{"id":3345,"date":"2025-06-15T16:55:49","date_gmt":"2025-06-15T08:55:49","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=3345"},"modified":"2025-06-15T16:59:20","modified_gmt":"2025-06-15T08:59:20","slug":"202506%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-promptbreederself-referential-self-improvement-via-prompt-evolution","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202506%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-promptbreederself-referential-self-improvement-via-prompt-evolution\/","title":{"rendered":"202506 \u8bba\u6587\u7814\u8bfb-PromptBreeder:Self-referential Self-improvement via Prompt Evolution"},"content":{"rendered":"\n<p>\u4f5c\u8005\uff1aChrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, Tim Rockta \u0308schel<br>\u5355\u4f4d\uff1aGoogle DeepMind<br>\u6765\u6e90\uff1aICML 2024<\/p>\n\n\n\n<p>\u8fd1\u5e74\u6765\uff0c\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u7684\u53d1\u5c55\uff0c\u7279\u522b\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u8fdb\u6b65\uff0c\u81ea\u7136\u8bed\u8a00\u5904\u7406\u7684\u80fd\u529b\u5f97\u5230\u4e86\u663e\u8457\u63d0\u5347\u3002\u8fd9\u4e9b\u6a21\u578b\u80fd\u591f\u5728\u5404\u79cd\u4efb\u52a1\u4e0a\u5b9e\u73b0\u5148\u8fdb\u7684\u6027\u80fd\uff0c\u5305\u62ec\u6587\u672c\u751f\u6210\u3001\u95ee\u7b54\u7cfb\u7edf\u7b49\u3002<br>\u63d0\u793a\u662f\u57fa\u7840\u6a21\u578b\u4e0b\u6e38\u6027\u80fd\u7684\u6838\u5fc3\u3002\u8be5\u6587\u7ae0\u63d0\u51fa\u4e00\u79cd\u901a\u8fc7\u591a\u6837\u6027\u4fdd\u6301\u8fdb\u5316\u7b97\u6cd5\u6765\u89e3\u51b3\u6536\u76ca\u9012\u51cf\u95ee\u9898\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u7528\u4e8e LLM \u63d0\u793a\u7684\u81ea\u6211\u5f15\u7528\u81ea\u6211\u5b8c\u5584\u3002<br>\u5373Promptbreeder\uff0c\u4e00\u4e2a\u63d0\u793a\u8fdb\u5316\u7cfb\u7edf\uff0c\u53ef\u4ee5\u81ea\u52a8\u63a2\u7d22\u7ed9\u5b9a\u9886\u57df\u7684\u63d0\u793a\uff0c\u5e76\u4e14\u80fd\u591f\u627e\u5230\u4efb\u52a1\u63d0\u793a\uff0c\u4ece\u800c\u63d0\u9ad8 LLM \u83b7\u5f97\u8be5\u9886\u57df\u95ee\u9898\u7684\u7b54\u6848\u7684\u80fd\u529b\u3002Promptbreeder \u662f\u901a\u7528\u7684\uff0c\u56e0\u4e3a\u540c\u4e00\u4e2a\u7cfb\u7edf\u80fd\u591f\u9002\u5e94\u8bb8\u591a\u4e0d\u540c\u7684\u9886\u57df\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"722\" height=\"450\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-1.png\" alt=\"\" class=\"wp-image-3346\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-1.png 722w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-1-300x187.png 