
{"id":11467,"date":"2026-08-26T16:59:47","date_gmt":"2026-08-26T08:59:47","guid":{"rendered":"https:\/\/infernews.com\/blog\/orarl-annotations-as-rollouts-for-efficient-scalable-reinforcement-learning-of-u-2\/"},"modified":"2026-08-26T17:05:11","modified_gmt":"2026-08-26T09:05:11","slug":"orarl-annotations-as-rollouts-for-efficient-scalable-reinforcement-learning-of-u-2","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/orarl-annotations-as-rollouts-for-efficient-scalable-reinforcement-learning-of-u-2\/","title":{"rendered":"OraRL\uff1a\u6a19\u8a3b\u5373 Rollout\uff0c\u7d71\u4e00\u591a\u6a21\u614b\u6a21\u578b\u5f37\u5316\u5b78\u7fd2"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-3b7423e906db.jpg\" alt=\"Animated OraRL method preview\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9762\u5c0d\u4e00\u6bb5\u9577\u5f71\u7247\uff0c\u6a21\u578b\u8981\u540c\u6642\u56de\u7b54\u5167\u5bb9\u554f\u984c\u3001\u627e\u51fa\u6642\u9593\u7247\u6bb5\u3001\u5b9a\u4f4d\u756b\u9762\u5340\u57df\uff0c\u5f80\u5f80\u8981\u5728\u591a\u5957\u4efb\u52d9\u65b9\u6cd5\u4e4b\u9593\u53d6\u6368\u3002OraRL \u662f\u4e00\u500b\u7528\u65bc\u7d71\u4e00\u5f71\u7247\u591a\u6a21\u614b\u6a21\u578b\uff08video MLLMs\uff09\u5f37\u5316\u5b78\u7fd2\u7684\u7814\u7a76\u9805\u76ee\uff0c\u5c07\u539f\u672c\u53ea\u7528\u4f86\u8a55\u5206\u7b54\u6848\u7684\u6a19\u8a3b\uff0c\u5e8f\u5217\u5316\u6210\u6a21\u578b\u53ef\u76f4\u63a5\u5b78\u7fd2\u7684 oracle rollout\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u6838\u5fc3\u505a\u6cd5\u662f\u628a\u4e00\u689d\u6a19\u8a3b\u7b54\u6848\u52a0\u5165\u540c\u4e00\u63d0\u793a\u7684 policy samples\uff0c\u518d\u53ea\u7528 policy rewards \u8a08\u7b97 on-policy baseline\uff0c\u907f\u514d\u6e96\u78ba\u6a19\u8a3b\u626d\u66f2\u6a21\u578b\u539f\u672c\u7684\u76f8\u5c0d\u6bd4\u8f03\uff1bannotation-policy reward gap \u5247\u7528\u65bc\u65b9\u5411\u6027\u4fee\u6b63\uff0c\u6700\u5f8c\u4ee5 sign-balanced pruning \u7be9\u9078\u66f4\u65b0\u8a0a\u865f\u3002\u9019\u500b\u5b89\u6392\u4fdd\u7559\u63a2\u7d22\uff0c\u540c\u6642\u4e0d\u9700\u8981 chain-of-thought supervision \u6216\u984d\u5916\u89e3\u78bc\u3002<\/p>\n\n\n<div\n    \tclass=\"align wp-block-vpb-video\"    id='vpbpVideoPlayer-1'\n    data-attributes='{&quot;source&quot;:&quot;https:\\\/\\\/orarl.github.io\\\/assets\\\/orarl-teaser.mp4&quot;,&quot;repeat&quot;:true,&quot;autoplay&quot;:true,&quot;muted&quot;:true,&quot;align&quot;:&quot;&quot;,&quot;poster&quot;:&quot;&quot;,&quot;controls&quot;:{&quot;play-large&quot;:true,&quot;restart&quot;:false,&quot;rewind&quot;:true,&quot;play&quot;:true,&quot;fast-forward&quot;:true,&quot;progress&quot;:true,&quot;current-time&quot;:true,&quot;duration&quot;:false,&quot;mute&quot;:true,&quot;volume&quot;:true,&quot;pip&quot;:false,&quot;airplay&quot;:false,&quot;captions&quot;:false,&quot;settings&quot;:true,&quot;download&quot;:false,&quot;fullscreen&quot;:true},&quot;captions&quot;:[],&quot;ratio&quot;:&quot;16:9&quot;,&quot;accentColor&quot;:&quot;&quot;,&quot;lazyLoad&quot;:false,&quot;width&quot;:&quot;100%&quot;,&quot;radius&quot;:&quot;0px&quot;,&quot;resetOnEnd&quot;:false,&quot;autoHideControl&quot;:true,&quot;isSetup&quot;:false}'\n><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u540c\u4e00\u5957\u66f4\u65b0\u898f\u5247\u8986\u84cb temporal grounding\u3001spatial grounding\u3001segmentation\u3001tracking\u3001spatial-temporal grounding\u3001video QA \u53ca spatial intelligence\uff0cVideo-ORA-9B \u5247\u4ee5\u4e00\u500b\u6a21\u578b\u8655\u7406\u4e03\u985e\u5f71\u7247\u7406\u89e3\u4efb\u52d9\u30024B \u8a2d\u5b9a\u7684\u66f4\u65b0\u6642\u9593\u7531\u6bcf\u6b65 92.5 \u79d2\u964d\u81f3 62.4 \u79d2\uff0c\u901f\u5ea6\u63d0\u5347 1.48 \u500d\uff0c\u55ae\u5f35 GPU \u5cf0\u503c\u8a18\u61b6\u9ad4\u4ea6\u7531 62.4 GB \u964d\u81f3 50.9 GB\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u63a8\u7406\u90e8\u5206\u63d0\u4f9b\u8f03\u5be6\u7528\u7684\u53d6\u6368\u53c3\u8003\uff1aH20 \u4ee5 vLLM \u548c BF16 \u8f09\u5165\u6642\uff0c4B \u53ca 9B \u5206\u5225\u4f54 8.6 GiB \u548c 17.6 GiB\uff1b\u5341\u5206\u9418\u3001\u6bcf\u79d2\u5169\u5e40\u7684\u5f71\u7247\u63a1\u7528 answer-only decoding \u5f8c\uff0c\u7e3d\u5ef6\u9072\u7531 29.03 \u79d2\u964d\u81f3 24.30 \u79d2\u3002\u5132\u5b58\u5eab\u5217\u51fa Environment\u3001Training \u53ca Evaluation \u6587\u4ef6\uff0c\u4f46\u63d0\u4f9b\u7684\u8cc7\u6599\u672a\u5305\u542b\u5b8c\u6574\u5b89\u88dd\u6b65\u9a5f\u6216\u53ef\u76f4\u63a5\u4e0b\u8f09\u6a21\u578b\u7684\u7d30\u7bc0\uff0c\u8f03\u9069\u5408\u7814\u7a76\u5718\u968a\u6309\u6587\u4ef6\u6aa2\u67e5\u74b0\u5883\u5f8c\u6e2c\u8a66\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8a13\u7df4\u65b9\u6cd5\uff1a<\/strong>\u6a19\u8a3b\u540c\u6642\u4f5c\u70ba\u6b63\u5411 rollout \u548c\u4efb\u52d9\u6709\u6548\u7684\u5b78\u7fd2\u76ee\u6a19\u3002<\/li>\n\n\n\n<li><strong>\u6db5\u84cb\u7bc4\u570d\uff1a<\/strong>\u4e00\u5957 RL recipe \u652f\u63f4\u4e03\u985e\u5f71\u7247\u611f\u77e5\u4efb\u52d9\u3002<\/li>\n\n\n\n<li><strong>\u6548\u7387\u6539\u5584\uff1a<\/strong>4B \u66f4\u65b0\u901f\u5ea6\u63d0\u5347 1.48 \u500d\uff0c\u8a18\u61b6\u9ad4\u9700\u6c42\u4e0b\u964d\u3002<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u512a\u5316\uff1a<\/strong>\u652f\u63f4\u5f71\u7247\u5feb\u53d6\u3001\u4e00\u6b21\u89e3\u78bc\u91cd\u7528\u53ca answer-only decoding\u3002<\/li>\n\n\n\n<li><strong>\u9069\u5408\u60c5\u5883\uff1a<\/strong>\u9700\u8981\u7d71\u4e00\u8655\u7406\u5f71\u7247\u5b9a\u4f4d\u3001\u8ffd\u8e64\u3001\u5206\u5272\u3001\u554f\u7b54\u548c\u7a7a\u9593\u63a8\u7406\u7684\u7814\u7a76\u6216\u5de5\u7a0b\u5718\u968a\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/orarl.github.io\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/HVision-NKU\/OraRL\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/OraRL\/Video-ORA-9B\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>OraRL \u628a\u5f71\u7247\u6a19\u8a3b\u8f49\u6210\u53ef\u9760\u7684\u6b63\u5411 rollout\uff0c\u4ee4\u540c\u4e00\u6a21\u578b\u517c\u9867\u5b9a\u4f4d\u3001\u5206\u5272\u3001\u8ffd\u8e64\u3001\u554f\u7b54\u8207\u7a7a\u9593\u7406\u89e3\u3002<\/p>\n","protected":false},"author":8,"featured_media":11466,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,76,119,186,127,31,120,197,199],"tags":[],"class_list":["post-11467","post","type-post","status-publish","format-standard","hentry","category-133","category-76","category-119","category-186","category-127","category-video","category-120","category-framework","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11467","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=11467"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11467\/revisions"}],"predecessor-version":[{"id":11471,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11467\/revisions\/11471"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11466"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11467"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11467"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11467"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}