{"id":7793,"date":"2026-09-10T09:00:26","date_gmt":"2026-09-10T01:00:26","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/09\/10\/gpt-6%e5%b8%a6%e7%81%ab%e5%be%aa%e7%8e%aftransformer%ef%bc%8c%e9%98%bf%e9%87%8c%e6%97%a9%e5%b7%b2%e5%b8%83%e5%b1%80\/"},"modified":"2026-09-10T09:00:26","modified_gmt":"2026-09-10T01:00:26","slug":"gpt-6%e5%b8%a6%e7%81%ab%e5%be%aa%e7%8e%aftransformer%ef%bc%8c%e9%98%bf%e9%87%8c%e6%97%a9%e5%b7%b2%e5%b8%83%e5%b1%80","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/09\/10\/gpt-6%e5%b8%a6%e7%81%ab%e5%be%aa%e7%8e%aftransformer%ef%bc%8c%e9%98%bf%e9%87%8c%e6%97%a9%e5%b7%b2%e5%b8%83%e5%b1%80\/","title":{"rendered":"GPT-6\u5e26\u706b\u5faa\u73afTransformer\uff0c\u963f\u91cc\u65e9\u5df2\u5e03\u5c40"},"content":{"rendered":"<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/ad1d66434887f2aba2675a46bff6a214.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/ad1d66434887f2aba2675a46bff6a214.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"GPT-6\u5e26\u706b\u5faa\u73afTransformer\uff0c\u963f\u91cc\u65e9\u5df2\u5e03\u5c40\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/994829922ff43ef9517acfd9f8b4e1fa.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/994829922ff43ef9517acfd9f8b4e1fa.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"GPT-6\u5e26\u706b\u5faa\u73afTransformer\uff0c\u963f\u91cc\u65e9\u5df2\u5e03\u5c40\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/1d0c7219b01b0cb51aff305643a72460.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/1d0c7219b01b0cb51aff305643a72460.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"GPT-6\u5e26\u706b\u5faa\u73afTransformer\uff0c\u963f\u91cc\u65e9\u5df2\u5e03\u5c40\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/cda9116e615ed73c9e54a62acb89eb73.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/09\/cda9116e615ed73c9e54a62acb89eb73.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"GPT-6\u5e26\u706b\u5faa\u73afTransformer\uff0c\u963f\u91cc\u65e9\u5df2\u5e03\u5c40\"\/><\/div><\/figure>\n<p>GPT-6 Astra\u4e00\u53d1\u5e03\uff0crecurrent depth\u7a81\u7136\u4e00\u591c\u706b\u4e86\uff01<\/p>\n<p>\u6700\u65e9\u662fThe Information\u7206\u6599\uff0cAstra\u4f7f\u7528\u4e86\u8fd9\u79cd\u300c\u5faa\u73af\u6df1\u5ea6\u300d\u6280\u672f\uff1a<\/p>\n<p>\u8ba9\u540c\u4e00\u7ec4Transformer\u5c42\u53cd\u590d\u8fd0\u884c\uff0c\u5728\u4e0d\u589e\u52a0\u540c\u7b49\u89c4\u6a21\u53c2\u6570\u7684\u60c5\u51b5\u4e0b\uff0c\u628a\u8ba1\u7b97\u505a\u5f97\u66f4\u6df1\u3002<\/p>\n<p>\uff1f\uff1f\u6a21\u578b\u4e0d\u5fc5\u7ee7\u7eed\u75af\u72c2\u300c\u957f\u8089\u300d\uff0c\u4e5f\u80fd\u8ba9\u73b0\u6709\u53c2\u6570\u591a\u5e72\u51e0\u8f6e\u6d3b\uff0c\u8fd8\u6709\u8fd9\u597d\u4e8b\u3002<\/p>\n<p>\u4e0d\u51fa\u6240\u6599\uff0c\u4e89\u8bae\u5f88\u5feb\u5230\u6765\u3002\u3002\u3002<\/p>\n<p>\u7531\u4e8e\u5faa\u73af\u53d1\u751f\u5728hidden state\u91cc\uff0c\u4e2d\u95f4\u8fc7\u7a0b\u672a\u5fc5\u4f1a\u5199\u6210\u4eba\u7c7b\u53ef\u8bfb\u7684\u601d\u7ef4\u94fe\uff0c\u6a21\u578b\u4f1a\u4e0d\u4f1a\u56e0\u6b64\u66f4\u96be\u76d1\u6d4b\uff1f<\/p>\n<p>OpenAI\u4e5f\u7531\u6b64\u88abAI\u5b89\u5168\u4e13\u5bb6\u96c6\u4f53\u8ba8\u4f10\uff0c\u751a\u81f3\u4e00\u5ea6\u903c\u5f97\u9996\u5e2d\u79d1\u5b66\u5bb6Jakub Pachocki\u4eb2\u81ea\u4e0b\u573a\u56de\u5e94\u3002<\/p>\n<p>Jakub