
{"id":8408,"date":"2026-05-20T05:46:05","date_gmt":"2026-05-19T21:46:05","guid":{"rendered":"https:\/\/infernews.com\/blog\/post-trained-moe-can-skip-half-experts-via-self-distillation\/"},"modified":"2026-05-20T05:46:05","modified_gmt":"2026-05-19T21:46:05","slug":"post-trained-moe-can-skip-half-experts-via-self-distillation","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/post-trained-moe-can-skip-half-experts-via-self-distillation\/","title":{"rendered":"ZEDA \u9ede\u6a23\u4ee4 MoE \u63a8\u7406\u66f4\u6173\u529b"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/zeda-59c88a1050db.jpg\" alt=\"Overview of Unified Post-Training Framework.\"><\/figure>\n<p>ZEDA \u91dd\u5c0d\u7684\u662f\u4e00\u500b\u5f88\u5be6\u969b\u7684\u75db\u9ede\uff1a\u5927\u578b MoE \u6a21\u578b\u96d6\u7136\u5f37\uff0c\u4f46\u90e8\u7f72\u6642\u6bcf\u6b21\u56de\u61c9\u90fd\u8981\u52d5\u7528\u5514\u5c11\u8a08\u7b97\u8cc7\u6e90\uff0c\u6210\u672c\u9ad8\u3001\u901f\u5ea6\u4ea6\u53d7\u5f71\u97ff\u3002\u5462\u500b\u5c08\u6848\u7684\u76ee\u6a19\uff0c\u4fc2\u5514\u4f7f\u7531\u982d\u518d\u8a13\u7df4\u6a21\u578b\uff0c\u800c\u4fc2\u5728\u73fe\u6709\u3001\u5df2\u505a\u5b8c\u5f8c\u8a13\u7df4\u7684 MoE \u4e4b\u4e0a\uff0c\u518d\u6539\u9020\u6210\u66f4\u9748\u6d3b\u7684\u52d5\u614b\u7248\u672c\u3002<\/p>\n<p>\u5b83\u7684\u505a\u6cd5\u6709\u9ede\u4f3c\u300c\u8001\u5e2b\u5e36\u5b78\u751f\u300d\uff1a\u5148\u7528\u539f\u672c\u7684 MoE \u7576\u56fa\u5b9a\u8001\u5e2b\uff0c\u518d\u8a13\u7df4\u65b0\u7684\u5b78\u751f\u6a21\u578b\u53bb\u5b78\u7fd2\u8f38\u51fa\uff0c\u540c\u6642\u52a0\u5165\u4e00\u7a2e\u96f6\u8f38\u51fa\u7684\u5c08\u5bb6\uff0c\u8b93\u90e8\u5206\u8f03\u7c21\u55ae\u7684 token \u53ef\u4ee5\u7565\u904e\u4e0d\u5fc5\u8981\u8a08\u7b97\u3002\u6839\u64da\u8ad6\u6587\u8207\u5009\u5eab\u8cc7\u8a0a\uff0c\u9019\u7a2e\u65b9\u6cd5\u53ef\u6e1b\u5c11\u8d85\u904e\u4e00\u534a expert FLOPs\uff0c\u6574\u9ad4\u8868\u73fe\u53ea\u5c6c\u8f15\u5fae\u4e0b\u8dcc\uff0c\u4e26\u5e36\u4f86\u7d04 1.20 \u500d\u7aef\u5230\u7aef\u63a8\u7406\u52a0\u901f\u3002<\/p>\n<p>ZEDA \u4e0d\u662f\u901a\u7528\u958b\u767c\u6846\u67b6\uff1b\u5b83\u662f\u6e05\u83ef C3I \u5718\u968a\u7684\u4e00\u500b\u7814\u7a76\u5c08\u6848\uff0c\u5f9e\u516c\u958b\u8ad6\u6587\u6458\u8981\u770b\uff0c\u5168\u540d\u662f Zero-Expert Self-Distillation Adaptation\uff0c\u76ee\u6a19\u662f\u628a\u975c\u614b MoE \u6a21\u578b\u8f49\u6210\u66f4\u9ad8\u6548\u7684\u52d5\u614b MoE \u6a21\u578b\uff0c\u4ee5\u964d\u4f4e\u63a8\u7406\u6210\u672c\u4e26\u63d0\u5347\u901f\u5ea6\u3002\u9019\u500b\u5c08\u6848\u5c0d\u61c9\u7684 GitHub \u5009\u5eab\u5c31\u662f TsinghuaC3I\/ZEDA\uff0c\u800c\u8ad6\u6587\u9801\u9762\u4e5f\u660e\u78ba\u6307\u5411\u8a72 repo\u3002<\/p>\n<p>\u5be6\u969b\u52d5\u624b\u6642\uff0c\u6d41\u7a0b\u5927\u81f4\u5206\u5169\u6b65\uff1a\u5148\u505a SFT\uff0c\u5229\u7528\u8001\u5e2b\u6a21\u578b\u7522\u751f\u7684\u56de\u61c9\u6216\u5df2\u91cb\u51fa\u7684 rollout \u7d50\u679c\u8a13\u7df4\u5b78\u751f\uff1b\u4e4b\u5f8c\u518d\u505a OPD\uff0c\u6539\u70ba\u7531\u5b78\u751f\u81ea\u5df1\u751f\u6210\uff0c\u518d\u7531\u8001\u5e2b\u63d0\u4f9b token \u7d1a\u5225\u76ee\u6a19\u53bb\u5fae\u8abf\u3002\u5009\u5eab\u4ea6\u63d0\u5230\u53ef\u914d\u5408\u5df2\u516c\u958b\u7684 prompts \u8207 