{"id":7807,"date":"2025-11-27T05:35:00","date_gmt":"2025-11-26T20:35:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7807"},"modified":"2025-11-23T08:41:06","modified_gmt":"2025-11-22T23:41:06","slug":"virtual-width-networks","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7807","title":{"rendered":"Virtual Width Networks"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Virtual Width Networks\u00a0<\/strong>[130.5]<br>VWN(Virtual Width Networks)\u306f,\u96a0\u308c\u305f\u30b5\u30a4\u30ba\u3092\u5897\u5927\u3055\u305b\u308b\u3053\u3068\u306a\u304f,\u3088\u308a\u5e83\u3044\u8868\u73fe\u306e\u5229\u70b9\u3092\u63d0\u4f9b\u3059\u308b\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3067\u3042\u308b\u3002 \u5927\u898f\u6a21\u5b9f\u9a13\u3067\u306f,8\u500d\u62e1\u5f35\u306b\u3088\u308a,\u6b21\u306e2\u500d\u306e2\u500d,\u6b21\u306e2\u500d\u306e3\u500d\u306e2\u500d\u306e\u6700\u9069\u5316\u304c\u53ef\u80fd\u3068\u306a\u3063\u305f\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2511.11238v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2511.11238v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Fri, 14 Nov 2025 12:41:57 GMT)<\/li>\n\n\n\n<li>Transfomer\u306b\u7d71\u5408\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u306a\u6539\u5584\u306e\u63d0\u6848\u3001\u300cWe introduced Virtual Width Networks (VWN) as a practical mechanism to decouple representational width from the quadratic compute typically associated with widening. With a modest 1.5\u00d7 expansion, we observe consistent improvements. When scaling to 8\u00d7 virtual width, optimization accelerates markedly: next-token prediction loss converges more than 2\u00d7 faster and multi-token prediction loss more than 3\u00d7 faster relative to the baseline width. Beyond these discrete points, the performance of VWN exhibits a clear scaling behavior.\u300d\u3001\u901a\u4fe1\u3084\u30e1\u30e2\u30ea\u90e8\u5206\u3067\u306e\u5236\u7d04\u304c\u3042\u308b\u3068\u306e\u3053\u3068\u3060\u304c\u3001\u300cIn practice, virtual width expansions in the 1.5\u00d7\u20134\u00d7 range are more feasible on today\u2019s stacks,\u300d\u3068\u3044\u3046\u8a18\u8f09\u306b\u306f\u671f\u5f85\u304c\u6301\u3066\u308b\u3002<\/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":[415,551],"class_list":["post-7807","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-transformer","tag-551"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7807","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=7807"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7807\/revisions"}],"predecessor-version":[{"id":7808,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7807\/revisions\/7808"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7807"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7807"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7807"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}