{"id":7158,"date":"2025-07-24T05:43:00","date_gmt":"2025-07-23T20:43:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7158"},"modified":"2025-07-20T22:49:51","modified_gmt":"2025-07-20T13:49:51","slug":"mixture-of-recursions-learning-dynamic-recursive-depths-for-adaptive-token-level-computation","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7158","title":{"rendered":"Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation\u00a0<\/strong>[50.0]<br>\u672c\u7814\u7a76\u3067\u306f,Mixture-of-Recursions (MoR)\u3092\u5c0e\u5165\u3057\u305f\u3002 MoR\u306f\u30d1\u30e9\u30e1\u30fc\u30bf\u52b9\u7387\u3092\u9054\u6210\u3059\u308b\u305f\u3081\u306b\u518d\u5e30\u30b9\u30c6\u30c3\u30d7\u3092\u307e\u305f\u3044\u3060\u5171\u6709\u30ec\u30a4\u30e4\u306e\u30b9\u30bf\u30c3\u30af\u3092\u518d\u5229\u7528\u3057\u3001\u8efd\u91cf\u30eb\u30fc\u30bf\u306f\u9069\u5fdc\u30c8\u30fc\u30af\u30f3\u30ec\u30d9\u30eb\u306e\u601d\u8003\u3092\u53ef\u80fd\u306b\u3059\u308b\u3002 \u307e\u305f\u3001KV\u30da\u30a2\u3092\u6700\u521d\u306e\u518d\u5e30\u304b\u3089\u518d\u5229\u7528\u3059\u308bKV\u5171\u6709\u5909\u7a2e\u3092\u63d0\u6848\u3057\u3001\u7279\u306b\u30d7\u30ea\u30d5\u30a3\u30eb\u30ec\u30a4\u30c6\u30f3\u30b7\u3068\u30e1\u30e2\u30ea\u30d5\u30c3\u30c8\u30d7\u30ea\u30f3\u30c8\u306e\u524a\u6e1b\u3092\u56f3\u3063\u3066\u3044\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.10524v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.10524v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Mon, 14 Jul 2025 17:49:00 GMT)<\/li>\n\n\n\n<li>\u300cWe propose Mixture-of-Recursions (MoR)\u2014a framework that dynamically adjusts recursion step for each token during pretraining and inference. The core of MoR lies in two components: a routing mechanism that assigns token-specific recursion steps to adaptively concentrate computation on more challenging tokens, and a KV caching strategy that defines how KV pairs are stored and selectively utilized for attention at each recursive step.\u300d\u3068\u3044\u3046\u69cb\u9020\u306e\u63d0\u6848\u3002\u300cMoR consistently outperforms recursive baselines and matches or exceeds the standard Transformers at larger scales, despite using significantly fewer parameters (approximately one-third due to layer tying with \ud835\udc41\ud835\udc45= 3).\u300d\u3068\u306e\u3053\u3068\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/raymin0223\/mixture_of_recursions\">GitHub &#8211; raymin0223\/mixture_of_recursions: Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Thinking<\/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":[1],"tags":[],"class_list":["post-7158","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7158","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=7158"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7158\/revisions"}],"predecessor-version":[{"id":7159,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7158\/revisions\/7159"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7158"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7158"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7158"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}