
{"id":13495,"date":"2026-10-17T01:01:10","date_gmt":"2026-10-16T17:01:10","guid":{"rendered":"https:\/\/infernews.com\/blog\/scope-3\/"},"modified":"2026-10-17T01:03:25","modified_gmt":"2026-10-16T17:03:25","slug":"scope-3","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/scope-3\/","title":{"rendered":"SCOPE \u8b93\u6587\u5b57\u63d0\u793a\u8a5e\u6309\u8a9e\u610f\u5b78\u7fd2"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/SCOPE.jpg\" alt=\"Uniform token GRPO versus SCOPE span credit.\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7576\u63d0\u793a\u8a5e\u5305\u542b\u7269\u4ef6\u3001\u6578\u91cf\u3001\u5c6c\u6027\u548c\u95dc\u4fc2\u6642\uff0c\u9010\u500b subword token \u5957\u7528\u540c\u4e00\u500b\u5716\u50cf\u5206\u6578\uff0c\u5f80\u5f80\u96e3\u4ee5\u77e5\u9053\u54ea\u6bb5\u6587\u5b57\u771f\u6b63\u9700\u8981\u4fee\u6539\u3002SCOPE \u662f\u4e00\u500b\u958b\u6e90\u7684\u6587\u5b57\u8f49\u5716\u50cf\u63d0\u793a\u8a5e\u512a\u5316\u9805\u76ee\uff0c\u5229\u7528 GRPO\uff08Group Relative Policy Optimization\uff09\u8b93\u8a9e\u8a00\u6a21\u578b\u91cd\u5beb\u63d0\u793a\u8a5e\uff0c\u540c\u6642\u4fdd\u7559\u51cd\u7d50\u7684\u5716\u50cf\u751f\u6210\u5668\u4f5c\u8a55\u5206\u57fa\u6e96\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u7d44\u63d0\u793a\u8a5e\u6703\u7531 pi_theta \u7522\u751f N-1 \u500b\u6539\u5beb\u7248\u672c\uff0c\u4e26\u52a0\u5165\u539f\u59cb\u63d0\u793a\u8a5e\u4f5c\u7b2c N \u500b\u6210\u54e1\uff1b\u9810\u8a2d N \u70ba 8\u3002Qwen-Image \u4ee5\u540c\u4e00\u500b diffusion seed \u751f\u6210 512 x 512 \u5716\u50cf\uff0c\u4f7f\u7528 20 steps \u548c CFG 4.0\uff0c\u518d\u7531 GenEval \u56de\u50b3 0 \u6216 1\uff0c\u6e1b\u5c11\u4e0d\u540c\u96a8\u6a5f\u7d50\u679c\u5c0d\u6bd4\u8f03\u9020\u6210\u7684\u5e72\u64fe\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img data-dominant-color=\"b5a89b\" data-has-transparency=\"false\" style=\"--dominant-color: #b5a89b;\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"961\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/SCOPE-1.jpg\" alt=\"\" class=\"wp-image-13497 not-transparent\" srcset=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/SCOPE-1.jpg 1024w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/SCOPE-1-300x282.jpg 300w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/SCOPE-1-767x720.jpg 767w\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">SCOPE \u7684\u5dee\u7570\u5728\u65bc\u628a\u56de\u994b\u7531 token \u5c64\u9762\u63d0\u5347\u81f3\u8a9e\u610f\u7247\u8a9e\u3002\u9805\u76ee\u6703\u5728\u6bcf\u500b phrase \u5167\u6c60\u5316 log pi_theta \u8207 log pi_old \u7684\u5dee\u503c\uff0c\u5c07 span coefficient \u9650\u5236\u5728 0.8 \u81f3 1.2\uff0c\u518d\u628a\u7247\u8a9e token \u6b0a\u91cd\u6b63\u898f\u5316\u81f3\u5e73\u5747\u503c 1\uff1btags \u548c\u5782\u76f4\u7dda\u5247\u7dad\u6301\u6b0a\u91cd 1\u3002\u539f\u59cb\u63d0\u793a\u8a5e\u5206\u6578\u4f4e\u65bc\u9810\u8a2d tau 1 \u6642\uff0c\u63a1\u7528\u622a\u65b7 reward