{"id":523,"date":"2025-10-31T07:00:00","date_gmt":"2025-10-30T23:00:00","guid":{"rendered":"https:\/\/smartaiunion.com\/?p=523"},"modified":"2025-10-29T11:25:06","modified_gmt":"2025-10-29T03:25:06","slug":"thinking-machines-lab%e5%8d%9a%e5%ae%a2%e6%8f%90%e5%87%ba%e5%9c%a8%e7%ad%96%e7%95%a5%e8%92%b8%e9%a6%8f%ef%bc%8cqwen%e8%a2%abcue-38%e6%ac%a1","status":"publish","type":"post","link":"https:\/\/smartaiunion.com\/index.php\/2025\/10\/31\/523\/","title":{"rendered":"Thinking Machines Lab\u535a\u5ba2\u63d0\u51fa\u5728\u7b56\u7565\u84b8\u998f\uff0cQwen\u88abcue 38\u6b21"},"content":{"rendered":"\n<p>\u5728\u7b56\u7565\u84b8\u998f\u662f\u4e00\u79cd\u5c06\u5f3a\u5316\u5b66\u4e60 (RL) \u7684\u7ea0\u9519\u76f8\u5173\u6027\u4e0e SFT \u7684\u5956\u52b1\u5bc6\u5ea6\u76f8\u7ed3\u5408\u7684\u8bad\u7ec3\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u521a\u521a\uff0c\u4e0d\u53d1\u8bba\u6587\u3001\u7231\u53d1\u535a\u5ba2\u7684 Thinking Machines Lab \uff08\u4ee5\u4e0b\u7b80\u79f0 TML\uff09\u518d\u6b21\u66f4\u65b0\uff0c\u53d1\u5e03\u4e86\u4e00\u7bc7\u9898\u4e3a\u300a\u5728\u7b56\u7565\u84b8\u998f\u300b\u7684\u535a\u5ba2\u3002<\/p>\n\n\n\n<p><strong>\u5728\u7b56\u7565\u84b8\u998f\uff08on-policy distillation\uff09<\/strong>\u662f\u4e00\u79cd\u5c06\u5f3a\u5316\u5b66\u4e60 (RL) \u7684\u7ea0\u9519\u76f8\u5173\u6027\u4e0e SFT \u7684\u5956\u52b1\u5bc6\u5ea6\u76f8\u7ed3\u5408\u7684\u8bad\u7ec3\u65b9\u6cd5\u3002\u5728\u5c06\u5176\u7528\u4e8e\u6570\u5b66\u63a8\u7406\u548c\u5185\u90e8\u804a\u5929\u52a9\u624b\u65f6\uff0cTML \u53d1\u73b0\u5728\u7b56\u7565\u84b8\u998f\u53ef\u4ee5\u6781\u4f4e\u7684\u6210\u672c\u8d85\u8d8a\u5176\u4ed6\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u8be5\u516c\u53f8 CEO Mira Murati \u8868\u793a\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53ef\u7528\u4e8e\u5c0f\u6a21\u578b\uff0c\u4f7f\u5176\u5177\u5907\u5f3a\u5927\u7684\u9886\u57df\u6027\u80fd\u548c\u6301\u7eed\u5b66\u4e60\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_b95136d532144cbc89d4e4073c1f1e67@5091053_oswg130734oswg867oswg918_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u8fd9\u7bc7\u65b0\u535a\u5ba2\u4e2d\uff0cTML \u660e\u786e\u8868\u793a\u8fd9\u9879\u65b0\u6210\u679c\u53d7\u5230\u4e86 Qwen \u56e2\u961f\u7814\u7a76\u7684\u542f\u53d1\uff0c\u5e76\u4e14\u5176\u5b9e\u9a8c\u8fc7\u7a0b\u4e2d\u4e5f\u5927\u91cf\u7528\u5230\u4e86 Qwen3 \u7cfb\u5217\u6a21\u578b\u3002\u4e8b\u5b9e\u4e0a\uff0c\u5728\u539f\u82f1\u6587\u535a\u5ba2\u4e2d\uff0c\u300cQwen\u300d\u8fd9\u4e2a\u5173\u952e\u8bcd\u4e00\u5171\u51fa\u73b0\u4e86 38 \u6b21\u4e4b\u591a\uff01\u6bd4\u5c0f\u7c73 17 \u7cfb\u5217\u53d1\u5e03\u4f1a\u96f7\u603b\u63d0\u5230\u300c\u82f9\u679c\u300d\u7684 37 \u6b21\u8fd8\u591a\u4e00\u6b21\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_7903d0cb7726488285766dde5f75ba9b@5091053_oswg74878oswg1080oswg145_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4f5c\u4e3a\u4e00\u5bb6\u660e\u661f\u521b\u4e1a\u516c\u53f8\uff0cTML \u7684\u66f4\u65b0\u4e5f\u5438\u5f15\u4e86\u5e7f\u6cdb\u5173\u6ce8\u3002\u6709\u4eba\u603b\u7ed3\u5176\u4f18\u52bf\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_9a203882b2904d5983d9a2c992c01ac2@5091053_oswg163328oswg869oswg528_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u66f4\u662f\u6709\u7f51\u53cb\u76db\u8d5e\uff0cTML \u624d\u662f\u771f Open AI\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_35d5f34cd1a94132ab4162a73f505120@5091053_oswg19837oswg562oswg131_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_a07d01f511e5435b9eef1ca0d3bf47f7@5091053_oswg22739oswg745oswg207_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u535a\u5ba2\u5730\u5740\uff1ahttps:\/\/thinkingmachines.ai\/blog\/on-policy-distillation\/<\/p>\n\n\n\n<p>\u8fd9\u7bc7\u535a\u5ba2\u7684\u4e3b\u8981\u4f5c\u8005\u662f Thinking Machines Lab \u7814\u7a76\u8005 Kevin Lu\u3002\u4ed6\u4e4b\u524d\u66fe\u5728 OpenAI \u5de5\u4f5c\uff0c\u9886\u5bfc\u4e86 4o-mini \u7684\u53d1\u5e03\uff0c\u5e76\u53c2\u4e0e\u8fc7 GPT-5 series\u3001GPT-oss\u3001o3 &amp; o4-mini\u30014.1-nano &amp; 4.1-mini\u3001o1-mini\u3001o3-mini \u7b49\u6a21\u578b\u7684\u7814\u53d1\u5de5\u4f5c\u3002<\/p>\n\n\n\n<p>\u4e0b\u9762\u6211\u4eec\u5c31\u6765\u8be6\u7ec6\u770b\u770b\u8fd9\u7bc7\u535a\u5ba2\u7684\u5185\u5bb9\u3002<\/p>\n\n\n\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u80fd\u591f\u5728\u7279\u5b9a\u9886\u57df\u5c55\u73b0\u51fa\u4e13\u5bb6\u7ea7\u7684\u6c34\u5e73\u3002\u8fd9\u662f\u51e0\u79cd\u80fd\u529b\u5171\u540c\u4f5c\u7528\u7684\u7ed3\u679c\uff0c\u5305\u62ec\uff1a\u5bf9\u8f93\u5165\u7684\u611f\u77e5\u3001\u77e5\u8bc6\u68c0\u7d22\u3001\u89c4\u5212\u9009\u62e9\u548c\u53ef\u9760\u6267\u884c\u3002<\/p>\n\n\n\n<p>\u8981\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u9700\u8981\u4e00\u7cfb\u5217\u7684\u8bad\u7ec3\u65b9\u6cd5\u3002\u6211\u4eec\u53ef\u4ee5\u5c06\u5176\u5927\u81f4\u5206\u4e3a\u4e09\u4e2a\u9636\u6bb5\uff1a<\/p>\n\n\n\n<p><strong>\u9884\u8bad\u7ec3\uff08Pre-training\uff09<\/strong>\uff1a\u6559\u6388\u901a\u7528\u80fd\u529b\uff0c\u4f8b\u5982\u8bed\u8a00\u4f7f\u7528\u3001\u5bbd\u6cdb\u7684\u63a8\u7406\u548c\u4e16\u754c\u77e5\u8bc6\u3002<\/p>\n\n\n\n<p><strong>\u4e2d\u8bad\u7ec3\uff08Mid-training\uff09<\/strong>\uff1a\u4f20\u6388\u9886\u57df\u77e5\u8bc6\uff0c\u4f8b\u5982\u4ee3\u7801\u3001\u533b\u7597\u6570\u636e\u5e93\u6216\u516c\u53f8\u5185\u90e8\u6587\u4ef6\u3002<\/p>\n\n\n\n<p><strong>\u540e\u8bad\u7ec3\uff08Post-training\uff09<\/strong>\uff1a\u5f15\u5bfc\u51fa\u76ee\u6807\u884c\u4e3a\uff0c\u4f8b\u5982\u9075\u5faa\u6307\u4ee4\u3001\u89e3\u51b3\u6570\u5b66\u95ee\u9898\u6216\u804a\u5929\u3002<\/p>\n\n\n\n<p>\u5728\u7279\u5b9a\u4e13\u4e1a\u9886\u57df\uff0c\u7ecf\u8fc7\u5f3a\u5316\u8bad\u7ec3\u7684\u5c0f\u578b\u6a21\u578b\uff0c\u5176\u8868\u73b0\u5f80\u5f80\u4f18\u4e8e\u90a3\u4e9b\u5927\u578b\u7684\u901a\u7528\u6a21\u578b\u3002\u4f7f\u7528\u5c0f\u578b\u6a21\u578b\u6709\u5f88\u591a\u597d\u5904\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u51fa\u4e8e\u9690\u79c1\u6216\u5b89\u5168\u8003\u8651\uff0c\u5b83\u4eec\u53ef\u4ee5\u8fdb\u884c\u672c\u5730\u90e8\u7f72\u3002<\/p>\n\n\n\n<p>\u5b83\u4eec\u53ef\u4ee5\u66f4\u8f7b\u677e\u5730\u6301\u7eed\u8bad\u7ec3\u548c\u66f4\u65b0\u3002<\/p>\n\n\n\n<p>\u5b83\u4eec\u8fd8\u80fd\u8282\u7701\u63a8\u7406\u6210\u672c\u3002<\/p>\n<\/blockquote>\n\n\n\n<p>\u60f3\u8981\u5229\u7528\u8fd9\u4e9b\u4f18\u52bf\uff0c\u5c31\u9700\u8981\u4e3a\u8bad\u7ec3\u7684\u540e\u7eed\u9636\u6bb5\u9009\u62e9\u6b63\u786e\u7684\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u540e\u8bad\u7ec3\u300c\u5b66\u751f\u300d\u6a21\u578b\u7684\u65b9\u6cd5\u53ef\u4ee5\u5206\u4e3a\u4e24\u79cd\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u5728\u7b56\u7565\uff08On-policy\uff09\u8bad\u7ec3\uff1a\u4ece\u5b66\u751f\u6a21\u578b\u81ea\u8eab\u91c7\u6837\u8f68\u8ff9\uff08rollouts\uff09\uff0c\u5e76\u4e3a\u8fd9\u4e9b\u8f68\u8ff9\u5206\u914d\u67d0\u79cd\u5956\u52b1\u3002<\/p>\n\n\n\n<p>\u79bb\u7b56\u7565\uff08Off-policy\uff09\u8bad\u7ec3\uff1a\u4f9d\u8d56\u4e8e\u67d0\u4e2a\u5916\u90e8\u6765\u6e90\u7684\u76ee\u6807\u8f93\u51fa\uff0c\u5b66\u751f\u6a21\u578b\u9700\u8981\u5b66\u4e60\u6a21\u4eff\u8fd9\u4e9b\u8f93\u51fa\u3002<\/p>\n<\/blockquote>\n\n\n\n<p>\u4f8b\u5982\uff0c\u6211\u4eec\u53ef\u80fd\u5e0c\u671b\u8bad\u7ec3\u