{"id":7077,"date":"2025-07-14T06:13:00","date_gmt":"2025-07-13T21:13:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7077"},"modified":"2025-07-13T21:30:58","modified_gmt":"2025-07-13T12:30:58","slug":"grok-4-phi4-mini-flash-reasoning-smollm3-kimi-k2-t5gemma","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7077","title":{"rendered":"Grok 4, Phi4-mini-Flash-Reasoning, SmolLM3, Kimi-K2, T5Gemma"},"content":{"rendered":"\n<p>\u5148\u9031\u3082\u69d8\u3005\u306a\u30e2\u30c7\u30eb\u304c\u767a\u8868\u3055\u308c\u305f\u304c\u3001\u6ce8\u76ee\u306f\u69d8\u3005\u306a\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3067\u5f37\u529b\u306a\u6027\u80fd\u3092\u4e3b\u5f35\u3059\u308bGrok 4\u3060\u308d\u3046\uff08<a href=\"https:\/\/x.ai\/news\/grok-4\">Grok 4 | xAI<\/a>\uff09\u3002Humanity&#8217;s Last Exam\u306744.4%\u3068\u975e\u5e38\u306b\u5f37\u529b\u306b\u898b\u3048\u308b\u3002<\/p>\n\n\n\n<p>\u30aa\u30fc\u30d7\u30f3\u306a\u30e2\u30c7\u30eb\u3068\u3057\u3066\u306f\u30e2\u30c7\u30eb\u69cb\u9020\u304c\u9762\u767d\u3044Phi4-mini-Flash-Reasoning\uff08<a href=\"https:\/\/azure.microsoft.com\/en-us\/blog\/reasoning-reimagined-introducing-phi-4-mini-flash-reasoning\/?msockid=3249593f820a60fe1bf04ad683e06139\">Reasoning reimagined: Introducing Phi-4-mini-flash-reasoning | Microsoft Azure Blog<\/a>\u3001\u8ad6\u6587\u306f\u5f8c\u8ff0\uff09\u3001HuggingFace\u306e\u5c0f\u578b\u30e2\u30c7\u30ebSmolLM3\uff08<a href=\"https:\/\/huggingface.co\/docs\/transformers\/model_doc\/smollm3\">SmolLM3<\/a>, <a href=\"https:\/\/github.com\/huggingface\/smollm\">GitHub &#8211; huggingface\/smollm: Everything about the SmolLM and SmolVLM family of models<\/a>\uff09\u3001\u7dcf\u30d1\u30e9\u30e1\u30fc\u30bf1T \/ 32 B Active\u3068\u6975\u7aef\u306aMoE\u69cb\u6210\u3067\u975e\u5e38\u306b\u9ad8\u6027\u80fd\u306aKimi-K2\uff08<a href=\"https:\/\/github.com\/MoonshotAI\/Kimi-K2\">GitHub &#8211; MoonshotAI\/Kimi-K2: Kimi K2 is the large language model series developed by Moonshot AI team<\/a>\u3001<a href=\"https:\/\/moonshotai.github.io\/Kimi-K2\/\">Kimi K2<\/a>\uff09\u306a\u3069\u8208\u5473\u6df1\u3044\u767a\u8868\u304c\u76f8\u6b21\u3044\u3060\u3002\u307e\u305f\u3001<a href=\"https:\/\/developers.googleblog.com\/en\/t5gemma\/\">T5Gemma: A new collection of encoder-decoder Gemma models &#8211; Google Developers Blog<\/a>\u306b\u3082\u8981\u6ce8\u76ee\u3002Decoder only\u3067\u306a\u3044\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306e\u826f\u3055\u304c\u73fe\u308c\u308b\u30bf\u30b9\u30af\u3082\u591a\u305d\u3046\u306b\u601d\u3046\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Encoder-Decoder Gemma: Improving the Quality-Efficiency Trade-Off via Adaptation&nbsp;<\/strong>[52.2]<br>\u6211\u3005\u306f,\u30c7\u30b3\u30fc\u30c0\u306e\u307f\u306e\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\u3092\u30a8\u30f3\u30b3\u30fc\u30c0-\u30c7\u30b3\u30fc\u30c0\u30e2\u30c7\u30eb\u306b\u9069\u5fdc\u3055\u305b\u308b\u3068\u3044\u3046,\u65b0\u3057\u3044\u554f\u984c\u3092\u7814\u7a76\u3059\u308b\u3002 \u9069\u5fdc\u306f\u30c7\u30b3\u30fc\u30c0\u306e\u307f\u306eLLM\u306e\u80fd\u529b\u3092\u7d99\u627f\u3059\u308b\u3060\u3051\u3067\u306a\u304f\u3001\u8a08\u7b97\u306e\u9700\u8981\u3092\u6e1b\u3089\u3059\u3053\u3068\u304c\u3067\u304d\u308b\u3068\u4e3b\u5f35\u3057\u3066\u3044\u308b\u3002 \u540c\u69d8\u306e\u63a8\u8ad6\u4e88\u7b97\u306e\u4e0b\u3067\u306f\u3001\u30a8\u30f3\u30b3\u30fc\u30c0-\u30c7\u30b3\u30fc\u30c0 LLM \u306f(\u3057\u3070\u3057\u3070\u3088\u308a\u512a\u308c\u305f)\u4e8b\u524d\u8a13\u7df4\u6027\u80fd\u3092\u9054\u6210\u3067\u304d\u308b\u304c\u3001\u30c7\u30b3\u30fc\u30c0\u306e\u307f\u306e\u6027\u80fd\u3088\u308a\u3082\u306f\u308b\u304b\u306b\u512a\u308c\u305f\u5fae\u8abf\u6574\u6027\u80fd\u304c\u5f97\u3089\u308c\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2504.06225v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2504.06225v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Tue, 08 Apr 2025 17:13:41 GMT)<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation&nbsp;<\/strong>[129.5]<br>\u6211\u3005\u306f\u3001\u30ec\u30a4\u30e4\u9593\u306e\u52b9\u7387\u7684\u306a\u30e1\u30e2\u30ea\u5171\u6709\u306e\u305f\u3081\u306e\u30b7\u30f3\u30d7\u30eb\u3067\u52b9\u679c\u7684\u306a\u30e1\u30ab\u30cb\u30ba\u30e0\u3067\u3042\u308bGated Memory Unit(GMU)\u3092\u7d39\u4ecb\u3057\u305f\u3002 \u3053\u308c\u306f\u3001GMU\u3092\u7d44\u307f\u8fbc\u3093\u3067Samba\u30d9\u30fc\u30b9\u306e\u30bb\u30eb\u30d5\u30c7\u30b3\u30fc\u30c0\u304b\u3089\u30e1\u30e2\u30ea\u8aad\u307f\u51fa\u3057\u72b6\u614b\u3092\u5171\u6709\u3059\u308b\u30c7\u30b3\u30fc\u30c0\u30fb\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u30fb\u30c7\u30b3\u30fc\u30c0\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3067\u3042\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.06607v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.06607v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Wed, 09 Jul 2025 07:27:00 GMT)<\/li>\n\n\n\n<li>Phi4-mini-Flash-Reasoning\u306e\u8ad6\u6587<\/li>\n\n\n\n<li>\u300cOur decoder-hybrid-decoder architecture taking Samba [RLL+25] as the self-decoder. Gated Memory Units (GMUs) are interleaved with the cross-attention layers in the cross-decoder to reduce the decoding complexity. As in YOCO [SDZ+24], the full attention layer only need to compute the KV cache during prefilling with the self-decoder, leading to linear computation complexity for the prefill stage.\u300d\u3068\u8a08\u7b97\u91cf\u7684\u306b\u6709\u5229\u306a\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3067LRM\u306b\u9069\u3057\u3066\u3044\u308b\u3088\u3046\u306b\u898b\u3048\u308b\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities&nbsp;<\/strong>[1584.5]<br>Gemini 2.5 Pro\u306f\u79c1\u305f\u3061\u306e\u6700\u3082\u6709\u80fd\u306a\u30e2\u30c7\u30eb\u3067\u3042\u308a\u3001\u30d5\u30ed\u30f3\u30c6\u30a3\u30a2\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u3068\u63a8\u8ad6\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3067SoTA\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u5b9f\u73fe\u3057\u3066\u3044\u307e\u3059\u3002 Gemini 2.5 Flash\u306f\u8a08\u7b97\u3068\u30ec\u30a4\u30c6\u30f3\u30b7\u306e\u8981\u6c42\u306e\u3054\u304f\u4e00\u90e8\u3067\u512a\u308c\u305f\u63a8\u8ad6\u6a5f\u80fd\u3092\u63d0\u4f9b\u3059\u308b\u3002 Gemini 2.0 