{"id":20690,"date":"2025-02-09T14:42:36","date_gmt":"2025-02-09T06:42:36","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=20690"},"modified":"2025-02-09T14:44:55","modified_gmt":"2025-02-09T06:44:55","slug":"shenrujiedu-titans","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/shenrujiedu-titans\/","title":{"rendered":"\u6df1\u5165\u89e3\u8bfb Titans\uff1a\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def"},"content":{"rendered":"<blockquote><p>\u300aTitans: Learning to Memorize at Test Time\u300b\u539f\u6587\uff1ahttps:\/\/arxiv.org\/pdf\/2501.00663v1<\/p>\n<p>Titans \u67b6\u6784 \u975e\u5b98\u65b9\u5b9e\u73b0\uff1ahttps:\/\/github.com\/lucidrains\/titans-pytorch<\/p><\/blockquote>\n<p>&nbsp;<\/p>\n<h3>\u4e00\u3001\u00a0<strong>\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a\uff1aTransformer \u7684\u5c40\u9650\u6027\u4e0e\u4eba\u7c7b\u8bb0\u5fc6\u7684\u542f\u53d1<\/strong><\/h3>\n<h4>1.\u00a0<strong>Transformer \u7684\u5c40\u9650\u6027\uff1a\u957f\u5e8f\u5217\u5904\u7406\u7684\u74f6\u9888<\/strong><\/h4>\n<p><a href=\"https:\/\/www.kdjingpai.com\/de\/transformer\/\">Transformer<\/a> \u6a21\u578b\u81ea\u63d0\u51fa\u4ee5\u6765\uff0c\u51ed\u501f\u5176\u5f3a\u5927\u7684<strong>\u81ea\u6ce8\u610f\u529b\u673a\u5236<\/strong>\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u7b49\u9886\u57df\u53d6\u5f97\u4e86\u9769\u547d\u6027\u7684\u8fdb\u5c55\u3002\u7136\u800c\uff0c\u968f\u7740\u4efb\u52a1\u590d\u6742\u5ea6\u7684\u63d0\u5347\uff0cTransformer \u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\u9010\u6e10\u66b4\u9732\u51fa\u4e00\u4e9b\u5173\u952e\u95ee\u9898\uff1a<\/p>\n<ul>\n<li><strong>\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8\uff0c\u9650\u5236\u6a21\u578b\u6269\u5c55\u6027:<\/strong>\n<ul>\n<li>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u9700\u8981\u8ba1\u7b97\u6bcf\u4e2a <a href=\"https:\/\/www.kdjingpai.com\/de\/tokenization\/\">token<\/a> \u4e0e\u5e8f\u5217\u4e2d\u6240\u6709\u5176\u4ed6 token \u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\uff0c\u5176\u65f6\u95f4\u590d\u6742\u5ea6\u548c\u7a7a\u95f4\u590d\u6742\u5ea6\u5747\u4e3a\u00a0<strong>O(N\u00b2)<\/strong>\uff0c\u5176\u4e2d\u00a0<strong>N<\/strong>\u00a0\u4e3a\u5e8f\u5217\u957f\u5ea6\u3002<\/li>\n<li>\u8fd9\u610f\u5473\u7740\u5f53\u5e8f\u5217\u957f\u5ea6\u589e\u52a0\u65f6\uff0c\u8ba1\u7b97\u91cf\u548c\u5185\u5b58\u6d88\u8017\u5448\u5e73\u65b9\u7ea7\u589e\u957f\uff0c\u4e25\u91cd\u9650\u5236\u4e86\u6a21\u578b\u5904\u7406\u957f\u5e8f\u5217\u7684\u80fd\u529b\uff0c\u4f8b\u5982\u5728\u5904\u7406\u957f\u6587\u672c\u3001\u89c6\u9891\u7406\u89e3\u6216\u957f\u671f\u65f6\u95f4\u5e8f\u5217\u9884\u6d4b\u7b49\u4efb\u52a1\u65f6\uff0cTransformer \u5f80\u5f80\u529b\u4e0d\u4ece\u5fc3\u3002<\/li>\n<\/ul>\n<p><div id=\"attachment_20691\" style=\"width: 1177px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-20691\" class=\"wp-image-20691 size-full\" title=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/19dce977de9b69e.jpg\" alt=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-1\" width=\"1167\" height=\"261\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/19dce977de9b69e.jpg 1167w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/19dce977de9b69e-768x172.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/19dce977de9b69e-18x4.jpg 18w\" sizes=\"auto, (max-width: 1167px) 100vw, 1167px\" \/><p id=\"caption-attachment-20691\" class=\"wp-caption-text\">\u56fe 1\uff1a\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u8fc7\u7a0b\u3002<\/p><\/div><\/li>\n<li><strong>\u4e0a\u4e0b\u6587\u7a97\u53e3\u6709\u9650\uff0c\u96be\u4ee5\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb:<\/strong>\n<ul>\n<li>\u4e3a\u4e86\u7f13\u89e3\u8ba1\u7b97\u8d1f\u62c5\uff0cTransformer \u901a\u5e38\u91c7\u7528\u56fa\u5b9a\u957f\u5ea6\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\uff08\u4f8b\u5982 512 \u6216 1024\uff09\uff0c\u8fd9\u610f\u5473\u7740\u6a21\u578b\u53ea\u80fd\u5173\u6ce8\u5230\u5f53\u524d\u7a97\u53e3\u5185\u7684\u4fe1\u606f\u3002<\/li>\n<li>\u7136\u800c\uff0c\u8bb8\u591a\u73b0\u5b9e\u4e16\u754c\u7684\u4efb\u52a1\u90fd\u9700\u8981\u6a21\u578b\u80fd\u591f\u6355\u6349\u5230\u66f4\u957f\u65f6\u95f4\u8303\u56f4\u5185\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u4f8b\u5982\u7406\u89e3\u957f\u7bc7\u6587\u7ae0\u6216\u5bf9\u8bdd\u4e2d\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u6574\u5408\u89c6\u9891\u4e2d\u4e0d\u540c\u65f6\u95f4\u70b9\u7684\u4fe1\u606f\uff0c\u4ee5\u53ca\u5229\u7528\u5386\u53f2\u6570\u636e\u4e2d\u7684\u957f\u671f\u8d8b\u52bf\u548c\u6a21\u5f0f\u8fdb\u884c\u9884\u6d4b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>2.