{"id":7118,"date":"2025-07-14T05:37:00","date_gmt":"2025-07-13T20:37:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7118"},"modified":"2025-07-13T21:22:27","modified_gmt":"2025-07-13T12:22:27","slug":"frontier-llms-still-struggle-with-simple-reasoning-tasks","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7118","title":{"rendered":"Frontier LLMs Still Struggle with Simple Reasoning Tasks\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>Frontier LLMs Still Struggle with Simple Reasoning Tasks&nbsp;<\/strong>[53.5]<br>\u3053\u306e\u7814\u7a76\u306f\u3001\u30d5\u30ed\u30f3\u30c6\u30a3\u30a2\u8a00\u8a9e\u30e2\u30c7\u30eb\u306e\u6027\u80fd\u3092\u3001\u5e45\u5e83\u3044\u300c\u5bb9\u6613\u300d\u63a8\u8ad6\u554f\u984c\u306b\u5bfe\u3057\u3066\u7814\u7a76\u3059\u308b\u3002 \u8a08\u7b97,\u4e00\u968e\u8ff0\u8a9e\u8ad6\u7406,\u8a3c\u660e\u6728,\u65c5\u884c\u8a08\u753b\u306a\u3069,\u624b\u7d9a\u304d\u7684\u306b\u751f\u6210\u3055\u308c\u305f\u5358\u7d14\u306a\u63a8\u8ad6\u30bf\u30b9\u30af\u306e\u30b9\u30a4\u30fc\u30c8\u3092\u4f5c\u6210\u3057\u307e\u3059\u3002 \u6700\u5148\u7aef\u306e\u601d\u8003\u30e2\u30c7\u30eb\u3067\u3055\u3048\u3001\u3053\u306e\u3088\u3046\u306a\u554f\u984c\u3084\u540c\u69d8\u306e\u7406\u7531\u3067\u4e00\u8cab\u3057\u3066\u5931\u6557\u3059\u308b\u3053\u3068\u3092\u793a\u3057\u307e\u3059\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.07313v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.07313v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Wed, 09 Jul 2025 22:22:49 GMT)<\/li>\n\n\n\n<li>\u300cBy extending previous work in the literature, we create a suite of procedurally generated simple reasoning tasks, including counting, first-order logic, proof trees, and travel planning, with changeable parameters (such as document length. or the number of variables in a math problem) that can arbitrarily increase the amount of computation required to produce the answer while preserving the fundamental difficulty. While previous work showed that traditional, non-thinking models can be made to fail on such problems, we demonstrate that even state-of-the-art thinking models consistently fail on such problems and for similar reasons (e g , statistical shortcuts, errors in intermediate steps, and difficulties in processing long contexts).\u300d\u3068\u7c21\u5358\u3060\u304cLLM\/LRM\u306b\u3088\u3063\u3066\u89e3\u304d\u306b\u304f\u3044\u30bf\u30b9\u30af\u3092\u4f5c\u6210\u3002<\/li>\n\n\n\n<li>\u300cSimilarly to other recent works, our results suggest that LLMs mimic training data rather than performing true reasoning, making it relatively easy to find out-of-distribution problems where the models fail, and this problem is also present at the newest thinking models. This suggests that users remain careful when relying on the output of LLMs.\u300d\u3068\u6307\u6458\u3057\u3066\u3044\u308b\u3002\u4e0b\u8a18\u306eCatAttack\u306e\u6642\u3082\u611f\u3058\u305f\u304cLLM\/LRM\u306f\u4eba\u9593\u306e\u80fd\u529b\u3068\u306f\u304b\u306a\u308a\u7570\u306a\u3063\u3066\u3044\u308b\u3053\u3068\u306f\u610f\u8b58\u3057\u305f\u307b\u3046\u304c\u826f\u3044\u3068\u601d\u3046\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/google-deepmind\/unpuzzles_and_simple_reasoning\/\">https:\/\/github.com\/google-deepmind\/unpuzzles_and_simple_reasoning\/<\/a>\u3068\u306e\u3053\u3068<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Cats Confuse Reasoning LLM: Query Agnostic Adversarial Triggers for Reasoning Models&nbsp;<\/strong>[25.1]<br>\u672c\u7a3f\u3067\u306f,\u554f\u5408\u305b\u306b\u4f9d\u5b58\u3057\u306a\u3044\u9006\u5f15\u304d\u91d1\u3092\u5c0e\u5165\u3059\u308b\u3053\u3068\u3067,\u6bb5\u968e\u3054\u3068\u306e\u554f\u984c\u89e3\u6c7a\u3092\u8a13\u7df4\u3057\u305f\u63a8\u8ad6\u30e2\u30c7\u30eb\u306e\u30ed\u30d0\u30b9\u30c8\u6027\u306b\u3064\u3044\u3066\u691c\u8a0e\u3059\u308b\u3002 \u3088\u308a\u5f31\u304f\u5b89\u4fa1\u306a\u30d7\u30ed\u30ad\u30b7\u30e2\u30c7\u30eb\u4e0a\u3067\u30c8\u30ea\u30ac\u3092\u751f\u6210\u3059\u308b\u81ea\u52d5\u53cd\u5fa9\u653b\u6483\u30d1\u30a4\u30d7\u30e9\u30a4\u30f3\u3067\u3042\u308bCatAttack\u3092\u63d0\u6848\u3059\u308b\u3002 \u6211\u3005\u306e\u7814\u7a76\u7d50\u679c\u306f\u3001\u63a8\u8ad6\u30e2\u30c7\u30eb\u306b\u304a\u3051\u308b\u91cd\u5927\u306a\u8106\u5f31\u6027\u3092\u6d6e\u304d\u5f6b\u308a\u306b\u3057\u3066\u3001\u6700\u5148\u7aef\u30e2\u30c7\u30eb\u3067\u3055\u3048\u3001\u5fae\u5999\u306a\u6575\u306e\u5165\u529b\u306b\u5f71\u97ff\u3092\u53d7\u3051\u3084\u3059\u3044\u3053\u3068\u3092\u660e\u3089\u304b\u306b\u3057\u305f\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2503.01781v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2503.01781v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Mon, 03 Mar 2025 18:10:54 GMT)<\/li>\n\n\n\n<li>\u300cFor example, appending, Interesting fact: cats sleep most of their lives, to any math problem leads to more than doubling the chances of a model getting the answer wrong. Our findings highlight critical vulnerabilities in reasoning models, revealing that even state-of- the-art models remain susceptible to subtle adversarial inputs, raising security and reliability concerns.\u300d\u3068\u3044\u3046\u9762\u767d\u3044\u653b\u6483\u3002\u4e00\u65b9\u3067\u3001\u30ce\u30a4\u30ba\uff08\u7121\u95a2\u4fc2\uff09\u306a\u4e8b\u4f8b\u304cRAG\u306e\u6539\u5584\u306b\u6709\u52b9\u3068\u3044\u3046\u8a71\u3082\u3042\u308a\u52d5\u4f5c\u306f\u672c\u5f53\u306b\u8b0e\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/huggingface.co\/datasets\/collinear-ai\/cat-attack-adversarial-triggers\">collinear-ai\/cat-attack-adversarial-triggers \u00b7 Datasets at Hugging Face<\/a><\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>The Power of Noise: Redefining Retrieval for RAG Systems\u00a0<\/strong>[19.4]<br>Retrieval-Augmented Generation (RAG) \u306f\u3001\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\u306e\u4e8b\u524d\u5b66\u7fd2\u77e5\u8b58\u3092\u8d85\u3048\u3066\u62e1\u5f35\u3059\u308b\u65b9\u6cd5\u3068\u3057\u3066\u767b\u5834\u3057\u305f\u3002 \u6211\u3005\u306f\u3001RAG\u30bd\u30ea\u30e5\u30fc\u30b7\u30e7\u30f3\u304c\u53d6\u5f97\u3059\u3079\u304d\u30d1\u30b9IR\u30b7\u30b9\u30c6\u30e0\u306e\u7a2e\u985e\u306b\u7126\u70b9\u3092\u5f53\u3066\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2401.14887v4\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2401.14887v4\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 1 May 2024 08:15:07 GMT)<\/li>\n\n\n\n<li>\u300cFinally, and even more surprisingly, random, noisy documents are actually helpful in increasing the accuracy of these systems when correctly positioned within a prompt.\u300d\u3068\u7121\u95a2\u4fc2\u306a\u4e8b\u4f8b\u304c\u6709\u52b9\u306a\u306e\u306f\u8208\u5473\u6df1\u3044<\/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":[223,232],"class_list":["post-7118","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-llm","tag-lrm"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7118","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=7118"}],"version-history":[{"count":2,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7118\/revisions"}],"predecessor-version":[{"id":7121,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7118\/revisions\/7121"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7118"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7118"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7118"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}