{"id":8352,"date":"2026-03-25T05:36:00","date_gmt":"2026-03-24T20:36:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=8352"},"modified":"2026-03-21T14:38:37","modified_gmt":"2026-03-21T05:38:37","slug":"prism-demystifying-retention-and-interaction-in-mid-training","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=8352","title":{"rendered":"PRISM: Demystifying Retention and Interaction in Mid-Training"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>PRISM: Demystifying Retention and Interaction in Mid-Training\u00a0<\/strong>[20.2]<br>PRISM\u306f\u3001\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\u306b\u304a\u3051\u308b\u4e2d\u7d1a\u5b66\u7fd2\u8a2d\u8a08\u306e\u9078\u629e\u306b\u95a2\u3059\u308b\u7dcf\u5408\u7684\u306a\u5b9f\u8a3c\u7684\u7814\u7a76\u3067\u3042\u308b\u3002 \u7d0427B\u306e\u9ad8\u54c1\u4f4d\u30c8\u30fc\u30af\u30f3\u306e\u4e2d\u9593\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3067\u306f, \u6570\u5b66\u3067\u306f+15\u304b\u3089+40\u70b9, \u30b3\u30fc\u30c9\u3067\u306f+5\u304b\u3089+12\u70b9, \u79d1\u5b66\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3067\u306f+6\u304b\u3089+13\u70b9, \u4e00\u822c\u6027\u80fd\u306f+6\u304b\u3089+13\u70b9\u3068\u306a\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2603.17074v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2603.17074v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Tue, 17 Mar 2026 19:04:33 GMT)<\/li>\n\n\n\n<li>\u300cState-of-the-art models now incorporate an additional intermediate stage, mid-training, in which higher-quality, domain-focused data mixtures are used to imbue reasoning capabilities before downstream fine-tuning and reinforcement learning (RL) (Team et al , 2025; Olmo et al , 2025).\u300d\u3068\u306e\u3053\u3068\u3067\u3001Mid training\u306b\u95a2\u3059\u308b\u6709\u52b9\u6027\u306e\u5206\u6790\u3002<\/li>\n\n\n\n<li>\u30d7\u30ed\u30b8\u30a7\u30af\u30c8\u30b5\u30a4\u30c8\u306f<a href=\"https:\/\/bharat-runwal.github.io\/PRISM\/\">PRISM: Demystifying Retention and Interaction in Mid-Training<\/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":[718],"class_list":["post-8352","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-mid-training"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8352","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=8352"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8352\/revisions"}],"predecessor-version":[{"id":8353,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8352\/revisions\/8353"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8352"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8352"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8352"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}