\u00a0<strong>\u7ebf\u6027 Transformer \u7684\u6298\u4e2d\u65b9\u6848\uff1a\u6548\u7387\u4e0e\u6027\u80fd\u7684\u77db\u76fe<\/strong><\/h4>\n<p>\u4e3a\u4e86\u89e3\u51b3 Transformer \u7684\u8ba1\u7b97\u74f6\u9888\uff0c\u7814\u7a76\u4eba\u5458\u63d0\u51fa\u4e86<strong>\u7ebf\u6027 Transformer<\/strong>\uff0c\u5176\u4e3b\u8981\u6539\u8fdb\u5728\u4e8e\uff1a<\/p>\n<ul>\n<li><strong>\u7528\u6838\u51fd\u6570\u66ff\u4ee3 softmax:<\/strong>\u00a0\u5c06\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u7684 softmax \u8ba1\u7b97\u66ff\u6362\u4e3a\u6838\u51fd\u6570\uff0c\u4ece\u800c\u5c06\u8ba1\u7b97\u590d\u6742\u5ea6\u964d\u4f4e\u4e3a\u00a0<strong>O(N)<\/strong>\u3002<\/li>\n<li><strong>\u53ef\u5e76\u884c\u5316\u63a8\u7406:<\/strong>\u00a0\u7ebf\u6027 Transformer \u7684\u8ba1\u7b97\u8fc7\u7a0b\u53ef\u4ee5\u8868\u793a\u4e3a\u5faa\u73af\u5f62\u5f0f\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u9ad8\u6548\u7684\u63a8\u7406\u3002<\/li>\n<\/ul>\n<p>\u7136\u800c\uff0c\u7ebf\u6027 Transformer \u4e5f\u5b58\u5728\u4e00\u4e9b\u5c40\u9650\u6027\uff1a<\/p>\n<ul>\n<li><strong>\u6027\u80fd\u4e0b\u964d:<\/strong>\n<ul>\n<li>\u6838\u6280\u5de7\u4f7f\u5f97\u6a21\u578b\u9000\u5316\u4e3a\u7ebf\u6027\u5faa\u73af\u7f51\u7edc\uff0c\u6570\u636e\u88ab\u538b\u7f29\u4e3a\u77e9\u9635\u503c\u72b6\u6001\uff0c\u5bfc\u81f4\u5176\u6027\u80fd\u4e0d\u5982\u6807\u51c6 Transformer\u3002<\/li>\n<li>\u8fd9\u79cd\u538b\u7f29\u65b9\u5f0f\u96be\u4ee5\u6709\u6548\u6355\u6349\u590d\u6742\u7684\u975e\u7ebf\u6027\u4f9d\u8d56\u5173\u7cfb\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5185\u5b58\u7ba1\u7406\u95ee\u9898:<\/strong>\n<ul>\n<li>\u7ebf\u6027 Transformer \u5c06\u5386\u53f2\u6570\u636e\u538b\u7f29\u5230\u56fa\u5b9a\u5927\u5c0f\u7684\u77e9\u9635\u4e2d\uff0c\u4f46\u5f53\u5904\u7406\u975e\u5e38\u957f\u7684\u4e0a\u4e0b\u6587\u65f6\uff0c\u8fd9\u79cd\u538b\u7f29\u65b9\u5f0f\u4f1a\u5bfc\u81f4<strong>\u5185\u5b58\u6ea2\u51fa<\/strong>\uff0c\u4ece\u800c\u5f71\u54cd\u6a21\u578b\u6027\u80fd\u3002<\/li>\n<\/ul>\n<p><div id=\"attachment_20692\" style=\"width: 1051px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-20692\" class=\"wp-image-20692 size-full\" title=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-2\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e.jpg\" alt=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-2\" width=\"1041\" height=\"795\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e.jpg 1041w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e-768x587.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e-16x12.jpg 16w\" sizes=\"auto, (max-width: 1041px) 100vw, 1041px\" \/><p id=\"caption-attachment-20692\" class=\"wp-caption-text\">\u56fe 2\uff1a\u7ebf\u6027 Transformer \u7684\u5185\u5b58\u66f4\u65b0\u8fc7\u7a0b\u3002<\/p><\/div><\/li>\n<\/ul>\n<h4>3.\u00a0<strong>\u4eba\u7c7b\u8bb0\u5fc6\u7cfb\u7edf\u7684\u542f\u53d1\uff1a\u6784\u5efa\u66f4\u5f3a\u5927\u7684\u957f\u65f6\u8bb0\u5fc6\u673a\u5236<\/strong><\/h4>\n<p>\u4e3a\u4e86\u514b\u670d\u4e0a\u8ff0\u6311\u6218\uff0c\u4f5c\u8005\u4ece<strong>\u4eba\u7c7b\u8bb0\u5fc6\u7cfb\u7edf<\/strong>\u4e2d\u6c72\u53d6\u7075\u611f\uff1a<\/p>\n<ul>\n<li><strong>\u8bb0\u5fc6\u4e0e\u5b66\u4e60\u7684\u5173\u7cfb:<\/strong>\n<ul>\n<li>\u8bba\u6587\u501f\u9274\u4e86\u795e\u7ecf\u5fc3\u7406\u5b66\u6587\u732e\u4e2d\u5173\u4e8e\u8bb0\u5fc6\u548c\u5b66\u4e60\u7684\u5b9a\u4e49\uff0c\u5c06\u8bb0\u5fc6\u89c6\u4e3a\u7531\u8f93\u5165\u5f15\u8d77\u7684\u795e\u7ecf\u66f4\u65b0\uff0c\u5c06\u5b66\u4e60\u5b9a\u4e49\u4e3a\u5728\u7ed9\u5b9a\u76ee\u6807\u7684\u60c5\u51b5\u4e0b\u83b7\u53d6\u6709\u6548\u4e14\u6709\u7528\u7684\u8bb0\u5fc6\u7684\u8fc7\u7a0b\u3002<\/li>\n<li>\u8fd9\u610f\u5473\u7740\uff0c<strong>\u6709\u6548\u7684\u5b66\u4e60\u79bb\u4e0d\u5f00\u5f3a\u5927\u7684\u8bb0\u5fc6\u673a\u5236<\/strong>\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4eba\u7c7b\u8bb0\u5fc6\u7684\u591a\u5c42\u6b21\u6027:<\/strong>\n<ul>\n<li>\u4eba\u7c7b\u8bb0\u5fc6\u7cfb\u7edf\u5e76\u975e\u5355\u4e00\u7ed3\u6784\uff0c\u800c\u662f\u7531\u591a\u4e2a\u5b50\u7cfb\u7edf\u7ec4\u6210\uff0c\u4f8b\u5982\u77ed\u671f\u8bb0\u5fc6\u3001\u5de5\u4f5c\u8bb0\u5fc6\u548c\u957f\u671f\u8bb0\u5fc6\uff0c\u6bcf\u4e2a\u5b50\u7cfb\u7edf\u90fd\u6709\u4e0d\u540c\u7684\u529f\u80fd\u548c\u7ec4\u7ec7\u7ed3\u6784\uff0c\u5e76\u80fd\u591f\u72ec\u7acb\u8fd0\u4f5c\u3002<\/li>\n<li>\u8fd9\u79cd\u591a\u5c42\u6b21\u6027\u4f7f\u5f97\u4eba\u7c7b\u80fd\u591f\u9ad8\u6548\u5730\u5b58\u50a8\u3001\u68c0\u7d22\u548c\u7ba1\u7406\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u73b0\u6709\u6a21\u578b\u7684\u4e0d\u8db3:<\/strong>\n<ul>\n<li>\u73b0\u6709\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff08\u4ece Hopfield \u7f51\u7edc\u5230 LSTM \u548c