4e00\u4e2a\u7d27\u51d1\u6a21\u578b\u6765\u89e3\u51b3\u5982\u4e0b\u7684\u6570\u5b66\u95ee\u9898\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_92e8f99a6aaa4cbf8b3165eb92319f8c@5091053_oswg8660oswg469oswg81_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6765\u8fdb\u884c\u5728\u7b56\u7565\u8bad\u7ec3\u3002\u5177\u4f53\u505a\u6cd5\u662f\u6839\u636e\u5b66\u751f\u6a21\u578b\u7684\u6bcf\u4e2a\u8f68\u8ff9\u662f\u5426\u89e3\u51b3\u4e86\u95ee\u9898\u6765\u4e3a\u5176\u8bc4\u5206\u3002\u8fd9\u4e2a\u8bc4\u5206\u53ef\u4ee5\u7531\u4eba\u5de5\u5b8c\u6210\uff0c\u4e5f\u53ef\u4ee5\u7531\u4e00\u4e2a\u80fd\u53ef\u9760\u7ed9\u51fa\u6b63\u786e\u7b54\u6848\u7684\u300c\u6559\u5e08\u300d\u6a21\u578b\u6765\u5b8c\u6210\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_4e8b58989a6c48d7b0c2cca093d02323@5091053_oswg16821oswg1016oswg155_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u5728\u7b56\u7565\u8bad\u7ec3\u7684\u4f18\u52bf\u5728\u4e8e\uff0c\u5b66\u751f\u901a\u8fc7\u5728\u81ea\u5df1\u7684\u6837\u672c\u4e0a\u8bad\u7ec3\uff0c\u80fd\u66f4\u76f4\u63a5\u5730\u5b66\u4f1a\u907f\u514d\u9519\u8bef\u3002<\/p>\n\n\n\n<p>\u4f46 RL \u6709\u4e00\u4e2a\u4e3b\u8981\u7f3a\u70b9\uff1a\u5b83\u63d0\u4f9b\u7684\u53cd\u9988\u975e\u5e38\u7a00\u758f\uff08sparse feedback\uff09\u3002\u65e0\u8bba\u4f7f\u7528\u591a\u5c11 token\uff0c\u5b83\u5728\u6bcf\u4e2a\u8bad\u7ec3\u56de\u5408\uff08episode\uff09\u4e2d\u6559\u6388\u7684\u6bd4\u7279\u6570\u662f\u56fa\u5b9a\u7684\u3002<\/p>\n\n\n\n<p>\u5728\u6211\u4eec\u4e0a\u9762\u7684\u4f8b\u5b50\u4e2d\uff0c\u5b66\u751f\u53ea\u77e5\u9053\u300c21\u300d\u662f\u9519\u8bef\u7b54\u6848\uff0c\u5e76\u66f4\u65b0\u6a21\u578b\u4ee5\u907f\u514d\u4ea7\u751f\u8fd9\u4e2a\u8f68\u8ff9\u3002\u4f46\u5b83\u5e76\u6ca1\u6709\u5b66\u5230\u7a76\u7adf\u9519\u5728\u54ea\u91cc \u2014\u2014 \u662f\u641e\u9519\u4e86\u8fd0\u7b97\u987a\u5e8f\uff0c\u8fd8\u662f\u7b97\u672f\u672c\u8eab\u51fa\u4e86\u9519\u3002\u8fd9\u79cd\u53cd\u9988\u7684\u7a00\u758f\u6027\u4f7f\u5f97 RL \u5728\u8bb8\u591a\u5e94\u7528\u4e2d\u6548\u7387\u4f4e\u4e0b\u3002<\/p>\n\n\n\n<p>\u79bb\u7b56\u7565\u8bad\u7ec3\u901a\u5e38\u901a\u8fc7\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u6765\u5b8c\u6210\uff0c\u5373\u5728\u4e00\u7ec4\u7cbe\u5fc3\u7b56\u5212\u7684\u3001\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u7684\u6709\u6807\u6ce8\u793a\u4f8b\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u8fd9\u4e9b\u6709\u6807\u6ce8\u793a\u4f8b\u7684\u6765\u6e90\u53ef\u4ee5\u662f\u4e00\u4e2a\u5728\u5f53\u524d\u4efb\u52a1\u4e0a\u8868\u73b0\u51fa\u8272\u7684\u6559\u5e08\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u4e00\u79cd\u79f0\u4e3a\u84b8\u998f\uff08distillation\uff09\u7684\u673a\u5236\uff1a\u8bad\u7ec3\u5b66\u751f\u6a21\u578b\u6765\u5339\u914d\u6559\u5e08\u6a21\u578b\u7684\u8f93\u51fa\u5206\u5e03\u3002\u6211\u4eec\u5728\u6559\u5e08\u7684\u8f68\u8ff9\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u8fd9\u4e9b\u8f68\u8ff9\u662f\u751f\u6210\u7684 token \u7684\u5b8c\u6574\u5e8f\u5217\uff0c\u5305\u62ec\u4e2d\u95f4\u7684\u601d\u8003\u6b65\u9aa4\u3002<\/p>\n\n\n\n<p>\u5728\u6bcf\u4e00\u6b65\uff0c\u6211\u4eec\u65e2\u53ef\u4ee5\u4f7f\u7528\u6559\u5e08\u5b8c\u6574\u7684\u300c\u4e0b\u4e00\u4e2a token \u5206\u5e03\u300d\uff08\u5e38\u88ab\u79f0\u4e3a \u300clogit \u84b8\u998f\u300d\uff09\uff0c\u4e5f\u53ef\u4ee5\u53ea\u91c7\u6837\u7ed9\u5b9a\u7684\u5e8f\u5217\u3002\u5b9e\u8df5\u8bc1\u660e\uff0c\u91c7\u6837\u5e8f\u5217\u63d0\u4f9b\u4e86\u5bf9\u6559\u5e08\u5206\u5e03\u7684\u65e0\u504f\u4f30\u8ba1\uff0c\u5e76\u80fd\u8fbe\u5230\u76f8\u540c\u7684\u76ee\u6807\u3002\u5b66\u751f\u6a21\u578b\u4f1a\u6839\u636e\u81ea\u5df1\u751f\u6210\u8be5 token \u7684\u6982\u7387\u6709\u591a\u4f4e\uff0c\u6765\u76f8\u5e94\u5730\u66f4\u65b0\u5bf9\u5e8f\u5217\u4e2d\u6bcf\u4e2a token \u7684\u5b66\u4e60\uff08\u5728\u4e0b\u4f8b\u4e2d\u7528\u6df1\u8272\u8868\u793a\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_9ea6684c94984954a91b445cac713bf7@5091053_oswg36798oswg1060oswg165_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4e8b\u5b9e\u8bc1\u660e\uff0c\u84b8\u998f\u5927\u578b\u6559\u5e08\u6a21\u578b\uff0c\u5728\u8bad\u7ec3\u5c0f\u578b\u6a21\u578b\u65b9\u9762\u975e\u5e38\u6709\u6548\uff0c\u4f7f\u5176\u80fd\u591f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u9075\u5faa\u6307\u4ee4<\/p>\n\n\n\n<p>\u8fdb\u884c\u6570\u5b66\u548c\u79d1\u5b66\u63a8\u7406<\/p>\n\n\n\n<p>\u4ece\u533b\u7597\u7b14\u8bb0\u4e2d\u63d0\u53d6\u4e34\u5e8a\u4fe1\u606f<\/p>\n\n\n\n<p>\u4ee5\u53ca\u53c2\u4e0e\u591a\u8f6e\u804a\u5929\u5bf9\u8bdd<\/p>\n<\/blockquote>\n\n\n\n<p>\u7528\u4e8e\u8fd9\u4e9b\u5e94\u7528\u548c\u5176\u4ed6\u5e94\u7528\u7684\u84b8\u998f\u6570\u636e\u96c6\u901a\u5e38\u662f\u5f00\u6e90\u548c\u516c\u5f00\u53d1\u5e03\u7684\u3002<\/p>\n\n\n\n<p>\u79bb\u7b56\u7565\u8bad\u7ec3\u7684\u7f3a\u70b9\u662f\uff0c\u5b66\u751f\u662f\u5728\u6559\u5e08\u7ecf\u5e38\u9047\u5230\u7684\u4e0a\u4e0b\u6587\u4e2d\u5b66\u4e60\uff0c\u800c\u4e0d\u662f\u5728\u5b66\u751f\u81ea\u5df1\u5c06\u6765\u4f1a\u7ecf\u5e38\u9047\u5230\u7684\u4e0a\u4e0b\u6587\u4e2d\u5b66\u4e60\u3002<\/p>\n\n\n\n<p>\u8fd9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u590d\u5408\u9519\u8bef\uff08compounding error\uff09\uff1a\u5982\u679c\u5b66\u751f\u65e9\u671f\u72af\u4e86\u4e00\u4e2a\u6559\u5e08\u4ece\u672a\u72af\u8fc7\u7684\u9519\u8bef\uff0c\u5b83\u4f1a\u53d1\u73b0\u81ea\u5df1\u8d8a\u6765\u8d8a\u504f\u79bb\u5728\u8bad\u7ec3\u4e2d\u89c2\u5bdf\u5230\u7684\u72b6\u6001\u3002<\/p>\n\n\n\n<p>\u5f53\u6211\u4eec\u5173\u5fc3\u5b66\u751f\u5728\u957f\u5e8f\u5217\u4e0a\u7684\u8868\u73b0\u65f6\uff0c\u8fd9\u4e2a\u95ee\u9898\u53d8\u5f97\u5c24\u4e3a\u7a81\u51fa\u3002\u4e3a\u4e86\u907f\u514d\u8fd9\u79cd\u504f\u79bb\uff0c\u5b66\u751f\u5fc5\u987b\u5b66\u4f1a\u4ece\u81ea\u5df1\u7684\u9519\u8bef\u4e2d\u6062\u590d\u3002<\/p>\n\n\n\n<p>\u79bb\u7b56\u7565\u84b8\u998f\u89c2\u5bdf\u5230\u7684\u53e6\u4e00\u4e2a\u95ee\u9898\u662f\uff0c\u5b66\u751f\u53ef\u4ee5\u5b66\u4f1a\u6a21\u4eff\u6559\u5e08\u7684\u98ce\u683c\u548c\u81ea\u4fe1\uff0c\u4f46\u4e0d\u4e00\u5b9a\u80fd\u5b66\u4f1a\u5176\u4e8b\u5b9e\u7684\u51c6\u786e\u6027\u3002<\/p>\n\n\n\n<p>\u6253\u4e2a\u6bd4\u65b9\uff1a\u5982\u679c\u4f60\u5728\u5b66\u4e60\u56fd\u9645\u8c61\u68cb\uff0c\u5728\u7b56\u7565 RL \u5c31\u597d\u6bd4\u5728\u6ca1\u6709\u6559\u7ec3\u6307\u5bfc\u7684\u60c5\u51b5\u4e0b\u81ea\u5df1\u4e0b\u68cb\u3002\u8d62\u68cb\u6216\u8f93\u68cb\u7684\u53cd\u9988\u4e0e\u4f60\u81ea\u5df1\u7684\u4e0b\u6cd5\u76f4\u63a5\u76f8\u5173\uff0c\u4f46\u6bcf\u5c40\u53ea\u6536\u5230\u4e00\u6b21\u53cd\u9988\uff0c\u800c\u4e14\u4e0d\u4f1a\u544a\u8bc9\u4f60\u54ea\u4e9b\u68cb\u6b65\u5bf9\u7ed3\u679c\u8d21\u732e\u6700\u5927\u3002\u79bb\u7b56\u7565\u84b8\u998f\u5219\u7c7b\u4f3c\u4e8e\u89c2\u770b\u4e00\u4f4d\u7279\u7ea7\u5927\u5e08\u4e0b\u68cb \u2014\u2014 \u4f60\u89c2\u5bdf\u5230\u7684\u662f\u975e\u5e38\u9ad8\u8d85\u7684\u68cb\u6b65\uff0c\u4f46\u8fd9\u4e9b\u68cb\u6b65\u662f\u5728\u65b0\u624b\u73a9\u5bb6\u5f88\u5c11\u4f1a\u9047\u5230\u7684\u68cb\u5c40\u72b6\u6001\u4e0b\u8d70\u51fa\u7684\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5e0c\u671b\u80fd\u5c06 RL \u7684\u5728\u7b56\u7565\u76f8\u5173\u6027\u4e0e\u84b8\u998f\u7684\u5bc6\u96c6\u5956\u52b1\u4fe1\u53f7\u7ed3\u5408\u8d77\u6765\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u5b66\u4e60\u56fd\u9645\u8c61\u68cb\u6765\u8bf4\uff0c\u8fd9\u5c31\u597d\u6bd4\u6709\u4e00\u4f4d\u8001\u5e08\u6765\u4e3a\u4f60\u81ea\u5df1\u7684\u6bcf\u4e00\u6b65\u68cb\u6253\u5206\uff0c\u4ece\u300c\u5927\u9519\u7279\u9519\u300d\u5230\u300c\u5999\u4e0d\u53ef\u8a00\u300d\u3002\u5bf9\u4e8e LLM \u7684\u540e\u8bad\u7ec3\u6765\u8bf4\uff0c\u8fd9\u5c31\u662f\u5728\u7b56\u7565\u84b8\u998f\uff08on-policy distillation\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_1248c9c7d05f4740ac6e13e33e934c1c@5091053_oswg674184oswg1080oswg641_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u5728\u7b56\u7565\u84b8\u998f \u2014\u2014 \u96c6\u4e24\u8005\u4e4b\u957f<\/strong><\/h2>\n\n\n\n<p>\u5728\u7b56\u7565\u84b8\u998f\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a\u4ece\u5b66\u751f\u6a21\u578b\u4e2d\u91c7\u6837\u8f68\u8ff9\uff0c\u5e76\u4f7f\u7528\u4e00\u4e2a\u9ad8\u6027\u80fd\u7684\u6559\u5e08\u6a21\u578b\u6765\u4e3a\u6bcf\u4e2a\u8f68\u8ff9\u7684\u6bcf\u4e00\u4e2a token \u8bc4\u5206\u3002<\/p>\n\n\n\n<p>\u56de\u5230\u6211\u4eec\u4e0a\u9762\u7684\u6570\u5b66\u4f8b\u5b50\uff0c\u5728\u7b56\u7565\u84b8\u998f\u4f1a\u7ed9\u89e3\u9898\u7684\u6bcf\u4e00\u6b65\u6253\u5206\uff0c\u60e9\u7f5a\u90a3\u4e9b\u5bfc\u81f4\u5b66\u751f\u5f97\u51fa\u9519\u8bef\u7b54\u6848\u7684\u9519\u8bef\u6b65\u9aa4\uff0c\u540c\u65f6\u5f3a\u5316\u90a3\u4e9b\u6267\u884c\u6b63\u786e\u7684\u6b65\u9aa4\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_2624d349a5ac40ae820b41169afe43a2@5091053_oswg32257oswg1026oswg171_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u63a2\u8ba8\u4e86\u5728\u7b56\u7565\u84b8\u998f\u5728\u4ee5\u4e0b\u4efb\u52a1\u4e0a\u7684\u5e94\u7528\uff1a<\/p>\n\n\n\n<p>1. \u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u6570\u5b66\u63a8\u7406\u3002<\/p>\n\n\n\n<p>2. \u8bad\u7ec3\u4e00\u4e2a\u517c\u5177\u9886\u57df\u77e5\u8bc6\u548c\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u7684\u52a9\u624b\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5728\u5df2\u7ecf\u5177\u5907\u9884\u8bad\u7ec3\u548c\u4e2d\u8bad\u7ec3\u57fa\u7840\u80fd\u529b\u7684\u6a21\u578b\u4e0a\u5e94\u7528\u5728\u7b56\u7565\u84b8\u998f\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u8fd9\u662f\u4e00\u79cd\u5ec9\u4ef7\u800c\u5f3a\u5927\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5b83\u6210\u529f\u5c06\u5728\u7b56\u7565\u8bad\u7ec3\u7684\u4f18\u52bf\u548c\u5bc6\u96c6\u5956\u52b1\u4fe1\u53f7\u7ed3\u5408\u5230\u4e86\u4e00\u8d77\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_4eb7f19f9d824739b8e4338a8081d133@5091053_oswg34393oswg785oswg283_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6211\u4eec\u7684\u5728\u7b56\u7565\u84b8\u998f\u5de5\u4f5c\u501f\u9274\u4e86 DAGGER\uff08Ross et al, 2010\uff09\uff0c\u8fd9\u662f\u4e00\u79cd\u8fed\u4ee3\u5f0f\u7684 SFT \u7b97\u6cd5\uff0c\u5b83\u5305\u542b\u4e86\u6559\u5e08\u5bf9\u5b66\u751f\u8bbf\u95ee\u8fc7\u7684\u72b6\u6001\u7684\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p>\u5b83\u4e5f\u7c7b\u4f3c\u4e8e\u8fc7\u7a0b\u5956\u52b1\u5efa\u6a21\uff08Lightman et al, 2023\uff09\uff0c\u8fd9\u662f\u4e00\u79cd RL \u65b9\u6cd5\uff0c\u4f1a\u5bf9\u5b66\u751f\u6a21\u578b\u601d\u7ef4\u94fe\u4e2d\u7684\u6bcf\u4e00\u6b65\u90fd\u8fdb\u884c\u8bc4\u5206\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u6269\u5c55\u4e86 Agarwal et al.\uff082023\uff09\u548c Qwen3 \u56e2\u961f\uff082025\uff09\u4e4b\u524d\u7684\u5728\u7b56\u7565\u84b8\u998f\u5de5\u4f5c\u3002\u4f7f\u7528 Tinker \u8bad\u7ec3 API\uff0c\u6211\u4eec\u590d\u523b\u4e86 Qwen3 \u7684\u6210\u679c\uff0c\u5373\u901a\u8fc7\u5728\u7b56\u7565\u84b8\u998f\u5728\u63a8\u7406\u57fa\u51c6\u4e0a\u5b9e\u73b0\u4e86\u540c\u7b49\u6027\u80fd\uff0c\u800c\u6210\u672c\u4ec5\u4e3a RL \u7684\u4e00\u5c0f\u90e8\u5206\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u5b9e\u73b0<\/strong><\/h2>\n\n\n\n<p>\u4f60\u53ef\u4ee5\u5728\u8fd9\u4e2a Tinker cookbook \u4e2d\u8ddf\u7740\u5b66\u4e60\u5b9e\u73b0\u7684\u6bcf\u4e00\u6b65\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-embed\"><div class=\"wp-block-embed__wrapper\">\nhttps:\/\/github.com\/thinking-machines-lab\/tinker-cookbook\/tree\/main\/tinker_cookbook\/recipes\/distillation\n<\/div><\/figure>\n\n\n\n<p><strong>\u635f\u5931\u51fd\u6570\uff1a\u53cd\u5411 KL<\/strong><\/p>\n\n\n\n<p>\u5728\u7b56\u7565\u84b8\u998f\u53ef\u4ee5\u4f7f\u7528\u591a\u79cd\u635f\u5931\u51fd\u6570\u6765\u4e3a\u5b66\u751f\u7684\u8f68\u8ff9\u8bc4\u5206\u3002\u4e3a\u7b80\u5355\u8d77\u89c1\uff0c\u6211\u4eec\u9009\u62e9\u9010 token \u7684\u53cd\u5411 KL\uff08reverse KL\uff09\u2014\u2014 \u5373\u5728\u7ed9\u5b9a\u76f8\u540c\u5148\u524d\u8f68\u8ff9\u7684\u6761\u4ef6\u4e0b\uff0c\u5b66\u751f\uff08\u03c0_\u03b8\uff09\u548c\u6559\u5e08\uff08\u03c0_teacher\uff09\u5728\u6bcf\u4e2a token \u4e0a\u7684\u5206\u5e03\u4e4b\u95f4\u7684\u6563\u5ea6\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_bbd2604dcece4893bc5175792417f24c@5091053_oswg11756oswg824oswg74_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6211\u4eec\u7684\u5956\u52b1\u51fd\u6570\u4f1a\u6700\u5c0f\u5316\u53cd\u5411 KL\uff0c\u8fd9\u4f1a\u4fc3\u4f7f\u5b66\u751f\u5728\u81ea\u5df1\u6240\u5904\u7684\u6bcf\u79cd\u72b6\u6001\u4e0b\u90fd\u53bb\u8fd1\u4f3c\u6559\u5e08\u7684\u884c\u4e3a\u3002\u5f53\u5b66\u751f\u7684\u884c\u4e3a\u4e0e\u6559\u5e08\u5b8c\u5168\u4e00\u81f4\u65f6\uff0c\u53cd\u5411 KL \u4e3a\u96f6\u3002\u4e3a\u7b80\u5355\u8d77\u89c1\uff0c\u6211\u4eec\u4f7f\u7528\u7684\u6298\u6263\u56e0\u5b50\u4e3a\u96f6\uff1a\u5728\u4efb\u4f55\u7ed9\u5b9a\u7684\u65f6\u95f4\u6b65\uff0c\u5b66\u751f\u53ea\u4f18\u5316\u773c\u524d\u7684\u4e0b\u4e00\u4e2a token\uff0c\u4e0d\u8003\u8651\u672a\u6765\u7684 token\u3002<\/p>\n\n\n\n<p>\u53cd\u5411 KL \u4e0e RL \u6709\u7740\u5929\u7136\u7684\u534f\u540c\u4f5c\u7528\uff0cRL \u901a\u5e38\u4f18\u5316\u7531\u5956\u52b1\u6a21\u578b\u5f15\u5bfc\u7684\u67d0\u79cd\u5e8f\u5217\u7ea7\u53cd\u5411 KL\u3002\u7136\u800c\uff0c\u4e0e\u5b9e\u8df5\u4e2d\u7684\u5927\u591a\u6570\u5956\u52b1\u6a21\u578b\u4e0d\u540c\uff0c\u53cd\u5411 KL \u662f\u300c\u4e0d\u53ef\u7834\u89e3\u7684\u300d\uff08unhackable\uff09\uff0c\u56e0\u4e3a\u4ece\u6559\u5e08\u6a21\u578b\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u4f4e KL \u603b\u662f\u5bf9\u5e94\u7740\u9ad8\u6982\u7387\u7684\u671f\u671b\u884c\u4e3a\u3002\u53cd\u5411 KL \u7684\u53e6\u4e00\u4e2a\u6709\u7528\u7279\u6027\u662f\u5b83\u662f\u300c\u5bfb\u627e\u4f17\u6570\uff08mode seeking\uff09\u300d\u7684 \u2014\u2014 \u5b83\u5b66\u4e60\u4e00\u79cd\u7279\u5b9a\u884c\u4e3a\uff08\u6559\u5e08\u7684\u884c\u4e3a\uff09\uff0c\u800c\u4e0d\u662f\u5c06\u5176\u5206\u5e03\u5206\u6563\u5728\u51e0\u4e2a\u6b21\u4f18\u9009\u9879\u4e0a\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u65b9\u6cd5\u53ef\u8282\u7701\u5927\u91cf\u8ba1\u7b97\u8d44\u6e90\u3002\u56e0\u4e3a\u5b83\u4e0d\u9700\u8981\u7b49\u5f85\u4e00\u4e2a\u8f68\u8ff9\u5b8c\u6210\u91c7\u6837\u624d\u80fd\u8ba1\u7b97\u5956\u52b1\uff0c\u6240\u4ee5\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u66f4\u77ed\u6216\u90e8\u5206\u7684\u8f68\u8ff9\u8fdb\u884c\u8bad\u7ec3\u3002\u67e5\u8be2\u6559\u5e08\u7684\u5bf9\u6570\u6982\u7387\u4e5f\u53ea\u9700\u8981\u5927\u578b\u6a21\u578b\u8fdb\u884c\u4e00\u6b21\u524d\u5411\u4f20\u64ad\uff0c\u800c\u8f68\u8ff9\u5219\u662f\u7531\u66f4\u5c0f\u3001\u66f4\u5ec9\u4ef7\u7684\u5b66\u751f\u6a21\u578b\u751f\u6210\u7684\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u4e5f\u4e0d\u9700\u8981\u5355\u72ec\u7684\u5956\u52b1\u6216\u6807\u6ce8\u6a21\u578b\u3002\u5c06\u57fa\u4e8e\u84b8\u998f\u7684\u9010 