{"id":4905,"date":"2024-10-18T17:58:07","date_gmt":"2024-10-18T09:58:07","guid":{"rendered":"https:\/\/nullthought.net\/?p=4905"},"modified":"2024-11-07T11:02:56","modified_gmt":"2024-11-07T03:02:56","slug":"%e6%a8%a1%e5%9e%8b%e7%be%a4%ef%bc%88model-swarms%ef%bc%89%ef%bc%8c%e7%81%b5%e6%b4%bb%e7%bb%84%e5%90%88%e5%a4%9a%e4%b8%aallm%ef%bc%88%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%ef%bc%89%e4%b8%93","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=4905","title":{"rendered":"\u6a21\u578b\u7fa4\uff08Model Swarms\uff09\uff0c\u7075\u6d3b\u7ec4\u5408\u591a\u4e2aLLM\uff08\u5927\u8bed\u8a00\u6a21\u578b\uff09\u4e13\u5bb6\u7684\u7fa4\u4f53\u667a\u80fd"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2410.11163\" target=\"_blank\" rel=\"noreferrer noopener\">Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence<\/a><\/strong>\uff08\u300aModel Swarms\uff1a\u901a\u8fc7\u7fa4\u4f53\u667a\u80fd\u5bf9LLM\u4e13\u5bb6\u8fdb\u884c\u534f\u4f5c\u641c\u7d22\u9002\u5e94\u300b\uff09\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u7fa4\u4f53\u667a\u80fd\u534f\u4f5c\u641c\u7d22\u7b97\u6cd5\u6765\u4f18\u5316\u548c\u9002\u5e94\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4e13\u5bb6\u7fa4\u4f53\u3002\u8be5\u65b9\u6cd5\u57fa\u4e8e\u7c92\u5b50\u7fa4\u4f18\u5316\uff08PSO\uff09\u7684\u601d\u60f3\uff0c\u5c06\u591a\u4e2aLLM\u4e13\u5bb6\u89c6\u4e3a\u201c\u7c92\u5b50\u201d\uff0c\u5728\u6743\u91cd\u7a7a\u95f4\u4e2d\u8fdb\u884c\u534f\u4f5c\u641c\u7d22\uff0c\u4ee5\u627e\u5230\u9002\u5408\u4e0d\u540c\u4efb\u52a1\u6216\u9886\u57df\u7684\u6700\u4f73\u6a21\u578b\u7ec4\u5408\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aShangbin Feng, Zifeng Wang, Yike Wang, Sayna Ebrahimi, Hamid Palangi, Lesly Miculicich, Achin Kulshrestha, Nathalie Rauschmayr, Yejin Choi, Yulia Tsvetkov, Chen-Yu Lee, Tomas Pfister\uff0c\u6765\u81eaUniversity of Washington\u548cGoogle\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"762\" height=\"235\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-14.png\" alt=\"\" class=\"wp-image-4906\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-14.png 762w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-14-300x93.png 300w\" sizes=\"auto, (max-width: 762px) 100vw, 762px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2410.11163\" target=\"_blank\" rel=\"noreferrer noopener\">Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u5982\u4e0b\u4e3a\u8bba\u6587\u6982\u8981\u5185\u5bb9\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e00\u3001\u8bba\u6587\u80cc\u666f\u4e0e\u52a8\u673a<\/h3>\n\n\n\n<p>\u76ee\u524d\u7684\u7814\u7a76\u503e\u5411\u4e8e\u5f00\u53d1\u5927\u89c4\u6a21\u7684\u901a\u7528\u578b\u8bed\u8a00\u6a21\u578b\uff08\u5982GPT-3\uff09\uff0c\u8fd9\u79cd\u6a21\u578b\u901a\u5e38\u901a\u8fc7\u5171\u4eab\u53c2\u6570\u6765\u5e94\u5bf9\u6240\u6709\u8bed\u8a00\u548c\u4efb\u52a1\u3002\u7136\u800c\uff0c\u901a\u7528\u6a21\u578b\u9762\u4e34\u7684\u6311\u6218\u662f\u5176\u5728\u7279\u5b9a\u4efb\u52a1\u548c\u9886\u57df\u4e2d\u8868\u73b