Transformer\uff09\u5728\u5904\u7406\u6cdb\u5316\u3001\u957f\u5ea6\u5916\u63a8\u548c\u63a8\u7406\u65b9\u9762\u90fd\u5b58\u5728\u6311\u6218\uff0c\u800c\u8fd9\u4e9b\u80fd\u529b\u5bf9\u4e8e\u8bb8\u591a\u73b0\u5b9e\u4e16\u754c\u7684\u590d\u6742\u4efb\u52a1\u81f3\u5173\u91cd\u8981\u3002<\/li>\n<li>\u8fd9\u4e9b\u67b6\u6784\u867d\u7136\u4ece\u4eba\u8111\u4e2d\u6c72\u53d6\u7075\u611f\uff0c\u4f46\u90fd\u7f3a\u4e4f\u5bf9<strong>\u957f\u65f6\u8bb0\u5fc6<\/strong>\u7684\u6709\u6548\u5efa\u6a21\uff0c\u4ee5\u53ca\u5bf9<strong>\u8bb0\u5fc6\u7cfb\u7edf\u591a\u5c42\u6b21\u6027<\/strong>\u7684\u6a21\u62df\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>\u4e8c\u3001\u00a0<strong>\u6838\u5fc3\u521b\u65b0\uff1a\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u4e0e Titans \u67b6\u6784<\/strong><\/h3>\n<p>\u57fa\u4e8e\u4ee5\u4e0a\u601d\u8003\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u4ee5\u4e0b\u521b\u65b0\u70b9\uff1a<\/p>\n<h4>1.\u00a0<strong>\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757 (Neural Long-term Memory Module)<\/strong><\/h4>\n<p><strong>(1) \u8bbe\u8ba1\u7406\u5ff5:<\/strong><\/p>\n<ul>\n<li><strong>\u5143\u4e0a\u4e0b\u6587\u5b66\u4e60\u673a\u5236:<\/strong>\n<ul>\n<li>\u8be5\u6a21\u5757\u88ab\u8bbe\u8ba1\u4e3a\u4e00\u4e2a<strong>\u5143\u6a21\u578b<\/strong>\uff0c\u5728\u6d4b\u8bd5\u65f6\u5b66\u4e60\u5982\u4f55\u5c06\u6570\u636e\u8bb0\u5fc6\/\u5b58\u50a8\u5230\u5176\u53c2\u6570\u4e2d\u3002<\/li>\n<li>\u8fd9\u79cd<strong>\u5728\u7ebf\u5b66\u4e60<\/strong>\u7684\u65b9\u5f0f\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u6839\u636e\u5f53\u524d\u8f93\u5165\u52a8\u6001\u8c03\u6574\u8bb0\u5fc6\uff0c\u800c\u4e0d\u662f\u4f9d\u8d56\u4e8e\u9884\u8bad\u7ec3\u65f6\u56fa\u5b9a\u7684\u8bb0\u5fc6\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u57fa\u4e8e\u60ca\u8bb6\u5ea6\u7684\u8bb0\u5fc6\u66f4\u65b0:<\/strong>\n<ul>\n<li>\u4f5c\u8005\u501f\u9274\u4e86\u4eba\u7c7b\u8bb0\u5fc6\u673a\u5236\u4e2d\u201c\u4ee4\u4eba\u60ca\u8bb6\u7684\u4e8b\u4ef6\u66f4\u5bb9\u6613\u88ab\u8bb0\u4f4f\u201d\u7684\u7279\u70b9\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e<strong>\u60ca\u8bb6\u5ea6<\/strong>\u7684\u8bb0\u5fc6\u66f4\u65b0\u673a\u5236\u3002<\/li>\n<li><strong>\u60ca\u8bb6\u5ea6<\/strong>\u901a\u8fc7\u8ba1\u7b97\u795e\u7ecf\u7f51\u7edc\u76f8\u5bf9\u4e8e\u8f93\u5165\u7684<strong>\u68af\u5ea6<\/strong>\u6765\u8861\u91cf\uff0c\u68af\u5ea6\u8d8a\u5927\uff0c\u8bf4\u660e\u8f93\u5165\u6570\u636e\u4e0e\u5386\u53f2\u6570\u636e\u5dee\u5f02\u8d8a\u5927\uff0c\u66f4\u503c\u5f97\u88ab\u8bb0\u4f4f\u3002\n<ul>\n<li>\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u6709\u6548\u5730\u6355\u6349\u5230\u6570\u636e\u4e2d\u7684\u5173\u952e\u4fe1\u606f\uff0c\u5e76\u5c06\u5176\u5b58\u50a8\u5230\u957f\u65f6\u8bb0\u5fc6\u4e2d\u3002<\/li>\n<li>\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u7ebf\u6027 Transformer \u53ea\u80fd\u57fa\u4e8e\u5f53\u524d\u7684\u8f93\u5165\u6570\u636e\u8fdb\u884c\u7ebf\u6027\u53d8\u6362\uff0c\u96be\u4ee5\u6709\u6548\u6355\u6349\u957f\u8ddd\u79bb\u7684\u4f9d\u8d56\u5173\u7cfb\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><div id=\"attachment_20693\" style=\"width: 1513px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-20693\" class=\"wp-image-20693 size-full\" title=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-3\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/a9ad2c7d24993c4.jpg\" alt=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-3\" width=\"1503\" height=\"351\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/a9ad2c7d24993c4.jpg 1503w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/a9ad2c7d24993c4-768x179.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/a9ad2c7d24993c4-18x4.jpg 18w\" sizes=\"auto, (max-width: 1503px) 100vw, 1503px\" \/><p id=\"caption-attachment-20693\" class=\"wp-caption-text\">\u56fe 3\uff1a\u57fa\u4e8e\u60ca\u8bb6\u5ea6\u7684\u8bb0\u5fc6\u66f4\u65b0\u673a\u5236\u3002<\/p><\/div><\/li>\n<\/ul>\n<p><strong>(2) \u5173\u952e\u6280\u672f:<\/strong><\/p>\n<ul>\n<li><strong>\u52a8\u91cf\u673a\u5236:<\/strong>\n<ul>\n<li>\u4e3a\u4e86\u9632\u6b62\u6a21\u578b\u88ab\u5355\u4e2a\u4ee4\u4eba\u60ca\u8bb6\u7684\u4e8b\u4ef6\u8fc7\u5ea6\u5f71\u54cd\uff0c\u4f5c\u8005\u5f15\u5165\u4e86<strong>\u52a8\u91cf\u673a\u5236<\/strong>\uff0c\u5c06\u8fc7\u53bb\u65f6\u523b\u7684\u60ca\u8bb6\u5ea6\u4e5f\u8003\u8651\u5728\u5185\u3002<\/li>\n<li>\u8fd9\u610f\u5473\u7740\u6a21\u578b\u4f1a\u7efc\u5408\u8003\u8651\u5f53\u524d\u8f93\u5165\u548c\u5386\u53f2\u8f93\u5165\u7684\u60ca\u8bb6\u5ea6\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u5e73\u6ed1\u7684\u8bb0\u5fc6\u66f4\u65b0\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u8870\u51cf\u673a\u5236:<\/strong>\n<ul