Pachocki\u8fd8\u900f\u9732\uff0c\u81ea\u5df1\u6b63\u5728\u618b\u4e00\u7bc7\u5c0f\u4f5c\u6587\u4ee5\u5b8c\u6574\u89e3\u91ca\u6574\u4ef6\u4e8b\u60c5\u3002<\/p>\n<p>\u8fb9\u7b49\u5c0f\u4f5c\u6587\uff0c\u8fb9\u8ba9\u6211\u4eec\u628a\u76ee\u5149\u62c9\u56de\u5faa\u73afTransformer\u672c\u8eab\uff1a<\/p>\n<p>\u591a\u5faa\u73af\u51e0\u5708\uff0c\u6a21\u578b\u771f\u7684\u5c31\u4f1a\u66f4\u5f3a\u5417\uff1f<\/p>\n<p>\u5b66\u754c\u5176\u5b9e\u65e9\u5df2\u8e29\u5751\u3002<\/p>\n<p>\u540c\u7b49\u7b97\u529b\u4e0b\uff0c\u7701\u4e86\u53c2\u6570\u7684\u5faa\u73af\u6a21\u578b\uff0c\u7ecf\u5e38\u6253\u4e0d\u8fc7\u666e\u901aTransformer\u3002<\/p>\n<p>\u5361\u4f4f\u5b83\u7684\uff0c\u662f\u4e00\u4e2a\u76f8\u5f53\u73b0\u5b9e\u7684\u95ee\u9898\uff1a<\/p>\n<p>\u300c\u8ba1\u7b97\u5197\u4f59\u300d\u3002<\/p>\n<p>\u6709\u610f\u601d\u7684\u662f\uff0c\u963f\u91cc\u65e9\u572811\u4e2a\u6708\u524d\u5c31\u53d1\u4e86\u76f8\u5173\u8bba\u6587\uff0c\u9776\u5b50\u6b63\u662f\u76f4\u6307\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n<p>\u4e00\u7bc7MeSH\uff0c\u89e3\u51b3\u5faa\u73af\u5185\u90e8\u600e\u4e48\u7ba1\u7406\u4fe1\u606f\uff0c\u5411\u884c\u4e1a\u8001\u5927\u96be\u300c\u5faa\u73af\u7a7a\u8f6c\u300d\u53d1\u8d77\u51b2\u51fb\u3002<\/p>\n<p>\u4e00\u7bc7SpiralFormer\uff0c\u6b7b\u78d5\u5faa\u73af\u4e4b\u95f4\u300c\u4ee5\u4ec0\u4e48\u7cbe\u5ea6\u8ba1\u7b97\u300d\u7684\u7c92\u5ea6\u95ee\u9898\uff0c\u8ba9\u6bcf\u4e00\u8f6e\u7b97\u529b\u90fd\u82b1\u5728\u5200\u5203\u4e0a\u3002<\/p>\n<p>\u5faa\u73afTransformer\u4e4b\u6240\u4ee5\u8ba9\u5927\u5bb6\u4e0a\u5934\uff0c\u662f\u56e0\u4e3a\u5b83\u7ed9\u5927\u6a21\u578b\u63d0\u4f9b\u4e86\u4e00\u79cd\u65b0\u7684\u300c\u53d8\u5f3a\u65b9\u5f0f\u300d\u3002<\/p>\n<p>\u8fc7\u53bbScaling\u9760\u5806\u53c2\u6570\uff0c24\u5c42\u5c31\u5f9724\u5957\u53c2\u6570\uff0c\u73b0\u57288\u5c42\u53c2\u6570\u8dd13\u904d\uff0c\u4e5f\u80fd\u5b8c\u621024\u5c42\u6df1\u5ea6\u7684\u8ba1\u7b97\u3002<\/p>\n<p>\u6548\u679c\u80fd\u4e0d\u80fd\u8ffd\u4e0a\uff0c\u8fd8\u9700\u8981\u5b9e\u9a8c\u8bf4\u8bdd\uff0c\u4f46\u9700\u8981\u5b58\u50a8\u7684\u53c2\u6570\u786e\u5b9e\u5c11\u4e86\u3002<\/p>\n<p>\u7528\u8ba1\u7b97\u6df1\u5ea6\u6362\u53c2\u6570\u89c4\u6a21\uff0c\u8fd9\u7b14\u8d26\u8981\u662f\u7b97\u5f97\u8fc7\u6765\uff0c\u60f3\u8c61\u7a7a\u95f4\u53ef\u5c31\u5927\u4e86\u3002<\/p>\n<p>\u5176\u5b9e\u5728Astra\u4e4b\u524d\uff0cClaude Mythos\u5c31\u66fe\u628a\u5faa\u73af\u67b6\u6784\u5e26\u706b\u8fc7\u4e00\u8f6e\u3002<\/p>\n<p>\u5f53\u65f6\u6700\u6293\u773c\u7403\u7684\uff0c\u662f\u4e00\u9879\u56fe\u641c\u7d22\u6210\u7ee9\u3002<\/p>\n<p>\u5728GraphWalks BFS\u7684256K\u81f31M\u4e0a\u4e0b\u6587\u6d4b\u8bd5\u4e2d\uff0cMythos Preview\u62ff\u5230\u4e8680.0%\uff0cGPT-5.4\u662f21.4%\uff0c\u81ea\u5bb6\u7684Opus 4.6\u5219\u662f38.7%\u3002<\/p>\n<p>Mythos Preview\u5f97\u5206\u63a5\u8fd1GPT-5.4\u7684\u56db\u500d\u3002<\/p>\n<p>\u8fd9\u4e2a\u6d4b\u8bd5\uff0c\u53ef\u4ee5\u7406\u89e3\u6210\u7ed9\u6a21\u578b\u4e00\u5f20\u590d\u6742\u7684\u5173\u7cfb\u7f51\uff0c\u8ba9\u5b83\u4ece\u8d77\u70b9\u51fa\u53d1\uff0c\u4e00\u5c42\u5c42\u627e\u51fa\u76f8\u8fde\u7684\u8282\u70b9\u3002<\/p>\n<p>\u627e\u5230\u4e00\u6279\uff0c\u8bb0\u4f4f\u7ed3\u679c\uff0c\u518d\u7ee7\u7eed\u5f80\u4e0b\u627e\uff0c\u5f88\u8003\u9a8c\u8fde\u7eed\u591a\u6b65\u5904\u7406\u4fe1\u606f\u7684\u80fd\u529b\u3002<\/p>\n<p>\u5de7\u7684\u662f\uff0c\u5faa\u73af\u67b6\u6784\u64c5\u957f\u7684\uff0c\u6070\u6070\u4e5f\u662f\u5728hidden state\u91cc\u8fde\u7eed\u5904\u7406\u591a\u8f6e\u4fe1\u606f\u3002<\/p>\n<p>\u4e8e\u662fMythos\u7684\u6210\u7ee9\u4e00\u51fa\u6765\uff0c\u5916\u754c\u5f88\u5feb\u4fbf\u628a\u4e24\u8005\u8054\u7cfb\u5230\u4e86\u4e00\u8d77\u3002<\/p>\n<p>\u81f3\u4e8e\u5b83\u5230\u5e95\u6709\u6ca1\u6709\u91c7\u7528\u5faa\u73af\u67b6\u6784\uff0c\u5b98\u65b9\u6ca1\u6709\u660e\u8bf4\uff0c\u7b54\u6848\u4e5f\u5c31\u7559\u7ed9\u4e86\u5916\u754c\u731c\u6d4b\u3002<\/p>\n<p>\u4e0d\u8fc7\u4eceMythos\u5230Astra\uff0c\u5927\u5bb6\u671f\u5f85\u7684\u5176\u5b9e\u662f\u540c\u4e00\u4ef6\u4e8b\uff1a<\/p>\n<p>\u6a21\u578b\u4e0d\u7528\u4e00\u76f4\u957f\u5927\uff0c\u4e5f\u80fd\u9760\u5185\u90e8\u591a\u7b97\u51e0\u8f6e\uff0c\u628a\u80fd\u529b\u63d0\u4e0a\u53bb\u3002<\/p>\n<p>\u8fd9\u4e5f\u8ba9\u5faa\u73af\u67b6\u6784\u6210\u4e3a\u4e0b\u4e00\u4ee3\u9ad8\u6548\u5927\u6a21\u578b\u7684\u5019\u9009\u8def\u7ebf\u4e4b\u4e00\u3002<\/p>\n<p>\u4e0d\u8fc7\u95ee\u9898\u6765\u4e86\uff0c\u591a\u7b97\u51e0\u8f6e\uff0c\u548c\u591a\u7b97\u51fa\u70b9\u4e1c\u897f\uff0c\u8fd8\u771f\u662f\u4e24\u56de\u4e8b\uff01\uff01<\/p>\n<p>\u770b\u770b\u8fd9\u5f20\u56fe\u5c31\u660e\u767d\u4e86\u3002<\/p>\n<p>\u6a2a\u8f74\u4ece\u5de6\u5230\u53f3\uff0c\u662f\u4fe1\u606f\u7ecf\u8fc7\u6a21\u578b\u5404\u4e2a\u6a21\u5757\u7684\u8fc7\u7a0b\uff0c\u4e2d\u95f4\u4e09\u4e2a\u84dd\u8272\u67f1\u5b50\u5bf9\u5e94\u4e09\u6b21\u6838\u5fc3\u5faa\u73af\uff0c\u4e5f\u5c31\u662fcore