{"id":60125,"date":"2025-10-25T06:29:41","date_gmt":"2025-10-24T22:29:41","guid":{"rendered":"https:\/\/www.kdjingpai.com\/?p=60125"},"modified":"2025-10-25T06:29:41","modified_gmt":"2025-10-24T22:29:41","slug":"gepa","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/gepa\/","title":{"rendered":"GEPA\uff1a\u901a\u8fc7\u53cd\u601d\u6027\u6587\u672c\u8fdb\u5316\u5b9e\u73b0AI\u7cfb\u7edf\u4f18\u5316"},"content":{"rendered":"<p>GEPA (Genetic-Pareto) \u662f\u4e00\u4e2a\u7528\u4e8e\u4f18\u5316AI\u7cfb\u7edf\u4e2d\u5404\u7c7b\u6587\u672c\u7ec4\u4ef6\u7684\u6846\u67b6\u3002\u8fd9\u4e9b\u6587\u672c\u7ec4\u4ef6\u53ef\u4ee5AI\u6a21\u578b\u7684\u63d0\u793a\u8bcd\u3001\u4ee3\u7801\u7247\u6bb5\u6216\u914d\u7f6e\u6587\u4ef6\u3002\u5b83\u91c7\u7528\u4e86\u4e00\u79cd\u540d\u4e3a\u201c\u53cd\u601d\u6027\u6587\u672c\u8fdb\u5316\u201d\u7684\u65b9\u6cd5\uff0c\u901a\u8fc7\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u6765\u5206\u6790\u548c\u53cd\u601dAI\u7cfb\u7edf\u7684\u884c\u4e3a\u3002\u5177\u4f53\u6765\u8bf4\uff0cGEPA\u4f1a\u68c0\u89c6\u7cfb\u7edf\u8fd0\u884c\u8fc7\u7a0b\u4e2d\u4ea7\u751f\u7684\u6267\u884c\u548c\u8bc4\u4f30\u8bb0\u5f55\uff0c\u5e76\u5229\u7528\u8fd9\u4e9b\u4fe1\u606f\u6765\u8fdb\u884c\u9488\u5bf9\u6027\u7684\u6539\u8fdb\u3002\u8be5\u6846\u67b6\u7ed3\u5408\u4e86\u8fed\u4ee3\u53d8\u5f02\u3001\u53cd\u601d\u548c\u5e15\u7d2f\u6258\u6700\u4f18\u9009\u62e9\u7b49\u7b56\u7565\uff0c\u80fd\u591f\u5728\u8bc4\u4f30\u6b21\u6570\u6709\u9650\u7684\u60c5\u51b5\u4e0b\uff0c\u6f14\u5316\u51fa\u6027\u80fd\u66f4\u5f3a\u7684\u7cfb\u7edf\u7248\u672c\u3002GEPA\u4e0d\u4ec5\u53ef\u4ee5\u4f18\u5316\u5355\u4e2a\u7ec4\u4ef6\uff0c\u8fd8\u80fd\u534f\u540c\u6f14\u5316\u6a21\u5757\u5316\u7cfb\u7edf\u4e2d\u7684\u591a\u4e2a\u7ec4\u4ef6\uff0c\u4ece\u800c\u5728\u7279\u5b9a\u9886\u57df\u83b7\u5f97\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002\u6839\u636e\u5176\u7814\u7a76\u8bba\u6587\u300aGEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning\u300b\u7684\u9610\u8ff0\uff0c\u76f8\u8f83\u4e8e\u4f20\u7edf\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff0cGEPA\u5728\u63d0\u5347\u6027\u80fd\u7684\u540c\u65f6\uff0c\u6240\u9700\u6837\u672c\u6570\u91cf\u4e5f\u5927\u5e45\u51cf\u5c11\uff0c\u5c55\u73b0\u4e86\u66f4\u9ad8\u7684\u6548\u7387\u3002<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u53cd\u601d\u6027\u6587\u672c\u8fdb\u5316<\/strong>: \u5229\u7528\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5206\u6790\u7cfb\u7edf\u6267\u884c\u8f68\u8ff9\uff08\u5982\u63a8\u7406\u8fc7\u7a0b\u3001\u5de5\u5177\u8c03\u7528\u548c\u8f93\u51fa\uff09\uff0c\u4ee5\u81ea\u7136\u8bed\u8a00\u5f62\u5f0f\u8bca\u65ad\u95ee\u9898\u5e76\u63d0\u51fa\u6539\u8fdb\u65b9\u6848\u3002<\/li>\n<li><strong>\u591a\u76ee\u6807\u4f18\u5316<\/strong>: \u91c7\u7528\u5e15\u7d2f\u6258\u6700\u4f18\u9009\u62e9\u673a\u5236\uff0c\u53ef\u4ee5\u540c\u65f6\u4f18\u5316\u591a\u4e2a\u76ee\u6807\uff08\u4f8b\u5982\uff0c\u5728\u4fdd\u8bc1\u51c6\u786e\u6027\u7684\u540c\u65f6\u7f29\u77ed\u63d0\u793a\u8bcd\u957f\u5ea6\uff09\uff0c\u5e76\u4fdd\u7559\u591a\u6837\u5316\u7684\u4f18\u826f\u5019\u9009\u65b9\u6848\u3002<\/li>\n<li><strong>\u9ad8\u6837\u672c\u6548\u7387<\/strong>: \u4e0e\u9700\u8981\u6570\u5343\u6b21\u5c1d\u8bd5\u7684\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u76f8\u6bd4\uff0cGEPA\u80fd\u7528\u6781\u5c11\u7684\u6837\u672c\uff08\u201crollouts\u201d\uff09\u5b9e\u73b0\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff0c\u6700\u591a\u53ef\u5c06\u6240\u9700\u6837\u672c\u91cf\u51cf\u5c1135\u500d\u3002<\/li>\n<li><strong>\u5e7f\u6cdb\u7684\u9002\u7528\u6027<\/strong>: \u4e0d\u4ec5\u80fd\u4f18\u5316AI\u63d0\u793a\u8bcd\uff0c\u8fd8\u80fd\u4f18\u5316\u4ee3\u7801\u3001\u6307\u4ee4\u548c\u5b8c\u6574\u7684AI\u7a0b\u5e8f\uff0c\u4f8b\u5982<code>DSPy<\/code>\u7a0b\u5e8f\u4e2d\u7684\u7b7e\u540d\u3001\u6a21\u5757\u548c\u63a7\u5236\u6d41\u3002<\/li>\n<li><strong>\u7075\u6d3b\u7684\u9002\u914d\u5668\u63a5\u53e3<\/strong>: \u901a\u8fc7\u5b9e\u73b0<code>GEPAAdapter<\/code>\u63a5\u53e3\uff0c\u7528\u6237\u53ef\u4ee5\u5c06GEPA\u96c6\u6210\u5230\u4efb\u4f55\u5305\u542b\u6587\u672c\u7ec4\u4ef6\u7684\u7cfb\u7edf\u4e2d\u3002\u7cfb\u7edf\u96c6\u6210\u7684\u6838\u5fc3\u662f\u5b9a\u4e49<code>Evaluate<\/code>\uff08\u8bc4\u4f30\uff09\u548c<code>Extract Traces for <a href=\"https:\/\/www.kdjingpai.com\/pt\/reflection-2\/\">Reflection<\/a><\/code>\uff08\u63d0\u53d6\u53cd\u601d\u8f68\u8ff9\uff09\u4e24\u4e2a\u65b9\u6cd5\u3002<\/li>\n<li><strong>\u4e0eDSPy\u6846\u67b6\u96c6\u6210<\/strong>: GEPA\u5df2\u76f4\u63a5\u96c6\u6210\u5230<code>DSPy<\/code>\u6846\u67b6\u4e2d\uff0c\u7528\u6237\u53ef\u4ee5\u901a\u8fc7<code>dspy.GEPA<\/code>\u00a0API\u8f7b\u677e\u8c03\u7528\uff0c\u8fd9\u662f\u4f7f\u7528GEPA\u6700\u7b80\u5355\u4e14\u529f\u80fd\u6700\u5f3a\u5927\u7684\u65b9\u5f0f\u3002<\/li>\n<li><strong>\u652f\u6301\u590d\u6742\u7cfb\u7edf\u4f18\u5316<\/strong>: GEPA\u80fd\u591f\u4f18\u5316\u590d\u6742\u7684AI\u7cfb\u7edf\uff0c\u4f8b\u5982\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u7cfb\u7edf\u3001\u591a\u8f6e\u5bf9\u8bdd\u667a\u80fd\u4f53\u4ee5\u53ca\u5728\u5916\u90e8\u73af\u5883\u4e2d\u8fd0\u884c\u7684\u667a\u80fd\u4f53\uff08\u5982<code>terminal-bench<\/code>\uff09\u3002<\/li>\n<\/ul>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<p>GEPA\u662f\u4e00\u4e2a\u529f\u80fd\u5f3a\u5927\u7684\u6846\u67b6\uff0c\u65e8\u5728\u901a\u8fc7\u6a21\u62df\u4eba\u7c7b\u201c\u53cd\u601d-\u6539\u8fdb\u201d\u7684\u5b66\u4e60\u6a21\u5f0f\u6765\u81ea\u52a8\u4f18\u5316AI\u7cfb\u7edf\u4e2d\u7684\u6587\u672c\u7ec4\u4ef6\uff0c\u5982\u63d0\u793a\u8bcd\u6216\u4ee3\u7801\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u7684\u4f7f\u7528\u8bf4\u660e\u3002<\/p>\n<h3>\u5b89\u88c5<\/h3>\n<p>GEPA\u53ef\u4ee5\u901a\u8fc7Python\u7684\u5305\u7ba1\u7406\u5668pip\u8f7b\u677e\u5b89\u88c5\u3002<\/p>\n<p><strong>\u7a33\u5b9a\u7248\u5b89\u88c5\uff1a<\/strong><br \/>\n\u6253\u5f00\u7ec8\u7aef\u6216\u547d\u4ee4\u884c\u5de5\u5177\uff0c\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4\uff1a<\/p>\n<pre><code>pip install gepa\r\n<\/code><\/pre>\n<p><strong>\u6700\u65b0\u5f00\u53d1\u7248\u5b89\u88c5\uff1a<\/strong><br \/>\n\u5982\u679c\u4f60\u5e0c\u671b\u4f53\u9a8c\u6700\u65b0\u7684\u529f\u80fd\uff0c\u53ef\u4ee5\u76f4\u63a5\u4eceGitHub\u4ed3\u5e93\u5b89\u88c5\uff1a<\/p>\n<pre><code>pip install git+https:\/\/github.com\/gepa-ai\/gepa.git\r\n<\/code><\/pre>\n<h3>\u6838\u5fc3\u6982\u5ff5<\/h3>\n<p>\u8981\u6709\u6548\u4f7f\u7528GEPA\uff0c\u9700\u8981\u7406\u89e3\u5176\u4e24\u4e2a\u6838\u5fc3\u6982\u5ff5\uff1a<\/p>\n<ol>\n<li><strong>\u53cd\u601d (Reflection)<\/strong>: GEPA\u7684\u6838\u5fc3\u673a\u5236\u3002\u5b83\u4e0d\u53ea\u662f\u770b\u4e00\u4e2a\u4efb\u52a1\u6700\u7ec8\u662f\u5426\u6210\u529f\uff08\u5373\u4e00\u4e2a\u7b80\u5355\u7684\u5206\u6570\uff09\uff0c\u800c\u662f\u8ba9\u4e00\u4e2a\u5f3a\u5927\u7684\u8bed\u8a00\u6a21\u578b\uff08\u79f0\u4e3a\u201c\u53cd\u601d\u6a21\u578b\u201d\uff09\u53bb\u9605\u8bfb\u6574\u4e2a\u4efb\u52a1\u7684\u6267\u884c\u8bb0\u5f55\uff08trace\uff09\u3002\u8fd9\u4e2a\u8bb0\u5f55\u5305\u542b\u4e86AI\u7684\u6240\u6709\u201c\u601d\u8003\u201d\u6b65\u9aa4\u3001\u4e2d\u95f4\u8f93\u51fa\u3001\u9047\u5230\u7684\u9519\u8bef\u7b49\u3002\u901a\u8fc7\u9605\u8bfb\u8fd9\u4e9b\u8be6\u7ec6\u7684\u8bb0\u5f55\uff0c\u53cd\u601d\u6a21\u578b\u80fd\u4ee5\u81ea\u7136\u8bed\u8a00\u7684\u5f62\u5f0f\u63d0\u51fa\u5177\u4f53\u7684\u3001\u6709\u9488\u5bf9\u6027\u7684\u6539\u8fdb\u5efa\u8bae\u3002<\/li>\n<li><strong>\u8fdb\u5316 (Evolution)<\/strong>: GEPA\u501f\u9274\u4e86\u9057\u4f20\u7b97\u6cd5\u7684\u601d\u60f3\u3002\u5b83\u4ece\u4e00\u4e2a\u521d\u59cb\u7684\u63d0\u793a\u8bcd\uff08\u201c\u79cd\u5b50\u201d\uff09\u5f00\u59cb\uff0c\u901a\u8fc7\u53cd\u601d\u751f\u6210\u4e00\u4e9b\u65b0\u7684\u3001\u53ef\u80fd\u66f4\u597d\u7684\u63d0\u793a\u8bcd\u7248\u672c\uff08\u201c\u53d8\u5f02\u201d\uff09\u3002\u7136\u540e\uff0c\u5b83\u4f1a\u6d4b\u8bd5\u8fd9\u4e9b\u65b0\u7248\u672c\uff0c\u5e76\u4fdd\u7559\u8868\u73b0\u6700\u597d\u7684\u90a3\u6279\uff08\u201c\u9009\u62e9\u201d\uff09\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u4f1a\u4e0d\u65ad\u91cd\u590d\uff0c\u6bcf\u4e00\u4ee3\u90fd\u4f1a\u5728\u524d\u4e00\u4ee3\u7684\u57fa\u7840\u4e0a\u8fdb\u884c\u4f18\u5316\uff0c\u6700\u7ec8\u6f14\u5316\u51fa\u9ad8\u6027\u80fd\u7684\u63d0\u793a\u8bcd\u3002<\/li>\n<\/ol>\n<h3>\u6700\u7b80\u5355\u7684\u4f7f\u7528\u65b9\u5f0f\uff1a\u901a\u8fc7DSPy\u6846\u67b6<\/h3>\n<p>\u5bf9\u4e8e\u5927\u591a\u6570\u7528\u6237\u6765\u8bf4\uff0c\u5c06GEPA\u4e0e<code>DSPy<\/code>\u6846\u67b6\u7ed3\u5408\u4f7f\u7528\u662f\u6700\u63a8\u8350\u7684\u65b9\u5f0f\u3002<code>DSPy<\/code>\u53ef\u4ee5\u5e2e\u52a9\u4f60\u6784\u5efa\u6a21\u5757\u5316\u7684\u8bed\u8a00\u6a21\u578b\u7a0b\u5e8f\uff0c\u800cGEPA\u5219\u4f5c\u4e3a\u4f18\u5316\u5668\u6765\u63d0\u5347\u8fd9\u4e9b\u7a0b\u5e8f\u7684\u6027\u80fd\u3002<\/p>\n<p>\u4ee5\u4e0b\u662f\u4e00\u4e2a\u4f18\u5316\u6570\u5b66\u89e3\u9898\u63d0\u793a\u8bcd\u7684\u7b80\u5355\u793a\u4f8b\uff1a<\/p>\n<p><strong>\u6b65\u9aa41\uff1a\u51c6\u5907\u73af\u5883\u548c\u6570\u636e<\/strong><br \/>\n\u786e\u4fdd\u4f60\u5df2\u7ecf\u5b89\u88c5\u4e86<code>gepa<\/code>\u548c<code>dspy-ai<\/code>\uff0c\u5e76\u8bbe\u7f6e\u4e86\u4f60\u7684OpenAI API\u5bc6\u94a5\u3002<\/p>\n<pre><code>import gepa\r\nimport dspy\r\n# \u8bbe\u7f6e\u5927\u8bed\u8a00\u6a21\u578b\r\ntask_lm = dspy.OpenAI(model='gpt-4.1-mini', max_tokens=1000)\r\n# \u8bbe\u7f6e\u4e00\u4e2a\u66f4\u5f3a\u5927\u7684\u6a21\u578b\u7528\u4e8e\u53cd\u601d\r\nreflection_lm = dspy.OpenAI(model='gpt-5', max_tokens=3500)\r\ndspy.settings.configure(lm=task_lm)\r\n# \u52a0\u8f7d\u6570\u636e\u96c6\uff08\u8fd9\u91cc\u4f7f\u7528\u5185\u7f6e\u7684AIME\u6570\u5b66\u7ade\u8d5b\u9898\u793a\u4f8b\uff09\r\ntrainset, valset, _ = gepa.examples.aime.init_dataset()\r\n<\/code><\/pre>\n<p><strong>\u6b65\u9aa42\uff1a\u5b9a\u4e49\u521d\u59cb\u7684\u7a0b\u5e8f\uff08\u6216\u63d0\u793a\u8bcd\uff09<\/strong><br \/>\n\u5728<code>DSPy<\/code>\u4e2d\uff0c\u4f60\u53ef\u4ee5\u5b9a\u4e49\u4e00\u4e2a\u7b80\u5355\u7684<code>Signature<\/code>\u6765\u63cf\u8ff0\u4efb\u52a1\u7684\u8f93\u5165\u548c\u8f93\u51fa\uff0c\u7136\u540e\u7528\u4e00\u4e2a<code>Module<\/code>\u6765\u5b9e\u73b0\u5b83\u3002<\/p>\n<pre><code>class CoT(dspy.Module):\r\ndef __init__(self):\r\nsuper().__init__()\r\nself.prog = dspy.ChainOfThought(\"problem -&gt; reasoning, answer\")\r\ndef forward(self, problem):\r\nreturn self.prog(problem=problem)\r\n<\/code><\/pre>\n<p><strong>\u6b65\u9aa43\uff1a\u5b9a\u4e49\u8bc4\u4f30\u6307\u6807<\/strong><br \/>\n\u4f60\u9700\u8981\u544a\u8bc9GEPA\u5982\u4f55\u5224\u65ad\u4e00\u4e2a\u8f93\u51fa\u7684\u597d\u574f\u3002\u8fd9\u91cc\u6211\u4eec\u5b9a\u4e49\u4e00\u4e2a\u7b80\u5355\u7684\u6307\u6807\uff0c\u68c0\u67e5\u6a21\u578b\u8f93\u51fa\u7684\u7b54\u6848\u662f\u5426\u6b63\u786e\u3002<\/p>\n<pre><code>def aime_metric(gold, pred, trace=None):\r\n# gold\u662f\u6807\u51c6\u7b54\u6848\uff0cpred\u662f\u6a21\u578b\u7684\u9884\u6d4b\u8f93\u51fa\r\nreturn gold.answer == pred.answer\r\n<\/code><\/pre>\n<p><strong>\u6b65\u9aa44\uff1a\u8fd0\u884cGEPA\u4f18\u5316\u5668<\/strong><br \/>\n\u73b0\u5728\uff0c\u4f60\u53ef\u4ee5\u914d\u7f6e\u5e76\u8fd0\u884c<code>dspy.GEPA<\/code>\u4f18\u5316\u5668\u4e86\u3002<\/p>\n<pre><code>from dspy.teleprompt import GEPA\r\n# \u914d\u7f6e\u4f18\u5316\u5668\r\n# dspy_program\u662f\u4f60\u8981\u4f18\u5316\u7684DSPy\u7a0b\u5e8f\r\n# trainset\u662f\u8bad\u7ec3\u6570\u636e\r\n# valset\u662f\u9a8c\u8bc1\u6570\u636e\r\n# metric\u662f\u8bc4\u4f30\u51fd\u6570\r\n# reflection_lm\u662f\u7528\u4e8e\u53cd\u601d\u7684\u6a21\u578b\r\noptimizer = GEPA(dspy_program=CoT(),\r\ntrainset=trainset,\r\nvalset=valset,\r\nmetric=aime_metric,\r\nreflection_lm=reflection_lm)\r\n# \u8fd0\u884c\u4f18\u5316\uff0c\u8bbe\u7f6e\u4f18\u5316\u9884\u7b97\uff08\u4f8b\u5982\uff0c\u6700\u591a\u8c03\u7528\u8bc4\u4f30\u6307\u6807150\u6b21\uff09\r\noptimized_program = optimizer.compile(max_metric_calls=150)\r\n<\/code><\/pre>\n<p>\u6267\u884c\u5b8c\u6bd5\u540e\uff0c<code>optimized_program<\/code>\u5185\u90e8\u7684\u63d0\u793a\u8bcd\u5c31\u5df2\u7ecf\u88abGEPA\u4f18\u5316\u8fc7\u4e86\u3002\u4f60\u4f1a\u53d1\u73b0\uff0c\u4f18\u5316\u540e\u7684\u63d0\u793a\u8bcd\u5305\u542b\u4e86\u975e\u5e38\u5177\u4f53\u548c\u8be6\u7ec6\u7684\u89e3\u9898\u7b56\u7565\u548c\u6ce8\u610f\u4e8b\u9879\uff0c\u8fd9\u4e9b\u90fd\u662fGEPA\u901a\u8fc7\u53cd\u601d\u5386\u53f2\u9519\u8bef\u81ea\u52a8\u5b66\u4e60\u5230\u7684\u3002<\/p>\n<h3>\u72ec\u7acb\u4f7f\u7528GEPA\uff08\u9ad8\u7ea7\u7528\u6cd5\uff09<\/h3>\n<p>\u5982\u679c\u4f60\u6ca1\u6709\u4f7f\u7528<code>DSPy<\/code>\u6846\u67b6\uff0c\u4e5f\u53ef\u4ee5\u72ec\u7acb\u4f7f\u7528GEPA\u3002\u8fd9\u65f6\uff0c\u4f60\u9700\u8981\u81ea\u5df1\u5b9e\u73b0\u4e00\u4e2a<code>GEPAAdapter<\/code>\uff0c\u4f5c\u4e3aGEPA\u4e0e\u4f60\u7684\u7cfb\u7edf\u4e4b\u95f4\u7684\u6865\u6881\u3002<\/p>\n<p><code>GEPAAdapter<\/code>\u9700\u8981\u5b9e\u73b0\u4e24\u4e2a\u5173\u952e\u65b9\u6cd5\uff1a<\/p>\n<ol>\n<li><code>Evaluate(self, candidate, trainset_sample)<\/code>:\n<ul>\n<li>\u8fd9\u4e2a\u65b9\u6cd5\u63a5\u6536GEPA\u751f\u6210\u7684\u4e00\u4e2a\u5019\u9009\u6587\u672c\u7ec4\u4ef6\uff08<code>candidate<\/code>\uff09\u548c\u4e00\u90e8\u5206\u8bad\u7ec3\u6570\u636e\uff08<code>trainset_sample<\/code>\uff09\u3002<\/li>\n<li>\u4f60\u9700\u8981\u7528\u8fd9\u4e2a\u5019\u9009\u7ec4\u4ef6\u6765\u8fd0\u884c\u4f60\u7684\u7cfb\u7edf\uff0c\u5e76\u8fd4\u56de\u7cfb\u7edf\u7684\u6267\u884c\u5f97\u5206\u548c\u8be6\u7ec6\u7684\u6267\u884c\u8f68\u8ff9\uff08traces\uff09\u3002\u8f68\u8ff9\u53ef\u4ee5\u662f\u4efb\u4f55\u6709\u52a9\u4e8e\u53cd\u601d\u7684\u6587\u672c\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<\/li>\n<li><code>ExtractTracesforReflection(self, traces, component_name)<\/code>:\n<ul>\n<li>\u8fd9\u4e2a\u65b9\u6cd5\u63a5\u6536<code>Evaluate<\/code>\u65b9\u6cd5\u8fd4\u56de\u7684\u8f68\u8ff9\uff0c\u5e76\u4ece\u4e2d\u63d0\u53d6\u4e0e\u7279\u5b9a\u7ec4\u4ef6\uff08<code>component_name<\/code>\uff09\u76f8\u5173\u7684\u90e8\u5206\u3002<\/li>\n<li>\u63d0\u53d6\u51fa\u7684\u6587\u672c\u5c06\u4ea4\u7ed9\u53cd\u601d\u6a21\u578b\u8fdb\u884c\u5206\u6790\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p>\u8fd9\u662f\u4e00\u4e2a\u6982\u5ff5\u6027\u7684\u793a\u4f8b\u7ed3\u6784\uff1a<\/p>\n<pre><code>from gepa.core import GEPAAdapter\r\nclass MyCustomAdapter(GEPAAdapter):\r\ndef Evaluate(self, candidate, trainset_sample):\r\n# \u4f60\u7684\u7cfb\u7edf\u903b\u8f91\uff1a\u4f7f\u7528candidate\u4e2d\u7684\u63d0\u793a\u8bcd\u5904\u7406trainset_sample\u4e2d\u7684\u6570\u636e\r\n# ...\r\nscores = [...]  # \u8ba1\u7b97\u5f97\u5206\r\ntraces = [...]  # \u6536\u96c6\u8be6\u7ec6\u7684\u65e5\u5fd7\u6216\u4e2d\u95f4\u6b65\u9aa4\r\nreturn scores, traces\r\ndef ExtractTracesforReflection(self, traces, component_name):\r\n# \u4ecetraces\u4e2d\u63d0\u53d6\u548ccomponent_name\u76f8\u5173\u7684\u6587\u672c\u4fe1\u606f\r\n# ...