>\n<li>\u4e3a\u4e86\u9632\u6b62\u5185\u5b58\u6ea2\u51fa\uff0c\u4f5c\u8005\u8fd8\u5f15\u5165\u4e86<strong>\u8870\u51cf\u673a\u5236<\/strong>\uff0c\u901a\u8fc7<strong>\u6743\u91cd\u8870\u51cf<\/strong>\u7684\u65b9\u5f0f\u9010\u6e10\u9057\u5fd8\u4e0d\u91cd\u8981\u7684\u4fe1\u606f\u3002<\/li>\n<li>\u8be5\u673a\u5236\u53ef\u4ee5\u770b\u4f5c\u662f\u4e00\u79cd<strong>\u95e8\u63a7\u673a\u5236<\/strong>\uff0c\u53ef\u4ee5\u6839\u636e\u9700\u8981\u9009\u62e9\u6027\u5730\u6e05\u9664\u8bb0\u5fc6\u3002<\/li>\n<li>\u4f5c\u8005\u6307\u51fa\uff0c\u8fd9\u79cd\u8870\u51cf\u673a\u5236\u662f\u73b0\u4ee3\u5faa\u73af\u6a21\u578b\u4e2d\u9057\u5fd8\u673a\u5236\u7684\u6cdb\u5316\uff0c\u5e76\u4e14\u4e0e\u5143\u795e\u7ecf\u7f51\u7edc\u5728\u5c0f\u6279\u91cf\u68af\u5ea6\u4e0b\u964d\u3001\u52a8\u91cf\u548c\u6743\u91cd\u8870\u51cf\u4e0b\u7684\u4f18\u5316\u7b49\u4ef7\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><strong>(3) \u8bb0\u5fc6\u7ed3\u6784:<\/strong><\/p>\n<ul>\n<li>\u4e0e\u4f20\u7edf\u7ebf\u6027\u8bb0\u5fc6\u6a21\u578b\u4e0d\u540c\uff0c\u4f5c\u8005\u91c7\u7528\u4e86<strong>\u591a\u5c42\u611f\u77e5\u673a (MLP)<\/strong>\u00a0\u4f5c\u4e3a\u8bb0\u5fc6\u6a21\u5757\u3002\n<ul>\n<li>MLP \u5177\u6709\u66f4\u5f3a\u7684\u975e\u7ebf\u6027\u8868\u8fbe\u80fd\u529b\uff0c\u80fd\u591f\u66f4\u6709\u6548\u5730\u5b58\u50a8\u548c\u68c0\u7d22\u590d\u6742\u7684\u4fe1\u606f\u3002<\/li>\n<li>\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u7ebf\u6027 Transformer \u53ea\u80fd\u4f7f\u7528\u77e9\u9635\u503c\u72b6\u6001\u6765\u5b58\u50a8\u4fe1\u606f\uff0c\u96be\u4ee5\u6355\u6349\u590d\u6742\u7684\u975e\u7ebf\u6027\u5173\u7cfb\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>2.\u00a0<strong>Titans \u67b6\u6784\uff1a\u6574\u5408\u957f\u65f6\u8bb0\u5fc6\u4e0e\u77ed\u671f\u8bb0\u5fc6<\/strong><\/h4>\n<p>\u5728\u8bbe\u8ba1\u51fa\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u4e4b\u540e\uff0c\u4f5c\u8005\u8fdb\u4e00\u6b65\u601d\u8003\u5982\u4f55\u5c06\u5176\u6709\u6548\u5730\u6574\u5408\u5230\u6df1\u5ea6\u5b66\u4e60\u67b6\u6784\u4e2d\uff0c\u5e76\u63d0\u51fa\u4e86 Titans \u67b6\u6784\uff0c\u5176\u4e3b\u8981\u7279\u70b9\u5982\u4e0b\uff1a<\/p>\n<p><strong>(1) \u4e09\u4e2a\u8d85\u5934\u534f\u540c\u5de5\u4f5c:<\/strong><\/p>\n<ul>\n<li><strong>\u6838\u5fc3 (Core):<\/strong>\n<ul>\n<li>\u7531\u77ed\u671f\u8bb0\u5fc6\u7ec4\u6210\uff0c\u8d1f\u8d23\u5904\u7406\u6570\u636e\u7684\u4e3b\u8981\u6d41\u7a0b\u3002<\/li>\n<li>\u4f7f\u7528<strong>\u6709\u9650\u7a97\u53e3\u5927\u5c0f\u7684\u6ce8\u610f\u529b\u673a\u5236<\/strong>\uff0c\u4f8b\u5982\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b (SWA) \u6216\u5168\u8fde\u63a5\u6ce8\u610f\u529b (FCA)\u3002<\/li>\n<li>\u77ed\u671f\u8bb0\u5fc6\u53ef\u4ee5\u770b\u4f5c\u662f<strong>\u77ed\u65f6\u8bb0\u5fc6<\/strong>\uff0c\u7528\u4e8e\u6355\u6349\u5f53\u524d\u4e0a\u4e0b\u6587\u4e2d\u7684\u4f9d\u8d56\u5173\u7cfb\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u957f\u671f\u8bb0\u5fc6 (Long-term Memory):<\/strong>\n<ul>\n<li>\u8d1f\u8d23\u5b58\u50a8\/\u8bb0\u4f4f\u957f\u671f\u8fc7\u53bb\u7684\u4fe1\u606f\u3002<\/li>\n<li>\u91c7\u7528\u4e0a\u8ff0\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u3002<\/li>\n<li>\u957f\u671f\u8bb0\u5fc6\u53ef\u4ee5\u770b\u4f5c\u662f<strong>\u957f\u65f6\u8bb0\u5fc6<\/strong>\uff0c\u7528\u4e8e\u5b58\u50a8\u548c\u68c0\u7d22\u66f4\u957f\u65f6\u95f4\u8303\u56f4\u5185\u7684\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6301\u4e45\u8bb0\u5fc6 (Persistent Memory):<\/strong>\n<ul>\n<li>\u662f\u4e00\u7ec4\u53ef\u5b66\u4e60\u4f46\u4e0e\u6570\u636e\u65e0\u5173\u7684\u53c2\u6570\uff0c\u7528\u4e8e\u7f16\u7801\u6709\u5173\u4efb\u52a1\u7684\u5148\u9a8c\u77e5\u8bc6\u3002<\/li>\n<li>\u7c7b\u4f3c\u4e8e Transformer \u4e2d\u5168\u8fde\u63a5\u5c42\u7684\u53c2\u6570\uff0c\u4f46\u5177\u6709\u4e0d\u540c\u7684\u529f\u80fd\u3002<\/li>\n<li>\u6301\u4e45\u8bb0\u5fc6\u53ef\u4ee5\u770b\u4f5c\u662f<strong>\u5143\u8bb0\u5fc6<\/strong>\uff0c\u7528\u4e8e\u5b58\u50a8\u4efb\u52a1\u76f8\u5173\u7684\u77e5\u8bc6\uff0c\u4f8b\u5982\u8bed\u6cd5\u89c4\u5219\u3001\u5e38\u8bc6\u77e5\u8bc6\u7b49\u3002<\/li>\n<\/ul>\n<p><div id=\"attachment_20694\" style=\"width: 1315px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-20694\" class=\"wp-image-20694 size-full\" title=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-4\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/0bdc98da76e61b5.jpg\" alt=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-4\" width=\"1305\" height=\"426\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/0bdc98da76e61b5.jpg 1305w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/0bdc98da76e61b5-768x251.