loop\u3002<\/p>\n<p>\u7eb5\u8f74\u8868\u793a\u8ba1\u7b97\u524d\u540ehidden state\u7684\u76f8\u5bf9\u53d8\u5316\u5e45\u5ea6\u3002<\/p>\n<p>\u7b2c\u4e00\u4e2acore loop\u66f4\u65b0\u5f88\u5927\uff0c\u540e\u9762\u4e24\u4e2a\u5374\u8fc5\u901f\u7f29\u5c0f\u3002<\/p>\n<p>\u6ce8\u610f\uff0c\u67f1\u5b50\u53d8\u77ee\uff0c\u5e76\u4e0d\u610f\u5473\u7740\u6a21\u578b\u5c11\u82b1\u4e86\u7b97\u529b\uff0c\u77e9\u9635\u8fd0\u7b97\u548cAttention\u7167\u6837\u8981\u8dd1\u3002<\/p>\n<p>\u8fd9\u6b63\u662f\u8ba1\u7b97\u5197\u4f59\u7684\u5c34\u5c2c\uff1a<\/p>\n<p>\u5faa\u73af\u6b21\u6570\u589e\u52a0\u4e86\uff0c\u540e\u7eed\u8ba1\u7b97\u5e26\u6765\u7684\u589e\u91cf\u5374\u8d8a\u6765\u8d8a\u5c0f\u3002<\/p>\n<p>\u6765\u81ea\u963f\u91cc\u53ca\u5408\u4f5c\u9ad8\u6821\u7684\u7814\u7a76\u56e2\u961f\uff0c\u6b63\u662f\u4ece\u8fd9\u91cc\u5165\u624b\u4e00\u8def\u6df1\u6316\u3002<\/p>\n<p>\u4ed6\u4eec\u5148\u67e5\u6e05\u5faa\u73af\u4e3a\u4ec0\u4e48\u7a7a\u8f6c\uff0c\u518d\u60f3\u529e\u6cd5\u8ba9\u540e\u9762\u51e0\u8f6e\u771f\u6b63\u5e72\u4e0a\u6d3b\u3002<\/p>\n<p>\u56e2\u961f\u9996\u5148\u7ed9\u51fa\u7684\u89e3\u6cd5\u53ebMeSH\uff0c\u73b0\u5df2\u88abICLR 2026\u63a5\u6536\u3002<\/p>\n<p>\u8bba\u6587\u9996\u7248\u57282025\u5e7410\u6708\u516c\u5f00\uff0c\u7b49\u4e8e11\u4e2a\u6708\u524d\uff0c\u4ed6\u4eec\u5c31\u5728\u7ed9\u5faa\u73afTransformer\u505a\u300c\u4f53\u68c0\u300d\u4e86\u3002<\/p>\n<p>\u8bba\u6587\u62c9\u5230\u5e95\uff0c\u5148\u770b\u6548\u679c\u3002<\/p>\n<p>\u5728Pythia-1.4B\u7ea7\u522b\u7684\u5b9e\u9a8c\u4e2d\uff0c\u5faa\u73af\u7ed3\u6784\u901a\u8fc7\u6743\u91cd\u5171\u4eab\uff0c\u51cf\u5c11\u4e86\u7ea633%\u7684\u975e\u5d4c\u5165\u53c2\u6570\u3002\u52a0\u5165MeSH\u540e\uff0c\u96f6\u6837\u672c\u4e0b\u6e38\u5e73\u5747\u51c6\u786e\u7387\u53cd\u800c\u4ece\u666e\u901aTransformer\u768449.50%\u63d0\u9ad8\u523050.56%\uff0c\u9886\u51481.06\u4e2a\u767e\u5206\u70b9\u3002<\/p>\n<p>\u5c11\u4e86\u4e09\u5206\u4e4b\u4e00\u7684\u975e\u5d4c\u5165\u53c2\u6570\uff0c\u6210\u7ee9\u53cd\u800c\u66f4\u597d\u3002<\/p>\n<p>\u800cMeSH\u65b0\u589e\u7684\u8def\u7531\u5668\u53c2\u6570\uff0c\u4ec5\u76f8\u5f53\u4e8e\u666e\u901aTransformer\u975e\u5d4c\u5165\u53c2\u6570\u7684\u7ea60.005%\u3002<\/p>\n<p>\u76f8\u8f83\u57fa\u7840\u5faa\u73af\u6a21\u578b\uff0c\u989d\u5916\u8ba1\u7b97\u5f00\u9500\u4e5f\u53ea\u6709\u7ea60.014%\uff0c\u5c5e\u4e8e\u82b1\u5c0f\u94b1\uff0c\u529e\u5927\u4e8b\u3002<\/p>\n<p>MeSH\u662f\u600e\u4e48\u505a\u5230\u7684\uff1f<\/p>\n<p>\u56e2\u961f\u76f4\u63a5\u94bb\u8fdb\u6a21\u578b\u5185\u90e8\uff0c\u68c0\u67e5\u6a21\u578b\u6bcf\u4e00\u8f6e\u5230\u5e95\u5e72\u4e86\u4ec0\u4e48\uff0c\u7ed3\u679c\u4e00\u4e0b\u5b50\u627e\u5230\u4e86\u4e09\u4e2a\u300c\u6478\u9c7c\u8bc1\u636e\u300d\u3002<\/p>\n<p>\u7b2c\u4e00\u4e2a\u8bc1\u636e\uff0c\u662f\u540e\u9762\u7684\u5faa\u73af\u8d8a\u6765\u8d8a\u4e0d\u7231\u52a8\u3002<\/p>\n<p>\u56e2\u961f\u6bd4\u8f83\u4e86hidden state\u7ecf\u8fc7\u6bcf\u8f6e\u8ba1\u7b97\u524d\u540e\u7684\u53d8\u5316\u5e45\u5ea6\uff0c\u7ed3\u679c\u53d1\u73b0\u7b2c\u4e00\u4e2acore loop\u627f\u62c5\u4e86\u7edd\u5927\u90e8\u5206\u66f4\u65b0\uff0c\u5230\u4e86\u540e\u9762\u51e0\u4e2acore loop\uff0c\u72b6\u6001\u53d8\u5316\u8fc5\u901f\u7f29\u5c0f\u3002<\/p>\n<p>\u7b2c\u4e8c\u4e2a\u8bc1\u636e\uff0c\u662f\u524d\u540e\u51e0\u8f6e\u8d8a\u6765\u8d8a\u50cf\u3002<\/p>\n<p>\u56e2\u961f\u4f7f\u7528CKA\u5206\u6790\u4e0d\u540c\u5faa\u73af\u4e4b\u95f4\u7684\u8868\u793a\u76f8\u4f3c\u5ea6\uff0c\u7136\u540e\u53d1\u73b0\u76f8\u90bb\u8f6e\u6b21\u7684hidden