{"id":11135,"date":"2026-08-18T02:28:50","date_gmt":"2026-08-17T18:28:50","guid":{"rendered":"https:\/\/infernews.com\/blog\/qwen3-5-4b\/"},"modified":"2026-08-18T02:28:50","modified_gmt":"2026-08-17T18:28:50","slug":"qwen3-5-4b","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/qwen3-5-4b\/","title":{"rendered":"S\u00b2VOPD \u8b93\u5c0f\u6a21\u578b\u770b\u5c11\u4e00\u9ede\uff0cQwen3.5-4B \u53cd\u800c\u770b\u5f97\u66f4\u6e96"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-1f2d8bb963fd.jpg\" alt=\"UC San Diego\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0f\u6a21\u578b\u9762\u5c0d\u5716\u7247\u6642\uff0c\u6e1b\u5c11\u5b83\u80fd\u770b\u5230\u7684\u8cc7\u8a0a\uff0c\u7adf\u7136\u53ef\u4ee5\u5e36\u4f86\u66f4\u597d\u7684\u5b78\u7fd2\u8a0a\u865f\u3002Self-Supervised Visual On-Policy Distillation\uff08S\u00b2VOPD\uff09\u662f\u4e00\u7a2e\u8996\u89ba\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\uff0c\u900f\u904e\u8b93\u5b78\u751f\u6a21\u578b\u89c0\u770b\u4f4e\u89e3\u6790\u5ea6\u3001\u52a0\u5165\u96a8\u6a5f Gaussian noise \u7684\u5716\u7247\uff0c\u8655\u7406\u6c92\u6709\u6a19\u8a3b\u3001\u734e\u52f5\u6216\u66f4\u5f37\u6559\u5e2b\u6a21\u578b\u53ef\u7528\u7684\u9650\u5236\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8a13\u7df4\u671f\u9593\uff0c\u5b78\u751f\u6a21\u578b\u6703\u6839\u64da\u53d7\u5e72\u64fe\u7684\u5716\u7247\u7522\u751f\u56de\u7b54\uff1b\u63a1\u7528 Exponential Moving Average\uff08EMA\uff09\u7684\u6559\u5e2b\u6a21\u578b\u5247\u4ee5\u539f\u59cb\u5716\u7247\u8a55\u5206\u76f8\u540c\u7684\u751f\u6210\u524d\u7db4\uff0c\u518d\u5229\u7528 top-k generalized Jensen-Shannon divergence\uff08JSD\uff09\u628a\u8f03\u5b8c\u6574\u7684 token \u5206\u4f48\u50b3\u7d66\u5b78\u751f\u3002\u6559\u5e2b\u8207\u5b78\u751f\u4e4b\u9593\u7684\u8cc7\u8a0a\u5dee\u7570\uff0c\u4f86\u81ea\u5b78\u751f\u5c11\u770b\u4e00\u9ede\uff0c\u800c\u4e0d\u662f\u984d\u5916\u52a0\u5165 privileged information\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2022 Qwen3.5-4B \u5728\u516d\u9805\u7d30\u7c92\u5ea6\u611f\u77e5\u57fa\u6e96\u7684\u5e73\u5747\u6e96\u78ba\u7387\uff0c\u7531 70.7% \u5347\u81f3 77.4%<br \/>\n\u2022 4B \u6a21\u578b\u8868\u73fe\u9ad8\u65bc Qwen3-VL-Instruct-235B \u7684 75.8% \u548c GPT-5.4 \u7684 72.8%<br \/>\n\u2022 \u4e0d\u9700\u8981 labels\u3001rewards\u3001region annotations \u6216\u66f4\u5f37\u6559\u5e2b\u6a21\u578b<br \/>\n\u2022 4B \u540c\u6642\u63d0\u5347\u611f\u77e5\u8868\u73fe 5.7% \u548c\u6578\u5b78\u63a8\u7406 3.7%<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">S\u00b2VOPD \u4f7f\u7528\u5b78\u751f\u8996\u89d2\u7e2e\u653e\u81f3\u539f\u5716 0.3 \u81f3 0.6 \u500d\uff0c\u4e26\u52a0\u5165 stochastic Gaussian noise\u3002\u7d50\u679c\u986f\u793a\uff0c\u9019\u7a2e\u505a\u6cd5\u4e0d\u53ea\u6539\u5584\u8996\u89ba\u611f\u77e5\uff0c\u4e5f\u80fd\u907f\u514d\u90e8\u5206 privileged-information \u65b9\u6cd5\u5728\u6578\u5b78\u63a8\u7406\u4e0a\u7684\u9000\u6b65\uff1b\u4f8b\u5982 O\u209a\u209bd \u548c ZwZ \u5728\u90e8\u5206 MathVision\u3001MathVerse \u6e2c\u8a66\u4e2d\u51fa\u73fe\u660e\u986f\u4e0b\u964d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u9805\u65b9\u6cd5\u8f03\u9069\u5408\u7814\u7a76\u5c0f\u578b\u591a\u6a21\u614b\u6a21\u578b\u8a13\u7df4\u3001\u8996\u89ba\u63a8\u7406\u548c\u81ea\u76e3\u7763\u5b78\u7fd2\u7684\u8b80\u8005\u3002\u73fe\u6709\u7d50\u679c\u96c6\u4e2d\u65bc\u516d\u9805\u7d30\u7c92\u5ea6\u611f\u77e5\u57fa\u6e96\u53ca\u6578\u5b78\u63a8\u7406\u6e2c\u8a66\uff0c\u80fd\u5426\u7a69\u5b9a\u5ef6\u4f38\u81f3\u5176\u4ed6\u8cc7\u6599\u3001\u6a21\u578b\u67b6\u69cb\u548c\u66f4\u8907\u96dc\u5f71\u50cf\u4efb\u52d9\uff0c\u4ecd\u9700\u8981\u9032\u4e00\u6b65\u9a57\u8b49\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/williamium3000.github.io\/s2vopd\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>S\u00b2VOPD \u900f\u904e\u524a\u5f31\u5b78\u751f\u6a21\u578b\u7684\u8996\u89ba\u8cc7\u8a0a\uff0c\u5728\u96f6\u6a19\u8a3b\u4e0b\u63d0\u5347\u7d30\u7c92\u5ea6\u611f\u77e5\u8207\u6578\u5b78\u63a8\u7406\u8868\u73fe\u3002<\/p>\n","protected":false},"author":8,"featured_media":11134,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,30,127,199],"tags":[],"class_list":["post-11135","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-image","category-127","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11135","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=11135"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11135\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11134"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11135"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11135"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11135"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}