token \u5956\u52b1\u4e0e\u5e8f\u5217\u7ea7\u7684\u73af\u5883\u5956\u52b1\u7ed3\u5408\u8d77\u6765\u53ef\u80fd\u4f1a\u6709\u597d\u5904\uff1b\u8fd9\u662f\u672a\u6765\u4e00\u4e2a\u6709\u8da3\u7684\u6f5c\u5728\u7814\u7a76\u9886\u57df\u3002<\/p>\n\n\n\n<p><strong>\u56fe\u89e3<\/strong><\/p>\n\n\n\n<p>\u4e0b\u9762\u6211\u4eec\u6765\u770b\u4e00\u4e2a\u771f\u5b9e\u7684\u4f8b\u5b50\uff0c\u8fd9\u662f\u4e00\u4e2a\u9519\u8bef\u7684\u5b66\u751f\u8f68\u8ff9\uff0c\u7531\u6559\u5e08\u6a21\u578b\u8fdb\u884c\u8bc4\u5206\u3002\u8fd9\u4e2a\u4f8b\u5b50\u6765\u81ea SimpleBench\uff0c\u5b83\u8981\u6c42\u6a21\u578b\u505a\u51fa\u4e00\u4e2a\u5173\u952e\u89c2\u5bdf\uff1a\u95ee\u9898\u7684\u524d\u63d0\u5f88\u91cd\u8981\u3002\u6b63\u786e\u7b54\u6848\u662f \u300cB. 0\u300d\uff0c\u56e0\u4e3a\u51b0\u5757\u5728\u714e\u9505\u91cc\u4f1a\u878d\u5316\u3002\u800c\u5b66\u751f\u6a21\u578b\uff08Qwen3-4B-Instruct-2507\uff09\u9519\u8bef\u5730\u5c06\u5176\u89c6\u4e3a\u4e00\u4e2a\u7eaf\u7cb9\u7684\u6570\u5b66\u95ee\u9898\uff0c\u6ca1\u6709\u8003\u8651\u7269\u7406\u80cc\u666f\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_11051dec820f4009ade6c8b1aacc7ab8@5091053_oswg150828oswg1080oswg543_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u989c\u8272\u8d8a\u6df1\uff0c\u4ee3\u8868\u8be5 token \u53d7\u5230\u6559\u5e08\u6a21\u578b\uff08Qwen3-235B-A22B-Instruct-2507\uff09\u7684\u60e9\u7f5a\u8d8a\u9ad8\uff08\u6559\u5e08\u6a21\u578b\u6b63\u786e\u89e3\u51b3\u4e86\u8fd9\u4e2a\u95ee\u9898\uff09\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u770b\u5230\uff0c\u5b83\u60e9\u7f5a\u4e86\u90a3\u4e9b\u5f15\u5bfc\u5b66\u751f\u8bef\u5165\u6b67\u9014\u7684\u77ed\u8bed\u7684\u8d77\u59cb token\uff0c\u8fd9\u76f4\u89c2\u4e0a\u5bf9\u5e94\u4e86\u5f15\u5bfc\u63a8\u7406\u7684\u91cd\u8981\u300c\u5206\u53c9 token\u300d\uff08forking tokens\uff09\u3002\u6700\u7ec8\u7b54\u6848\uff08\u867d\u7136\u662f\u9519\u7684\uff09\u5e76\u6ca1\u6709\u53d7\u5230\u60e9\u7f5a \u2014\u2014 \u56e0\u4e3a\u5728\u7ed9\u5b9a\u524d\u9762\u6240\u6709\u5e8f\u5217\u7684\u6761\u4ef6\u4e0b\uff0c\u8fd9\u4e2a\u7b54\u6848\u662f\u5b8c\u5168\u53ef\u9884\u6d4b\u7684\u3002<\/p>\n\n\n\n<p><strong>\u4f2a\u4ee3\u7801<\/strong><\/p>\n\n\n\n<p>\u6211\u4eec\u5728 Tinker \u7684 RL \u811a\u672c\u4e4b\u4e0a\u5b9e\u73b0\u4e86\u5728\u7b56\u7565\u84b8\u998f\uff0c\u8be5\u811a\u672c\u5df2\u7ecf\u5b9e\u73b0\u4e86\u91c7\u6837\u3001\u5956\u52b1\u8ba1\u7b97\u548c\u7b56\u7565\u68af\u5ea6\u5f0f\u7684\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u5730\u5740\uff1ahttps:\/\/thinkingmachines.ai\/blog\/on-policy-distillation\/(https:\/\/github.com\/thinking-machines-lab\/tinker-cookbook\/blob\/main\/tinker_cookbook\/rl\/train.py)<\/p>\n\n\n\n<p><strong>\u521d\u59cb\u5316\u6559\u5e08\u5ba2\u6237\u7aef<\/strong>\uff1aTinker API \u53ef\u4ee5\u8f7b\u677e\u5730\u4e3a\u4e0d\u540c\u6a21\u578b\u521b\u5efa\u4e0d\u540c\u7684\u5ba2\u6237\u7aef\u3002\u6211\u4eec\u4f7f\u7528\u91c7\u6837\u5ba2\u6237\u7aef\uff0c\u56e0\u4e3a\u6211\u4eec\u4e0d\u9700\u8981\u901a\u8fc7\u6559\u5e08\u6a21\u578b\u4f20\u64ad\u5bf9\u6570\u6982\u7387\u3002<\/p>\n\n\n\n<p><strong>\u91c7\u6837\u8f68\u8ff9<\/strong>\uff1a\u6211\u4eec\u50cf\u5728 RL \u4e2d\u4e00\u6837\u4ece\u5b66\u751f\u6a21\u578b\u4e2d\u91c7\u6837\u8f68\u8ff9\u3002\u5728\u91c7\u6837\u671f\u95f4\uff0cRL \u5df2\u7ecf\u8ba1\u7b97\u4e86\u5b66\u751f\u7684\u5bf9\u6570\u6982\u7387 log \u03c0_\u03b8(x\uff09\uff0c\u7528\u4f5c\u91cd\u8981\u6027\u91c7\u6837\u635f\u5931\u7684\u4e00\u90e8\u5206\u3002<\/p>\n\n\n\n<p><strong>\u8ba1\u7b97\u5956\u52b1<\/strong>\uff1a\u6211\u4eec\u7528 compute_logprobs \u51fd\u6570\u5728\u91c7\u6837\u51fa\u7684\u8f68\u8ff9\u4e0a\u67e5\u8be2\u6559\u5e08\u5ba2\u6237\u7aef\uff0c\u5b83\u4f1a\u8fd4\u56de\u6559\u5e08\u5728\u5b66\u751f\u91c7\u6837\u7684 token x \u4e0a\u7684\u5bf9\u6570\u6982\u7387 log \u03c0_teacher (x\uff09\u3002\u7136\u540e\u6211\u4eec\u7528\u8fd9\u4e2a\u6765\u8ba1\u7b97\u53cd\u5411 KL\u3002<\/p>\n\n\n\n<p><strong>\u4f7f\u7528 RL \u8fdb\u884c\u8bad\u7ec3<\/strong>\uff1a\u6211\u4eec\u5c06\u9010 token \u7684\u4f18\u52bf\uff08advantage\uff09\u8bbe\u7f6e\u4e3a\u8d1f\u7684\u53cd\u5411 KL\uff0c\u5e76\u8c03\u7528 RL \u7684\u91cd\u8981\u6027\u91c7\u6837\u635f\u5931\u51fd\u6570\u6765\u5bf9\u5b66\u751f\u6a21\u578b\u6267\u884c\u8bad\u7ec3\u66f4\u65b0\u3002<\/p>\n\n\n\n<p>\u4f2a\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_71b7073426b648c0b9d2a175c883326c@5091053_oswg83648oswg879oswg576_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u5728\u4e0b\u9762\u7684\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u901a\u5e38\u5c06\u5728\u7b56\u7565\u84b8\u998f\u5e94\u7528\u4e8e\u5df2\u7ecf\u8fc7\u7279\u5b9a\u9886\u57df\u77e5\u8bc6\u4e2d\u8bad\u7ec3\u7684\u6a21\u578b\u3002\u8fd9\u79cd\u8bad\u7ec3\u63d0\u9ad8\u4e86\u5b66\u751f\u751f\u6210\u6559\u5e08\u5206\u5e03\u8303\u56f4\u5185\u7684 token \u7684\u6982\u7387\uff0c\u5c3d\u7ba1\u8fd9\u901a\u5e38\u8fdc\u4e0d\u8db3\u4ee5\u590d\u523b\u6559\u5e08\u7684\u6027\u80fd\u3002\u901a\u5e38\uff0c\u6b63\u5982\u6211\u4eec\u5c06\u5728\u4e2a\u6027\u5316\u793a\u4f8b\u4e2d\u770b\u5230\u7684\uff0c\u751f\u6210\u76f8\u5173 token \u7684\u6982\u7387\u5f00\u59cb\u65f6\u4e3a\u96f6\uff0c\u56e0\u4e3a\u5b66\u751f\u7f3a\u4e4f\u4efb\u4f55\u76f8\u5173\u7684\u9886\u57df\u77e5\u8bc6\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5c06\u4f7f\u7528\u5728\u7b56\u7565\u84b8\u998f\u8fdb\u884c\u540e\u8bad\u7ec3\uff0c\u5e76\u5c06\u5176\u4e0e\u8bad\u7ec3\u4e13\u5bb6\u6a21\u578b\u7684\u5176\u4ed6\u6700\u540e\u5173\u952e\u9636\u6bb5\u7684\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u84b8\u998f\u4ee5\u83b7\u5f97\u63a8\u7406\u80fd\u529b<\/strong><\/h2>\n\n\n\n<p>\u6211\u4eec\u4f7f\u7528\u84b8\u998f\u6765\u8bad\u7ec3 Qwen3-8B-Base \u6a21\u578b\u7684\u6570\u5b66\u63a8\u7406\u80fd\u529b\uff0c\u5e76\u4f7f\u7528 Qwen3-32B \u4f5c\u4e3a\u6559\u5e08\u6a21\u578b\u3002\u6559\u5e08\uff08Qwen3-32B\uff09\u548c\u5b66\u751f\uff08Qwen3-8B-Base\uff09\u90fd\u662f\u76ee\u524d Tinker \u4e0a\u652f\u6301\u7684\u6a21\u578b\uff0c\u56e0\u6b64\u4f60\u53ef\u4ee5\u4f7f\u7528 Tinker cookbook \u590d\u73b0\u6211\u4eec\u7684\u5b9e\u9a8c\u3002<\/p>\n\n\n\n<p><strong>\u79bb\u7b56\u7565\u84b8\u998f<\/strong><\/p>\n\n\n\n<p>\u5982\u524d\u6240\u8ff0\uff0c\u6211\u4eec\u6240\u6709\u7684\u5b9e\u9a8c\u90fd\u4ee5\u79bb\u7b56\u7565\u84b8\u998f\uff08\u5373\u5728\u6559\u5e08\u751f\u6210\u7684\u793a\u4f8b\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff09\u7684\u5f62\u5f0f\u4f5c\u4e3a\u4e2d\u8bad\u7ec3\u7684\u8d77\u70b9\u3002\u7528\u4e8e\u6570\u5b66\u63a8\u7406\u7684\u6570\u636e\u96c6\u662f OpenThoughts-3\uff0c\u8fd9\u662f\u4e00\u4e2a\u7531 QwQ-32B\uff08\u4e00\u4e2a\u7c7b\u4f3c\u4e8e Qwen3-32B \u7684\u63a8\u7406\u6a21\u578b\uff09\u751f\u6210\u7684\u63a8\u7406\u63d0\u793a\u548c\u54cd\u5e94\u7684\u96c6\u5408\u3002<\/p>\n\n\n\n<p>\u5728 40 \u4e07\u4e2a\u63d0\u793a\u4e0a\u5bf9\u5b66\u751f\u6a21\u578b\uff08Qwen3-8B-Base\uff09\u8fdb\u884c\u5168\u53c2\u6570\u5fae\u8c03\uff08full fine-tuning\uff09\uff0c\u5728 