state\u9ad8\u5ea6\u63a5\u8fd1\u3002<\/p>\n<p>\u540c\u4e00\u7ec4\u7f51\u7edc\u8f6c\u4e86\u4e00\u5708\u53c8\u4e00\u5708\uff0c\u5410\u51fa\u6765\u7684\u4e1c\u897f\u5374\u5dee\u4e0d\u591a\uff0c\u554a\u8fd9\uff01\uff01<\/p>\n<p>\u66f4\u624e\u5fc3\u7684\u662f\uff0c\u7b2c\u4e09\u4e2a\u8bc1\u636e\u663e\u793ahidden state\u8fd8\u8d8a\u7b97\u8d8a\u300c\u5355\u4e00\u300d\u3002<\/p>\n<p>\u5947\u5f02\u503c\u8c31\u5206\u6790\u663e\u793a\uff0c\u5faa\u73af\u8fdb\u884c\u4e4b\u540e\uff0c\u6a21\u578b\u8868\u793a\u9010\u6e10\u6324\u8fdb\u4e00\u4e2a\u4f4e\u7ef4\u7a7a\u95f4\u3002<\/p>\n<p>\u539f\u672c\u53ef\u4ee5\u4ece\u591a\u4e2a\u65b9\u5411\u8868\u8fbe\u7684\u4fe1\u606f\uff0c\u88ab\u538b\u7f29\u5230\u5c11\u6570\u51e0\u4e2a\u65b9\u5411\u4e0a\uff0c\u8868\u8fbe\u8d8a\u6765\u8d8a\u5355\u4e00\u3002<\/p>\n<p>\u4e09\u4e2a\u8bc1\u636e\u6446\u5728\u4e00\u8d77\uff0c\u6a21\u578b\u7684\u6478\u9c7c\u5957\u8def\u5c31\u5f88\u6e05\u695a\u4e86\uff1a<\/p>\n<p>\u7b2c\u4e00\u4e2acore loop\u731b\u5e72\uff0c\u540e\u9762\u51e0\u4e2acore loop\u5408\u7740\u7eaf\u7eaf\u966a\u8dd1\u3002<\/p>\n<p>\u4e3a\u4ec0\u4e48\u4f1a\u8fd9\u6837\uff1f\u8fd9\u7bc7\u8bba\u6587\u7ed9\u51fa\u4e86\u4e24\u4e2a\u8bca\u65ad\u3002<\/p>\n<p>\u4e00\u4e2a\u662f\u300c\u8ba1\u7b97\u65e0\u5206\u5316\u300d\u3002\u540c\u4e00\u7ec4\u7f51\u7edc\u53cd\u590d\u4f7f\u7528\uff0c\u53c8\u7f3a\u5c11\u660e\u786e\u7684\u8f6e\u6b21\u4fe1\u606f\uff0c\u5f88\u96be\u5728\u4e0d\u540c\u9636\u6bb5\u5f62\u6210\u5206\u5de5\uff0c\u5bb9\u6613\u6cbf\u7528\u76f8\u4f3c\u7684\u8ba1\u7b97\u65b9\u5f0f\u3002<\/p>\n<p>\u53e6\u4e00\u4e2a\u662f\u300c\u4fe1\u606f\u8fc7\u8f7d\u300d\u3002\u540c\u4e00\u4efdhidden state\u65e2\u8981\u4fdd\u5b58\u539f\u59cb\u8f93\u5165\uff0c\u53c8\u8981\u88c5\u4e0b\u524d\u51e0\u8f6e\u7ed3\u679c\uff0c\u8fd8\u5f97\u627f\u62c5\u5f53\u524d\u8ba1\u7b97\u3002\u957f\u671f\u8bb0\u5fc6\u548c\u4e34\u65f6\u52a0\u5de5\u5168\u6324\u5728\u4e00\u8d77\uff0c\u6700\u540e\u8c01\u4e5f\u5e72\u4e0d\u597d\u3002<\/p>\n<p>\u8bf4\u5230\u5e95\uff0c\u5f97\u91cd\u65b0\u7814\u7a76\u5faa\u73af\u4e4b\u95f4\u600e\u4e48\u300c\u4ea4\u63a5\u5de5\u4f5c\u300d\u3002<\/p>\n<p>\u4f20\u7edf\u5faa\u73af\u6a21\u578b\u91cc\uff0c\u4e0a\u4e00\u8f6e\u7b97\u5b8c\uff0c\u76f4\u63a5\u628ahidden state\u6574\u4efd\u9012\u7ed9\u4e0b\u4e00\u8f6e\uff0c\u518d\u7531\u540c\u4e00\u7ec4\u7f51\u7edc\u63a5\u7740\u7b97\u3002<\/p>\n<p>\u867d\u7136\u540e\u6765\u5927\u5bb6\u6253\u8fc7\u8865\u4e01\uff0c\u6015\u5b83\u7b97\u7740\u7b97\u7740\u5fd8\u4e86\u539f\u9898\uff0c\u5c31\u628a\u6700\u521d\u7684\u72b6\u6001\u91cd\u65b0\u52a0\u56de\u6765\uff0c\u4f46\u4e0d\u7ba1\u7b97\u5230\u7b2c\u51e0\u8f6e\u3001\u5f53\u524d\u5904\u7406\u54ea\u4e2atoken\uff0c\u4fe1\u606f\u7684\u642d\u914d\u89c4\u5219\u90fd\u662f\u5199\u6b7b\u7684\u3002<\/p>\n<p>MeSH\u60f3\u4e86\u4e00\u62db\uff1a<\/p>\n<p>\u4e0d\u5982\u5e72\u8106\u7ed9\u6a21\u578b\u52a0\u4e00\u4e2a\u53ef\u4ee5\u52a8\u6001\u5b58\u53d6\u7684\u300c\u8d44\u6599\u67dc\u300d\uff1f<\/p>\n<p>\u8fd9\u4e2a\u8d44\u6599\u67dc\u53ebMemory Buffer\uff0c\u7531\u591a\u4e2a\u8bb0\u5fc6\u69fd\u7ec4\u6210\uff0c\u7528\u6765\u4fdd\u5b58\u539f\u59cb\u8f93\u5165\u548c\u5faa\u73af\u8fc7\u7a0b\u4e2d\u79ef\u7d2f\u7684\u4fe1\u606f\uff0c\u4e0d\u5fc5\u5168\u90e8\u6324\u5728\u5f53\u524dhidden state\u91cc\u3002<\/p>\n<p>\u63a5\u7740\uff0cMeSH\u7ed9\u8d44\u6599\u67dc\u914d\u4e86\u4e24\u4e2a\u7ba1\u7406\u5458\u3002<\/p>\n<p>\u800c\u4e14\uff0c\u8fd9\u4e9b\u7ba1\u7406\u5458\u5404\u6709\u5404\u7684\u5de5\u4f5c\u65b9\u5f0f\u3002<\/p>\n<p>\u6bcf\u8f6e\u5faa\u73af\u90fd\u6709\u81ea\u5df1\u7684\u8def\u7531\u5668\uff0c\u6bcf\u4e2atoken\u4e5f\u80fd\u5f97\u5230\u4e0d\u540c\u7684\u8bfb\u5199\u6743\u91cd\u3002<\/p>\n<p>\u65b0\u4fe1\u606f\u5404\u5b58\u591a\u5c11\uff0c\u5386\u53f2\u4fe1\u606f\u5404\u53d6\u591a\u5c11\uff0c\u90fd\u7531\u6a21\u578b\u81ea\u5df1\u5b66\u4e60\u3002<\/p>\n