Flash\u3068Flash-Lite\u306f\u4f4e\u30ec\u30a4\u30c6\u30f3\u30b7\u3068\u4f4e\u30b3\u30b9\u30c8\u3067\u30cf\u30a4\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u63d0\u4f9b\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.06261v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.06261v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Mon, 07 Jul 2025 17:36:04 GMT)<\/li>\n\n\n\n<li>Gemini 2.5\u306e\u8ad6\u6587\u3082\u51fa\u3066\u3044\u305f\u3002\u5171\u8457\u8005\u306e\u4eba\u6570\u304c\u3059\u3054\u3044\uff083300\u4eba\u4ee5\u4e0a\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>SciMaster: Towards General-Purpose Scientific AI Agents, Part I. X-Master as Foundation: Can We Lead on Humanity&#8217;s Last Exam?&nbsp;<\/strong>[47.2]<br>\u672c\u7a3f\u3067\u306f,\u4eba\u9593\u7814\u7a76\u8005\u3092\u30a8\u30df\u30e5\u30ec\u30fc\u30c8\u3059\u308b\u30c4\u30fc\u30eb\u5f37\u5316\u63a8\u8ad6\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u3067\u3042\u308bX-Master\u3092\u7d39\u4ecb\u3059\u308b\u3002 X\u30de\u30b9\u30bf\u30fc\u30ba\u306fHumanity&#8217;s Last Exam\u306b32.1%\u306e\u30b9\u30b3\u30a2\u3067\u6700\u65b0\u8a18\u9332\u3092\u6a39\u7acb\u3057\u305f\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.05241v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.05241v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Mon, 07 Jul 2025 17:50:52 GMT)<\/li>\n\n\n\n<li>Agentic\u306a\u30a2\u30d7\u30ed\u30fc\u30c1\uff0bDeepSeek-R1-0528\u3067Humanity&#8217;s Last Exam 32.1%\u3092\u9054\u6210\u3068\u3044\u3046\u5831\u544a\u3002\u30d9\u30fc\u30b9\u30e2\u30c7\u30eb\u3068\u3057\u3066Grok 4\u3092\u4f7f\u3063\u305f\u5834\u5408\u306e\u30b9\u30b3\u30a2\u304c\u6c17\u306b\u306a\u308b\u3068\u3053\u308d\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/sjtu-sai-agents\/X-Master\">GitHub &#8211; sjtu-sai-agents\/X-Master: Official implementation of X-Master, a general-purpose tool-augmented reasoning agent.<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u5148\u9031\u3082\u69d8\u3005\u306a\u30e2\u30c7\u30eb\u304c\u767a\u8868\u3055\u308c\u305f\u304c\u3001\u6ce8\u76ee\u306f\u69d8\u3005\u306a\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3067\u5f37\u529b\u306a\u6027\u80fd\u3092\u4e3b\u5f35\u3059\u308bGrok 4\u3060\u308d\u3046\uff08Grok 4 | xAI\uff09\u3002Humanity&#8217;s Last Exam\u306744.4%\u3068\u975e\u5e38\u306b\u5f37\u529b\u306b\u898b\u3048\u308b\u3002 \u30aa\u30fc &hellip; <a href=\"https:\/\/devneko.jp\/wordpress\/?p=7077\" class=\"more-link\"><span class=\"screen-reader-text\">&#8220;Grok 4, Phi4-mini-Flash-Reasoning, SmolLM3, Kimi-K2, T5Gemma&#8221; \u306e<\/span>\u7d9a\u304d\u3092\u8aad\u3080<\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[223,232,376],"class_list":["post-7077","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-llm","tag-lrm","tag-ssm"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7077","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=7077"}],"version-history":[{"count":6,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7077\/revisions"}],"predecessor-version":[{"id":7123,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7077\/revisions\/7123"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7077"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7077"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7077"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}