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/0bdc98da76e61b5-18x6.jpg 18w\" sizes=\"auto, (max-width: 1305px) 100vw, 1305px\" \/><p id=\"caption-attachment-20694\" class=\"wp-caption-text\">\u56fe 4\uff1aTitans \u67b6\u6784\u793a\u610f\u56fe (MAC \u53d8\u4f53)\u3002<\/p><\/div><\/li>\n<\/ul>\n<p><strong>(2) \u4e09\u79cd\u4e0d\u540c\u7684\u6574\u5408\u65b9\u5f0f:<\/strong><\/p>\n<ul>\n<li><strong>\u8bb0\u5fc6\u4f5c\u4e3a\u4e0a\u4e0b\u6587 (MAC):<\/strong>\n<ul>\n<li>\u5c06\u957f\u65f6\u8bb0\u5fc6\u548c\u6301\u4e45\u8bb0\u5fc6\u4e0e\u8f93\u5165\u5e8f\u5217\u8fde\u63a5\u8d77\u6765\uff0c\u4f5c\u4e3a\u5f53\u524d\u4e0a\u4e0b\u6587\u7684\u8865\u5145\u4fe1\u606f\u3002<\/li>\n<li>\u6ce8\u610f\u529b\u673a\u5236\u51b3\u5b9a\u54ea\u4e9b\u4fe1\u606f\u9700\u8981\u5b58\u50a8\u5230\u957f\u65f6\u8bb0\u5fc6\u4e2d\u3002<\/li>\n<li>\u5728\u6d4b\u8bd5\u65f6\uff0c\u6301\u4e45\u8bb0\u5fc6\u53c2\u6570\u4fdd\u6301\u56fa\u5b9a\uff0c\u6ce8\u610f\u529b\u6a21\u5757\u6743\u91cd\u8fdb\u884c\u4e0a\u4e0b\u6587\u5b66\u4e60\uff0c\u800c\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u7ee7\u7eed\u5b66\u4e60\/\u8bb0\u5fc6\u4fe1\u606f\u3002<\/li>\n<li>\u8fd9\u79cd\u8bbe\u8ba1\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u6839\u636e\u5f53\u524d\u8f93\u5165\u7075\u6d3b\u5730\u5229\u7528\u957f\u65f6\u8bb0\u5fc6\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u8bb0\u5fc6\u4f5c\u4e3a\u95e8\u63a7 (MAG):<\/strong>\n<ul>\n<li>\u4f7f\u7528\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u4f5c\u4e3a\u77ed\u671f\u8bb0\u5fc6\uff0c\u795e\u7ecf\u8bb0\u5fc6\u6a21\u5757\u4f5c\u4e3a\u957f\u671f\u8bb0\u5fc6\u3002<\/li>\n<li>\u901a\u8fc7\u95e8\u63a7\u673a\u5236\u5c06\u4e24\u8005\u7ed3\u5408\u8d77\u6765\uff0c\u4f8b\u5982\u4f7f\u7528\u53ef\u5b66\u4e60\u7684\u5411\u91cf\u503c\u6743\u91cd\u5bf9\u4e24\u8005\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u7136\u540e\u5e94\u7528\u975e\u7ebf\u6027\u6fc0\u6d3b\u51fd\u6570\u3002<\/li>\n<li>\u8fd9\u79cd\u8bbe\u8ba1\u53ef\u4ee5\u770b\u4f5c\u662f\u4e00\u79cd\u591a\u5934\u67b6\u6784\uff0c\u5176\u4e2d\u4e0d\u540c\u5934\u7684\u7ed3\u6784\u4e0d\u540c\u3002<\/li>\n<\/ul>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-20695\" title=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-5\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/d19b949031bb706.jpg\" alt=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-5\" width=\"711\" height=\"303\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/d19b949031bb706.jpg 711w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/d19b949031bb706-18x8.jpg 18w\" sizes=\"auto, (max-width: 711px) 100vw, 711px\" \/><br \/>\n<em>\u56fe 5\uff1aTitans \u67b6\u6784\u7684\u4e0d\u540c\u53d8\u4f53 (MAC \u548c MAG)\u3002<\/em><\/li>\n<li><strong>\u8bb0\u5fc6\u4f5c\u4e3a\u5c42 (MAL):<\/strong>\n<ul>\n<li>\u5c06\u795e\u7ecf\u8bb0\u5fc6\u6a21\u5757\u4f5c\u4e3a\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u7684\u5c42\u4f7f\u7528\uff0c\u5728\u6ce8\u610f\u529b\u6a21\u5757\u4e4b\u524d\u538b\u7f29\u8fc7\u53bb\u548c\u5f53\u524d\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/li>\n<li>\u8fd9\u79cd\u8bbe\u8ba1\u5728\u6587\u732e\u4e2d\u66f4\u4e3a\u5e38\u89c1\uff0c\u4f8b\u5982 H3 \u6a21\u578b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><strong>(3) \u4f18\u52bf:<\/strong><\/p>\n<ul>\n<li><strong>\u66f4\u7075\u6d3b\u7684\u8bb0\u5fc6\u7ba1\u7406:<\/strong>\n<ul>\n<li>\u901a\u8fc7\u5c06\u8bb0\u5fc6\u6a21\u5757\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u6216\u95e8\u63a7\u5206\u652f\uff0cTitans \u67b6\u6784\u80fd\u591f\u6839\u636e\u5f53\u524d\u8f93\u5165\u52a8\u6001\u5730\u5229\u7528\u957f\u65f6\u8bb0\u5fc6\u4fe1\u606f\u3002<\/li>\n<li>\u8fd9\u4e0e\u5c06\u8bb0\u5fc6\u6a21\u5757\u4f5c\u4e3a\u5c42\u4f7f\u7528\u7684\u4f20\u7edf\u65b9\u6cd5\u76f8\u6bd4\uff0c\u66f4\u5177\u7075\u6d3b\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u66f4\u5f3a\u7684\u8868\u8fbe\u80fd\u529b:<\/strong>\n<ul>\n<li>\u4e09\u4e2a\u8d85\u5934\u7684\u534f\u540c\u5de5\u4f5c\uff0c\u4f7f\u5f97 Titans \u67b6\u6784\u80fd\u591f\u66f4\u6709\u6548\u5730\u5904\u7406\u957f\u5e8f\u5217\u6570\u636e\uff0c\u5e76\u6574\u5408\u77ed\u671f\u8bb0\u5fc6\u3001\u957f\u671f\u8bb0\u5fc6\u548c\u6301\u4e45\u8bb0\u5fc6\u7684\u4f18\u52bf\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u53ef\u6269\u5c55\u6027:<\/strong>\n<ul>\n<li>\u76f8\u6bd4 Transformer\uff0cTitans \u67b6\u6784\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\u5177\u6709\u66f4\u597d\u7684\u53ef\u6269\u5c55\u6027\uff0c\u80fd\u591f\u5728\u66f4\u5927\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e0b\u4fdd\u6301\u9ad8\u6027\u80fd\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>\u4e09\u3001\u00a0<strong>\u5b9e\u9a8c\u7ed3\u679c\u4e0e\u5206\u6790\uff1a\u9a8c\u8bc1 Titans \u67b6\u6784\u7684\u6709\u6548\u6027<\/strong><\/h3>\n<p>\u4f5c\u8005\u5728\u591a\u4e2a\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u5b9e\u9a8c\uff0c\u4ee5\u8bc4\u4f30 Titans \u67b6\u6784\u53ca\u5176\u53d8\u4f53\u7684\u6027\u80fd\uff1a<\/p>\n<h4>1.