{"id":8190,"date":"2026-02-03T05:35:00","date_gmt":"2026-02-02T20:35:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=8190"},"modified":"2026-02-01T10:42:43","modified_gmt":"2026-02-01T01:42:43","slug":"self-improving-pretraining-using-post-trained-models-to-pretrain-better-models","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=8190","title":{"rendered":"Self-Improving Pretraining: using post-trained models to pretrain better models"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Self-Improving Pretraining: using post-trained models to pretrain better models\u00a0<\/strong>[40.2]<br>\u672c\u7a3f\u3067\u306f\u3001\u6587\u66f8\u3092\u30b9\u30c8\u30ea\u30fc\u30e0\u3057\u3001\u5f37\u5316\u5b66\u7fd2(RL)\u3092\u7528\u3044\u3066\u6b21\u306eK\u751f\u6210\u30c8\u30fc\u30af\u30f3\u3092\u5404\u30b9\u30c6\u30c3\u30d7\u3067\u6539\u5584\u3059\u308b\u65b0\u3057\u3044\u4e8b\u524d\u5b66\u7fd2\u624b\u6cd5\u3092\u63d0\u6848\u3059\u308b\u3002 \u5b9f\u9a13\u3067\u306f, \u5b9f\u52b9\u6027\u3068\u5b89\u5168\u6027\u306e\u70b9\u3067\u6a19\u6e96\u4e8b\u524d\u8a13\u7df4\u3088\u308a\u308236.2%\u306818.5%\u306e\u76f8\u5bfe\u7684\u306a\u6539\u5584\u304c\u5f97\u3089\u308c, \u7dcf\u751f\u7523\u54c1\u8cea\u306e86.3%\u307e\u3067\u5411\u4e0a\u3057\u305f\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2601.21343v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2601.21343v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Thu, 29 Jan 2026 07:09:30 GMT)<\/li>\n\n\n\n<li>\u300cOur work re-envisions pretraining by using a strong post-trained model to provide superior supervision signals. This works in two ways: (i) by providing rewrites on the original streaming pretrain data; and (ii) by acting as a judge. (i) We showed that such a self-improving setup can improve the factuality, safety and overall generation quality of pretrained models.\u300d\u3068\u3044\u3046\u30d5\u30ec\u30fc\u30e0\u306e\u63d0\u6848\u3002\u52b9\u679c\u306f\u3042\u308a\u305d\u3046\u3068\u601d\u3044\u3064\u3064\u3001\u3053\u308c\u3092\u5b9f\u884c\u3067\u304d\u308b\u7814\u7a76\u6a5f\u95a2\u304c\u3069\u308c\u304f\u3089\u3044\u3042\u308b\u304b\u306f\u6c17\u306b\u306a\u308b\u3068\u3053\u308d\u3002Discussion\u306e\u300cGoing further, there are other aspects of a powerful model one may wish for pretraining to also capture, i.e. other skills! \u2013 an obvious one being stronger reasoning ability.\u300d\u3092\u542b\u3081\u3066\u30fb\u30fb\u30fb\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":[356],"class_list":["post-8190","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-self-x"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8190","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=8190"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8190\/revisions"}],"predecessor-version":[{"id":8191,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8190\/revisions\/8191"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8190"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8190"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8190"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}