{"id":6148,"date":"2025-02-03T04:52:00","date_gmt":"2025-02-02T19:52:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=6148"},"modified":"2025-02-03T04:52:00","modified_gmt":"2025-02-02T19:52:00","slug":"sft-memorizes-rl-generalizes-a-comparative-study-of-foundation-model-post-training","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=6148","title":{"rendered":"SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training\u00a0<\/strong>[127.5]<br>\u30d5\u30a1\u30a6\u30f3\u30c7\u30fc\u30b7\u30e7\u30f3\u30e2\u30c7\u30eb\u3067\u306f\u3001\u6559\u5e2b\u4ed8\u304d\u5fae\u8abf\u6574(SFT)\u3068\u5f37\u5316\u5b66\u7fd2(RL)\u304c\u30dd\u30b9\u30c8\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u6280\u8853\u3068\u3057\u3066\u5e83\u304f\u4f7f\u308f\u308c\u3066\u3044\u308b\u3002 \u672c\u7a3f\u3067\u306f,\u4e00\u822c\u5316\u3068\u8a18\u61b6\u306b\u304a\u3051\u308bSFT\u3068RL\u306e\u9055\u3044\u306b\u3064\u3044\u3066\u691c\u8a0e\u3059\u308b\u3002 RL\u306f\u3001\u7279\u306b\u7d50\u679c\u306b\u57fa\u3065\u304f\u5831\u916c\u3067\u8a13\u7df4\u3055\u308c\u305f\u5834\u5408\u3001\u30eb\u30fc\u30eb\u30d9\u30fc\u30b9\u306e\u30c6\u30ad\u30b9\u30c8\u3068\u8996\u899a\u7684\u30d0\u30ea\u30a8\u30fc\u30b7\u30e7\u30f3\u306e\u4e21\u65b9\u3067\u4e00\u822c\u5316\u3055\u308c\u308b\u3053\u3068\u3092\u793a\u3059\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2501.17161v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2501.17161v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Tue, 28 Jan 2025 18:59:44 GMT)<\/li>\n\n\n\n<li>\u307e\u3055\u306b\u4eca\u77e5\u308a\u305f\u3044\u60c5\u5831\u3068\u3044\u3046\u611f\u3058\u306e\u8ad6\u6587\u3001\u300cThrough extensive experiments on the GeneralPoints and V-IRL tasks, we demonstrated that RL exhibits superior performance in learning generalizable knowledge, while SFT tends to merely memorize the training data, across both the rule and visual variations.\u300d\u3068\u306e\u3053\u3068\u3002<\/li>\n\n\n\n<li>\u4e0a\u8a18\u306b\u52a0\u3048\u3001\u300cSFT is necessary for RL training when the backbone model does not follow instructions.\u300d\u306f\u3068\u3066\u3082\u8208\u5473\u6df1\u3044\u3002\u57fa\u790e\u6027\u80fd\u306b\u3088\u3063\u3066\u52b9\u679c\u7684\u306a\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u65b9\u91dd\u304c\u7570\u306a\u308b\u3068\u3044\u3046\u306e\u306f\u4ed6\u306e\u4e8b\u4f8b\u3067\u3082\u3088\u304f\u898b\u3089\u308c\u308b\u5370\u8c61\u304c\u3042\u308a\uff08\u76f4\u611f\u7684\u306b\u3082\u305d\u3046\u3060\u308d\u3046\u3068\u3082\u601d\u3044\uff09\u3001\u3053\u306e\u3042\u305f\u308a\u306f\u91cd\u8981\u306a\u30ce\u30a6\u30cf\u30a6\u3067\u3042\u308a\u305d\u3046\u3002<\/li>\n\n\n\n<li>\u30d7\u30ed\u30b8\u30a7\u30af\u30c8\u30b5\u30a4\u30c8\u306f<a href=\"https:\/\/tianzhechu.com\/SFTvsRL\/\">SFT Memorizes, RL Generalizes<\/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":[150,576],"class_list":["post-6148","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-fine-tuning","tag-576"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6148","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=6148"}],"version-history":[{"count":0,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6148\/revisions"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6148"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6148"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6148"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}