\u00a0<strong>\u8bed\u8a00\u5efa\u6a21\u4e0e\u5e38\u8bc6\u63a8\u7406:<\/strong><\/h4>\n<ul>\n<li><strong>\u5b9e\u9a8c\u8bbe\u7f6e:<\/strong>\n<ul>\n<li>\u4f7f\u7528\u4e86\u4e09\u79cd\u4e0d\u540c\u89c4\u6a21\u7684 Titans \u6a21\u578b (340M, 400M, 760M \u53c2\u6570) \u4ee5\u53ca\u591a\u4e2a\u57fa\u7ebf\u6a21\u578b\uff0c\u5305\u62ec Transformer++, RetNet, GLA, Mamba, Mamba2, DeltaNet, TTT \u548c Gated DeltaNet\u3002<\/li>\n<li>\u8bad\u7ec3\u6570\u636e\u91c7\u7528 FineWeb-Edu \u6570\u636e\u96c6\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4e3b\u8981\u7ed3\u679c:<\/strong>\n<ul>\n<li>\u5728\u975e\u6df7\u5408\u6a21\u578b\u4e2d\uff0c\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u5728\u56f0\u60d1\u5ea6\u548c\u51c6\u786e\u7387\u6307\u6807\u4e0a\u5747\u53d6\u5f97\u4e86\u6700\u4f73\u6027\u80fd\u3002<\/li>\n<li>Titans \u7684\u4e09\u79cd\u53d8\u4f53 (MAC, MAG, MAL) \u5747\u4f18\u4e8e Samba (Mamba + \u6ce8\u610f\u529b) \u548c Gated DeltaNet-H2 (Gated DeltaNet + \u6ce8\u610f\u529b)\u3002<\/li>\n<li>MAC \u5728\u5904\u7406\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb\u65f6\u8868\u73b0\u66f4\u4f73\uff0c\u800c MAG \u548c MAC \u5747\u4f18\u4e8e MAL \u53d8\u4f53\u3002<\/li>\n<\/ul>\n<p><div id=\"attachment_20696\" style=\"width: 109px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-20696\" class=\"wp-image-20696 size-full\" title=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-6\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/d6c46ebab9d8f58.jpg\" alt=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-6\" width=\"99\" height=\"90\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/d6c46ebab9d8f58.jpg 99w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/d6c46ebab9d8f58-13x12.jpg 13w\" sizes=\"auto, (max-width: 99px) 100vw, 99px\" \/><p id=\"caption-attachment-20696\" class=\"wp-caption-text\">\u56fe 6\uff1aTitans \u4e0e\u57fa\u7ebf\u6a21\u578b\u5728\u8bed\u8a00\u5efa\u6a21\u548c\u5e38\u8bc6\u63a8\u7406\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u5bf9\u6bd4\u3002<\/p><\/div><\/li>\n<\/ul>\n<h4>2.\u00a0<strong>\u201c\u9488\u5728\u5e72\u8349\u5806\u4e2d\u201d\u4efb\u52a1:<\/strong><\/h4>\n<ul>\n<li><strong>\u5b9e\u9a8c\u8bbe\u7f6e:<\/strong>\n<ul>\n<li>\u4f7f\u7528 RULER \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684 Single NIAH (S-NIAH) \u4efb\u52a1\uff0c\u8bc4\u4f30\u6a21\u578b\u5728 2K, 4K, 8K \u548c 16K \u957f\u5ea6\u5e8f\u5217\u4e0a\u7684\u68c0\u7d22\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4e3b\u8981\u7ed3\u679c:<\/strong>\n<ul>\n<li>\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u5728\u6240\u6709\u4e09\u4e2a\u4efb\u52a1\u4e2d\u5747\u53d6\u5f97\u4e86\u6700\u4f73\u7ed3\u679c\u3002<\/li>\n<li>Titans \u53d8\u4f53\u4e5f\u8868\u73b0\u51fa\u8272\uff0c\u5176\u4e2d MAC \u53d8\u4f53\u8868\u73b0\u6700\u4f73\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>3.\u00a0<strong>BABILong \u57fa\u51c6\u6d4b\u8bd5:<\/strong><\/h4>\n<ul>\n<li><strong>\u5b9e\u9a8c\u8bbe\u7f6e:<\/strong>\n<ul>\n<li>\u8be5\u4efb\u52a1\u8981\u6c42\u6a21\u578b\u5728\u6781\u957f\u7684\u6587\u6863\u4e2d\u63a8\u7406\u5206\u5e03\u7684\u4e8b\u5b9e\u4fe1\u606f\u3002<\/li>\n<li>\u5206\u4e3a\u5c11\u6837\u672c\u8bbe\u7f6e\u548c\u5fae\u8c03\u8bbe\u7f6e\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4e3b\u8981\u7ed3\u679c:<\/strong>\n<ul>\n<li>\u5728\u5c11\u6837\u672c\u8bbe\u7f6e\u4e2d\uff0cTitans \u4f18\u4e8e\u6240\u6709\u57fa\u7ebf\uff0c\u5305\u62ec\u53c2\u6570\u6570\u91cf\u66f4\u5927\u7684\u6a21\u578b\uff0c\u4f8b\u5982 GPT-4 \u548c GPT4o-mini\u3002<\/li>\n<li>\u5728\u5fae\u8c03\u8bbe\u7f6e\u4e2d\uff0cTitans \u4e5f\u4f18\u4e8e\u6240\u6709\u6a21\u578b\uff0c\u5373\u4f7f\u662f\u50cf GPT-4 \u8fd9\u6837\u7684\u8d85\u5927\u6a21\u578b\u3002<\/li>\n<li>\u4e0e\u57fa\u4e8e Transformer \u7684\u8bb0\u5fc6\u6a21\u578b (RMT) \u76f8\u6bd4\uff0cTitans \u8868\u73b0\u51fa\u66f4\u597d\u7684\u6027\u80fd\uff0c\u4e3b\u8981\u5f52\u529f\u4e8e\u5176\u5f3a\u5927\u7684\u8bb0\u5fc6\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p><div id=\"attachment_20692\" style=\"width: 1051px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-20692\" class=\"wp-image-20692 size-full\" title=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-2\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e.jpg\" alt=\"\u6df1\u5165\u89e3\u8bfb\u300aTitans: Learning to Memorize at Test Time\u300b\u2014\u2014\u957f\u65f6\u8bb0\u5fc6\u4e0e\u9ad8\u6548\u5e8f\u5217\u5efa\u6a21\u7684\u878d\u5408\u4e4b\u8def-2\" width=\"1041\" height=\"795\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e.jpg 1041w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e-768x587.