AIME&#8217;24\uff08\u4e00\u4e2a\u6570\u5b66\u95ee\u9898\u57fa\u51c6\u6d4b\u8bd5\uff09\u4e0a\u83b7\u5f97\u4e86 60% \u7684\u5206\u6570\u3002\u6211\u4eec\u4e5f\u53ef\u4ee5\u4f7f\u7528 LoRA \u8fdb\u884c\u8bad\u7ec3\uff0c\u4f46\u5728\u9ad8\u5bb9\u91cf\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u65f6\uff0c\u5b83\u843d\u540e\u4e8e\u5168\u53c2\u6570\u5fae\u8c03\u3002\u5728\u6240\u6709\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u90fd\u770b\u5230\u6027\u80fd\u5448\u5bf9\u6570\u7ebf\u6027\u589e\u957f \u2014\u2014 \u6700\u521d\u7684\u6027\u80fd\u63d0\u5347\u5f88\u5ec9\u4ef7\uff0c\u4f46\u540e\u671f\u7684\u63d0\u5347\u6210\u672c\u9ad8\u6602\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_8fd93bebd3df4564887cb25e680ea030@5091053_oswg97985oswg1080oswg640_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6211\u4eec\u53ef\u4ee5\u5c06\u5728 40 \u4e07\u4e2a\u63d0\u793a\u4e0a\u5fae\u8c03\u8fc7\u7684\u6a21\u578b\u89c6\u4e3a\u4e00\u4e2a\u68c0\u67e5\u70b9\uff0c\u7136\u540e\u5c1d\u8bd5\u5404\u79cd\u540e\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5c06\u5176\u5728 AIME\u201924 \u57fa\u51c6\u4e0a\u7684\u5206\u6570\u4ece 60% \u63d0\u9ad8\u5230 70%\u3002<\/p>\n\n\n\n<p>\u9ed8\u8ba4\u65b9\u6cd5\u662f\u5728\u66f4\u591a\u63d0\u793a\u4e0a\u8fdb\u884c\u5fae\u8c03\uff0c\u5373\u7ee7\u7eed\u79bb\u7b56\u7565\u84b8\u998f\u7684\u8fc7\u7a0b\u3002\u6839\u636e\u5bf9\u6570\u7ebf\u6027\u8d8b\u52bf\u63a8\u65ad\uff0c\u6211\u4eec\u4f30\u8ba1\u6a21\u578b\u5728\u5927\u7ea6 200 \u4e07\u4e2a\u63d0\u793a\u4e0a\u80fd\u8fbe\u5230 70% \u7684\u5206\u6570\u3002\u8fd9\u4e2a\u63a8\u65ad\u9700\u8981 scaling law \u6301\u7eed\u6709\u6548\u800c\u4e0d\u505c\u6ede\uff0c\u8fd9\u5e76\u4e0d\u7b80\u5355\u3002<\/p>\n\n\n\n<p><strong>\u5f3a\u5316\u5b66\u4e60<\/strong><\/p>\n\n\n\n<p>Qwen3 \u6280\u672f\u62a5\u544a\u79f0\uff0c\u5728\u7c7b\u4f3c\u7684 SFT \u521d\u59cb\u5316\u57fa\u7840\u4e0a\uff0c\u901a\u8fc7 17,920 \u4e2a GPU \u5c0f\u65f6\u7684 RL\uff0c\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u8fbe\u5230\u4e86 67.6% \u7684\u6027\u80fd\u3002\u8fd9\u5f88\u96be\u4e0e\u84b8\u998f\u7684\u6210\u672c\u76f4\u63a5\u6bd4\u8f83\uff0c\u4f46\u57fa\u4e8e\u5bf9 SFT \u8bad\u7ec3\u5806\u6808\u7684\u4e00\u4e9b\u5408\u7406\u5047\u8bbe\uff0c\u8fd9\u4e0e\u5728 200 \u4e07\u4e2a\u79bb\u7b56\u7565\u84b8\u998f\u63d0\u793a\u4e0a\u8bad\u7ec3\u7684\u6210\u672c\u76f8\u4f3c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_736cab8417c84fb29118f91d0b0315ce@5091053_oswg39892oswg1032oswg287_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p><strong>Qwen \u56e2\u961f\u8fd8\u62a5\u544a\u79f0\uff0c\u4f7f\u7528\u5728\u7b56\u7565\u84b8\u998f\uff0c\u80fd\u4ee5 RL \u6210\u672c\u7684\u5341\u5206\u4e4b\u4e00\uff0c\u5728 AIME\u201924 \u4e0a\u8fbe\u5230\u4e86 74.4% \u7684\u66f4\u9ad8\u5206\u6570\u3002\u8fd9\u4e5f\u542f\u53d1\u4e86\u6211\u4eec\u7684\u5de5\u4f5c\u3002<\/strong><\/p>\n\n\n\n<p><strong>\u5728\u7b56\u7565\u84b8\u998f<\/strong><\/p>\n\n\n\n<p>\u4f5c\u4e3a\u66ff\u4ee3\u65b9\u6848\uff0c\u6211\u4eec\u8fd0\u884c\u4e86\u5728\u7b56\u7565\u84b8\u998f\u3002\u4ece 40 \u4e07 SFT \u68c0\u67e5\u70b9\u5f00\u59cb\uff0c\u5728\u7b56\u7565\u84b8\u998f\u5728\u5927\u7ea6 150 \u4e2a\u6b65\u9aa4\u5185\u5c31\u8fbe\u5230\u4e86 AIME\u201924 70% \u7684\u6210\u7ee9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_615cec51e4d341b0a5818d51f14e3d8c@5091053_oswg147891oswg1080oswg447_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u8de8\u65b9\u6cd5\u6bd4\u8f83\u8ba1\u7b97\u6210\u672c\u5e76\u975e\u6613\u4e8b\u3002\u4e0b\u9762\uff0c\u6211\u4eec\u7528 FLOPs\uff08\u6d6e\u70b9\u8fd0\u7b97\u6b21\u6570\uff09\u6765\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_bd645dd375274c97a92cb728c69dbfd9@5091053_oswg101039oswg1080oswg531_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6211\u4eec\u53d1\u73b0\uff0c\u5f53 SFT \u6570\u636e\u96c6\u662f\u73b0\u6210\u7684\uff08\u5982\u6211\u4eec\u7684 OpenThoughts-3 \u793a\u4f8b\uff09\u6216\u5728\u591a\u6b21\u8bad\u7ec3\u4e2d\u88ab\u644a\u9500\u65f6\uff0c\u57fa\u7ebf\u6210\u672c\u964d\u4f4e\u4e86 9 \u500d\u3002<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u6211\u4eec\u7ecf\u5e38\u5e0c\u671b\u5728\u4e00\u4e2a\u6ca1\u6709\u73b0\u6210\u79bb\u7b56\u7565\u84b8\u998f\u6570\u636e\u96c6\u7684\u65b0\u4efb\u52a1\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u5c0f\u6a21\u578b\u3002\u5982\u679c\u6211\u4eec\u5c06\u6559\u5e08\u6a21\u578b\u7684\u5168\u90e8\u6210\u672c\uff08\u5373\u5305\u62ec\u4ece\u6559\u5e08\u6a21\u578b\u91c7\u6837\u7684\u989d\u5916\u6210\u672c\uff09\u8ba1\u7b97\u5728\u79bb\u7b56\u7565\u84b8\u998f\u4e2d\uff0c\u90a3\u4e48\u603b\u6210\u672c\u53ef\u964d\u4f4e\u7ea6 30 \u500d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u7528\u4e8e\u4e2a\u6027\u5316\u7684\u84b8\u998f<\/strong><\/h2>\n\n\n\n<p>\u9664\u4e86\u5c06\u5c0f\u578b\u6a21\u578b\u8bad\u7ec3\u5230\u5728\u901a\u7528\u4efb\u52a1\u4e0a\u8868\u73b0\u51fa\u8272\u4e4b\u5916\uff0c\u84b8\u998f\u7684\u53e6\u4e00\u4e2a\u7528\u4f8b\u662f\u4e2a\u6027\u5316\u3002\u4f8b\u5b50\u5305\u62ec\u5728\u5bf9\u8bdd\u4e2d\u9075\u5faa\u7279\u5b9a\u7684\u8bed\u6c14\u548c\u8f93\u51fa\u683c\u5f0f\uff0c\u6216\u8005\u50cf\u5de5\u5177\u4f7f\u7528\u548c\u6210\u672c\u9884\u7b97\u8fd9\u6837\u7684\u80fd\u529b\u3002\u6211\u4eec\u7ecf\u5e38\u5e0c\u671b\u5728\u4f20\u6388\u65b0\u9886\u57df\u77e5\u8bc6\u7684\u540c\u65f6\u8bad\u7ec3\u8fd9\u79cd\u884c\u4e3a\u3002<\/p>\n\n\n\n<p>\u540c\u65f6\u8bad\u7ec3\u8fd9\u4e24\u8005\u901a\u5e38\u5f88\u56f0\u96be\uff0c\u8f7b\u91cf\u7ea7\u5fae\u8c03\uff08\u5982 LoRA\uff09\u5f80\u5f80\u4e0d\u8db3\u4ee5\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u56e0\u6b64\u9700\u8981\u66f4\u5927\u89c4\u6a21\u7684\u4e2d\u8bad\u7ec3\u3002\u5728\u638c\u63e1\u65b0\u77e5\u8bc6\u7684\u57fa\u7840\u4e0a\u5b66\u4e60\u540e\u8bad\u7ec3\u884c\u4e3a\uff0c\u9700\u8981\u4e00\u4e2a\u590d\u6742\u7684\u540e\u8bad\u7ec3\u5806\u6808\uff0c\u901a\u5e38\u7531\u4e13\u6709\u6570\u636e\u548c\u5956\u52b1\u6a21\u578b\u7ec4\u6210\u3002\u867d\u7136\u524d\u6cbf\u5b9e\u9a8c\u5ba4\u53ef\u4ee5\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u4f46\u5176\u4ed6\u4ece\u4e1a\u8005\u8981\u590d\u523b\u53ef\u80fd\u5f88\u56f0\u96be\u6216\u6210\u672c\u9ad8\u6602\u3002<\/p>\n\n\n\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u5728\u7b56\u7565\u84b8\u998f\u53ef\u4ee5\u6709\u6548\u5730\u7528\u4e8e\u540e\u8bad\u7ec3\u4e13\u4e1a\u5316\u884c\u4e3a\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e5f\u9002\u7528\u4e8e\u6301\u7eed\u5b66\u4e60\u6216\u300c\u6d4b\u8bd5\u65f6\u8bad\u7ec3\u300d\uff1a\u5373\u5728\u6a21\u578b\u90e8\u7f72\u540e\u66f4\u65b0\u5b83\u4eec\uff0c\u800c\u4e0d\u4f1a\u5bfc\u81f4\u57fa\u7840\u6027\u80fd\u4e0b\u964d\u3002\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a\u5728\u516c\u53f8\u5185\u90e8\u6587\u6863\u4e0a\u8fdb\u884c\u4e2d\u8bad\u7ec3\u7684\u6a21\u578b\u4f5c\u4e3a\u5e94\u7528\u793a\u4f8b\u3002<\/p>\n\n\n\n<p><strong>\u8bad\u7ec3\u4e00\u4e2a\u5185\u90e8\u52a9\u624b<\/strong><\/p>\n\n\n\n<p>\u5b9a\u5236\u6a21\u578b\u7684\u4e00\u4e2a\u5e38\u89c1\u76ee\u6807\u662f\u5145\u5f53\u52a9\u624b\uff1a\u5728\u67d0\u4e2a\u9886\u57df\u62e5\u6709\u4e13\u5bb6\u77e5\u8bc6\uff0c\u5e76\u4e14\u5177\u6709\u53ef\u9760\u7684\u52a9\u624b\u5f0f\u884c\u4e3a\u3002\u6211\u4eec\u53ef\u80fd\u9700\u8981\u5bf9\u8fd9\u4e24\u8005\u8fdb\u884c\u5355\u72ec\u8bad\u7ec3\uff0c\u5c24\u5176\u662f\u5f53\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u65e0\u6cd5\u4ec5\u4ece\u9884\u8bad\u7ec3\u6570\u636e\u4e2d\u5b66\u5230\uff0c\u6216\u8005\u5b66\u4e60\u5b83\u4f1a\u5e72\u6270\u884c\u4e3a\u65f6\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u7684\u4f8b\u5b50\u662f\u4e00\u4e2a\u516c\u53f8\u5185\u90e8\u52a9\u624b\uff0c\u6211\u4eec\u6709\u4e24\u4e2a\u671f\u671b\uff1a<\/p>\n\n\n\n<p>1. \u6a21\u578b\u5bf9\u8be5\u9886\u57df\uff08\u516c\u53f8\u6587\u6863\uff09\u77e5\u8bc6\u6e0a\u535a\u3002\u9884\u8bad\u7ec3\u6a21\u578b\u6ca1\u6709\u89c1\u8fc7\u4efb\u4f55\u516c\u53f8\u5185\u90e8\u6587\u6863\uff0c\u56e0\u6b64\u65e0\u8bba\u6a21\u578b\u89c4\u6a21\u591a\u5927\uff0c\u90fd\u53ea\u80fd\u731c\u6d4b\u3002\u6211\u4eec\u5c06\u4f7f\u7528\u5185\u90e8\u77e5\u8bc6\u53ec\u56de\u8bc4\u4f30\uff08\u300c\u5185\u90e8 