{"id":19051,"date":"2025-01-22T12:30:57","date_gmt":"2025-01-22T04:30:57","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=19051"},"modified":"2025-01-22T12:31:11","modified_gmt":"2025-01-22T04:31:11","slug":"buxingyudeepseek-r1","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/buxingyudeepseek-r1\/","title":{"rendered":"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a\uff0c\u63d0\u5347\u957f\u4e0a\u4e0b\u6587\u548c\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b"},"content":{"rendered":"<h3>Kimi k1.5\u6280\u672f\u62a5\u544a\u901f\u8bfb<\/h3>\n<p><strong>1. \u5f3a\u5927\u7684\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\uff1a<\/strong><\/p>\n<p><a href=\"https:\/\/www.kdjingpai.com\/ja\/kimi-ai\/\">Kimi<\/a> k1.5 \u6a21\u578b\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u548c\u6a21\u6001\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u63a8\u7406\u6027\u80fd\uff0c\u5305\u62ec\u6570\u5b66\u3001\u4ee3\u7801\u3001\u6587\u672c\u548c\u89c6\u89c9\u63a8\u7406\u7b49\u4efb\u52a1\u3002<\/p>\n<p>\u5b83\u4e0d\u4ec5\u80fd\u5904\u7406\u7eaf\u6587\u672c\uff0c\u8fd8\u80fd\u7406\u89e3\u56fe\u50cf\u548c\u6587\u672c\u7684\u7ec4\u5408\uff0c\u5b9e\u73b0\u771f\u6b63\u7684\u591a\u6a21\u6001\u63a8\u7406\u3002<\/p>\n<p>long-CoT \u7248\u672c\u548c short-CoT \u7248\u672c\u90fd\u5c55\u73b0\u4e86\u5f3a\u5927\u7684\u6027\u80fd\u3002<\/p>\n<p>\u5728AIME\u3001MATH 500\u3001Codeforces\u7b49\u6570\u636e\u96c6\u4e0a\u5747\u8fbe\u5230\u4e86\u9886\u5148\u6c34\u5e73\u3002<\/p>\n<p>\u5728\u89c6\u89c9\u57fa\u51c6\u6d4b\u8bd5\u5982MathVista\u4e0a\u4e5f\u6709\u8f83\u5f3a\u7684\u8868\u73b0\u3002<\/p>\n<p><strong>2. \u957f\u4e0a\u4e0b\u6587 RL \u8bad\u7ec3\u7684\u7a81\u7834\uff1a<\/strong><\/p>\n<p>\u901a\u8fc7\u6269\u5c55\u4e0a\u4e0b\u6587\u7a97\u53e3\u5230 128k\uff0c\u6a21\u578b\u6027\u80fd\u6301\u7eed\u63d0\u9ad8\uff0c\u9a8c\u8bc1\u4e86\u4e0a\u4e0b\u6587\u957f\u5ea6\u662f\u5f3a\u5316\u5b66\u4e60\u6269\u5c55\u7684\u5173\u952e\u7ef4\u5ea6\u3002<\/p>\n<ul>\n<li>\u91c7\u7528\u4e86\u90e8\u5206 rollout \u6280\u672f\uff0c\u663e\u8457\u63d0\u9ad8\u4e86\u957f\u4e0a\u4e0b\u6587 RL \u8bad\u7ec3\u7684\u6548\u7387\uff0c\u4f7f\u5f97\u8bad\u7ec3\u66f4\u957f\u3001\u66f4\u590d\u6742\u7684\u63a8\u7406\u8fc7\u7a0b\u6210\u4e3a\u53ef\u80fd\u3002<\/li>\n<li>\u8bc1\u660e\u4e86\u957f\u4e0a\u4e0b\u6587\u5bf9\u590d\u6742\u95ee\u9898\u7684\u89e3\u51b3\u80fd\u529b\u81f3\u5173\u91cd\u8981\u3002<\/li>\n<\/ul>\n<p><strong>3. \u6709\u6548\u7684 long2short \u65b9\u6cd5\uff1a<\/strong><\/p>\n<ul>\n<li>\u63d0\u51fa\u4e86\u6709\u6548\u7684 long2short \u65b9\u6cd5\uff0c\u5229\u7528\u957f\u4e0a\u4e0b\u6587\u6a21\u578b\u7684\u77e5\u8bc6\u6765\u6539\u8fdb\u77ed\u4e0a\u4e0b\u6587\u6a21\u578b\u7684\u6027\u80fd\uff0c\u5e76\u63d0\u9ad8 <a href=\"https:\/\/www.kdjingpai.com\/ja\/tokenization\/\">Token<\/a> \u6548\u7387\u3002<\/li>\n<li>\u901a\u8fc7\u6a21\u578b\u5408\u5e76\u3001\u6700\u77ed\u62d2\u7edd\u91c7\u6837\u3001DPO \u7b49\u6280\u672f\uff0c\u5b9e\u73b0\u4e86\u77ed\u6a21\u578b\u7684\u6027\u80fd\u63d0\u5347\uff0c\u5e76\u4e14\u6bd4\u76f4\u63a5\u8bad\u7ec3\u77ed\u6a21\u578b\u66f4\u9ad8\u6548\u3002<\/li>\n<\/ul>\n<p>\u8bc1\u660e\u4e86\u53ef\u4ee5\u901a\u8fc7\u8fc1\u79fb long-CoT \u6a21\u578b\u7684\u601d\u7ef4\u5148\u9a8c\u77e5\u8bc6\u6765\u63d0\u9ad8\u77ed\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n<p><strong>4. \u4f18\u5316\u7684 RL \u8bad\u7ec3\u6846\u67b6\uff1a<\/strong><\/p>\n<p>\u63d0\u51fa\u4e86\u6539\u8fdb\u7684\u7b56\u7565\u4f18\u5316\u65b9\u6cd5\uff0c\u65e0\u9700\u4f9d\u8d56\u590d\u6742\u6280\u672f\uff0c\u5373\u53ef\u5b9e\u73b0\u5f3a\u5927\u7684\u6027\u80fd\u3002<\/p>\n<p>\u63a2\u7d22\u4e86\u591a\u79cd\u91c7\u6837\u7b56\u7565\u3001\u957f\u5ea6\u60e9\u7f5a\u548c\u6570\u636e\u914d\u65b9\u4f18\u5316\uff0c\u63d0\u5347\u4e86 RL \u8bad\u7ec3\u7684\u6548\u7387\u548c\u6548\u679c\u3002<\/p>\n<p>\u901a\u8fc7\u6d88\u878d\u7814\u7a76\uff0c\u9a8c\u8bc1\u4e86\u6240\u63d0\u51fa\u7684\u65b9\u6cd5\u548c\u8bad\u7ec3\u6280\u5de7\u7684\u6709\u6548\u6027\u3002<\/p>\n<p><strong>5. \u8be6\u7ec6\u7684\u8bad\u7ec3\u8fc7\u7a0b\u548c\u7cfb\u7edf\u8bbe\u8ba1\uff1a<\/strong><\/p>\n<ul>\n<li>\u516c\u5f00\u4e86 Kimi k1.5 \u7684\u8be6\u7ec6\u8bad\u7ec3\u914d\u65b9\uff0c\u5305\u62ec\u9884\u8bad\u7ec3\u3001\u5fae\u8c03\u548c\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\u3002<\/li>\n<li>\u4ecb\u7ecd\u4e86\u6df7\u5408\u90e8\u7f72\u7b56\u7565\uff0c\u4f18\u5316\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u8d44\u6e90\u5229\u7528\u7387\u3002<\/li>\n<li>\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u6570\u636e\u7ba1\u9053\u548c\u8d28\u91cf\u63a7\u5236\u673a\u5236\uff0c\u786e\u4fdd\u8bad\u7ec3\u6570\u636e\u7684\u8d28\u91cf\u3002<\/li>\n<li>\u5f00\u53d1\u4e86\u5b89\u5168\u7684\u4ee3\u7801\u6c99\u7bb1\uff0c\u7528\u4e8e\u6267\u884c\u548c\u8bc4\u4f30\u751f\u6210\u7684\u4ee3\u7801\u3002<\/li>\n<\/ul>\n<p><strong>6. \u6a21\u578b\u89c4\u6a21\u4e0e\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u63a2\u7d22\uff1a<\/strong><br \/>\n\u7814\u7a76\u4e86\u6a21\u578b\u89c4\u6a21\u548c\u4e0a\u4e0b\u6587\u957f\u5ea6\u5bf9\u6027\u80fd\u7684\u5f71\u54cd\uff0c\u53d1\u73b0\u89c4\u6a21\u8f83\u5927\u7684\u6a21\u578b\u6027\u80fd\u66f4\u597d\uff0c\u4f46\u4f7f\u7528\u66f4\u957f\u4e0a\u4e0b\u6587\u7684\u5c0f\u578b\u6a21\u578b\u4e5f\u80fd\u63a5\u8fd1\u751a\u81f3\u5ab2\u7f8e\u5927\u6a21\u578b\u7684\u6027\u80fd\uff0c\u8bc1\u660e\u4e86 RL \u8bad\u7ec3\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u589e\u5f3a\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>\u6458\u8981<\/h3>\n<p>\u4f7f\u7528\u4e0b\u4e00\u4e2a Token \u9884\u6d4b\u8fdb\u884c\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u5df2\u88ab\u8bc1\u660e\u53ef\u4ee5\u6709\u6548\u5730\u6269\u5c55\u8ba1\u7b97\u91cf\uff0c\u4f46\u53d7\u9650\u4e8e\u53ef\u7528\u7684\u8bad\u7ec3\u6570\u636e\u91cf\u3002\u6269\u5c55\u5f3a\u5316\u5b66\u4e60 (RL) \u4e3a\u4eba\u5de5\u667a\u80fd\u7684\u6301\u7eed\u6539\u8fdb\u5f00\u8f9f\u4e86\u4e00\u4e2a\u65b0\u7684\u7ef4\u5ea6\uff0c\u6709\u671b\u901a\u8fc7\u5b66\u4e60\u63a2\u7d22\u5956\u52b1\u6765\u6269\u5c55\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u8bad\u7ec3\u6570\u636e\u3002\u7136\u800c\uff0c\u5148\u524d\u53d1\u8868\u7684\u5de5\u4f5c\u5e76\u672a\u4ea7\u751f\u5177\u6709\u7ade\u4e89\u529b\u7684\u7ed3\u679c\u3002\u9274\u4e8e\u6b64\uff0c\u6211\u4eec\u62a5\u544a\u4e86 Kimi k1.5 \u7684\u8bad\u7ec3\u5b9e\u8df5\uff0cKimi k1.5 \u662f\u6211\u4eec\u6700\u65b0\u7684\u4f7f\u7528 RL \u8bad\u7ec3\u7684\u591a\u6a21\u6001 LLM\uff0c\u5305\u62ec\u5176 RL \u8bad\u7ec3\u6280\u672f\u3001\u591a\u6a21\u6001\u6570\u636e\u914d\u65b9\u548c\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\u3002\u957f\u4e0a\u4e0b\u6587\u6269\u5c55\u548c\u6539\u8fdb\u7684\u7b56\u7565\u4f18\u5316\u65b9\u6cd5\u662f\u6211\u4eec\u65b9\u6cd5\u7684\u5173\u952e\u8981\u7d20\uff0c\u5b83\u5efa\u7acb\u4e86\u4e00\u4e2a\u7b80\u5355\u6709\u6548\u7684 RL \u6846\u67b6\uff0c\u800c\u65e0\u9700\u4f9d\u8d56\u66f4\u590d\u6742\u7684\u6280\u672f\uff0c\u4f8b\u5982\u8499\u7279\u5361\u7f57\u6811\u641c\u7d22\u3001\u4ef7\u503c\u51fd\u6570\u548c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6211\u4eec\u7684\u7cfb\u7edf\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u548c\u6a21\u6001\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u63a8\u7406\u6027\u80fd\u2014\u2014\u4f8b\u5982\uff0c\u5728 AIME \u4e0a\u8fbe\u5230 77.5\uff0c\u5728 MATH 500 \u4e0a\u8fbe\u5230 96.2\uff0c\u5728 Codeforces \u4e0a\u8fbe\u5230\u7b2c 94 \u4e2a\u767e\u5206\u4f4d\u6570\uff0c\u5728 MathVista \u4e0a\u8fbe\u5230 74.9\uff0c\u4e0e OpenAI \u7684 01 \u76f8\u5339\u914d\u3002\u6b64\u5916\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u6709\u6548\u7684 long2short \u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u4f7f\u7528\u957f CoT \u6280\u672f\u6765\u6539\u8fdb short-CoT \u6a21\u578b\uff0c\u4ea7\u751f\u6700\u5148\u8fdb\u7684 short-CoT \u63a8\u7406\u7ed3\u679c\u2014\u2014\u4f8b\u5982\uff0c\u5728 AIME \u4e0a\u8fbe\u5230 60.8\uff0c\u5728 MATH500 \u4e0a\u8fbe\u5230 94.6\uff0c\u5728 LiveCodeBench \u4e0a\u8fbe\u5230 47.3\uff0c\u5176\u6027\u80fd\u5927\u5e45\u4f18\u4e8e\u73b0\u6709\u7684 short-CoT \u6a21\u578b\uff0c\u4f8b\u5982 GPT-40 \u548c <a href=\"https:\/\/www.kdjingpai.com\/ja\/claudeanquanfubai\/\">Claude<\/a> Sonnet 3.5\uff08\u9ad8\u8fbe +550%\uff09\u3002<\/p>\n<div id=\"attachment_19052\" style=\"width: 864px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19052\" class=\"wp-image-19052 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/bdda9e3c4a897c7.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"854\" height=\"395\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/bdda9e3c4a897c7.png 854w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/bdda9e3c4a897c7-300x139.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/bdda9e3c4a897c7-768x355.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/bdda9e3c4a897c7-18x8.png 18w\" sizes=\"auto, (max-width: 854px) 100vw, 854px\" \/><p id=\"caption-attachment-19052\" class=\"wp-caption-text\">\u56fe 1\uff1aKimi k1.5 long-CoT \u7ed3\u679c\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_19053\" style=\"width: 1081px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19053\" class=\"wp-image-19053 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/3201ac5ce1235ee.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1071\" height=\"586\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/3201ac5ce1235ee.png 1071w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/3201ac5ce1235ee-300x164.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/3201ac5ce1235ee-1024x560.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/3201ac5ce1235ee-768x420.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/3201ac5ce1235ee-18x10.png 18w\" sizes=\"auto, (max-width: 1071px) 100vw, 1071px\" \/><p id=\"caption-attachment-19053\" class=\"wp-caption-text\">\u56fe 2\uff1aKimi k1.5 short-CoT \u7ed3\u679c\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<h3>1 \u7b80\u4ecb<\/h3>\n<p>\u5728\u6269\u5c55\u5b9a\u5f8b\u7684\u80cc\u666f\u4e0b\uff0c\u8bed\u8a00\u6a21\u578b\u4f7f\u7528\u4e0b\u4e00\u4e2a Token \u9884\u6d4b\u8fdb\u884c\u9884\u8bad\u7ec3\u5df2\u88ab\u7814\u7a76\uff0c\u5176\u4e2d\u6309\u6bd4\u4f8b\u7f29\u653e\u6a21\u578b\u53c2\u6570\u548c\u6570\u636e\u5927\u5c0f\u4f1a\u5bfc\u81f4\u667a\u80fd\u7684\u6301\u7eed\u6539\u8fdb\u3002(Kaplan et al. 2020; Hoffmann et al. 2022) \u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53d7\u9650\u4e8e\u53ef\u7528\u7684\u9ad8\u8d28\u91cf\u8bad\u7ec3\u6570\u636e\u91cf\u3002(Villalobos et al. 2024; Muennighoff et al. 2023) \u5728\u672c\u62a5\u544a\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86 Kimi k1.5 \u7684\u8bad\u7ec3\u914d\u65b9\uff0cKimi k1.5 \u662f\u6211\u4eec\u6700\u65b0\u7684\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60 (RL) \u8bad\u7ec3\u7684\u591a\u6a21\u6001 LLM\u3002\u76ee\u6807\u662f\u63a2\u7d22\u4e00\u79cd\u53ef\u80fd\u7684\u65b0\u7684\u6301\u7eed\u6269\u5c55\u7ef4\u5ea6\u3002\u4f7f\u7528\u5e26\u6709 LLM \u7684 RL\uff0c\u6a21\u578b\u5b66\u4e60\u63a2\u7d22\u5956\u52b1\uff0c\u56e0\u6b64\u4e0d\u9650\u4e8e\u9884\u5148\u5b58\u5728\u7684\u9759\u6001\u6570\u636e\u96c6\u3002<\/p>\n<p>\u5173\u4e8e k1.5 \u7684\u8bbe\u8ba1\u548c\u8bad\u7ec3\uff0c\u6709\u51e0\u4e2a\u5173\u952e\u8981\u7d20\u3002<\/p>\n<ul>\n<li><strong>\u957f\u4e0a\u4e0b\u6587\u6269\u5c55<\/strong>: \u6211\u4eec\u5c06 RL \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u6269\u5c55\u5230 128k\uff0c\u5e76\u89c2\u5bdf\u5230\u968f\u7740\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u589e\u52a0\uff0c\u6027\u80fd\u6301\u7eed\u63d0\u9ad8\u3002\u6211\u4eec\u65b9\u6cd5\u80cc\u540e\u7684\u4e00\u4e2a\u5173\u952e\u601d\u60f3\u662f\u4f7f\u7528\u90e8\u5206 rollout \u6765\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u2014\u2014\u5373\uff0c\u901a\u8fc7\u91cd\u7528\u5148\u524d\u8f68\u8ff9\u7684\u5927\u90e8\u5206\u6765\u91c7\u6837\u65b0\u8f68\u8ff9\uff0c\u4ece\u800c\u907f\u514d\u4ece\u5934\u5f00\u59cb\u91cd\u65b0\u751f\u6210\u65b0\u8f68\u8ff9\u7684\u6210\u672c\u3002\u6211\u4eec\u7684\u89c2\u5bdf\u7ed3\u679c\u8868\u660e\uff0c\u4e0a\u4e0b\u6587\u957f\u5ea6\u662f LLM \u4e2d RL \u6301\u7eed\u6269\u5c55\u7684\u5173\u952e\u7ef4\u5ea6\u3002<\/li>\n<li><strong>\u6539\u8fdb\u7684\u7b56\u7565\u4f18\u5316<\/strong>: \u6211\u4eec\u63a8\u5bfc\u4e86\u5177\u6709 long-CoT \u7684 RL \u516c\u5f0f\uff0c\u5e76\u91c7\u7528\u5728\u7ebf\u955c\u50cf\u4e0b\u964d\u7684\u53d8\u4f53\u6765\u5b9e\u73b0\u7a33\u5065\u7684\u7b56\u7565\u4f18\u5316\u3002\u901a\u8fc7\u6211\u4eec\u6709\u6548\u7684\u91c7\u6837\u7b56\u7565\u3001\u957f\u5ea6\u60e9\u7f5a\u548c\u6570\u636e\u914d\u65b9\u7684\u4f18\u5316\uff0c\u8fdb\u4e00\u6b65\u6539\u8fdb\u4e86\u8be5\u7b97\u6cd5\u3002<\/li>\n<li><strong>\u7b80\u5316\u7684\u6846\u67b6<\/strong>: \u957f\u4e0a\u4e0b\u6587\u6269\u5c55\u4e0e\u6539\u8fdb\u7684\u7b56\u7565\u4f18\u5316\u65b9\u6cd5\u76f8\u7ed3\u5408\uff0c\u5efa\u7acb\u4e86\u4e00\u4e2a\u4f7f\u7528 LLM \u8fdb\u884c\u5b66\u4e60\u7684\u7b80\u5316 RL \u6846\u67b6\u3002\u7531\u4e8e\u6211\u4eec\u80fd\u591f\u6269\u5c55\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u56e0\u6b64\u5b66\u4e60\u5230\u7684 CoT \u8868\u73b0\u51fa\u89c4\u5212\u3001\u53cd\u601d\u548c\u7ea0\u6b63\u7684\u7279\u6027\u3002\u589e\u52a0\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u4f1a\u589e\u52a0\u641c\u7d22\u6b65\u9aa4\u7684\u6570\u91cf\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u8868\u660e\uff0c\u65e0\u9700\u4f9d\u8d56\u66f4\u590d\u6742\u7684\u6280\u672f\uff08\u4f8b\u5982\u8499\u7279\u5361\u7f57\u6811\u641c\u7d22\u3001\u4ef7\u503c\u51fd\u6570\u548c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff09\u5373\u53ef\u5b9e\u73b0\u5f3a\u5927\u7684\u6027\u80fd\u3002<\/li>\n<li><strong>\u591a\u6a21\u6001<\/strong>: \u6211\u4eec\u7684\u6a21\u578b\u662f\u4f7f\u7528\u6587\u672c\u548c\u89c6\u89c9\u6570\u636e\u8054\u5408\u8bad\u7ec3\u7684\uff0c\u8fd9\u4f7f\u5176\u5177\u6709\u5bf9\u8fd9\u4e24\u79cd\u6a21\u6001\u8fdb\u884c\u8054\u5408\u63a8\u7406\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u6709\u6548\u7684 long2short \u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u4f7f\u7528 long-CoT \u6280\u672f\u6765\u6539\u8fdb short-CoT \u6a21\u578b\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5305\u62ec\u5e94\u7528\u5e26\u6709 long-CoT \u6fc0\u6d3b\u7684\u957f\u5ea6\u60e9\u7f5a\u548c\u6a21\u578b\u5408\u5e76\u3002<\/p>\n<p>\u6211\u4eec\u7684 long-CoT \u7248\u672c\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u548c\u6a21\u6001\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u63a8\u7406\u6027\u80fd\u2014\u2014\u4f8b\u5982\uff0c\u5728 AIME \u4e0a\u8fbe\u5230 77.5\uff0c\u5728 MATH 500 \u4e0a\u8fbe\u5230 96.2\uff0c\u5728 Codeforces \u4e0a\u8fbe\u5230\u7b2c 94 \u4e2a\u767e\u5206\u4f4d\u6570\uff0c\u5728 MathVista \u4e0a\u8fbe\u5230 74.9\uff0c\u4e0e OpenAI \u7684 01 \u76f8\u5339\u914d\u3002\u6211\u4eec\u7684\u6a21\u578b\u8fd8\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684 short-CoT \u63a8\u7406\u7ed3\u679c\u2014\u2014\u4f8b\u5982\uff0c\u5728 AIME \u4e0a\u8fbe\u5230 60.8\uff0c\u5728 MATH500 \u4e0a\u8fbe\u5230 94.6\uff0c\u5728 LiveCodeBench \u4e0a\u8fbe\u5230 47.3\uff0c\u5176\u6027\u80fd\u5927\u5e45\u4f18\u4e8e\u73b0\u6709\u7684 short-CoT \u6a21\u578b\uff0c\u4f8b\u5982 GPT-40 \u548c Claude Sonnet 3.5\uff08\u9ad8\u8fbe +550%\uff09\u3002\u7ed3\u679c\u5982\u56fe 1 \u548c\u56fe 2 \u6240\u793a\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>2 \u65b9\u6cd5\uff1a\u4f7f\u7528 LLM \u8fdb\u884c\u5f3a\u5316\u5b66\u4e60<\/h3>\n<p>Kimi k1.5 \u7684\u5f00\u53d1\u5305\u62ec\u51e0\u4e2a\u9636\u6bb5\uff1a\u9884\u8bad\u7ec3\u3001vanilla \u76d1\u7763\u5fae\u8c03 (SFT)\u3001long-CoT \u76d1\u7763\u5fae\u8c03\u548c\u5f3a\u5316\u5b66\u4e60 (RL)\u3002\u672c\u62a5\u544a\u4fa7\u91cd\u4e8e RL\uff0c\u9996\u5148\u6982\u8ff0 RL \u63d0\u793a\u96c6\u7ba1\u7406\uff08\u7b2c 2.1 \u8282\uff09\u548c long-CoT \u76d1\u7763\u5fae\u8c03\uff08\u7b2c 2.2 \u8282\uff09\uff0c\u7136\u540e\u6df1\u5165\u8ba8\u8bba\u7b2c 2.3 \u8282\u4e2d\u7684 RL \u8bad\u7ec3\u7b56\u7565\u3002\u6709\u5173\u9884\u8bad\u7ec3\u548c vanilla \u76d1\u7763\u5fae\u8c03\u7684\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\u53ef\u4ee5\u5728\u7b2c 2.5 \u8282\u4e2d\u627e\u5230\u3002<\/p>\n<h3>2.1 RL \u63d0\u793a\u96c6\u7ba1\u7406<\/h3>\n<p>\u901a\u8fc7\u6211\u4eec\u7684\u521d\u6b65\u5b9e\u9a8c\uff0c\u6211\u4eec\u53d1\u73b0 RL \u63d0\u793a\u96c6\u7684\u8d28\u91cf\u548c\u591a\u6837\u6027\u5728\u786e\u4fdd\u5f3a\u5316\u5b66\u4e60\u7684\u6709\u6548\u6027\u65b9\u9762\u8d77\u7740\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u3002\u7cbe\u5fc3\u6784\u5efa\u7684\u63d0\u793a\u96c6\u4e0d\u4ec5\u5f15\u5bfc\u6a21\u578b\u8fdb\u884c\u7a33\u5065\u7684\u63a8\u7406\uff0c\u800c\u4e14\u8fd8\u51cf\u8f7b\u4e86\u5956\u52b1\u653b\u51fb\u548c\u8fc7\u5ea6\u62df\u5408\u8868\u9762\u6a21\u5f0f\u7684\u98ce\u9669\u3002\u5177\u4f53\u800c\u8a00\uff0c\u4e09\u4e2a\u5173\u952e\u5c5e\u6027\u5b9a\u4e49\u4e86\u9ad8\u8d28\u91cf\u7684 RL \u63d0\u793a\u96c6\uff1a<\/p>\n<ul>\n<li><strong>\u591a\u6837\u5316\u7684\u8986\u76d6\u8303\u56f4<\/strong>: \u63d0\u793a\u5e94\u6db5\u76d6\u5404\u79cd\u5b66\u79d1\uff0c\u4f8b\u5982 STEM\u3001\u7f16\u7801\u548c\u4e00\u822c\u63a8\u7406\uff0c\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u9002\u5e94\u6027\u5e76\u786e\u4fdd\u5176\u5728\u4e0d\u540c\u9886\u57df\u4e2d\u7684\u5e7f\u6cdb\u9002\u7528\u6027\u3002<\/li>\n<li><strong>\u5747\u8861\u7684\u96be\u5ea6<\/strong>: \u63d0\u793a\u96c6\u5e94\u5305\u62ec\u5206\u5e03\u5747\u5300\u7684\u7b80\u5355\u3001\u4e2d\u7b49\u548c\u56f0\u96be\u95ee\u9898\uff0c\u4ee5\u4fc3\u8fdb\u6e10\u8fdb\u5f0f\u5b66\u4e60\u5e76\u9632\u6b62\u8fc7\u5ea6\u62df\u5408\u7279\u5b9a\u7684\u590d\u6742\u6027\u7ea7\u522b\u3002<\/li>\n<li><strong>\u51c6\u786e\u7684\u53ef\u8bc4\u4f30\u6027<\/strong>: \u63d0\u793a\u5e94\u5141\u8bb8\u9a8c\u8bc1\u8005\u8fdb\u884c\u5ba2\u89c2\u548c\u53ef\u9760\u7684\u8bc4\u4f30\uff0c\u4ee5\u786e\u4fdd\u6a21\u578b\u6027\u80fd\u662f\u57fa\u4e8e\u6b63\u786e\u7684\u63a8\u7406\u800c\u4e0d\u662f\u8868\u9762\u6a21\u5f0f\u6216\u968f\u673a\u731c\u6d4b\u6765\u8861\u91cf\u7684\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4e86\u5728\u63d0\u793a\u96c6\u4e2d\u5b9e\u73b0\u591a\u6837\u5316\u7684\u8986\u76d6\u8303\u56f4\uff0c\u6211\u4eec\u91c7\u7528\u81ea\u52a8\u8fc7\u6ee4\u5668\u6765\u9009\u62e9\u9700\u8981\u4e30\u5bcc\u63a8\u7406\u4e14\u6613\u4e8e\u8bc4\u4f30\u7684\u95ee\u9898\u3002\u6211\u4eec\u7684\u6570\u636e\u96c6\u5305\u62ec\u6765\u81ea\u5404\u4e2a\u9886\u57df\u7684\u95ee\u9898\uff0c\u4f8b\u5982 STEM \u9886\u57df\u3001\u7ade\u8d5b\u548c\u4e00\u822c\u63a8\u7406\u4efb\u52a1\uff0c\u5305\u542b\u7eaf\u6587\u672c\u548c\u56fe\u50cf\u6587\u672c\u95ee\u7b54\u6570\u636e\u3002\u6b64\u5916\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u6807\u8bb0\u7cfb\u7edf\u6765\u6309\u9886\u57df\u548c\u5b66\u79d1\u5bf9\u63d0\u793a\u8fdb\u884c\u5206\u7c7b\uff0c\u4ee5\u786e\u4fdd\u4e0d\u540c\u4e3b\u9898\u9886\u57df\u7684\u5747\u8861\u8868\u793a (M. Li et al. 2023; W. Liu et al. 2023)\u3002<\/p>\n<p>\u6211\u4eec\u91c7\u7528\u4e86\u4e00\u79cd\u57fa\u4e8e\u6a21\u578b\u7684\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u5229\u7528\u6a21\u578b\u81ea\u8eab\u7684\u80fd\u529b\u6765\u81ea\u9002\u5e94\u5730\u8bc4\u4f30\u6bcf\u4e2a\u63d0\u793a\u7684\u96be\u5ea6\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u63d0\u793a\uff0cSFT \u6a21\u578b\u4f7f\u7528\u76f8\u5bf9\u8f83\u9ad8\u7684\u91c7\u6837\u6e29\u5ea6\u751f\u6210\u5341\u6b21\u7b54\u6848\u3002\u7136\u540e\u8ba1\u7b97\u901a\u8fc7\u7387\uff0c\u5e76\u5c06\u5176\u7528\u4f5c\u63d0\u793a\u96be\u5ea6\u7684\u4ee3\u7406\u2014\u2014\u901a\u8fc7\u7387\u8d8a\u4f4e\uff0c\u96be\u5ea6\u8d8a\u9ad8\u3002\u8fd9\u79cd\u65b9\u6cd5\u5141\u8bb8\u96be\u5ea6\u8bc4\u4f30\u4e0e\u6a21\u578b\u7684\u5185\u5728\u80fd\u529b\u4fdd\u6301\u4e00\u81f4\uff0c\u4f7f\u5176\u5728 RL \u8bad\u7ec3\u4e2d\u975e\u5e38\u6709\u6548\u3002\u901a\u8fc7\u5229\u7528\u8fd9\u79cd\u65b9\u6cd5\uff0c\u6211\u4eec\u53ef\u4ee5\u9884\u5148\u7b5b\u9009\u6389\u5927\u591a\u6570\u7410\u788e\u7684\u6848\u4f8b\uff0c\u5e76\u5728 RL \u8bad\u7ec3\u671f\u95f4\u8f7b\u677e\u63a2\u7d22\u4e0d\u540c\u7684\u91c7\u6837\u7b56\u7565\u3002<\/p>\n<p>\u4e3a\u4e86\u907f\u514d\u6f5c\u5728\u7684\u5956\u52b1\u653b\u51fb (Everitt et al. 2021; Pan et al. 2022)\uff0c\u6211\u4eec\u9700\u8981\u786e\u4fdd\u53ef\u4ee5\u51c6\u786e\u9a8c\u8bc1\u6bcf\u4e2a\u63d0\u793a\u7684\u63a8\u7406\u8fc7\u7a0b\u548c\u6700\u7ec8\u7b54\u6848\u3002\u7ecf\u9a8c\u89c2\u5bdf\u8868\u660e\uff0c\u4e00\u4e9b\u590d\u6742\u7684\u63a8\u7406\u95ee\u9898\u53ef\u80fd\u5177\u6709\u76f8\u5bf9\u7b80\u5355\u4e14\u6613\u4e8e\u731c\u6d4b\u7684\u7b54\u6848\uff0c\u4ece\u800c\u5bfc\u81f4\u865a\u5047\u7684\u6b63\u5411\u9a8c\u8bc1\u2014\u2014\u6a21\u578b\u901a\u8fc7\u4e0d\u6b63\u786e\u7684\u63a8\u7406\u8fc7\u7a0b\u8fbe\u5230\u6b63\u786e\u7684\u7b54\u6848\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u6392\u9664\u4e86\u5bb9\u6613\u51fa\u73b0\u6b64\u7c7b\u9519\u8bef\u7684\u95ee\u9898\uff0c\u4f8b\u5982\u591a\u9879\u9009\u62e9\u9898\u3001\u771f\/\u5047\u9898\u548c\u57fa\u4e8e\u8bc1\u660e\u7684\u95ee\u9898\u3002\u6b64\u5916\uff0c\u5bf9\u4e8e\u4e00\u822c\u7684\u95ee\u7b54\u4efb\u52a1\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u65b9\u6cd5\u6765\u8bc6\u522b\u548c\u5220\u9664\u6613\u4e8e\u653b\u51fb\u7684\u63d0\u793a\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u63d0\u793a\u6a21\u578b\u5728\u6ca1\u6709\u4efb\u4f55 CoT \u63a8\u7406\u6b65\u9aa4\u7684\u60c5\u51b5\u4e0b\u731c\u6d4b\u53ef\u80fd\u7684\u7b54\u6848\u3002\u5982\u679c\u6a21\u578b\u5728 N \u6b21\u5c1d\u8bd5\u4e2d\u9884\u6d4b\u51fa\u6b63\u786e\u7684\u7b54\u6848\uff0c\u5219\u8be5\u63d0\u793a\u88ab\u8ba4\u4e3a\u592a\u5bb9\u6613\u653b\u51fb\u800c\u88ab\u5220\u9664\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u8bbe\u7f6e N = 8 \u53ef\u4ee5\u5220\u9664\u5927\u591a\u6570\u5bb9\u6613\u653b\u51fb\u7684\u63d0\u793a\u3002\u5f00\u53d1\u66f4\u9ad8\u7ea7\u7684\u9a8c\u8bc1\u6a21\u578b\u4ecd\u7136\u662f\u672a\u6765\u7814\u7a76\u7684\u4e00\u4e2a\u5f00\u653e\u65b9\u5411\u3002<\/p>\n<h3>2.2 Long-CoT \u76d1\u7763\u5fae\u8c03<\/h3>\n<p>\u901a\u8fc7\u6539\u8fdb\u7684 RL \u63d0\u793a\u96c6\uff0c\u6211\u4eec\u91c7\u7528\u63d0\u793a\u5de5\u7a0b\u6765\u6784\u5efa\u4e00\u4e2a\u5c0f\u800c\u9ad8\u8d28\u91cf\u7684 long-CoT \u70ed\u8eab\u6570\u636e\u96c6\uff0c\u5176\u4e2d\u5305\u542b\u9488\u5bf9\u6587\u672c\u548c\u56fe\u50cf\u8f93\u5165\u7684\u51c6\u786e\u9a8c\u8bc1\u7684\u63a8\u7406\u8def\u5f84\u3002\u8fd9\u79cd\u65b9\u6cd5\u7c7b\u4f3c\u4e8e\u62d2\u7edd\u91c7\u6837 (RS)\uff0c\u4f46\u4fa7\u91cd\u4e8e\u901a\u8fc7\u63d0\u793a\u5de5\u7a0b\u751f\u6210 long-CoT \u63a8\u7406\u8def\u5f84\u3002\u7531\u6b64\u4ea7\u751f\u7684\u70ed\u8eab\u6570\u636e\u96c6\u65e8\u5728\u5c01\u88c5\u5bf9\u7c7b\u4eba\u63a8\u7406\u81f3\u5173\u91cd\u8981\u7684\u5173\u952e\u8ba4\u77e5\u8fc7\u7a0b\uff0c\u4f8b\u5982\u89c4\u5212\uff08\u6a21\u578b\u7cfb\u7edf\u5730\u6982\u8ff0\u6267\u884c\u524d\u7684\u6b65\u9aa4\uff09\uff1b\u8bc4\u4f30\uff08\u5305\u62ec\u5bf9\u4e2d\u95f4\u6b65\u9aa4\u7684\u6279\u5224\u6027\u8bc4\u4f30\uff09\uff1b\u53cd\u601d\uff08\u4f7f\u6a21\u578b\u80fd\u591f\u91cd\u65b0\u8003\u8651\u548c\u6539\u8fdb\u5176\u65b9\u6cd5\uff09\uff1b\u4ee5\u53ca\u63a2\u7d22\uff08\u9f13\u52b1\u8003\u8651\u66ff\u4ee3\u89e3\u51b3\u65b9\u6848\uff09\u3002\u901a\u8fc7\u5728\u8fd9\u4e2a\u70ed\u8eab\u6570\u636e\u96c6\u4e0a\u6267\u884c\u8f7b\u91cf\u7ea7\u7684 SFT\uff0c\u6211\u4eec\u53ef\u4ee5\u6709\u6548\u5730\u5f15\u5bfc\u6a21\u578b\u5185\u5316\u8fd9\u4e9b\u63a8\u7406\u7b56\u7565\u3002\u56e0\u6b64\uff0c\u7ecf\u8fc7\u5fae\u8c03\u7684 long-CoT \u6a21\u578b\u8868\u73b0\u51fa\u751f\u6210\u66f4\u8be6\u7ec6\u4e14\u903b\u8f91\u4e0a\u8fde\u8d2f\u7684\u54cd\u5e94\u7684\u6539\u8fdb\u80fd\u529b\uff0c\u4ece\u800c\u589e\u5f3a\u4e86\u5176\u5728\u5404\u79cd\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u3002<\/p>\n<h3>2.3 \u5f3a\u5316\u5b66\u4e60<\/h3>\n<h4>2.3.1 \u95ee\u9898\u8bbe\u5b9a<\/h4>\n<p>\u7ed9\u5b9a\u4e00\u4e2a\u8bad\u7ec3\u6570\u636e\u96c6 D = {(xi, y<em>i)}^n_i=1\uff0c\u5176\u4e2d\u5305\u542b\u95ee\u9898 xi \u548c\u5bf9\u5e94\u7684\u771f\u5b9e\u7b54\u6848 y<\/em>i\uff0c\u6211\u4eec\u7684\u76ee\u6807\u662f\u8bad\u7ec3\u7b56\u7565\u6a21\u578b \u03c0\u03b8 \u4ee5\u51c6\u786e\u5730\u89e3\u51b3\u6d4b\u8bd5\u95ee\u9898\u3002\u5728\u590d\u6742\u7684\u63a8\u7406\u80cc\u666f\u4e0b\uff0c\u95ee\u9898 x \u5230\u89e3\u51b3\u65b9\u6848 y \u7684\u6620\u5c04\u5e76\u975e\u6613\u4e8b\u3002\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u601d\u7ef4\u94fe (CoT) \u65b9\u6cd5\u5efa\u8bae\u4f7f\u7528\u4e00\u7cfb\u5217\u4e2d\u95f4\u6b65\u9aa4 z = (z1, z2, &#8230;, zm) \u6765\u6865\u63a5 x \u548c y\uff0c\u5176\u4e2d\u6bcf\u4e2a zi \u90fd\u662f\u4e00\u4e2a\u8fde\u8d2f\u7684 Token \u5e8f\u5217\uff0c\u53ef\u4f5c\u4e3a\u89e3\u51b3\u95ee\u9898\u7684\u91cd\u8981\u7684\u4e2d\u95f4\u6b65\u9aa4 (J. Wei et al. 2022)\u3002\u5728\u89e3\u51b3\u95ee\u9898 x \u65f6\uff0c\u4ee5\u81ea\u56de\u5f52\u65b9\u5f0f\u91c7\u6837\u601d\u7ef4 zt ~ \u03c0\u03b8(\u00b7|x, z1, &#8230;, zt\u22121)\uff0c\u7136\u540e\u91c7\u6837\u6700\u7ec8\u7b54\u6848 y ~ \u03c0\u03b8(\u00b7|x, z1,&#8230;, zm)\u3002\u6211\u4eec\u4f7f\u7528 y, z ~ \u03c0\u03b8 \u6765\u8868\u793a\u6b64\u91c7\u6837\u8fc7\u7a0b\u3002\u8bf7\u6ce8\u610f\uff0c\u601d\u7ef4\u548c\u6700\u7ec8\u7b54\u6848\u90fd\u662f\u4f5c\u4e3a\u8bed\u8a00\u5e8f\u5217\u91c7\u6837\u7684\u3002<\/p>\n<p>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u589e\u5f3a\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u89c4\u5212\u7b97\u6cd5\u6765\u63a2\u7d22\u5404\u79cd\u601d\u7ef4\u8fc7\u7a0b\uff0c\u4ece\u800c\u5728\u63a8\u7406\u65f6\u751f\u6210\u6539\u8fdb\u7684 CoT (Yao et al. 2024; Y. Wu et al. 2024; Snell et al. 2024)\u3002\u8fd9\u4e9b\u65b9\u6cd5\u7684\u6838\u5fc3\u89c1\u89e3\u662f\u663e\u5f0f\u6784\u5efa\u7531\u503c\u4f30\u8ba1\u5f15\u5bfc\u7684\u601d\u7ef4\u641c\u7d22\u6811\u3002\u8fd9\u5141\u8bb8\u6a21\u578b\u63a2\u7d22\u601d\u7ef4\u8fc7\u7a0b\u7684\u5404\u79cd\u5ef6\u7eed\uff0c\u6216\u8005\u5728\u9047\u5230\u6b7b\u80e1\u540c\u65f6\u56de\u6eaf\u4ee5\u8c03\u67e5\u65b0\u7684\u65b9\u5411\u3002\u66f4\u8be6\u7ec6\u5730\u8bf4\uff0c\u4ee4 T \u4e3a\u641c\u7d22\u6811\uff0c\u5176\u4e2d\u6bcf\u4e2a\u8282\u70b9\u8868\u793a\u90e8\u5206\u89e3\u51b3\u65b9\u6848 s = (x, z1:|s|)\u3002\u8fd9\u91cc s \u7531\u95ee\u9898 x \u548c\u4e00\u7cfb\u5217\u601d\u7ef4 z1:|s| = (z1, &#8230;, z|s|) \u7ec4\u6210\uff0c\u8fd9\u4e9b\u601d\u7ef4\u901a\u5411\u8be5\u8282\u70b9\uff0c\u5176\u4e2d |s| \u8868\u793a\u5e8f\u5217\u4e2d\u601d\u7ef4\u7684\u6570\u91cf\u3002\u89c4\u5212\u7b97\u6cd5\u4f7f\u7528\u8bc4\u8bba\u6a21\u578b v \u6765\u63d0\u4f9b\u53cd\u9988 v(x, z1:|s|)\uff0c\u8fd9\u6709\u52a9\u4e8e\u8bc4\u4f30\u5f53\u524d\u89e3\u51b3\u95ee\u9898\u7684\u8fdb\u5c55\u5e76\u8bc6\u522b\u73b0\u6709\u90e8\u5206\u89e3\u51b3\u65b9\u6848\u4e2d\u7684\u4efb\u4f55\u9519\u8bef\u3002\u6211\u4eec\u6ce8\u610f\u5230\uff0c\u53cd\u9988\u53ef\u4ee5\u7531\u5224\u522b\u5206\u6570\u6216\u8bed\u8a00\u5e8f\u5217\u63d0\u4f9b (L. Zhang et al. 2024)\u3002\u5728\u6240\u6709 s \u2208 T \u7684\u53cd\u9988\u7684\u6307\u5bfc\u4e0b\uff0c\u89c4\u5212\u7b97\u6cd5\u9009\u62e9\u6700\u6709\u53ef\u80fd\u6269\u5c55\u7684\u8282\u70b9\uff0c\u4ece\u800c\u6269\u5927\u641c\u7d22\u6811\u3002\u4e0a\u8ff0\u8fc7\u7a0b\u4f1a\u91cd\u590d\u8fed\u4ee3\uff0c\u76f4\u5230\u5f97\u51fa\u5b8c\u6574\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u53ef\u4ee5\u4ece\u7b97\u6cd5\u7684\u89d2\u5ea6\u6765\u5904\u7406\u89c4\u5212\u7b97\u6cd5\u3002\u7ed9\u5b9a\u7b2c t \u6b21\u8fed\u4ee3\u65f6\u53ef\u7528\u7684\u8fc7\u53bb\u641c\u7d22\u5386\u53f2 (s1, v(s1), &#8230;, st\u22121, v(st\u22121))\uff0c\u89c4\u5212\u7b97\u6cd5 A \u8fed\u4ee3\u5730\u786e\u5b9a\u4e0b\u4e00\u4e2a\u641c\u7d22\u65b9\u5411 A(st|s1, v(s1),&#8230;, st\u22121, v(st\u22121)) \u5e76\u4e3a\u5f53\u524d\u641c\u7d22\u8fdb\u5ea6 A(v(st)|s1, v(s1),&#8230;, st) \u63d0\u4f9b\u53cd\u9988\u3002\u7531\u4e8e\u601d\u7ef4\u548c\u53cd\u9988\u90fd\u53ef\u4ee5\u88ab\u89c6\u4e3a\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\uff0c\u5e76\u4e14\u8fd9\u4e9b\u7ec4\u4ef6\u90fd\u53ef\u4ee5\u8868\u793a\u4e3a\u8bed\u8a00 Token \u5e8f\u5217\uff0c\u56e0\u6b64\u6211\u4eec\u4f7f\u7528 z \u6765\u66ff\u6362 s \u548c v \u4ee5\u7b80\u5316\u8868\u793a\u6cd5\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5c06\u89c4\u5212\u7b97\u6cd5\u89c6\u4e3a\u76f4\u63a5\u4f5c\u7528\u4e8e\u4e00\u7cfb\u5217\u63a8\u7406\u6b65\u9aa4 A(\u00b7|z1, z2, &#8230; ) \u7684\u6620\u5c04\u3002\u5728\u8fd9\u4e2a\u6846\u67b6\u4e2d\uff0c\u89c4\u5212\u7b97\u6cd5\u4f7f\u7528\u7684\u641c\u7d22\u6811\u4e2d\u5b58\u50a8\u7684\u6240\u6709\u4fe1\u606f\u90fd\u88ab\u5c55\u5e73\u5230\u63d0\u4f9b\u7ed9\u7b97\u6cd5\u7684\u5b8c\u6574\u4e0a\u4e0b\u6587\u4e2d\u3002\u8fd9\u4e3a\u751f\u6210\u9ad8\u8d28\u91cf CoT \u63d0\u4f9b\u4e86\u4e00\u4e2a\u6709\u8da3\u7684\u89c6\u89d2\uff1a\u6211\u4eec\u53ef\u4ee5\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\u6765\u8fd1\u4f3c\u6b64\u8fc7\u7a0b\uff0c\u800c\u4e0d\u662f\u663e\u5f0f\u6784\u5efa\u641c\u7d22\u6811\u5e76\u5b9e\u73b0\u89c4\u5212\u7b97\u6cd5\u3002\u5728\u8fd9\u91cc\uff0c\u601d\u7ef4\u7684\u6570\u91cf\uff08\u5373\uff0c\u8bed\u8a00 Token\uff09\u7c7b\u4f3c\u4e8e\u4f20\u7edf\u4e0a\u5206\u914d\u7ed9\u89c4\u5212\u7b97\u6cd5\u7684\u8ba1\u7b97\u9884\u7b97\u3002\u6700\u8fd1\u5728\u957f\u4e0a\u4e0b\u6587\u7a97\u53e3\u65b9\u9762\u7684\u8fdb\u6b65\u4fc3\u8fdb\u4e86\u8bad\u7ec3\u548c\u6d4b\u8bd5\u9636\u6bb5\u7684\u65e0\u7f1d\u53ef\u6269\u5c55\u6027\u3002\u5982\u679c\u53ef\u884c\uff0c\u6b64\u65b9\u6cd5\u4f7f\u6a21\u578b\u80fd\u591f\u901a\u8fc7\u81ea\u56de\u5f52\u9884\u6d4b\u76f4\u63a5\u5728\u63a8\u7406\u7a7a\u95f4\u4e0a\u8fd0\u884c\u9690\u5f0f\u641c\u7d22\u3002\u56e0\u6b64\uff0c\u6a21\u578b\u4e0d\u4ec5\u5b66\u4e60\u89e3\u51b3\u4e00\u7ec4\u8bad\u7ec3\u95ee\u9898\uff0c\u800c\u4e14\u8fd8\u5f00\u53d1\u4e86\u6709\u6548\u89e3\u51b3\u5355\u4e2a\u95ee\u9898\u7684\u80fd\u529b\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u5bf9\u672a\u89c1\u6d4b\u8bd5\u95ee\u9898\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u6211\u4eec\u8003\u8651\u8bad\u7ec3\u6a21\u578b\u4ee5\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60 (RL) \u751f\u6210 CoT (OpenAI 2024)\u3002\u4ee4 r \u4e3a\u5956\u52b1\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u6839\u636e\u771f\u5b9e\u7b54\u6848 y* \u4e3a\u7ed9\u5b9a\u95ee\u9898 x \u63d0\u51fa\u7684\u7b54\u6848 y \u7684\u6b63\u786e\u6027\u63d0\u4f9b\u4f9d\u636e\uff0c\u65b9\u6cd5\u662f\u5206\u914d\u4e00\u4e2a\u503c r(x, y, y*) \u2208 {0, 1}\u3002\u5bf9\u4e8e\u53ef\u9a8c\u8bc1\u7684\u95ee\u9898\uff0c\u5956\u52b1\u76f4\u63a5\u7531\u9884\u5b9a\u4e49\u7684\u6807\u51c6\u6216\u89c4\u5219\u786e\u5b9a\u3002\u4f8b\u5982\uff0c\u5728\u7f16\u7801\u95ee\u9898\u4e2d\uff0c\u6211\u4eec\u4f1a\u8bc4\u4f30\u7b54\u6848\u662f\u5426\u901a\u8fc7\u4e86\u6d4b\u8bd5\u7528\u4f8b\u3002\u5bf9\u4e8e\u5177\u6709\u81ea\u7531\u5f62\u5f0f\u771f\u5b9e\u7b54\u6848\u7684\u95ee\u9898\uff0c\u6211\u4eec\u8bad\u7ec3\u5956\u52b1\u6a21\u578b r(x, y, y*) \u6765\u9884\u6d4b\u7b54\u6848\u662f\u5426\u4e0e\u771f\u5b9e\u7b54\u6848\u5339\u914d\u3002\u7ed9\u5b9a\u95ee\u9898 x\uff0c\u6a21\u578b \u03c0\u03b8 \u901a\u8fc7\u91c7\u6837\u8fc7\u7a0b z ~ \u03c0\u03b8(\u00b7|x)\uff0cy ~ \u03c0\u03b8(\u00b7|x, z) \u751f\u6210 CoT \u548c\u6700\u7ec8\u7b54\u6848\u3002\u751f\u6210\u7684 CoT \u7684\u8d28\u91cf\u53d6\u51b3\u4e8e\u5b83\u662f\u5426\u53ef\u4ee5\u4ea7\u751f\u6b63\u786e\u7684\u6700\u7ec8\u7b54\u6848\u3002\u603b\u4e4b\uff0c\u6211\u4eec\u8003\u8651\u4ee5\u4e0b\u76ee\u6807\u6765\u4f18\u5316\u7b56\u7565<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19054\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c83e1eb19f10ce.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"483\" height=\"74\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c83e1eb19f10ce.png 483w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c83e1eb19f10ce-300x46.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c83e1eb19f10ce-18x3.png 18w\" sizes=\"auto, (max-width: 483px) 100vw, 483px\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>\u901a\u8fc7\u6269\u5c55 RL \u8bad\u7ec3\uff0c\u6211\u4eec\u65e8\u5728\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u53ef\u4ee5\u5229\u7528\u7b80\u5355\u57fa\u4e8e\u63d0\u793a\u7684 CoT \u548c\u89c4\u5212\u589e\u5f3a\u7684 CoT \u7684\u4f18\u52bf\u3002\u8be5\u6a21\u578b\u4ecd\u7136\u5728\u63a8\u7406\u671f\u95f4\u81ea\u56de\u5f52\u5730\u91c7\u6837\u8bed\u8a00\u5e8f\u5217\uff0c\u4ece\u800c\u907f\u514d\u4e86\u90e8\u7f72\u671f\u95f4\u9ad8\u7ea7\u89c4\u5212\u7b97\u6cd5\u6240\u9700\u7684\u590d\u6742\u5e76\u884c\u5316\u3002\u7136\u800c\uff0c\u4e0e\u7b80\u5355\u7684\u57fa\u4e8e\u63d0\u793a\u7684\u65b9\u6cd5\u7684\u4e00\u4e2a\u5173\u952e\u533a\u522b\u5728\u4e8e\uff0c\u6a21\u578b\u4e0d\u5e94\u4ec5\u4ec5\u9075\u5faa\u4e00\u7cfb\u5217\u63a8\u7406\u6b65\u9aa4\u3002\u76f8\u53cd\uff0c\u5b83\u8fd8\u5e94\u8be5\u5b66\u4e60\u5173\u952e\u7684\u89c4\u5212\u6280\u80fd\uff0c\u5305\u62ec\u9519\u8bef\u8bc6\u522b\u3001\u56de\u6eaf\u548c\u89e3\u51b3\u65b9\u6848\u6539\u8fdb\uff0c\u65b9\u6cd5\u662f\u5229\u7528\u6574\u4e2a\u63a2\u7d22\u601d\u7ef4\u96c6\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/p>\n<h4>2.3.2 \u7b56\u7565\u4f18\u5316<\/h4>\n<p>\u6211\u4eec\u5e94\u7528\u5728\u7ebf\u7b56\u7565\u955c\u50cf\u4e0b\u964d\u7684\u53d8\u4f53\u4f5c\u4e3a\u6211\u4eec\u7684\u8bad\u7ec3\u7b97\u6cd5 (Abbasi-Yadkori et al. 2019; Mei et al. 2019; Tomar et al. 2020)\u3002\u8be5\u7b97\u6cd5\u8fed\u4ee3\u6267\u884c\u3002\u5728\u7b2c i \u6b21\u8fed\u4ee3\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u5f53\u524d\u6a21\u578b \u03c0\u03b8i \u4f5c\u4e3a\u53c2\u8003\u6a21\u578b\uff0c\u5e76\u4f18\u5316\u4ee5\u4e0b\u76f8\u5bf9\u71b5\u6b63\u5219\u5316\u7b56\u7565\u4f18\u5316\u95ee\u9898\uff0c<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19055\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/398b67e4320f86e.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"812\" height=\"65\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/398b67e4320f86e.png 812w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/398b67e4320f86e-300x24.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/398b67e4320f86e-768x61.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/398b67e4320f86e-18x1.png 18w\" sizes=\"auto, (max-width: 812px) 100vw, 812px\" \/><\/p>\n<p>\u5176\u4e2d \u03c4 &gt; 0 \u662f\u63a7\u5236\u6b63\u5219\u5316\u7a0b\u5ea6\u7684\u53c2\u6570\u3002\u6b64\u76ee\u6807\u5177\u6709\u5c01\u95ed\u5f62\u5f0f\u7684\u89e3<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19056\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f08f1ad5bfe77bc.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"602\" height=\"52\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f08f1ad5bfe77bc.png 602w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f08f1ad5bfe77bc-300x26.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f08f1ad5bfe77bc-18x2.png 18w\" sizes=\"auto, (max-width: 602px) 100vw, 602px\" \/><\/p>\n<p>\u5176\u4e2d Z = \u03a3y&#8217;,z&#8217; \u03c0\u03b8i (y&#8217;, z&#8217;|x) exp(r(x, y&#8217;, y*)\/\u03c4) \u662f\u5f52\u4e00\u5316\u56e0\u5b50\u3002\u53d6\u4e24\u8fb9\u7684\u5bf9\u6570\uff0c\u6211\u4eec\u5f97\u5230\u5bf9\u4e8e\u4efb\u4f55 (y, z)\uff0c\u4ee5\u4e0b\u7ea6\u675f\u90fd\u6ee1\u8db3\uff0c\u8fd9\u5141\u8bb8\u6211\u4eec\u5728\u4f18\u5316\u671f\u95f4\u5229\u7528\u79bb\u7ebf\u7b56\u7565\u6570\u636e<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19057\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/878d6132f5ff5b2.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"551\" height=\"98\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/878d6132f5ff5b2.png 551w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/878d6132f5ff5b2-300x53.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/878d6132f5ff5b2-18x3.png 18w\" sizes=\"auto, (max-width: 551px) 100vw, 551px\" \/><\/p>\n<p>\u8fd9\u6fc0\u53d1\u4e86\u4ee5\u4e0b\u66ff\u4ee3\u635f\u5931<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19058\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/fa96ca22e5fc2ac.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1015\" height=\"113\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/fa96ca22e5fc2ac.png 1015w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/fa96ca22e5fc2ac-300x33.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/fa96ca22e5fc2ac-768x86.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/fa96ca22e5fc2ac-18x2.png 18w\" sizes=\"auto, (max-width: 1015px) 100vw, 1015px\" \/><\/p>\n<p>\u4e3a\u4e86\u8fd1\u4f3c \u03c4 log Z\uff0c\u6211\u4eec\u4f7f\u7528\u6837\u672c (y1, z1), &#8230;, (yk, zk) ~ \u03c0\u03b8\uff1a\u03c4 log Z \u2248 log 1\/k \u03a3^k_j=1 exp(r(x, yj, y*)\/\u03c4)\u3002\u6211\u4eec\u8fd8\u53d1\u73b0\u4f7f\u7528\u91c7\u6837\u5956\u52b1\u7684\u7ecf\u9a8c\u5e73\u5747\u503c \u03c4 = mean(r(x, y1, y*), &#8230;, r(x, yk, y*)) \u4f1a\u4ea7\u751f\u6709\u6548\u7684\u5b9e\u9645\u7ed3\u679c\u3002\u8fd9\u662f\u5408\u7406\u7684\uff0c\u56e0\u4e3a\u5f53 \u03c4 \u2192 \u221e \u65f6\uff0c\u03c4 log Z \u63a5\u8fd1 \u03c0\u03b8i \u4e0b\u7684\u9884\u671f\u5956\u52b1\u3002\u6700\u540e\uff0c\u6211\u4eec\u901a\u8fc7\u53d6\u66ff\u4ee3\u635f\u5931\u7684\u68af\u5ea6\u6765\u5f97\u51fa\u6211\u4eec\u7684\u5b66\u4e60\u7b97\u6cd5\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u95ee\u9898 x\uff0c\u4f7f\u7528\u53c2\u8003\u7b56\u7565 \u03c0\u03b8i \u91c7\u6837 k \u4e2a\u54cd\u5e94\uff0c\u5e76\u4e14\u68af\u5ea6\u7531\u4e0b\u5f0f\u7ed9\u51fa<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19059\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/97e2d2068c0b04a.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"994\" height=\"124\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/97e2d2068c0b04a.png 994w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/97e2d2068c0b04a-300x37.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/97e2d2068c0b04a-768x96.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/97e2d2068c0b04a-18x2.png 18w\" sizes=\"auto, (max-width: 994px) 100vw, 994px\" \/><\/p>\n<p>\u5bf9\u4e8e\u90a3\u4e9b\u719f\u6089\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u7684\u4eba\u6765\u8bf4\uff0c\u6b64\u68af\u5ea6\u7c7b\u4f3c\u4e8e\u4f7f\u7528\u91c7\u6837\u5956\u52b1\u7684\u5e73\u5747\u503c\u4f5c\u4e3a\u57fa\u7ebf\u7684 (2) \u7684\u7b56\u7565\u68af\u5ea6 (Kool et al. 2019; Ahmadian et al. 2024)\u3002\u4e3b\u8981\u533a\u522b\u5728\u4e8e\u54cd\u5e94\u662f\u4ece \u03c0\u03b8i \u800c\u4e0d\u662f\u5728\u7ebf\u7b56\u7565\u4e2d\u91c7\u6837\u7684\uff0c\u5e76\u4e14\u5e94\u7528\u4e86 l2 \u6b63\u5219\u5316\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u5176\u89c6\u4e3a\u901a\u5e38\u7684\u5728\u7ebf\u6b63\u5219\u5316\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\u5230\u79bb\u7ebf\u7b56\u7565\u60c5\u51b5\u7684\u81ea\u7136\u6269\u5c55 (Nachum et al. 2017)\u3002\u6211\u4eec\u4ece D \u4e2d\u91c7\u6837\u4e00\u6279\u95ee\u9898\uff0c\u5e76\u5c06\u53c2\u6570\u66f4\u65b0\u4e3a \u03b8i+1\uff0c\u968f\u540e\u8be5\u53c2\u6570\u5c06\u4f5c\u4e3a\u4e0b\u4e00\u6b21\u8fed\u4ee3\u7684\u53c2\u8003\u7b56\u7565\u3002\u7531\u4e8e\u6bcf\u6b21\u8fed\u4ee3\u90fd\u4f1a\u7531\u4e8e\u53c2\u8003\u7b56\u7565\u7684\u53d8\u5316\u800c\u8003\u8651\u4e0d\u540c\u7684\u4f18\u5316\u95ee\u9898\uff0c\u56e0\u6b64\u6211\u4eec\u8fd8\u5728\u6bcf\u6b21\u8fed\u4ee3\u5f00\u59cb\u65f6\u91cd\u7f6e\u4f18\u5316\u5668\u3002<\/p>\n<p>\u6211\u4eec\u6392\u9664\u4e86\u8bad\u7ec3\u7cfb\u7edf\u4e2d\u7684\u4ef7\u503c\u7f51\u7edc\uff0c\u8be5\u7f51\u7edc\u4e5f\u5728\u4e4b\u524d\u7684\u7814\u7a76\u4e2d\u88ab\u5229\u7528 (Ahmadian et al. 2024)\u3002\u867d\u7136\u8fd9\u79cd\u8bbe\u8ba1\u9009\u62e9\u663e\u7740\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\uff0c\u4f46\u6211\u4eec\u4e5f\u5047\u8bbe\u7ecf\u5178 RL \u4e2d\u7528\u4e8e\u4fe1\u7528\u5206\u914d\u7684\u4ef7\u503c\u51fd\u6570\u7684\u4f20\u7edf\u4f7f\u7528\u53ef\u80fd\u4e0d\u9002\u5408\u6211\u4eec\u7684\u80cc\u666f\u3002\u8003\u8651\u8fd9\u6837\u4e00\u79cd\u60c5\u51b5\uff1a\u6a21\u578b\u751f\u6210\u4e86\u90e8\u5206 CoT (z1, z2, &#8230;, zt) \u5e76\u4e14\u6709\u4e24\u4e2a\u6f5c\u5728\u7684\u4e0b\u4e00\u4e2a\u63a8\u7406\u6b65\u9aa4\uff1azt+1 \u548c z&#8217;t+1\u3002\u5047\u8bbe zt+1 \u76f4\u63a5\u5bfc\u81f4\u4e86\u6b63\u786e\u7684\u7b54\u6848\uff0c\u800c z&#8217;t+1 \u5305\u542b\u4e00\u4e9b\u9519\u8bef\u3002\u5982\u679c\u53ef\u4ee5\u8bbf\u95ee oracle \u4ef7\u503c\u51fd\u6570\uff0c\u5b83\u5c06\u8868\u660e zt+1 \u6bd4 z&#8217;t+1 \u4fdd\u7559\u66f4\u9ad8\u7684\u4ef7\u503c\u3002\u6839\u636e\u6807\u51c6\u4fe1\u7528\u5206\u914d\u539f\u5219\uff0c\u9009\u62e9 z&#8217;t+1 \u5c06\u53d7\u5230\u60e9\u7f5a\uff0c\u56e0\u4e3a\u5b83\u76f8\u5bf9\u4e8e\u5f53\u524d\u7b56\u7565\u5177\u6709\u8d1f\u4f18\u52bf\u3002\u4f46\u662f\uff0c\u63a2\u7d22 z&#8217;t+1 \u5bf9\u4e8e\u8bad\u7ec3\u6a21\u578b\u751f\u6210 long CoT \u975e\u5e38\u6709\u4ef7\u503c\u3002\u901a\u8fc7\u4f7f\u7528\u4ece long CoT \u5f97\u51fa\u7684\u6700\u7ec8\u7b54\u6848\u7684\u7406\u7531\u4f5c\u4e3a\u5956\u52b1\u4fe1\u53f7\uff0c\u6a21\u578b\u53ef\u4ee5\u4ece\u91c7\u7528 z&#8217;t+1 \u7684\u8bd5\u9519\u6a21\u5f0f\u4e2d\u5b66\u4e60\uff0c\u53ea\u8981\u5b83\u6210\u529f\u6062\u590d\u5e76\u5f97\u51fa\u6b63\u786e\u7684\u7b54\u6848\u3002\u4ece\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\u5f97\u51fa\u7684\u5173\u952e\u7ed3\u8bba\u662f\uff0c\u6211\u4eec\u5e94\u8be5\u9f13\u52b1\u6a21\u578b\u63a2\u7d22\u5404\u79cd\u63a8\u7406\u8def\u5f84\uff0c\u4ee5\u589e\u5f3a\u5176\u89e3\u51b3\u590d\u6742\u95ee\u9898\u7684\u80fd\u529b\u3002\u8fd9\u79cd\u63a2\u7d22\u6027\u65b9\u6cd5\u4f1a\u4ea7\u751f\u5927\u91cf\u7684\u7ecf\u9a8c\uff0c\u8fd9\u4e9b\u7ecf\u9a8c\u652f\u6301\u5173\u952e\u89c4\u5212\u6280\u80fd\u7684\u5f00\u53d1\u3002\u6211\u4eec\u7684\u4e3b\u8981\u76ee\u6807\u4e0d\u662f\u5c06\u8bad\u7ec3\u95ee\u9898\u7684\u51c6\u786e\u6027\u63d0\u9ad8\u5230\u6700\u9ad8\u6c34\u5e73\uff0c\u800c\u662f\u4fa7\u91cd\u4e8e\u4f7f\u6a21\u578b\u638c\u63e1\u6709\u6548\u7684\u95ee\u9898\u89e3\u51b3\u7b56\u7565\uff0c\u4ece\u800c\u6700\u7ec8\u63d0\u9ad8\u5176\u5728\u6d4b\u8bd5\u95ee\u9898\u4e0a\u7684\u6027\u80fd\u3002<\/p>\n<h4>2.3.3 \u957f\u5ea6\u60e9\u7f5a<\/h4>\n<p>\u6211\u4eec\u89c2\u5bdf\u5230\u4e00\u79cd\u8fc7\u5ea6\u601d\u8003\u73b0\u8c61\uff0c\u5373\u6a21\u578b\u5728 RL \u8bad\u7ec3\u671f\u95f4\u7684\u54cd\u5e94\u957f\u5ea6\u663e\u7740\u589e\u52a0\u3002\u867d\u7136\u8fd9\u4f1a\u5bfc\u81f4\u66f4\u597d\u7684\u6027\u80fd\uff0c\u4f46\u8fc7\u5ea6\u5197\u957f\u7684\u63a8\u7406\u8fc7\u7a0b\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u671f\u95f4\u6210\u672c\u9ad8\u6602\uff0c\u5e76\u4e14\u4eba\u7c7b\u901a\u5e38\u4e0d\u559c\u6b22\u8fc7\u5ea6\u601d\u8003\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u957f\u5ea6\u5956\u52b1\u6765\u9650\u5236 Token \u957f\u5ea6\u7684\u5feb\u901f\u589e\u957f\uff0c\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u7684 Token \u6548\u7387\u3002\u7ed9\u5b9a\u95ee\u9898 x \u7684 k \u4e2a\u91c7\u6837\u54cd\u5e94 (y1, z1), &#8230;, (yk, zk) \u548c\u771f\u5b9e\u7b54\u6848 y*\uff0c\u4ee4 len(i) \u4e3a (yi, zi) \u7684\u957f\u5ea6\uff0cmin_len = min_i len(i) \u4e14 max_len = max_i len(i)\u3002\u5982\u679c max_len = min_len\uff0c\u6211\u4eec\u4e3a\u6240\u6709\u54cd\u5e94\u8bbe\u7f6e\u957f\u5ea6\u5956\u52b1\u4e3a\u96f6\uff0c\u56e0\u4e3a\u5b83\u4eec\u5177\u6709\u76f8\u540c\u7684\u957f\u5ea6\u3002\u5426\u5219\uff0c\u957f\u5ea6\u5956\u52b1\u7531\u4e0b\u5f0f\u7ed9\u51fa<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19060\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/13c9f98384e3169.