{"id":8251,"date":"2026-05-17T01:05:13","date_gmt":"2026-05-16T17:05:13","guid":{"rendered":"https:\/\/infernews.com\/blog\/alphagrpo\/"},"modified":"2026-05-17T01:05:13","modified_gmt":"2026-05-16T17:05:13","slug":"alphagrpo","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/alphagrpo\/","title":{"rendered":"AlphaGRPO\uff1a\u80fd\u81ea\u6211\u4fee\u6b63\u7684\u591a\u6a21\u614b\u751f\u6210"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/appendix_editing-5e996dc899bb.jpg\" alt=\"Og image\"><\/figure>\n<p>AlphaGRPO \u4fc2\u4e00\u500b\u7528\u55ba\u539f\u751f\u7d71\u4e00\u591a\u6a21\u614b\u6a21\u578b\u5605\u8a13\u7df4\u6846\u67b6\uff0c\u91cd\u9ede\u4fc2\u4ee4\u6a21\u578b\u5514\u53ea\u4fc2\u300c\u751f\u6210\u300d\uff0c\u800c\u4fc2\u6703\u6839\u64da\u63d0\u793a\u4e3b\u52d5\u63a8\u7406\uff0c\u4e26\u55ba\u8f38\u51fa\u6709\u504f\u5dee\u6642\u5617\u8a66\u81ea\u884c\u4fee\u6b63\u3002\u7db2\u9801\u5167\u5bb9\u6307\u51fa\uff0c\u4f62\u4e3b\u8981\u9762\u5411\u6587\u5b57\u751f\u5716\u540c\u76f8\u95dc\u7de8\u8f2f\u5834\u666f\uff0c\u76ee\u6a19\u4fc2\u6539\u5584\u7d30\u7bc0\u7406\u89e3\u3001\u69cb\u5716\u4e00\u81f4\u6027\uff0c\u540c\u57cb\u5c0d\u96b1\u542b\u8981\u6c42\u5605\u638c\u63e1\u3002<\/p>\n<p>\u5462\u500b\u65b9\u6cd5\u7279\u5225\u4e4b\u8655\uff0c\u5728\u65bc\u5c07 GRPO \u5f15\u5165 AR-Diffusion \u985e\u578b\u5605\u7d71\u4e00\u6a21\u578b\uff0c\u800c\u4e14\u5514\u9700\u8981\u984d\u5916 cold-start \u968e\u6bb5\u3002\u53e6\u4e00\u500b\u6838\u5fc3\u8a2d\u8a08\u4fc2 DVReward\uff1a\u5148\u5c07\u8907\u96dc\u6307\u4ee4\u62c6\u6210\u591a\u500b\u53ef\u6838\u5be6\u5605\u7d30\u554f\u984c\uff0c\u518d\u7531\u958b\u6e90\u591a\u6a21\u614b\u6a21\u578b\u6309\u8a9e\u610f\u5c0d\u9f4a\u540c\u756b\u9762\u54c1\u8cea\u63d0\u4f9b\u8f03\u7a69\u5b9a\u3001\u53ef\u89e3\u91cb\u5605\u56de\u994b\uff0c\u907f\u514d\u53ea\u9760\u55ae\u4e00\u5206\u6578\u592a\u904e\u7c60\u7d71\u3002<\/p>\n<p>\u5982\u679c\u4f60\u60f3\u7406\u89e3\u9ede\u6a23\u4f7f\u7528\uff0c\u6982\u5ff5\u4e0a\u53ef\u4ee5\u7576\u4f62\u4fc2\u4e00\u7a2e\u8a13\u7df4\u6216\u5f37\u5316\u73fe\u6709\u591a\u6a21\u614b\u751f\u6210\u6a21\u578b\u5605\u65b9\u6cd5\uff0c\u800c\u5514\u4fc2\u4e00\u822c\u7d42\u7aef\u7528\u5bb6\u5373\u958b\u5373\u7528\u5605 App\u3002\u8f03\u9069\u5408\u7814\u7a76\u4eba\u54e1\u3001\u6a21\u578b\u5de5\u7a0b\u5e2b\uff0c\u6216\u8005\u9700\u8981\u6539\u5584\u6587\u5b57\u751f\u5716\u3001\u7d30\u7c92\u5ea6\u5c6c\u6027\u63a7\u5236\u3001\u5f71\u50cf\u7de8\u8f2f\u6cdb\u5316\u80fd\u529b\u5605\u5718\u968a\u53c3\u8003\u540c\u5be6\u4f5c\u3002<\/p>\n<ul>\n<li>\u652f\u63f4\u63a8\u7406\u578b\u6587\u5b57\u751f\u5716\uff0c\u80fd\u66f4\u4e3b\u52d5\u88dc\u8db3\u7528\u