<p>\u770b\u5230\u8fd9\u91cc\uff0c\u5173\u6ce8\u8fc7\u524d\u6bb5\u65f6\u95f4Hyper-Connections\u3001mHC\u8ba8\u8bba\u7684\u670b\u53cb\uff0c\u53ef\u80fd\u5df2\u7ecf\u6709\u70b9\u773c\u719f\u4e86\u3002<\/p>\n<p>\u5b83\u4eec\u5728\u8ba1\u7b97\u5f62\u5f0f\u4e0a\uff0c\u786e\u5b9e\u6709\u76f8\u4f3c\u4e4b\u5904\u3002<\/p>\n<p>Hyper-Connections\u628a\u539f\u5148\u5355\u8def\u4f20\u9012\u7684\u6b8b\u5dee\u4fe1\u606f\u6269\u5c55\u6210\u591a\u8def\uff0c\u518d\u901a\u8fc7\u53ef\u5b66\u4e60\u7684\u6620\u5c04\uff0c\u7ec4\u5408\u51fa\u4ea4\u7ed9\u8ba1\u7b97\u5c42\u7684\u8f93\u5165\uff0c\u5e76\u628a\u8ba1\u7b97\u7ed3\u679c\u5206\u914d\u56de\u591a\u8def\u72b6\u6001\u3002mHC\u8fdb\u4e00\u6b65\u7ea6\u675f\u8fd9\u4e9b\u8fde\u63a5\uff0c\u6539\u5584\u6df1\u5c42\u4f20\u64ad\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n<p>MeSH\u5219\u628a\u7c7b\u4f3c\u7684\u4fe1\u606f\u8c03\u914d\u65b9\u5f0f\u7528\u5728\u5faa\u73af\u4e4b\u95f4\uff0c\u540c\u65f6\u7f13\u89e3\u4e86\u524d\u9762\u4e24\u4e2a\u95ee\u9898\uff1a<\/p>\n<p>Memory Buffer\u628a\u5386\u53f2\u4fe1\u606f\u5206\u6d41\u51fa\u53bb\uff0chidden state\u7ec8\u4e8e\u80fd\u4e13\u5fc3\u5904\u7406\u5f53\u524d\u8f6e\u6b21\u3002<\/p>\n<p>\u52a8\u6001\u8bfb\u5199\u8def\u7531\u5668\u5219\u4e3a\u6bcf\u8f6e\u642d\u914d\u4e0d\u540c\u7684\u4fe1\u606f\uff0c\u8ba9\u5171\u4eab\u7f51\u7edc\u9010\u6e10\u5f62\u6210\u5206\u5de5\uff0c\u51cf\u5c11\u91cd\u590d\u52b3\u52a8\u3002<\/p>\n<p>\u8fd9\u65f6\u5019\u518d\u7528\u540c\u6837\u7684\u4e09\u9879\u6307\u6807\u91cd\u65b0\u68c0\u67e5\uff0c\u6a21\u578b\u6478\u9c7c\u65f6\u5448\u73b0\u7684\u4e09\u4e2a\u75c7\u72b6\uff0c\u679c\u7136\u90fd\u6709\u4e86\u660e\u663e\u6539\u5584\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0c\u5faa\u73af\u4e4b\u95f4\u6709\u4e86\u5206\u5de5\uff0c\u8fd8\u6709\u4e00\u4ef6\u4e8b\u6ca1\u53d8\uff1a<\/p>\n<p>\u6bcf\u4e00\u8f6e\u4f9d\u7136\u8981\u5728\u5b8c\u6574token\u5e8f\u5217\u4e0a\u8ba1\u7b97\u3002<\/p>\n<p>\u8fd9\u5c31\u50cf\u660e\u660e\u53ea\u60f3\u5148\u770b\u6e05\u4e00\u5ea7\u57ce\u5e02\u7684\u6574\u4f53\u5e03\u5c40\uff0c\u5374\u8fd8\u662f\u5f97\u6cbf\u7740\u6bcf\u6761\u8857\u6328\u5bb6\u6328\u6237\u8d70\u4e0a\u4e00\u904d\u3002<\/p>\n<p>\u4fe1\u606f\u4f1a\u5206\u5de5\u4e86\uff0c\u8ba1\u7b97\u7684\u5c3a\u5ea6\u5374\u8fd8\u662f\u56fa\u5b9a\u7684\u3002<\/p>\n<p>\u8fd9\u5c31\u8f6e\u5230\u7b2c\u4e8c\u7bc7\u8bba\u6587SpiralFormer\u767b\u53f0\u4e86\u3002<\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u7ee7\u7eed\u5f80\u4e0b\u8ffd\u95ee\uff1a<\/p>\n<p>\u65e2\u7136\u4e0d\u540c\u5faa\u73af\u53ef\u4ee5\u5904\u7406\u4e0d\u540c\u4fe1\u606f\uff0c\u4e3a\u4ec0\u4e48\u8fd8\u8981\u8ba9\u6bcf\u4e00\u8f6e\u90fd\u76ef\u7740\u5b8c\u6574\u7684token\u5e8f\u5217\uff1f<\/p>\n<p>SpiralFormer\u7ed9\u51fa\u7684\u7b54\u6848\uff0c\u662f\u628a\u6bcf\u8f6e\u5faa\u73af\u7684\u5e8f\u5217\u957f\u5ea6seq_len\u4e5f\u53d8\u6210\u53ef\u8c03\u7684\u3002\u8fd9\u4e00\u6210\u679c\u521a\u88abEMNLP 2026\u63a5\u6536\u3002<\/p>\n<p>\u7b80\u5355\u6765\u8bf4\uff0c\u4ed6\u4eec\u53d7Coconut\u7b49\u9690\u5f0f\u601d\u8003\u5de5\u4f5c\u7684\u542f\u53d1\uff0c\u610f\u8bc6\u5230\u6a21\u578b\u4e2d\u95f4\u7684\u300c\u601d\u8003\u300d\u5e76\u4e0d\u4e00\u5b9a\u6bcf\u4e00\u6b65\u90fd\u8981\u7ef4\u6301\u5b8c\u6574\u3001\u663e\u5f0f\u7684token\u5e8f\u5217\uff0c\u4e5f\u53ef\u4ee5\u5148\u538b\u7f29\u6210\u66f4\u7d27\u51d1\u7684\u8868\u793a\uff0c\u5728\u6f5c\u7a7a\u95f4\u4e2d\u7ee7\u7eed\u8ba1\u7b97\u3002<\/p>\n<p>\u5c06\u76f8\u5173\u601d\u8def\u8fd0\u7528\u5230SpiralFormer\u91cc\uff0c\u6548\u679c\u5c45\u7136\u8fd8\u4e0d\u9519\u3002<\/p>\n<p>\u5728Pythia-1.4B\u7ea7\u522b\u7684\u5b9e\u9a8c\u4e2d\uff0cSpiralFormer-L\u548c\u666e\u901aTransformer\u7684\u53c2\u6570\u91cf\u57fa\u672c\u76f8\u540c\uff0c\u8ba1\u7b97\u91cf\u5374\u4ece14.08T