QA\u300d\uff09\u6765\u8861\u91cf\u8fd9\u4e00\u70b9\u3002<\/p>\n\n\n\n<p>2. \u6a21\u578b\u8868\u73b0\u51fa\u5f3a\u5927\u7684\u540e\u8bad\u7ec3\u884c\u4e3a\uff0c\u5373\u9075\u5faa\u6307\u4ee4\u3002\u6211\u4eec\u5c06\u4f7f\u7528\u5e38\u7528\u7684 IF-eval \u6765\u8861\u91cf\u8fd9\u4e00\u70b9\u3002<\/p>\n\n\n\n<p><strong>\u8bad\u7ec3\u65b0\u77e5\u8bc6\u4f1a\u964d\u4f4e\u5df2\u5b66\u5230\u7684\u884c\u4e3a<\/strong><\/p>\n\n\n\n<p>\u6211\u4eec\u5c06\u4ece Qwen3-8B\uff08\u800c\u4e0d\u662f\u57fa\u7840\u6a21\u578b\uff09\u5f00\u59cb\u3002Qwen3-8B \u5df2\u7ecf\u901a\u8fc7 RL \u8fdb\u884c\u4e86\u540e\u8bad\u7ec3\uff0c\u638c\u63e1\u4e86\u4f5c\u4e3a\u52a9\u624b\u6709\u7528\u7684\u6280\u80fd\uff0c\u5982\u6307\u4ee4\u9075\u5faa\u548c\u63a8\u7406\u3002\u5148\u524d\u7684\u7814\u7a76\u8868\u660e\uff0c\u8fd9\u79cd\u5f3a\u5316\u5b66\u4e60\u53ea\u8bad\u7ec3\u4e86\u539f\u59cb\u6a21\u578b\u7684\u4e00\u5c0f\u90e8\u5206\u5b50\u7f51\u7edc\uff0c\u56e0\u6b64\u5f53\u7f51\u7edc\u5728\u5927\u91cf\u6570\u636e\u4e0a\u8fdb\u4e00\u6b65\u8bad\u7ec3\u65f6\uff0c\u53ef\u80fd\u4f1a\u53d8\u5f97\u5f88\u8106\u5f31\u3002\u6211\u4eec\u7814\u7a76\u4e86\u8fd9\u79cd\u60c5\u51b5\u53d1\u751f\u7684\u7a0b\u5ea6\uff0c\u4ee5\u53ca\u5982\u4f55\u6062\u590d\u6240\u9700\u7684\u884c\u4e3a\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u51cf\u5c11\u8fd9\u79cd\u707e\u96be\u6027\u9057\u5fd8\uff08catastrophic forgetting\uff09\uff0c\u4e2d\u8bad\u7ec3\u4e2d\u4e00\u79cd\u5e38\u89c1\u7684\u505a\u6cd5\u662f\u6df7\u5165\u6765\u81ea\u6a21\u578b\u539f\u59cb\u9884\u8bad\u7ec3\u5206\u5e03\u7684\u300c\u80cc\u666f\u6570\u636e\u300d\u3002<\/p>\n\n\n\n<p>\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u65e0\u6cd5\u8bbf\u95ee Qwen3 \u7684\u9884\u8bad\u7ec3\u5206\u5e03\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u8003\u8651\u4e00\u4e2a\u66f4\u5f3a\u3001\u6210\u672c\u66f4\u9ad8\u7684\u57fa\u7ebf\uff1a\u6211\u4eec\u83b7\u53d6 Tulu3 \u63d0\u793a\uff08\u4e00\u4e2a\u5e7f\u6cdb\u7684\u804a\u5929\u548c\u6307\u4ee4\u9075\u5faa\u6570\u636e\u96c6\uff09\uff0c\u5e76\u4f7f\u7528 Qwen3-8B \u91cd\u65b0\u91c7\u6837\u5b83\u4eec\uff0c\u4ee5\u5145\u5f53\u804a\u5929\u80cc\u666f\u6570\u636e\u3002<\/p>\n\n\n\n<p>\u7136\u540e\uff0c\u6211\u4eec\u5728\u5185\u90e8\u6587\u6863\u548c\u804a\u5929\u6570\u636e\u7684\u4e0d\u540c\u6df7\u5408\u6bd4\u4f8b\u4e0a\u5fae\u8c03 Qwen3-8B\u3002\u63d0\u9ad8\u6587\u6863\u6570\u636e\u6bd4\u4f8b\u4f1a\u76f4\u63a5\u63d0\u5347\u6a21\u578b\u7684\u77e5\u8bc6\u6c34\u5e73\u3002\u7136\u800c\uff0c\u5c3d\u7ba1\u6df7\u5165\u81f3\u5c11 30% \u7684\u804a\u5929\u6570\u636e\u6709\u52a9\u4e8e\u4fdd\u7559\u5927\u90e8\u5206\u6307\u4ee4\u9075\u5faa\u80fd\u529b\uff0c\u4f46\u6ca1\u6709\u4efb\u4f55\u4e00\u79cd\u6743\u91cd\u914d\u6bd4\u80fd\u7ef4\u6301\u5728 IF-eval \u4e0a\u7684\u539f\u59cb\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_6c0de6889fe44003856f7cfb4de22876@5091053_oswg147152oswg1080oswg537_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u5bf9\u4e8e\u4efb\u4f55\u7ed9\u5b9a\u7684\u6df7\u5408\u6bd4\u4f8b\uff0c\u6211\u4eec\u90fd\u89c2\u5bdf\u5230 IF-eval \u6027\u80fd\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u4e0b\u964d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_7798a1bf72644758890787be3b31fb6d@5091053_oswg177936oswg1080oswg680_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u53e6\u4e00\u79cd\u5e38\u7528\u7684\u65b9\u6cd5\u662f\u4f7f\u7528 LoRA \u6765\u7ea6\u675f\u53c2\u6570\u66f4\u65b0\uff0c\u4ece\u800c\u51cf\u5c11\u707e\u96be\u6027\u9057\u5fd8\u7684\u53ef\u80fd\u6027\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4ecd\u7136\u4e0d\u8db3\u4ee5\u4fdd\u7559 IF-eval\uff0c\u800c\u4e14 LoRA \u5b66\u5230\u7684\u77e5\u8bc6\u4e5f\u66f4\u5c11\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_7045f7f628cb490983bfb6e2e5b3326b@5091053_oswg171331oswg1080oswg533_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p><strong>\u5728\u7b56\u7565\u84b8\u998f\u6062\u590d\u540e\u8bad\u7ec3\u884c\u4e3a<\/strong><\/p>\n\n\n\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u8bd5\u56fe\u5728\u5bf9\u5185\u90e8\u6587\u6863\u8fdb\u884c\u5fae\u8c03\u540e\uff0c\u6062\u590d\u6307\u4ee4\u9075\u5faa\u884c\u4e3a\u3002\u8fd9\u79cd\u884c\u4e3a\u6700\u521d\u662f\u7528 RL \u8bad\u7ec3\u7684\uff0c\u6210\u672c\u9ad8\u6602\uff0c\u800c\u4e14\u6b63\u5982\u6211\u4eec\u6240\u89c1\uff0c\u5f88\u8106\u5f31\u3002<\/p>\n\n\n\n<p>\u53d6\u800c\u4ee3\u4e4b\u7684\u662f\uff0c\u6211\u4eec\u5728 Tulu3 \u63d0\u793a\u4e0a\uff0c\u4f7f\u7528\u6a21\u578b\u7684\u65e9\u671f\u7248\u672c Qwen3-8B \u4f5c\u4e3a\u6559\u5e08\uff0c\u6765\u8fdb\u884c\u5728\u7b56\u7565\u84b8\u998f\u3002\u8bf7\u6ce8\u610f\uff0c\u8fd9\u4e2a\u8bad\u7ec3\u9636\u6bb5\u4e0e\u5185\u90e8\u6587\u6863\u6570\u636e\u65e0\u5173\uff0c\u5176\u552f\u4e00\u76ee\u7684\u662f\u6062\u590d\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u4f7f\u7528\u6a21\u578b\u7684\u65e9\u671f\u7248\u672c\u4f5c\u4e3a\u6559\u5e08\u6765\u300c\u91cd\u65b0\u5524\u8d77\u300d\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u4e22\u5931\u7684\u80fd\u529b\uff0c\u8fd9\u4f7f\u5f97\u5728\u7b56\u7565\u84b8\u998f\u5728\u6301\u7eed\u5b66\u4e60\uff08continuous learning\uff09\u65b9\u9762\u975e\u5e38\u6709\u524d\u666f\u3002\u6211\u4eec\u53ef\u4ee5\u4ea4\u66ff\u8fdb\u884c\u300c\u5728\u65b0\u6570\u636e\u4e0a\u5fae\u8c03\u300d\u548c\u300c\u84b8\u998f\u4ee5\u6062\u590d\u884c\u4e3a\u300d\u8fd9\u4e24\u4e2a\u9636\u6bb5\uff0c\u4f7f\u6211\u4eec\u7684\u6a21\u578b\u80fd\u591f\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\u5b66\u4e60\u5e76\u4fdd\u6301\u77e5\u8bc6\u7684\u6700\u65b0\u72b6\u6001\u3002<\/p>\n\n\n\n<p>\u5728 70-30 \u6df7\u5408\u7684\u5185\u90e8\u6587\u6863\u6570\u636e\u548c\u804a\u5929\u6570\u636e\u4e0a\u5fae\u8c03\u540e\uff0c\u5728\u7b56\u7565\u84b8\u998f\u51e0\u4e4e\u5b8c\u5168\u6062\u590d\u4e86\u5728 IF-eval \u4e0a\u7684\u6027\u80fd\uff0c\u4e14\u6ca1\u6709\u635f\u5931\u4efb\u4f55\u77e5\u8bc6\uff1b\u6211\u4eec\u8fd8\u89c2\u5bdf\u5230\u804a\u5929\u80fd\u529b\u548c\u6a21\u578b\u5728\u5185\u90e8 QA \u8bc4\u4f30\u4e2d\u7684\u300c\u77e5\u8bc6\u300d\u6027\u80fd\u4e4b\u95f4\u5b58\u5728\u4e00\u4e9b\u6b63\u5411\u8fc1\u79fb\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_1dfe9134ecea42329d6e3e459fbc0b5a@5091053_oswg59344oswg1080oswg355_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4ece\u672c\u8d28\u4e0a\u8bb2\uff0c\u6211\u4eec\u5c06\u8bed\u8a00\u6a21\u578b\u672c\u8eab\u89c6\u4e3a\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\uff0c\u9ad8\u6982\u7387\u7684\u884c\u4e3a\u4f1a\u53d7\u5230\u5956\u52b1\u3002\u8fd9\u4e0e\u9006\u5411 