{"id":11148,"date":"2026-08-18T13:51:39","date_gmt":"2026-08-18T05:51:39","guid":{"rendered":"https:\/\/infernews.com\/blog\/modular-cognitive-architecture-emerges-in-large-language-models\/"},"modified":"2026-08-18T13:51:39","modified_gmt":"2026-08-18T05:51:39","slug":"modular-cognitive-architecture-emerges-in-large-language-models","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/modular-cognitive-architecture-emerges-in-large-language-models\/","title":{"rendered":"LLM_Modularity\uff1aMIT \u7814\u7a76\u62c6\u89e3 LLM \u6a21\u7d44\u5316\u8a8d\u77e5\u67b6\u69cb"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-b714a677282b.jpg\" alt=\"probe.py types a reasoning prompt for each cognitive domain; the model\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4eba\u8166\u5c55\u73fe\u51fa\u986f\u8457\u7684\u529f\u80fd\u7279\u5316\u7a0b\u5ea6\uff0c\u4e0d\u540c\u7684\u795e\u7d93\u7db2\u8def\u5206\u5225\u652f\u6301\u8a9e\u8a00\u3001\u5f62\u5f0f\u63a8\u7406\u3001\u5c0d\u4ed6\u4eba\u5fc3\u667a\u7684\u63a8\u7406\u4ee5\u53ca\u5c0d\u7269\u7406\u4e16\u754c\u7684\u63a8\u7406\u3002\u9019\u7a2e\u6a21\u7d44\u5316\u7d44\u7e54\u662f\u667a\u6167\u7cfb\u7d71\u5efa\u69cb\u7684\u57fa\u672c\u539f\u5247\uff0c\u9084\u662f\u751f\u7269\u5927\u8166\u7279\u6709\u7684\u9032\u5316\u5076\u7136\u73fe\u8c61\uff1f\u672c\u6587\u6aa2\u9a57\u4e86\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u2014\u2014\u53e6\u4e00\u985e\u900f\u904e\u622a\u7136\u4e0d\u540c\u7684\u6700\u4f73\u5316\u904e\u7a0b\u6240\u5efa\u69cb\u7684\u667a\u6167\u7cfb\u7d71\u2014\u2014\u662f\u5426\u4e5f\u5448\u73fe\u51fa\u985e\u4f3c\u7684\u7d44\u7e54\u7d50\u69cb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u7559\u610f\u7684\u4e0d\u662f\u5b83\u518d\u8b1b\u4e00\u6b21\u6a21\u578b\u6703\u63a8\u7406\uff0c\u800c\u662f\u5b83\u5617\u8a66\u56de\u7b54\uff1a\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u505a\u8a9e\u8a00\u3001\u6578\u7406\u3001\u7269\u7406\u540c\u793e\u4ea4\u63a8\u7406\u6642\uff0c\u5167\u90e8\u662f\u5426\u771f\u7684\u6709\u8fd1\u4f3c\u4eba\u8166\u529f\u80fd\u7db2\u7d61\u7684\u5206\u5de5\u3002Pengrui-Han\/LLM_Modularity \u5c6c\u65bc\u7814\u7a76\u5206\u6790\u5de5\u5177\u8207\u8cc7\u6599\u6d41\u7a0b\u9805\u76ee\uff0c\u6838\u5fc3\u5de5\u4f5c\u662f\u7528 attribution patching\u3001neuron overlap \u540c causal ablation\uff0c\u5b9a\u4f4d 46 \u500b\u4efb\u52d9\u5728\u56db\u500b\u8a8d\u77e5\u9818\u57df\u5404\u81ea\u4f9d\u8cf4\u7684\u795e\u7d93\u5143\u7fa4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7d50\u8ad6\u5514\u6b62\u505c\u7559\u5728\u53ef\u8996\u5316\u3002\u9805\u76ee\u6574\u7406\u51fa\u516d\u500b 24B \u81f3 123B frontier LLMs \u7684\u4e00\u81f4\u7d50\u679c\uff1a\u540c\u9818\u57df\u4efb\u52d9\u7684\u795e\u7d93\u5143\u91cd\u758a\u9ad8\u51fa 4.3 \u500d\uff0cARI = 0.78\uff1b\u628a\u67d0\u500b\u9818\u57df\u76f8\u95dc\u795e\u7d93\u5143\u505a lesioning \u6216 ablation\uff0c\u8a72\u9818\u57df\u6e96\u78ba\u7387\u4e0b\u964d\u5e45\u5ea6\u6bd4\u8de8\u9818\u57df\u9ad8 10.3 \u500d\u3002\u63db\u53e5\u8a71\u8aaa\uff0c\u6a21\u7d44\u5316\u5514\u4fc2\u55ae\u9760\u5716\u50cf\u89e3\u8b80\uff0c\u800c\u4fc2\u6709\u56e0\u679c\u5e72\u9810\u652f\u6301\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ddf\u4e00\u822c\u53ea\u770b attention pattern \u6216 embedding \u805a\u985e\u7684\u505a\u6cd5\u76f8\u6bd4\uff0c\u9019\u500b\u9805\u76ee\u66f4\u8457\u91cd\u300c\u54ea\u4e9b\u55ae\u5143\u771f\u7684\u652f\u6490\u4efb\u52d9\u5b8c\u6210\u300d\uff0c\u6240\u4ee5\u6703\u7528 full-sequence teacher-forcing metrics\u3001corrupted-activation ablation\uff0c\u540c\u6309\u6a21\u578b\u5bb6\u65cf\u8abf\u6574\u7684 nnsight \u8f09\u5165\u6d41\u7a0b\u3002\u4ee3\u50f9\u4ea6\u5f88\u76f4\u63a5\uff1a\u5b83\u4e0d\u662f\u8f15\u91cf\u7d1a demo\uff0c\u9700\u8981 Linux\u3001torch 2.10.0\u3001nnsight 0.4.1\uff0c\u540c NVIDIA GPU\uff1b24B \u81f3 123B \u6a21\u578b\u66f4\u8981\u591a GPU \u624d\u8f03\u53ef\u884c\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8986\u84cb Language\u3001Formal reasoning (Multiple-Demand)\u3001Physical reasoning\u3001Social reasoning (Theory of Mind) \u56db\u985e\u63a8\u7406<\/li>\n<li>\u63d0\u4f9b 46 \u500b\u4efb\u52d9\u914d\u7f6e\u3001clean\/corrupted data pair\u3001\u5206\u6790\u8173\u672c\u8207\u5df2\u6574\u7406 results CSV<\/li>\n<li>\u91cd\u9ede\u4e0d\u5728\u8a13\u7df4\u65b0\u6a21\u578b\uff0c\u800c\u5728\u6aa2\u67e5\u73fe\u6709 LLM \u5167\u90e8\u795e\u7d93\u5143\u662f\u5426\u51fa\u73fe\u7a69\u5b9a\u5206\u5de5<\/li>\n<li>\u5df2\u9644 figures\u3001overlap matrix \u540c ablation analysis\uff0c\u9069\u5408\u5ef6\u4f38\u505a\u5fa9\u73fe\u6216\u4e8c\u6b21\u7814\u7a76<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8f03\u9069\u5408\u505a mechanistic interpretability\u3001\u8a8d\u77e5\u79d1\u5b78\u3001AI \u5b89\u5168\uff0c\u6216\u8005\u60f3\u6bd4\u8f03 Qwen\u3001\u5176\u4ed6\u5927\u578b\u6a21\u578b\u5167\u90e8\u63a8\u7406\u7d50\u69cb\u7684\u7814\u7a76\u5718\u968a\u3002\u5b83\u66ab\u6642\u66f4\u63a5\u8fd1\u7814\u7a76\u7ba1\u7dda\uff0c\u800c\u4e0d\u662f\u5373\u88dd\u5373\u7528\u7522\u54c1\uff1b\u4e0d\u904e\u8cc7\u6599\u593e\u7d50\u69cb\u3001config\u3001scripts \u540c results \u5df2\u7d93\u8db3\u5920\u5b8c\u6574\uff0c\u5c0d\u60f3\u91cd\u8dd1\u5be6\u9a57\u3001\u6539\u4efb\u52d9\u96c6\uff0c\u6216\u8005\u628a OSWorld \u4ee5\u5916\u7684\u8907\u96dc\u63a8\u7406\u57fa\u6e96\u63a5\u5165\u5206\u6790\u6d41\u7a0b\u7684\u4eba\uff0c\u53c3\u8003\u50f9\u503c\u76f8\u7576\u9ad8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/pengrui-han.github.io\/LLM_Modularity_Page\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/Pengrui-Han\/LLM_Modularity\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u540c\u4e00\u985e\u63a8\u7406\u6703\u53cd\u8986\u52d5\u54e1\u540c\u4e00\u6279\u795e\u7d93\u5143\uff0c\u8de8\u9818\u57df\u5247\u660e\u986f\u5206\u958b\u3002\u9019\u500b\u9805\u76ee\u628a LLM \u5167\u90e8\u7684\u300c\u5206\u5de5\u300d\u505a\u6210\u53ef\u91cd\u8dd1\u7684\u5206\u6790\u6d41\u7a0b\u3002<\/p>\n","protected":false},"author":8,"featured_media":11147,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,179,131,168,170],"tags":[],"class_list":["post-11148","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-nvidia","category-embedding","category-linux","category-170"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11148","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=11148"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11148\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11147"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11148"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11148"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11148"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}