0\u7684\u5c40\u9650\u6027\uff0c\u5c24\u5176\u662f\u5728\u6570\u636e\u91cf\u8f83\u5c11\u7684\u60c5\u51b5\u4e0b\u3002\u56e0\u6b64\uff0c\u8fd1\u5e74\u6765\u7684\u7814\u7a76\u9010\u6e10\u8f6c\u5411\u6a21\u5757\u5316\u7684\u591a\u6a21\u578b\u534f\u4f5c\uff0c\u5373\u591a\u4e2a\u4e13\u5bb6\u6a21\u578b\u76f8\u4e92\u4ea4\u4e92\u4e0e\u8865\u5145\uff0c\u4ee5\u53d1\u6325\u5404\u81ea\u7684\u7279\u957f\u3002\u8fd9\u79cd\u65b9\u6cd5\u5305\u62ec\u4e13\u5bb6\u6df7\u5408\uff08Mixture-of-Experts\uff0cMoE\uff09\u7b49\u65b9\u6cd5\uff0c\u901a\u8fc7\u4e0d\u540c\u9886\u57df\u4e13\u5bb6\u7684\u5408\u4f5c\u6765\u63d0\u9ad8\u4efb\u52a1\u7684\u5904\u7406\u80fd\u529b\uff0c\u4f46\u8fd9\u4e9b\u65b9\u6cd5\u901a\u5e38\u5728\u6a21\u578b\u7ec4\u6210\u65b9\u5f0f\u4e0a\u6709\u5f88\u5f3a\u7684\u5047\u8bbe\u548c\u9650\u5236\uff0c\u7f3a\u4e4f\u7075\u6d3b\u6027\uff0c\u5c24\u5176\u662f\u5f53\u9762\u5bf9\u672a\u77e5\u9886\u57df\u6216\u590d\u6742\u4efb\u52a1\u65f6\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u534f\u4f5c\u641c\u7d22\u7b97\u6cd5\u2014\u2014<strong>Model Swarms<\/strong>\uff0c\u8be5\u7b97\u6cd5\u57fa\u4e8e\u7fa4\u4f53\u667a\u80fd\uff0c\u901a\u8fc7\u5728\u6743\u91cd\u7a7a\u95f4\u4e2d\u7684\u534f\u4f5c\u641c\u7d22\u5b9e\u73b0\u6a21\u578b\u7684\u9002\u5e94\u4e0e\u4f18\u5316\u3002Model Swarms \u7684\u76ee\u6807\u662f\u901a\u8fc7\u7075\u6d3b\u7ec4\u5408\u591a\u4e2aLLM\uff08\u5927\u8bed\u8a00\u6a21\u578b\uff09\u4e13\u5bb6\uff0c\u4f7f\u5b83\u4eec\u80fd\u9002\u5e94\u591a\u6837\u5316\u7684\u4efb\u52a1\u548c\u7528\u6237\u9700\u6c42\uff0c\u7279\u522b\u662f\u5728\u4f4e\u6570\u636e\u91cf\u7684\u573a\u666f\u4e0b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e8c\u3001Model Swarms \u7684\u6838\u5fc3\u601d\u60f3\u4e0e\u673a\u5236<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u7fa4\u4f53\u667a\u80fd\u4e0e\u7c92\u5b50\u7fa4\u4f18\u5316\uff08PSO\uff09\u7684\u542f\u53d1<\/h4>\n\n\n\n<p>Model Swarms \u7684\u8bbe\u8ba1\u7075\u611f\u6765\u6e90\u4e8e\u7fa4\u4f53\u667a\u80fd\uff0c\u5c24\u5176\u662f\u7c92\u5b50\u7fa4\u4f18\u5316\uff08Particle Swarm Optimization, PSO\uff09\u3002PSO\u662f\u4e00\u79cd\u57fa\u4e8e\u81ea\u7136\u754c\u7fa4\u4f53\u884c\u4e3a\u7684\u4f18\u5316\u7b97\u6cd5\uff0c\u5982\u9e1f\u7fa4\u89c5\u98df\u3001\u9c7c\u7fa4\u6e38\u52a8\u7b49\u3002PSO\u4e2d\uff0c\u6240\u6709\u7c92\u5b50\uff08\u5373\u5019\u9009\u89e3\uff09\u5728\u591a\u7ef4\u7a7a\u95f4\u4e2d\u534f\u4f5c\u5bfb\u627e\u6700\u4f18\u89e3\uff0c\u6bcf\u4e2a\u7c92\u5b50\u6839\u636e\u81ea\u8eab\u548c\u7fa4\u4f53\u7684\u7ecf\u9a8c\u8c03\u6574\u81ea\u5df1\u7684\u4f4d\u7f6e\uff0c\u4ee5\u6b64\u8fbe\u5230\u5168\u5c40\u4f18\u5316\u3002<\/p>\n\n\n\n<p>\u5728 Model Swarms \u4e2d\uff0c\u591a\u4e2aLLM\u4e13\u5bb6\u88ab\u89c6\u4e3a\u7fa4\u4f53\u4e2d\u7684\u7c92\u5b50\uff0c\u6bcf\u4e2a\u7c92\u5b50\u90fd\u6709\u81ea\u5df1\u7684\u201c\u4f4d\u7f6e\u201d\uff08\u6a21\u578b\u6743\u91cd\uff09\u548c\u201c\u901f\u5ea6\u201d\uff08\u5728\u6743\u91cd\u7a7a\u95f4\u4e2d\u7684\u79fb\u52a8\u65b9\u5411\uff09\u3002\u8fd9\u4e9b\u4e13\u5bb6\u6a21\u578b\u901a\u8fc7\u534f\u4f5c\u8c03\u6574\u5176\u6743\u91cd\u548c\u53c2\u6570\uff0c\u6700\u7ec8\u4f18\u5316\u4e00\u4e2a\u6548\u7528\u51fd\u6570\uff0c\u8be5\u6548\u7528\u51fd\u6570\u53ef\u4ee5\u4ee3\u8868\u4efb\u52a1\u7684\u5177\u4f53\u76ee\u6807\uff0c\u5982\u6a21\u578b\u7684\u51c6\u786e\u7387\u6216\u5bf9\u7279\u5b9a\u7528\u6237\u504f\u597d\u7684\u9002\u5e94\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. \u7b97\u6cd5\u7684\u57fa\u672c\u6d41\u7a0b<\/h4>\n\n\n\n<h5 class=\"wp-block-heading\">\uff081\uff09\u521d\u59cb\u7fa4\u4f53\u6784\u5efa<\/h5>\n\n\n\n<p>Model Swarms \u5f00\u59cb\u65f6\u9700\u8981\u4e00\u4e2a\u521d\u59cb\u7684LLM\u4e13\u5bb6\u7fa4\u4f53\uff0c\u8fd9\u4e9b\u4e13\u5bb6\u53ef\u4ee5\u662f\u5b8c\u6574\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u6216\u8005\u901a\u8fc7LoRA\uff08\u4f4e\u79e9\u9002\u914d\u5668\uff09\u5728\u4e0d\u540c\u4efb\u52a1\u4e0a\u5fae\u8c03\u8fc7\u7684\u6a21\u578b\u3002\u521d\u59cb\u4e13\u5bb6\u7fa4\u4f53\u901a\u8fc7\u4e24\u4e24\u63d2\u503c\uff08pairwise interpolation\uff09\u6269\u5c55\u4e3a\u66f4\u5927\u7684\u6a21\u578b\u96c6\u5408\uff0c\u4ee5\u589e\u52a0\u641c\u7d22\u8fc7\u7a0b\u4e2d\u7684\u591a\u6837\u6027\u548c\u63a2\u7d22\u80fd\u529b\u3002\u5177\u4f53\u5730\u8bf4\uff0c\u901a\u8fc7\u968f\u673a\u9009\u62e9\u4e24\u4e2a\u4e13\u5bb6\u5e76\u5bf9\u5176\u6743\u91cd\u8fdb\u884c\u7ebf\u6027\u7ec4\u5408\u6765\u751f\u6210\u65b0\u7684\u521d\u59cb\u6a21\u578b\uff0c\u4f7f\u5f97\u641c\u7d22\u8fc7\u7a0b\u4ece\u66f4\u591a\u6837\u5316\u7684\u8d77\u70b9\u5f00\u59cb\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\uff082\uff09\u901f\u5ea6\u548c\u4f4d\u7f6e\u66f4\u65b0<\/h5>\n\n\n\n<p>\u6bcf\u4e2a\u7c92\u5b50\u6a21\u578b\u7684\u79fb\u52a8\u7531\u5176\u901f\u5ea6\u51b3\u5b9a\uff0c\u800c\u901f\u5ea6\u7684\u66f4\u65b0\u53d7\u5230\u4e09\u4e2a\u56e0\u7d20\u7684\u5f71\u54cd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u60ef\u6027\u9879<\/strong>\uff1a\u4fdd\u6301\u7c92\u5b50\u5f53\u524d\u7684\u901f\u5ea6\uff0c\u786e\u4fdd\u5176\u6cbf\u65e2\u5b9a\u65b9\u5411\u524d\u8fdb\uff0c\u63a2\u7d22\u672a\u77e5\u533a\u57df\u3002<\/li>\n\n\n\n<li><strong>\u4e2a\u4f53\u6700\u4f18\uff08Personal Best, p_i\uff09<\/strong>\uff1a\u6bcf\u4e2a\u7c92\u5b50\u6839\u636e\u5176\u641c\u7d22\u5386\u53f2\u4e2d\u53d1\u73b0\u7684\u6700\u4f73\u4f4d\u7f6e\u8fdb\u884c\u8c03\u6574\uff0c\u671d\u5411\u5386\u53f2\u6700\u4f73\u7684\u4f4d\u7f6e\u79fb\u52a8\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63a2\u7d22\u6709\u524d\u666f\u7684\u533a\u57df\u3002<\/li>\n\n\n\n<li><strong>\u5168\u5c40\u6700\u4f18\uff08Global Best, g\uff09\u548c\u5168\u5c40\u6700\u5dee\uff08Global Worst, g_w\uff09<\/strong>\uff1a\u7c92\u5b50\u4e5f\u4f1a\u53d7\u5230\u7fa4\u4f53\u6574\u4f53\u6700\u4f18\u4f4d\u7f6e\u7684\u5438\u5f15\u548c\u6700\u5dee\u4f4d\u7f6e\u7684\u6392\u65a5\uff0c\u4ece\u800c\u4fc3\u8fdb\u63a2\u7d22\u6709\u6f5c\u529b\u7684\u533a\u57df\u5e76\u8fdc\u79bb\u52a3\u8d28\u6a21\u578b\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u901f\u5ea6\u7684\u66f4\u65b0\u516c\u5f0f\u4e3a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"147\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-15.png\" alt=\"\" class=\"wp-image-4907\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-15.png 865w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-15-300x51.