{"id":7085,"date":"2025-07-16T05:13:00","date_gmt":"2025-07-15T20:13:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7085"},"modified":"2025-07-12T15:18:09","modified_gmt":"2025-07-12T06:18:09","slug":"scaling-rl-to-long-videos","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7085","title":{"rendered":"Scaling RL to Long Videos"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Scaling RL to Long Videos&nbsp;<\/strong>[107.4]<br>LongVILA-R1-7B \u306f VideoMME \u306a\u3069\u306e\u9577\u3044\u30d3\u30c7\u30aa QA \u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3067\u9ad8\u3044\u6027\u80fd\u3092\u767a\u63ee\u3059\u308b\u3002 LongVILA-R1\u306f\u3001\u8996\u899a\u8a00\u8a9e\u30e2\u30c7\u30eb\u306b\u304a\u3051\u308b\u30ed\u30f3\u30b0\u30d3\u30c7\u30aa\u63a8\u8ad6\u306b\u5411\u3051\u3066\u306e\u7b2c\u4e00\u6b69\u3068\u306a\u308b\u3002 \u5404\u7a2e\u30e2\u30c0\u30ea\u30c6\u30a3\u306eRL\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3092\u30b5\u30dd\u30fc\u30c8\u3059\u308b,\u4e00\u822c\u516c\u958b\u306e\u305f\u3081\u306e\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u30b7\u30b9\u30c6\u30e0\u3092\u30ea\u30ea\u30fc\u30b9\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.07966v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.07966v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Thu, 10 Jul 2025 17:47:40 GMT)<\/li>\n\n\n\n<li>\u300c(1) a large-scale dataset, LongVideo-Reason, comprising 52K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling.\u300d\u3092\u4f7f\u7528\u3057\u3066\u306e\u9577\u3044\u52d5\u753b\u3092\u7406\u89e3\u3059\u308b\u305f\u3081\u306e\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306e\u63d0\u6848<\/li>\n\n\n\n<li>\u300cUnlike domains such as math or code reasoning, where structured supervision and benchmarks are readily available [7, 8], long video reasoning requires annotating complex temporal dynamics, goals, spatial relations, and narrative elements\u2014often across minutes or hours of footage\u300d\u3068\u3001\u30b3\u30fc\u30c9\u751f\u6210\u3084\u6570\u5b66\u7684\u63a8\u8ad6\u3068\u306f\u7570\u306a\u308b\u96e3\u3057\u3055\u304c\u3042\u308b\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/NVlabs\/Long-RL\">GitHub &#8211; NVlabs\/Long-RL: Long-RL: Scaling RL to Long Sequences<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[552,576],"class_list":["post-7085","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-552","tag-576"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7085","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7085"}],"version-history":[{"count":2,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7085\/revisions"}],"predecessor-version":[{"id":7087,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7085\/revisions\/7087"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7085"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7085"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7085"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}