{"id":6248,"date":"2025-10-09T12:14:31","date_gmt":"2025-10-09T04:14:31","guid":{"rendered":"https:\/\/nullthought.net\/?p=6248"},"modified":"2025-10-09T12:14:32","modified_gmt":"2025-10-09T04:14:32","slug":"7m%e5%8f%82%e6%95%b0%e5%b0%8f%e6%a8%a1%e5%9e%8b%ef%bc%88tiny-recursive-model%ef%bc%8ctrm%ef%bc%89%ef%bc%8c%e5%9c%a8%e6%95%b0%e7%8b%ac%e3%80%81%e8%bf%b7%e5%ae%ab%e3%80%81arc-agi-%e7%ad%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6248","title":{"rendered":"7M\u53c2\u6570\u5c0f\u6a21\u578b\uff08Tiny Recursive Model\uff0cTRM\uff09\uff0c\u5728\u6570\u72ec\u3001\u8ff7\u5bab\u3001ARC-AGI \u7b49\u201c\u786c\u63a8\u7406\u201d\u4efb\u52a1\u4e0a\u8868\u73b0\u4f18\u4e8e\u5927\u53c2\u6570LLM"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2510.04871\" target=\"_blank\" rel=\"noreferrer noopener\">Less is More: Recursive Reasoning with Tiny Networks<\/a><\/strong>\u9488\u5bf9\u5728\u6570\u72ec\u3001\u8ff7\u5bab\u3001ARC-AGI \u7b49\u201c\u786c\u63a8\u7406\u201d\u4efb\u52a1\u4e0a\uff0c\u4f20\u7edf\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5373\u4fbf\u7ed3\u5408\u94fe\u5f0f\u601d\u7ef4\uff08CoT\uff09\u4e0e\u6d4b\u8bd5\u65f6\u8ba1\u7b97\uff08TTC\uff09\u4ecd\u96be\u4ee5\u8fbe\u6210\u4eba\u7c7b\u7ea7\u8868\u73b0\u7684\u95ee\u9898\uff0c\u63d0\u51fa\u4e86\u4e00\u4e2a\u663e\u8457\u66f4\u7b80\u6d01\u4e14\u66f4\u5f3a\u6cdb\u5316\u7684\u9012\u5f52\u63a8\u7406\u8303\u5f0f\u2014\u2014Tiny Recursive Model\uff08TRM\uff09\u3002TRM\u7528<strong>\u5355\u4e2a\u4ec5\u4e24\u5c42\u7684\u5c0f\u7f51\u7edc<\/strong>\uff0c\u901a\u8fc7<strong>\u5bf9\u201c\u6f5c\u5728\u63a8\u7406\u72b6\u6001 z\u201d\u4e0e\u201c\u5f53\u524d\u89e3 y\u201d\u4ea4\u66ff\u9012\u5f52\u66f4\u65b0<\/strong>\u7684\u65b9\u5f0f\uff0c\u5728\u5c0f\u6570\u636e\uff08~1K\uff09\u548c\u5c0f\u6a21\u578b\uff08~7M \u53c2\u6570\uff09\u6761\u4ef6\u4e0b\uff0c\u8d85\u8d8a\u4e86\u4f7f\u7528\u4e24\u5957\u7f51\u7edc\u3001\u4f9d\u8d56\u56fa\u5b9a\u70b9\u8fd1\u4f3c\u4e0e Q-learning \u5f0f ACT \u7684\u5c42\u7ea7\u9012\u5f52\u6a21\u578b HRM\uff0c\u5e76\u5728\u591a\u4e2a\u57fa\u51c6\u4e0a\u4f18\u4e8e\u5927\u91cf\u53c2\u6570\u7ea7\u522b\u8fdc\u8d85\u7684 LLM\u3002\u4f5c\u8005\u62a5\u544a\uff1aSudoku-Extreme 87.4%\u3001Maze-Hard 85.3%\u3001ARC-AGI-1 44.6%\u3001ARC-AGI-2 7.8%\uff0c\u5176\u4e2d\u5bf9\u6bd4 HRM \u7684 55.0%\u300174.5%\u300140.3%\u30015.0%\u5747\u6709\u63d0\u5347\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aAlexia Jolicoeur-Martineau\uff0c\u6765\u81eaSamsung\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"435\" height=\"825\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-2.png\" alt=\"\" class=\"wp-image-6250\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-2.png 435w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-2-158x300.png 158w\" sizes=\"auto, (max-width: 435px) 100vw, 