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-15-768x131.png 768w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\">\uff083\uff09\u6743\u91cd\u66f4\u65b0<\/h5>\n\n\n\n<p>\u6839\u636e\u66f4\u65b0\u540e\u7684\u901f\u5ea6\uff0c\u7c92\u5b50\u7684\u6743\u91cd\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u66f4\u65b0\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"204\" height=\"60\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-17.png\" alt=\"\" class=\"wp-image-4908\"\/><\/figure>\n\n\n\n<p>\u5176\u4e2d\uff0c\u03bb\u662f\u6b65\u957f\u53c2\u6570\uff0c\u63a7\u5236\u6bcf\u6b21\u79fb\u52a8\u7684\u5e45\u5ea6\u3002\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\uff0c\u6a21\u578b\u6839\u636e\u6548\u7528\u51fd\u6570f\u7684\u503c\u6765\u66f4\u65b0\u5168\u5c40\u548c\u4e2a\u4eba\u6700\u4f18\u7684\u4f4d\u7f6e\u3002\u5982\u679c\u67d0\u4e2a\u7c92\u5b50\u5728\u8fde\u7eed\u82e5\u5e72\u6b21\u8fed\u4ee3\u4e2d\u672a\u80fd\u627e\u5230\u66f4\u4f18\u7684\u4f4d\u7f6e\uff0c\u5219\u8be5\u7c92\u5b50\u4f1a\u88ab\u201c\u91cd\u542f\u201d\u56de\u5176\u4e2a\u4eba\u6700\u4f73\u4f4d\u7f6e\uff0c\u5e76\u5c06\u901f\u5ea6\u8bbe\u4e3a0\uff0c\u7ed9\u4e88\u5176\u65b0\u7684\u63a2\u7d22\u673a\u4f1a\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\uff084\uff09\u534f\u4f5c\u641c\u7d22\u7ec8\u6b62\u6761\u4ef6<\/h5>\n\n\n\n<p>\u534f\u4f5c\u641c\u7d22\u5728\u6ee1\u8db3\u4ee5\u4e0b\u6761\u4ef6\u4e4b\u4e00\u65f6\u7ec8\u6b62\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5168\u5c40\u6700\u4f18\u4f4d\u7f6e\u5728 c \u6b21\u8fed\u4ee3\u4e2d\u672a\u53d1\u751f\u53d8\u5316\u3002<\/li>\n\n\n\n<li>\u8fbe\u5230\u6700\u5927\u8fed\u4ee3\u6b21\u6570 K\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u641c\u7d22\u7ed3\u675f\u65f6\uff0c\u5168\u5c40\u6700\u4f18\u7684\u6a21\u578b\u4f4d\u7f6e\u88ab\u4f5c\u4e3a\u6700\u7ec8\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e09\u3001\u5b9e\u9a8c\u8bbe\u7f6e\u4e0e\u7ed3\u679c\u5206\u6790<\/h3>\n\n\n\n<p>\u8bba\u6587\u7684\u5b9e\u9a8c\u6db5\u76d6\u4e86\u56db\u79cd\u7c7b\u578b\u7684LLM\u9002\u5e94\u4efb\u52a1\uff1a\u5355\u4efb\u52a1\u9002\u5e94\u3001\u591a\u4efb\u52a1\u9886\u57df\u9002\u5e94\u3001\u5956\u52b1\u6a21\u578b\u4f18\u5316\u3001\u4ee5\u53ca\u4eba\u7c7b\u5174\u8da3\u9002\u5e94\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u5355\u4efb\u52a1\u9002\u5e94<\/h4>\n\n\n\n<p>\u5728\u5355\u4efb\u52a1\u9002\u5e94\u4e2d\uff0cModel Swarms \u88ab\u5e94\u7528\u4e8e9\u4e2a\u6570\u636e\u96c6\uff0c\u5305\u62ec\u77e5\u8bc6\u7c7b\uff08\u5982MMLU\uff09\u3001\u63a8\u7406\u7c7b\uff08\u5982GSM8k\uff09\u4ee5\u53ca\u5b89\u5168\u6027\u7c7b\u4efb\u52a1\uff08\u5982TruthfulQA\uff09\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cModel