FLOPs\u964d\u523013.13T\uff0c5-shot\u4e0b\u6e38\u5e73\u5747\u51c6\u786e\u7387\u5219\u4ece51.93%\u63d0\u9ad8\u523054.37%\u3002<\/p>\n<p>\u7b97\u5f97\u66f4\u5c11\uff0c\u6210\u7ee9\u53cd\u800c\u66f4\u9ad8\u3002<\/p>\n<p>\u5177\u4f53\u601d\u8def\u662f\uff0c\u522b\u8ba9\u6a21\u578b\u6bcf\u4e00\u8f6e\u90fd\u7528\u540c\u6837\u7684\u5206\u8fa8\u7387\u770b\u95ee\u9898\u3002<\/p>\n<p>\u5b83\u4f1a\u5148\u628a\u76f8\u90bbtoken\u538b\u7f29\u6210\u66f4\u77ed\u7684\u5e8f\u5217\uff0c\u518d\u6309\u7167\u4ece\u7c97\u5230\u7ec6\u7684\u987a\u5e8f\u5faa\u73af\u3002<\/p>\n<p>\u8bba\u6587\u91c7\u7528\u7684\u4e00\u79cd\u5178\u578b\u65b9\u6848\uff0c\u662f\u5148\u4ece\u539f\u5e8f\u5217\u76841\/8\u957f\u5ea6\u5f00\u59cb\uff0c\u4e4b\u540e\u4f9d\u6b21\u6269\u5c55\u52301\/4\u30011\/2\uff0c\u6700\u540e\u56de\u5230\u5b8c\u6574\u5e8f\u5217\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u8fd9\u5c31\u50cf\u770b\u5730\u56fe\u3002<\/p>\n<p>\u7b2c\u4e00\u8f6e\u5148\u770b\u57ce\u5e02\u5168\u8c8c\uff0c\u5f04\u6e05\u4e0d\u540c\u533a\u57df\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u7136\u540e\u9010\u6b65\u653e\u5927\uff0c\u770b\u5230\u8857\u533a\u548c\u9053\u8def\uff0c\u6700\u540e\u518d\u843d\u5230\u5177\u4f53\u4f4d\u7f6e\u3002<\/p>\n<p>\u56de\u5230\u8bba\u6587\u5c31\u662f\uff0cSpiralFormer\u5728\u6bcf\u8f6e\u5faa\u73af\u5f00\u59cb\u524d\uff0c\u90fd\u4f1a\u5bf9\u5e8f\u5217\u505a\u4e00\u6b21\u300c\u7f29\u653e\u300d\uff1a<\/p>\n<p>\u628a\u76f8\u90bbtoken\u5212\u6210\u4e00\u7ec4\uff0c\u901a\u8fc7\u53ef\u5b66\u4e60\u7684\u6743\u91cd\u805a\u5408\u6210\u4e00\u4e2a\u66f4\u7c97\u7684\u8868\u793a\u3002\u538b\u7f29\u5b8c\u7684\u5e8f\u5217\u66f4\u77ed\uff0cAttention\u8981\u7b97\u7684token\u5c11\u4e86\uff0c\u8ba1\u7b97\u91cf\u81ea\u7136\u5f80\u4e0b\u6389\u3002<\/p>\n<p>\u7b97\u5b8c\u4e4b\u540e\uff0c\u518d\u628a\u7ed3\u679c\u5206\u914d\u56de\u539f\u6765\u7684token\u4f4d\u7f6e\uff0c\u4ea4\u7ed9\u4e0b\u4e00\u8f6e\u7ee7\u7eed\u5904\u7406\u3002<\/p>\n<p>\u8fd9\u91cc\u9700\u8981\u6ce8\u610f\uff0c\u8003\u8651\u5230\u6a21\u578b\u751f\u6210\u6587\u672c\u65f6\u4e0d\u80fd\u5077\u770b\u540e\u9762\u7684\u5185\u5bb9\uff0cSpiralFormer\u8fd8\u5bf9\u5199\u56de\u7ed3\u679c\u505a\u4e86\u53f3\u79fb\uff0c\u907f\u514d\u538b\u7f29\u64cd\u4f5c\u6cc4\u9732\u672a\u6765\u4fe1\u606f\u3002<\/p>\n<p>\u5c31\u8fd9\u6837\uff0c\u538b\u7f29\u3001\u8ba1\u7b97\u3001\u8fd8\u539f\u2026\u2026<\/p>\n<p>\u6bcf\u5faa\u73af\u4e00\u8f6e\uff0c\u5e8f\u5217\u5206\u8fa8\u7387\u5c31\u63d0\u9ad8\u4e00\u7ea7\uff0c\u6a21\u578b\u4e5f\u4ece\u89c2\u5bdf\u6574\u4f53\u9010\u6e10\u8f6c\u5411\u5904\u7406\u7ec6\u8282\u3002<\/p>\n<p>\u4e0d\u8fc7\u8fd9\u91cc\u8fd8\u6709\u4e00\u4e2a\u7591\u95ee\uff1a<\/p>\n<p>\u8fd9\u5957\u300c\u5148\u770b\u5168\u5c40\u3001\u518d\u62a0\u7ec6\u8282\u300d\u7684\u8bf4\u6cd5\uff0c\u5230\u5e95\u53ea\u662f\u8bbe\u8ba1\u8005\u7684\u4e00\u53a2\u60c5\u613f\uff0c\u8fd8\u662f\u6a21\u578b\u771f\u7684\u5b66\u4f1a\u4e86\uff1f<\/p>\n<p>\u56e2\u961f\u63a5\u7740\u94bb\u8fdbAttention\u91cc\uff0c\u505a\u4e86\u4e24\u9879probing\u5b9e\u9a8c\u3002<\/p>\n<p>\u5148\u662f\u89c2\u5bdfAttention\u770b\u5f97\u6709\u591a\u6563\u3002<\/p>\n<p>\u7ed3\u679c\u663e\u793a\uff0c\u65e9\u671f\u7684\u4f4e\u5206\u8fa8\u7387\u5faa\u73af\u4e2d\uff0c\u6ce8\u610f\u529b\u5206\u5e03\u5f97\u66f4\u5e7f\uff0c\u4f1a\u540c\u65f6\u5173\u6ce8\u8f83\u5927\u8303\u56f4\u7684\u4fe1\u606f\uff0c\u800c\u5230\u4e86\u540e\u671f\u7684\u9ad8\u5206\u8fa8\u7387\u5faa\u73af\uff0c\u6ce8\u610f\u529b\u5f00\u59cb\u96c6\u4e2d\u5230\u5c11\u6570\u5173\u952e\u4f4d\u7f6e\u3002<\/p>\n<p>\u7136\u540e\u76f4\u63a5\u68c0\u67e5Attention\u6709\u591a\u5173\u6ce8\u9644\u8fd1\u7684token\u3002<\/p>\n<p>\u968f\u7740\u5206\u8fa8\u7387\u63d0\u9ad8\uff0c\u4ed6\u4eec\u53d1\u73b0\u843d\u5728\u5c40\u90e8\u8303\u56f4\u5185\u7684\u6ce8\u610f\u529b\u8d8a\u6765\u8d8a\u591a\u3002<\/p>\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c\u6a21\u578b\u5230\u4e86\u540e\u9762\u51e0\u8f6e\uff0c\u786e\u5b9e\u66f4\u503e\u5411\u4e8e\u5904\u7406\u5c40\u90e8\u5173\u7cfb\u3002<\/p>\n<p>\u8fd9\u4e5f\u5c31\u5f15\u51fa\u4e86SpiralFormer\u7684\u6838\u5fc3\u53d1\u73b0\uff1a<\/p>\n<p>\u5faa\u73af\u521d\u671f\u5efa\u7acbglobal