300w\" sizes=\"auto, (max-width: 722px) 100vw, 722px\" \/><\/figure>\n\n\n\n<p>\u5177\u4f53\u7684\u4f8b\u5b50\uff0c\u8003\u8651\u7528\u4e8e\u4e3a GSM8K\uff08\u4e00\u4e2a \u201c\u5c0f\u5b66\u6570\u5b66 \u201d\u5355\u8bcd\u95ee\u9898\u6570\u636e\u96c6\uff09\u751f\u6210\u4efb\u52a1\u63d0\u793a\u548c\u7a81\u53d8\u63d0\u793a\u7684\u521d\u59cb\u5316\u6b65\u9aa4\uff1a<br>\u95ee\u9898\u63cf\u8ff0\u662f\u201c\u89e3\u51b3\u6570\u5b66\u5355\u8bcd\u95ee\u9898\uff0c\u4ee5\u963f\u62c9\u4f2f\u6570\u5b57\u5f62\u5f0f\u7ed9\u51fa\u60a8\u7684\u7b54\u6848\u201d\u3002<br>\u5728\u6bcf\u4e2a\u8fdb\u5316\u5355\u4f4d\u8fdb\u5316\u4e86\u4e24\u4e2a\u4efb\u52a1\u63d0\u793a\uff08\u52a0\u4e0a\u4e00\u4e2a\u7a81\u53d8\u63d0\u793a\uff09\u3002\u4e3a\u4e86\u4fc3\u8fdb\u521d\u59cb\u63d0\u793a\u7684\u591a\u6837\u6027\uff0c\u901a\u8fc7\u5c06\u968f\u673a\u7ed8\u5236\u7684\u201c\u7a81\u53d8\u63d0\u793a\u201d\uff08\u4f8b\u5982\u201cMake a variant of the prompt\u201d\uff09\u548c\u968f\u673a\u7ed8\u5236\u7684\u201c\u601d\u7ef4\u98ce\u683c\u201d\uff08\u4f8b\u5982\u201cLet&#8217;s think step by step\u201d\uff09\u8fde\u63a5\u5230\u95ee\u9898\u63cf\u8ff0\u6765\u751f\u6210\u521d\u59cb\u4efb\u52a1\u63d0\u793a\uff0c\u5e76\u5c06\u5176\u63d0\u4f9b\u7ed9 LLM \u4ee5\u4ea7\u751f\u5ef6\u7eed\uff0c\u4ece\u800c\u4ea7\u751f\u521d\u59cb\u4efb\u52a1\u63d0\u793a\u3002<br>\u6211\u4eec\u8fd9\u6837\u505a\u4e24\u6b21\uff0c\u4ee5\u4fbf\u6bcf\u4e2a\u5355\u5143\u751f\u6210\u4e24\u4e2a\u521d\u59cb\u4efb\u52a1\u63d0\u793a\u3002\u7a81\u53d8\u63d0\u793a\u548c\u601d\u7ef4\u65b9\u5f0f\u90fd\u662f\u4ece\u4e00\u7ec4\u521d\u59cb\u7a81\u53d8\u63d0\u793a\u548c\u4e00\u7ec4\u601d\u7ef4\u65b9\u5f0f\u4e2d\u968f\u673a\u62bd\u6837\u7684\u3002<br>mutation-prompt \u88ab\u6dfb\u52a0\u5230\u8fdb\u5316\u5355\u5143\u4e2d\uff0c\u56e0\u6b64\u5728\u6574\u4e2a\u8fdb\u5316\u8fd0\u884c\u8fc7\u7a0b\u4e2d\u4e0e\u5176\u7279\u5b9a\u7684 task-prompt \u76f8\u5173\u8054\u3002<br>\u5728\u4e0a\u9762\u7684\u793a\u4f8b\u4e2d\uff0c\u7528\u4e8e\u751f\u6210\u521d\u59cb\u4efb\u52a1\u63d0\u793a\u7684 LLM \u7684\u5b8c\u6574\u8f93\u5165\u5b57\u7b26\u4e32\u53ef\u4ee5\u662f \u201cMake a variant of the prompt.