Swarms\u76f8\u8f83\u4e8e12\u4e2a\u57fa\u51c6\u65b9\u6cd5\uff0c\u5e73\u5747\u6027\u80fd\u63d0\u5347\u4e8613.3%\uff0c\u5c24\u5176\u662f\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\uff0c\u8868\u73b0\u63d0\u5347\u8fbe\u5230\u4e8621.0%\u3002\u8fd9\u8868\u660e\u8be5\u65b9\u6cd5\u5728\u63a8\u7406\u5bc6\u96c6\u578b\u7684\u4efb\u52a1\u4e2d\u6709\u663e\u8457\u7684\u4f18\u52bf\uff0c\u53ef\u80fd\u56e0\u4e3a\u534f\u4f5c\u641c\u7d22\u589e\u5f3a\u4e86\u4e13\u5bb6\u6a21\u578b\u5728\u590d\u6742\u903b\u8f91\u63a8\u7406\u4e0a\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. \u591a\u4efb\u52a1\u9886\u57df\u9002\u5e94<\/h4>\n\n\n\n<p>\u5728\u591a\u4efb\u52a1\u9886\u57df\u9002\u5e94\u4e2d\uff0cModel Swarms \u88ab\u7528\u4e8e\u540c\u65f6\u4f18\u5316\u533b\u7597\u3001\u6cd5\u5f8b\u3001\u79d1\u5b66\u548c\u6587\u5316\u7b49\u9886\u57df\u4e2d\u7684\u591a\u4e2a\u4efb\u52a1\u3002\u7ed3\u679c\u663e\u793a\uff0c\u5728\u6240\u6709\u6d4b\u8bd5\u4efb\u52a1\u4e2d\uff0cModel Swarms \u7684\u5e73\u5747\u6027\u80fd\u63d0\u5347\u4e865.7%\uff0c\u7279\u522b\u662f\u5728\u6cd5\u5f8b\u9886\u57df\uff0c\u8868\u73b0\u63d0\u5347\u6700\u4e3a\u663e\u8457\uff08\u5206\u522b\u8fbe\u5230\u4e8611.3%\u548c10.5%\uff09\u3002\u8fd9\u8868\u660e\u6a21\u578b\u7fa4\u4f53\u534f\u4f5c\u80fd\u591f\u6709\u6548\u5904\u7406\u5305\u542b\u591a\u79cd\u4efb\u52a1\u7684\u590d\u6742\u9886\u57df\uff0c\u5e76\u4e14\u5728\u540c\u4e00\u4e2a\u9886\u57df\u4e2d\uff0c\u6a21\u578b\u7684\u534f\u4f5c\u9002\u5e94\u901a\u5e38\u4f18\u4e8e\u4ec5\u9488\u5bf9\u5355\u4e2a\u4efb\u52a1\u7684\u9002\u5e94\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3. \u5956\u52b1\u6a21\u578b\u4f18\u5316<\/h4>\n\n\n\n<p>Model Swarms \u8fd8\u88ab\u5e94\u7528\u4e8e\u4f18\u5316\u5956\u52b1\u6a21\u578b\uff0c\u76ee\u6807\u662f\u4f7f\u6a21\u578b\u9002\u5e94\u4e00\u822c\u6027\u548c\u5bf9\u7acb\u6027\u504f\u597d\uff0c\u4f8b\u5982\u957f\u7bc7\u8be6\u7ec6\u63cf\u8ff0\u4e0e\u7b80\u6d01\u660e\u4e86\u7684\u56de\u7b54\u4e4b\u95f4\u7684\u51b2\u7a81\u3002\u5b9e\u9a8c\u4e2d\uff0cModel Swarms \u76f8\u8f83\u4e8e\u5305\u62ecPPO\u548cDPO\u5728\u5185\u768414\u4e2a\u57fa\u51c6\u65b9\u6cd5\uff0c\u5728\u4f4e\u6570\u636e\u60c5\u666f\uff08\u4ec5200\u4e2a\u793a\u4f8b\uff09\u4e0b\uff0c\u6027\u80fd\u5e73\u5747\u63d0\u5347\u4e866.7%\u3002\u5c24\u5176\u662f\u5bf9\u591a\u6837\u5316\u7528\u6237\u504f\u597d\u7684\u9002\u5e94\uff0cModel Swarms \u5c55\u73b0\u4e86\u5353\u8d8a\u7684\u7075\u6d3b\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">4. \u4eba\u7c7b\u5174\u8da3\u9002\u5e94<\/h4>\n\n\n\n<p>\u572816\u4e2a\u4eba\u7c7b\u5174\u8da3\u9886\u57df\u4e2d\uff0cModel Swarms \u7684\u4f18\u5316\u4f7f\u5f97\u6a21\u578b\u5728\u201cLLM-as-a-judge\u201d\u8bc4\u5206\u4e2d\u63d0\u5347\u4e8617.6%\uff0c\u5728\u4e8b\u5b9e\u6027\u9a8c\u8bc1\u4e0a\u63d0\u5347\u4e8617.0%\uff0c\u5e76\u4e14\u5728\u4eba\u7c7b\u8bc4\u4f30\u4e2d\u53d6\u5f97\u4e8670.8%\u7684\u80dc\u7387\u3002\u8fd9\u8868\u660e\uff0cModel