\u70bc (Refine)
\u5c06\u6279\u8bc4\u5bb6\u63d0\u4f9b\u7684\u5177\u4f53\u53cd\u9988\u3001\u539f\u59cb\u7528\u6237\u6307\u4ee4\u4ee5\u53ca\u4e0a\u4e00\u8f6e\u7684\u54cd\u5e94\u4e00\u8d77\u8f93\u5165\u7ed9\u5e95\u5c42\u7684LLM\u3002LLM\u6839\u636e\u8fd9\u4e9b\u4fe1\u606f\u751f\u6210\u4e00\u4e2a\u6539\u8fdb\u540e\u7684\u65b0\u54cd\u5e94\uff0c\u76ee\u6807\u662f\u4fee\u6b63\u4e4b\u524d\u672a\u6ee1\u8db3\u7684\u7ea6\u675f\u3002
\u5faa\u73af\u8fed\u4ee3\uff1a\u65b0\u751f\u6210\u7684\u54cd\u5e94\u4f1a\u518d\u6b21\u8fdb\u5165\u201c\u6279\u5224\u201d\u6b65\u9aa4\uff0c\u63a5\u53d7\u65b0\u4e00\u8f6e\u7684\u68c0\u67e5\u3002\u8fd9\u4e2a\u201c\u6279\u5224-\u7cbe\u70bc\u201d\u7684\u5faa\u73af\u4f1a\u6301\u7eed\u8fdb\u884c\uff0c\u76f4\u5230\u54cd\u5e94\u901a\u8fc7\u6240\u6709\u7ea6\u675f\u7684\u68c0\u9a8c\uff0c\u6216\u8005\u8fbe\u5230\u9884\u8bbe\u7684\u6700\u5927\u8fed\u4ee3\u6b21\u6570\u3002<\/p>\n\n\n\n

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

\u5b9e\u9a8c\u56f4\u7ed5\u4e09\u5927\u8d21\u732e\u5c55\u5f00\uff1a
\u9a8c\u8bc1 LLM-as-a-Judge\uff1a
\u6570\u636e\u96c6\uff1a\u521b\u5efa\u4e86EvalJudge\u6570\u636e\u96c6\uff0c\u5305\u542b\u8fd11000\u4e2a\u7ecf\u8fc7\u4e13\u5bb6\u4eba\u5de5\u9a8c\u8bc1\u7684\uff08\u6307\u4ee4\uff0c\u7ea6\u675f\uff0c\u54cd\u5e94\uff09\u4e09\u5143\u7ec4\u3002
\u6a21\u578b\uff1a\u6d4b\u8bd5\u4e86GPT-4\u3001GPT-4-Turbo\u3001GPT-3.5-Turbo\u7b49\u95ed\u6e90\u6a21\u578b\uff0c\u4ee5\u53caMistral\u3001Vicuna\u3001Zephyr\u7b49\u5f00\u6e90\u6a21\u578b\u3002
\u7b56\u7565\uff1a\u6bd4\u8f83\u4e86\u4e0d\u540c\u7684\u63d0\u793a\u7b56\u7565\uff08\u5982\u6574\u4f53\u8bc4\u4f30 vs. \u9010\u6761\u8bc4\u4f30\uff0c\u662f\u5426\u4f7f\u7528\u601d\u7ef4\u94feCoT\uff09\u548c\u5fae\u8c03\u7b56\u7565\uff08\u7528GPT-4-Turbo\u7684\u63a8\u7406\u8fc7\u7a0b\u5f31\u76d1\u7763\u8bad\u7ec3Mistral\uff09\u3002
REALINSTRUCT \u57fa\u51c6\u6d4b\u8bd5\uff1a
\u5728REALINSTRUCT\u548c\u6d41\u884c\u7684\u5408\u6210\u57fa\u51c6IFEval\u4e0a\uff0c\u5bf9\u591a\u4e2a\u5f00\u6e90\u548c\u95ed\u6e90\u6a21\u578b\u8fdb\u884c\u4e86\u591a\u7ea6\u675f\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u7684\u57fa\u51c6\u6d4b\u8bd5\u3002
\u4f7f\u7528\u7ecf\u8fc7\u9a8c\u8bc1\u7684\u6700\u4f73\u8bc4\u4f30\u5668\uff08GPT-4-Turbo + CoT\uff09\u8fdb\u884c\u6253\u5206\u3002
DECRIM \u6d41\u6c34\u7ebf\u8bc4\u4f30\uff1a
\u4e3b\u6a21\u578b\uff1a\u4ee5Mistral 7B v0.2\u4f5c\u4e3a\u5e95\u5c42LLM\u3002
\u57fa\u7ebf\uff1a\u4e0e\u5e38\u89c4\u63d0\u793a\u3001\u201cMake sure to follow all constraints\u201d\u63d0\u793a\u3001\u4ee5\u53caSelf-Refine\u7b49\u73b0\u6709\u81ea\u4fee\u6b63\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\u3002
\u6d88\u878d\u7814\u7a76\uff1a\u5206\u522b\u6d4b\u8bd5\u4e86\u4e0d\u540c\u5f3a\u5ea6\u7684\u201c\u5206\u89e3\u5668\u201d\uff08Self-Decomposer vs. Oracle Decomposer\uff09\u548c\u201c\u6279\u8bc4\u5bb6\u201d\uff08Self-Critic, Supervised Critic, GPT-4, Oracle Critic\uff09\u5bf9\u6700\u7ec8\u6027\u80fd\u7684\u5f71\u54cd\u3002
\u8d28\u91cf\u8bc4\u4f30\uff1a\u4f7f\u7528Prometheus-2\u8bc4\u4f30\u7cbe\u70bc\u524d\u540e\u54cd\u5e94\u7684\u6574\u4f53\u8d28\u91cf\uff0c\u786e\u4fdd\u4fee\u6b63\u8fc7\u7a0b\u6ca1\u6709\u635f\u5bb3\u56de\u7b54\u7684\u6d41\u7545\u6027\u548c\u76f8\u5173\u6027\u3002<\/p>\n\n\n\n