\u8ba9\u6211\u4eec\u4e00\u6b65\u4e00\u6b65\u6765\u601d\u8003\u3002\u8bf4\u660e\uff1a\u89e3\u51b3\u6570\u5b66\u5355\u8bcd\u95ee\u9898\uff0c\u4ee5\u963f\u62c9\u4f2f\u6570\u5b57\u5f62\u5f0f\u7ed9\u51fa\u60a8\u7684\u7b54\u6848\u3002\u6307\u4ee4\u7a81\u53d8\uff1a\u201c\u3002\u8bf7\u6ce8\u610f\u5982\u4f55\u6dfb\u52a0\u63a7\u5236\u5b57\u7b26\u4e32 \u201cINSTRUCTION\u201d \u548c \u201cINSTRUCTION MUTANT\u201d \u4ee5\u9f13\u52b1\u9002\u5f53\u7684\u5ef6\u7eed\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"698\" height=\"228\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-2.png\" alt=\"\" class=\"wp-image-3348\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-2.png 698w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-2-300x98.png 300w\" sizes=\"auto, (max-width: 698px) 100vw, 698px\" \/><\/figure>\n\n\n\n<p>\u4ece\u4e00\u4e2a\u73b0\u6709\u7684\u4efb\u52a1\u63d0\u793a P\uff08\u4e00\u9636\u63d0\u793a\u751f\u6210\uff09\u6216\u9f13\u52b1\u81ea\u7531\u751f\u6210\u65b0\u4efb\u52a1\u63d0\u793a\u7684\u901a\u7528\u63d0\u793a\u751f\u6210\u4e00\u4e2a\u65b0\u7684\u4efb\u52a1\u63d0\u793a P \u2032 &#8211; \u5373\u4e0d\u4f7f\u7528\u73b0\u6709\u7684\u7236\u7ea7\uff0c\u56e0\u6b64\u662f\u96f6\u9636\u63d0\u793a\u751f\u6210\u3002<br>\u96f6\u9636\u63d0\u793a\u751f\u6210\uff1a\u901a\u8fc7\u5c06\u95ee\u9898\u63cf\u8ff0 D\u4e0e\u63d0\u793a\u8fde\u63a5\u8d77\u6765\u6765\u751f\u6210\u4e00\u4e2a\u65b0\u7684\u4efb\u52a1\u63d0\u793a\uff0c\u5b83\u9080\u8bf7 LLM \u63d0\u51fa\u4e00\u4e2a\u53ef\u4ee5\u5e2e\u52a9\u89e3\u51b3\u7ed9\u5b9a\u95ee\u9898\u57df\u4e2d\u7684\u95ee\u9898\u7684\u65b0\u63d0\u793a\u3002\u5c06\u7b2c\u4e00\u4e2a\u751f\u6210\u7684 hint \u63d0\u53d6\u4e3a\u65b0\u7684\u4efb\u52a1\u63d0\u793a\u7b26\u3002\u8fd9\u4e2a\u65b0\u7684\u4efb\u52a1\u63d0\u793a\u7b26\u4e0d\u4f9d\u8d56\u4e8e\u4ee5\u524d\u627e\u5230\u7684\u4efb\u4f55\u4efb\u52a1\u63d0\u793a\u7b26\u3002\u76f8\u53cd\uff0c\u6bcf\u6b21\u90fd\u4f1a\u4ece\u95ee\u9898\u63cf\u8ff0\u4e2d\u91cd\u65b0\u751f\u6210\u5b83\u3002<br>\u5305\u542b\u8fd9\u4e2a\u96f6\u9636\u8fd0\u7b97\u7b26\u7684\u57fa\u672c\u539f\u7406\u662f\uff0c\u5f53\u63d0\u793a\u8fdb\u5316\u53d1\u6563\u65f6\uff0c\u8fd9\u4e2a\u8fd0\u7b97\u7b26\u5141\u8bb8\u751f\u6210\u4e0e\u539f\u59cb\u95ee\u9898\u63cf\u8ff0\u5bc6\u5207\u76f8\u5173\u7684\u65b0\u4efb\u52a1\u63d0\u793a\uff0c\u7c7b\u4f3c\u4e8e\u81ea\u52a8\u5316\u8bfe\u7a0b\u5b66\u4e60\u65b9\u6cd5\u4e2d\u7684\u7edf\u4e00\u91cd\u65b0\u91c7\u6837\u3002<br>\u4e00\u9636\u63d0\u793a\u751f\u6210\uff1a\u5c06 mutation-prompt\uff08\u7ea2\u8272\uff09\u8fde\u63a5\u5230\u7236 task-prompt\uff08\u84dd\u8272\uff09\uff0c\u5e76\u5c06\u5176\u4f20\u9012\u7ed9 LLM \u4ee5\u751f\u6210 muted task-prompt\u3002\u4f8b\u5982\uff0c\u201c\u4ee5\u53e6\u4e00\u79cd\u65b9\u5f0f\u518d\u6b21\u8bf4\u51fa\u8be5\u6307\u4ee4\u3002\u4e0d\u8981\u4f7f\u7528\u539f\u59cb\u8bf4\u660e\u4e2d\u7684\u4efb\u4f55\u5355\u8bcd\uff0c\u8fd9\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u5c0f\u4f19\u5b50\u3002\u8bf4\u660e\uff1a\u89e3\u51b3\u6570\u5b66\u5355\u8bcd\u95ee\u9898\uff0c\u4ee5\u963f\u62c9\u4f2f\u6570\u5b57\u7ed9\u51fa\u60a8\u7684\u7b54\u6848\u3002\u6307\u4ee4\u7a81\u53d8\u4f53\uff1a\u201c.