gap\uff1b\u9054\u5230\u9580\u6abb\u5f8c\uff0c\u6539\u7528\u6539\u5beb\u7fa4\u7d44\u7684 clipped z-score\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8a13\u7df4\u6d41\u7a0b<\/strong>\uff1a\u539f\u59cb\u63d0\u793a\u8a5e\u6703\u53c3\u8207\u8a55\u5206\u53ca advantage \u8a08\u7b97\uff0c\u4f46\u4e0d\u76f4\u63a5\u66f4\u65b0\u53c3\u6578\u3002<\/li>\n\n\n\n<li><strong>\u7b56\u7565\u7a69\u5b9a\u6027<\/strong>\uff1api_old \u662f\u4e0a\u4e00\u500b optimizer step \u7684\u51cd\u7d50\u526f\u672c\uff0c\u7b2c\u4e00\u6b65\u6703\u8207 pi_theta \u76f8\u540c\u3002<\/li>\n\n\n\n<li><strong>\u9069\u5408\u60c5\u5883<\/strong>\uff1a\u9700\u8981\u6539\u5584\u7269\u4ef6\u6578\u91cf\u3001\u5c6c\u6027\u6216\u7a7a\u9593\u95dc\u4fc2\u8868\u9054\u7684\u6587\u5b57\u8f49\u5716\u50cf\u7814\u7a76\u8207\u5be6\u9a57\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6e2c\u8a66\u6642\u53ef\u89c0\u5bdf reward service \u7522\u751f\u7684\u5716\u50cf\u3001logs\/node-RANK\/\u5167\u7684\u5b89\u88dd\u8207\u8a13\u7df4\u8a18\u9304\uff0c\u4ee5\u53ca\u6bcf\u9805\u5de5\u4f5c\u4f7f\u7528\u7684 service.token\uff1b\u8a72 token \u6b0a\u9650\u8a2d\u5b9a\u70ba 0600\u3002\u9805\u76ee\u66f4\u63a5\u8fd1\u7814\u7a76\u578b\u8a13\u7df4\u6846\u67b6\uff0c\u800c\u4e0d\u662f\u5373\u958b\u5373\u7528\u7684\u5716\u50cf\u751f\u6210\u4ecb\u9762\uff0c\u4f7f\u7528\u8005\u9700\u8981\u6e96\u5099 Qwen-Image\u3001\u8a55\u5206\u670d\u52d9\u53ca\u8db3\u5920\u7684\u751f\u6210\u8cc7\u6e90\uff0c\u4ea6\u8981\u7559\u610f GenEval \u7684\u4e8c\u5143\u5206\u6578\u672a\u5fc5\u80fd\u6db5\u84cb\u6240\u6709\u7f8e\u5b78\u6216\u8a9e\u610f\u54c1\u8cea\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/MAC-AutoML\/SCOPE\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u5b57\u8f49\u5716\u50cf\u4e0d\u518d\u628a\u540c\u4e00\u5206\u6578\u5e73\u5747\u5206\u7d66\u6bcf\u500b\u5b57\u8a5e\uff0cSCOPE \u6309\u7269\u4ef6\u3001\u6578\u91cf\u8207\u95dc\u4fc2\u8abf\u6574\u63d0\u793a\u8a5e\u512a\u5316\u65b9\u5411\u3002<\/p>\n","protected":false},"author":8,"featured_media":13494,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,76,127,185,30,161,197,166,152,213],"tags":[],"class_list":["post-13495","post","type-post","status-publish","format-standard","hentry","category-133","category-76","category-127","category-qwen","category-image","category-161","category-framework","category-mac","category-python","category-tokenize"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13495","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=13495"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13495\/revisions"}],"predecessor-version":[{"id":13498,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13495\/revisions\/13498"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13494"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13495"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13495"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13495"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}