{"id":7160,"date":"2025-07-23T05:51:00","date_gmt":"2025-07-22T20:51:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7160"},"modified":"2025-07-20T23:01:33","modified_gmt":"2025-07-20T14:01:33","slug":"probing-for-arithmetic-errors-in-language-models","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7160","title":{"rendered":"Probing for Arithmetic Errors in Language Models\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Probing for Arithmetic Errors in Language Models\u00a0<\/strong>[86.8]<br>\u8a00\u8a9e\u30e2\u30c7\u30eb\u306e\u5185\u90e8\u30a2\u30af\u30c6\u30a3\u30d9\u30fc\u30b7\u30e7\u30f3\u306f\u3001\u7b97\u8853\u8aa4\u5dee\u3092\u691c\u51fa\u3059\u308b\u305f\u3081\u306b\u4f7f\u7528\u3067\u304d\u308b\u3002 \u5358\u7d14\u306a\u30d7\u30ed\u30fc\u30d6\u306f\u30e2\u30c7\u30eb\u304c\u4e88\u6e2c\u3057\u305f\u51fa\u529b\u3068\u6b63\u89e3\u306e\u4e21\u65b9\u3092\u96a0\u853d\u72b6\u614b\u304b\u3089\u6b63\u78ba\u306b\u5fa9\u53f7\u3067\u304d\u308b\u3053\u3068\u3092\u793a\u3059\u3002 \u30e2\u30c7\u30eb\u7cbe\u5ea6\u309290%\u4ee5\u4e0a\u306e\u7cbe\u5ea6\u3067\u4e88\u6e2c\u3059\u308b\u8efd\u91cf\u30a8\u30e9\u30fc\u691c\u51fa\u5668\u3092\u8a13\u7df4\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.12379v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.12379v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 16 Jul 2025 16:27:50 GMT)<\/li>\n\n\n\n<li>\u300cStarting with a controlled set- ting of 3-digit addition, we show that simple probes can accurately decode both the model\u2019s predicted output and the correct an- swer from hidden states, regardless of whether the model\u2019s output is correct.\u300d\u306f\u307e\u3041\u3067\u304d\u308b\u3060\u308d\u3046\u3068\u3057\u3066\u3001\u300cWe then extend this analysis to a more complex setting, where the model is asked to solve math word problems only requiring addition (Cobbe et al , 2021) using a structured chain-of-thought (CoT) format (Wei et al , 2022), in which intermediate steps are expressed as equations (e g , &lt;a+b=c>).  Remarkably, we find that the same probes trained on simple arithmetic queries can be applied directly to this setting, maintaining over 80% accuracy in detecting whether the model is producing correct intermediate results.\u300d\u3084self correlation\u306b\u5f79\u7acb\u3063\u305f\u308a\u306f\u9762\u767d\u3044\u7d50\u679c\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":[356,588,589],"class_list":["post-7160","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-self-x","tag-588","tag-589"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7160","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=7160"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7160\/revisions"}],"predecessor-version":[{"id":7161,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7160\/revisions\/7161"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7160"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7160"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7160"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}