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1174\" height=\"98\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/13c9f98384e3169.png 1174w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/13c9f98384e3169-300x25.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/13c9f98384e3169-1024x85.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/13c9f98384e3169-768x64.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/13c9f98384e3169-18x2.png 18w\" sizes=\"auto, (max-width: 1174px) 100vw, 1174px\" \/><\/p>\n<p>\u672c\u8d28\u4e0a\uff0c\u6211\u4eec\u9f13\u52b1\u8f83\u77ed\u7684\u54cd\u5e94\uff0c\u5e76\u60e9\u7f5a\u6b63\u786e\u7b54\u6848\u4e2d\u8f83\u957f\u7684\u54cd\u5e94\uff0c\u540c\u65f6\u660e\u786e\u60e9\u7f5a\u9519\u8bef\u7b54\u6848\u4e2d\u8f83\u957f\u7684\u54cd\u5e94\u3002\u7136\u540e\u5c06\u6b64\u57fa\u4e8e\u957f\u5ea6\u7684\u5956\u52b1\u4e0e\u6743\u91cd\u53c2\u6570\u6dfb\u52a0\u5230\u539f\u59cb\u5956\u52b1\u4e2d\u3002<\/p>\n<p>\u5728\u6211\u4eec\u7684\u521d\u6b65\u5b9e\u9a8c\u4e2d\uff0c\u957f\u5ea6\u60e9\u7f5a\u53ef\u80fd\u4f1a\u51cf\u6162\u521d\u59cb\u9636\u6bb5\u7684\u8bad\u7ec3\u901f\u5ea6\u3002\u4e3a\u4e86\u7f13\u89e3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5efa\u8bae\u5728\u8bad\u7ec3\u7684\u521d\u59cb\u9636\u6bb5\u9010\u6b65\u70ed\u8eab\u957f\u5ea6\u60e9\u7f5a\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5728\u6ca1\u6709\u957f\u5ea6\u60e9\u7f5a\u7684\u60c5\u51b5\u4e0b\u91c7\u7528\u6807\u51c6\u7b56\u7565\u4f18\u5316\uff0c\u7136\u540e\u5728\u8bad\u7ec3\u7684\u5176\u4f59\u90e8\u5206\u91c7\u7528\u6052\u5b9a\u7684\u957f\u5ea6\u60e9\u7f5a\u3002<\/p>\n<h4>2.3.4 \u91c7\u6837\u7b56\u7565<\/h4>\n<p>\u5c3d\u7ba1 RL \u7b97\u6cd5\u672c\u8eab\u5177\u6709\u76f8\u5bf9\u8f83\u597d\u7684\u91c7\u6837\u5c5e\u6027\uff08\u96be\u5ea6\u66f4\u5927\u7684\u95ee\u9898\u63d0\u4f9b\u66f4\u5927\u7684\u68af\u5ea6\uff09\uff0c\u4f46\u5176\u8bad\u7ec3\u6548\u7387\u53d7\u5230\u9650\u5236\u3002\u56e0\u6b64\uff0c\u4e00\u4e9b\u5b9a\u4e49\u660e\u786e\u7684\u5148\u9a8c\u91c7\u6837\u65b9\u6cd5\u53ef\u4ee5\u4ea7\u751f\u66f4\u5927\u7684\u6027\u80fd\u63d0\u5347\u3002\u6211\u4eec\u5229\u7528\u591a\u79cd\u4fe1\u53f7\u6765\u8fdb\u4e00\u6b65\u6539\u8fdb\u91c7\u6837\u7b56\u7565\u3002\u9996\u5148\uff0c\u6211\u4eec\u6536\u96c6\u7684 RL \u8bad\u7ec3\u6570\u636e\u81ea\u7136\u5e26\u6709\u4e0d\u540c\u7684\u96be\u5ea6\u6807\u7b7e\u3002\u4f8b\u5982\uff0c\u6570\u5b66\u7ade\u8d5b\u95ee\u9898\u6bd4\u5c0f\u5b66\u6570\u5b66\u9898\u66f4\u96be\u3002\u5176\u6b21\uff0c\u7531\u4e8e RL \u8bad\u7ec3\u8fc7\u7a0b\u591a\u6b21\u91c7\u6837\u76f8\u540c\u7684\u95ee\u9898\uff0c\u6211\u4eec\u8fd8\u53ef\u4ee5\u8ddf\u8e2a\u6bcf\u4e2a\u95ee\u9898\u7684\u6210\u529f\u7387\u4f5c\u4e3a\u96be\u5ea6\u7684\u6307\u6807\u3002\u6211\u4eec\u63d0\u51fa\u4e86\u4e24\u79cd\u91c7\u6837\u65b9\u6cd5\u6765\u5229\u7528\u8fd9\u4e9b\u5148\u9a8c\u77e5\u8bc6\u6765\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n<p><strong>\u8bfe\u7a0b\u91c7\u6837<\/strong>\u00a0\u6211\u4eec\u9996\u5148\u5728\u8f83\u7b80\u5355\u7684\u4efb\u52a1\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u7136\u540e\u9010\u6b65\u8fc7\u6e21\u5230\u66f4\u5177\u6311\u6218\u6027\u7684\u4efb\u52a1\u3002\u7531\u4e8e\u521d\u59cb RL \u6a21\u578b\u7684\u6027\u80fd\u6709\u9650\uff0c\u56e0\u6b64\u5c06\u6709\u9650\u7684\u8ba1\u7b97\u9884\u7b97\u82b1\u8d39\u5728\u975e\u5e38\u56f0\u96be\u7684\u95ee\u9898\u4e0a\u901a\u5e38\u4f1a\u4ea7\u751f\u5f88\u5c11\u7684\u6b63\u786e\u6837\u672c\uff0c\u4ece\u800c\u5bfc\u81f4\u8f83\u4f4e\u7684\u8bad\u7ec3\u6548\u7387\u3002\u540c\u65f6\uff0c\u6211\u4eec\u6536\u96c6\u7684\u6570\u636e\u81ea\u7136\u5305\u542b\u7b49\u7ea7\u548c\u96be\u5ea6\u6807\u7b7e\uff0c\u8fd9\u4f7f\u5f97\u57fa\u4e8e\u96be\u5ea6\u7684\u91c7\u6837\u6210\u4e3a\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u7684\u4e00\u79cd\u76f4\u89c2\u6709\u6548\u7684\u65b9\u6cd5\u3002<\/p>\n<p><strong>\u4f18\u5148\u91c7\u6837<\/strong>\u00a0\u9664\u4e86\u8bfe\u7a0b\u91c7\u6837\u5916\uff0c\u6211\u4eec\u8fd8\u4f7f\u7528\u4f18\u5148\u91c7\u6837\u7b56\u7565\u6765\u5173\u6ce8\u6a21\u578b\u6027\u80fd\u4e0d\u4f73\u7684\u95ee\u9898\u3002\u6211\u4eec\u8ddf\u8e2a\u6bcf\u4e2a\u95ee\u9898 i \u7684\u6210\u529f\u7387 si\uff0c\u5e76\u4ee5\u4e0e 1\/si \u6210\u6bd4\u4f8b\u7684\u65b9\u5f0f\u91c7\u6837\u95ee\u9898\uff0c\u4ee5\u4fbf\u6210\u529f\u7387\u8f83\u4f4e\u7684\u95ee\u9898\u83b7\u5f97\u66f4\u9ad8\u7684\u91c7\u6837\u6982\u7387\u3002\u8fd9\u4f1a\u5c06\u6a21\u578b\u7684\u52aa\u529b\u5bfc\u5411\u5176\u6700\u8584\u5f31\u7684\u9886\u57df\uff0c\u4ece\u800c\u52a0\u5feb\u5b66\u4e60\u901f\u5ea6\u5e76\u63d0\u9ad8\u6574\u4f53\u6027\u80fd\u3002<\/p>\n<h4>2.3.5 \u6709\u5173\u8bad\u7ec3\u914d\u65b9\u7684\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f<\/h4>\n<p><strong>\u7528\u4e8e\u7f16\u7801\u7684\u6d4b\u8bd5\u7528\u4f8b\u751f\u6210<\/strong>\u00a0\u7531\u4e8e\u7f51\u4e0a\u8bb8\u591a\u7f16\u7801\u95ee\u9898\u6ca1\u6709\u53ef\u7528\u7684\u6d4b\u8bd5\u7528\u4f8b\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u81ea\u52a8\u751f\u6210\u6d4b\u8bd5\u7528\u4f8b\u7684\u65b9\u6cd5\uff0c\u8fd9\u4e9b\u6d4b\u8bd5\u7528\u4f8b\u53ef\u4ee5\u7528\u4f5c\u4f7f\u7528 RL \u8bad\u7ec3\u6a21\u578b\u7684\u5956\u52b1\u3002\u6211\u4eec\u7684\u91cd\u70b9\u4e3b\u8981\u653e\u5728\u4e0d\u9700\u8981\u7279\u6b8a\u88c1\u5224\u7684\u95ee\u9898\u4e0a\u3002\u6211\u4eec\u8fd8\u5047\u8bbe\u8fd9\u4e9b\u95ee\u9898\u7684\u771f\u5b9e\u7b54\u6848\u662f\u53ef\u7528\u7684\uff0c\u4ee5\u4fbf\u6211\u4eec\u53ef\u4ee5\u5229\u7528\u8fd9\u4e9b\u7b54\u6848\u6765\u751f\u6210\u66f4\u9ad8\u8d28\u91cf\u7684\u6d4b\u8bd5\u7528\u4f8b\u3002<\/p>\n<p>\u6211\u4eec\u5229\u7528\u5e7f\u6cdb\u8ba4\u53ef\u7684\u6d4b\u8bd5\u7528\u4f8b\u751f\u6210\u5e93 CYaRon\u00b9 \u6765\u589e\u5f3a\u6211\u4eec\u7684\u65b9\u6cd5\u3002\u6211\u4eec\u4f7f\u7528\u6211\u4eec\u7684\u57fa\u7840 Kimi k1.5 \u57fa\u4e8e\u95ee\u9898\u9648\u8ff0\u751f\u6210\u6d4b\u8bd5\u7528\u4f8b\u3002CYaRon \u7684\u4f7f\u7528\u8bf4\u660e\u548c\u95ee\u9898\u63cf\u8ff0\u4f5c\u4e3a\u751f\u6210\u5668\u7684\u8f93\u5165\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u9996\u5148\u4f7f\u7528\u751f\u6210\u5668\u751f\u6210 50 \u4e2a\u6d4b\u8bd5\u7528\u4f8b\uff0c\u5e76\u968f\u673a\u91c7\u6837\u6bcf\u4e2a\u6d4b\u8bd5\u7528\u4f8b\u7684 10 \u4e2a\u771f\u5b9e\u7b54\u6848\u63d0\u4ea4\u3002\u6211\u4eec\u9488\u5bf9\u63d0\u4ea4\u8fd0\u884c\u6d4b\u8bd5\u7528\u4f8b\u3002\u5982\u679c\u81f3\u5c11 7 \u4e2a\u63d0\u4ea4\u4e2d\u6709 10 \u4e2a\u4ea7\u751f\u5339\u914d\u7684\u7ed3\u679c\uff0c\u5219\u8ba4\u4e3a\u6d4b\u8bd5\u7528\u4f8b\u6709\u6548\u3002\u7ecf\u8fc7\u6b64\u8f6e\u7b5b\u9009\u540e\uff0c\u6211\u4eec\u83b7\u5f97\u4e86\u4e00\u7ec4\u9009\u5b9a\u7684\u6d4b\u8bd5\u7528\u4f8b\u3002\u5982\u679c\u81f3\u5c11 9 \u4e2a\u63d0\u4ea4\u4e2d\u6709 10 \u4e2a\u901a\u8fc7\u4e86\u6574\u5957\u9009\u5b9a\u7684\u6d4b\u8bd5\u7528\u4f8b\uff0c\u5219\u5c06\u95ee\u9898\u53ca\u5176\u5173\u8054\u7684\u9009\u5b9a\u6d4b\u8bd5\u7528\u4f8b\u6dfb\u52a0\u5230\u6211\u4eec\u7684\u8bad\u7ec3\u96c6\u4e2d\u3002<\/p>\n<p>\u5728\u7edf\u8ba1\u65b9\u9762\uff0c\u4ece 1,000 \u4e2a\u5728\u7ebf\u7ade\u8d5b\u95ee\u9898\u7684\u6837\u672c\u4e2d\uff0c\u5927\u7ea6\u6709 614 \u4e2a\u95ee\u9898\u4e0d\u9700\u8981\u7279\u6b8a\u88c1\u5224\u3002\u6211\u4eec\u5f00\u53d1\u4e86 463 \u4e2a\u6d4b\u8bd5\u7528\u4f8b\u751f\u6210\u5668\uff0c\u8fd9\u4e9b\u751f\u6210\u5668\u4ea7\u751f\u4e86\u81f3\u5c11 40 \u4e2a\u6709\u6548\u7684\u6d4b\u8bd5\u7528\u4f8b\uff0c\u4ece\u800c\u5bfc\u81f4\u6211\u4eec\u7684\u8bad\u7ec3\u96c6\u4e2d\u5305\u542b 323 \u4e2a\u95ee\u9898\u3002<\/p>\n<p><strong>\u6570\u5b66\u7684\u5956\u52b1\u5efa\u6a21<\/strong>\u00a0\u8bc4\u4f30\u6570\u5b66\u89e3\u51b3\u65b9\u6848\u7684\u4e00\u4e2a\u6311\u6218\u662f\uff0c\u4e0d\u540c\u7684\u4e66\u5199\u5f62\u5f0f\u53ef\u4ee5\u8868\u793a\u76f8\u540c\u7684\u57fa\u672c\u7b54\u6848\u3002\u4f8b\u5982\uff0ca\u00b2 &#8211; 4 \u548c (a + 2)(a &#8211; 2) \u53ef\u80fd\u90fd\u662f\u6709\u6548\u89e3\u51b3\u65b9\u6848\u3002\u6211\u4eec\u91c7\u7528\u4e86\u4e24\u79cd\u65b9\u6cd5\u6765\u63d0\u9ad8\u5956\u52b1\u6a21\u578b\u7684\u8bc4\u5206\u51c6\u786e\u6027\uff1a<\/p>\n<ol>\n<li>\u7ecf\u5178 RM\uff1a\u4ece InstructGPT (Ouyang et al. 2022) \u65b9\u6cd5\u4e2d\u6c72\u53d6\u7075\u611f\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86\u4e00\u4e2a\u57fa\u4e8e\u503c\u5934\u7684\u5956\u52b1\u6a21\u578b\uff0c\u5e76\u6536\u96c6\u4e86\u5927\u7ea6 80 \u4e07\u4e2a\u7528\u4e8e\u5fae\u8c03\u7684\u6570\u636e\u70b9\u3002\u8be5\u6a21\u578b\u6700\u7ec8\u5c06\u201c\u95ee\u9898\u201d\u3001\u201c\u53c2\u8003\u7b54\u6848\u201d\u548c\u201c\u54cd\u5e94\u201d\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u8f93\u51fa\u4e00\u4e2a\u8868\u793a\u54cd\u5e94\u662f\u5426\u6b63\u786e\u7684\u6807\u91cf\u3002<\/li>\n<li>\u601d\u7ef4\u94fe RM\uff1a\u6700\u8fd1\u7684\u7814\u7a76 (Ankner et al. 2024; McAleese et al. 2024) \u8868\u660e\uff0c\u4f7f\u7528\u601d\u7ef4\u94fe (CoT) \u63a8\u7406\u589e\u5f3a\u7684\u5956\u52b1\u6a21\u578b\u53ef\u4ee5\u663e\u7740\u4f18\u4e8e\u7ecf\u5178\u65b9\u6cd5\uff0c\u5c24\u5176\u662f\u5728\u9700\u8981\u7ec6\u5fae\u7684\u6b63\u786e\u6027\u6807\u51c6\uff08\u4f8b\u5982\u6570\u5b66\uff09\u7684\u4efb\u52a1\u4e2d\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u6536\u96c6\u4e86\u5927\u7ea6 80 \u4e07\u4e2a CoT \u6807\u8bb0\u7684\u793a\u4f8b\u7684\u540c\u7b49\u5927\u5c0f\u7684\u6570\u636e\u96c6\u6765\u5fae\u8c03 Kimi \u6a21\u578b\u3002\u5728\u4e0e\u7ecf\u5178 RM \u76f8\u540c\u7684\u8f93\u5165\u57fa\u7840\u4e0a\uff0c\u601d\u7ef4\u94fe\u65b9\u6cd5\u5728 JSON \u683c\u5f0f\u4e2d\u660e\u786e\u751f\u6210\u9010\u6b65\u63a8\u7406\u8fc7\u7a0b\uff0c\u7136\u540e\u63d0\u4f9b\u6700\u7ec8\u7684\u6b63\u786e\u6027\u5224\u65ad\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u5f3a\u5927\u4e14\u53ef\u89e3\u91ca\u7684\u5956\u52b1\u4fe1\u53f7\u3002<\/li>\n<\/ol>\n<p>\u5728\u6211\u4eec\u7684\u624b\u52a8\u62bd\u67e5\u4e2d\uff0c\u7ecf\u5178 RM \u7684\u51c6\u786e\u7387\u7ea6\u4e3a 84.4\uff0c\u800c\u601d\u7ef4\u94fe RM \u7684\u51c6\u786e\u7387\u8fbe\u5230 98.5\u3002\u5728 RL \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u601d\u7ef4\u94fe RM \u6765\u786e\u4fdd\u66f4\u6b63\u786e\u7684\u53cd\u9988\u3002<\/p>\n<p><strong>\u89c6\u89c9\u6570\u636e<\/strong>\u00a0\u4e3a\u4e86\u63d0\u9ad8\u6a21\u578b\u7684\u771f\u5b9e\u4e16\u754c\u56fe\u50cf\u63a8\u7406\u80fd\u529b\u5e76\u5b9e\u73b0\u89c6\u89c9\u8f93\u5165\u4e0e\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u4e4b\u95f4\u66f4\u6709\u6548\u7684\u5bf9\u9f50\uff0c\u6211\u4eec\u7684\u89c6\u89c9\u5f3a\u5316\u5b66\u4e60 (Vision