{"id":1072,"date":"2026-06-25T03:30:35","date_gmt":"2026-06-24T18:30:35","guid":{"rendered":"https:\/\/itexplore.org\/jp\/columns\/ai-cybersecurity-gene-editing-small-llm-reasoning-frontiers\/"},"modified":"2026-06-25T03:30:35","modified_gmt":"2026-06-24T18:30:35","slug":"ai-cybersecurity-gene-editing-small-llm-reasoning-frontiers","status":"publish","type":"post","link":"https:\/\/itexplore.org\/jp\/columns\/ai-cybersecurity-gene-editing-small-llm-reasoning-frontiers\/","title":{"rendered":"AI\u30b5\u30a4\u30d0\u30fc\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u3001\u907a\u4f1d\u5b50\u7de8\u96c6\u3001\u5c0f\u578bLLM\u63a8\u8ad6\u306e\u6700\u524d\u7dda"},"content":{"rendered":"<p>\u672c\u65e5\u306e\u6ce8\u76eeAI\u30fb\u30c6\u30c3\u30af\u30cb\u30e5\u30fc\u30b9\u3092\u3001\u5c02\u9580\u7684\u306a\u5206\u6790\u3068\u5171\u306b\u304a\u5c4a\u3051\u3057\u307e\u3059\u3002<\/p>\n<div class=\"wp-block-vk-blocks-alert vk_alert alert alert-warning has-alert-icon\">\n<div class=\"vk_alert_icon\">\n<div class=\"vk_alert_icon_icon\"><i class=\"fa-solid fa-triangle-exclamation\" aria-hidden=\"true\"><\/i><\/div>\n<div class=\"vk_alert_icon_text\"><span>Warning<\/span><\/div>\n<\/div>\n<div class=\"vk_alert_content\">\n<p>\u3053\u306e\u8a18\u4e8b\u306fAI\u306b\u3088\u3063\u3066\u81ea\u52d5\u751f\u6210\u30fb\u5206\u6790\u3055\u308c\u305f\u3082\u306e\u3067\u3059\u3002AI\u306e\u6027\u8cea\u4e0a\u3001\u4e8b\u5b9f\u8aa4\u8a8d\u304c\u542b\u307e\u308c\u308b\u53ef\u80fd\u6027\u304c\u3042\u308b\u305f\u3081\u3001\u91cd\u8981\u306a\u5224\u65ad\u3092\u4e0b\u3059\u969b\u306f\u5fc5\u305a\u30ea\u30f3\u30af\u5148\u306e\u4e00\u6b21\u30bd\u30fc\u30b9\u3092\u3054\u78ba\u8a8d\u304f\u3060\u3055\u3044\u3002<\/p>\n<\/div>\n<\/div>\n<div class=\"wp-block-group\" style=\"margin-top:40px;margin-bottom:40px\">\n<h2 class=\"wp-block-heading\">Daybreak: \u4e16\u754c\u4e2d\u306e\u7d44\u7e54\u3092\u4fdd\u8b77\u3059\u308b\u305f\u3081\u306e\u30c4\u30fc\u30eb | OpenAI<\/h2>\n<ul>\n<li><strong>\u539f\u984c:<\/strong> Daybreak: Tools for securing every organization in the world | OpenAI<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">\u5c02\u9580\u30a2\u30ca\u30ea\u30b9\u30c8\u306e\u5206\u6790<\/h3>\n<div class=\"ai-summary-content\">\n<p><strong>OpenAI<\/strong>\u306f\u3001\u30b5\u30a4\u30d0\u30fc\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u306e\u8ab2\u984c\u306b\u5bfe\u51e6\u3059\u308b\u305f\u3081\u3001\u65b0\u3057\u3044\u30c4\u30fc\u30eb\u3068\u30d1\u30fc\u30c8\u30ca\u30fc\u30b7\u30c3\u30d7\u3001\u305d\u3057\u3066\u300c<strong>GPT-5.5-Cyber<\/strong>\u300d\u306e\u5b8c\u5168\u7248\u3092\u767a\u8868\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u306f\u3001\u8106\u5f31\u6027\u767a\u898b\u304b\u3089\u30a8\u30f3\u30c9\u30c4\u30fc\u30a8\u30f3\u30c9\u306e\u30d1\u30c3\u30c1\u81ea\u52