\u6b64\u8fc7\u7a0b\u4e0e\u521d\u59cb\u5316\u65b9\u6cd5\u76f8\u540c\uff0c\u53ea\u662f\u4e0d\u4f7f\u7528\u968f\u673a\u91c7\u6837\u7684\u601d\u7ef4\u6837\u5f0f\u5b57\u7b26\u4e32\u3002\u4e00\u9636\u63d0\u793a\u751f\u6210\u662f Promptbreeder \u7684\u6807\u51c6\u65e0\u6027\u7a81\u53d8\u8fd0\u7b97\u7b26\uff0c\u5b83\u662f\u6bcf\u4e2a\u9057\u4f20\u7b97\u6cd5\u7684\u6838\u5fc3\u2014\u2014\u91c7\u7528\u4e00\u4e2a\u4eb2\u672c\u57fa\u56e0\u578b\uff08\u4efb\u52a1\u63d0\u793a\uff09\u5e76\u5c06\u7a81\u53d8\u5e94\u7528\u4e8e\u5b83\uff08\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u53d7\u7a81\u53d8\u63d0\u793a\u7684\u5f71\u54cd\uff09\u3002<\/p>\n\n\n\n<p>\u5206\u5e03\u7a81\u53d8\u7684\u4f30\u8ba1\u4e0b\u4e00\u7c7b\u7a81\u53d8\u8fd0\u7b97\u7b26\u4e0d\u4ec5\u4ee5\u96f6\u4e2a\u6216\u4e00\u4e2a\u7236\u7ea7\u4e3a\u6761\u4ef6\uff0c\u8fd8\u4ee5\u4e00\u7ec4\u7236\u7ea7\u4e3a\u6761\u4ef6\u3002<br>\u56e0\u6b64\uff0c\u901a\u8fc7\u8003\u8651\u603b\u4f53\u6a21\u5f0f\uff0c\u5b83\u4eec\u53ef\u80fd\u66f4\u5177\u8868\u73b0\u529b\u3002 \u5206\u5e03\u4f30\u8ba1\uff08EDA\uff09\u7a81\u53d8\uff1a\u53d7Hauschild\u548cPelikan\uff082011\uff09\u7684\u542f\u53d1\uff0c\u6211\u4eec\u5411LLM\u63d0\u4f9b\u4e86\u4e00\u4efd\u8fc7\u6ee4\u548c\u7f16\u53f7\u7684\u5f53\u524d\u4efb\u52a1\u63d0\u793a\u5217\u8868\uff0c\u5e76\u8981\u6c42\u5b83\u7ee7\u7eed\u8fd9\u4e2a\u5217\u8868\uff0c\u4f7f\u7528\u65b0\u7684\u4efb\u52a1\u63d0\u793a\u3002\u6839\u636eBERT\u8fc7\u6ee4\u63d0\u793a\u7684\u7fa4\u4f53\uff0c\u5d4c\u5165\u5f7c\u6b64\u4e4b\u95f4\u7684\u4f59\u5f26\u76f8\u4f3c\u6027\u2014\u2014\u5982\u679c\u4e00\u4e2a\u4e2a\u4f53\u4e0e\u5217\u8868\u4e2d\u7684\u4efb\u4f55\u5176\u4ed6\u6761\u76ee\u7684\u76f8\u4f3c\u5ea6\u8d85\u8fc70.95\uff0c\u5219\u4e0d\u4f1a\u88ab\u5305\u542b\u5728\u5217\u8868\u4e2d\uff0c\u4ece\u800c\u9f13\u52b1\u591a\u6837\u6027\u3002\u63d0\u793a\u4ee5\u968f\u673a\u987a\u5e8f\u5217\u51fa\uff0c\u6211\u4eec\u4e0d\u4f1a\u8ba9 LLM \u8bbf\u95ee\u603b\u4f53\u4e2d\u4e2a\u4f53\u7684\u9002\u5e94\u5ea6\u503c\u2014\u2014\u6211\u4eec\u5728\u521d\u6b65\u5b9e\u9a8c\u4e2d\u53d1\u73b0 LLM \u4e0d\u7406\u89e3\u8fd9\u4e9b\u9002\u5e94\u5ea6\u503c3\uff0c\u5e76\u6c42\u52a9\u4e8e\u5728\u5217\u8868\u4e2d\u751f\u6210\u6761\u76ee\u7684\u526f\u672c\u3002<\/p>\n\n\n\n<p>\u5b9e\u9a8c\u8bbe\u7f6e<\/p>\n\n\n\n<p>\u6570\u636e\u96c6\uff1a<br>GSM8K, SVAMP, MultiArith, AddSub, AQuA-RAT, SingleEq\u7b49\u7b97\u672f\u63a8\u7406\u6570\u636e\u96c6\u3002<br>CommonsenseQA, StrategyQA\u7528\u4e8e\u5e38\u8bc6\u63a8\u7406\u3002<br>\u6307\u4ee4\u5f52\u7eb3\u4efb\u52a1\u548c\u5176\u4ed6\u5206\u7c7b\u4efb\u52a1\u3002<br>\u5b9e\u9a8c\u8bbe\u8ba1\uff1a<br>\u4f7f\u7528\u7684\u79cd\u7fa4\u5927\u5c0f\u3001\u4e16\u4ee3\u6570\u53ca\u8bc4\u4f30\u65b9\u6cd5\u3002<br>LLM\u91c7\u6837\u7684\u4e09\u79cd\u4e0d\u540c\u4e0a\u4e0b\u6587\uff1aRedescriber, Inducer, Evaluator\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"893\" height=\"535\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-3.png\" alt=\"\" class=\"wp-image-3349\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-3.png 893w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-3-300x180.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/06\/image-3-768x460.png 768w\" sizes=\"auto, (max-width: 893px) 100vw, 893px\" \/><\/figure>\n\n\n\n<p>\u6838\u5fc3\u5185\u5bb9\u603b\u7ed3\uff1a<br>1 \u63d0\u51fa\u4e86Promptbreeder\uff0c\u4e00\u79cd\u7528\u4e8e LLM \u7684\u81ea\u6211\u5f15\u7528\u81ea\u6211\u6539\u8fdb\u65b9\u6cd5\uff0c\u5b83\u4e3a\u76ee\u524d\u7684\u9886\u57df\u6539\u8fdb\u63d0\u793a\uff0c\u5e76\u6539\u8fdb\u4e86\u5b83\u6539\u8fdb\u8fd9\u4e9b\u63d0\u793a\u7684\u65b9\u5f0f\uff0c\u5e76\u4e14\u80fd\u591f\u5ef6\u4f38\u4e0a\u4e0b\u6587\u3002<br>3 \u8be5\u65b9\u6cd5\u6838\u5fc3\u662f\u901a\u8fc7\u4e00\u4e2a\u9057\u4f20\u7b97\u6cd5\u6765\u751f\u6210\u548c\u4f18\u5316\u63d0\u793a\u3002<br>3 \u6587\u7ae0\u4e2d\u5bf9\u90e8\u5206\u5e7f\u6cdb\u7684\u5e38\u7528\u7b97\u6cd5\u548c\u5e38\u8bc6\u63a8\u7406\u57fa\u51c6\u4e0a\u5bf9\u6700\u5148\u8fdb\u7684\u63d0\u793a\u7b56\u7565\u63d0\u51fa\u4e86\u6539\u8fdb\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1aChrisantha Fernando, Dylan Banarse, Henryk Michalews [&hellip;]<\/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-3345","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3345","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=3345"}],"version-history":[{"count":3,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3345\/revisions"}],"predecessor-version":[{"id":3354,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3345\/revisions\/3354"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=3345"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=3345"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=3345"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}