{"id":8077,"date":"2026-01-22T05:35:00","date_gmt":"2026-01-21T20:35:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=8077"},"modified":"2026-01-11T14:40:52","modified_gmt":"2026-01-11T05:40:52","slug":"what-matters-for-safety-alignment","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=8077","title":{"rendered":"What Matters For Safety Alignment?\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>What Matters For Safety Alignment?\u00a0<\/strong>[38.9]<br>\u672c\u7a3f\u3067\u306f,AI\u30b7\u30b9\u30c6\u30e0\u306e\u5b89\u5168\u30a2\u30e9\u30a4\u30e1\u30f3\u30c8\u80fd\u529b\u306b\u95a2\u3059\u308b\u7dcf\u5408\u7684\u7814\u7a76\u306b\u3064\u3044\u3066\u8ff0\u3079\u308b\u3002 \u672c\u7814\u7a76\u3067\u306f,6\u3064\u306e\u91cd\u8981\u306a\u5185\u5728\u30e2\u30c7\u30eb\u7279\u6027\u30683\u3064\u306e\u5916\u90e8\u653b\u6483\u624b\u6cd5\u306e\u5f71\u97ff\u3092\u7cfb\u7d71\u7684\u306b\u691c\u8a0e\u3057,\u6bd4\u8f03\u3057\u305f\u3002 LRMs GPT-OSS-20B, Qwen3-Next-80B-A3B-Thinking, GPT-OSS-120B\u3092\u6700\u3082\u5b89\u5168\u306a3\u3064\u306e\u30e2\u30c7\u30eb\u3068\u3057\u3066\u540c\u5b9a\u3057\u305f\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2601.03868v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2601.03868v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 07 Jan 2026 12:31:52 GMT)<\/li>\n\n\n\n<li>\u300cWe systematically investigate and compare the influence of six critical intrinsic model characteristics and three external attack techniques. Our large-scale evaluation is conducted using 32 recent, popular LLMs and LRMs across thirteen distinct model families, spanning a parameter scale from 3B to 235B.\u300d\u3068\u5b89\u5168\u6027\u304b\u3089\u306eLLM\/LRM\u306e\u8a55\u4fa1\u3002\u300cThe top-three safest families OpenAI GPT-OSS [5], Alibaba Qwen3-Next [27], and Google Gemma-3 [28] are dramatically safer than the top-three most vulnerable Deepseek- R1Distilled [2], Mistral-v0.3 [29], and Seed-OSS [30] families.  These disparities can be interpreted as indicators of varying institutional investments in safety research and relative maturity in model training pipelines and infrastructures.\u300d\u3068\u3057\u3066\u3044\u308b\u3002<\/li>\n\n\n\n<li>\u57fa\u672c\u7684\u306e\u306f\u516c\u958b\u30e2\u30c7\u30eb\u304c\u5bfe\u8c61\u306e\u3088\u3046\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":[32,223,387],"class_list":["post-8077","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-attack","tag-llm","tag-survey"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8077","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=8077"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8077\/revisions"}],"predecessor-version":[{"id":8078,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/8077\/revisions\/8078"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8077"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8077"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8077"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}