d5\u5316\u307e\u3067\u3092\u52a0\u901f\u3055\u305b\u308b\u3053\u3068\u3092\u76ee\u7684\u3068\u3057\u3066\u3044\u307e\u3059\u3002AI\u304c\u8106\u5f31\u6027\u767a\u898b\u3092\u52a0\u901f\u3055\u305b\u308b\u4e00\u65b9\u3067\u3001\u30d1\u30c3\u30c1\u9069\u7528\u304c\u65b0\u305f\u306a\u30dc\u30c8\u30eb\u30cd\u30c3\u30af\u3068\u306a\u3063\u3066\u3044\u308b\u73fe\u72b6\u306b\u5bfe\u5fdc\u3059\u308b\u305f\u3081\u3001\u300c<strong>Daybreak<\/strong>\u300d\u306f\u9632\u5fa1\u5074\u304c\u6a5f\u68b0\u901f\u5ea6\u3067\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u306e\u8106\u5f31\u6027\u3092\u4fee\u6b63\u3067\u304d\u308b\u3088\u3046\u652f\u63f4\u3057\u307e\u3059\u3002<\/p>\n<p>\u4e3b\u8981\u306a\u53d6\u308a\u7d44\u307f\u3068\u3057\u3066\u3001\u300c<strong>Codex Security<\/strong>\u300d\u30d7\u30e9\u30b0\u30a4\u30f3\u306e\u66f4\u65b0\u304c\u3042\u308a\u3001\u3053\u308c\u306f\u65e2\u5b58\u30b7\u30b9\u30c6\u30e0\u306e\u8106\u5f31\u6027\u767a\u898b\u3068\u30d1\u30c3\u30c1\u9069\u7528\u3092\u52a0\u901f\u3057\u3001\u65b0\u305f\u306a\u8106\u5f31\u6027\u304c\u672c\u756a\u74b0\u5883\u306b\u5230\u9054\u3059\u308b\u306e\u3092\u9632\u304e\u307e\u3059\u3002\u307e\u305f\u3001\u300c<strong>GPT-5.5-Cyber<\/strong>\u300d\u306f\u3001<strong>CyberGym<\/strong>\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u306785.6%\u306e\u30b9\u30b3\u30a2\u3092\u9054\u6210\u3057\u3001\u5358\u4e00\u30e2\u30c7\u30eb\u3068\u3057\u3066\u306f\u6700\u9ad8\u306e\u6027\u80fd\u3092\u793a\u3057\u307e\u3057\u305f\u3002\u3055\u3089\u306b\u3001<strong>ExploitGym<\/strong>\u3084<strong>SEC-bench Pro<\/strong>\u3068\u3044\u3063\u305f\u5b9f\u4e16\u754c\u306e\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3067\u3082<strong>GPT-5.5<\/strong>\u3092\u4e0a\u56de\u308b\u6027\u80fd\u3092\u767a\u63ee\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u300c<strong>Daybreak Cyber Partner Program<\/strong>\u300d\u3092\u901a\u3058\u3066\u3001\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u30d1\u30fc\u30c8\u30ca\u30fc\u306f<strong>GPT-5.5<\/strong>\u3092\u9867\u5ba2\u306e\u88fd\u54c1\u3084\u30b5\u30fc\u30d3\u30b9\u306b\u7d71\u5408\u3057\u3001\u9632\u5fa1\u80fd\u529b\u3092\u5f37\u5316\u3067\u304d\u307e\u3059\u3002\u307e\u305f\u3001\u300c<strong>Patch the Planet<\/strong>\u300d\u30a4\u30cb\u30b7\u30a2\u30c1\u30d6\u306f\u3001<strong>Trail of Bits<\/strong>\u3001<strong>HackerOne<\/strong>\u3001<strong>Calif<\/strong>\u3068\u5354\u529b\u3057\u3001\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u30d7\u30ed\u30b8\u30a7\u30af\u30c8\u306e\u30e1\u30f3\u30c6\u30ca\u30fc\u304c\u8106\u5f31\u6027\u3092\u4fee\u6b63\u3059\u308b\u306e\u3092\u652f\u63f4\u3057\u307e\u3059\u3002\u3053\u306e\u53d6\u308a\u7d44\u307f\u306f\u3001<strong>Firefox<\/strong>\u3001<strong>V8<\/strong>\u3001<strong>Safari<\/strong>\u3001<strong>OpenBSD<\/strong>\u3001<strong>FreeBSD<\/strong>\u306a\u3069\u306e\u5e83\u304f\u4f7f\u7528\u3055\u308c\u3066\u3044\u308b\u30b7\u30b9\u30c6\u30e0\u3067\u8106\u5f31\u6027\u3092\u7279\u5b9a\u3057\u3001\u691c\u8a3c\u3059\u308b\u306e\u306b\u5f79\u7acb\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\ud83d\udc49 <strong><a href=\"https:\/\/openai.com\/index\/daybreak-securing-the-world\/\" target=\"_blank\" rel=\"noopener\">OpenAI \u3067\u8a18\u4e8b\u5168\u6587\u3092\u8aad\u3080<\/a><\/strong><\/p>\n<\/div>\n<ul>\n<li><strong>\u8981\u70b9:<\/strong> OpenAI's Daybreak initiative, powered by GPT-5.5-Cyber and Codex Security, aims to democratize and accelerate vulnerability patching at machine speed, shifting the cybersecurity focus from discovery to rapid remediation through advanced AI models and ecosystem partnerships.<\/li>\n<li><strong>\u8457\u8005:<\/strong> OpenAI<\/li>\n<\/ul>\n<blockquote class=\"wp-block-quote\"><p><span>English Summary:<\/span><\/p>\n<p><strong>OpenAI<\/strong> has announced new tools, partnerships, and the full version of <strong>GPT-5.5-Cyber<\/strong> under its <strong>Daybreak<\/strong> initiative, aiming to accelerate the process from vulnerability discovery to end-to-end patch automation. Recognizing that while AI has accelerated vulnerability discovery, patching has become the new bottleneck, <strong>Daybreak<\/strong> seeks to empower defenders to fix vulnerable software at machine speed.<\/p>\n<p>Key components include an update to the <strong>Codex Security<\/strong> plugin, designed to accelerate the discovery and patching of vulnerabilities in existing systems and prevent new ones from reaching production. The updated <strong>GPT-5.5-Cyber<\/strong> model achieved a state-of-the-art score of 85.6% on the <strong>CyberGym<\/strong> benchmark, the highest measured from a single model, and also outperformed <strong>GPT-5.5<\/strong> on real-world security benchmarks like <strong>ExploitGym<\/strong> and <strong>SEC-bench Pro<\/strong>.<\/p>\n<p>Through the <strong>Daybreak Cyber Partner Program<\/strong>, security partners can integrate <strong>GPT-5.5<\/strong> with Trusted Access for Cyber into their products and services to enhance defensive capabilities for their customers. Additionally, the <strong>Patch the Planet<\/strong> initiative, founded with <strong>Trail of Bits<\/strong>, <strong>HackerOne<\/strong>, and <strong>Calif<\/strong>, supports open-source project maintainers in moving from findings to fixes. This work has already helped identify and validate vulnerabilities in widely used systems such as <strong>Firefox<\/strong>, <strong>V8<\/strong>, <strong>Safari<\/strong>, <strong>OpenBSD<\/strong>, and <strong>FreeBSD<\/strong>.