\r\nreturn relevant_textual_traces\r\n# \u7136\u540e\u8c03\u7528gepa.optimize\r\ngepa_result = gepa.optimize(\r\nseed_candidate={\"my_prompt\": \"Initial prompt here...\"},\r\nadapter=MyCustomAdapter(),\r\ntrainset=my_train_data,\r\nvalset=my_val_data,\r\n# ... \u5176\u4ed6\u53c2\u6570\r\n)\r\n<\/code><\/pre>\n<p>\u8fd9\u79cd\u65b9\u5f0f\u867d\u7136\u66f4\u590d\u6742\uff0c\u4f46\u5b83\u63d0\u4f9b\u4e86\u6781\u5927\u7684\u7075\u6d3b\u6027\uff0c\u8ba9GEPA\u53ef\u4ee5\u7528\u4e8e\u4f18\u5316\u4efb\u4f55\u57fa\u4e8e\u6587\u672c\u7684\u7cfb\u7edf\u3002<\/p>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u590d\u6742\u63a8\u7406\u4efb\u52a1\u63d0\u793a\u8bcd\u4f18\u5316<\/strong><br \/>\n\u5bf9\u4e8e\u9700\u8981\u591a\u6b65\u63a8\u7406\u7684\u590d\u6742\u4efb\u52a1\uff08\u5982\u6570\u5b66\u3001\u903b\u8f91\u548c\u7b56\u7565\u89c4\u5212\uff09\uff0c\u4e00\u4e2a\u5fae\u5c0f\u7684\u63d0\u793a\u8bcd\u6539\u52a8\u5c31\u53ef\u80fd\u5bfc\u81f4\u7ed3\u679c\u7684\u5de8\u5927\u5dee\u5f02\u3002GEPA\u80fd\u591f\u901a\u8fc7\u5206\u6790\u6a21\u578b\u7684\u63a8\u7406\u94fe\u6761\uff0c\u81ea\u52a8\u53d1\u73b0\u5e76\u7ea0\u6b63\u5176\u4e2d\u7684\u903b\u8f91\u7f3a\u9677\uff0c\u751f\u6210\u9ad8\u5ea6\u4f18\u5316\u7684\u6307\u4ee4\uff0c\u5f15\u5bfc\u6a21\u578b\u91c7\u7528\u66f4\u6709\u6548\u7684\u89e3\u9898\u7b56\u7565\u3002<\/li>\n<li><strong>\u4ee3\u7801\u751f\u6210\u4e0e\u4f18\u5316<\/strong><br \/>\nGEPA\u4e0d\u4ec5\u53ef\u4ee5\u751f\u6210\u4ee3\u7801\uff0c\u8fd8\u80fd\u6839\u636e\u7f16\u8bd1\u9519\u8bef\u3001\u6027\u80fd\u5206\u6790\u62a5\u544a\u6216\u4ee3\u7801\u5ba1\u67e5\u6ce8\u91ca\u7b49\u6587\u672c\u53cd\u9988\u6765\u4f18\u5316\u4ee3\u7801\u3002\u4f8b\u5982\uff0c\u5b83\u53ef\u4ee5\u5c06\u4e00\u4e2a\u901a\u7528\u7684\u4ee3\u7801\u7247\u6bb5\uff0c\u6839\u636e\u7279\u5b9a\u786c\u4ef6\uff08\u5982GPU\uff09\u7684\u6587\u6863\u548c\u9519\u8bef\u4fe1\u606f\uff0c\u8fed\u4ee3\u4fee\u6539\u6210\u4e00\u4e2a\u9ad8\u5ea6\u4f18\u5316\u7684\u7248\u672c\u3002<\/li>\n<li><strong>\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u7cfb\u7edf\u8c03\u4f18<\/strong><br \/>\nRAG\u7cfb\u7edf\u5305\u542b\u591a\u4e2a\u73af\u8282\uff08\u67e5\u8be2\u91cd\u6784\u3001\u6587\u6863\u68c0\u7d22\u3001\u7b54\u6848\u5408\u6210\u7b49\uff09\uff0c\u6bcf\u4e2a\u73af\u8282\u90fd\u7531\u63d0\u793a\u8bcd\u9a71\u52a8\u3002GEPA\u53ef\u4ee5\u540c\u65f6\u4f18\u5316\u6240\u6709\u8fd9\u4e9b\u63d0\u793a\u8bcd\uff0c\u901a\u8fc7\u5206\u6790\u6574\u4e2aRAG\u7cfb\u7edf\u7684\u6267\u884c\u8f68\u8ff9\uff0c\u63d0\u5347\u68c0\u7d22\u7684\u7cbe\u51c6\u5ea6\u548c\u7b54\u6848\u7684\u8d28\u91cf\u3002<\/li>\n<li><strong>\u667a\u80fd\u4f53\uff08Agent\uff09\u884c\u4e3a\u6307\u4ee4\u5fae\u8c03<\/strong><br \/>\n\u5bf9\u4e8e\u9700\u8981\u4e0e\u5916\u90e8\u5de5\u5177\u6216\u73af\u5883\u4ea4\u4e92\u7684\u667a\u80fd\u4f53\uff0cGEPA\u53ef\u4ee5\u901a\u8fc7\u5206\u6790\u667a\u80fd\u4f53\u7684\u884c\u4e3a\u65e5\u5fd7\uff08\u5305\u62ecAPI\u8c03\u7528\u3001\u5de5\u5177\u8fd4\u56de\u7ed3\u679c\u548c\u73af\u5883\u53cd\u9988\uff09\uff0c\u4f18\u5316\u5176\u6838\u5fc3\u6307\u4ee4\uff08\u5373\u7cfb\u7edf\u63d0\u793a\u8bcd\uff09\uff0c\u8ba9\u667a\u80fd\u4f53\u66f4\u9ad8\u6548\u3001\u66f4\u53ef\u9760\u5730\u5b8c\u6210\u4efb\u52a1\u3002<\/li>\n<li><strong>\u7279\u5b9a\u9886\u57df\u77e5\u8bc6\u7684\u6307\u4ee4\u5b66\u4e60<\/strong><br \/>\n\u5728\u4e13\u4e1a\u9886\u57df\uff08\u5982\u533b\u7597\u3001\u6cd5\u5f8b\u3001\u91d1\u878d\uff09\uff0cAI\u7cfb\u7edf\u9700\u8981\u4e25\u683c\u9075\u5faa\u7279\u5b9a\u7684\u6307\u5357\u548c\u89c4\u8303\u3002GEPA\u53ef\u4ee5\u5c06\u8fd9\u4e9b\u6307\u5357\u6587\u6863\u4f5c\u4e3a\u53cd\u601d\u7684\u4f9d\u636e\uff0c\u5f53\u7cfb\u7edf\u8f93\u51fa\u4e0d\u7b26\u5408\u89c4\u8303\u65f6\uff0cGEPA\u80fd\u81ea\u52a8\u5c06\u76f8\u5173\u89c4\u5219\u878d\u5165\u5230\u63d0\u793a\u8bcd\u4e2d\uff0c\u4f7f\u7cfb\u7edf\u8f93\u51fa\u66f4\u5408\u89c4\u3002<\/li>\n<\/ol>\n<h2>QA<\/h2>\n<ol>\n<li><strong>GEPA\u4e0e\u4f20\u7edf\u7684\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u4f18\u5316\u65b9\u6cd5\u6709\u4f55\u4e0d\u540c\uff1f<\/strong><br \/>\n\u4e3b\u8981\u533a\u522b\u5728\u4e8e\u5b66\u4e60\u4fe1\u53f7\u7684\u4e30\u5bcc\u7a0b\u5ea6\u3002\u4f20\u7edf\u7684RL\u65b9\u6cd5\u901a\u5e38\u4f9d\u8d56\u4e00\u4e2a\u5355\u4e00\u7684\u3001\u7a00\u758f\u7684\u5956\u52b1\u5206\u6570\uff08\u6bd4\u5982\u4efb\u52a1\u6210\u529f\u5f971\u5206\uff0c\u5931\u8d25\u5f970\u5206\uff09\uff0c\u6a21\u578b\u9700\u8981\u5927\u91cf\u5c1d\u8bd5\u624d\u80fd\u5b66\u5230\u6709\u6548\u7684\u7b56\u7565\u3002\u800cGEPA\u5229\u7528\u7684\u662f\u4e30\u5bcc\u7684\u81ea\u7136\u8bed\u8a00\u53cd\u9988\uff0c\u901a\u8fc7LLM\u201c\u9605\u8bfb\u201d\u8be6\u7ec6\u7684\u6267\u884c\u8fc7\u7a0b\u8bb0\u5f55\u6765\u7406\u89e3\u5931\u8d25\u7684\u5177\u4f53\u539f\u56e0\uff0c\u4ece\u800c\u80fd\u7528\u66f4\u5c11\u7684\u6837\u672c\u505a\u51fa\u66f4\u7cbe\u786e\u7684\u6539\u8fdb\u3002<\/li>\n<li><strong>\u4f7f\u7528GEPA\u662f\u5426\u9700\u8981\u975e\u5e38\u5f3a\u5927\u7684\u8bed\u8a00\u6a21\u578b\uff1f<\/strong><br \/>\nGEPA\u7684\u8bbe\u8ba1\u4e2d\u5305\u542b\u4e24\u79cd\u6a21\u578b\uff1a\u4e00\u4e2a\u662f\u88ab\u4f18\u5316\u7684\u201c\u4efb\u52a1\u6a21\u578b\u201d\uff0c\u53e6\u4e00\u4e2a\u662f\u8fdb\u884c\u5206\u6790\u7684\u201c\u53cd\u601d\u6a21\u578b\u201d\u3002\u901a\u5e38\u5efa\u8bae\u4f7f\u7528\u4e00\u4e2a\u80fd\u529b\u5c3d\u53ef\u80fd\u5f3a\u7684\u6a21\u578b\u4f5c\u4e3a\u201c\u53cd\u601d\u6a21\u578b\u201d\uff08\u5982GPT-4\u6216\u66f4\u9ad8\u7ea7\u7684\u6a21\u578b\uff09\uff0c\u56e0\u4e3a\u5b83\u9700\u8981\u6df1\u523b\u7406\u89e3\u590d\u6742\u7684\u6267\u884c\u8f68\u8ff9\u548c\u4e0a\u4e0b\u6587\u3002\u800c\u88ab\u4f18\u5316\u7684\u201c\u4efb\u52a1\u6a21\u578b\u201d\u5219\u53ef\u4ee5\u662f\u4efb\u4f55\u4f60\u9700\u8981\u63d0\u5347\u6027\u80fd\u7684\u6a21\u578b\uff0c\u5305\u62ec\u4e00\u4e9b\u66f4\u5c0f\u3001\u66f4\u7ecf\u6d4e\u7684\u6a21\u578b\u3002<\/li>\n<li><strong>GEPA\u4e2d\u7684\u201cPareto\u201d\uff08\u5e15\u7d2f\u6258\uff09\u662f\u4ec0\u4e48\u610f\u601d\uff1f<\/strong><br \/>\n\u201c\u5e15\u7d2f\u6258\u201d\u6765\u6e90\u4e8e\u5e15\u7d2f\u6258\u6700\u4f18\u7684\u6982\u5ff5\uff0c\u7528\u4e8e\u591a\u76ee\u6807\u4f18\u5316\u3002\u5728GEPA\u4e2d\uff0c\u8fd9\u610f\u5473\u7740\u4f18\u5316\u8fc7\u7a0b\u4e0d\u4ec5\u4ec5\u8ffd\u6c42\u5355\u4e00\u6307\u6807\u7684\u6700\u9ad8\u5206\uff08\u5982\u51c6\u786e\u7387\uff09\uff0c\u5b83\u8fd8\u53ef\u4ee5\u540c\u65f6\u8003\u8651\u5176\u4ed6\u76ee\u6807\uff0c\u6bd4\u5982\u63d0\u793a\u8bcd\u7684\u957f\u5ea6\u3001API\u8c03\u7528\u6210\u672c\u6216\u54cd\u5e94\u5ef6\u8fdf\u3002GEPA\u4f1a\u4fdd\u7559\u4e00\u4e2a\u201c\u5e15\u7d2f\u6258\u524d\u6cbf\u201d\uff0c\u5373\u4e00\u7ec4\u5728\u4e0d\u540c\u76ee\u6807\u4e0a\u53d6\u5f97\u826f\u597d\u5e73\u8861\u7684\u5019\u9009\u65b9\u6848\uff0c\u800c\u4e0d\u662f\u4ec5\u4ec5\u4fdd\u7559\u4e00\u4e2a\u5355\u4e00\u7684\u201c\u6700\u4f73\u201d\u65b9\u6848\u3002<\/li>\n<li><strong>GEPA\u662f\u5426\u53ea\u80fd\u4f18\u5316\u82f1\u6587\u63d0\u793a\u8bcd\uff1f<\/strong><br \/>\n\u4e0d\u662f\u3002GEPA\u7684\u5e95\u5c42\u673a\u5236\u662f\u57fa\u4e8e\u8bed\u8a00\u6a21\u578b\u5bf9\u6587\u672c\u7684\u7406\u89e3\u548c\u751f\u6210\u80fd\u529b\uff0c\u56e0\u6b64\u5b83\u5929\u7136\u652f\u6301\u591a\u8bed\u8a00\u3002\u53ea\u8981\u4f60\u63d0\u4f9b\u7684\u8bad\u7ec3\u6570\u636e\u3001\u8bc4\u4f30\u6307\u6807\u548c\u53cd\u601d\u6a21\u578b\u652f\u6301\u76f8\u5e94\u7684\u8bed\u8a00\uff08\u4f8b\u5982\u4e2d\u6587\uff09\uff0cGEPA\u5c31\u53ef\u4ee5\u7528\u6765\u4f18\u5316\u8be5\u8bed\u8a00\u7684\u6587\u672c\u7ec4\u4ef6\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>GEPA (Genetic-Pareto) \u662f\u4e00\u4e2a\u7528\u4e8e\u4f18\u5316AI\u7cfb\u7edf\u4e2d\u5404\u7c7b\u6587\u672c\u7ec4\u4ef6\u7684\u6846\u67b6\u3002\u8fd9\u4e9b\u6587\u672c\u7ec4\u4ef6\u53ef\u4ee5AI\u6a21\u578b\u7684\u63d0\u793a\u8bcd\u3001\u4ee3\u7801\u7247\u6bb5\u6216\u914d\u7f6e\u6587\u4ef6\u3002\u5b83\u91c7\u7528\u4e86\u4e00\u79cd\u540d\u4e3a\u201c\u53cd\u601d\u6027\u6587\u672c\u8fdb\u5316\u201d\u7684\u65b9\u6cd5\uff0c\u901a\u8fc7\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u6765\u5206\u6790\u548c\u53cd\u601dAI\u7cfb\u7edf\u7684\u884c\u4e3a\u3002\u5177&#8230;<\/p>\n","protected":false},"author":1,"featured_media":32782,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20,501],"tags":[230,227],"class_list":["post-60125","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","category-prompt-aids","tag-aikaiyuanxiangmu","tag-promptstishizhilinga"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/60125","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=60125"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/60125\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/32782"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=60125"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=60125"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=60125"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}