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

\u6838\u5fc3\u5185\u5bb9\u603b\u7ed3\uff1a
1.\u63d0\u51faREALINSTRUCT\u57fa\u51c6\uff1a\u9996\u4e2a\u57fa\u4e8e\u771f\u5b9e\u7528\u6237\u4ea4\u4e92\u7684\u591a\u7ea6\u675f\u6307\u4ee4\u9075\u5faa\u8bc4\u4f30\u96c6\uff0c\u586b\u8865\u4e86\u5408\u6210\u6570\u636e\u4e0e\u73b0\u5b9e\u5e94\u7528\u4e4b\u95f4\u7684\u9e3f\u6c9f\u3002
2.\u63d0\u51faDECRIM\u6846\u67b6\uff1a\u4e00\u4e2a\u4e0d\u4f9d\u8d56\u4e8e\u7ea6\u675f\u5148\u9a8c\u77e5\u8bc6\u7684\u3001\u57fa\u4e8e\u7ec6\u7c92\u5ea6\u53cd\u9988\u7684\u8fed\u4ee3\u5f0f\u81ea\u4fee\u6b63\u8303\u5f0f\u3002\u5176\u6838\u5fc3\u521b\u65b0\u5728\u4e8e\u5c06\u590d\u6742\u7684\u5168\u5c40\u4f18\u5316\u95ee\u9898\uff0c\u8f6c\u5316\u4e3a\u4e00\u7cfb\u5217\u7531\u6279\u8bc4\u5bb6\u6307\u5bfc\u7684\u3001\u76ee\u6807\u660e\u786e\u7684\u5c40\u90e8\u4fee\u6b63\u4efb\u52a1\u3002
3.\u5bf9LLM-as-a-Judge\u7684\u7cfb\u7edf\u6027\u5206\u6790\uff1a\u9996\u6b21\u5728\u591a\u7ea6\u675f\u9075\u5faa\u4efb\u52a1\u4e0a\uff0c\u5168\u9762\u6bd4\u8f83\u4e86\u4e0d\u540c\u6a21\u578b\u548c\u7b56\u7565\u4f5c\u4e3a\u8bc4\u4f30\u5668\u7684\u6548\u679c\uff0c\u4e3a\u793e\u533a\u63d0\u4f9b\u4e86\u5b9d\u8d35\u7684\u5b9e\u8df5\u6307\u5357\u3002<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1aThomas Palmeira Ferraz, Kartik Mehta, Yu-Hsiang Lin\u7b49 […]<\/p>\n","protected":false},"author":66,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4733","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4733","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/66"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=4733"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4733\/revisions"}],"predecessor-version":[{"id":4738,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4733\/revisions\/4738"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4733"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4733"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4733"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}