pattern\uff0c\u968f\u7740\u5faa\u73af\u63a8\u8fdb\uff0c\u518d\u9010\u6e10\u8f6c\u5411local pattern\u3002<\/p>\n<p>\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u59cb\u7ec8\u5904\u7406\u5b8c\u6574token\u5e8f\u5217\u7684\u666e\u901aLoopedFormer\uff0c\u867d\u7136\u4e5f\u4f1a\u51fa\u73b0\u7c7b\u4f3c\u53d8\u5316\uff0c\u4f46\u8d8b\u52bf\u66f4\u5f31\uff0c\u4e5f\u4e0d\u591f\u7a33\u5b9a\u3002<\/p>\n<p>\u8fd9\u8bf4\u660e\uff0c\u4ece\u5168\u5c40\u5230\u5c40\u90e8\u7684\u5206\u5de5\u5e76\u4e0d\u4f1a\u968f\u7740\u5faa\u73af\u81ea\u52a8\u51fa\u73b0\uff0c\u591a\u5206\u8fa8\u7387\u8bbe\u8ba1\u624d\u662f\u5173\u952e\u9a71\u52a8\u56e0\u7d20\u3002<\/p>\n<p>\u56e2\u961f\u8fd8\u4fdd\u6301\u53c2\u6570\u91cf\u548c\u8bad\u7ec3\u8ba1\u7b97\u9884\u7b97\u4e0d\u53d8\uff0c\u53ea\u8c03\u6574\u5faa\u73af\u5c42\u6240\u5360\u7684\u6bd4\u4f8b\u3002<\/p>\n<p>\u7ed3\u679c\u5f62\u6210\u4e86\u4e00\u6761U\u5f62\u66f2\u7ebf\uff1a<\/p>\n<p>\u968f\u7740\u5faa\u73af\u6bd4\u4f8b\u63d0\u9ad8\uff0c\u6a21\u578b\u8868\u73b0\u5148\u53d8\u597d\uff0c\u572830%\u523040%\u5de6\u53f3\u8fbe\u5230\u6700\u4f73\uff0c\u7ee7\u7eed\u589e\u52a0\u5faa\u73af\uff0c\u8fc7\u5ea6\u5171\u4eab\u53c2\u6570\u53c8\u4f1a\u62d6\u7d2f\u6548\u679c\u3002<\/p>\n<p>SpiralFormer\u4e5f\u7531\u6b64\u6253\u5f00\u4e86\u53e6\u4e00\u4e2a\u53ef\u4ee5\u63a2\u7d22\u7684Scaling\u7ef4\u5ea6\uff1a<\/p>\n<p>\u9664\u4e86\u5faa\u73af\u591a\u5c11\u6b21\uff0c\u8fd8\u80fd\u8c03\u6574\u6bcf\u8f6e\u628a\u5e8f\u5217\u538b\u7f29\u5230\u4ec0\u4e48\u5c3a\u5ea6\u3002<\/p>\n<p>\u56de\u5934\u518d\u770b\uff0c\u8fd9\u4e24\u7bc7\u8bba\u6587\u5176\u5b9e\u8865\u4e0a\u4e86\u5faa\u73afTransformer\u6700\u7f3a\u7684\u4e1c\u897f\uff1a<\/p>\n<p>\u6ca1\u9519\uff0c\u6b63\u662f\u5206\u5de5\uff01\uff01<\/p>\n<p>MeSH\u89e3\u51b3\u6bcf\u8f6e\u300c\u62ff\u4ec0\u4e48\u7b97\u300d\uff0cSpiralFormer\u89e3\u51b3\u6bcf\u8f6e\u300c\u4ee5\u4ec0\u4e48\u7c92\u5ea6\u7b97\u300d\u3002<\/p>\n<p>\u4e24\u5757\u62fc\u8d77\u6765\uff0c\u4e5f\u8ba9\u5faa\u73af\u67b6\u6784\u8d70\u5411\u4e0b\u4e00\u4ee3\u9ad8\u6548LLM\uff0c\u591a\u4e86\u51e0\u5206\u5e95\u6c14\u3002<\/p>\n<p>\u8fc7\u53bb\u5927\u6a21\u578b\u8981\u60f3Scaling\uff0c\u4e3b\u8981\u9760\u53c2\u6570\u3001\u6570\u636e\u548c\u8bad\u7ec3\u7b97\u529b\u8fd9\u4e09\u5927\u652f\u67f1\u3002\u63a8\u7406\u6a21\u578b\u5174\u8d77\u540e\uff0c\u53c8\u591a\u4e86\u4e00\u79cd\u505a\u6cd5\uff0c\u5373\u8ba9\u6a21\u578b\u751f\u6210\u66f4\u957f\u7684\u601d\u7ef4\u94fe\uff0c\u7528\u66f4\u591atoken\u6362\u66f4\u597d\u7684\u7b54\u6848\u3002<\/p>\n<p>\u73b0\u5728\u5462\uff1f\u5faa\u73afTransformer\u53c8\u63d0\u4f9b\u4e86\u4e00\u79cd\u9009\u62e9\uff1a<\/p>\n<p>\u8ba1\u7b97\u53ef\u4ee5\u5728hidden state\u91cc\u7ee7\u7eed\u8fdb\u884c\uff0c\u540c\u4e00\u7ec4\u53c2\u6570\u53cd\u590d\u4f7f\u7528\uff0c\u6a21\u578b\u89c4\u6a21\u53ef\u4ee5\u4e0d\u53d8\uff0c\u5185\u90e8\u8ba1\u7b97\u5374\u80fd\u4e00\u5c42\u5c42\u52a0\u6df1\u3002<\/p>\n<p>\u867d\u7136\u8fd9\u5e76\u4e0d\u610f\u5473\u7740\u53ef\u4ee5\u767d\u6361\u7b97\u529b\uff0c\u6bd5\u7adf\u5faa\u73af\u7ed3\u6784\u7701\u4e0b\u7684\u662f\u53c2\u6570\uff0c\u4e0d\u662f\u8ba1\u7b97\u3002<\/p>\n<p>\u4f46\u53ea\u8981\u8ba9\u591a\u51fa\u6765\u7684\u6bcf\u4e00\u8f6e\u90fd\u7b97\u5f97\u503c\uff0c\u7528\u66f4\u5c11\u7684\u53c2\u6570\u6491\u8d77\u66f4\u5f3a\u7684\u6a21\u578b\uff0c\u5c31\u6709\u4e86\u53ef\u80fd\u3002<\/p>\n<p>\u8fd9\u4e5f\u662fMeSH\u548cSpiralFormer\u5b58\u5728\u7684\u610f\u4e49\uff0c\u5b83\u4eec\u6b63\u662f\u5728\u628a\u8fd9\u79cd\u53ef\u80fd\u4e00\u6b65\u6b65\u505a\u5b9e\u3002<\/p>\n<p>\u56f4\u7ed5Astra\u548cMythos\uff0c\u5916\u754c\u8fd8\u5728\u70ed\u8bae\u300c\u5230\u5e95\u6709\u6ca1\u6709\u7528\u5faa\u73af\u67b6\u6784\u300d\u300c\u8fd9\u4f1a\u4e0d\u4f1a\u6210\u4e3a\u4e0b\u4e00\u4ee3\u8def\u7ebf\u300d\u3002<\/p>\n<p>\u800c\u963f\u91cc\u53ca\u5408\u4f5c\u9ad8\u6821\u56e2\u961f\u5df2\u7ecf\u628a\u95ee\u9898\u62c6\u5230\u4e86\u5faa\u73af\u5185\u90e8\uff1a<\/p>\n<p>\u4ece\u67e5\u75c5\u56e0\uff0c\u5230\u6539\u7ed3\u6784\uff0c\u518d\u5230\u9a8c\u8bc1\u6548\u679c\uff0c\u8fd9\u6761\u8def\u7ebf\u4e0a\u7684\u5177\u4f53\u5361\u70b9\uff0c\u5c31\u8fd9\u6837\u88ab\u9010\u4e2a\u62c6\u5f00\u4e86\u3002<\/p>\n<p>MeSH\u83b7ICLR 