<\/p>\n<\/blockquote>\n<\/div>\n<div class=\"wp-block-group\" style=\"margin-top:40px;margin-bottom:40px\">\n<h2 class=\"wp-block-heading\">CRISPR\u306f\u4e8c\u91cd\u3089\u305b\u3093\u3092\u5207\u65ad\u3059\u308b\u3053\u3068\u306a\u304f\u5b8c\u5168\u306a\u907a\u4f1d\u5b50\u3092\u633f\u5165\u53ef\u80fd\u306b\u3002\u5b9f\u9a13\u5ba4\u3067\u9032\u5316\u3057\u305f\u9175\u7d20\u304c10,000\u4ee5\u4e0a\u306eDNA\u6587\u5b57\u3092\u30d2\u30c8\u7d30\u80de\u306b\u5c0e\u5165\u3002<\/h2>\n<ul>\n<li><strong>\u539f\u984c:<\/strong> CRISPR no longer needs to break the double helix to insert a complete gene. A laboratory-evolved enzyme has just introduced more than 10,000 DNA letters into human cells.<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">\u5c02\u9580\u30a2\u30ca\u30ea\u30b9\u30c8\u306e\u5206\u6790<\/h3>\n<div class=\"ai-summary-content\">\n<p>\u907a\u4f1d\u5b50\u7de8\u96c6\u6280\u8853\u300c<strong>CRISPR<\/strong>\u300d\u306b\u6b74\u53f2\u7684\u306a\u9032\u5c55\u304c\u3042\u308a\u3001<strong>DNA<\/strong>\u306e\u4e8c\u91cd\u3089\u305b\u3093\u3092\u5207\u65ad\u3059\u308b\u3053\u3068\u306a\u304f\u3001\u5b8c\u5168\u306a\u907a\u4f1d\u5b50\u3092\u30d2\u30c8\u7d30\u80de\u306b\u633f\u5165\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3057\u305f\u3002\u3053\u308c\u307e\u3067<strong>CRISPR-Cas9<\/strong>\u306f<strong>DNA<\/strong>\u3092\u5207\u65ad\u3057\u3001\u7d30\u80de\u306e\u4fee\u5fa9\u30b7\u30b9\u30c6\u30e0\u306b\u4f9d\u5b58\u3057\u3066\u3044\u307e\u3057\u305f\u304c\u3001\u3053\u306e\u65b0\u3057\u3044\u6280\u8853\u306f\u3088\u308a\u5b89\u5168\u3067\u5f37\u529b\u306a\u907a\u4f1d\u5b50\u7de8\u96c6\u306e\u9053\u3092\u958b\u304d\u307e\u3059\u3002<\/p>\n<p><strong>\u30b3\u30ed\u30f3\u30d3\u30a2\u5927\u5b66<\/strong>\u306e\u751f\u5316\u5b66\u8005<strong>Samuel Sternberg<\/strong>\u6c0f\u304c\u7387\u3044\u308b\u7814\u7a76\u30c1\u30fc\u30e0\u306f\u3001\u300c<strong>CAST (CRISPR-Associated Transposase)<\/strong>\u300d\u3068\u547c\u3070\u308c\u308b\u7570\u306a\u308b\u6280\u8853\u3092\u958b\u767a\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u306f\u3001<strong>CRISPR<\/strong>\u30b7\u30b9\u30c6\u30e0\u3092\u5229\u7528\u3057\u3066\u76ee\u7684\u306e\u914d\u5217\u3092\u7279\u5b9a\u3057\u3064\u3064\u3001<strong>\u30c8\u30e9\u30f3\u30b9\u30dd\u30b6\u30fc\u30bc<\/strong>\uff08\u30b2\u30ce\u30e0\u5185\u3067<strong>DNA<\/strong>\u65ad\u7247\u3092\u79fb\u52d5\u3067\u304d\u308b\u5929\u7136\u9175\u7d20\uff09\u3092\u6d3b\u7528\u3057\u3066\u3001<strong>DNA<\/strong>\u69cb\u9020\u3092\u640d\u50b7\u3059\u308b\u3053\u3068\u306a\u304f\u5b8c\u5168\u306a\u907a\u4f1d\u5b50\u3092\u633f\u5165\u3057\u307e\u3059\u3002<\/p>\n<p>\u7814\u7a76\u8005\u305f\u3061\u306f\u3001\u30a6\u30a4\u30eb\u30b9\u3092\u30ad\u30e3\u30ea\u30a2\u3068\u3057\u3066\u4f7f\u7528\u3059\u308b\u6307\u5411\u6027\u9032\u5316\u30d7\u30ed\u30bb\u30b9\u3092\u901a\u3058\u3066\u3001\u3053\u308c\u3089\u306