Swarms\u80fd\u591f\u663e\u8457\u63d0\u9ad8\u6a21\u578b\u5bf9\u7279\u5b9a\u9886\u57df\u548c\u7279\u5b9a\u793e\u533a\u5174\u8da3\u7684\u9002\u5e94\u6027\uff0c\u751f\u6210\u66f4\u52a0\u7b26\u5408\u4eba\u7c7b\u671f\u671b\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u56db\u3001\u6df1\u5165\u5206\u6790\u4e0e\u8ba8\u8bba<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u6b63\u786e\u6027\u7684\u65b0\u5174\u80fd\u529b<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u8bc4\u4f30LLM\u4e13\u5bb6\u662f\u5426\u5728\u534f\u4f5c\u641c\u7d22\u8fc7\u7a0b\u4e2d\u5b66\u5230\u4e86\u65b0\u7684\u6280\u80fd\uff0c\u8bba\u6587\u5b9a\u4e49\u4e86\u56db\u4e2a\u6b63\u786e\u6027\u7b49\u7ea7\uff0c\u5e76\u5f15\u5165\u4e86\u201c\u6b63\u786e\u6027\u63d0\u5347\u201d\uff08C-surge\uff09\u548c\u201c\u6b63\u786e\u6027\u65b0\u5174\u201d\uff08C-emerge\uff09\u4e24\u4e2a\u6307\u6807\u3002\u5b9e\u9a8c\u8868\u660e\uff0cModel Swarms \u5728\u56db\u4e2a\u6570\u636e\u96c6\u4e0a\u7684\u5e73\u5747C-surge\u4e3a48.2%\uff0cC-emerge\u4e3a36.0%\u523053.5%\u3002\u8fd9\u610f\u5473\u7740\uff0c\u901a\u8fc7\u534f\u4f5c\u641c\u7d22\uff0cLLM\u4e13\u5bb6\u7fa4\u4f53\u80fd\u591f\u89e3\u51b3\u5148\u524d\u88ab\u89c6\u4e3a\u201c\u65e0\u6cd5\u89e3\u51b3\u201d\u7684\u95ee\u9898\uff0c\u4ece\u800c\u5c55\u73b0\u51fa\u65b0\u5174\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. \u5f31\u5230\u5f3a\u7684\u8f6c\u53d8<\/h4>\n\n\n\n<p>\u5728\u534f\u4f5c\u641c\u7d22\u8fc7\u7a0b\u4e2d\uff0c\u5b9e\u9a8c\u53d1\u73b0\u5e76\u975e\u6240\u6709\u6700\u4f18\u7684\u6700\u7ec8\u4e13\u5bb6\u5728\u521d\u59cb\u65f6\u5c31\u662f\u6700\u4f73\u7684\u3002\u76f8\u53cd\uff0c\u5f88\u591a\u8d77\u521d\u8868\u73b0\u8f83\u5dee\u7684\u4e13\u5bb6\u901a\u8fc7\u534f\u4f5c\u8f6c\u53d8\u6210\u4e86\u6700\u4f18\u7684\u4e13\u5bb6\uff0c\u8fd9\u79cd\u73b0\u8c61\u88ab\u79f0\u4e3a\u201c\u4ece\u5f31\u5230\u5f3a\u7684\u8f6c\u53d8\u201d\u3002\u8bba\u6587\u7684\u5b9e\u9a8c\u6570\u636e\u663e\u793a\uff0c\u5728\u6700\u7ec8\u8868\u73b0\u6700\u4f18\u7684\u4e13\u5bb6\u4e2d\uff0c\u53ea\u670910.4%\u6700\u521d\u6392\u540d\u7b2c\u4e00\uff0c\u800c56.9%\u7684\u4e13\u5bb6\u539f\u672c\u4f4d\u4e8e\u7fa4\u4f53\u7684\u4e0b\u534a\u90e8\u5206\u3002\u8fd9\u8bc1\u660eModel Swarms\u80fd\u591f\u6709\u6548\u6fc0\u53d1\u5f31\u6a21\u578b\u7684\u6f5c\u5728\u80fd\u529b\uff0c\u4f7f\u5176\u901a\u8fc7\u534f\u4f5c\u83b7\u5f97\u6700\u4f73\u8868\u73b0\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3. \u591a\u6837\u6027\u5bf9\u7fa4\u4f53\u534f\u4f5c\u7684\u5f71\u54cd<\/h4>\n\n\n\n<p>\u591a\u6837\u6027\u5728 Model Swarms \u7684\u6210\u529f\u4e2d\u8d77\u7740\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u3002\u5b9e\u9a8c\u901a\u8fc7\u63a7\u5236\u521d\u59cb\u4e13\u5bb6\u7684\u6570\u91cf\u548c\u591a\u6837\u6027\uff0c\u53d1\u73b0\u591a\u6837\u6027\u8d8a\u9ad8\u7684\u4e13\u5bb6\u7fa4\u4f53\uff0c\u5176\u534f\u4f5c\u641c\u7d22\u7684\u6548\u679c\u8d8a\u597d\uff0c\u6027\u80fd\u5e73\u5747\u63d0\u5347\u4e8635.3%\u3002\u8fd9\u8868\u660e\uff0c\u62e5\u6709\u591a\u6837\u5316\u7684\u4e13\u5bb6\u80fd\u591f\u63d0\u4f9b\u66f4\u5e7f\u6cdb\u7684\u63a2\u7d22\u7a7a\u95f4\uff0c\u4fc3\u8fdb\u6a21\u578b\u5728\u9002\u5e94\u4efb\u52a1\u65f6\u83b7\u5f97\u66f4\u4f18\