435px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2510.04871\" target=\"_blank\" rel=\"noreferrer noopener\">Less is More: Recursive Reasoning with Tiny Networks<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e00\u3001\u80cc\u666f\uff1aHRM \u7684\u7406\u5ff5\u4e0e\u74f6\u9888<\/h5>\n\n\n\n<p>HRM\uff08Hierarchical Reasoning Model\uff09\u63d0\u51fa\u4ee5<strong>\u4e24\u5957\u5c0f\u7f51\u7edc<\/strong>\u5728<strong>\u4e0d\u540c\u9891\u7387<\/strong>\uff08\u4f4e\u5c42 fL \u9ad8\u9891\u3001 \u9ad8\u5c42 fH \u4f4e\u9891\uff09\u9012\u5f52\u66f4\u65b0\u4e24\u4e2a\u6f5c\u53d8\u91cf\uff08zL\u3001zH\uff09\uff0c\u5e76\u5728<strong>\u6df1\u5ea6\u76d1\u7763<\/strong>\uff08deep supervision\uff09\u4e0b\u591a\u8f6e\u6539\u8fdb\u7b54\u6848\uff1b\u8bad\u7ec3\u65f6\u5f15\u5165<strong>\u81ea\u9002\u5e94\u8ba1\u7b97\u65f6\u95f4 ACT<\/strong>\u4ee5\u65e9\u505c\uff1b\u4e3a\u964d\u4f4e\u5185\u5b58\uff0c\u91c7\u7528<strong>\u9690\u51fd\u6570\u5b9a\u7406\uff08IFT\uff09+ \u4e00\u6b65\u68af\u5ea6\u8fd1\u4f3c<\/strong>\uff0c\u53ea\u5bf9\u6700\u540e\u5c11\u91cf\u9012\u5f52\u6b65\u56de\u4f20\u68af\u5ea6\u3002\u8be5\u65b9\u6cd5\u867d\u5728 Sudoku\u3001Maze\u3001ARC-AGI \u6709\u7a81\u7834\uff0c\u4f46\u5b58\u5728\u4e09\u7c7b\u5173\u952e\u95ee\u9898\uff1a<br>1\uff09<strong>\u56fa\u5b9a\u70b9\u5047\u8bbe\u5b58\u7591<\/strong>\uff1a\u5b9e\u9645\u9012\u5f52\u6b65\u6570\u5c0f\uff08\u5e38\u7528 n=2\u3001T=2\uff09\uff0c\u96be\u4fdd\u8bc1\u5df2\u8fbe\u56fa\u5b9a\u70b9\u5374\u4f7f\u7528\u4e00\u6b65\u68af\u5ea6\u8fd1\u4f3c\uff1b\u8bba\u6587\u4e5f\u63f4\u5f15\u7b2c\u4e09\u65b9\u5206\u6790\u6307\u51fa\uff0c<strong>\u6df1\u5ea6\u76d1\u7763<\/strong>\u800c\u975e\u201c\u5c42\u7ea7\u9012\u5f52\u201d\u4f3c\u662f\u4e3b\u8981\u6536\u76ca\u6765\u6e90\u3002<br>2\uff09<strong>ACT \u8bad\u7ec3\u6210\u672c\u504f\u9ad8<\/strong>\uff1a\u91c7\u7528 Q-learning \u5f62\u5f0f\u7684 halt\/continue \u4fe1\u53f7\uff0c\u5176\u4e2d\u201ccontinue\u201d\u5206\u652f\u9700\u8981<strong>\u989d\u5916\u4e00\u6b21\u5b8c\u6574\u524d\u5411<\/strong>\uff0c\u5bfc\u81f4\u6bcf\u6b65\u4f18\u5316\u9700\u4e24\u6b21\u524d\u5411\u3002<br>3\uff09<strong>\u751f\u7269\u5b66\u9690\u55bb\u590d\u6742\u4e14\u96be\u4ee5\u505a\u56e0\u679c\u5f52\u56e0<\/strong>\uff1a\u4e24\u5c42\u7ea7\u3001\u4e24\u6f5c\u53d8\u91cf\u7684\u8bbe\u8ba1\u66f4\u50cf\u662f\u5de5\u7a0b\u8bbe\u5b9a\uff0c\u7f3a\u4e4f\u6e05\u6670\u7b80\u5316\u89e3\u91ca\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e8c\u3001TRM \u7684\u601d\u60f3\u91cd\u6784\uff1a\u628a\u201c\u5c42\u7ea7\u201d\u5316\u7e41\u4e3a\u7b80<\/h5>\n\n\n\n<p>TRM\u629b\u5f03\u201c\u4e8c\u7f51\u7edc\u4e8c\u5c42\u7ea7\u201d\u7684\u751f\u7269\u5b66\u9690\u55bb\uff0c\u8f6c\u800c\u7ed9\u51fa<strong>\u66f4\u76f4\u89c2\u7684\u4e8c\u53d8\u91cf\u89e3\u91ca<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>y<\/strong>\uff1a\u5f53\u524d\uff08\u5d4c\u5165\u540e\u7684\uff09<strong>\u5019\u9009\u89