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/c3237112ec43a9e-16x12.jpg 16w\" sizes=\"auto, (max-width: 1041px) 100vw, 1041px\" \/><p id=\"caption-attachment-20692\" class=\"wp-caption-text\">\u56fe 7\uff1aTitans \u4e0e\u57fa\u7ebf\u6a21\u578b\u5728 BABILong \u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u6027\u80fd\u5bf9\u6bd4\u3002<\/p><\/div><\/li>\n<\/ul>\n<h4>4.\u00a0<strong>\u65f6\u95f4\u5e8f\u5217\u9884\u6d4b:<\/strong><\/h4>\n<ul>\n<li><strong>\u5b9e\u9a8c\u8bbe\u7f6e:<\/strong>\n<ul>\n<li>\u4f7f\u7528 Simba \u6846\u67b6\uff0c\u5c06 Mamba \u6a21\u5757\u66ff\u6362\u4e3a\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u3002<\/li>\n<li>\u5728 ETT, ECL, Traffic \u548c Weather \u57fa\u51c6\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bc4\u4f30\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4e3b\u8981\u7ed3\u679c:<\/strong>\n<ul>\n<li>\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u4f18\u4e8e\u6240\u6709\u57fa\u7ebf\uff0c\u5305\u62ec\u57fa\u4e8e Mamba, \u7ebf\u6027\u6a21\u578b\u548c Transformer \u7684\u67b6\u6784\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>5.\u00a0<strong>DNA \u5efa\u6a21:<\/strong><\/h4>\n<ul>\n<li><strong>\u5b9e\u9a8c\u8bbe\u7f6e:<\/strong>\n<ul>\n<li>\u5728 GenomicsBenchmarks \u4e0a\u8bc4\u4f30\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u4e3b\u8981\u7ed3\u679c:<\/strong>\n<ul>\n<li>Titans (LMM) \u5728\u4e0d\u540c\u7684\u4e0b\u6e38\u57fa\u56e0\u7ec4\u5b66\u4efb\u52a1\u4e2d\u5747\u5177\u6709\u7ade\u4e89\u529b\uff0c\u4e0e\u6700\u5148\u8fdb\u7684\u65b9\u6cd5\u4e0d\u76f8\u4e0a\u4e0b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>6.\u00a0<strong>\u6548\u7387\u5206\u6790:<\/strong><\/h4>\n<ul>\n<li><strong>\u4e3b\u8981\u7ed3\u679c:<\/strong>\n<ul>\n<li>\u4e0e\u5176\u4ed6\u5faa\u73af\u6a21\u578b\u76f8\u6bd4\uff0c\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757\u7684\u8bad\u7ec3\u901f\u5ea6\u7565\u6162\uff0c\u4e3b\u8981\u539f\u56e0\u662f\u5176\u5177\u6709\u66f4\u6df1\u7684\u8bb0\u5fc6\u548c\u66f4\u590d\u6742\u7684\u8f6c\u6362\u8fc7\u7a0b\uff0c\u4ee5\u53ca Mamba2 \u5b9e\u73b0\u4e86\u9ad8\u5ea6\u4f18\u5316\u7684\u5185\u6838\u3002<\/li>\n<li>Titans (MAL) \u6bd4\u57fa\u7ebf\u4ee5\u53ca\u8bb0\u5fc6\u6a21\u5757\u66f4\u5feb\uff0c\u4e3b\u8981\u539f\u56e0\u662f\u4f7f\u7528\u4e86 FlashAttention \u7684\u9ad8\u5ea6\u4f18\u5316\u7684\u5185\u6838\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>7.\u00a0<strong>\u6d88\u878d\u7814\u7a76:<\/strong><\/h4>\n<ul>\n<li><strong>\u4e3b\u8981\u7ed3\u679c:<\/strong>\n<ul>\n<li>\u795e\u7ecf\u8bb0\u5fc6\u8bbe\u8ba1\u7684\u6240\u6709\u7ec4\u4ef6\u5747\u5bf9\u6027\u80fd\u6709\u79ef\u6781\u8d21\u732e\uff0c\u5176\u4e2d\u6743\u91cd\u8870\u51cf, \u52a8\u91cf, \u5377\u79ef\u548c\u6301\u4e45\u8bb0\u5fc6\u7684\u8d21\u732e\u6700\u5927\u3002<\/li>\n<li>\u67b6\u6784\u8bbe\u8ba1\u5bf9\u6027\u80fd\u4e5f\u6709\u91cd\u8981\u5f71\u54cd\uff0cMAC \u548c MAG \u5728\u8bed\u8a00\u5efa\u6a21\u548c\u5e38\u8bc6\u63a8\u7406\u4efb\u52a1\u4e2d\u8868\u73b0\u63a5\u8fd1\uff0c\u800c MAC \u5728\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e2d\u8868\u73b0\u66f4\u4f18\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h3>\u56db\u3001\u00a0<strong>\u8bba\u6587\u7684\u521b\u65b0\u70b9\u4e0e\u4f18\u52bf<\/strong><\/h3>\n<ol>\n<li><strong>\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757:<\/strong>\n<ul>\n<li>\u501f\u9274\u4e86\u4eba\u7c7b\u8bb0\u5fc6\u673a\u5236\u4e2d\u7684\u5173\u952e\u8981\u7d20\uff0c\u4f8b\u5982\u60ca\u8bb6\u5ea6, \u52a8\u91cf\u548c\u9057\u5fd8\u673a\u5236\uff0c\u5b9e\u73b0\u4e86\u66f4\u6709\u6548\u7684\u8bb0\u5fc6\u66f4\u65b0\u548c\u5b58\u50a8\u3002<\/li>\n<li>\u91c7\u7528\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u4f5c\u4e3a\u8bb0\u5fc6\u6a21\u5757\uff0c\u8d4b\u4e88\u6a21\u578b\u66f4\u5f3a\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u8bbe\u8ba1\u4e86 Titans \u67b6\u6784\uff0c\u5c06\u957f\u65f6\u8bb0\u5fc6\u4e0e\u77ed\u671f\u8bb0\u5fc6\u6709\u673a\u7ed3\u5408:<\/strong>\n<ul>\n<li>\u63d0\u51fa\u4e86\u4e09\u79cd\u4e0d\u540c\u7684\u6574\u5408\u65b9\u5f0f\uff0c\u4e3a\u4e0d\u540c\u5e94\u7528\u573a\u666f\u63d0\u4f9b\u4e86\u7075\u6d3b\u7684\u9009\u62e9\u3002<\/li>\n<li>\u6838\u5fc3, \u957f\u671f\u8bb0\u5fc6\u548c\u6301\u4e45\u8bb0\u5fc6\u4e09\u4e2a\u8d85\u5934\u7684\u534f\u540c\u5de5\u4f5c\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u66f4\u6709\u6548\u5730\u5904\u7406\u957f\u5e8f\u5217\u6570\u636e\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5728\u591a\u4e2a\u4efb\u52a1\u4e0a\u5747\u8868\u73b0\u51fa\u8272:<\/strong>\n<ul>\n<li>\u65e0\u8bba\u662f\u8bed\u8a00\u5efa\u6a21, \u5e38\u8bc6\u63a8\u7406, \u8fd8\u662f\u65f6\u95f4\u5e8f\u5217\u9884\u6d4b\u548c