2026\u63a5\u6536\uff0cSpiralFormer\u83b7EMNLP 2026\u63a5\u6536\uff0c\u9876\u4f1a\u7684\u8ba4\u53ef\u4e5f\u7b97\u662f\u7ed9\u8fd9\u756a\u6301\u7eed\u6df1\u6316\u589e\u52a0\u4e86\u66f4\u591a\u542b\u91d1\u91cf\u3002<\/p>\n<p>\u53ea\u80fd\u8bf4\uff0c\u4e3a\u4e86\u8d70\u5230\u4eca\u5929\u8fd9\u4e00\u6b65\uff0c\u524d\u4eba\u5df2\u7ecf\u5728\u8fd9\u6761\u8def\u4e0a\u8e29\u4e86\u4e0d\u5c11\u5751\uff0c\u4e5f\u586b\u4e86\u4e0d\u5c11\u5751\u3002<\/p>\n<p>\u65e9\u57282018\u5e74\u7684Universal Transformer\u4e2d\uff0c\u7814\u7a76\u8005\u5c31\u5f00\u59cb\u5c1d\u8bd5\u53cd\u590d\u4f7f\u7528\u5171\u4eab\u5c42\u3002<\/p>\n<p>\u8fd1\u51e0\u5e74\uff0cLooped Transformer\u3001recurrent depth\u548clatent reasoning\u7b49\u7814\u7a76\u53c8\u9646\u7eed\u51fa\u73b0\uff0c\u80cc\u540e\u90fd\u662f\u540c\u4e00\u4e2a\u5ff5\u5934\uff1a<\/p>\n<p>\u6a21\u578b\u60f3\u53d8\u5f3a\uff0c\u9664\u4e86\u7ee7\u7eed\u957f\u5927\uff0c\u80fd\u4e0d\u80fd\u4e5f\u8ba9\u73b0\u6709\u53c2\u6570\u591a\u60f3\u51e0\u8f6e\uff1f<\/p>\n<p>Astra\u7684\u51fa\u73b0\uff0c\u66f4\u662f\u628a\u8fd9\u6761\u539f\u672c\u504f\u5b66\u672f\u7684\u8def\u7ebf\uff0c\u4e00\u628a\u63a8\u5230\u4e86\u884c\u4e1a\u805a\u5149\u706f\u4e0b\u3002<\/p>\n<p>\u81f3\u4e8e\u4e0b\u4e00\u4ee3\u9ad8\u6548\u5927\u6a21\u578b\u4f1a\u4e0d\u4f1a\u4ece\u8fd9\u91cc\u957f\u51fa\u6765\uff0c\u73b0\u5728\u4e0b\u7ed3\u8bba\u786e\u5b9e\u8fd8\u6709\u70b9\u65e9\u3002<\/p>\n<p>\u90a3\u5c31\uff0c\u8ba9\u5b50\u5f39\u518d\u98de\u4e00\u4f1a\u513f\u5427\u3002<\/p>\n<p>MeSH\uff1ahttps:\/\/arxiv.org\/abs\/2510.07739SpiralFormer\uff1a https:\/\/arxiv.org\/abs\/2602.11698<\/p>\n<blockquote>\n<p><strong>\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n<p>\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n<p>\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n<p>\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>GPT-6 Astra\u4e00\u53d1\u5e03\uff0crecurrent depth\u7a81\u7136\u4e00\u591c\u706b\u4e86\uff01<\/p>\n<p>\u6700\u65e9\u662fThe Information\u7206\u6599\uff0cAstra\u4f7f\u7528\u4e86\u8fd9\u79cd\u300c\u5faa\u73af\u6df1\u5ea6\u300d\u6280\u672f\uff1a<\/p>\n<p>\u8ba9\u540c\u4e00\u7ec4Transformer\u5c42\u53cd\u590d\u8fd0\u884c\uff0c\u5728\u4e0d\u589e\u52a0\u540c\u7b49\u89c4\u6a21\u53c2\u6570\u7684\u60c5\u51b5\u4e0b\uff0c\u628a\u8ba1\u7b97\u505a\u5f97\u66f4\u6df1\u3002<\/p>\n<p>\uff1f\uff1f\u6a21\u578b\u4e0d\u5fc5\u7ee7\u7eed\u75af\u72c2\u300c\u957f\u8089\u300d\uff0c\u4e5f\u80fd\u8ba9\u73b0\u6709\u53c2\u6570\u591a\u5e72\u51e0\u8f6e\u6d3b\uff0c\u8fd8\u6709\u8fd9\u597d\u4e8b\u3002<\/p>\n<p>\u4e0d\u51fa\u6240\u6599\uff0c\u4e89\u8bae\u5f88\u5feb\u5230\u6765\u3002\u3002\u3002<\/p>\n<p>\u7531\u4e8e\u5faa\u73af\u53d1\u751f\u5728hidden state\u91cc\uff0c<\/p>\n","protected":false},"author":1,"featured_media":7789,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[25],"class_list":["post-7793","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-25"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/7793","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=7793"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/7793\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/7789"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=7793"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=7793"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=7793"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}