e3<\/strong>\uff1b<\/li>\n\n\n\n<li><strong>z<\/strong>\uff1a\u4e0d\u53ef\u76f4\u63a5\u89e3\u7801\u7684<strong>\u6f5c\u5728\u63a8\u7406\u72b6\u6001<\/strong>\uff08\u7c7b\u4f3c\u201c\u53ef\u88ab\u8fed\u4ee3\u7684\u601d\u8003\u75d5\u8ff9\/\u4e2d\u95f4\u8bc1\u636e\u201d\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u9012\u5f52\u8fc7\u7a0b\u5728\u6bcf\u4e2a\u201c\u6539\u8fdb\u56de\u5408\u201d\u91cc\u5206\u4e24\u6b65\u4ea4\u66ff\uff1a<br>1\uff09\u7528 (x,y,z) \u66f4\u65b0 z\uff08\u201c\u5148\u60f3\u6e05\u695a\u201d\uff09\uff1b<br>2\uff09\u7528 (y,z) \u66f4\u65b0 y\uff08\u201c\u518d\u6539\u7b54\u6848\u201d\uff09\u3002<\/p>\n\n\n\n<p>\u5173\u952e\u70b9\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53ea\u7528<strong>\u4e00\u4e2a\u4e24\u5c42\u7f51\u7edc<\/strong>\u540c\u65f6\u627f\u62c5\u201c\u66f4\u65b0 z\u201d\u548c\u201c\u66f4\u65b0 y\u201d\u7684\u529f\u80fd\uff1b\u662f\u5426\u5305\u542b\u8f93\u5165 x \u51b3\u5b9a\u5b50\u4efb\u52a1\uff08\u66f4\u65b0 z \u7684\u5206\u652f\u8bfb\u5165 x\uff0c\u66f4\u65b0 y \u7684\u5206\u652f\u4e0d\u8bfb\u5165 x\uff09\u3002<\/li>\n\n\n\n<li>\u5728\u6df1\u5ea6\u76d1\u7763\u4e0b\uff0c\u6a21\u578b\u5b66\u4f1a\u201c\u63a5\u8fc7\u4efb\u610f (y,z)\u201d\u518d\u6539\u8fdb\u4e00\u70b9\uff0c\u56e0\u6b64\u53ef\u4ee5\u5148\u505a T\u22121 \u6b21<strong>\u65e0\u68af\u5ea6<\/strong>\u9012\u5f52\u9760\u8fd1\u89e3\u9762\uff0c\u518d\u505a 1 \u6b21<strong>\u5e26\u68af\u5ea6<\/strong>\u9012\u5f52\u8fdb\u884c\u5b66\u4e60\uff0c<strong>\u65e0\u9700 IFT\/\u4e00\u6b65\u8fd1\u4f3c<\/strong>\u3002<\/li>\n\n\n\n<li><strong>ACT \u7b80\u5316<\/strong>\uff1a\u53ea\u4fdd\u7559\u201c\u662f\u5426\u5df2\u8db3\u591f\u597d\u53ef\u505c\u201d\u7684<strong>\u5355\u4e00\u4e8c\u5206\u7c7b<\/strong>\uff08\u4e0e\u201c\u9884\u6d4b\u6b63\u786e\u201d\u76f8\u5173\u8054\uff09\uff0c\u53bb\u6389\u201ccontinue\u201d\u635f\u5931\uff0c\u4ece\u800c<strong>\u907f\u514d\u989d\u5916\u524d\u5411<\/strong>\u3002<\/li>\n\n\n\n<li><strong>EMA<\/strong>\uff08\u6307\u6570\u6ed1\u52a8\u5e73\u5747\uff09\u7a33\u5b9a\u8bad\u7ec3\uff0c\u5c24\u5176\u5c0f\u6570\u636e\u573a\u666f\u4e0b\u663e\u8457\u6291\u5236\u8fc7\u62df\u5408\u53d1\u6563\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e09\u3001\u7b97\u6cd5\u673a\u5236\u4e0e\u8bad\u7ec3\u7ec6\u8282\uff08\u76f4\u89c2\u89e3\u8bfb\uff09<\/h5>\n\n\n\n<p>1\uff09<strong>\u9012\u5f52\u5355\u5143\uff08latent recursion\uff09<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u91cd\u590d n \u6b21\uff1az\u2190net(x,y,z)\uff1b<\/li>\n\n\n\n<li>\u7136\u540e\u4e00\u6b21\uff1ay\u2190net(y,z)\uff1b<br>\u8fd9\u6784\u6210\u4e00\u6b21\u201c\u5b8c\u6574\u9012\u5f52\u8fc7\u7a0b\uff08full recursion\uff09\u201d\u3002<\/li>\n<\/ul>\n\n\n\n<p>2\uff09<strong>\u6df1\u5ea6\u76d1\u7763\uff08N<sub>sup<\/sub> \u2264 16\uff09<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a\u76d1\u7763\u6b65\uff1a\u5148\u8dd1 