DNA \u5efa\u6a21\uff0cTitans \u67b6\u6784\u5747\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u6027\u80fd\uff0c\u4f18\u4e8e\u73b0\u6709\u7684 Transformer \u548c\u7ebf\u6027\u5faa\u73af\u6a21\u578b\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u53ef\u6269\u5c55\u6027\u5f3a:<\/strong>\n<ul>\n<li>\u80fd\u591f\u5728\u66f4\u5927\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e0b\u4fdd\u6301\u9ad8\u6027\u80fd\uff0c\u4e3a\u5904\u7406\u8d85\u957f\u5e8f\u5217\u63d0\u4f9b\u4e86\u53ef\u80fd\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h3>\u4e94\u3001\u00a0<strong>\u672a\u6765\u5c55\u671b<\/strong><\/h3>\n<p>\u5c3d\u7ba1 Titans \u67b6\u6784\u5728\u591a\u4e2a\u65b9\u9762\u53d6\u5f97\u4e86\u4ee4\u4eba\u77a9\u76ee\u7684\u6210\u679c\uff0c\u4f46\u4ecd\u6709\u4ee5\u4e0b\u65b9\u5411\u503c\u5f97\u8fdb\u4e00\u6b65\u63a2\u7d22:<\/p>\n<ol>\n<li><strong>\u63a2\u7d22\u66f4\u590d\u6742\u7684\u8bb0\u5fc6\u6a21\u5757\u67b6\u6784:<\/strong>\n<ul>\n<li>\u4f8b\u5982\uff0c\u5f15\u5165\u5c42\u6b21\u5316\u8bb0\u5fc6\u7ed3\u6784\uff0c\u6216\u8005\u5c06\u8bb0\u5fc6\u6a21\u5757\u4e0e\u56fe\u795e\u7ecf\u7f51\u7edc\u7b49\u5176\u4ed6\u6a21\u578b\u76f8\u7ed3\u5408\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5f00\u53d1\u66f4\u9ad8\u6548\u7684\u8bb0\u5fc6\u66f4\u65b0\u548c\u5b58\u50a8\u673a\u5236:<\/strong>\n<ul>\n<li>\u4f8b\u5982\uff0c\u5229\u7528\u7a00\u758f\u5316\u6280\u672f\u6216\u91cf\u5316\u6280\u672f\u6765\u964d\u4f4e\u5185\u5b58\u6d88\u8017\u548c\u8ba1\u7b97\u6210\u672c\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5c06 Titans \u67b6\u6784\u5e94\u7528\u4e8e\u66f4\u5e7f\u6cdb\u7684\u9886\u57df:<\/strong>\n<ul>\n<li>\u4f8b\u5982\uff0c\u89c6\u9891\u7406\u89e3, \u673a\u5668\u4eba\u63a7\u5236, \u63a8\u8350\u7cfb\u7edf\u7b49\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u63a2\u7d22\u66f4\u6709\u6548\u7684\u8bad\u7ec3\u7b56\u7565:<\/strong>\n<ul>\n<li>\u4f8b\u5982\uff0c\u5f15\u5165\u66f4\u5148\u8fdb\u7684\u4f18\u5316\u7b97\u6cd5\uff0c\u6216\u8005\u5229\u7528\u5143\u5b66\u4e60\u6765\u52a0\u901f\u6a21\u578b\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u7814\u7a76 Titans \u67b6\u6784\u7684\u53ef\u89e3\u91ca\u6027:<\/strong>\n<ul>\n<li>\u6df1\u5165\u7406\u89e3 Titans \u5982\u4f55\u5b58\u50a8\u548c\u5229\u7528\u957f\u65f6\u8bb0\u5fc6\u4fe1\u606f\uff0c\u53ef\u4ee5\u4e3a\u6784\u5efa\u66f4\u5f3a\u5927\u7684 AI \u7cfb\u7edf\u63d0\u4f9b\u65b0\u7684\u601d\u8def\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h3>\u516d\u3001\u00a0<strong>\u603b\u7ed3<\/strong><\/h3>\n<p>\u8fd9\u7bc7\u8bba\u6587\u7684\u6838\u5fc3\u8d21\u732e\u5728\u4e8e\uff1a<\/p>\n<ul>\n<li><strong>\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u578b\u7684\u795e\u7ecf\u957f\u65f6\u8bb0\u5fc6\u6a21\u5757<\/strong>\uff0c\u5176\u8bbe\u8ba1\u7075\u611f\u6765\u6e90\u4e8e\u4eba\u7c7b\u8bb0\u5fc6\u7cfb\u7edf\uff0c\u5e76\u7ed3\u5408\u4e86\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u5173\u952e\u6982\u5ff5\uff0c\u4f8b\u5982\u68af\u5ea6\u4e0b\u964d, \u52a8\u91cf\u548c\u6743\u91cd\u8870\u51cf\u3002<\/li>\n<li><strong>\u6784\u5efa\u4e86 Titans \u67b6\u6784<\/strong>\uff0c\u5c06\u957f\u65f6\u8bb0\u5fc6\u4e0e\u77ed\u671f\u8bb0\u5fc6\u6709\u673a\u7ed3\u5408\uff0c\u5e76\u63a2\u7d22\u4e86\u4e09\u79cd\u4e0d\u540c\u7684\u6574\u5408\u65b9\u5f0f\uff0c\u4e3a\u4e0d\u540c\u5e94\u7528\u573a\u666f\u63d0\u4f9b\u4e86\u7075\u6d3b\u7684\u9009\u62e9\u3002<\/li>\n<li><strong>\u901a\u8fc7\u4e25\u8c28\u7684\u5b9e\u9a8c\u9a8c\u8bc1\u4e86 Titans \u7684\u4f18\u8d8a\u6027\u80fd<\/strong>\uff0c\u5728\u591a\u4e2a\u4efb\u52a1\u4e0a\u5747\u8868\u73b0\u51fa\u8272\uff0c\u7279\u522b\u662f\u5728\u5904\u7406\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u65f6\uff0c\u5c55\u73b0\u4e86\u5f3a\u5927\u7684\u6269\u5c55\u80fd\u529b\u548c\u66f4\u9ad8\u7684\u51c6\u786e\u7387\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u300aTitans: Learning to Memorize at Test Time\u300b\u539f\u6587\uff1ahttps:\/\/arxiv.org\/pdf\/2501.00663v1 Titans \u67b6\u6784 \u975e\u5b98\u65b9\u5b9e\u73b0\uff1ahttps:\/\/github.com\/luci&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-20690","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/20690","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=20690"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/20690\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=20690"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=20690"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=20690"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}