RL\uff08inverse RL\uff09\u6709\u5173\uff1a\u9ad8\u6982\u7387\u7684\u884c\u4e3a\u5bf9\u5e94\u4e8e\u5047\u5b9a\u7684\u6f5c\u5728\u504f\u597d\u6a21\u578b\u4e2d\u7684\u6709\u5229\u5956\u52b1\u3002\u4efb\u4f55\u7ecf\u8fc7\u6307\u4ee4\u8c03\u4f18\u7684\u5f00\u6e90\u6743\u91cd\u6a21\u578b\u90fd\u53ef\u4ee5\u5728\u8fd9\u4e2a\u610f\u4e49\u4e0a\u7528\u4f5c\u5956\u52b1\u6a21\u578b\uff1b\u6211\u4eec\u53ea\u9700\u8981\u80fd\u8bbf\u95ee compute_logprobs \u51fd\u6570\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u8ba8\u8bba<\/strong><\/h2>\n\n\n\n<p><strong>\u5bc6\u96c6\u76d1\u7763\u53ef\u6781\u5927\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387<\/strong><\/p>\n\n\n\n<p>\u5f3a\u5316\u5b66\u4e60\u548c\u5728\u7b56\u7565\u84b8\u998f\u90fd\u901a\u8fc7\u53cd\u5411 KL \u8fdb\u884c\u5b66\u4e60\uff0c\u4fee\u526a\u57fa\u7840\u7b56\u7565\u4e2d\u5b58\u5728\u7684\u52a8\u4f5c\u7a7a\u95f4\u3002\u533a\u522b\u5728\u4e8e\u5956\u52b1\u7684\u5bc6\u5ea6\u3002<\/p>\n\n\n\n<p>\u5728 LoRA Without Regret \u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4fe1\u606f\u8bba\u7684\u89c2\u70b9\uff0c\u5373\u5f3a\u5316\u5b66\u4e60\u6bcf\u4e2a\u56de\u5408\u53ea\u6559\u6388 O (1\uff09 \u7684\u6bd4\u7279\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u84b8\u998f\u6bcf\u4e2a\u56de\u5408\u6559\u6388 O (N\uff09 \u6bd4\u7279\uff0c\u5176\u4e2d N \u662f token \u7684\u6570\u91cf\u3002\u901a\u8fc7\u66f4\u5bc6\u96c6\u7684\u5956\u52b1\uff0c\u6211\u4eec\u5230\u5e95\u80fd\u83b7\u5f97\u591a\u5c11\u8bad\u7ec3\u6548\u7387\u7684\u63d0\u5347\uff1f<\/p>\n\n\n\n<p>\u6211\u4eec\u505a\u4e86\u4e00\u4e2a\u5b9e\u9a8c\u6765\u76f4\u63a5\u6bd4\u8f83\u4e24\u8005\uff1a<\/p>\n\n\n\n<p>\u4ece Qwen3-8B-Base\uff08\u6ca1\u6709\u989d\u5916\u7684 SFT\uff09\u5f00\u59cb\u3002<\/p>\n\n\n\n<p>\u5728 DeepMath \u4e0a\u8fd0\u884c RL\u3002\u6211\u4eec\u4f7f\u7528 128 \u7684 LoRA rank\u3002\u751f\u6210\u7684\u6a21\u578b\u662f\u84b8\u998f\u7684\u6559\u5e08\u3002<\/p>\n\n\n\n<p>\u4ece RL \u8bad\u7ec3\u7684\u6a21\u578b\uff082\uff09\u5728\u7b56\u7565\u84b8\u998f\u56de\u57fa\u7840\u6a21\u578b\uff081\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_1f793d04689e4814acddbfb8bb85fe11@5091053_oswg125699oswg1080oswg474_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4ece\u76f8\u540c\u7684\u521d\u59cb\u5316\u5f00\u59cb\uff0c\u5728\u7b56\u7565\u84b8\u998f\u5b66\u4e60 RL \u8bad\u7ec3\u7684\u7b56\u7565\u6240\u9700\u7684\u68af\u5ea6\u6b65\u6570\u5927\u7ea6\u5c11 7-10 \u500d\uff0c\u8fd9\u5bf9\u5e94\u4e8e 50-100 \u500d\u7684\u8ba1\u7b97\u6548\u7387\u63d0\u5347\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u770b\u5230\uff0c\u84b8\u998f\u8fbe\u5230\u6559\u5e08\u6027\u80fd\u6c34\u5e73\u7684\u901f\u5ea6\u6bd4 RL \u5feb\u4e86\u5927\u7ea6 7-10 \u500d\u3002\u53cd\u5411 KL \u4e0b\u964d\u5230\u63a5\u8fd1\u96f6\uff0cAIME \u5206\u6570\u5728 10 \u4e2a\u68af\u5ea6\u6b65\u5185\u5c31\u5f97\u4ee5\u6062\u590d\uff0c\u800c RL \u5219\u9700\u8981 70 \u6b65\u624d\u80fd\u8fbe\u5230\u8be5\u6c34\u5e73\u3002<\/p>\n\n\n\n<p><strong>\u84b8\u998f\u53ef\u4ee5\u6709\u6548\u5730\u91cd\u7528\u8bad\u7ec3\u6570\u636e\u4ee5\u63d0\u9ad8\u6570\u636e\u6548\u7387<\/strong><\/p>\n\n\n\n<p>\u5bf9\u4e8e\u4ece\u4e1a\u8005\u6765\u8bf4\uff0c\u6536\u96c6\u5927\u91cf\u7684\u8bad\u7ec3\u63d0\u793a\u53ef\u80fd\u65e2\u56f0\u96be\u53c8\u8017\u65f6\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5e0c\u671b\u80fd\u591f\u5728\u8bad\u7ec3\u4e2d\u591a\u6b21\u91cd\u7528\u63d0\u793a\u3002<\/p>\n\n\n\n<p>\u4f7f\u7528 RL \u65f6\uff0c\u5728\u540c\u4e00\u4e2a\u63d0\u793a\u4e0a\u8bad\u7ec3\u591a\u4e2a\u8f6e\u6b21\uff08epochs\uff09\u5e38\u5e38\u5bfc\u81f4\u5bf9\u6700\u7ec8\u7b54\u6848\u7684\u7b80\u5355\u8bb0\u5fc6\u3002<\/p>\n\n\n\n<p>\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5728\u7b56\u7565\u84b8\u998f\u5b66\u4e60\u7684\u662f\u8fd1\u4f3c\u6559\u5e08\u7684\u5b8c\u6574\u5206\u5e03\uff0c\u800c\u4e0d\u662f\u8bb0\u5fc6\u5355\u4e2a\u7b54\u6848\u3002\u8fd9\u4f7f\u6211\u4eec\u80fd\u591f\u7528\u540c\u4e00\u4e2a\u63d0\u793a\u8bad\u7ec3\u591a\u4e2a\u6837\u672c\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u91cd\u590d\u4e86\u4e0a\u8ff0\u5728\u6570\u5b66\u4e0a\u8bad\u7ec3 Qwen3-8B-Base \u7684\u5b9e\u9a8c\uff0c\u4f46\u8fd9\u6b21\u53ea\u4f7f\u7528\u6570\u636e\u96c6\u4e2d\u968f\u673a\u9009\u62e9\u7684\u4e00\u4e2a\u63d0\u793a\u3002\u6211\u4eec\u5728\u8fd9\u4e2a\u63d0\u793a\u4e0a\u8fde\u7eed\u8bad\u7ec3\u4e86 20 \u6b65\u3002\u5c3d\u7ba1\u6211\u4eec\u53ea\u5728\u4e00\u4e2a\u63d0\u793a\u4e0a\u8bad\u7ec3\uff0c\u4f46\u6211\u4eec\u786e\u5b9e\u8fbe\u5230\u4e86\u4e0e\u6559\u5e08\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_1ceb82679734475d82e1b7f5d35f4c25@5091053_oswg143605oswg1080oswg476_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p><strong>RL \u5728\u8bed\u4e49\u7b56\u7565\u7a7a\u95f4\u4e2d\u641c\u7d22<\/strong><\/p>\n\n\n\n<p>\u6211\u4eec\u5df2\u7ecf\u770b\u5230\uff0c\u5728\u7b56\u7565\u84b8\u998f\u53ef\u4ee5\u7528\u5c11\u5f97\u591a\u7684\u8bad\u7ec3\u6b65\u9aa4\u6765\u590d\u523b RL \u63d0\u4f9b\u7684\u5b66\u4e60\u6548\u679c\u3002<\/p>\n\n\n\n<p>\u4e00\u79cd\u89e3\u91ca\u662f\uff0c\u4e0e\u9884\u8bad\u7ec3\u4e0d\u540c\uff0cRL \u5e76\u672a\u5728\u68af\u5ea6\u6b65\u9aa4\u672c\u8eab\u4e0a\u82b1\u8d39\u5927\u91cf\u8ba1\u7b97\u3002\u6211\u4eec\u5e94\u8be5\u8ba4\u4e3a RL \u5c06\u5176\u5927\u90e8\u5206\u8ba1\u7b97\u7528\u5728\u4e86\u641c\u7d22\uff08search\uff09\u4e0a \u2014\u2014 \u5373\u63a8\u51fa\u4e00\u4e2a\u7b56\u7565\u5e76\u5206\u914d\u529f\u52b3 \u2014\u2014 \u800c\u4e0d\u662f\u8fdb\u884c\u66f4\u65b0\u3002<\/p>\n\n\n\n<p>\u9884\u8bad\u7ec3\u901a\u8fc7\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u63a2\u7d22\u9ad8\u7ef4\u53c2\u6570\u7a7a\u95f4\u3002\u9884\u8bad\u7ec3\u9700\u8981\u6d77\u91cf\u4fe1\u606f\uff0c\u5e76\u4e14\u975e\u5e38\u96be\u4ee5\u84b8\u998f\u3002<\/p>\n\n\n\n<p>\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u6211\u4eec\u5e94\u8be5\u8ba4\u4e3a RL \u662f\u5728\u63a2\u7d22\u8bed\u4e49\u7b56\u7565\uff08semantic strategies\uff09\u7684\u7a7a\u95f4\u3002RL \u4e0d\u662f\u5728\u53c2\u6570\u7a7a\u95f4\u4e2d\u63a2\u7d22\uff0c\u800c\u662f\u9760\u8fd0\u6c14\u300c\u5076\u7136\u300d\u53d1\u73b0\u65b0\u7b56\u7565 \u2014\u2014 \u5b83\u53ea\u662f\u4ece\u5b83\u5df2\u6709\u7684\u6743\u91cd\u96c6\u5408\u4e2d\u968f\u673a\u62bd\u6837\u3002<\/p>\n\n\n\n<p>\u4e00\u65e6\u627e\u5230\u4e86\u4e00\u4e2a\u597d\u7684\u7b56\u7565\uff0c\u84b8\u998f\u5c31\u6210\u4e86\u5b66\u4e60\u5b83\u7684\u6377\u5f84\uff1a<strong>\u5728\u7b56\u7565\u84b8\u998f\u4e0d\u9700\u8981\u5bf9 RL \u8bfe\u7a0b\u4e2d\u7684\u4e2d\u95f4\u7b56\u7565\u8fdb\u884c\u5efa\u6a21\uff0c\u800c\u53ea\u9700\u8981\u5b66\u4e60\u6700\u7ec8\u7684\u7b56\u7565\u3002<\/strong><\/p>\n\n\n\n<p>\u6253\u4e2a\u6bd4\u65b9\uff1a\u5728\u79d1\u5b66\u7814\u7a76\u4e2d\uff0c\u6211\u4eec\u82b1\u8d39\u5927\u91cf\u65f6\u95f4\u548c\u8d44\u6e90\u6765\u5bfb\u627e\u7b54\u6848\u548c\u63a2\u7d22\u65b0\u601d\u60f3\u3002\u4e00\u65e6\u53d1\u73b0\u4e00\u4e2a\u7ed3\u679c\uff0c\u7528\u81ea\u7136\u8bed\u8a00\u5c06\u5176\u6559\u7ed9\u522b\u4eba\u5c31\u7b80\u5355\u5f97\u591a\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u50cf\u8fd0\u52a8\u8fd9\u6837\u7684\u76f4\u89c9\u6027\u8eab\u4f53\u6280\u80fd\uff0c\u5c31\u5f88\u96be\u6559\u7ed9\u522b\u4eba\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u77e5\u8bc6\u5b58\u5728\u4e8e\u4e00\u79cd\u5929\u751f\u7684\u8bed\u8a00\u4e2d\uff08\u4f8b\u5982\uff0c\u808c\u8089\u8bb0\u5fc6\uff09\uff0c\u53ea\u6709\u6211\u4eec\u81ea\u5df1\u624d\u80fd\u8f7b\u6613\u7406\u89e3\u3002\u8fd0\u52a8\u53ea\u80fd\u901a\u8fc7\u53cd\u590d\u7ec3\u4e60\u6765\u5b66\u4e60\u3002<\/p>\n\n\n\n<p><strong>\u4f5c\u4e3a\u6301\u7eed\u5b66\u4e60\u5de5\u5177\u7684\u5728\u7b56\u7565\u5b66\u4e60<\/strong><\/p>\n\n\n\n<p>\u5728\u5173\u4e8e\u4e2a\u6027\u5316\u7684\u90e8\u5206\uff0c\u6211\u4eec\u63a2\u8ba8\u4e86\u5728\u7b56\u7565\u84b8\u998f\u5c06\u4e13\u4e1a\u8bad\u7ec3\u884c\u4e3a\u91cd\u65b0\u5f15\u5165\u6a21\u578b\u7684\u80fd\u529b\u3002\u8fd9\u53ef\u4ee5\u63a8\u5e7f\u5230\u66f4\u5e7f\u6cdb\u7684\u6301\u7eed\u5b66\u4e60\uff08continual learning\uff09\u4efb\u52a1\uff0c\u8fd9\u4e9b\u4efb\u52a1\u8981\u6c42\u5728\u4e0d\u964d\u4f4e\u5148\u524d\u80fd\u529b\u7684\u60c5\u51b5\u4e0b\u83b7\u53d6\u65b0\u77e5\u8bc6\u3002<\/p>\n\n\n\n<p>\u5148\u524d\u7684\u5de5\u4f5c\u53d1\u73b0\uff0c\u5728\u7b56\u7565\u5b66\u4e60\uff08RL\uff09\u6bd4\u8f83\u5c11\u5730\u9057\u5fd8\u3002\u7136\u800c\uff0cRL \u53ea\u80fd\u5851\u9020\u884c\u4e3a \u2014\u2014 \u5b83\u4e0d\u80fd\u5f88\u597d\u5730\u6559\u6388\u65b0\u77e5\u8bc6\uff0c\u56e0\u6b64\u4e0d\u8db3\u4ee5\u7528\u4e8e\u6301\u7eed\u5b66\u4e60\u3002<\/p>\n\n\n\n<p>\u5728\u4e0a\u4e00\u8282\u4e2d\uff0c\u6211\u4eec\u770b\u5230 SFT\uff08\u5305\u62ec\u79bb\u7b56\u7565\u84b8\u998f\uff09\u5728\u652f\u6301\u6301\u7eed\u5b66\u4e60\u65b9\u9762\u662f\u5931\u8d25\u7684\uff0c\u56e0\u4e3a\u5b83\u4f1a\u964d\u4f4e\u884c\u4e3a\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u66f4\u6df1\u5165\u5730\u7814\u7a76\u4e86\u8fd9\u4e2a\u95ee\u9898\u3002\u5f53\u6211\u4eec\u5728\u6a21\u578b\u81ea\u5df1\u7684\u6837\u672c\u6570\u636e\u96c6\u4e0a\u8fd0\u884c SFT \u65f6\u4f1a\u53d1\u751f\u4ec0\u4e48\uff1f\u6211\u4eec\u770b\u5230\uff0c\u4efb\u4f55\u5927\u4e8e\u96f6\u7684\u5b9e\u7528\u5b66\u4e60\u7387\u90fd\u4f1a\u5bfc\u81f4\u6307\u4ee4\u9075\u5faa\u8bc4\u4f30\u7684\u6027\u80fd\u4e0b\u964d\uff01<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/img.36krcdn.com\/hsossms\/20251028\/v2_35701789703149b5bf29376f976502cf@5091053_oswg111861oswg1080oswg472_img_000?x-oss-process=image\/format,jpg\/interlace,1\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4e00\u4e2a\u53ef\u80fd\u7684\u89e3\u91ca\u662f\uff0c\u867d\u7136 KL \u6563\u5ea6\u5728\u671f\u671b\u4e0a\u4e3a 0\uff0c\u4f46\u6bcf\u4e2a\u6709\u9650\u7684\u6279\u6b21\uff08batch\uff09\u5728\u5b9e\u8df5\u4e2d\u90fd\u4f1a\u8868\u73b0\u51fa\u7565\u5fae\u4e0d\u540c\u7684\u5206\u5e03\u3002\u5728\u8fd9\u4e9b\u6709\u9650\u6279\u6b21\u4e0a\u8bad\u7ec3\u4f1a\u5bfc\u81f4\u975e\u96f6\u7684\u68af\u5ea6\u66f4\u65b0\uff0c\u8fd9\u4f1a\u4f7f\u66f4\u65b0\u540e\u7684\u6a21\u578b\u7b56\u7565\u504f\u79bb\u5176\u539f\u59cb\u72b6\u6001\u3002<\/p>\n\n\n\n<p>\u5728\u7b56\u7565\u84b8\u998f\u59cb\u7ec8\u4fdd\u6301\u5728\u5728\u7b56\u7565\u72b6\u6001\uff0c\u5e76\u4e14\u7531\u4e8e\u6559\u5e08\u4fdd\u6301\u4e0d\u53d8\uff0c\u5b66\u751f\u4f1a\u6536\u655b\u4e8e\u6559\u5e08\u7684\u671f\u671b\u884c\u4e3a\uff0c\u800c\u4e0d\u4f1a\u50cf SFT \u90a3\u6837\u5728\u81ea\u84b8\u998f\u8bbe\u7f6e\u4e2d\u51fa\u73b0\u6027\u80fd\u8870\u9000\u3002\u8fd9\u4f7f\u5f97<strong>\u5728\u7b56\u7565\u84b8\u998f\u6210\u4e3a\u4e00\u79cd\u975e\u5e38\u6709\u524d\u666f\u7684\u6301\u7eed\u5b66\u4e60\u5de5\u5177<\/strong>\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u603b\u7ed3<\/strong><\/h2>\n\n\n\n<p>\u6211\u4eec\u63a2\u8ba8\u4e86\u5728\u7b56\u7565\u84b8\u998f\u5728\u8bad\u7ec3\u5c0f\u578b\u6a21\u578b\u8fdb\u884c\u6570\u5b66\u63a8\u7406\u6216\u6301\u7eed\u5b66\u4e60\u52a9\u624b\u7b49\u65b9\u9762\u7684\u5e94\u7528\u3002\u6211\u4eec\u5c06\u5728\u7b56\u7565\u84b8\u998f\u4e0e\u5176\u4ed6\u4e24\u79cd\u540e\u8bad\u7ec3\u65b9\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\uff1a\u79bb\u7b56\u7565\u84b8\u998f\u548c\u5728\u7b56\u7565 RL\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u53d1\u73b0\uff0c<strong>\u5728\u7b56\u7565\u84b8\u998f\u7ed3\u5408\u4e86\u4e24\u8005\u7684\u4f18\u70b9\uff1a\u5728\u7b56\u7565\u8bad\u7ec3\u7684\u53ef\u9760\u6027\u80fd\u4ee5\u53ca\u5bc6\u96c6\u5956\u52b1\u4fe1\u53f7\u5e26\u6765\u7684\u6210\u672c\u6548\u76ca\u3002<\/strong><\/p>\n\n\n\n<p>\u540e\u8bad\u7ec3\u662f\u8fbe\u5230\u524d\u6cbf\u6a21\u578b\u80fd\u529b\u7684\u5173\u952e\u90e8\u5206\u3002\u901a\u8fc7\u5229\u7528\u6765\u81ea\u5b66\u751f\u7684\u5728\u7b56\u7565\u91c7\u6837\u548c\u6765\u81ea\u6559\u5e08\u7684\u5bc6\u96c6\u76d1\u7763\uff0c\u5728\u7b56\u7565\u84b8\u998f\u65b9\u6848\u80fd\u591f\u4ee5\u524d\u6cbf\u9ad8\u8ba1\u7b97\u91cf RL \u8fd0\u884c\u6210\u672c\u7684\u4e00\u5c0f\u90e8\u5206\uff0c\u8fbe\u5230\u8fd9\u4e9b\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u7684\u5b9e\u73b0\u53ef\u4ee5\u5728 Tinker cookbook \u4e2d\u627e\u5230\u3002\u6211\u4eec\u7684\u5de5\u4f5c\u63a2\u7d22\u4e86\u5728\u7b56\u7565\u84b8\u998f\u7684\u7b80\u5355\u76f4\u63a5\u7684\u5b9e\u4f8b\u5316\uff0c\u4ee5\u6e05\u6670\u5730\u5c55\u793a\u5176\u4f18\u52bf\u3002\u6211\u4eec\u5e0c\u671b\u7ee7\u7eed\u7814\u7a76\u84b8\u998f\u7684\u65b0\u5e94\u7528\u3001\u6539\u8fdb\u6559\u5e08\u76d1\u7763\u7684\u65b0\u65b9\u6cd5\uff0c\u4ee5\u53ca\u63d0\u9ad8\u6570\u636e\u6548\u7387\u548c\u6301\u7eed\u5b66\u4e60\u7684\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u5728 Thinking Machines\uff0c\u6211\u4eec\u7684\u4f7f\u547d\u662f\u4e3a\u4eba\u4eec\u63d0\u4f9b\u517c\u5177\u524d\u6cbf\u6027\u80fd\u3001\u9002\u5e94\u6027\u548c\u4e2a\u6027\u5316\u7684 AI \u6a21\u578b\u3002\u5728\u7b56\u7565\u84b8\u998f\u662f\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\u7684\u6709\u529b\u5de5\u5177\u3002<\/p>\n\n\n\n<p>\u6587\u7ae0\u94fe\u63a5\uff1a<a href=\"https:\/\/36kr.com\/p\/3528135837817733\">https:\/\/36kr.com\/p\/3528135837817733<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u7b56\u7565\u84b8\u998f\u662f\u4e00\u79cd\u5c06\u5f3a\u5316\u5b66\u4e60 (RL) \u7684\u7ea0\u9519\u76f8\u5173\u6027\u4e0e SFT \u7684\u5956\u52b1&#46;&#46;&#46;<\/p>\n","protected":false},"author":2,"featured_media":488,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[18],"tags":[],"class_list":["post-523","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tools"],"_links":{"self":[{"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/posts\/523","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/comments?post=523"}],"version-history":[{"count":1,"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/posts\/523\/revisions"}],"predecessor-version":[{"id":524,"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/posts\/523\/revisions\/524"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/media\/488"}],"wp:attachment":[{"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/media?parent=523"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/categories?post=523"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/smartaiunion.com\/index.php\/wp-json\/wp\/v2\/tags?post=523"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}