T\u22121 \u6b21<strong>\u65e0\u68af\u5ea6<\/strong>\u5b8c\u6574\u9012\u5f52\uff0c\u8fdb\u4e00\u6b65\u201c\u6696\u673a\u201d(y,z)\uff1b\u518d\u8dd1 1 \u6b21<strong>\u5e26\u68af\u5ea6<\/strong>\u5b8c\u6574\u9012\u5f52\u5e76\u8ba1\u7b97\u4e3b\u635f\u5931\uff08\u4ea4\u53c9\u71b5 w.r.t. \u771f\u503c\uff09\u3002<\/li>\n\n\n\n<li>\u8bad\u7ec3\u65f6\u9644\u5e26\u4e00\u4e2a<strong>halt \u6982\u7387<\/strong>\uff08BCE \u635f\u5931\uff09\u2192 \u82e5\u201c\u770b\u8d77\u6765\u5df2\u6b63\u786e\u201d\uff0c\u53ef<strong>\u65e9\u505c<\/strong>\u7ee7\u7eed\u4e0b\u4e00\u4e2a\u6837\u672c\uff0c\u4ece\u800c\u5728\u56fa\u5b9a\u603b\u8fed\u4ee3\u9884\u7b97\u4e0b\u8986\u76d6\u66f4\u591a\u6837\u672c\u3002<\/li>\n<\/ul>\n\n\n\n<p>3\uff09<strong>\u4e3a\u4ec0\u4e48\u53ea\u8981\u4e24\u4e2a\u7279\u5f81\uff08y \u4e0e z\uff09\u5c31\u591f\uff1f<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u82e5<strong>\u6ca1\u6709 y<\/strong>\uff0c\u6a21\u578b\u88ab\u8feb\u5728 z \u5185\u540c\u65f6\u4fdd\u5b58\u201c\u89e3\u201d\u548c\u201c\u63a8\u7406\u201d\uff0c\u51b2\u7a81\u5927\u3001\u96be\u8bad\u7ec3\uff1b<\/li>\n\n\n\n<li>\u82e5\u628a z \u518d<strong>\u62c6\u591a\u4efd<\/strong>\uff08\u591a\u5c3a\u5ea6 z\uff09\uff0c\u5f15\u5165\u5197\u4f59\u8026\u5408\u53cd\u800c\u964d\u4f4e\u6cdb\u5316\uff1b<\/li>\n\n\n\n<li>\u5b9e\u9a8c\u8bc1\u660e\u201c\u4e24\u53d8\u91cf\u6700\u4f73\u201d\uff0c\u5355 z \u6216\u591a z \u5747\u663e\u8457\u6389\u70b9\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u56db\u3001\u4e0e HRM \u7684\u7cfb\u7edf\u6027\u5bf9\u6bd4\uff08\u8bbe\u8ba1\u2192\u8bad\u7ec3\u2192\u4ee3\u4ef7\uff09<\/h4>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u7ef4\u5ea6<\/th><th>HRM<\/th><th>TRM<\/th><\/tr><\/thead><tbody><tr><td>\u7ed3\u6784<\/td><td>\u4e24\u7f51\u7edc\uff08fL \u9ad8\u9891\u3001fH \u4f4e\u9891\uff09\uff1bzL\u3001zH<\/td><td>\u5355\u7f51\u7edc\uff1b\u201cy=\u5019\u9009\u89e3\u3001z=\u6f5c\u5728\u63a8\u7406\u201d<\/td><\/tr><tr><td>\u9012\u5f52\u56de\u4f20<\/td><td>\u501a\u8d56 IFT \u7684\u201c\u4e00\u6b65\u68af\u5ea6\u8fd1\u4f3c\u201d\uff0c\u4ec5\u672b\u4e24\u6b65\u56de\u4f20<\/td><td>\u5bf9<strong>\u5b8c\u6574\u9012\u5f52<\/strong>\u56de\u4f20\u68af\u5ea6\uff0c\u65e0\u9700\u56fa\u5b9a\u70b9\u5047\u8bbe<\/td><\/tr><tr><td>\u6df1\u5ea6\u76d1\u7763<\/td><td>\u662f\uff08N<sub>sup<\/sub>\u226416\uff09<\/td><td>\u662f\uff08\u540c\uff09<\/td><\/tr><tr><td>ACT<\/td><td>Q-learning\uff1ahalt\/continue\uff0c<strong>\u9700\u989d\u5916\u4e00\u6b21\u524d\u5411<\/strong><\/td><td>\u4ec5 halt \u4e8c\u5206\u7c7b\uff0c<strong>\u4e0d\u9700\u989d\u5916\u524d\u5411<\/strong><\/td><\/tr><tr><td>\u5f52\u7eb3\u504f\u7f6e<\/td><td>\u751f\u7269\u5c42\u7ea7\u89e3\u91ca<\/td><td>\u4efb\u52a1\u76f4\u89c9\u89e3\u91ca\uff08\u5148\u60f3\u540e\u6539\uff09<\/td><\/tr><tr><td>\u53c2\u6570\u89c4\u6a21<\/td><td>\u5178\u578b 