RL) \u6570\u636e\u4e3b\u8981\u6765\u81ea\u4e09\u4e2a\u4e0d\u540c\u7684\u7c7b\u522b\uff1a\u771f\u5b9e\u4e16\u754c\u6570\u636e\u3001\u5408\u6210\u89c6\u89c9\u63a8\u7406\u6570\u636e\u548c\u6587\u672c\u6e32\u67d3\u6570\u636e\u3002<\/p>\n<ol>\n<li>\u771f\u5b9e\u4e16\u754c\u6570\u636e\u6db5\u76d6\u4e86\u5404\u4e2a\u5e74\u7ea7\u7684\u5404\u79cd\u79d1\u5b66\u95ee\u9898\uff0c\u8fd9\u4e9b\u95ee\u9898\u9700\u8981\u56fe\u5f62\u7406\u89e3\u548c\u63a8\u7406\uff1b\u9700\u8981\u89c6\u89c9\u611f\u77e5\u548c\u63a8\u7406\u7684\u5b9a\u4f4d\u731c\u6d4b\u4efb\u52a1\uff1b\u4ee5\u53ca\u6d89\u53ca\u7406\u89e3\u590d\u6742\u56fe\u8868\u7684\u5206\u6790\u4efb\u52a1\u7b49\u3002\u8fd9\u4e9b\u6570\u636e\u96c6\u63d0\u9ad8\u4e86\u6a21\u578b\u5728\u771f\u5b9e\u4e16\u754c\u573a\u666f\u4e2d\u6267\u884c\u89c6\u89c9\u63a8\u7406\u7684\u80fd\u529b\u3002<\/li>\n<li>\u5408\u6210\u89c6\u89c9\u63a8\u7406\u6570\u636e\u662f\u4eba\u4e3a\u751f\u6210\u7684\uff0c\u5305\u62ec\u65e8\u5728\u63d0\u9ad8\u7279\u5b9a\u89c6\u89c9\u63a8\u7406\u6280\u80fd\u7684\u7a0b\u5e8f\u521b\u5efa\u7684\u56fe\u50cf\u548c\u573a\u666f\uff0c\u4f8b\u5982\u7406\u89e3\u7a7a\u95f4\u5173\u7cfb\u3001\u51e0\u4f55\u56fe\u6848\u548c\u5bf9\u8c61\u4ea4\u4e92\u3002\u8fd9\u4e9b\u5408\u6210\u6570\u636e\u96c6\u4e3a\u6d4b\u8bd5\u6a21\u578b\u7684\u89c6\u89c9\u63a8\u7406\u80fd\u529b\u63d0\u4f9b\u4e86\u53d7\u63a7\u73af\u5883\uff0c\u5e76\u63d0\u4f9b\u4e86\u65e0\u7a77\u65e0\u5c3d\u7684\u8bad\u7ec3\u793a\u4f8b\u3002<\/li>\n<li>\u6587\u672c\u6e32\u67d3\u6570\u636e\u662f\u901a\u8fc7\u5c06\u6587\u672c\u5185\u5bb9\u8f6c\u6362\u4e3a\u89c6\u89c9\u683c\u5f0f\u521b\u5efa\u7684\uff0c\u4f7f\u6a21\u578b\u5728\u5904\u7406\u8de8\u4e0d\u540c\u6a21\u6001\u7684\u57fa\u4e8e\u6587\u672c\u7684\u67e5\u8be2\u65f6\u4fdd\u6301\u4e00\u81f4\u6027\u3002\u901a\u8fc7\u5c06\u6587\u672c\u6587\u6863\u3001\u4ee3\u7801\u7247\u6bb5\u548c\u7ed3\u6784\u5316\u6570\u636e\u8f6c\u6362\u4e3a\u56fe\u50cf\uff0c\u6211\u4eec\u786e\u4fdd\u6a21\u578b\u65e0\u8bba\u8f93\u5165\u662f\u7eaf\u6587\u672c\u8fd8\u662f\u4ee5\u56fe\u50cf\u5f62\u5f0f\u6e32\u67d3\u7684\u6587\u672c\uff08\u5982\u5c4f\u5e55\u622a\u56fe\u6216\u7167\u7247\uff09\uff0c\u90fd\u80fd\u63d0\u4f9b\u4e00\u81f4\u7684\u54cd\u5e94\u3002\u8fd9\u4e5f\u6709\u52a9\u4e8e\u589e\u5f3a\u6a21\u578b\u5728\u5904\u7406\u6587\u672c\u7e41\u91cd\u56fe\u50cf\u65f6\u7684\u80fd\u529b\u3002<\/li>\n<\/ol>\n<p>\u6bcf\u79cd\u7c7b\u578b\u7684\u6570\u636e\u5bf9\u4e8e\u6784\u5efa\u4e00\u4e2a\u5168\u9762\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u81f3\u5173\u91cd\u8981\uff0c\u8be5\u6a21\u578b\u53ef\u4ee5\u6709\u6548\u5730\u7ba1\u7406\u5404\u79cd\u771f\u5b9e\u4e16\u754c\u7684\u5e94\u7528\uff0c\u540c\u65f6\u786e\u4fdd\u5728\u5404\u79cd\u8f93\u5165\u6a21\u5f0f\u4e0b\u4fdd\u6301\u4e00\u81f4\u7684\u6027\u80fd\u3002<\/p>\n<h3>2.4 Long2short\uff1a\u7528\u4e8e Short-CoT \u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u538b\u7f29<\/h3>\n<p>\u5c3d\u7ba1 long-CoT \u6a21\u578b\u5b9e\u73b0\u4e86\u5f3a\u5927\u7684\u6027\u80fd\uff0c\u4f46\u4e0e\u6807\u51c6\u7684 short-CoT LLM \u76f8\u6bd4\uff0c\u5b83\u6d88\u8017\u4e86\u66f4\u591a\u7684\u6d4b\u8bd5\u65f6 Token\u3002\u4f46\u662f\uff0c\u53ef\u4ee5\u5c06\u5728 long-CoT \u6a21\u578b\u4e2d\u7684\u601d\u7ef4\u5148\u9a8c\u77e5\u8bc6\u8f6c\u79fb\u5230 short-CoT \u6a21\u578b\u4e2d\uff0c\u4ee5\u4fbf\u5373\u4f7f\u5728\u6709\u9650\u7684\u6d4b\u8bd5\u65f6 Token \u9884\u7b97\u4e0b\u4e5f\u53ef\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002\u6211\u4eec\u9488\u5bf9\u6b64 long2short \u95ee\u9898\u63d0\u51fa\u4e86\u51e0\u79cd\u65b9\u6cd5\uff0c\u5305\u62ec\u6a21\u578b\u5408\u5e76 (Yang et al. 2024)\u3001\u6700\u77ed\u62d2\u7edd\u91c7\u6837\u3001DPO (Rafailov et al. 2024) \u548c long2short RL\u3002\u4ee5\u4e0b\u63d0\u4f9b\u4e86\u5bf9\u8fd9\u4e9b\u65b9\u6cd5\u7684\u8be6\u7ec6\u63cf\u8ff0\uff1a<\/p>\n<p><strong>\u6a21\u578b\u5408\u5e76<\/strong>\u00a0\u6a21\u578b\u5408\u5e76\u5df2\u88ab\u53d1\u73b0\u6709\u52a9\u4e8e\u4fdd\u6301\u6cdb\u5316\u80fd\u529b\u3002\u6211\u4eec\u8fd8\u53d1\u73b0\u5b83\u5728\u5408\u5e76 long-CoT \u6a21\u578b\u548c short-CoT \u6a21\u578b\u65f6\u63d0\u9ad8\u4e86 Token \u6548\u7387\u3002\u8fd9\u79cd\u65b9\u6cd5\u5c06 long-CoT \u6a21\u578b\u4e0e\u8f83\u77ed\u7684\u6a21\u578b\u76f8\u7ed3\u5408\uff0c\u65e0\u9700\u8bad\u7ec3\u5373\u53ef\u83b7\u5f97\u65b0\u6a21\u578b\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u901a\u8fc7\u7b80\u5355\u5730\u5e73\u5747\u5b83\u4eec\u7684\u6743\u91cd\u6765\u5408\u5e76\u8fd9\u4e24\u4e2a\u6a21\u578b\u3002<\/p>\n<p><strong>\u6700\u77ed\u62d2\u7edd\u91c7\u6837<\/strong>\u00a0\u6211\u4eec\u89c2\u5bdf\u5230\u6211\u4eec\u7684\u6a21\u578b\u4e3a\u540c\u4e00\u4e2a\u95ee\u9898\u751f\u6210\u957f\u5ea6\u5dee\u5f02\u5f88\u5927\u7684\u54cd\u5e94\u3002\u57fa\u4e8e\u6b64\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u6700\u77ed\u62d2\u7edd\u91c7\u6837\u65b9\u6cd5\u3002\u8fd9\u79cd\u65b9\u6cd5\u5bf9\u540c\u4e00\u4e2a\u95ee\u9898\u91c7\u6837 n \u6b21\uff08\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0cn = 8\uff09\uff0c\u5e76\u9009\u62e9\u6700\u77ed\u7684\u6b63\u786e\u54cd\u5e94\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\u3002<\/p>\n<p><strong>DPO<\/strong>\u00a0\u4e0e\u6700\u77ed\u62d2\u7edd\u91c7\u6837\u7c7b\u4f3c\uff0c\u6211\u4eec\u5229\u7528 Long CoT \u6a21\u578b\u6765\u751f\u6210\u591a\u4e2a\u54cd\u5e94\u6837\u672c\u3002\u9009\u62e9\u6700\u77ed\u7684\u6b63\u786e\u89e3\u51b3\u65b9\u6848\u4f5c\u4e3a\u6b63\u6837\u672c\uff0c\u800c\u8f83\u957f\u7684\u54cd\u5e94\u5219\u88ab\u89c6\u4e3a\u8d1f\u6837\u672c\uff0c\u5305\u62ec\u9519\u8bef\u8f83\u957f\u7684\u54cd\u5e94\u548c\u6b63\u786e\u8f83\u957f\u7684\u54cd\u5e94\uff08\u6bd4\u9009\u62e9\u7684\u6b63\u6837\u672c\u957f 1.5 \u500d\uff09\u3002\u8fd9\u4e9b\u6b63\u8d1f\u5bf9\u6784\u6210\u4e86\u7528\u4e8e DPO \u8bad\u7ec3\u7684\u6210\u5bf9\u504f\u597d\u6570\u636e\u3002<\/p>\n<p><strong>Long2short RL<\/strong>\u00a0\u5728\u6807\u51c6\u7684 RL \u8bad\u7ec3\u9636\u6bb5\u4e4b\u540e\uff0c\u6211\u4eec\u9009\u62e9\u4e00\u4e2a\u5728\u6027\u80fd\u548c Token \u6548\u7387\u4e4b\u95f4\u63d0\u4f9b\u6700\u4f73\u5e73\u8861\u7684\u6a21\u578b\u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\uff0c\u5e76\u8fdb\u884c\u5355\u72ec\u7684 long2short RL \u8bad\u7ec3\u9636\u6bb5\u3002\u5728\u7b2c\u4e8c\u9636\u6bb5\uff0c\u6211\u4eec\u5e94\u7528\u7b2c 2.3.3 \u8282\u4e2d\u4ecb\u7ecd\u7684\u957f\u5ea6\u60e9\u7f5a\uff0c\u5e76\u663e\u7740\u51cf\u5c11\u6700\u5927 rollout \u957f\u5ea6\uff0c\u4ee5\u8fdb\u4e00\u6b65\u60e9\u7f5a\u8d85\u51fa\u6240\u9700\u957f\u5ea6\u4f46\u53ef\u80fd\u6b63\u786e\u7684\u54cd\u5e94\u3002<\/p>\n<h3>2.5 \u5176\u4ed6\u8bad\u7ec3\u8be6\u60c5<\/h3>\n<h4>2.5.1 \u9884\u8bad\u7ec3<\/h4>\n<p>Kimi k1.5 \u57fa\u7840\u6a21\u578b\u662f\u5728\u4e00\u4e2a\u591a\u6837\u5316\u3001\u9ad8\u8d28\u91cf\u7684\u591a\u6a21\u6001\u8bed\u6599\u5e93\u4e0a\u8bad\u7ec3\u7684\u3002\u8bed\u8a00\u6570\u636e\u6db5\u76d6\u4e94\u4e2a\u9886\u57df\uff1a\u82f1\u8bed\u3001\u4e2d\u6587\u3001\u4ee3\u7801\u3001\u6570\u5b66\u63a8\u7406\u548c\u77e5\u8bc6\u3002\u591a\u6a21\u6001\u6570\u636e\uff0c\u5305\u62ec\u5b57\u5e55\u3001\u56fe\u50cf\u6587\u672c\u4ea4\u9519\u3001OCR\u3001\u77e5\u8bc6\u548c QA \u6570\u636e\u96c6\uff0c\u4f7f\u6211\u4eec\u7684\u6a21\u578b\u80fd\u591f\u83b7\u5f97\u89c6\u89c9\u8bed\u8a00\u80fd\u529b\u3002\u4e25\u683c\u7684\u8d28\u91cf\u63a7\u5236\u786e\u4fdd\u4e86\u6574\u4f53\u9884\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u76f8\u5173\u6027\u3001\u591a\u6837\u6027\u548c\u5e73\u8861\u6027\u3002\u6211\u4eec\u7684\u9884\u8bad\u7ec3\u5206\u4e09\u4e2a\u9636\u6bb5\u8fdb\u884c\uff1a(1) \u89c6\u89c9\u8bed\u8a00\u9884\u8bad\u7ec3\uff0c\u5176\u4e2d\u5efa\u7acb\u4e86\u5f3a\u5927\u7684\u8bed\u8a00\u57fa\u7840\uff0c\u7136\u540e\u9010\u6b65\u8fdb\u884c\u591a\u6a21\u6001\u96c6\u6210\uff1b(2) \u51b7\u9759\u671f\uff0c\u5229\u7528\u7cbe\u9009\u548c\u5408\u6210\u6570\u636e\u5de9\u56fa\u80fd\u529b\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u63a8\u7406\u548c\u57fa\u4e8e\u77e5\u8bc6\u7684\u4efb\u52a1\uff1b\u4ee5\u53ca (3) \u957f\u4e0a\u4e0b\u6587\u6fc0\u6d3b\uff0c\u5c06\u5e8f\u5217\u5904\u7406\u6269\u5c55\u5230 131,072 \u4e2a Token\u3002\u6709\u5173\u6211\u4eec\u9884\u8bad\u7ec3\u5de5\u4f5c\u7684\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u89c1\u9644\u5f55 B\u3002<\/p>\n<h4>2.5.2 Vanilla \u76d1\u7763\u5fae\u8c03<\/h4>\n<p>\u6211\u4eec\u521b\u5efa\u4e86\u6db5\u76d6\u591a\u4e2a\u9886\u57df\u7684 vanilla SFT \u8bed\u6599\u5e93\u3002\u5bf9\u4e8e\u975e\u63a8\u7406\u4efb\u52a1\uff0c\u5305\u62ec\u95ee\u7b54\u3001\u5199\u4f5c\u548c\u6587\u672c\u5904\u7406\uff0c\u6211\u4eec\u6700\u521d\u901a\u8fc7\u4eba\u5de5\u6ce8\u91ca\u6784\u5efa\u79cd\u5b50\u6570\u636e\u96c6\u3002\u6b64\u79cd\u5b50\u6570\u636e\u96c6\u7528\u4e8e\u8bad\u7ec3\u79cd\u5b50\u6a21\u578b\u3002\u968f\u540e\uff0c\u6211\u4eec\u6536\u96c6\u5404\u79cd\u63d0\u793a\uff0c\u5e76\u4f7f\u7528\u79cd\u5b50\u6a21\u578b\u4e3a\u6bcf\u4e2a\u63d0\u793a\u751f\u6210\u591a\u4e2a\u54cd\u5e94\u3002\u7136\u540e\uff0c\u6ce8\u91ca\u8005\u5bf9\u8fd9\u4e9b\u54cd\u5e94\u8fdb\u884c\u6392\u540d\uff0c\u5e76\u6539\u8fdb\u6392\u540d\u6700\u9ad8\u7684\u54cd\u5e94\u4ee5\u751f\u6210\u6700\u7ec8\u7248\u672c\u3002\u5bf9\u4e8e\u8bf8\u5982\u6570\u5b66\u548c\u7f16\u7801\u95ee\u9898\u4e4b\u7c7b\u7684\u63a8\u7406\u4efb\u52a1\uff0c\u5176\u4e2d\u57fa\u4e8e\u89c4\u5219\u548c\u5956\u52b1\u5efa\u6a21\u7684\u9a8c\u8bc1\u6bd4\u4eba\u5de5\u5224\u65ad\u66f4\u51c6\u786e\u3001\u66f4\u9ad8\u6548\uff0c\u6211\u4eec\u5229\u7528\u62d2\u7edd\u91c7\u6837\u6765\u6269\u5c55 SFT \u6570\u636e\u96c6\u3002<\/p>\n<p>\u6211\u4eec\u7684 vanilla SFT \u6570\u636e\u96c6\u5305\u542b\u5927\u7ea6 100 \u4e07\u4e2a\u6587\u672c\u793a\u4f8b\u3002\u5177\u4f53\u6765\u8bf4\uff0c50 \u4e07\u4e2a\u793a\u4f8b\u7528\u4e8e\u4e00\u822c\u95ee\u7b54\uff0c20 \u4e07\u4e2a\u7528\u4e8e\u7f16\u7801\uff0c20 \u4e07\u4e2a\u7528\u4e8e\u6570\u5b66\u548c\u79d1\u5b66\uff0c5 \u5343\u4e2a\u7528\u4e8e\u521b\u610f\u5199\u4f5c\uff0c\u4ee5\u53ca 2 \u4e07\u4e2a\u7528\u4e8e\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\uff0c\u4f8b\u5982\u6458\u8981\u3001\u6587\u6863\u95ee\u7b54\u3001\u7ffb\u8bd1\u548c\u5199\u4f5c\u3002\u6b64\u5916\uff0c\u6211\u4eec\u6784\u5efa\u4e86 100 \u4e07\u4e2a\u6587\u672c\u89c6\u89c9\u793a\u4f8b\uff0c\u6db5\u76d6\u4e86\u5404\u79cd\u7c7b\u522b\uff0c\u5305\u62ec\u56fe\u8868\u89e3\u91ca\u3001OCR\u3001\u57fa\u4e8e\u56fe\u50cf\u7684\u5bf9\u8bdd\u3001\u89c6\u89c9\u7f16\u7801\u3001\u89c6\u89c9\u63a8\u7406\u4ee5\u53ca\u5177\u6709\u89c6\u89c9\u8f85\u52a9\u7684\u6570\u5b66\/\u79d1\u5b66\u95ee\u9898\u3002<\/p>\n<p>\u6211\u4eec\u9996\u5148\u5728 32k \u4e2a Token \u7684\u5e8f\u5217\u957f\u5ea6\u4e0b\u8bad\u7ec3\u6a21\u578b 1 \u4e2a epoch\uff0c\u7136\u540e\u5728 128k \u4e2a Token \u7684\u5e8f\u5217\u957f\u5ea6\u4e0b\u518d\u8bad\u7ec3 1 \u4e2a epoch\u3002\u5728\u7b2c\u4e00\u9636\u6bb5 (32k)\uff0c\u5b66\u4e60\u7387\u4ece 2 \u00d7 10-5 \u8870\u51cf\u5230 2 \u00d7 10-6\uff0c\u7136\u540e\u5728\u7b2c\u4e8c\u9636\u6bb5 (128k) \u91cd\u65b0\u5347\u6e29\u5230 1 \u00d7 10-5\uff0c\u6700\u540e\u8870\u51cf\u5230 1 \u00d7 10-6\u3002\u4e3a\u4e86\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\uff0c\u6211\u4eec\u5c06\u591a\u4e2a\u8bad\u7ec3\u793a\u4f8b\u6253\u5305\u5230\u6bcf\u4e2a\u8bad\u7ec3\u5e8f\u5217\u4e2d\u3002<\/p>\n<h4>2.6 RL \u57fa\u7840\u8bbe\u65bd<\/h4>\n<div id=\"attachment_19061\" style=\"width: 1125px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19061\" class=\"wp-image-19061 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c48ce5acc850f4.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1115\" height=\"481\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c48ce5acc850f4.png 1115w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c48ce5acc850f4-300x129.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c48ce5acc850f4-1024x442.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c48ce5acc850f4-768x331.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5c48ce5acc850f4-18x8.png 18w\" sizes=\"auto, (max-width: 1115px) 100vw, 1115px\" \/><p id=\"caption-attachment-19061\" class=\"wp-caption-text\">\u56fe 3\uff1a\u7528\u4e8e LLM \u7684\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u7cfb\u7edf<\/p><\/div>\n<p>&nbsp;<\/p>\n<h5>(a) \u7cfb\u7edf\u6982\u8ff0<\/h5>\n<h5>(b) \u90e8\u5206 rollout<\/h5>\n<h5>2.6.1 \u7528\u4e8e LLM \u7684\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u7cfb\u7edf<\/h5>\n<p>\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u5f3a\u5316\u5b66\u4e60 (RL) \u5df2\u7ecf\u6210\u4e3a\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u4e00\u79cd\u5173\u952e\u8bad\u7ec3\u65b9\u6cd5 (Ouyang et al. 2022) (Jaech et al. 2024)\uff0c\u5b83\u4ece\u5176\u5728\u638c\u63e1\u590d\u6742\u6e38\u620f\uff08\u5982\u56f4\u68cb\u3001\u661f\u9645\u4e89\u9738 II \u548c Dota 2\uff09\u65b9\u9762\u7684\u6210\u529f\u4e2d\u6c72\u53d6\u7075\u611f\uff0c\u4f8b\u5982 AlphaGo (Silver et al. 2017)\u3001AlphaStar (Vinyals et al. 2019) \u548c OpenAI Dota Five (Berner et al. 2019) \u7b49\u7cfb\u7edf\u3002\u9075\u5faa\u8fd9\u4e00\u4f20\u7edf\uff0cKimi k1.5 \u7cfb\u7edf\u91c7\u7528\u4e86\u8fed\u4ee3\u540c\u6b65 RL \u6846\u67b6\uff0c\u8be5\u6846\u67b6\u7ecf\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\uff0c\u53ef\u901a\u8fc7\u6301\u4e45\u5b66\u4e60\u548c\u9002\u5e94\u6765\u589e\u5f3a\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002\u6b64\u7cfb\u7edf\u4e2d\u7684\u4e00\u9879\u5173\u952e\u521b\u65b0\u662f\u5f15\u5165\u4e86\u90e8\u5206 rollout \u6280\u672f\uff0c\u8be5\u6280\u672f\u65e8\u5728\u4f18\u5316\u590d\u6742\u63a8\u7406\u8f68\u8ff9\u7684\u5904\u7406\u3002<\/p>\n<p>\u5982\u56fe 3a \u6240\u793a\uff0cRL \u8bad\u7ec3\u7cfb\u7edf\u901a\u8fc7\u8fed\u4ee3\u540c\u6b65\u65b9\u6cd5\u8fd0\u884c\uff0c\u6bcf\u6b21\u8fed\u4ee3\u90fd\u5305\u542b\u4e00\u4e2a rollout \u9636\u6bb5\u548c\u4e00\u4e2a\u8bad\u7ec3\u9636\u6bb5\u3002\u5728 rollout \u9636\u6bb5\uff0c\u7531\u4e2d\u592e\u4e3b\u670d\u52a1\u5668\u534f\u8c03\u7684 rollout \u5de5\u4f5c\u7a0b\u5e8f\u901a\u8fc7\u4e0e\u6a21\u578b\u4ea4\u4e92\u6765\u751f\u6210 rollout \u8f68\u8ff9\uff0c\u4ece\u800c\u4ea7\u751f\u9488\u5bf9\u5404\u79cd\u8f93\u5165\u7684\u54cd\u5e94\u5e8f\u5217\u3002\u7136\u540e\uff0c\u8fd9\u4e9b\u8f68\u8ff9\u5b58\u50a8\u5728\u91cd\u653e\u7f13\u51b2\u533a\u4e2d\uff0c\u8fd9\u786e\u4fdd\u4e86\u4e00\u4e2a\u591a\u6837\u5316\u4e14\u65e0\u504f\u5dee\u7684\u8bad\u7ec3\u6570\u636e\u96c6\uff0c\u65b9\u6cd5\u662f\u4e2d\u65ad\u65f6\u95f4\u76f8\u5173\u6027\u3002\u5728\u968f\u540e\u7684\u8bad\u7ec3\u9636\u6bb5\uff0c\u8bad\u7ec3\u5de5\u4f5c\u7a0b\u5e8f\u4f1a\u8bbf\u95ee\u8fd9\u4e9b\u7ecf\u9a8c\u4ee5\u66f4\u65b0\u6a21\u578b\u7684\u6743\u91cd\u3002\u6b64\u5faa\u73af\u8fc7\u7a0b\u5141\u8bb8\u6a21\u578b\u4ece\u5176\u64cd\u4f5c\u4e2d\u4e0d\u65ad\u5b66\u4e60\uff0c\u5e76\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\u8c03\u6574\u5176\u7b56\u7565\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002<\/p>\n<p>\u4e2d\u592e\u4e3b\u670d\u52a1\u5668\u5145\u5f53\u4e2d\u592e\u6307\u6325\uff0c\u7ba1\u7406 rollout \u5de5\u4f5c\u7a0b\u5e8f\u3001\u8bad\u7ec3\u5de5\u4f5c\u7a0b\u5e8f\u3001\u5177\u6709\u5956\u52b1\u6a21\u578b\u7684\u8bc4\u4f30\u548c\u91cd\u653e\u7f13\u51b2\u533a\u4e4b\u95f4\u7684\u6570\u636e\u6d41\u548c\u901a\u4fe1\u3002\u5b83\u786e\u4fdd\u7cfb\u7edf\u534f\u8c03\u8fd0\u884c\uff0c\u5e73\u8861\u8d1f\u8f7d\u5e76\u4fc3\u8fdb\u9ad8\u6548\u7684\u6570\u636e\u5904\u7406\u3002<\/p>\n<p>\u8bad\u7ec3\u5de5\u4f5c\u7a0b\u5e8f\u8bbf\u95ee\u8fd9\u4e9b rollout \u8f68\u8ff9\uff0c\u65e0\u8bba\u662f\u5728\u5355\u4e2a\u8fed\u4ee3\u4e2d\u5b8c\u6210\u8fd8\u662f\u5728\u591a\u4e2a\u8fed\u4ee3\u4e2d\u5212\u5206\uff0c\u4ee5\u8ba1\u7b97\u68af\u5ea6\u66f4\u65b0\uff0c\u4ece\u800c\u6539\u8fdb\u6a21\u578b\u7684\u53c2\u6570\u5e76\u589e\u5f3a\u5176\u6027\u80fd\u3002\u6b64\u8fc7\u7a0b\u7531\u5956\u52b1\u6a21\u578b\u76d1\u7763\uff0c\u8be5\u6a21\u578b\u8bc4\u4f30\u6a21\u578b\u8f93\u51fa\u7684\u8d28\u91cf\u5e76\u63d0\u4f9b\u5fc5\u8981\u7684\u53cd\u9988\u6765\u6307\u5bfc\u8bad\u7ec3\u8fc7\u7a0b\u3002\u5956\u52b1\u6a21\u578b\u7684\u8bc4\u4f30\u5bf9\u4e8e\u786e\u5b9a\u6a21\u578b\u7b56\u7565\u7684\u6709\u6548\u6027\u548c\u5f15\u5bfc\u6a21\u578b\u8d70\u5411\u6700\u4f73\u6027\u80fd\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u8be5\u7cfb\u7edf\u8fd8\u5305\u542b\u4ee3\u7801\u6267\u884c\u670d\u52a1\uff0c\u8be5\u670d\u52a1\u4e13\u95e8\u8bbe\u8ba1\u7528\u4e8e\u5904\u7406\u4ee3\u7801\u76f8\u5173\u7684\u95ee\u9898\uff0c\u5e76\u4e14\u662f\u5956\u52b1\u6a21\u578b\u7684\u7ec4\u6210\u90e8\u5206\u3002\u6b64\u670d\u52a1\u5728\u5b9e\u9645\u7f16\u7801\u573a\u666f\u4e2d\u8bc4\u4f30\u6a21\u578b\u7684\u8f93\u51fa\uff0c\u786e\u4fdd\u6a21\u578b\u7684\u5b66\u4e60\u4e0e\u771f\u5b9e\u4e16\u754c\u7684\u7f16\u7a0b\u6311\u6218\u7d27\u5bc6\u7ed3\u5408\u3002\u901a\u8fc7\u9488\u5bf9\u5b9e\u9645\u4ee3\u7801\u6267\u884c\u9a8c\u8bc1\u6a21\u578b\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u6b64\u53cd\u9988\u5faa\u73af\u5bf9\u4e8e\u6539\u8fdb\u6a21\u578b\u7684\u7b56\u7565\u548c\u589e\u5f3a\u5176\u5728\u4ee3\u7801\u76f8\u5173\u4efb\u52a1\u4e2d\u7684\u6027\u80fd\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<h5>2.6.2 \u7528\u4e8e Long CoT RL \u7684\u90e8\u5206 Rollout<\/h5>\n<p>\u6211\u4eec\u5de5\u4f5c\u7684\u4e3b\u8981\u601d\u60f3\u4e4b\u4e00\u662f\u6269\u5c55\u957f\u4e0a\u4e0b\u6587 RL \u8bad\u7ec3\u3002\u90e8\u5206 rollout \u662f\u4e00\u79cd\u5173\u952e\u6280\u672f\uff0c\u5b83\u901a\u8fc7\u7ba1\u7406\u957f\u8f68\u8ff9\u548c\u77ed\u8f68\u8ff9\u7684 rollout\uff0c\u6709\u6548\u5730\u89e3\u51b3\u4e86\u5904\u7406 long-CoT \u7279\u6027\u7684\u6311\u6218\u3002\u6b64\u6280\u672f\u5efa\u7acb\u4e86\u56fa\u5b9a\u7684\u8f93\u51fa Token \u9884\u7b97\uff0c\u9650\u5236\u6bcf\u4e2a rollout \u8f68\u8ff9\u7684\u957f\u5ea6\u3002\u5982\u679c\u8f68\u8ff9\u5728 rollout \u9636\u6bb5\u8d85\u51fa Token \u9650\u5236\uff0c\u5219\u5c06\u672a\u5b8c\u6210\u7684\u90e8\u5206\u4fdd\u5b58\u5230\u91cd\u653e\u7f13\u51b2\u533a\u4e2d\uff0c\u5e76\u5728\u4e0b\u4e00\u6b21\u8fed\u4ee3\u4e2d\u7ee7\u7eed\u3002\u5b83\u786e\u4fdd\u6ca1\u6709\u5355\u4e2a\u5197\u957f\u7684\u8f68\u8ff9\u5784\u65ad\u7cfb\u7edf\u7684\u8d44\u6e90\u3002\u6b64\u5916\uff0c\u7531\u4e8e rollout \u5de5\u4f5c\u7a0b\u5e8f\u662f\u5f02\u6b65\u8fd0\u884c\u7684\uff0c\u5f53\u67d0\u4e9b\u5de5\u4f5c\u7a0b\u5e8f\u6b63\u5728\u5904\u7406\u957f\u8f68\u8ff9\u65f6\uff0c\u5176\u4ed6\u5de5\u4f5c\u7a0b\u5e8f\u53ef\u4ee5\u72ec\u7acb\u5904\u7406\u65b0\u7684\u3001\u8f83\u77ed\u7684 rollout \u4efb\u52a1\u3002\u5f02\u6b65\u64cd\u4f5c\u901a\u8fc7\u786e\u4fdd\u6240\u6709 rollout \u5de5\u4f5c\u7a0b\u5e8f\u90fd\u79ef\u6781\u4e3a\u8bad\u7ec3\u8fc7\u7a0b\u505a\u51fa\u8d21\u732e\u6765\u6700\u5927\u5316\u8ba1\u7b97\u6548\u7387\uff0c\u4ece\u800c\u4f18\u5316\u7cfb\u7edf\u7684\u6574\u4f53\u6027\u80fd\u3002<\/p>\n<p>\u5982\u56fe 3b \u6240\u793a\uff0c\u90e8\u5206 rollout \u7cfb\u7edf\u901a\u8fc7\u5c06\u957f\u54cd\u5e94\u5206\u89e3\u4e3a\u8de8\u8fed\u4ee3\u7684\u7247\u6bb5\uff08\u4ece\u8fed\u4ee3 n-m \u5230\u8fed\u4ee3 n\uff09\u6765\u5de5\u4f5c\u3002\u91cd\u653e\u7f13\u51b2\u533a\u5145\u5f53\u4e2d\u592e\u5b58\u50a8\u673a\u5236\uff0c\u7528\u4e8e\u7ef4\u62a4\u8fd9\u4e9b\u54cd\u5e94\u7247\u6bb5\uff0c\u5176\u4e2d\u53ea\u6709\u5f53\u524d\u8fed\u4ee3\uff08\u8fed\u4ee3 n\uff09\u9700\u8981\u5728\u7ebf\u8ba1\u7b97\u3002\u4ee5\u524d\u7684\u7247\u6bb5\uff08\u8fed\u4ee3 n-m \u5230 n-1\uff09\u53ef\u4ee5\u4ece\u7f13\u51b2\u533a\u4e2d\u6709\u6548\u5730\u91cd\u7528\uff0c\u4ece\u800c\u65e0\u9700\u91cd\u590d rollout\u3002\u8fd9\u79cd\u5206\u6bb5\u65b9\u6cd5\u663e\u7740\u51cf\u5c11\u4e86\u8ba1\u7b97\u5f00\u9500\uff1a\u7cfb\u7edf\u4e0d\u662f\u4e00\u6b21\u6027 rollout \u6574\u4e2a\u54cd\u5e94\uff0c\u800c\u662f\u4ee5\u589e\u91cf\u65b9\u5f0f\u5904\u7406\u548c\u5b58\u50a8\u7247\u6bb5\uff0c\u4ece\u800c\u53ef\u4ee5\u5728\u4fdd\u6301\u5feb\u901f\u8fed\u4ee3\u65f6\u95f4\u7684\u540c\u65f6\u751f\u6210\u66f4\u957f\u7684\u54cd\u5e94\u3002\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u67d0\u4e9b\u7247\u6bb5\u53ef\u4ee5\u4ece\u635f\u5931\u8ba1\u7b97\u4e2d\u6392\u9664\uff0c\u4ee5\u8fdb\u4e00\u6b65\u4f18\u5316\u5b66\u4e60\u8fc7\u7a0b\uff0c\u4ece\u800c\u4f7f\u6574\u4e2a\u7cfb\u7edf\u65e2\u9ad8\u6548\u53c8\u53ef\u6269\u5c55\u3002<\/p>\n<p>\u90e8\u5206 rollout \u7684\u5b9e\u73b0\u8fd8\u63d0\u4f9b\u4e86\u91cd\u590d\u68c0\u6d4b\u3002\u8be5\u7cfb\u7edf\u4f1a\u8bc6\u522b\u751f\u6210\u5185\u5bb9\u4e2d\u91cd\u590d\u7684\u5e8f\u5217\uff0c\u5e76\u63d0\u524d\u7ec8\u6b62\u5b83\u4eec\uff0c\u4ece\u800c\u5728\u4fdd\u6301\u8f93\u51fa\u8d28\u91cf\u7684\u540c\u65f6\u51cf\u5c11\u4e0d\u5fc5\u8981\u7684\u8ba1\u7b97\u3002\u68c0\u6d4b\u5230\u7684\u91cd\u590d\u9879\u53ef\u80fd\u4f1a\u88ab\u5206\u914d\u989d\u5916\u7684\u60e9\u7f5a\uff0c\u4ece\u800c\u6709\u6548\u5730\u963b\u6b62\u63d0\u793a\u96c6\u4e2d\u51fa\u73b0\u5197\u4f59\u7684\u5185\u5bb9\u751f\u6210\u3002<\/p>\n<h5>2.6.3 \u8bad\u7ec3\u548c\u63a8\u7406\u7684\u6df7\u5408\u90e8\u7f72<\/h5>\n<p>RL \u8bad\u7ec3\u8fc7\u7a0b\u5305\u62ec\u4ee5\u4e0b\u9636\u6bb5\uff1a<\/p>\n<div id=\"attachment_19062\" style=\"width: 1091px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19062\" class=\"wp-image-19062 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5155dc14861b074.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1081\" height=\"604\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5155dc14861b074.png 1081w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5155dc14861b074-300x168.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5155dc14861b074-1024x572.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5155dc14861b074-768x429.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5155dc14861b074-18x10.png 18w\" sizes=\"auto, (max-width: 1081px) 100vw, 1081px\" \/><p id=\"caption-attachment-19062\" class=\"wp-caption-text\">\u56fe 4\uff1a\u6df7\u5408\u90e8\u7f72\u6846\u67b6<\/p><\/div>\n<p>&nbsp;<\/p>\n<ul>\n<li><strong>\u8bad\u7ec3\u9636\u6bb5<\/strong>\uff1a\u9996\u5148\uff0cMegatron (Shoeybi et al. 2020) \u548c <a href=\"https:\/\/www.kdjingpai.com\/ja\/vllm\/\">vLLM<\/a> (Kwon et al. 2023) \u5728\u5355\u72ec\u7684\u5bb9\u5668\u4e2d\u6267\u884c\uff0c\u8fd9\u4e9b\u5bb9\u5668\u7531\u79f0\u4e3a\u68c0\u67e5\u70b9\u5f15\u64ce\uff08\u7b2c 2.6.3 \u8282\uff09\u7684 shim \u8fdb\u7a0b\u5c01\u88c5\u3002Megatron \u5f00\u59cb\u8bad\u7ec3\u8fc7\u7a0b\u3002\u8bad\u7ec3\u5b8c\u6210\u540e\uff0cMegatron \u4f1a\u5378\u8f7d GPU \u5185\u5b58\uff0c\u5e76\u51c6\u5907\u5c06\u5f53\u524d\u6743\u91cd\u4f20\u8f93\u5230 vLLM\u3002<\/li>\n<li><strong>\u63a8\u7406\u9636\u6bb5<\/strong>\uff1a\u5728 Megatron \u5378\u8f7d\u540e\uff0cvLLM \u4ece\u865a\u62df\u6a21\u578b\u6743\u91cd\u5f00\u59cb\uff0c\u5e76\u4f7f\u7528\u901a\u8fc7 Mooncake (Qin et al. 2024) \u4ece Megatron \u4f20\u8f93\u7684\u6700\u65b0\u6743\u91cd\u66f4\u65b0\u5b83\u4eec\u3002\u5728 rollout \u5b8c\u6210\u540e\uff0c\u68c0\u67e5\u70b9\u5f15\u64ce\u5c06\u505c\u6b62\u6240\u6709 vLLM \u8fdb\u7a0b\u3002<\/li>\n<li><strong>\u540e\u7eed\u8bad\u7ec3\u9636\u6bb5<\/strong>\uff1a\u4e00\u65e6\u5206\u914d\u7ed9 vLLM \u7684\u5185\u5b58\u88ab\u91ca\u653e\uff0cMegatron \u5c31\u4f1a\u5378\u8f7d\u5185\u5b58\u5e76\u542f\u52a8\u53e6\u4e00\u8f6e\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<p>\u6211\u4eec\u53d1\u73b0\u73b0\u6709\u5de5\u4f5c\u96be\u4ee5\u540c\u65f6\u652f\u6301\u4ee5\u4e0b\u6240\u6709\u7279\u6027\u3002<\/p>\n<ul>\n<li><strong>\u590d\u6742\u7684\u5e76\u884c\u7b56\u7565<\/strong>\uff1aMegatron \u53ef\u80fd\u5177\u6709\u4e0e vLLM \u4e0d\u540c\u7684\u5e76\u884c\u7b56\u7565\u3002\u5728 Megatron \u7684\u591a\u4e2a\u8282\u70b9\u4e2d\u5206\u914d\u7684\u8bad\u7ec3\u6743\u91cd\u53ef\u80fd\u96be\u4ee5\u4e0e vLLM \u5171\u4eab\u3002<\/li>\n<li><strong>\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u7a7a\u95f2 GPU \u8d44\u6e90<\/strong>\uff1a\u5bf9\u4e8e\u5728\u7ebf\u7b56\u7565 RL\uff0c\u6700\u8fd1\u7684\u5de5\u4f5c\uff08\u5982 SGLang (L. Zheng et al. 2024) \u548c VLLM\uff09\u53ef\u80fd\u4f1a\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4fdd\u7559\u4e00\u4e9b GPU\uff0c\u8fd9\u53cd\u8fc7\u6765\u53ef\u80fd\u4f1a\u5bfc\u81f4\u7a7a\u95f2\u8bad\u7ec3 GPU\u3002\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u4e4b\u95f4\u5171\u4eab\u76f8\u540c\u7684\u8bbe\u5907\u4f1a\u66f4\u6709\u6548\u7387\u3002<\/li>\n<li><strong>\u52a8\u6001\u6269\u5c55\u80fd\u529b<\/strong>\uff1a\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u53ef\u4ee5\u901a\u8fc7\u589e\u52a0\u63a8\u7406\u8282\u70b9\u7684\u6570\u91cf\uff0c\u540c\u65f6\u4fdd\u6301\u8bad\u7ec3\u8fc7\u7a0b\u4e0d\u53d8\u6765\u5b9e\u73b0\u663e\u7740\u52a0\u901f\u3002\u6211\u4eec\u7684\u7cfb\u7edf\u53ef\u4ee5\u6709\u6548\u5730\u5229\u7528\u6240\u9700\u7684\u7a7a\u95f2 GPU \u8282\u70b9\u3002<\/li>\n<\/ul>\n<p>\u5982\u56fe 4 \u6240\u793a\uff0c\u6211\u4eec\u5728 Megatron \u548c vLLM \u4e4b\u4e0a\u5b9e\u73b0\u4e86\u6b64\u6df7\u5408\u90e8\u7f72\u6846\u67b6\uff08\u7b2c 2.6.3 \u8282\uff09\uff0c\u4ece\u800c\u5b9e\u73b0\u4e86\u4ece\u8bad\u7ec3\u5230\u63a8\u7406\u9636\u6bb5\u4e0d\u5230\u4e00\u5206\u949f\u7684\u65f6\u95f4\uff0c\u53cd\u4e4b\u5219\u7ea6\u4e3a 10 \u79d2\u3002<\/p>\n<p><strong>\u6df7\u5408\u90e8\u7f72\u7b56\u7565<\/strong>\u00a0\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7528\u4e8e\u8bad\u7ec3\u548c\u63a8\u7406\u4efb\u52a1\u7684\u6df7\u5408\u90e8\u7f72\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u5229\u7528 Kubernetes Sidecar \u5bb9\u5668\u5171\u4eab\u6240\u6709\u53ef\u7528\u7684 GPU\uff0c\u4ee5\u5c06\u4e24\u79cd\u5de5\u4f5c\u8d1f\u8f7d\u5e76\u7f6e\u5728\u4e00\u4e2a pod \u4e2d\u3002\u6b64\u7b56\u7565\u7684\u4e3b\u8981\u4f18\u70b9\u662f\uff1a<\/p>\n<ul>\n<li><strong>\u5b83\u6709\u52a9\u4e8e\u9ad8\u6548\u7684\u8d44\u6e90\u5171\u4eab\u548c\u7ba1\u7406\uff0c\u9632\u6b62\u8bad\u7ec3\u8282\u70b9\u5728\u7b49\u5f85\u63a8\u7406\u8282\u70b9\u65f6\u95f2\u7f6e\uff08\u5f53\u4e24\u8005\u90e8\u7f72\u5728\u5355\u72ec\u7684\u8282\u70b9\u4e0a\u65f6\uff09\u3002<\/strong><\/li>\n<li><strong>\u5229\u7528\u4e0d\u540c\u7684\u90e8\u7f72\u6620\u50cf\uff0c\u8bad\u7ec3\u548c\u63a8\u7406\u53ef\u4ee5\u72ec\u7acb\u8fed\u4ee3\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002<\/strong><\/li>\n<li><strong>\u8be5\u67b6\u6784\u4e0d\u9650\u4e8e vLLM\uff0c\u5176\u4ed6\u6846\u67b6\u4e5f\u53ef\u4ee5\u65b9\u4fbf\u5730\u96c6\u6210\u3002<\/strong><\/li>\n<\/ul>\n<p><strong>\u68c0\u67e5\u70b9\u5f15\u64ce<\/strong>\u00a0\u68c0\u67e5\u70b9\u5f15\u64ce\u8d1f\u8d23\u7ba1\u7406 vLLM \u8fdb\u7a0b\u7684\u751f\u547d\u5468\u671f\uff0c\u516c\u5f00 HTTP API\uff0c\u8fd9\u4e9b API \u80fd\u591f\u89e6\u53d1 vLLM \u7684\u5404\u79cd\u64cd\u4f5c\u3002\u4e3a\u4e86\u4fdd\u8bc1\u6574\u4f53\u4e00\u81f4\u6027\u548c\u53ef\u9760\u6027\uff0c\u6211\u4eec\u5229\u7528\u7531 etcd \u670d\u52a1\u7ba1\u7406\u7684\u5168\u5c40\u5143\u6570\u636e\u7cfb\u7edf\u6765\u5e7f\u64ad\u64cd\u4f5c\u548c\u72b6\u6001\u3002<\/p>\n<p>\u5f53 vLLM \u5378\u8f7d\u65f6\uff0c\u5b83\u53ef\u80fd\u96be\u4ee5\u5b8c\u5168\u91ca\u653e GPU \u5185\u5b58\uff0c\u8fd9\u4e3b\u8981\u662f\u7531\u4e8e CUDA \u56fe\u3001NCCL \u7f13\u51b2\u533a\u548c NVIDIA \u9a71\u52a8\u7a0b\u5e8f\u3002\u4e3a\u4e86\u6700\u5927\u7a0b\u5ea6\u5730\u51cf\u5c11\u5bf9 vLLM \u7684\u4fee\u6539\uff0c\u6211\u4eec\u4f1a\u5728\u9700\u8981\u65f6\u7ec8\u6b62\u5e76\u91cd\u65b0\u542f\u52a8\u5b83\uff0c\u4ee5\u83b7\u5f97\u66f4\u597d\u7684 GPU \u5229\u7528\u7387\u548c\u5bb9\u9519\u80fd\u529b\u3002<\/p>\n<p>Megatron \u4e2d\u7684\u5de5\u4f5c\u7a0b\u5e8f\u5c06\u62e5\u6709\u7684\u68c0\u67e5\u70b9\u8f6c\u6362\u4e3a\u5171\u4eab\u5185\u5b58\u4e2d\u7684 Hugging Face \u683c\u5f0f\u3002\u6b64\u8f6c\u6362\u8fd8\u8003\u8651\u4e86\u7ba1\u9053\u5e76\u884c\u6027\u548c\u4e13\u5bb6\u5e76\u884c\u6027\uff0c\u4ee5\u4fbf\u5728\u8fd9\u4e9b\u68c0\u67e5\u70b9\u4e2d\u4ec5\u4fdd\u7559\u5f20\u91cf\u5e76\u884c\u6027\u3002\u5171\u4eab\u5185\u5b58\u4e2d\u7684\u68c0\u67e5\u70b9\u968f\u540e\u88ab\u5206\u6210\u788e\u7247\u5e76\u6ce8\u518c\u5230\u5168\u5c40\u5143\u6570\u636e\u7cfb\u7edf\u4e2d\u3002\u6211\u4eec\u4f7f\u7528 Mooncake \u901a\u8fc7 RDMA \u5728\u5bf9\u7b49\u8282\u70b9\u4e4b\u95f4\u4f20\u8f93\u68c0\u67e5\u70b9\u3002\u9700\u8981\u5bf9 vLLM \u8fdb\u884c\u4e00\u4e9b\u4fee\u6539\u624d\u80fd\u52a0\u8f7d\u6743\u91cd\u6587\u4ef6\u5e76\u6267\u884c\u5f20\u91cf\u5e76\u884c\u8f6c\u6362\u3002<\/p>\n<p>\u6211\u4eec\u5f00\u53d1\u4e86\u6c99\u7bb1\u4f5c\u4e3a\u6267\u884c\u7528\u6237\u63d0\u4ea4\u4ee3\u7801\u7684\u5b89\u5168\u73af\u5883\uff0c\u8be5\u73af\u5883\u9488\u5bf9\u4ee3\u7801\u6267\u884c\u548c\u4ee3\u7801\u57fa\u51c6\u8bc4\u4f30\u8fdb\u884c\u4e86\u4f18\u5316\u3002\u901a\u8fc7\u52a8\u6001\u5207\u6362\u5bb9\u5668\u6620\u50cf\uff0c\u6c99\u7bb1\u901a\u8fc7 MultiPL-E (Cassano, Gouwar, D. Nguyen, S. Nguyen, et al. 2023)\u3001DMOJ Judge Server\u00b2\u3001Lean\u3001Jupyter Notebook \u548c\u5176\u4ed6\u6620\u50cf\u652f\u6301\u4e0d\u540c\u7684\u7528\u4f8b\u3002<\/p>\n<p>\u5bf9\u4e8e\u7f16\u7801\u4efb\u52a1\u4e2d\u7684 RL\uff0c\u6c99\u7bb1\u901a\u8fc7\u63d0\u4f9b\u4e00\u81f4\u4e14\u53ef\u91cd\u590d\u7684\u8bc4\u4f30\u673a\u5236\u6765\u786e\u4fdd\u8bad\u7ec3\u6570\u636e\u5224\u65ad\u7684\u53ef\u9760\u6027\u3002\u5b83\u7684\u53cd\u9988\u7cfb\u7edf\u652f\u6301\u591a\u9636\u6bb5\u8bc4\u4f30\uff0c\u4f8b\u5982\u4ee3\u7801\u6267\u884c\u53cd\u9988\u548c