
{"id":10437,"date":"2026-07-26T17:11:50","date_gmt":"2026-07-26T09:11:50","guid":{"rendered":"https:\/\/infernews.com\/blog\/vcsd\/"},"modified":"2026-07-26T17:11:50","modified_gmt":"2026-07-26T09:11:50","slug":"vcsd","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/vcsd\/","title":{"rendered":"VCSD \u9ede\u6a23\u903c\u53ef\u4ee5 Vision-Language Models \u771f\u4fc2\u7747\u5716"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/pasted-b10f8537f768.jpg\" alt=\"Cover Figure overview\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u5c11 Vision-Language Models \u6703\u8868\u9762\u4e0a\u8655\u7406\u5716\u7247\uff0c\u5be6\u969b\u537b\u6cbf\u4f4f\u8a9e\u8a00\u6163\u6027\u4f5c\u7b54\u3002VCSD \u5c6c\u65bc\u6a21\u578b\u8a13\u7df4\u65b9\u6cd5\uff0c\u91dd\u5c0d\u5605\u6b63\u6b63\u4fc2\u5462\u7a2e\u300c\u7b54\u6848\u4f3c\u4e4e\u5408\u7406\uff0c\u4f46\u672a\u5fc5\u771f\u4fc2\u7531\u5716\u50cf\u5e36\u52d5\u300d\u5605\u554f\u984c\uff1a\u5b83\u8b93\u540c\u4e00\u500b EMA teacher \u5206\u5225\u770b\u539f\u5716\u540c\u5167\u5bb9\u88ab\u62b9\u8d70\u5605 control input\uff0c\u518d\u7528\u5169\u8005\u5c0d\u6bcf\u500b response token \u5605\u5206\u4f48\u5dee\u7570\uff0c\u63d0\u7149\u51fa\u66f4\u4f9d\u8cf4\u8996\u89ba\u5167\u5bb9\u5605\u5b78\u7fd2\u76ee\u6a19\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709 on-policy self-distillation\uff08OPSD\uff09\u591a\u6578\u9760 privileged answers \u6216 visual evidence \u88fd\u9020 teacher \u6bd4 student \u66f4\u5f37\u5605\u8a0a\u865f\uff0cVCSD\u53cd\u904e\u4f86\u628a image-content removal \u8b8a\u6210\u975e\u5c0d\u7a31\u4f86\u6e90\u3002\u505a\u6cd5\u5514\u4fc2\u76f4\u63a5\u734e\u52f5\u67d0\u5e7e\u500b\u5b57\uff0c\u800c\u4fc2\u7528\u539f\u5716\u5206\u4f48 <code>p_hi<\/code> \u540c\u63a7\u5236\u8f38\u5165\u5206\u4f48 <code>p_ctrl<\/code> \u5605 log-probability \u5dee\uff0c\u914d\u5408 \u03b1 \u8abf\u6574\u5c0d\u6bd4\u5f37\u5ea6\uff0c\u518d\u7528 \u03b2-plausibility mask \u9650\u5236\u53ea\u55ba teacher \u539f\u672c\u5df2\u8996\u70ba\u53ef\u4fe1\u5605 token \u96c6\u5408\u5167\u91cd\u65b0\u5206\u914d\u6a5f\u7387\uff1bREADME \u4ea6\u8b1b\u660e \u03b2 \u8a2d\u6210 0.0 \u6703\u4ee4\u8a13\u7df4\u5d29\u6f70\uff0c\u4ee3\u8868\u5462\u500b\u8b77\u6b04\u5514\u4fc2\u88dd\u98fe\uff0c\u800c\u4fc2\u65b9\u6cd5\u6210\u7acb\u5605\u95dc\u9375\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u76ee\u524d\u4ecd\u7136\u4fc2 work in progress\uff0c\u4ee3\u78bc\u3001\u8a2d\u5b9a\u540c\u6587\u4ef6\u90fd\u53ef\u80fd\u518d\u6539\u3002\u5009\u5eab\u5df2\u653e\u51fa\u8a13\u7df4\u8cc7\u6599\u683c\u5f0f\u7dda\u7d22\uff0c\u4f8b\u5982 <code>train.parquet<\/code> \u9700\u8981 prompt \u540c image \u6b04\u4f4d\uff0c<code>train_answer.parquet<\/code>\u3001<code>val_answer.parquet<\/code> \u7528\u4f5c answer-conditioned validation\uff1b\u8a13\u7df4\u5247\u5efa\u57fa\u65bc\u7e7c\u627f\u81ea verl \u5605 GRPO\/PPO \u6d41\u7a0b\uff0cVCSD \u76f8\u95dc\u6539\u52d5\u96c6\u4e2d\u55ba <code>verl\/trainer\/ppo\/vcsd.py<\/code>\u3001<code>verl\/workers\/actor\/dp_actor.py<\/code> \u540c actor \u8a2d\u5b9a\u6a94\uff0c\u8868\u793a\u5b83\u6bd4\u8f03\u4f3c\u53ef\u63d2\u5165\u73fe\u6709 RL \u8a13\u7df4\u7ba1\u7dda\u5605\u9644\u52a0\u76ee\u6a19\uff0c\u800c\u5514\u4fc2\u4e00\u5957\u7368\u7acb\u6846\u67b6\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6838\u5fc3\u53d6\u5411\u4fc2\u7528 