e\u9175\u7d20\u306e\u52b9\u7387\u3092\u5411\u4e0a\u3055\u305b\u300110,000\u30cc\u30af\u30ec\u30aa\u30c1\u30c9\u3092\u8d85\u3048\u308b\u907a\u4f1d\u5b50\u65ad\u7247\u3092\u30d2\u30c8\u7d30\u80de\u306b\u633f\u5165\u3059\u308b\u3053\u3068\u306b\u6210\u529f\u3057\u307e\u3057\u305f\u3002\u3053\u306e\u6280\u8853\u306f\u3001\u907a\u4f1d\u6027\u75be\u60a3\u3084\u5e0c\u5c11\u306a\u4ee3\u8b1d\u6027\u75be\u60a3\u306e\u6cbb\u7642\u306b\u9769\u547d\u3092\u3082\u305f\u3089\u3059\u53ef\u80fd\u6027\u304c\u3042\u308a\u3001\u4e88\u671f\u305b\u306c\u5909\u7570\u3084\u6709\u5bb3\u907a\u4f1d\u5b50\u306e\u6d3b\u6027\u5316\u3068\u3044\u3063\u305f\u907a\u4f1d\u5b50\u7de8\u96c6\u306b\u4f34\u3046\u30ea\u30b9\u30af\u3092\u4f4e\u6e1b\u3057\u307e\u3059\u3002<\/p>\n<p>\ud83d\udc49 <strong><a href=\"https:\/\/es.gizmodo.com\/la-edicion-genetica-da-un-giro-historico-ya-es-posible-insertar-genes-completos-en-celulas-humanas-sin-cortar-adn-2000168128\" target=\"_blank\" rel=\"noopener\">Gizmodo en Espa\u00f1ol \u3067\u8a18\u4e8b\u5168\u6587\u3092\u8aad\u3080<\/a><\/strong><\/p>\n<\/div>\n<ul>\n<li><strong>\u8981\u70b9:<\/strong> A new CRISPR-associated transposase (CAST) system allows for the precise insertion of large gene sequences (over 10,000 nucleotides) into human cells without cutting the DNA double helix, offering a safer and more efficient gene editing method with revolutionary potential for treating genetic diseases.<\/li>\n<li><strong>\u8457\u8005:<\/strong> Mart\u00edn Nicol\u00e1s Parolari<\/li>\n<\/ul>\n<blockquote class=\"wp-block-quote\"><p><span>English Summary:<\/span><\/p>\n<p>A historic breakthrough in gene editing technology, <strong>CRISPR<\/strong>, now allows for the insertion of complete genes into human cells without cutting the <strong>DNA<\/strong> double helix. Previously, <strong>CRISPR-Cas9<\/strong> relied on cutting <strong>DNA<\/strong> and the cell's repair system, but this new technique opens a safer and more powerful path for gene editing.<\/p>\n<p>A research team led by biochemist <strong>Samuel Sternberg<\/strong> at <strong>Columbia University<\/strong> developed a different technique called <strong>CAST (CRISPR-Associated Transposase)<\/strong>. This method utilizes the <strong>CRISPR<\/strong> system to locate desired sequences while leveraging <strong>transposases<\/strong> (natural enzymes that can move <strong>DNA<\/strong> fragments within the genome) to insert complete genes without damaging the <strong>DNA<\/strong> structure.<\/p>\n<p>Through a process of directed evolution using viruses as carriers, researchers improved the efficiency of these enzymes, successfully inserting genetic fragments of over 10,000 nucleotides into human cells. This advancement has the potential to revolutionize treatments for hereditary diseases and rare metabolic disorders, reducing risks associated with gene editing such as unexpected mutations or the activation of harmful genes.