5bb6\u672a\u660e\u8b1b\u5605\u610f\u5716<\/li>\n<li>\u53ef\u505a\u81ea\u6211\u53cd\u601d\u5f0f\u4fee\u6b63\uff0c\u751f\u6210\u5f8c\u518d\u6aa2\u67e5\u4e26\u8abf\u6574\u932f\u914d\u5167\u5bb9<\/li>\n<li>\u56de\u994b\u6a5f\u5236\u8f03\u7d30\u7dfb\uff0c\u5c07\u8981\u6c42\u62c6\u89e3\u6210\u53ef\u9a57\u8b49\u9805\u76ee\u518d\u8a55\u4f30<\/li>\n<li>\u5728\u591a\u500b\u751f\u6210\u57fa\u6e96\u4e0a\u6709\u4e00\u81f4\u9032\u6b65\uff0c\u4ea6\u53ef\u9077\u79fb\u5230\u7de8\u8f2f\u4efb\u52d9<\/li>\n<li>\u63a8\u8ad6\u968e\u6bb5\u52a0\u5165\u81ea\u6211\u4fee\u6b63\u5f8c\uff0c\u6587\u4e2d\u6307\u6700\u9ad8\u53ef\u518d\u63d0\u5347 5.8%<\/li>\n<\/ul>\n<p>\u5c31\u8a55\u4f30\u7d50\u679c\u800c\u8a00\uff0c\u9801\u9762\u63d0\u5230 AlphaGRPO \u55ba GenEval\u3001TIIF-Bench\u3001DPG-Bench\u3001WISE \u7b49\u751f\u6210\u57fa\u6e96\uff0c\u4ee5\u53ca GEdit \u7de8\u8f2f\u4efb\u52d9\u90fd\u6709\u63d0\u5347\uff0c\u800c\u4e14\u7de8\u8f2f\u80fd\u529b\u4e26\u975e\u9760\u5c08\u9580\u7de8\u8f2f\u8a13\u7df4\u5f97\u4f86\uff0c\u53cd\u6620\u6cdb\u5316\u8868\u73fe\u4e0d\u4fd7\u3002\u4e0d\u904e\uff0c\u5177\u9ad4\u6548\u679c\u4ecd\u61c9\u6309\u6a21\u578b\u5e95\u5ea7\u3001\u8a55\u6e2c\u8a2d\u5b9a\u540c\u5be6\u969b\u8cc7\u6599\u800c\u5b9a\u3002<\/p>\n<p>\u8a13\u7df4\u7a0b\u5f0f\u78bc\u548c\u6a21\u578b\u6b0a\u91cd\u76ee\u524d\u6b63\u5728\u9032\u884c\u5167\u90e8\u5be9\u6838\uff0c\u5be9\u6838\u901a\u904e\u5f8c\u5c07\u4e88\u4ee5\u767c\u5e03\u3002<\/p>\n<p><strong>\u7db2\u5740\uff1a<\/strong> <a href=\"https:\/\/huangrh99.github.io\/AlphaGRPO\/\" rel=\"noopener noreferrer\">https:\/\/huangrh99.github.io\/AlphaGRPO\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5462\u500b\u6846\u67b6\u91dd\u5c0d\u6587\u5b57\u751f\u5716\u8207\u7de8\u8f2f\u4efb\u52d9\uff0c\u52a0\u5165\u53ef\u9a57\u8b49\u56de\u994b\uff0c\u63d0\u5347\u63a8\u7406\u540c\u81ea\u6211\u4fee\u6b63\u80fd\u529b\u3002<\/p>\n","protected":false},"author":8,"featured_media":8250,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[133,169,176,157,120,197],"tags":[],"class_list":["post-8251","post","type-post","status-publish","format-standard","hentry","category-133","category-169","category-176","category-157","category-120","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8251","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=8251"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8251\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8250"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8251"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8251"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8251"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}