repo \u7ea7\u7f16\u8f91\uff0c\u540c\u65f6\u4fdd\u6301\u7edf\u4e00\u7684\u4e0a\u4e0b\u6587\uff0c\u4ee5\u786e\u4fdd\u8de8\u7f16\u7a0b\u8bed\u8a00\u8fdb\u884c\u516c\u5e73\u548c\u516c\u6b63\u7684\u57fa\u51c6\u6bd4\u8f83\u3002<\/p>\n<p>\u6211\u4eec\u5c06\u8be5\u670d\u52a1\u90e8\u7f72\u5728 Kubernetes \u4e0a\u4ee5\u5b9e\u73b0\u53ef\u6269\u5c55\u6027\u548c\u5f39\u6027\uff0c\u901a\u8fc7 HTTP \u7aef\u70b9\u5c06\u5176\u516c\u5f00\u4ee5\u8fdb\u884c\u5916\u90e8\u96c6\u6210\u3002Kubernetes \u7684\u529f\u80fd\uff08\u5982\u81ea\u52a8\u91cd\u542f\u548c\u6eda\u52a8\u66f4\u65b0\uff09\u786e\u4fdd\u4e86\u53ef\u7528\u6027\u548c\u5bb9\u9519\u80fd\u529b\u3002<\/p>\n<p>\u4e3a\u4e86\u4f18\u5316\u6027\u80fd\u5e76\u652f\u6301 RL \u73af\u5883\uff0c\u6211\u4eec\u5c06\u51e0\u79cd\u6280\u672f\u96c6\u6210\u5230\u4ee3\u7801\u6267\u884c\u670d\u52a1\u4e2d\uff0c\u4ee5\u63d0\u9ad8\u6548\u7387\u3001\u901f\u5ea6\u548c\u53ef\u9760\u6027\u3002\u5176\u4e2d\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>\u4f7f\u7528 Crun<\/strong>\uff1a\u6211\u4eec\u4f7f\u7528 crun \u4f5c\u4e3a\u5bb9\u5668\u8fd0\u884c\u65f6\u800c\u4e0d\u662f Docker\uff0c\u4ece\u800c\u663e\u7740\u51cf\u5c11\u4e86\u5bb9\u5668\u542f\u52a8\u65f6\u95f4\u3002<\/li>\n<li><strong>Cgroup \u91cd\u7528<\/strong>\uff1a\u6211\u4eec\u9884\u5148\u521b\u5efa\u7528\u4e8e\u5bb9\u5668\u4f7f\u7528\u7684 cgroup\uff0c\u8fd9\u5728\u9ad8\u5e76\u53d1\u573a\u666f\u4e2d\u81f3\u5173\u91cd\u8981\uff0c\u5728\u8fd9\u4e9b\u573a\u666f\u4e2d\uff0c\u4e3a\u6bcf\u4e2a\u5bb9\u5668\u521b\u5efa\u548c\u9500\u6bc1 cgroup \u53ef\u80fd\u4f1a\u6210\u4e3a\u74f6\u9888\u3002<\/li>\n<li><strong>\u78c1\u76d8\u4f7f\u7528\u4f18\u5316<\/strong>\uff1a\u4f7f\u7528\u5b89\u88c5\u4e3a tmpfs \u7684\u4e0a\u5c42\u7684\u8986\u76d6\u6587\u4ef6\u7cfb\u7edf\u6765\u63a7\u5236\u78c1\u76d8\u5199\u5165\uff0c\u4ece\u800c\u63d0\u4f9b\u56fa\u5b9a\u5927\u5c0f\u7684\u9ad8\u901f\u5b58\u50a8\u7a7a\u95f4\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u5229\u4e8e\u4e34\u65f6\u5de5\u4f5c\u8d1f\u8f7d\u3002<\/li>\n<\/ul>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19063\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b90da930534e722.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1045\" height=\"134\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b90da930534e722.png 1045w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b90da930534e722-300x38.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b90da930534e722-1024x131.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b90da930534e722-768x98.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b90da930534e722-18x2.png 18w\" sizes=\"auto, (max-width: 1045px) 100vw, 1045px\" \/><\/p>\n<p>\u8fd9\u4e9b\u4f18\u5316\u63d0\u9ad8\u4e86\u4ee3\u7801\u6267\u884c\u4e2d\u7684 RL \u6548\u7387\uff0c\u4ece\u800c\u4e3a\u8bc4\u4f30 RL \u751f\u6210\u7684\u4ee3\u7801\u63d0\u4f9b\u4e86\u6301\u7eed\u53ef\u9760\u7684\u73af\u5883\uff0c\u8fd9\u5bf9\u4e8e\u8fed\u4ee3\u8bad\u7ec3\u548c\u6a21\u578b\u6539\u8fdb\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>&nbsp;<\/p>\n<h3>3 \u5b9e\u9a8c<\/h3>\n<h3>3.1 \u8bc4\u4f30<\/h3>\n<p><strong>\u7531\u4e8e k1.5 \u662f\u591a\u6a21\u6001\u6a21\u578b\uff0c\u6211\u4eec\u9488\u5bf9\u4e0d\u540c\u6a21\u6001\u5728\u5404\u79cd\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u4f30\u3002\u8be6\u7ec6\u7684\u8bc4\u4f30\u8bbe\u7f6e\u53ef\u5728\u9644\u5f55 C \u4e2d\u627e\u5230\u3002\u6211\u4eec\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e3b\u8981\u5305\u62ec\u4ee5\u4e0b\u4e09\u4e2a\u7c7b\u522b\uff1a<\/strong><\/p>\n<ul>\n<li><strong>\u6587\u672c\u57fa\u51c6\u6d4b\u8bd5<\/strong>\uff1aMMLU (Hendrycks et al. 2020)\u3001IF-Eval (J. Zhou et al. 2023)\u3001CLUEWSC (L. Xu et al. 2020)\u3001C-EVAL (Y. Huang et al. 2023)<\/li>\n<li><strong>\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5<\/strong>\uff1aHumanEval-Mul\u3001LiveCodeBench (Jain et al. 2024)\u3001Codeforces\u3001AIME 2024\u3001MATH-500 (Lightman et al. 2023)<\/li>\n<li><strong>\u89c6\u89c9\u57fa\u51c6\u6d4b\u8bd5<\/strong>\uff1aMMMU (Yue, Ni, et al. 2024)\u3001MATH-Vision (K. Wang et al. 2024)\u3001MathVista (Lu et al. 2023)<\/li>\n<\/ul>\n<h3>3.2 \u4e3b\u8981\u7ed3\u679c<\/h3>\n<p><strong>K1.5 long-CoT \u6a21\u578b<\/strong>\u00a0\u8868 2 \u4e2d\u663e\u793a\u4e86 Kimi k1.5 long-CoT \u6a21\u578b\u7684\u6027\u80fd\u3002\u901a\u8fc7 long-CoT \u76d1\u7763\u5fae\u8c03\uff08\u5728\u7b2c 2.2 \u8282\u4e2d\u63cf\u8ff0\uff09\u548c\u89c6\u89c9\u6587\u672c\u8054\u5408\u5f3a\u5316\u5b66\u4e60\uff08\u5728\u7b2c 2.3 \u8282\u4e2d\u8ba8\u8bba\uff09\uff0c\u6a21\u578b\u7684\u957f\u671f\u63a8\u7406\u80fd\u529b\u5f97\u5230\u663e\u7740\u589e\u5f3a\u3002\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u6269\u5c55\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86\u5176\u6027\u80fd\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u5404\u79cd\u6a21\u5f0f\u4e2d\u5b9e\u73b0\u6700\u5148\u8fdb\u7684\u7ed3\u679c\u3002\u6211\u4eec\u7684\u8bc4\u4f30\u63ed\u793a\u4e86\u6a21\u578b\u5728\u6269\u5c55\u4e0a\u4e0b\u6587\u4e2d\u63a8\u7406\u3001\u7406\u89e3\u548c\u7efc\u5408\u4fe1\u606f\u7684\u80fd\u529b\u65b9\u9762\u7684\u663e\u7740\u6539\u8fdb\uff0c\u4ee3\u8868\u4e86\u591a\u6a21\u6001 AI \u80fd\u529b\u7684\u8fdb\u6b65\u3002<\/p>\n<p><strong>K1.5 short-CoT \u6a21\u578b<\/strong>\u00a0\u8868 3 \u4e2d\u663e\u793a\u4e86 Kimi k1.5 short-CoT \u6a21\u578b\u7684\u6027\u80fd\u3002\u6b64\u6a21\u578b\u96c6\u6210\u4e86\u591a\u79cd\u6280\u672f\uff0c\u5305\u62ec\u4f20\u7edf\u7684\u76d1\u7763\u5fae\u8c03\uff08\u5728\u7b2c 2.5.2 \u8282\u4e2d\u8ba8\u8bba\uff09\u3001\u5f3a\u5316\u5b66\u4e60\uff08\u5728\u7b2c 2.3 \u8282\u4e2d\u63a2\u8ba8\uff09\u548c\u957f\u5230\u77ed\u7684\u63d0\u70bc\uff08\u5728\u7b2c 2.4 \u8282\u4e2d\u6982\u8ff0\uff09\u3002\u7ed3\u679c\u8868\u660e\uff0c\u4e0e\u9886\u5148\u7684\u5f00\u6e90\u548c\u4e13\u6709\u6a21\u578b\u76f8\u6bd4\uff0ck1.5 short-CoT \u6a21\u578b\u5728\u591a\u9879\u4efb\u52a1\u4e2d\u63d0\u4f9b\u4e86\u6709\u7ade\u4e89\u529b\u7684\u6216\u5353\u8d8a\u7684\u6027\u80fd\u3002\u5176\u4e2d\u5305\u62ec\u6587\u672c\u3001\u89c6\u89c9\u548c\u63a8\u7406\u6311\u6218\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u3001\u6570\u5b66\u3001\u7f16\u7801\u548c\u903b\u8f91\u63a8\u7406\u65b9\u9762\u5177\u6709\u663e\u8457\u4f18\u52bf\u3002<\/p>\n<div id=\"attachment_19064\" style=\"width: 997px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19064\" class=\"wp-image-19064 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5855ce764d239d9.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"987\" height=\"333\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5855ce764d239d9.png 987w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5855ce764d239d9-300x101.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5855ce764d239d9-768x259.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/5855ce764d239d9-18x6.png 18w\" sizes=\"auto, (max-width: 987px) 100vw, 987px\" \/><p id=\"caption-attachment-19064\" class=\"wp-caption-text\">\u8868 2\uff1aKimi k1.5 long-CoT \u548c\u65d7\u8230\u5f00\u6e90\u6a21\u578b\u548c\u4e13\u6709\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_19065\" style=\"width: 1084px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19065\" class=\"wp-image-19065 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/666aa6dcf519551.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1074\" height=\"427\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/666aa6dcf519551.png 1074w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/666aa6dcf519551-300x119.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/666aa6dcf519551-1024x407.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/666aa6dcf519551-768x305.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/666aa6dcf519551-18x7.png 18w\" sizes=\"auto, (max-width: 1074px) 100vw, 1074px\" \/><p id=\"caption-attachment-19065\" class=\"wp-caption-text\">\u8868 3\uff1aKimi k1.5 short-CoT \u548c\u65d7\u8230\u5f00\u6e90\u6a21\u578b\u548c\u4e13\u6709\u6a21\u578b\u7684\u6027\u80fd\u3002VLM \u6a21\u578b\u7684\u6027\u80fd\u662f\u4ece OpenCompass \u57fa\u51c6\u6d4b\u8bd5\u5e73\u53f0 (https:\/\/opencompass.org.cn\/) \u83b7\u5f97\u7684\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<h3>3.3 \u957f\u4e0a\u4e0b\u6587\u6269\u5c55<\/h3>\n<p>\u6211\u4eec\u91c7\u7528\u4e2d\u578b\u6a21\u578b\u6765\u7814\u7a76\u5177\u6709 LLM \u7684 RL \u7684\u6269\u5c55\u5c5e\u6027\u3002\u56fe 5 \u8bf4\u660e\u4e86\u5728\u6570\u5b66\u63d0\u793a\u96c6\u4e0a\u8bad\u7ec3\u7684\u5c0f\u6a21\u578b\u53d8\u4f53\u7684\u8bad\u7ec3\u8fed\u4ee3\u4e2d\u8bad\u7ec3\u51c6\u786e\u7387\u548c\u54cd\u5e94\u957f\u5ea6\u7684\u6f14\u53d8\u3002\u968f\u7740\u8bad\u7ec3\u7684\u8fdb\u884c\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u54cd\u5e94\u957f\u5ea6\u548c\u6027\u80fd\u51c6\u786e\u7387\u540c\u6b65\u589e\u52a0\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u66f4\u5177\u6311\u6218\u6027\u7684\u57fa\u51c6\u6d4b\u8bd5\u7684\u54cd\u5e94\u957f\u5ea6\u589e\u52a0\u5f97\u66f4\u5feb\uff0c\u8fd9\u8868\u660e\u8be5\u6a21\u578b\u5b66\u4e60\u4e3a\u590d\u6742\u95ee\u9898\u751f\u6210\u66f4\u8be6\u7ec6\u7684\u89e3\u51b3\u65b9\u6848\u3002\u56fe 6 \u8868\u660e\u6a21\u578b\u7684\u8f93\u51fa\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0e\u5176\u95ee\u9898\u89e3\u51b3\u80fd\u529b\u4e4b\u95f4\u5b58\u5728\u5f88\u5f3a\u7684\u76f8\u5173\u6027\u3002\u6211\u4eec\u6700\u7ec8\u8fd0\u884c\u7684 k1.5 \u5c06\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u5230 128k\uff0c\u5e76\u89c2\u5bdf\u5230\u5728\u786c\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u6301\u7eed\u6539\u8fdb\u3002<\/p>\n<div id=\"attachment_19066\" style=\"width: 1114px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19066\" class=\"wp-image-19066 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/82ff7fed3bd87d8.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1104\" height=\"820\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/82ff7fed3bd87d8.png 1104w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/82ff7fed3bd87d8-300x223.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/82ff7fed3bd87d8-1024x761.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/82ff7fed3bd87d8-768x570.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/82ff7fed3bd87d8-16x12.png 16w\" sizes=\"auto, (max-width: 1104px) 100vw, 1104px\" \/><p id=\"caption-attachment-19066\" class=\"wp-caption-text\">\u56fe 5\uff1a\u8bad\u7ec3\u8fed\u4ee3\u6b21\u6570\u589e\u52a0\u65f6\uff0c\u8bad\u7ec3\u51c6\u786e\u7387\u548c\u957f\u5ea6\u7684\u53d8\u5316\u3002\u8bf7\u6ce8\u610f\uff0c\u4ee5\u4e0a\u5206\u6570\u6765\u81ea\u5185\u90e8 long-cot \u6a21\u578b\uff0c\u5176\u6a21\u578b\u5927\u5c0f\u6bd4 k1.5 long-CoT \u6a21\u578b\u5c0f\u5f97\u591a\u3002\u9634\u5f71\u533a\u57df\u8868\u793a\u54cd\u5e94\u957f\u5ea6\u7684\u7b2c 95 \u4e2a\u767e\u5206\u4f4d\u6570\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<h3>3.4 Long2short<\/h3>\n<p>\u6211\u4eec\u5c06\u63d0\u8bae\u7684 long2short RL \u7b97\u6cd5\u4e0e\u7b2c 2.4 \u8282\u4e2d\u4ecb\u7ecd\u7684 DPO\u3001\u6700\u77ed\u62d2\u7edd\u91c7\u6837\u548c\u6a21\u578b\u5408\u5e76\u65b9\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\uff0c\u91cd\u70b9\u5173\u6ce8 long2short \u95ee\u9898\u7684 Token \u6548\u7387 (X. Chen et al. 2024)\uff0c\u7279\u522b\u662f\u5982\u4f55\u4f7f\u83b7\u5f97\u7684 long-cot \u6a21\u578b\u80fd\u591f\u4f7f\u77ed\u6a21\u578b\u53d7\u76ca\u3002\u5728\u56fe 7 \u4e2d\uff0ck1.5-long \u8868\u793a\u6211\u4eec\u9009\u62e9\u7528\u4e8e long2short \u8bad\u7ec3\u7684 long-cot \u6a21\u578b\u3002k1.5-short w\/ rl \u6307\u7684\u662f\u4f7f\u7528 long2short RL \u8bad\u7ec3\u83b7\u5f97\u7684\u77ed\u6a21\u578b\u3002k1.5-short w\/ dpo \u8868\u793a\u901a\u8fc7 DPO \u8bad\u7ec3\u63d0\u9ad8 Token \u6548\u7387\u7684\u77ed\u6a21\u578b\u3002k1.5-short w\/ merge \u8868\u793a\u6a21\u578b\u5408\u5e76\u540e\u7684\u6a21\u578b\uff0c\u800c k1.5-short w\/ merge + rs \u8868\u793a\u901a\u8fc7\u5bf9\u5408\u5e76\u7684\u6a21\u578b\u5e94\u7528\u6700\u77ed\u62d2\u7edd\u91c7\u6837\u83b7\u5f97\u7684\u77ed\u6a21\u578b\u3002k1.5-shortest \u8868\u793a\u6211\u4eec\u5728 long2short \u8bad\u7ec3\u671f\u95f4\u83b7\u5f97\u7684\u6700\u77ed\u6a21\u578b\u3002\u5982\u56fe 7 \u6240\u793a\uff0c\u4e0e DPO \u548c\u6a21\u578b\u5408\u5e76\u7b49\u5176\u4ed6\u65b9\u6cd5\u76f8\u6bd4\uff0c\u63d0\u8bae\u7684 long2short RL \u7b97\u6cd5\u8868\u73b0\u51fa\u6700\u9ad8\u7684 Token \u6548\u7387\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0ck1.5 \u7cfb\u5217\u4e2d\u7684\u6240\u6709\u6a21\u578b\uff08\u4ee5\u6a59\u8272\u6807\u8bb0\uff09\u90fd\u8868\u73b0\u51fa\u6bd4\u5176\u4ed6\u6a21\u578b\uff08\u4ee5\u84dd\u8272\u6807\u8bb0\uff09\u66f4\u597d\u7684 Token \u6548\u7387\u3002\u4f8b\u5982\uff0ck1.5-short w\/ rl \u5728 AIME2024 \u4e0a\u5b9e\u73b0\u4e86 60.8 \u7684 Pass@1 \u5206\u6570\uff08\u5e73\u5747 8 \u6b21\u8fd0\u884c\uff09\uff0c\u540c\u65f6\u5e73\u5747\u4ec5\u4f7f\u7528 3,272 \u4e2a Token\u3002\u540c\u6837\uff0ck1.5-shortest \u5728 MATH500 \u4e0a\u83b7\u5f97\u4e86 88.2 \u7684 Pass@1 \u5206\u6570\uff0c\u540c\u65f6\u6d88\u8017\u7684 Token \u6570\u91cf\u4e0e\u5176\u4ed6\u77ed\u6a21\u578b\u5927\u81f4\u76f8\u540c\u3002<\/p>\n<div id=\"attachment_19067\" style=\"width: 1382px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19067\" class=\"wp-image-19067 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b3d0cd8a3270627.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1372\" height=\"686\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b3d0cd8a3270627.png 1372w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b3d0cd8a3270627-300x150.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b3d0cd8a3270627-1024x512.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b3d0cd8a3270627-768x384.