<\/p>\n<\/blockquote>\n<\/div>\n<div class=\"wp-block-group\" style=\"margin-top:40px;margin-bottom:40px\">\n<h2 class=\"wp-block-heading\">VibeThinker-3B: \u5c0f\u578b\u8a00\u8a9e\u30e2\u30c7\u30eb\u306b\u304a\u3051\u308b\u691c\u8a3c\u53ef\u80fd\u306a\u63a8\u8ad6\u306e\u30d5\u30ed\u30f3\u30c6\u30a3\u30a2\u3092\u63a2\u6c42<\/h2>\n<ul>\n<li><strong>\u539f\u984c:<\/strong> VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models<\/li>\n<\/ul>\n<h3 class=\"wp-block-heading\">\u5c02\u9580\u30a2\u30ca\u30ea\u30b9\u30c8\u306e\u5206\u6790<\/h3>\n<div class=\"ai-summary-content\">\n<p>\u3053\u306e\u6280\u8853\u30ec\u30dd\u30fc\u30c8\u306f\u300130\u5104\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u30b3\u30f3\u30d1\u30af\u30c8\u306a\u9ad8\u5bc6\u5ea6\u30e2\u30c7\u30eb\u300c<strong>VibeThinker-3B<\/strong>\u300d\u3092\u7d39\u4ecb\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u30e2\u30c7\u30eb\u306f\u3001\u53b3\u5bc6\u306a\u5c0f\u898f\u6a21\u30e2\u30c7\u30eb\u306e\u67a0\u7d44\u307f\u5185\u3067\u691c\u8a3c\u53ef\u80fd\u306a\u63a8\u8ad6\u80fd\u529b\u3092\u3069\u3053\u307e\u3067\u9ad8\u3081\u3089\u308c\u308b\u304b\u3092\u63a2\u6c42\u3059\u308b\u305f\u3081\u306b\u958b\u767a\u3055\u308c\u307e\u3057\u305f\u3002<strong>Spectrum-to-Signal<\/strong>\u30dd\u30b9\u30c8\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u30d1\u30e9\u30c0\u30a4\u30e0\u306b\u57fa\u3065\u3044\u3066\u3001\u30ab\u30ea\u30ad\u30e5\u30e9\u30e0\u30d9\u30fc\u30b9\u306e\u6559\u5e2b\u3042\u308a\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3001\u30de\u30eb\u30c1\u30c9\u30e1\u30a4\u30f3\u5f37\u5316\u5b66\u7fd2\u3001\u30aa\u30d5\u30e9\u30a4\u30f3\u81ea\u5df1\u84b8\u7559\u3092\u542b\u3080\u6700\u9069\u5316\u3055\u308c\u305f\u30d1\u30a4\u30d7\u30e9\u30a4\u30f3\u3092\u901a\u3058\u3066\u30e2\u30c7\u30eb\u304c\u4f53\u7cfb\u7684\u306b\u5f37\u5316\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u5b9f\u9a13\u8a55\u4fa1\u3067\u306f\u3001<strong>VibeThinker-3B<\/strong>\u304c\u975e\u5e38\u306b\u8981\u6c42\u306e\u53b3\u3057\u3044\u691c\u8a3c\u53ef\u80fd\u30bf\u30b9\u30af\u306b\u304a\u3044\u3066\u6700\u5148\u7aef\u306e\u6027\u80fd\u3092\u9054\u6210\u3057\u305f\u3053\u3068\u304c\u793a\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u5177\u4f53\u7684\u306b\u306f\u3001<strong>AIME26<\/strong>\u306794.3\u70b9\uff08\u30af\u30ec\u30fc\u30e0\u30ec\u30d9\u30eb\u306e\u30c6\u30b9\u30c8\u6642\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u306797.1\u70b9\uff09\u3001<strong>LiveCodeBench v6<\/strong>\u306780.2 Pass@1\u3092\u9054\u6210\u3057\u3001\u6700\u8fd1\u306e\u672a\u898b\u306e<strong>LeetCode<\/strong>\u30b3\u30f3\u30c6\u30b9\u30c8\u3067\u306f96.1%\u306e\u5408\u683c\u7387\u3092\u793a\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u306b\u3088\u308a\u3001<strong>DeepSeek