rollout \u8cc7\u6599\u4f7f\u7528\uff0c\u5c0d\u60f3\u91cd\u73fe\u7d50\u679c\u6216\u5957\u7528\u5230\u6307\u5b9a MoE \u7684\u4eba\u6703\u65b9\u4fbf\u4e0d\u5c11\u3002<\/p>\n<ul>\n<li><strong>\u6838\u5fc3\u50f9\u503c<\/strong>\uff1a\u628a\u5df2\u5b8c\u6210\u8a13\u7df4\u7684\u975c\u614b MoE\uff0c\u6539\u6210\u63a8\u7406\u6642\u66f4\u6173\u7b97\u529b\u7684\u52d5\u614b MoE<\/li>\n<li><strong>\u65b9\u6cd5\u4eae\u9ede<\/strong>\uff1a\u52a0\u5165\u96f6\u8f38\u51fa\u5c08\u5bb6\uff0c\u518d\u7528\u5169\u968e\u6bb5\u81ea\u84b8\u993e\u7a69\u5b9a\u8f49\u63db\u904e\u7a0b<\/li>\n<li><strong>\u53ef\u9078\u6a21\u578b<\/strong>\uff1aQwen3-30B-A3B\u3001GLM-4.7-Flash<\/li>\n<li><strong>\u9069\u5408\u5834\u666f<\/strong>\uff1a\u6a21\u578b\u5df2\u5b9a\u7248\uff0c\u4f46\u4e0a\u7dda\u5f8c\u4ecd\u60f3\u518d\u58d3\u4f4e\u63a8\u7406\u6210\u672c<\/li>\n<li><strong>\u8cc7\u6599\u914d\u5957<\/strong>\uff1a\u63d0\u4f9b prompts \u96c6\u5408\uff0c\u4ea6\u91cb\u51fa\u90e8\u5206 rollout \u7d50\u679c\u53ef\u76f4\u63a5\u5229\u7528<\/li>\n<\/ul>\n<p>\u6574\u9ad4\u4f86\u770b\uff0cZEDA \u6700\u503c\u5f97\u7559\u610f\u7684\u5730\u65b9\uff0c\u4e0d\u662f\u55ae\u7d14\u8ffd\u6c42\u66f4\u9ad8\u5206\uff0c\u800c\u662f\u88dc\u4e0a\u300c\u6a21\u578b\u5df2\u7d93\u8a13\u7df4\u5b8c\uff0c\u4e4b\u5f8c\u4ef2\u53ef\u4ee5\u9ede\u6a23\u518d\u6173\u8cc7\u6e90\u300d\u9019\u4e00\u6b65\u3002\u5c0d\u7814\u7a76 MoE \u90e8\u7f72\u3001\u63a8\u7406\u512a\u5316\uff0c\u6216\u8005\u624b\u4e0a\u5df2\u6709\u5927\u578b\u5f8c\u8a13\u7df4\u6a21\u578b\u7684\u5718\u968a\uff0c\u9019\u500b\u65b9\u5411\u76f8\u7576\u6709\u53c3\u8003\u50f9\u503c\uff1b\u81f3\u65bc\u4e00\u822c\u8b80\u8005\uff0c\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\u4e00\u7a2e\u7528\u8f03\u5c11\u96fb\u8166\u529f\u592b\uff0c\u63db\u4f86\u5dee\u4e0d\u591a\u6548\u679c\u7684\u6539\u88dd\u65b9\u6848\u3002<\/p>\n<p><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/TsinghuaC3I\/ZEDA\" rel=\"noopener noreferrer\">https:\/\/github.com\/TsinghuaC3I\/ZEDA<\/a><\/p>\n<p><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2605.18643\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2605.18643<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>ZEDA\u91dd\u5c0d\u5df2\u8a13\u7df4\u5b8c\u6210\u7684 MoE \u6a21\u578b\uff0c\u5617\u8a66\u7528\u8f03\u4f4e\u6210\u672c\u63db\u53d6\u66f4\u5feb\u63a8\u7406\u3002\u91cd\u9ede\u662f\u6e1b\u5c11\u5c08\u5bb6\u8a08\u7b97\uff0c\u540c\u6642\u76e1\u91cf\u4fdd\u4f4f\u539f\u6709\u8868\u73fe\u3002<\/p>\n","protected":false},"author":8,"featured_media":8407,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[133,189,196,200],"tags":[],"class_list":["post-8408","post","type-post","status-publish","format-standard","hentry","category-133","category-189","category-196","category-200"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8408","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=8408"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8408\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8407"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8408"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8408"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8408"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}