27M<\/td><td>\u5178\u578b 5\u20137M\uff082 \u5c42\uff09<\/td><\/tr><tr><td>\u8bad\u7ec3\u7a33\u5b9a\u6027<\/td><td>\u5bb9\u6613\u8fc7\u62df\u5408\u5e76\u5d29\u6e83<\/td><td>\u5f15\u5165 EMA \u663e\u8457\u7a33\u5b9a<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e94\u3001\u5b9e\u9a8c\u7ed3\u679c\u4e0e\u6d88\u878d\u7ed3\u8bba<\/h5>\n\n\n\n<p>1\uff09<strong>\u603b\u4f53\u6210\u7ee9\uff08\u76f4\u63a5\u76d1\u7763\uff0c\u5c0f\u6837\u672c\uff09<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Sudoku-Extreme\uff1aTRM-MLP 87.4%\uff08HRM 55.0%\uff09<\/li>\n\n\n\n<li>Maze-Hard\uff0830\u00d730 \u957f\u4e0a\u4e0b\u6587\uff09\uff1aTRM-Att 85.3%\uff08HRM 74.5%\uff09<\/li>\n\n\n\n<li>ARC-AGI-1\uff1aTRM-Att 44.6%\uff08HRM 40.3%\uff09<\/li>\n\n\n\n<li>ARC-AGI-2\uff1aTRM-Att 7.8%\uff08HRM 5.0%\uff09<br>TRM \u5728\u53c2\u6570\u66f4\u5c0f\uff08~7M\uff09\u4e0b\u666e\u904d\u4f18\u4e8e HRM\uff08~27M\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>2\uff09<strong>\u81ea\u6ce8\u610f\u529b vs. MLP Mixer \u5f0f\u5e8f\u5217 MLP<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5c0f\u4e14\u56fa\u5b9a\u4e0a\u4e0b\u6587\uff08\u5982 9\u00d79 \u6570\u72ec\uff09<\/strong>\uff1a\u7528 MLP \u4ee3\u81ea\u6ce8\u610f\u529b\u66f4\u4f18\uff0887.4% vs 74.7%\uff09\uff0c\u56e0 L\u2264D \u65f6 MLP \u4ee5\u66f4\u5c11\u53c2\u6570\u8986\u76d6\u5168\u5e8f\u5217\u4ea4\u4e92\u3002<\/li>\n\n\n\n<li><strong>\u8f83\u5927\u7f51\u683c\uff0830\u00d730 \u8ff7\u5bab\u3001ARC\uff09<\/strong>\uff1a\u81ea\u6ce8\u610f\u529b\u66f4\u5408\u9002\uff0cMLP \u53cd\u800c\u6b20\u62df\u5408\/\u8fc7\u62df\u5408\u5931\u8861\u3002<\/li>\n<\/ul>\n\n\n\n<p>3\uff09<strong>\u5c42\u6570\u4e0e\u9012\u5f52\u6b65\u6570\uff08\u201c\u5c11\u5373\u662f\u591a\u201d\uff09<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u589e\u52a0\u5c42\u6570\uff084 \u5c42\uff09\u4f1a<strong>\u6cdb\u5316\u4e0b\u964d<\/strong>\uff0879.5%\uff09\uff0c\u800c<strong>2 \u5c42 + \u589e\u52a0\u9012\u5f52<\/strong>\u6700\u4f18\uff0887.4%\uff09\uff0c\u8bf4\u660e\u5728\u5c0f\u6570\u636e regime \u4e0b\uff0c<strong>\u628a\u5bb9\u91cf\u8f6c\u79fb\u5230\u201c\u6df1\u9012\u5f52\/\u6df1\u76d1\u7763\u201d<\/strong>\u6bd4\u201c\u52a0\u5c42\u6570\u201d\u66f4\u80fd\u6291\u5236\u8fc7\u62df\u5408\u3002<\/li>\n<\/ul>\n\n\n\n<p>4\uff09<strong>EMA \u4e0e ACT \u7b80\u5316<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53bb\u6389 EMA \u4f1a\u663e\u8457\u6389\u70b9\uff0879.9%\uff09\uff1b<\/li>\n\n\n\n<li>\u5c06 ACT \u4ece Q-learning\uff08\u4e24\u524d\u5411\uff09\u6539\u4e3a\u201c\u4ec5 halt\u201d\u4e0d\u5f71\u54cd\u6cdb\u5316\uff0c<strong>\u8fd8\u8282\u7701\u8ba1\u7b97<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p>5\uff09<strong>\u7279\u5f81\u5f62\u6001\u6d88\u878d<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5355 z \u6216\u591a\u5c3a\u5ea6 z\uff08\u628a\u6bcf\u6b21\u9012\u5f52\u4ea7\u51fa\u7684 