V3.2<\/strong>\u3001<strong>GLM-5<\/strong>\u3001<strong>Gemini 3 Pro<\/strong>\u3068\u3044\u3063\u305f\u3001\u6841\u9055\u3044\u306b\u5927\u304d\u306a\u30d5\u30e9\u30c3\u30b0\u30b7\u30c3\u30d7\u30e2\u30c7\u30eb\u3068\u540c\u7b49\u304b\u305d\u308c\u4ee5\u4e0a\u306e\u6027\u80fd\u3092\u767a\u63ee\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u3055\u3089\u306b\u3001<strong>IFEval<\/strong>\u306793.4\u70b9\u3092\u8a18\u9332\u3057\u3001\u3053\u306e\u6975\u7aef\u306a\u63a8\u8ad6\u80fd\u529b\u306e\u5f37\u5316\u304c\u53b3\u5bc6\u306a\u6307\u793a\u5236\u5fa1\u6027\u3092\u640d\u306a\u308f\u306a\u3044\u3053\u3068\u3092\u78ba\u8a8d\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u3089\u306e\u767a\u898b\u306f\u3001\u300c<strong>Parametric Compression-Coverage Hypothesis<\/strong>\u300d\u3092\u52d5\u6a5f\u3065\u3051\u308b\u3082\u306e\u3067\u3042\u308a\u3001\u691c\u8a3c\u53ef\u80fd\u306a\u63a8\u8ad6\u306f\u30b3\u30f3\u30d1\u30af\u30c8\u306a\u63a8\u8ad6\u30b3\u30a2\u306b\u5727\u7e2e\u53ef\u80fd\u3067\u3042\u308b\u4e00\u65b9\u3001\u30aa\u30fc\u30d7\u30f3\u306a\u30c9\u30e1\u30a4\u30f3\u77e5\u8b58\u3068\u6c4e\u7528\u7684\u306a\u80fd\u529b\u306b\u306f\u3001\u4e8b\u5b9f\u3001\u6982\u5ff5\u3001\u30ed\u30f3\u30b0\u30c6\u30fc\u30eb\u30b7\u30ca\u30ea\u30aa\u306b\u308f\u305f\u308b\u5e83\u7bc4\u306a\u30d1\u30e9\u30e1\u30fc\u30bf\u30ab\u30d0\u30ec\u30c3\u30b8\u304c\u5fc5\u8981\u3067\u3042\u308b\u3068\u3044\u3046\u898b\u65b9\u3092\u793a\u5506\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\ud83d\udc49 <strong><a href=\"https:\/\/arxiv.org\/abs\/2606.16140\" target=\"_blank\" rel=\"noopener\">arXiv \u3067\u8a18\u4e8b\u5168\u6587\u3092\u8aad\u3080<\/a><\/strong><\/p>\n<\/div>\n<ul>\n<li><strong>\u8981\u70b9:<\/strong> VibeThinker-3B, a 3-billion-parameter model, demonstrates frontier-level verifiable reasoning performance, matching or exceeding much larger flagship models, suggesting that complex reasoning capabilities can be efficiently compressed into compact AI models without sacrificing instruction controllability.<\/li>\n<li><strong>\u8457\u8005:<\/strong> Sen Xu, Shixi Liu, Wei Wang, Jixin Min, Yingwei Dai, Zhibin Yin, Yirong Chen, Xin Zhou, Junlin Zhang<\/li>\n<\/ul>\n<blockquote class=\"wp-block-quote\"><p><span>English Summary:<\/span><\/p>\n<p>This technical report introduces <strong>VibeThinker-3B<\/strong>, a compact dense model with 3 billion parameters, developed to explore the limits of verifiable reasoning within a strictly small-model regime. Building on the <strong>Spectrum-to-Signal<\/strong> post-training paradigm, the model is systematically enhanced through an optimized pipeline that includes curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation.<\/p>\n<p>Experimental evaluations demonstrate that <strong>VibeThinker-3B<\/strong> achieves frontier-level performance on highly demanding verifiable tasks. Specifically, it scored 94.3 on <strong>AIME26<\/strong> (improving to 97.1 with claim-level test-time scaling), 80.2 Pass@1 on <strong>LiveCodeBench v6<\/strong>, and exhibited strong out-of-distribution generalization with a 96.1% acceptance rate on recent unseen <strong>LeetCode<\/strong> contests. This performance places it in the band of first-tier reasoning systems, matching or exceeding flagship models that are orders of magnitude larger, such as <strong>DeepSeek V3.2<\/strong>, <strong>GLM-5<\/strong>, and <strong>Gemini 3 Pro<\/strong>.<\/p>\n<p>Furthermore, a score of 93.4 on <strong>IFEval<\/strong> confirms that this extreme reasoning enhancement does not compromise strict instruction controllability. These findings motivate the \"<strong>Parametric Compression-Coverage Hypothesis<\/strong>,\" which suggests that verifiable reasoning can be compressed into compact reasoning cores, while open-domain knowledge and general-purpose competence require broad parameter coverage over facts, concepts, and long-tail scenarios.<\/p>\n<\/blockquote>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>OpenAI\u306eDaybreak\u306b\u3088\u308bAI\u30b5\u30a4\u30d0\u30fc\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u5f37\u5316\u3001CRISPR\u306eDNA\u975e\u5207\u65ad\u907a\u4f1d\u5b50\u633f\u5165\u3001VibeThinker-3B\u306b\u3088\u308b\u5c0f\u578bLLM\u306e\u63a8\u8ad6\u80fd\u529b\u5411\u4e0a\u306b\u95a2\u3059\u308b\u6700\u65b0\u6280\u8853\u52d5\u5411\u3092\u6df1\u6398\u308a\u3002<\/p>\n","protected":false},"author":1,"featured_media":849,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"vkexunit_cta_each_option":"","footnotes":""},"categories":[3],"tags":[8,467,16,57,67,468,56,26,15,466],"class_list":{"0":"post-1072","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","6":"hentry","7":"category-columns","8":"tag-ai","9":"tag-crispr","10":"tag-llm","11":"tag-openai","12":"tag-vibe-coding","13":"tag-vibethinker","14":"tag-56","15":"tag-26","17":"tag-466"},"_links":{"self":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/posts\/1072","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/comments?post=1072"}],"version-history":[{"count":0,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/posts\/1072\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/media\/849"}],"wp:attachment":[{"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/media?parent=1072"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/categories?post=1072"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itexplore.org\/jp\/wp-json\/wp\/v2\/tags?post=1072"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}