{"id":7488,"date":"2025-09-22T03:34:00","date_gmt":"2025-09-21T18:34:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7488"},"modified":"2025-09-21T16:06:55","modified_gmt":"2025-09-21T07:06:55","slug":"compute-as-teacher-turning-inference-compute-into-reference-free-supervision","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7488","title":{"rendered":"Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision\u00a0<\/strong>[26.9]<br>\u6211\u3005\u306f\u3001\u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u30fb\u30a2\u30ba\u30fb\u6559\u5e2b(CaT)\u306b\u3088\u308b\u8abf\u67fb\u3092\u76e3\u7763\u306b\u8ee2\u63db\u3059\u308b\u3053\u3068\u3092\u63d0\u6848\u3059\u308b\u3002 CaT\u306f\u5e73\u884c\u30ed\u30fc\u30eb\u30a2\u30a6\u30c8\u306e\u30b0\u30eb\u30fc\u30d7\u304b\u3089\u5358\u4e00\u306e\u53c2\u7167\u3092\u5408\u6210\u3057\u3001\u305d\u308c\u306b\u5411\u3051\u3066\u6700\u9069\u5316\u3059\u308b\u3002 \u30c6\u30b9\u30c8\u30bf\u30a4\u30e0\u306e\u624b\u9806\u3068\u3057\u3066\u3001CaT\u306fGemma 3 4B\u3001Qwen 3 4B\u3001Llama 3.1 8B\u3092\u6539\u5584\u3057\u3066\u3044\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2509.14234v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2509.14234v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 17 Sep 2025 17:59:42 GMT)<\/li>\n\n\n\n<li>\u300c(i) verifiable tasks use programmatic equivalence on final answers; (ii) non-verifiable tasks use self-proposed rubrics\u2014binary, auditable criteria scored by an independent LLM judge, with reward given by the fraction satisfied.\u300d\u3068\u691c\u8a3c\u56f0\u96e3\u306a\u30bf\u30b9\u30af\u306b\u3082\u52b9\u679c\u304c\u3042\u308b\u306e\u304c\u8208\u5473\u6df1\u3044\u3002\u300cCaT can be applied at test time for inference-time gains or inside RL (CaT-RL) to improve the policy.\u300d\u3068\u306e\u3053\u3068\u3002\u5f37\u5316\u5b66\u7fd2\u3067\u3082\u52b9\u679c\u3092\u78ba\u8a8d\u3057\u3066\u3044\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":[356],"class_list":["post-7488","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\/7488","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=7488"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7488\/revisions"}],"predecessor-version":[{"id":7489,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7488\/revisions\/7489"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7488"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7488"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7488"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}