z_i \u90fd\u8de8\u6b65\u643a\u5e26\uff09\u90fd\u52a3\u4e8e\u201c\u4e24\u53d8\u91cf\uff08y,z\uff09\u201d\u57fa\u7ebf\uff0c\u9a8c\u8bc1\u201c\u4e8c\u800c\u4e0d\u7e41\u201d\u7684\u4f18\u52bf\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">\u516d\u3001\u65b9\u6cd5\u76f4\u89c9\u4e0e\u7406\u8bba\u601d\u8003<\/h5>\n\n\n\n<p>1\uff09<strong>\u4e3a\u4f55\u6df1\u9012\u5f52 + \u6df1\u76d1\u7763\u80fd\u6da8\u6cdb\u5316\uff1f<\/strong><br>\u5728\u5c0f\u6570\u636e\/\u5c0f\u6a21\u578b\u4e0b\uff0c\u76f4\u63a5\u201c\u52a0\u5bbd\u52a0\u6df1\u201d\u6613\u8fc7\u62df\u5408\uff1b\u53cd\u4e4b\uff0c<strong>\u901a\u8fc7\u591a\u6b21\u201c\u65e0\u68af\u5ea6\u81ea\u6821\u6b63\u201d\u628a\u72b6\u6001\u5e26\u5230\u66f4\u201c\u63a5\u8fd1\u53ef\u89e3\u6d41\u5f62\u201d\u7684\u533a\u57df<\/strong>\uff0c\u518d\u5728\u6700\u540e\u4e00\u6b21\u9012\u5f52\u8fdb\u884c\u5b66\u4e60\uff0c\u7b49\u4ef7\u4e8e\u201c\u628a\u68af\u5ea6\u6295\u5c04\u5230\u66f4\u6709\u4fe1\u606f\u7684\u5c40\u90e8\u90bb\u57df\u201d\uff0c\u8d77\u5230<strong>\u6b63\u5219\u5316<\/strong>\u4e0e<strong>\u4f18\u5316\u666f\u89c2\u91cd\u5851<\/strong>\u7684\u6548\u679c\u3002<\/p>\n\n\n\n<p>2\uff09<strong>y \u4e0e z \u7684\u89d2\u8272\u5206\u79bb<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>y \u5916\u663e\u3001\u53ef\u89e3\u7801\u4e14\u88ab\u5224\u5b9a\u5bf9\u9519\uff1bz \u5185\u9690\u3001\u7d2f\u79ef\u201c\u89e3\u51b3\u8be5\u9898\u7684\u601d\u8003\u7ebf\u7d22\u201d\u3002\u4e24\u8005\u5206\u79bb\u907f\u514d\u4fe1\u606f\u5185\u5377\u4e0e\u51b2\u7a81\u5b58\u50a8\uff0c\u8ba9\u201c\u60f3\u201d\uff08z\uff09\u4e0e\u201c\u6539\u201d\uff08y\uff09\u804c\u80fd\u6e05\u6670\uff0c\u9012\u5f52\u6548\u7387\u66f4\u9ad8\u3002<\/li>\n<\/ul>\n\n\n\n<p>3\uff09<strong>\u4e3a\u4f55\u65e0\u987b\u56fa\u5b9a\u70b9\u4e0e\u4e00\u6b65\u8fd1\u4f3c<\/strong><br>TRM \u4e0d\u662f\u8ffd\u6c42\u4e25\u683c\u56fa\u5b9a\u70b9\uff0c\u800c\u662f<strong>\u628a\u4e00\u6b21\u201c\u5b8c\u6574\u9012\u5f52\u8fc7\u7a0b\u201d\u89c6\u4e3a\u53ef\u5b66\u4e60\u7684\u6539\u8fdb\u7b97\u5b50<\/strong>\uff0c\u76f4\u63a5\u5bf9\u5176\u6c42\u68af\u5ea6\uff1b\u8fd9\u65e2\u89c4\u907f\u4e86\u56fa\u5b9a\u70b9\u662f\u5426\u5b58\u5728\/\u53ef\u6536\u655b\u7684\u524d\u63d0\uff0c\u53c8\u4fdd\u7559\u4e86\u201c\u591a\u6b21\u65e0\u68af\u5ea6\u6539\u8fdb\u201d\u7684\u4fbf\u5b9c\u6027\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e03\u3001\u5de5\u7a0b\u843d\u5730\u8981\u70b9\uff08\u590d\u73b0\u6307\u5357\u7ea7\u6458\u8981\uff09<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7f51\u7edc<\/strong>\uff1a2 \u5c42\uff08\u542b\u81ea\u6ce8\u610f\u529b\u6216\u5e8f\u5217 MLP\uff0c\u89c6\u4efb\u52a1\u4e0a\u4e0b\u6587\u957f\u5ea6\u800c\u5b9a\uff09\uff0c\u9690\u85cf\u7ef4 512\uff1b<\/li>\n\n\n\n<li><strong>\u9012\u5f52\u8d85\u53c2<\/strong>\uff1a\u63a8\u8350 T=3\u3001n=6\uff08Sudoku \u6700\u4f18\u70b9\uff09\uff0c\u6ce8\u610f n \u8fc7\u5927\u9700\u8003\u8651\u663e\u5b58\uff08TRM \u5bf9\u5b8c\u6574\u9012\u5f52\u56de\u4f20\uff09\uff1b<\/li>\n\n\n\n<li><strong>\u6df1\u5ea6\u76d1\u7763<\/strong>\uff1aN<sub>sup<\/sub> \u2264 16\uff1b\u8bad\u7ec3\u65f6\u6309\u76d1\u7763\u6b65\u5faa\u73af\uff0c\u201cT\u22121 \u6b21\u65e0\u68af\u5ea6 + 1 \u6b21\u6709\u68af\u5ea6\u201d\uff1b<\/li>\n\n\n\n<li><strong>ACT<\/strong>\uff1a\u4ec5\u8bad\u7ec3\u65f6\u9884\u6d4b halt \u6982\u7387\uff08BCE\uff09\uff0c\u4e3a\u8986\u76d6\u66f4\u591a\u6837\u672c\u53ef\u65e9\u505c\uff1b\u6d4b\u8bd5\u65f6\u901a\u5e38\u8dd1\u6ee1 N<sub>sup<\/sub>\uff1b<\/li>\n\n\n\n<li><strong>\u4f18\u5316<\/strong>\uff1aAdamW\uff08\u03b21=0.9, \u03b22=0.95\uff0cwarmup 2K\uff09\u3001\u7a33\u5b9a softmax\uff08stable-max\uff09\u3001\u9002\u5ea6\u6743\u8870\u51cf\uff1b<strong>EMA=0.999<\/strong>\uff1b<\/li>\n\n\n\n<li><strong>\u6570\u636e\u589e\u5f3a<\/strong>\uff1aSudoku\uff08\u89c4\u5219\u4fdd\u6301\u7684\u6d17\u724c \u00d71000\uff09\u3001Maze\uff08D8 \u7fa4 8 \u79cd\u53d8\u6362\uff09\u3001ARC\uff08\u989c\u8272\u7f6e\u6362+\u5e73\u79fb+ D8 \u00d71000\uff09\uff0c\u5e76\u5bf9\u6bcf\u6b21\u589e\u5f3a\u5206\u914d\u7279\u5b9a\u5d4c\u5165\u4ee5\u51cf\u5c0f\u6570\u636e\u7a00\u758f\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">\u516b\u3001\u9002\u7528\u8fb9\u754c\u4e0e\u5f00\u653e\u95ee\u9898<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4efb\u52a1\u4f9d\u8d56<\/strong>\uff1a\u5f53\u7f51\u683c\/\u4e0a\u4e0b\u6587\u8f83\u5927\u65f6\uff0c\u81ea\u6ce8\u610f\u529b\u4f18\u4e8e\u5e8f\u5217 MLP\uff1b\u5728\u56fa\u5b9a\u5c0f\u7f51\u683c\u4efb\u52a1\u4e2d\uff0c\u5e8f\u5217 MLP \u66f4\u7ecf\u6d4e\u9ad8\u6548\u3002<\/li>\n\n\n\n<li><strong>\u89c4\u6a21\u5f8b\u7f3a\u5931<\/strong>\uff1a\u4f5c\u8005\u6307\u51fa\u4ecd\u9700\u5efa\u7acb\u201c\u9012\u5f52\u63a8\u7406\u7f51\u7edc\u7684\u89c4\u6a21\u5f8b\u201d\uff08\u5c42\u6570\u3001\u9012\u5f52\u6b65\u6570\u3001\u5bbd\u5ea6\u4e0e\u6570\u636e\u91cf\u4e4b\u95f4\u7684\u6700\u4f18\u914d\u6bd4\uff09\u3002<\/li>\n\n\n\n<li><strong>\u751f\u6210\u5f0f\u6269\u5c55<\/strong>\uff1a\u5f53\u524d\u4e3a\u5224\u522b\u5f0f\/\u76d1\u7763\u5f0f\u4e00\u6b21\u4e00\u7b54\uff1b\u73b0\u5b9e\u4e2d\u591a\u89e3\u573a\u666f\u666e\u904d\uff0c\u503c\u5f97\u628a TRM \u6269\u5c55\u4e3a<strong>\u751f\u6210\u5f0f<\/strong>\uff08e.g., \u91c7\u6837\u591a\u6761\u63a8\u7406\u8f68\u8ff9\/\u591a\u89e3\uff09\u3002<\/li>\n\n\n\n<li><strong>\u7406\u8bba\u9610\u91ca<\/strong>\uff1a\u4e3a\u4f55\u201c\u6df1\u9012\u5f52+\u5c0f\u7f51\u7edc\u201d\u5728\u5c0f\u6570\u636e\u4e0b\u663e\u8457\u80dc\u4e8e\u201c\u6df1\u5927\u7f51\u7edc\u201d\u7684\u7cfb\u7edf\u6027\u7406\u8bba\u4ecd\u6b20\u7f3a\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e5d\u3001\u4e0e LLM\uff08CoT\/TTC\uff09\u7684\u5173\u7cfb\u4e0e\u542f\u793a<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li>TRM \u5e76\u975e\u57fa\u4e8e\u5927\u8bed\u6599\u9884\u8bad\u7ec3\uff0c\u800c\u662f<strong>\u5c0f\u6837\u672c\u76f4\u63a5\u76d1\u7763<\/strong>\uff1b\u5176\u5f3a\u9879\u5728<strong>\u7ed3\u6784\u6027\u63a8\u7406<\/strong>\u4e0e<strong>\u51e0\u4f55\u53d8\u6362\u4e0d\u53d8\u6027<\/strong>\u53ef\u901a\u8fc7\u6570\u636e\u589e\u5f3a\u4e0e\u9012\u5f52\u64cd\u4f5c\u663e\u5f0f\u6ce8\u5165\u3002<\/li>\n\n\n\n<li>\u5bf9 LLM \u7684\u542f\u793a\uff1a\u4e0e\u5176\u5728\u63a8\u7406\u65f6\u201c\u6269\u5c55\u91c7\u6837\u6df1\u5ea6\u201d\uff08\u591a\u6570\u6295\u7968\uff09\uff0c\u4e0d\u5982\u5728<strong>\u6a21\u578b\u7ed3\u6784\u4e2d\u5185\u751f\u201c\u5148\u601d\u540e\u6539\u201d\u7684\u9012\u5f52\u7b97\u5b50<\/strong>\uff0c\u7528<strong>\u5c11\u91cf\u53c2\u6570<\/strong>\u83b7\u5f97<strong>\u53ef\u590d\u5236\u7684\u7a33\u5b9a\u63a8\u7406\u8fc7\u7a0b<\/strong>\u3002<\/li>\n\n\n\n<li>\u5728\u5de5\u4e1a\/\u8fb9\u7f18\u63a8\u7406\u573a\u666f\u91cc\uff0cTRM \u98ce\u683c\u7684\u201c<strong>\u5c0f\u800c\u9012\u5f52<\/strong>\u201d\u53ef\u80fd\u66f4\u5177<strong>\u6210\u672c\/\u80fd\u6548\u4f18\u52bf<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">\u5341\u3001\u603b\u4f53\u8bc4\u4ef7<\/h5>\n\n\n\n<p>TRM \u4ee5<strong>\u66f4\u5c11\u7684\u5c42\u6570\u3001\u66f4\u5c0f\u7684\u53c2\u6570\u3001\u66f4\u7b80\u6d01\u7684\u8bad\u7ec3\u673a\u5236<\/strong>\uff0c\u5728\u591a\u4e2a\u786c\u63a8\u7406\u57fa\u51c6\u4e0a\u5b9e\u73b0\u4e86<strong>\u66f4\u5f3a\u7684\u6cdb\u5316<\/strong>\uff0c\u5e76\u7528\u4e00\u5957\u76f4\u89c2\u7684\u201cy\uff08\u89e3\uff09\/z\uff08\u601d\u8003\uff09\u201d\u4e8c\u53d8\u91cf\u89e3\u91ca\uff0c\u6f84\u6e05\u4e86 HRM \u4e2d\u201c\u5c42\u7ea7\u9012\u5f52\u201d\u7684\u771f\u5b9e\u6709\u6548\u56e0\u7d20\uff1a<strong>\u6df1\u5ea6\u76d1\u7763 + \u591a\u6b21\u65e0\u68af\u5ea6\u81ea\u6821\u6b63 + \u5b8c\u6574\u9012\u5f52\u56de\u4f20<\/strong>\u3002\u8fd9\u4e3a\u201c\u7528\u5c0f\u7f51\u7edc\u505a\u5927\u63a8\u7406\u201d\u6307\u660e\u4e86\u6e05\u6670\u3001\u53ef\u5b9e\u73b0\u7684\u6280\u672f\u8def\u5f84\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587Less is More: Recursive Reasoning with Tiny Networks\u9488 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39],"class_list":["post-6248","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Less is More: Recursive Reasoning with Tiny Networks\u9488&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6248","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6248"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6248\/revisions"}],"predecessor-version":[{"id":6251,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6248\/revisions\/6251"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6248"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6248"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6248"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}