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">4. \u4ee4\u5f31\u8005\u80dc\u4e8e\u5f3a\u8005\u7684\u534f\u4f5c<\/h4>\n\n\n\n<p>\u8bba\u6587\u8fd8\u63a2\u8ba8\u4e86\u5982\u679c\u4ec5\u4f7f\u7528\u521d\u59cb\u6a21\u578b\u4e2d\u7684\u5f31\u4e13\u5bb6\uff08\u800c\u975e\u6700\u4f18\u4e13\u5bb6\uff09\uff0c\u5b83\u4eec\u662f\u5426\u80fd\u591f\u901a\u8fc7\u534f\u4f5c\u8d85\u8fc7\u6700\u4f18\u4e13\u5bb6\u7684\u8868\u73b0\u3002\u7ed3\u679c\u8868\u660e\uff0c\u901a\u8fc7\u534f\u4f5c\uff0c\u5f31\u4e13\u5bb6\u7fa4\u4f53\u7684\u8868\u73b0\u53ef\u4ee5\u4f18\u4e8e\u6700\u5f3a\u7684\u5355\u4e2a\u4e13\u5bb6\uff0c\u8fd9\u5728\u5e73\u5747\u6027\u80fd\u4e0a\u63d0\u5347\u4e8635.4%\u3002\u8fd9\u8fdb\u4e00\u6b65\u8bc1\u5b9e\u4e86\u7fa4\u4f53\u667a\u80fd\u7684\u5f3a\u5927\u9002\u5e94\u80fd\u529b\uff0c\u901a\u8fc7\u534f\u4f5c\u5b9e\u73b0\u4e13\u5bb6\u7684\u201c\u5f31\u53d8\u5f3a\u201d\u8fc7\u7a0b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e94\u3001\u65b9\u6cd5\u7684\u52a0\u901f\u4e0e\u4f18\u5316<\/h3>\n\n\n\n<p>\u4e3a\u4e86\u51cf\u5c11\u641c\u7d22\u8fc7\u7a0b\u4e2d\u7684\u8ba1\u7b97\u5f00\u9500\uff0c\u8bba\u6587\u63d0\u51fa\u4e86<strong>Drop-K<\/strong>\u548c<strong>Drop-N<\/strong>\u4e24\u79cd\u7c7b\u4f3cdropout\u7684\u52a0\u901f\u65b9\u6cd5\uff0c\u5206\u522b\u662f\u968f\u673a\u8df3\u8fc7\u67d0\u4e9b\u8fed\u4ee3\u4e2d\u7684\u6a21\u578b\u8bc4\u4f30\u6216\u968f\u673a\u8df3\u8fc7\u67d0\u4e9b\u4e13\u5bb6\u7684\u8bc4\u4f30\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u901a\u8fc7\u8fd9\u4e24\u79cd\u52a0\u901f\u65b9\u6cd5\uff0c\u641c\u7d22\u7684\u8ba1\u7b97\u91cf\u6700\u591a\u53ef\u4ee5\u51cf\u5c1180%\uff0c\u800c\u6027\u80fd\u4ec5\u4e0b\u964d6.0%\uff0c\u8fd9\u4e3a\u672a\u6765\u66f4\u9ad8\u6548\u7684\u6a21\u578b\u534f\u4f5c\u63d0\u4f9b\u4e86\u53ef\u80fd\u6027\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u516d\u3001\u4e0e\u73b0\u6709\u5de5\u4f5c\u7684\u5bf9\u6bd4<\/h3>\n\n\n\n<p>\u8bba\u6587\u5bf9\u6bd4\u4e86Model Swarms\u4e0e\u73b0\u6709\u6a21\u578b\u7ec4\u5408\u65b9\u6cd5\u7684\u533a\u522b\u4e0e\u4f18\u52bf\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4e0e\u4e13\u5bb6\u6df7\u5408\uff08MoE\uff09\u65b9\u6cd5\u76f8\u6bd4<\/strong>\uff0cModel Swarms \u4e0d\u9700\u8981\u5bf9\u4e13\u5bb6\u7684\u7ec4\u5408\u65b9\u5f0f\u8fdb\u884c\u5047\u8bbe\uff0c\u800c\u662f\u901a\u8fc7\u534f\u4f5c\u641c\u7d22\u52a8\u6001\u9002\u5e94\u3002<\/li>\n\n\n\n<li><strong>\u4e0e\u8fdb\u5316\u7b97\u6cd5<\/strong>\uff08\u5982\u9057\u4f20\u7b97\u6cd5\uff09\u4e0d\u540c\uff0cModel Swarms \u4e0d\u9700\u8981\u590d\u6742\u7684\u624b\u5de5\u8bbe\u8ba1\u89c4\u5219\uff0c\u800c\u662f\u901a\u8fc7\u7fa4\u4f53\u667a\u80fd\u4e2d\u7684\u7b80\u5316\u7b56\u7565\u8fdb\u884c\u4f18\u5316\u3002<\/li>\n\n\n\n<li><strong>\u4e0e\u9759\u6001\u6a21\u578b\u7b97\u672f\u7ec4\u5408\u65b9\u6cd5<\/strong>\u4e0d\u540c\uff0cModel Swarms \u901a\u8fc7\u6548\u7528\u51fd\u6570\u52a8\u6001\u8c03\u6574\u6a21\u578b\u7684\u7ec4\u5408\u65b9\u5f0f\uff0c\u5177\u6709\u66f4\u9ad8\u7684\u9002\u5e94\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e03\u3001\u7ed3\u8bba\u4e0e\u672a\u6765\u5c55\u671b<\/h3>\n\n\n\n<p>Model