{"id":10312,"date":"2026-07-17T18:52:18","date_gmt":"2026-07-17T10:52:18","guid":{"rendered":"https:\/\/infernews.com\/blog\/univr\/"},"modified":"2026-07-17T18:52:18","modified_gmt":"2026-07-17T10:52:18","slug":"univr","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/univr\/","title":{"rendered":"UniVR\uff1a\u8996\u89ba\u63a8\u7406\u8a13\u7df4\u8b8a\u6210\u53ef\u63a7\u5de5\u4f5c\u6d41"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/pasted-23aa0477c62d.jpg\" alt=\"UniVR Overview\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">UniVR \u4fc2\u4e00\u500b\u80fd\u7406\u89e3\u6211\u5011\u8996\u89ba\u7a7a\u9593\u4e2d\u7684\u601d\u8003\u65b9\u5f0f\u53ca\u5176\u5728\u7d71\u4e00\u8996\u89ba\u63a8\u7406\u4e2d\u7684\u61c9\u7528\uff0c\u5b83\u91dd\u5c0d Emu3.5 unified generative models \u7684\u8a13\u7df4\u6846\u67b6\uff0c\u8655\u7406\u7684\u662f\u8996\u89ba\u63a8\u7406\u3001\u9577\u7a0b\u898f\u5283\u540c\u7d50\u679c\u5224\u65b7\u9ede\u6a23\u4e00\u9f4a\u5b78\u3002\u5b83\u5514\u4fc2\u62ff\u4f86\u76f4\u63a5\u505a\u63a8\u7406\u5c55\u793a\uff0c\u800c\u4fc2\u4ffe\u4f60\u7528\u81ea\u5df1\u7684\u8cc7\u6599\u540c\u734e\u52f5\u8a0a\u865f\uff0c\u53bb\u5fae\u8abf\u4e00\u500b\u5df2\u7d93\u61c2\u5f97\u8655\u7406\u5716\u50cf\u8207\u6587\u5b57\u7684\u5e95\u5ea7\u6a21\u578b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SFT\uff08supervised fine-tuning\uff09\u968e\u6bb5\u8981\u63d0\u4f9b\u7d71\u4e00\u683c\u5f0f\u7684\u6a23\u672c\uff1aquery image\u3001textual instruction\u3001visual reasoning trajectory\uff1bRL\uff08reinforcement learning\uff09\u968e\u6bb5\u5247\u6539\u6210\u900f\u904e HTTP reward server \u9001\u56de\u5206\u6578\u3002\u539f\u59cb\u8cc7\u6599\u6c92\u6709\u63d0\u4f9b\u5b8c\u6574\u5b89\u88dd\u6d41\u7a0b\uff0c\u6240\u4ee5\u76ee\u524d\u53ef\u78ba\u5b9a\u7684\u53ea\u6709\u8981\u628a\u81ea\u5b9a\u7fa9 PyTorch <code>Dataset<\/code> \u63a5\u5165 <code>UniVR_SFT\/train.py<\/code>\uff0c\u4ee5\u53ca\u628a reward function \u63db\u6210\u81ea\u5df1\u7684\u670d\u52d9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u548c\u4e00\u822c\u53ea\u505a\u55ae\u6b21\u5fae\u8abf\u7684\u505a\u6cd5\u76f8\u6bd4\uff0cUniVR \u7684\u53d6\u5411\u66f4\u504f\u5411\u300c\u5148\u6559\u683c\u5f0f\uff0c\u518d\u7528\u734e\u52f5\u4fee\u6b63\u63a8\u7406\u300d\u3002\u5b83\u5728 RL \u7aef\u7528 GRPO\uff0c\u4e26\u914d\u5408 HybridEngine \u8207 Emu3.5 \u7684 vLLM patch\uff0c\u5f37\u8abf rollout \u6548\u7387\uff1b\u540c\u6642\u4fdd\u7559 LoRA \u540c full-parameter training\uff0c\u9069\u5408\u8cc7\u6e90\u8207\u6539\u52d5\u5e45\u5ea6\u5514\u540c\u7684\u5718\u968a\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u652f\u63f4\u591a\u7bc0\u9ede SFT\uff0c\u517c\u5bb9 LoRA \u540c full-parameter training<\/li>\n<li>RL \u7aef\u57fa\u65bc verl\uff0c\u540c GRPO \u642d\u914d\u81ea\u8a02 HybridEngine<\/li>\n<li>Emu3.5 \u7684 vLLM no-CFG parallel inference \u53ef\u505a\u5230\u7d04 2 \u500d throughput<\/li>\n<li>\u734e\u52f5\u8a2d\u8a08\u5206\u6210 format reward\u3001global reward\uff0c\u540c step-level \u7684\u8996\u89ba\u63a8\u7406\u7d04\u675f<\/li>\n<li>\u76f8\u95dc\u6a21\u578b\u5305\u62ec Emu3.5 \u540c\u4f5c\u70ba\u8a55\u5206\u5668\u7684 Qwen3-VL-30B<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8f03\u5bb9\u6613\u53d7\u60e0\u7684\u60c5\u5883\u5305\u62ec\u505a\u8996\u89ba\u4ee3\u7406\u3001\u6a5f\u68b0\u81c2\/\u64cd\u4f5c\u898f\u5283\u3001\u9577\u7a0b\u4efb\u52d9\u63a8\u7406\uff0c\u6216\u8005\u60f3\u5c07\u73fe\u6210\u8996\u89ba\u6a21\u578b\u8f49\u6210\u81ea\u5df1\u5de5\u4f5c\u6d41\u7684\u5718\u968a\u3002\u5b83\u7684\u50f9\u503c\u5728\u65bc\u628a\u300c\u8cc7\u6599\u683c\u5f0f\u3001\u63a8\u7406\u8ecc\u8de1\u3001\u734e\u52f5\u5224\u65b7\u300d\u4e32\u6210\u540c\u4e00\u689d\u8a13\u7df4\u8def\u7dda\uff0c\u4ee4\u8996\u89ba\u4efb\u52d9\u5514\u518d\u53ea\u9760\u975c\u614b\u6a19\u8a3b\u53bb\u5b78\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/maverickren.github.io\/UniVR.github.io\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/bytedance\/UniVR\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.12800\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>UniVR \u628a Emu3.5 \u7684\u8996\u89ba\u63a8\u7406\u8a13\u7df4\u62c6\u6210 SFT \u8207 GRPO \u5169\u6bb5\uff0c\u8b93\u5718\u968a\u53ef\u7528\u81ea\u5df1\u7684\u8cc7\u6599\u540c\u734e\u52f5\u51fd\u6578\u63a5\u4e0a\u53bb\u3002\u5b83\u66f4\u50cf\u4e00\u500b\u9762\u5411\u8996\u89ba\u4efb\u52d9\u7684\u8a13\u7df4\u9aa8\u67b6\uff0c\u800c\u5514\u4fc2\u55ae\u4e00\u6a21\u578b\u3002<\/p>\n","protected":false},"author":8,"featured_media":10311,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,176,140,147,153,30,152,119,127,199],"tags":[],"class_list":["post-10312","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-176","category-gemini","category-deepseek","category-openai","category-image","category-python","category-119","category-127","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10312","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=10312"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10312\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10311"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10312"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10312"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10312"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}