{"id":11138,"date":"2026-08-18T02:43:35","date_gmt":"2026-08-17T18:43:35","guid":{"rendered":"https:\/\/infernews.com\/blog\/simpleopd\/"},"modified":"2026-08-18T02:43:35","modified_gmt":"2026-08-17T18:43:35","slug":"simpleopd","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/simpleopd\/","title":{"rendered":"SimpleOPD \u628a\u9577\u601d\u7dad\u84b8\u993e\u5e36\u5230\u7570 tokenizer \u5b78\u751f\u6a21\u578b"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-a54c21a0e982.jpg\" alt=\"Repository image for hhnqqq\/SimpleOPD\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u60f3\u628a\u9577\u4e0a\u4e0b\u6587\u63a8\u7406\u80fd\u529b\u4ea4\u5230\u8f03\u7d30\u3001\u8f03\u77ed\u4e0a\u4e0b\u6587\u7684\u5b78\u751f\u6a21\u578b\uff0c\u5361\u4f4d\u5f80\u5f80\u5514\u4fc2\u53ea\u6709\u6a21\u578b\u5f37\u5f31\uff0c\u800c\u662f tokenizer \u4e0d\u540c\u3001\u8f38\u51fa\u6108\u84b8\u993e\u6108\u9577\uff0c\u6700\u5f8c\u9023\u8a13\u7df4\u90fd\u8b8a\u5f97\u5514\u7a69\u3002SimpleOPD\u5c6c\u65bc\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\u9805\u76ee\uff0c\u8655\u7406\u7684\u662f long-context reasoning teacher \u5230 short-context student \u4e4b\u9593\u7684 On-policy distillation\uff08OPD\uff09\u8f49\u79fb\uff0c\u91cd\u9ede\u653e\u5728\u8de8 tokenizer \u4ecd\u7136\u53ef\u4ee5\u5b78\u5230\u63a8\u7406\u80fd\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u5187\u786c\u505a token-level \u5c0d\u9f4a\uff0c\u800c\u662f\u6539\u5728 shared text space \u5c0d\u9f4a\uff0c\u53ea\u76e3\u7763\u5e2b\u751f tokenizer \u525b\u597d\u5207\u51fa\u76f8\u540c\u6587\u5b57\u8de8\u5ea6\u7684\u4f4d\u7f6e\u3002\u5462\u500b\u53d6\u6368\u5f88\u52d9\u5be6\uff1a\u5c11\u4e86\u5f37\u884c\u5c0d\u9f4a\u5e36\u4f86\u7684\u566a\u97f3\uff0c\u4ee3\u50f9\u662f\u53ef\u76e3\u7763\u4f4d\u7f6e\u6703\u6536\u7a84\uff0c\u4f46\u63db\u4f86 arbitrary teacher-student tokenizer combinations \u4e5f\u80fd\u6210\u7acb\uff0c\u5c0d Qwen3\u3001Qwen3.5\u3001Intern-S2\u3001GLM-4.7\u3001Gemma4 \u9019\u985e\u4e0d\u540c\u5bb6\u65cf\u5b78\u751f\u6a21\u578b\u90fd\u66f4\u6709\u901a\u7528\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u500b\u95dc\u9375\uff0c\u662f\u5b83\u76f4\u63a5\u8655\u7406\u300c\u7b54\u6848\u6108\u751f\u6108\u9577\u300d\u9019\u500b\u8a13\u7df4\u73fe\u8c61\u3002SimpleOPD\u52a0\u5165 student-reference KL loss\uff0c\u518d\u5c0d <code>&lt;\/think&gt;<\/code> \u548c <code>&lt;|im_end|&gt;<\/code> \u9019\u985e\u7d42\u6b62 token \u505a advantage masking\uff0c\u76ee\u7684\u4e0d\u662f\u55ae\u7d14\u52a0\u6b63\u5247\u5316\uff0c\u800c\u662f\u58d3\u4f4f\u5b78\u751f\u6a21\u578b\u504f\u96e2\u521d\u59cb policy \u592a\u9060\uff0c\u6e1b\u5c11 teacher-student distribution mismatch\u3001\u622a\u65b7\u7387\u4e0a\u5347\u8207\u8a13\u7df4\u5931\u7a69\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u652f\u63f4 long-context teacher \u7522\u751f 100K+ token reasoning chains<\/li>\n<li>\u4ee5 shared text space \u5c0d\u9f4a\uff0c\u907f\u958b tokenizer \u4e0d\u4e00\u81f4\u5e36\u4f86\u7684\u84b8\u993e\u969c\u7919<\/li>\n<li>\u900f\u904e student-reference KL loss \u8207\u7d42\u6b62 token masking \u63a7\u5236\u9577\u5ea6\u81a8\u8139<\/li>\n<li>\u5728 ProofBench \u5c0d Intern-S2-Preview \u5e36\u4f86 +21.2 \u5206\uff0c\u5347\u81f3 55.2<\/li>\n<li>\u7d50\u679c\u5df2\u8d85\u904e Gemini-2.5-Pro\uff0c\u4e26\u63d0\u5230\u5c0d HLE\u3001HiPhO \u4e5f\u6709\u5916\u6ea2\u6536\u76ca<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u4e0d\u662f\u5373\u88dd\u5373\u7528\u7684\u7dda\u4e0a\u670d\u52d9\uff0c\u800c\u662f\u504f\u7814\u7a76\u8207\u8a13\u7df4\u6d41\u7a0b\u7684\u958b\u6e90\u9805\u76ee\uff1b\u74b0\u5883\u4f9d\u8cf4 SLIME 0.2.3\u3001Python 3.10+\u3001PyTorch 2.0+\uff0c\u4e26\u70ba GPU\u3001HTTP-based teacher inference\u3001timeout\u3001concurrency\u3001batch processing \u7559\u597d\u63a5\u53e3\u3002\u53d7\u76ca\u6700\u5927\u7684\u6703\u662f\u505a\u84b8\u993e\u8a13\u7df4\u3001\u60f3\u628a\u5f37\u6559\u5e2b\u63a8\u7406\u80fd\u529b\u58d3\u5230\u8f03\u6613\u90e8\u7f72\u5b78\u751f\u6a21\u578b\u7684\u5718\u968a\uff0c\u5c24\u5176\u662f\u8981\u8de8 tokenizer\u3001\u53c8\u4e0d\u60f3\u88ab\u8d85\u9577 reasoning chain \u62d6\u57ae\u8a13\u7df4\u7a69\u5b9a\u6027\u7684\u60c5\u5883\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/hhnqqq.github.io\/SimpleOPD-project-page\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/hhnqqq\/SimpleOPD\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9577\u4e0a\u4e0b\u6587\u6559\u5e2b\u6a21\u578b\u5514\u96e3\u627e\uff0c\u96e3\u7684\u662f\u9ede\u6a23\u7a69\u5b9a\u8f49\u79fb\u5230\u77ed\u4e0a\u4e0b\u6587\u5b78\u751f\u6a21\u578b\u3002SimpleOPD\u805a\u7126\u5462\u500b\u843d\u5dee\uff0c\u505a\u6cd5\u6bd4\u4e00\u822c\u84b8\u993e\u66f4\u8cbc\u8fd1\u8a13\u7df4\u73fe\u5834\u3002<\/p>\n","protected":false},"author":8,"featured_media":11137,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,185,153,140,152],"tags":[],"class_list":["post-11138","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-openai","category-gemini","category-python"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11138","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=11138"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11138\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11137"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11138"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11138"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11138"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}