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b3d0cd8a3270627-18x9.png 18w\" sizes=\"auto, (max-width: 1372px) 100vw, 1372px\" \/><p id=\"caption-attachment-19067\" class=\"wp-caption-text\">\u56fe 6\uff1a\u6a21\u578b\u6027\u80fd\u968f\u54cd\u5e94\u957f\u5ea6\u7684\u589e\u52a0\u800c\u63d0\u9ad8<\/p><\/div>\n<p>&nbsp;<\/p>\n<div id=\"attachment_19068\" style=\"width: 1383px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19068\" class=\"wp-image-19068 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f5c2654158fbd9d.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1373\" height=\"548\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f5c2654158fbd9d.png 1373w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f5c2654158fbd9d-300x120.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f5c2654158fbd9d-1024x409.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f5c2654158fbd9d-768x307.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/f5c2654158fbd9d-18x7.png 18w\" sizes=\"auto, (max-width: 1373px) 100vw, 1373px\" \/><p id=\"caption-attachment-19068\" class=\"wp-caption-text\">\u56fe 7\uff1aLong2Short \u6027\u80fd\u3002\u6240\u6709 k1.5 \u7cfb\u5217\u90fd\u5c55\u793a\u4e86\u6bd4\u5176\u4ed6\u6a21\u578b\u66f4\u597d\u7684 Token \u6548\u7387\u3002<\/p><\/div>\n<h3>3.5 \u6d88\u878d\u7814\u7a76<\/h3>\n<p><strong>\u6a21\u578b\u5927\u5c0f\u548c\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u6269\u5c55<\/strong>\u00a0\u6211\u4eec\u7684\u4e3b\u8981\u8d21\u732e\u662f\u5c06 RL \u5e94\u7528\u4e8e\u589e\u5f3a\u6a21\u578b\u751f\u6210\u6269\u5c55 CoT \u7684\u80fd\u529b\uff0c\u4ece\u800c\u63d0\u9ad8\u5176\u63a8\u7406\u80fd\u529b\u3002\u4e00\u4e2a\u81ea\u7136\u800c\u7136\u7684\u95ee\u9898\u662f\uff1a\u8fd9\u4e0e\u7b80\u5355\u5730\u589e\u52a0\u6a21\u578b\u5927\u5c0f\u76f8\u6bd4\u5982\u4f55\uff1f\u4e3a\u4e86\u8bc1\u660e\u6211\u4eec\u65b9\u6cd5\u7684\u6709\u6548\u6027\uff0c\u6211\u4eec\u4f7f\u7528\u76f8\u540c\u7684\u6570\u636e\u96c6\u8bad\u7ec3\u4e86\u4e24\u4e2a\u4e0d\u540c\u5927\u5c0f\u7684\u6a21\u578b\uff0c\u5e76\u8bb0\u5f55\u4e86\u5728 RL \u8bad\u7ec3\u671f\u95f4\u6240\u6709\u68c0\u67e5\u70b9\u7684\u8bc4\u4f30\u7ed3\u679c\u548c\u5e73\u5747\u63a8\u7406\u957f\u5ea6\u3002\u8fd9\u4e9b\u7ed3\u679c\u5982\u56fe 8 \u6240\u793a\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5c3d\u7ba1\u8f83\u5927\u7684\u6a21\u578b\u6700\u521d\u4f18\u4e8e\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u4f46\u8f83\u5c0f\u7684\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u5229\u7528\u901a\u8fc7 RL \u4f18\u5316\u7684\u8f83\u957f CoT \u6765\u5b9e\u73b0\u53ef\u6bd4\u7684\u6027\u80fd\u3002\u4f46\u662f\uff0c\u8f83\u5927\u7684\u6a21\u578b\u901a\u5e38\u6bd4\u8f83\u5c0f\u7684\u6a21\u578b\u8868\u73b0\u51fa\u66f4\u597d\u7684 Token \u6548\u7387\u3002\u8fd9\u4e5f\u8868\u660e\uff0c\u5982\u679c\u76ee\u6807\u662f\u83b7\u5f97\u5c3d\u53ef\u80fd\u597d\u7684\u6027\u80fd\uff0c\u5219\u6269\u5c55\u8f83\u5927\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u5177\u6709\u66f4\u9ad8\u7684\u4e0a\u9650\uff0c\u5e76\u4e14 Token \u6548\u7387\u66f4\u9ad8\u3002\u4f46\u662f\uff0c\u5982\u679c\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u5177\u6709\u9884\u7b97\uff0c\u5219\u8bad\u7ec3\u5177\u6709\u8f83\u957f\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u8f83\u5c0f\u6a21\u578b\u53ef\u80fd\u662f\u53ef\u884c\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n<p><strong>\u4f7f\u7528\u8d1f\u68af\u5ea6\u7684\u5f71\u54cd<\/strong>\u00a0\u6211\u4eec\u7814\u7a76\u4e86\u5728\u6211\u4eec\u7684\u8bbe\u7f6e\u4e2d\u4f7f\u7528 ReST (Gulcehre et al. 2023) \u4f5c\u4e3a\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\u7684\u6709\u6548\u6027\u3002ReST \u4e0e\u5176\u4ed6\u57fa\u4e8e RL \u7684\u65b9\u6cd5\uff08\u5305\u62ec\u6211\u4eec\u7684\u65b9\u6cd5\uff09\u4e4b\u95f4\u7684\u4e3b\u8981\u533a\u522b\u5728\u4e8e\uff0cReST \u901a\u8fc7\u62df\u5408\u4ece\u5f53\u524d\u6a21\u578b\u91c7\u6837\u7684\u6700\u4f73\u54cd\u5e94\u6765\u8fed\u4ee3\u5730\u6539\u8fdb\u6a21\u578b\uff0c\u800c\u4e0d\u4f1a\u5e94\u7528\u8d1f\u68af\u5ea6\u6765\u60e9\u7f5a\u4e0d\u6b63\u786e\u7684\u54cd\u5e94\u3002\u5982\u56fe 10 \u6240\u793a\uff0c\u4e0e ReST \u76f8\u6bd4\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u8868\u73b0\u51fa\u66f4\u9ad8\u7684\u6837\u672c\u590d\u6742\u6027\uff0c\u8868\u660e\u7eb3\u5165\u8d1f\u68af\u5ea6\u663e\u7740\u63d0\u9ad8\u4e86\u6a21\u578b\u751f\u6210 long-CoT \u7684\u6548\u7387\u3002\u6211\u4eec\u7684\u65b9\u6cd5\u4e0d\u4ec5\u63d0\u9ad8\u4e86\u63a8\u7406\u8d28\u91cf\uff0c\u800c\u4e14\u8fd8\u4f18\u5316\u4e86\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u4ece\u800c\u4f7f\u7528\u66f4\u5c11\u7684\u8bad\u7ec3\u6837\u672c\u5b9e\u73b0\u4e86\u7a33\u5065\u7684\u6027\u80fd\u3002\u8fd9\u4e00\u53d1\u73b0\u8868\u660e\uff0c\u7b56\u7565\u4f18\u5316\u7b97\u6cd5\u7684\u9009\u62e9\u5728\u6211\u4eec\u7684\u8bbe\u7f6e\u4e2d\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a ReST \u548c\u5176\u4ed6\u57fa\u4e8e RL \u7684\u65b9\u6cd5\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u5728\u5176\u4ed6\u9886\u57df\u5e76\u4e0d\u50cf (Gulcehre et al. 2023) \u4e2d\u90a3\u4e48\u660e\u663e\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7684\u7ed3\u679c\u7a81\u51fa\u4e86\u9009\u62e9\u9002\u5f53\u7684\u4f18\u5316\u7b56\u7565\u4ee5\u6700\u5927\u7a0b\u5ea6\u5730\u63d0\u9ad8\u751f\u6210 long-CoT \u7684\u6548\u7387\u7684\u91cd\u8981\u6027\u3002<\/p>\n<p><strong>\u91c7\u6837\u7b56\u7565<\/strong>\u00a0\u6211\u4eec\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u7b2c 2.3.4 \u8282\u4e2d\u4ecb\u7ecd\u7684\u8bfe\u7a0b\u91c7\u6837\u7b56\u7565\u7684\u6709\u6548\u6027\u3002\u6211\u4eec\u7684\u8bad\u7ec3\u6570\u636e\u96c6 D \u5305\u542b\u5404\u79cd\u96be\u5ea6\u7ea7\u522b\u7684\u95ee\u9898\u3002\u901a\u8fc7\u6211\u4eec\u7684\u8bfe\u7a0b\u91c7\u6837\u65b9\u6cd5\uff0c\u6211\u4eec\u6700\u521d\u4f7f\u7528 D \u8fdb\u884c\u70ed\u8eab\u9636\u6bb5\uff0c\u7136\u540e\u4ec5\u5173\u6ce8\u786c\u95ee\u9898\u6765\u8bad\u7ec3\u6a21\u578b\u3002\u5c06\u6b64\u65b9\u6cd5\u4e0e\u91c7\u7528\u4e0d\u8fdb\u884c\u4efb\u4f55\u8bfe\u7a0b\u8c03\u6574\u7684\u7edf\u4e00\u91c7\u6837\u7b56\u7565\u7684\u57fa\u7ebf\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\u3002\u5982\u56fe 9 \u6240\u793a\uff0c\u6211\u4eec\u7684\u7ed3\u679c\u6e05\u695a\u5730\u8868\u660e\uff0c\u63d0\u8bae\u7684\u8bfe\u7a0b\u91c7\u6837\u65b9\u6cd5\u663e\u7740\u63d0\u9ad8\u4e86\u6027\u80fd\u3002\u8fd9\u79cd\u6539\u8fdb\u53ef\u4ee5\u5f52\u56e0\u4e8e\u8be5\u65b9\u6cd5\u9010\u6b65\u6311\u6218\u6a21\u578b\u7684\u80fd\u529b\uff0c\u4f7f\u5176\u80fd\u591f\u5728\u5904\u7406\u590d\u6742\u95ee\u9898\u65f6\u5f62\u6210\u66f4\u5f3a\u5927\u7684\u7406\u89e3\u548c\u80fd\u529b\u3002\u901a\u8fc7\u5728\u521d\u6b65\u7684\u6982\u62ec\u6027\u4ecb\u7ecd\u4e4b\u540e\u5c06\u8bad\u7ec3\u7cbe\u529b\u96c6\u4e2d\u5728\u66f4\u56f0\u96be\u7684\u95ee\u9898\u4e0a\uff0c\u6a21\u578b\u53ef\u4ee5\u66f4\u597d\u5730\u52a0\u5f3a\u5176\u63a8\u7406\u548c\u95ee\u9898\u89e3\u51b3\u80fd\u529b\u3002<\/p>\n<div id=\"attachment_19069\" style=\"width: 1389px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19069\" class=\"wp-image-19069 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/626881786a74cd5.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1379\" height=\"346\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/626881786a74cd5.png 1379w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/626881786a74cd5-300x75.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/626881786a74cd5-1024x257.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/626881786a74cd5-768x193.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/626881786a74cd5-18x5.png 18w\" sizes=\"auto, (max-width: 1379px) 100vw, 1379px\" \/><p id=\"caption-attachment-19069\" class=\"wp-caption-text\">\u56fe 8\uff1a\u4e0d\u540c\u6a21\u578b\u5927\u5c0f\u7684\u6a21\u578b\u6027\u80fd\u4e0e\u54cd\u5e94\u957f\u5ea6\u7684\u5173\u7cfb<\/p><\/div>\n<div id=\"attachment_19070\" style=\"width: 848px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19070\" class=\"wp-image-19070 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8b8c5c49190b19e.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"838\" height=\"489\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8b8c5c49190b19e.png 838w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8b8c5c49190b19e-300x175.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8b8c5c49190b19e-768x448.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/8b8c5c49190b19e-18x12.png 18w\" sizes=\"auto, (max-width: 838px) 100vw, 838px\" \/><p id=\"caption-attachment-19070\" class=\"wp-caption-text\">\u56fe 9\uff1a\u5bf9\u6a21\u578b\u6027\u80fd\u7684\u8bfe\u7a0b\u5b66\u4e60\u65b9\u6cd5\u5206\u6790\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<h3>4 \u7ed3\u8bba<\/h3>\n<p>\u6211\u4eec\u4ecb\u7ecd\u4e86 k1.5\uff08\u6211\u4eec\u6700\u65b0\u7684\u4f7f\u7528 RL \u8bad\u7ec3\u7684\u591a\u6a21\u6001 LLM\uff09\u7684\u8bad\u7ec3\u914d\u65b9\u548c\u7cfb\u7edf\u8bbe\u8ba1\u3002\u6211\u4eec\u4ece\u5b9e\u8df5\u4e2d\u5f97\u51fa\u7684\u4e00\u4e2a\u5173\u952e\u89c1\u89e3\u662f\uff0c\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u6269\u5c55\u5bf9\u4e8e LLM \u7684\u6301\u7eed\u6539\u8fdb\u81f3\u5173\u91cd\u8981\u3002\u6211\u4eec\u91c7\u7528\u4e86\u4f18\u5316\u7684\u5b66\u4e60\u7b97\u6cd5\u548c\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\uff0c\u4f8b\u5982\u90e8\u5206 rollout\uff0c\u4ee5\u5b9e\u73b0\u9ad8\u6548\u7684\u957f\u4e0a\u4e0b\u6587 RL \u8bad\u7ec3\u3002\u5982\u4f55\u8fdb\u4e00\u6b65\u63d0\u9ad8\u957f\u4e0a\u4e0b\u6587 RL \u8bad\u7ec3\u7684\u6548\u7387\u548c\u53ef\u6269\u5c55\u6027\u4ecd\u7136\u662f\u4e00\u4e2a\u91cd\u8981\u7684\u524d\u8fdb\u65b9\u5411\u3002<\/p>\n<div id=\"attachment_19071\" style=\"width: 1119px\" class=\"wp-caption aligncenter\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-19071\" class=\"wp-image-19071 size-full\" title=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b4ecf1182c9592b.png\" alt=\"\u4e0d\u5e78\u4e0eDeepSeek-R1\u540c\u65f6\u53d1\u5e03\u7684Kimi k1.5\u6280\u672f\u62a5\u544a-1\" width=\"1109\" height=\"817\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b4ecf1182c9592b.png 1109w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b4ecf1182c9592b-300x221.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b4ecf1182c9592b-1024x754.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b4ecf1182c9592b-768x566.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/b4ecf1182c9592b-16x12.png 16w\" sizes=\"auto, (max-width: 1109px) 100vw, 1109px\" \/><p id=\"caption-attachment-19071\" class=\"wp-caption-text\">\u56fe 10\uff1a\u4f7f\u7528 ReST \u8fdb\u884c\u7b56\u7565\u4f18\u5316\u7684\u6bd4\u8f83\u3002<\/p><\/div>\n<p>&nbsp;<\/p>\n<p>\u6211\u4eec\u505a\u51fa\u7684\u53e6\u4e00\u4e2a\u8d21\u732e\u662f\u7ed3\u5408\u4e86\u591a\u79cd\u6280\u672f\uff0c\u4ece\u800c\u53ef\u4ee5\u6539\u8fdb\u7b56\u7565\u4f18\u5316\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u4f7f\u7528 LLM \u516c\u5f0f\u5316\u4e86 long-CoT RL\uff0c\u5e76\u63a8\u5bfc\u4e86\u7528\u4e8e\u9c81\u68d2\u4f18\u5316\u7684\u5728\u7ebf\u955c\u50cf\u4e0b\u964d\u7684\u53d8\u4f53\u3002\u6211\u4eec\u8fd8\u5c1d\u8bd5\u4e86\u91c7\u6837\u7b56\u7565\u3001\u957f\u5ea6\u60e9\u7f5a\u548c\u4f18\u5316\u6570\u636e\u914d\u65b9\uff0c\u4ee5\u5b9e\u73b0\u5f3a\u5927\u7684 RL \u6027\u80fd\u3002<\/p>\n<p>\u6211\u4eec\u8868\u660e\uff0c\u5373\u4f7f\u4e0d\u4f7f\u7528\u66f4\u590d\u6742\u7684\u6280\u672f\uff08\u4f8b\u5982\u8499\u7279\u5361\u7f57\u6811\u641c\u7d22\u3001\u4ef7\u503c\u51fd\u6570\u548c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff09\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7\u957f\u4e0a\u4e0b\u6587\u6269\u5c55\u548c\u6539\u8fdb\u7684\u7b56\u7565\u4f18\u5316\u6765\u5b9e\u73b0\u5f3a\u5927\u7684\u6027\u80fd\u3002\u5728\u672a\u6765\uff0c\u7814\u7a76\u5982\u4f55\u5728\u4e0d\u635f\u5bb3\u6a21\u578b\u63a2\u7d22\u80fd\u529b\u7684\u60c5\u51b5\u4e0b\u6539\u8fdb\u4fe1\u7528\u5206\u914d\u548c\u51cf\u5c11\u8fc7\u5ea6\u601d\u8003\u4e5f\u5c06\u5f88\u6709\u8da3\u3002<\/p>\n<p>\u6211\u4eec\u8fd8\u89c2\u5bdf\u5230\u4e86 long2short \u65b9\u6cd5\u7684\u6f5c\u529b\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u63d0\u9ad8\u4e86 short-CoT \u6a21\u578b\u7684\u6027\u80fd\u3002\u6b64\u5916\uff0c\u53ef\u4ee5\u901a\u8fc7\u8fed\u4ee3\u65b9\u5f0f\u5c06 long2short \u65b9\u6cd5\u4e0e long-CoT RL \u76f8\u7ed3\u5408\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8 Token \u6548\u7387\uff0c\u5e76\u5728\u7ed9\u5b9a\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u9884\u7b97\u4e2d\u63d0\u53d6\u6700\u4f73\u6027\u80fd\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Kimi k1.5\u6280\u672f\u62a5\u544a\u901f\u8bfb 1. \u5f3a\u5927\u7684\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\uff1a Kimi k1.5 \u6a21\u578b\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u548c\u6a21\u6001\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u63a8\u7406\u6027\u80fd\uff0c\u5305\u62ec\u6570\u5b66\u3001\u4ee3\u7801\u3001\u6587\u672c\u548c\u89c6\u89c9\u63a8\u7406\u7b49\u4efb\u52a1\u3002 \u5b83\u4e0d\u4ec5\u80fd\u5904\u7406\u7eaf\u6587\u672c\uff0c\u8fd8\u80fd\u7406\u89e3\u56fe\u50cf\u548c\u6587\u672c\u7684\u7ec4\u5408\uff0c\u5b9e\u73b0\u771f\u6b63\u7684\u591a\u6a21\u6001\u63a8\u7406\u3002&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[46],"tags":[],"class_list":["post-19051","post","type-post","status-publish","format-standard","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/19051","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=19051"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/19051\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=19051"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=19051"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=19051"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}