visual contrast \u4ee3\u66ff privileged answers \u6216 visual evidence<\/li>\n<li>\u5b78\u751f\u6a21\u578b\u5b78\u7fd2\u5605\u4fc2 full-vocab KL \u76ee\u6a19\uff0c\u5514\u4fc2\u9010 token \u624b\u52d5\u52a0\u6b0a<\/li>\n<li>control input \u53ef\u8a2d\u6210 black\u3001degrade \u6216 noimg\uff0c\u7528\u4f86\u6e2c\u8a66\u7b54\u6848\u6709\u5e7e\u591a\u771f\u4fc2\u9760\u5716\u50cf<\/li>\n<li>\u5df2\u516c\u958b\u7d50\u679c\u986f\u793a\uff0cVCSD \u5728 ViRL39K \u4e0a\u5c0d Qwen3-VL \u8207 Qwen3.5 \u7cfb\u5217\u5747\u6bd4 matched OPSD \u66f4\u597d<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f9e\u5df2\u516c\u958b\u6578\u5b57\u770b\uff0cQwen3-VL \u5728\u4e03\u500b benchmark aggregate \u4e0a\u7531 2B \u7684 62.27 \u5347\u5230 67.04\u30014B \u7531 71.30 \u5347\u5230 73.16\u30018B \u7531 72.51 \u5347\u5230 76.26\uff0c\u65b9\u5411\u76f8\u7576\u6e05\u695a\uff1a\u5b83\u60f3\u6539\u5584\u5605\u5514\u4fc2\u63a8\u7406\u6642\u8a08\u7b97\u91cf\uff0c\u800c\u4fc2\u8a13\u7df4\u671f\u9593\u9ede\u6a23\u628a\u300c\u5716\u7247\u771f\u6b63\u63d0\u4f9b\u5497\u4e5c\u5622\u300d\u8b8a\u6210\u66f4\u4e7e\u6de8\u5605\u76e3\u7763\u8a0a\u865f\u3002\u5c0d\u5df2\u7d93\u6709 Vision-Language Models RL \u8a13\u7df4\u6d41\u7a0b\u3001\u53c8\u60f3\u6e1b\u5c11\u5916\u90e8 teacher \u8207\u984d\u5916\u6a19\u8a3b\u4f9d\u8cf4\u5605\u7814\u7a76\u5718\u968a\uff0c\u5462\u500b\u9805\u76ee\u503c\u5f97\u8ddf\u9032\uff1b\u4e0d\u904e\u73fe\u968e\u6bb5\u4ecd\u8981\u63a5\u53d7\u6587\u4ef6\u672a\u9f4a\u3001\u4ecb\u9762\u53ef\u80fd\u8b8a\u52d5\uff0c\u4ee5\u53ca\u7d50\u679c\u4e3b\u8981\u4f86\u81ea\u8ad6\u6587\u8207\u9805\u76ee\u9801\u9762\u62ab\u9732\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/joliang17.github.io\/VisualCSD\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/joliang17\/VCSD\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.21556\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u540c\u4e00\u689d\u554f\u984c\u914d\u539f\u5716\u540c\u62b9\u8d70\u5167\u5bb9\u5605\u63a7\u5236\u8f38\u5165\uff0cVCSD\u7528\u5169\u6b21\u9810\u6e2c\u5dee\u7570\u6559\u6a21\u578b\u5206\u6e05\u908a\u5572\u7b54\u6848\u771f\u4fc2\u4f86\u81ea\u8996\u89ba\u8a0a\u606f\u3002\u5b83\u7784\u6e96\u5605\u5514\u4fc2\u52a0\u5927\u6a21\u578b\uff0c\u800c\u4fc2\u6e1b\u5c11\u6a21\u578b\u300c\u5187\u7747\u5716\u90fd\u7b54\u5230\u300d\u5605\u932f\u4f4d\u3002<\/p>\n","protected":false},"author":8,"featured_media":10436,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,30,119,149,184,197,199,204],"tags":[],"class_list":["post-10437","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-image","category-119","category-149","category-robotic","category-framework","category-dataset-","category-visionlanguageaction"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10437","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=10437"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10437\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10436"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10437"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10437"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10437"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}