Swarms \u63d0\u4f9b\u4e86\u4e00\u79cd\u57fa\u4e8e\u7fa4\u4f53\u667a\u80fd\u7684\u7075\u6d3b\u9002\u5e94\u548c\u4f18\u5316\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6846\u67b6\u3002\u5b83\u901a\u8fc7\u534f\u4f5c\u641c\u7d22\u7684\u65b9\u5f0f\uff0c\u65e0\u9700\u8fdb\u884c\u5927\u91cf\u7684\u8c03\u4f18\uff0c\u4fbf\u80fd\u5728\u4f4e\u6570\u636e\u91cf\u7684\u573a\u666f\u4e2d\u5b9e\u73b0\u5bf9\u4efb\u52a1\u7684\u6709\u6548\u9002\u5e94\u3002\u5b9e\u9a8c\u7ed3\u679c\u8bc1\u660e\uff0cModel Swarms \u80fd\u591f\u8d85\u8d8a\u73b0\u6709\u7684\u591a\u79cd\u6a21\u578b\u7ec4\u5408\u65b9\u6cd5\uff0c\u5c24\u5176\u5728\u590d\u6742\u4efb\u52a1\u548c\u591a\u6837\u5316\u9700\u6c42\u7684\u573a\u666f\u4e2d\u8868\u73b0\u51fa\u5353\u8d8a\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<p><strong>\u672a\u6765\u6311\u6218\u4e0e\u6539\u8fdb\u65b9\u5411<\/strong>\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u591a\u6837\u5316\u4e13\u5bb6\u9009\u62e9\u7684\u6311\u6218<\/strong>\uff1a\u5982\u4f55\u5728\u4f17\u591a\u516c\u5f00\u6a21\u578b\u4e2d\u9009\u62e9\u6700\u5408\u9002\u7684\u4e13\u5bb6\u7fa4\u4f53\u662f\u4e00\u4e2a\u5177\u6709\u6311\u6218\u6027\u7684\u4efb\u52a1\u3002<\/li>\n\n\n\n<li><strong>\u8ba1\u7b97\u5f00\u9500\u4e0e\u52a0\u901f\u4f18\u5316<\/strong>\uff1a\u534f\u4f5c\u641c\u7d22\u7684\u8ba1\u7b97\u6210\u672c\u8f83\u9ad8\uff0c\u5982\u4f55\u5728\u4fdd\u8bc1\u6027\u80fd\u7684\u540c\u65f6\u8fdb\u4e00\u6b65\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6\u662f\u672a\u6765\u7684\u91cd\u8981\u7814\u7a76\u65b9\u5411\u3002<\/li>\n\n\n\n<li><strong>\u6743\u91cd\u7fa4\u4f53\u4e0eToken\u7fa4\u4f53\u7684\u9002\u5e94\u6027<\/strong>\uff1a\u6743\u91cd\u7fa4\u4f53\u80fd\u591f\u66f4\u6df1\u5c42\u6b21\u5730\u6539\u53d8\u6a21\u578b\u7684\u80fd\u529b\uff0c\u4f46\u9700\u8981\u5171\u4eab\u67b6\u6784\uff0c\u800cToken\u7fa4\u4f53\u5177\u6709\u66f4\u9ad8\u7684\u7075\u6d3b\u6027\uff0c\u5982\u4f55\u5728\u8fd9\u4e24\u8005\u4e4b\u95f4\u627e\u5230\u5e73\u8861\u5c06\u662f\u672a\u6765\u7684\u91cd\u8981\u8bfe\u9898\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587Model Swarms: Collaborative Search to Adapt LLM Exper [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,86,69,66],"class_list":["post-4905","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-llm","tag-engineering","tag-google"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Model Swarms: Collaborative Search to Adapt LLM Exper&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4905","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4905"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4905\/revisions"}],"predecessor-version":[{"id":4909,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4905\/revisions\/4909"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4905"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4905"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4905"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}