{"id":1301,"date":"2025-11-01T14:51:47","date_gmt":"2025-11-01T06:51:47","guid":{"rendered":"https:\/\/lingbo.online\/?p=1301"},"modified":"2025-11-08T17:25:54","modified_gmt":"2025-11-08T09:25:54","slug":"a-survey-on-post-training-of-large-language-models","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/learning_note\/a-survey-on-post-training-of-large-language-models\/","title":{"rendered":"A Survey on Post-training of Large Language Models"},"content":{"rendered":"<h3>Meta Data<\/h3>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4 2025.08.1<\/li>\n<li>\u4f5c\u8005\uff1aGuiyao Tie, Zeli Zhao, Dingjie Song etc.<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2503.06072\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/abs\/2503.06072<\/a><\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/Mr-Tieguigui\/LLM-Post-Training\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/Mr-Tieguigui\/LLM-Post-Training<\/a><\/li>\n<li>\u9605\u8bfb\u601d\u7ef4\u5bfc\u56fe\uff1a<a href=\"_wp_link_placeholder\" data-wplink-edit=\"true\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/pan.baidu.com\/s\/11USr8_A1XIa6pX_GMkRxpA?pwd=gyRz<\/a><\/li>\n<\/ul>\n<h3><span style=\"color: revert; font-size: 1.75rem;\">\u6458\u8981<\/span><\/h3>\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b(LLM)\u7684\u51fa\u73b0\u4ece\u6839\u672c\u4e0a\u6539\u53d8\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0c\u4f7f\u5176\u5728\u4ece\u5bf9\u8bdd\u7cfb\u7edf\u5230\u79d1\u5b66\u63a2\u7d22\u7684\u5404\u4e2a\u9886\u57df\u90fd\u53d8\u5f97\u4e0d\u53ef\u6216\u7f3a\u3002\u7136\u800c\uff0c\u5b83\u4eec\u9884\u8bad\u7ec3\u7684\u67b6\u6784\u901a\u5e38\u4f1a\u5728\u4e13\u4e1a\u73af\u5883\u4e2d\u66b4\u9732\u51fa\u5c40\u9650\u6027\uff0c\u5305\u62ec\u53d7\u9650\u7684\u63a8\u7406\u80fd\u529b\u3001\u4f26\u7406\u4e0a\u7684\u4e0d\u786e\u5b9a\u6027(ethical uncertainties)\u4ee5\u53ca\u6b21\u4f18\u7684\u7279\u5b9a\u9886\u57df\u6027\u80fd\u3002 \u8fd9\u4e9b\u6311\u6218\u9700\u8981\u5148\u8fdb\u7684\u540e\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b (post-training language models, PoLMs) \u6765\u89e3\u51b3\u8fd9\u4e9b\u7f3a\u70b9\uff0c\u4f8b\u5982 OpenAI-o1\/o3 \u548c DeepSeek-R1\uff08\u7edf\u79f0\u4e3a\u5927\u578b\u63a8\u7406\u6a21\u578b\uff0cLarge Reasoning Models,LRMs\uff09\u3002 \u672c\u6587\u4ecb\u7ecd\u4e86\u9996\u4e2a\u5173\u4e8e PoLMs \u7684\u5168\u9762\u7efc\u8ff0\uff0c\u7cfb\u7edf\u5730\u8ffd\u8e2a\u4e86\u5b83\u4eec\u5728\u4e94\u4e2a\u6838\u5fc3\u8303\u5f0f\u4e2d\u7684\u6f14\u53d8\uff1a\u5fae\u8c03(Fine-tuning)\uff0c\u5b83\u63d0\u9ad8\u4e86\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u51c6\u786e\u6027\uff1b\u5bf9\u9f50(Alignment)\uff0c\u5b83\u786e\u4fdd\u4e86\u4f26\u7406\u4e00\u81f4\u6027\u548c\u4e0e\u4eba\u7c7b\u504f\u597d\u7684\u5bf9\u9f50\uff1b\u63a8\u7406(Reasoning)\uff0c\u5c3d\u7ba1\u5728\u5956\u52b1\u8bbe\u8ba1\u65b9\u9762\u9762\u4e34\u6311\u6218\uff0c\u4f46\u5b83\u63a8\u52a8\u4e86\u591a\u6b65\u63a8\u7406\uff1b\u6548\u7387(Efficiency)\uff0c\u5728\u65e5\u76ca\u590d\u6742\u7684\u73af\u5883\u4e2d\u4f18\u5316\u8d44\u6e90\u5229\u7528\uff1b\u4ee5\u53ca\u96c6\u6210\u548c\u9002\u5e94\uff0c\u5b83\u6269\u5c55\u4e86\u8de8\u591a\u79cd\u6a21\u6001\u7684\u80fd\u529b\uff0c\u540c\u65f6\u89e3\u51b3\u4e86\u8fde\u8d2f\u6027\u95ee\u9898\u3002 \u6211\u4eec\u901a\u8fc7\u7ed8\u5236\u4ece ChatGPT \u7684\u57fa\u7840\u5bf9\u9f50\u7b56\u7565\u5230 DeepSeek-R1 \u7684\u521b\u65b0\u63a8\u7406\u8fdb\u5c55\u7684\u8fdb\u6b65\uff0c\u9610\u8ff0\u4e86 PoLMs \u5982\u4f55\u5229\u7528\u6570\u636e\u96c6\u6765\u51cf\u8f7b\u504f\u5dee\uff0c\u6df1\u5316\u63a8\u7406\u80fd\u529b\u5e76\u589e\u5f3a\u9886\u57df\u9002\u5e94\u6027\u3002 \u6211\u4eec\u7684\u8d21\u732e\u5305\u62ec\u5bf9 PoLM \u6f14\u8fdb\u7684\u5f00\u521b\u6027\u7efc\u8ff0\uff0c\u4e00\u4e2a\u5bf9\u6280\u672f\u548c\u6570\u636e\u96c6\u8fdb\u884c\u5206\u7c7b\u7684\u7ed3\u6784\u5316\u5206\u7c7b\u6cd5\uff0c\u4ee5\u53ca\u4e00\u4e2a\u5f3a\u8c03 LRM \u5728\u63d0\u9ad8\u63a8\u7406\u80fd\u529b\u548c\u9886\u57df\u7075\u6d3b\u6027\u65b9\u9762\u4f5c\u7528\u7684\u6218\u7565\u8bae\u7a0b\u3002 \u4f5c\u4e3a\u540c\u7c7b\u9996\u4e2a\u7efc\u8ff0\uff0c\u8fd9\u9879\u5de5\u4f5c\u6574\u5408\u4e86\u6700\u8fd1\u7684 PoLM \u8fdb\u5c55\uff0c\u5e76\u4e3a\u672a\u6765\u7684\u7814\u7a76\u5efa\u7acb\u4e86\u4e00\u4e2a\u4e25\u683c\u7684\u77e5\u8bc6\u6846\u67b6\uff0c\u4fc3\u8fdb\u4e86 LLM \u7684\u53d1\u5c55\uff0c\u8fd9\u4e9b LLM \u5728\u79d1\u5b66\u548c\u793e\u4f1a\u5e94\u7528\u4e2d\u5177\u6709\u7cbe\u786e\u6027\u3001\u4f26\u7406\u9c81\u68d2\u6027\u548c\u591a\u529f\u80fd\u6027\u3002 \u9879\u76ee Github: <a href=\"https:\/\/github.com\/Mr-Tieguigui\/LLM-Post-Training\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/Mr-Tieguigui\/LLM-Post-Training<\/a>\u3002<\/p>\n<h3>Introduction<\/h3>\n<p>\u8bed\u8a00\u6a21\u578b (LMs) \u4ee3\u8868\u4e86\u65e8\u5728\u5bf9\u4eba\u7c7b\u8bed\u8a00\u8fdb\u884c\u5efa\u6a21\u548c\u751f\u6210\u7684\u590d\u6742\u8ba1\u7b97\u6846\u67b6\u3002 \u8fd9\u4e9b\u6a21\u578b\u5f7b\u5e95\u6539\u53d8\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406 (NLP) \u9886\u57df\uff0c\u4f7f\u673a\u5668\u80fd\u591f\u4ee5\u975e\u5e38\u63a5\u8fd1\u4eba\u7c7b\u8ba4\u77e5\u7684\u65b9\u5f0f\u7406\u89e3\u3001\u751f\u6210\u548c\u4e0e\u4eba\u7c7b\u8bed\u8a00\u4ea4\u4e92\u3002\u4e0e\u901a\u8fc7\u4e0e\u4e0a\u4e0b\u6587\u73af\u5883\u7684\u4ea4\u4e92\u548c\u63a5\u89e6\u81ea\u7136\u83b7\u5f97\u8bed\u8a00\u6280\u80fd\u7684\u4eba\u7c7b\u4e0d\u540c\uff0c\u673a\u5668\u5fc5\u987b\u7ecf\u8fc7\u5e7f\u6cdb\u7684\u3001\u6570\u636e\u9a71\u52a8\u7684\u8bad\u7ec3\u624d\u80fd\u53d1\u5c55\u7c7b\u4f3c\u7684\u80fd\u529b\u3002 \u8fd9\u63d0\u51fa\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u7814\u7a76\u6311\u6218\uff0c\u56e0\u4e3a\u4f7f\u673a\u5668\u80fd\u591f\u7406\u89e3\u548c\u751f\u6210\u4eba\u7c7b\u8bed\u8a00\uff0c\u540c\u65f6\u8fdb\u884c\u81ea\u7136\u7684\u3001\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u5bf9\u8bdd\uff0c\u4e0d\u4ec5\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u8d44\u6e90\uff0c\u8fd8\u9700\u8981\u6539\u8fdb\u7684\u6a21\u578b\u5f00\u53d1\u65b9\u6cd5\u3002<br \/>\n\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLMs) \u7684\u51fa\u73b0\uff0c\u5982 GPT-3 \u3001InstructGPT\u548cGPT-4\uff0c\u6807\u5fd7\u7740 LM \u53d1\u5c55\u7684\u4e00\u4e2a\u53d8\u9769\u9636\u6bb5\u3002\u8fd9\u4e9b\u6a21\u578b\u4ee5\u5176\u5e7f\u6cdb\u7684\u53c2\u6570\u5316\uff08extensive parameterization\uff09\u548c\u5148\u8fdb\u7684\u5b66\u4e60\u80fd\u529b\u800c\u8457\u79f0\uff0c\u65e8\u5728\u6355\u6349\u5927\u578b\u6570\u636e\u96c6\u4e2d\u590d\u6742\u7684\u8bed\u8a00\u7ed3\u6784(linguistic structures)\u3001\u4e0a\u4e0b\u6587\u5173\u7cfb(contextual relationships)\u548c\u6570\u636e\u96c6\u4e2d\u7684\u7ec6\u5fae\u6a21\u5f0f(nuanced patterns)\u3002\u8fd9\u4f7f\u5f97 LLMs \u4e0d\u4ec5\u53ef\u4ee5\u9884\u6d4b\u540e\u7eed\u5355\u8bcd\uff0c\u8fd8\u53ef\u4ee5\u5728\u5e7f\u6cdb\u7684\u4efb\u52a1\u4e2d\u751f\u6210\u8fde\u8d2f\u7684\u3001\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u6587\u672c\uff0c\u5305\u62ec\u7ffb\u8bd1\u3001\u95ee\u7b54\u548c\u6458\u8981\u3002 LLMs \u7684\u53d1\u5c55\u5f15\u53d1\u4e86\u91cd\u5927\u5b66\u672f\u5174\u8da3\uff0c\u53ef\u5206\u4e3a\u4e24\u4e2a\u4e3b\u8981\u9636\u6bb5\uff1a \u9884\u8bad\u7ec3 \u548c \u540e\u8bad\u7ec3\u3002<br \/>\n\u9884\u8bad\u7ec3:\u9884\u8bad\u7ec3\u7684\u6982\u5ff5\u6e90\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9 (CV) \u4efb\u52a1\u4e2d\u7684\u8fc1\u79fb\u5b66\u4e60\u3002\u5176\u4e3b\u8981\u76ee\u6807\u662f\u5229\u7528\u5e7f\u6cdb\u7684\u6570\u636e\u96c6\u5f00\u53d1\u901a\u7528\u6a21\u578b\uff0c\u8fd9\u6709\u52a9\u4e8e\u8f7b\u677e\u5fae\u8c03\u4ee5\u7528\u4e8e\u5404\u79cd\u4e0b\u6e38\u5e94\u7528\u3002\u9884\u8bad\u7ec3\u7684\u4e00\u4e2a\u663e\u8457\u4f18\u52bf\u662f\u5b83\u80fd\u591f\u5229\u7528\u4efb\u4f55\u672a\u6807\u8bb0\u7684\u6587\u672c\u8bed\u6599\u5e93\uff0c\u4ece\u800c\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u8bad\u7ec3\u6570\u636e\u6765\u6e90\u3002 \u7136\u800c\uff0c\u65e9\u671f\u7684\u9759\u6001\u9884\u8bad\u7ec3\u65b9\u6cd5\uff0c\u4f8b\u5982\u795e\u7ecf\u7f51\u7edc\u8bed\u8a00\u6a21\u578b (Neural Network Language Models,NNLM) \u548c Word2vec\uff0c\u96be\u4ee5\u9002\u5e94\u4e0d\u540c\u7684\u6587\u672c\u8bed\u4e49\u73af\u5883\uff0c\u4fc3\u4f7f\u4e86\u52a8\u6001\u9884\u8bad\u7ec3\u6280\u672f\u7684\u53d1\u5c55\uff0c\u5982 BERT\u548cXLNet\u3002 BERT\u901a\u8fc7\u5229\u7528 Transformer \u67b6\u6784\u5e76\u5728\u5927\u89c4\u6a21\u672a\u6807\u8bb0\u6570\u636e\u96c6\u4e0a\u91c7\u7528\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u6709\u6548\u5730\u89e3\u51b3\u4e86\u9759\u6001\u65b9\u6cd5\u7684\u5c40\u9650\u6027\u3002\u8fd9\u9879\u7814\u7a76\u786e\u7acb\u4e86\u201c\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u201d\u7684\u5b66\u4e60\u8303\u5f0f\uff0c\u542f\u53d1\u4e86\u4f17\u591a\u540e\u7eed\u7814\u7a76\uff0c\u8fd9\u4e9b\u7814\u7a76\u5f15\u5165\u4e86\u591a\u6837\u5316\u7684\u67b6\u6784\uff0c\u5305\u62ec GPT-2\u548c BART\u3002<br \/>\n\u540e\u8bad\u7ec3\uff1a\u540e\u8bad\u7ec3\u6307\u7684\u662f\u5728\u6a21\u578b\u7ecf\u8fc7\u9884\u8bad\u7ec3\u4e4b\u540e\u91c7\u7528\u7684\u6280\u672f\u548c\u65b9\u6cd5\uff0c\u65e8\u5728\u5b8c\u5584\u548c\u8c03\u6574\u6a21\u578b\u4ee5\u6ee1\u8db3\u7279\u5b9a\u4efb\u52a1\u6216\u7528\u6237\u9700\u6c42\u3002\u5728 GPT-3 \u53d1\u5e03\u4e4b\u540e\uff0c\u51ed\u501f\u5176 1750 \u4ebf\u4e2a\u53c2\u6570\uff0c\u540e\u8bad\u7ec3\u9886\u57df\u5728\u5174\u8da3\u548c\u521b\u65b0\u65b9\u9762\u7ecf\u5386\u4e86\u663e\u8457\u589e\u957f\u3002 \u51fa\u73b0\u4e86\u5404\u79cd\u65b9\u6cd5\u6765\u589e\u5f3a\u6a21\u578b\u6027\u80fd\uff0c\u5305\u62ec\u5fae\u8c03 \uff0c\u5b83\u4f7f\u7528\u6709\u6807\u7b7e\u7684\u6570\u636e\u96c6\u6216\u7279\u5b9a\u4efb\u52a1\u6570\u636e\u8c03\u6574\u6a21\u578b\u53c2\u6570\uff1b\u5bf9\u9f50\u7b56\u7565(alignment strategies)\uff0c\u5b83\u4f18\u5316\u6a21\u578b\u4ee5\u66f4\u597d\u5730\u4e0e\u7528\u6237\u504f\u597d\u5bf9\u9f50\uff1b\u77e5\u8bc6\u9002\u5e94\u6280\u672f(knowledge adaptation techniques)\uff0c\u5b83\u4f7f\u6a21\u578b\u80fd\u591f\u7ed3\u5408\u7279\u5b9a\u9886\u57df\u7684\u77e5\u8bc6\uff1b\u4ee5\u53ca\u63a8\u7406\u6539\u8fdb(reasoning improvements)\uff0c\u5b83\u589e\u5f3a\u6a21\u578b\u8fdb\u884c\u903b\u8f91\u63a8\u7406\u548c\u51b3\u7b56\u7684\u80fd\u529b\u3002\u8fd9\u4e9b\u6280\u672f\u7edf\u79f0\u4e3a\u540e\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b (PoLMs)\uff0c\u5bfc\u81f4\u4e86\u8bf8\u5982 GPT-4\u3001LLaMA-3\u3001Gemini-2.0\u548c Claude-3.5\u7b49\u6a21\u578b\u7684\u53d1\u5c55\uff0c\u6807\u5fd7\u7740 LLM \u80fd\u529b\u7684\u5b9e\u8d28\u6027\u8fdb\u6b65\u3002 \u7136\u800c\uff0c\u540e\u8bad\u7ec3\u6a21\u578b\u901a\u5e38\u96be\u4ee5\u5728\u6ca1\u6709\u91cd\u65b0\u8bad\u7ec3\u6216\u91cd\u5927\u53c2\u6570\u8c03\u6574\u7684\u60c5\u51b5\u4e0b\u9002\u5e94\u65b0\u4efb\u52a1\uff0c\u8fd9\u4f7f\u5f97 PTM \u5f00\u53d1\u6210\u4e3a\u4e00\u4e2a\u6d3b\u8dc3\u7684\u7814\u7a76\u9886\u57df\u3002<br \/>\n\u5982\u524d\u6240\u8ff0\uff0c\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b (PLMs) \u4e3b\u8981\u65e8\u5728\u63d0\u4f9b\u901a\u7528\u77e5\u8bc6\u548c\u80fd\u529b\uff0c\u800c PoLMs \u5219\u4fa7\u91cd\u4e8e\u4f7f\u8fd9\u4e9b\u6a21\u578b\u9002\u5e94\u7279\u5b9a\u4efb\u52a1\u548c\u9700\u6c42\u3002\u8fd9\u79cd\u9002\u5e94\u7684\u4e00\u4e2a\u663e\u8457\u4f8b\u5b50\u662f\u6700\u65b0\u7684 LLM\uff0cDeepSeek-R1\uff0c\u5b83\u8bf4\u660e\u4e86 PoLMs \u5728\u589e\u5f3a\u63a8\u7406\u80fd\u529b\u3001\u4e0e\u7528\u6237\u504f\u597d\u5bf9\u9f50\u4ee5\u53ca\u63d0\u9ad8\u8de8\u5404\u79cd\u9886\u57df\u7684\u9002\u5e94\u6027\u65b9\u9762\u7684\u6f14\u53d8\u3002 \u6b64\u5916\uff0c\u5f00\u6e90 LLM\uff08\u4f8b\u5982\uff0cLLaMA \u3001Gemma \u548c Nemotron\uff09\u548c\u7279\u5b9a\u9886\u57df\u5927\u578b\u6570\u636e\u96c6\uff08\u4f8b\u5982\uff0cPromptSource \u548c Flan\uff09\u7684\u65e5\u76ca\u666e\u53ca\uff0c\u6b63\u5728\u63a8\u52a8\u5b66\u672f\u7814\u7a76\u4eba\u5458\u548c\u884c\u4e1a\u4ece\u4e1a\u4eba\u5458\u5f00\u53d1 PoLMs \u7684\u8d8b\u52bf\u3002 \u8fd9\u4e00\u8d8b\u52bf\u7a81\u663e\u4e86\u4eba\u4eec\u5bf9 PoLMs \u9886\u57df\u4e2d\u91cf\u8eab\u5b9a\u5236\u7684\u9002\u5e94\u6027\u8d8a\u6765\u8d8a\u91cd\u89c6\u3002<br \/>\n\u5728\u73b0\u6709\u6587\u732e\u4e2d\uff0cPLMs \u5df2\u7ecf\u88ab\u5e7f\u6cdb\u8ba8\u8bba\u548c\u8c03\u67e5\uff0c\u800c PoLMs \u5f88\u5c11\u88ab\u7cfb\u7edf\u5730\u7efc\u8ff0\u3002 \u4e3a\u4e86\u63a8\u8fdb\u8fd9\u4e9b\u6280\u672f\uff0c\u5f7b\u5e95\u68c0\u67e5\u73b0\u6709\u7684\u7814\u7a76\u6210\u679c\uff0c\u4ee5\u786e\u5b9a\u5173\u952e\u6311\u6218\u3001\u5dee\u8ddd\u548c\u8fdb\u4e00\u6b65\u5b8c\u5584\u7684\u673a\u4f1a\uff0c\u81f3\u5173\u91cd\u8981\u3002\u672c\u8c03\u67e5\u65e8\u5728\u901a\u8fc7\u63d0\u4f9b\u4e00\u4e2a\u7ed3\u6784\u5316\u7684\u6846\u67b6\u6765\u586b\u8865\u8fd9\u4e00\u7a7a\u767d\uff0c\u4ee5\u5e94\u5bf9\u540e\u8bad\u7ec3\u7814\u7a76\u7684\u4e0d\u65ad\u53d1\u5c55\u3002\u5982\u56fe1\u6240\u793a \uff0c\u5b83\u63a2\u7d22\u4e86\u8bad\u7ec3\u540e\u7684\u591a\u4e2a\u9636\u6bb5\uff0c\u7279\u522b\u5173\u6ce8\u4ecechatgpt\u5230deepseek\u4f7f\u7528\u7684\u6280\u672f\u3002 \u8fd9\u4e9b\u6280\u672f\u6db5\u76d6\u4e86\u5e7f\u6cdb\u7684\u65b9\u6cd5\uff0c\u5305\u62ec\u5fae\u8c03\u3001LLM\u5bf9\u9f50\u3001\u63a8\u7406\u589e\u5f3a\u548c\u6548\u7387\u6539\u8fdb\u3002\u56fe\u4e2d\u84dd\u8272\u90e8\u5206\u7279\u522b\u7a81\u51fa\u4e86 DeepSeek \u5e94\u7528\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5f3a\u8c03\u4e86\u90a3\u4e9b\u6709\u52a9\u4e8e\u5176\u5728\u9002\u5e94\u7528\u6237\u504f\u597d\u548c\u7279\u5b9a\u9886\u57df\u9700\u6c42\u65b9\u9762\u53d6\u5f97\u6210\u529f\u7684\u521b\u65b0\u7b56\u7565\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101040018702.png\" alt=\"\" \/><\/p>\n<p><center>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u540e\u8bad\u7ec3\u6280\u672f\u7684\u53d1\u5c55\u6f14\u53d8\uff0c\u63cf\u7ed8\u4e86\u4ece\u6700\u521d\u7684\u65b9\u6cd5\u5230\u9ad8\u7ea7\u65b9\u6cd5\u7684\u6f14\u8fdb\uff0c\u91cd\u70b9\u4ecb\u7ecd\u4e86 DeepSeek \u6a21\u578b\u7684\u8d21\u732e\uff08\u4ee5\u84dd\u8272\u7a81\u51fa\u663e\u793a\uff09<\/center><\/p>\n<h4>\u4e3b\u8981\u8d21\u732e<\/h4>\n<p>\u672c\u6587\u662f\u5bf9 PoLMs \u7684\u9996\u6b21\u5168\u9762\u7efc\u8ff0\uff0c\u63d0\u4f9b\u4e86\u5bf9\u8be5\u9886\u57df\u6700\u65b0\u8fdb\u5c55\u7684\u900f\u5f7b\u3001\u7ed3\u6784\u5316\u7684\u63a2\u7d22\u3002\u867d\u7136\u4e4b\u524d\u7684\u8c03\u67e5\u901a\u5e38\u4fa7\u91cd\u4e8e LLM \u53d1\u5c55\u7684\u7279\u5b9a\u65b9\u9762\uff0c\u4f8b\u5982\u504f\u597d\u5bf9\u9f50\u3001\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u4ee5\u53ca LLMs \u7684\u57fa\u7840\u6280\u672f\uff0c\u4f46\u5b83\u4eec\u4e3b\u8981\u96c6\u4e2d\u5728\u72ed\u7a84\u7684\u5b50\u4e3b\u9898\u4e0a\u3002 \u76f8\u6bd4\u4e4b\u4e0b\uff0c\u672c\u6b21\u8c03\u67e5\u91c7\u7528\u4e86\u4e00\u79cd\u6574\u4f53\u65b9\u6cd5\uff0c\u5bf9\u540e\u8bad\u7ec3\u671f\u95f4\u5e38\u7528\u7684\u6838\u5fc3\u6280\u672f\u8fdb\u884c\u4e86\u5168\u9762\u7efc\u8ff0\uff0c\u5e76\u5bf9\u5176\u8fdb\u884c\u4e86\u7cfb\u7edf\u5206\u7c7b\u3002 \u6b64\u5916\uff0c\u6211\u4eec\u7814\u7a76\u4e86\u8fd9\u4e9b\u65b9\u6cd5\u4e0d\u53ef\u6216\u7f3a\u7684\u6570\u636e\u96c6\u548c\u73b0\u5b9e\u4e16\u754c\u5e94\u7528\u7a0b\u5e8f\uff0c \u5e76\u786e\u5b9a\u672a\u6765\u7814\u7a76\u7684\u5f00\u653e\u6311\u6218\u548c\u6709\u5e0c\u671b\u7684\u65b9\u5411\u3002 \u672c\u6b21\u8c03\u67e5\u7684\u4e3b\u8981\u8d21\u732e\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\u7efc\u5408\u7684\u5386\u53f2\u7efc\u8ff0\uff1a \u6211\u4eec\u9996\u6b21\u6df1\u5165\u7efc\u8ff0\u4e86PoLM\uff0c\u8ffd\u6eaf\u4e86\u5b83\u4eec\u4eceChatGPT\u6700\u521d\u7684\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u5230DeepSeek-R1\u521b\u65b0\u51b7\u542f\u52a8RL\u65b9\u6cd5\u7684\u6f14\u53d8\u8fc7\u7a0b\u3002 \u8fd9\u79cd\u7efc\u5408\u6db5\u76d6\u4e86\u5173\u952e\u6280\u672f\uff08\u5373\uff0c\u5fae\u8c03\u3001\u5bf9\u9f50\u3001\u63a8\u7406\u3001\u6548\u7387\u4ee5\u53ca\u96c6\u6210\u548c\u9002\u5e94\uff09\uff0c\u5206\u6790\u4e86\u5b83\u4eec\u7684\u53d1\u5c55\u53ca\u5176\u76f8\u5173\u7684\u6311\u6218\uff0c\u4f8b\u5982\u8ba1\u7b97\u590d\u6742\u6027\u548c\u4f26\u7406\u8003\u91cf\u3002\u901a\u8fc7\u5c06\u8fd9\u4e00\u8fdb\u5c55\u5448\u73b0\u4e3a\u8fde\u8d2f\u7684\u7efc\u8ff0\uff0c\u5e76\u8f85\u4ee5\u5fc5\u8981\u7684\u53c2\u8003\u6587\u732e\uff0c\u6211\u4eec\u4e3a\u7814\u7a76\u4eba\u5458\u63d0\u4f9b\u4e86\u4e00\u4e2a\u5bf9\u8fd1\u5e74\u6765\u540e\u8bad\u7ec3\u6f14\u53d8\u7684\u5168\u9762\u6982\u8ff0\uff0c\u4f5c\u4e3a\u8be5\u9886\u57df\u7684\u57fa\u7840\u8d44\u6e90\u3002<\/li>\n<li>\u7ed3\u6784\u5316\u5206\u7c7b\u6cd5\u548c\u6846\u67b6\uff1a\u6211\u4eec\u4ecb\u7ecd\u4e86\u56fe\u4e2d\u63cf\u8ff0\u7684\u7ed3\u6784\u6027\u5206\u7c7b\u6cd5\u3002 \u5c06\u540e\u8bad\u7ec3\u5206\u7c7b\u4e3a\u4e94\u4e2a\u4e0d\u540c\u7684\u7c7b\u522b\uff0c\u5e76\u5c06\u6570\u636e\u96c6\u7ec4\u7ec7\u4e3a\u4e03\u79cd\u7c7b\u578b\uff0c\u603b\u7ed3\u4e86\u5176\u5728\u4e13\u4e1a\uff0c\u6280\u672f\u548c\u4ea4\u4e92\u5f0f\u57df\u8fdb\u884c\u6784\u67b6\u5e94\u7528\u7a0b\u5e8f\u3002\u8be5\u6846\u67b6\u9610\u660e\u4e86\u8fd9\u4e9b\u65b9\u6cd5\u4e4b\u95f4\u7684\u76f8\u4e92\u5173\u7cfb\u548c\u5b9e\u9645\u5f71\u54cd\uff0c\u4e3a\u5b83\u4eec\u7684\u53d1\u5c55\u63d0\u4f9b\u4e86\u7cfb\u7edf\u7684\u89c6\u89d2\u3002\u901a\u8fc7\u63d0\u4f9b\u660e\u786e\u5b9a\u4e49\u7684\u7c7b\u522b\u548c\u5206\u6790\u89c1\u89e3\uff0c\u6211\u4eec\u63d0\u9ad8\u4e86\u65b0\u624b\u548c\u4e13\u5bb6\u5bf9\u540e\u8bad\u7ec3\u7814\u7a76\u590d\u6742\u6027\u7684\u53ef\u8bbf\u95ee\u6027\u548c\u7406\u89e3\uff0c\u4ece\u800c\u5efa\u7acb\u4e86\u4e00\u4e2a\u5168\u9762\u7684\u6307\u5357\u3002<\/li>\n<li>\u672a\u6765\u65b9\u5411\uff1a\u6211\u4eec\u91cd\u70b9\u4ecb\u7ecd\u4e86\u65b0\u5174\u8d8b\u52bf\uff0c\u7279\u522b\u662f\u5927\u578b\u63a8\u7406\u6a21\u578b\uff08LRM\uff09\u7684\u5174\u8d77\uff0c\u4f8b\u5982o1\u548cDeepSeek-R1\uff0c\u5b83\u4eec\u5229\u7528\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u6765\u7a81\u7834\u63a8\u7406\u7684\u754c\u9650\u3002\u6211\u4eec\u5f3a\u8c03\uff0c\u6301\u7eed\u7684\u8fdb\u6b65\u5bf9\u4e8e\u8fdb\u4e00\u6b65\u589e\u5f3a\u63a8\u7406\u80fd\u529b\u548c\u9886\u57df\u9002\u5e94\u6027\u81f3\u5173\u91cd\u8981\u3002\u6211\u4eec\u7684\u5206\u6790\u786e\u5b9a\u4e86\u5173\u952e\u6311\u6218\uff0c\u5305\u62ec\u53ef\u6269\u5c55\u6027\u9650\u5236\u3001\u4f26\u7406\u5bf9\u9f50\u98ce\u9669\u548c\u591a\u6a21\u6001\u96c6\u6210\u969c\u788d\u3002\u6211\u4eec\u63d0\u51fa\u4e86\u7814\u7a76\u9014\u5f84\uff0c\u4f8b\u5982\u81ea\u9002\u5e94RL\u6846\u67b6\u548c\u5173\u6ce8\u516c\u5e73\u6027\u7684\u4f18\u5316\u3002 \u8fd9\u4e9b\u65b9\u5411\u65e8\u5728\u63a8\u52a8\u540e\u8bad\u7ec3\u5411\u524d\u53d1\u5c55\uff0c\u786e\u4fddLLM\u5b9e\u73b0\u66f4\u9ad8\u7684\u7cbe\u5ea6\u548c\u53ef\u4fe1\u5ea6\uff0c\u4ee5\u6ee1\u8db3\u672a\u6765\u7684\u9700\u6c42\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101062823903.png\" alt=\"\" \/><br \/>\n<center>\u672c\u7814\u7a76\u4e2d\u7efc\u8ff0\u7684\u540e\u8bad\u7ec3\u6280\u672f\u7684\u7ed3\u6784\u6982\u8ff0\uff0c\u8bf4\u660e\u4e86\u65b9\u6cd5\u3001\u6570\u636e\u96c6\u548c\u5e94\u7528\u7684\u9886\u57df<\/center><\/li>\n<\/ul>\n<h4>\u7ec4\u7ec7<\/h4>\n<p>\u672c\u7efc\u8ff0\u7cfb\u7edf\u5730\u7ec4\u7ec7\u8d77\u6765\uff0c\u4ee5\u5168\u9762\u63a2\u8ba8\u540e\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff08PoLM\uff09\uff0c\u6db5\u76d6\u5176\u5386\u53f2\u6f14\u53d8\u3001\u65b9\u6cd5\u3001\u6570\u636e\u96c6\u3001\u5e94\u7528\u548c\u672a\u6765\u8f68\u8ff9\u3002 \u7b2c2\u8282\u63d0\u4f9b\u4e86PoLM\u7684\u5386\u53f2\u6982\u8ff0\u3002 \u7b2c 3 \u8282\u8003\u5bdf\u4e86\u5fae\u8c03\uff0c\u5305\u62ec\u7b2c 3.1 \u8282\u4e2d\u7684\u76d1\u7763\u5fae\u8c03\uff08Supervised Fine-Tuning\uff0cSFT\uff09\u548c\u7b2c 3.3 \u8282\u4e2d\u7684\u5f3a\u5316\u5fae\u8c03\uff08Reinforcement Fine-Tuning\uff0cRFT\uff09\u3002\u7b2c 4 \u8282\u8ba8\u8bba\u4e86\u5bf9\u9f50\uff0c\u6db5\u76d6\u4e86\u7b2c 4.1 \u8282\u4e2d\u7684\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u3001\u7b2c 4.2 \u8282\u4e2d\u7684\u57fa\u4e8e AI \u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning from Human Feedback \uff0cRLAIF\uff09\u548c\u7b2c 4.3 \u8282\u4e2d\u7684\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08Direct Preference Optimization\uff0cDPO\uff09\u3002\u7b2c 5 \u8282\u4fa7\u91cd\u4e8e\u63a8\u7406\uff0c\u5305\u62ec\u7b2c 5.1 \u8282\u4e2d\u7684\u81ea\u6211\u5b8c\u5584\uff08Self-Refinement\uff09\u65b9\u6cd5\u548c\u7b2c 5.2 \u8282\u4e2d\u7684\u7528\u4e8e\u63a8\u7406\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning for Reasoning\uff09\u3002\u7b2c 6 \u8282\u7efc\u8ff0\u4e86\u6548\u7387\u589e\u5f3a\u65b9\u6cd5\uff0c\u5305\u62ec\u7b2c 6.1 \u8282\u4e2d\u7684\u6a21\u578b\u538b\u7f29\u3001\u7b2c 6.2 \u8282\u4e2d\u7684\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff08Parameter-Efficient Fine-Tuning\uff0cPEFT\uff09\u548c\u7b2c 6.3 \u8282\u4e2d\u7684\u77e5\u8bc6\u84b8\u998f\u3002 \u7b2c 7 \u8282\u7814\u7a76\u4e86\u96c6\u6210\u548c\u9002\u5e94\uff0c\u8ba8\u8bba\u4e86\u591a\u6a21\u6001\u65b9\u6cd5\u3001\u9886\u57df\u81ea\u9002\u5e94\u548c\u6a21\u578b\u878d\u5408\u3002\u7b2c 8 \u8282\u7efc\u8ff0\u4e86\u5728\u540e\u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684\u6570\u636e\u96c6\u3002\u7b2c 9 \u8282\u63a2\u8ba8\u4e86 LLM \u5e94\u7528\u3002\u7b2c 10 \u8282\u8bc4\u4f30\u4e86\u5f00\u653e\u95ee\u9898\u548c\u672a\u6765\u65b9\u5411\u3002\u6700\u540e\uff0c\u7b2c 11 \u8282\u4ee5\u603b\u7ed3\u548c\u7814\u7a76\u5c55\u671b\u4f5c\u4e3a\u7ed3\u5c3e\u3002<\/p>\n<h3>\u6982\u8ff0<\/h3>\n<h4>PoLMs\u7684\u5386\u53f2<\/h4>\n<p>LLM \u7684\u53d1\u5c55\u662f\u81ea\u7136\u8bed\u8a00\u5904\u7406 (NLP) \u4e2d\u7684\u4e00\u4e2a\u5173\u952e\u7bc7\u7ae0\uff0c\u540e\u8bad\u7ec3\u65b9\u6cd5\u662f\u5b83\u4eec\u4ece\u5e7f\u4e49\u9884\u8bad\u7ec3\u67b6\u6784\u6f14\u53d8\u4e3a\u4e13\u95e8\u7684\u3001\u4efb\u52a1\u81ea\u9002\u5e94\u7cfb\u7edf\u7684\u5173\u952e\u50ac\u5316\u5242\u3002\u672c\u8282\u63cf\u8ff0\u4e86\u540e\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b (PoLMs) \u7684\u5386\u53f2\u8f68\u8ff9\uff0c\u8ffd\u6eaf\u4e86\u5b83\u4eec\u4ece\u4ee5BERT\u548cGPT\u4e3a\u4ee3\u8868\u7684\u57fa\u7840\u9884\u8bad\u7ec3\u91cc\u7a0b\u7891\u5230\u5f53\u4ee3\u6a21\u578b\u4e2d\u4f53\u73b0\u7684\u590d\u6742\u540e\u8bad\u7ec3\u8303\u5f0f\u7684\u6f14\u53d8\uff0c\u4f8b\u5982 o1 \u548c DeepSeek-R1\u3002 \u5982\u4e0b\u56fe\u6240\u793a\uff0c\u8fd9\u79cd\u53d1\u5c55\u53cd\u6620\u4e86\u4ece\u5efa\u7acb\u5e7f\u6cdb\u7684\u8bed\u8a00\u80fd\u529b\u5230\u589e\u5f3a\u7279\u5b9a\u4efb\u52a1\u7684\u9002\u5e94\u6027\u3001\u4f26\u7406\u5bf9\u9f50\u3001\u63a8\u7406\u590d\u6742\u6027\u548c\u591a\u6a21\u6001\u96c6\u6210\u7684\u8f6c\u53d8\uff0c\u6807\u5fd7\u7740 LLM \u80fd\u529b\u7684\u53d8\u9769\u4e4b\u65c5\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101063725392.png\" alt=\"\" \/><\/p>\n<p><center>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff082018-2025\uff09\u7684\u540e\u8bad\u7ec3\u6280\u672f\u53d1\u5c55\u65f6\u95f4\u7ebf\uff0c\u63cf\u7ed8\u4e86\u5176\u5386\u53f2\u8fdb\u7a0b\u4e2d\u7684\u5173\u952e\u91cc\u7a0b\u7891<\/center>\u73b0\u4ee3 PoLMs \u5386\u53f2\u7684\u5f00\u7aef\u4e0e 2018 \u5e74\u7684\u9884\u8bad\u7ec3\u9769\u547d\u76f8\u543b\u5408\uff0c\u4ee5 BERT \u548c GPT \u7684\u53d1\u5e03\u4e3a\u5148\u5bfc\uff0c\u5b83\u4eec\u91cd\u65b0\u5b9a\u4e49\u4e86 NLP \u57fa\u51c6\u3002BERT \u7684\u53cc\u5411\u81ea\u7f16\u7801\u6846\u67b6\uff0c\u5229\u7528\u4e86 Transformer \u67b6\u6784\u548c\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5728\u6355\u6349\u4e0a\u4e0b\u6587\u76f8\u5173\u6027\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u9002\u7528\u4e8e\u95ee\u9898\u56de\u7b54\u7b49\u4efb\u52a1\uff0c\u800c GPT \u7684\u81ea\u56de\u5f52\u8bbe\u8ba1\u5219\u4f18\u5148\u8003\u8651\u751f\u6210\u4e00\u81f4\u6027\uff0c\u4e3a\u6587\u672c\u751f\u6210\u5960\u5b9a\u4e86\u5148\u4f8b\u3002\u8fd9\u4e9b\u6a21\u578b\u786e\u7acb\u4e86\u201c\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u201d\u8303\u5f0f\uff0c\u968f\u540e\u57282019\u5e74\u901a\u8fc7T5\u8fdb\u884c\u4e86\u6539\u8fdb\uff0c\u5b83\u5c06\u5404\u79cd\u4efb\u52a1\u7edf\u4e00\u5728text-to-text\u6846\u67b6\u4e0b\uff0c\u4fc3\u8fdb\u4e86\u591a\u4efb\u52a1\u5b66\u4e60\uff0c\u5e76\u4e3a\u540e\u7eed\u7684\u8fdb\u6b65\u5960\u5b9a\u4e86\u575a\u5b9e\u7684\u57fa\u7840\u3002<br \/>\n\u4ece2020\u5e74\u5f00\u59cb\uff0cPoLM\u7684\u683c\u5c40\u5f00\u59cb\u53d1\u751f\u91cd\u5927\u53d8\u5316\uff0c\u8fd9\u662f\u7531\u4e8e\u8d8a\u6765\u8d8a\u9700\u8981\u6709\u6548\u5730\u5c06\u9884\u8bad\u7ec3\u6a21\u578b\u9002\u5e94\u4e8e\u5404\u79cd\u4efb\u52a1\uff0c\u540c\u65f6\u4ec5\u4f7f\u7528\u6709\u9650\u7684\u6570\u636e\u3002\u65e9\u671f\u521b\u65b0\uff0c\u5982prefix-tuning\u548cprompt-tuning\uff0c\u5f15\u5165\u4e86\u8f7b\u91cf\u7ea7\u7684\u9002\u5e94\u7b56\u7565\uff0c\u901a\u8fc7\u4fee\u6539\u6a21\u578b\u8f93\u5165\u800c\u4e0d\u662f\u91cd\u65b0\u8bad\u7ec3\u6574\u4e2a\u67b6\u6784\u6765\u5b9e\u73b0\u591a\u4efb\u52a1\u7075\u6d3b\u6027\uff0c\u4ece\u800c\u8282\u7701\u4e86\u8ba1\u7b97\u8d44\u6e90\uff0c\u540c\u65f6\u6269\u5927\u4e86\u9002\u7528\u8303\u56f4\u3002\u8fd9\u4e00\u65f6\u671f\u8fd8\u89c1\u8bc1\u4e86\u4ee5\u4eba\u4e3a\u4e2d\u5fc3\u7684\u4f18\u5316\u7684\u5173\u952e\u8f6c\u53d8\uff0c\u5373\u57282021\u5e74\u51fa\u73b0\u4e86\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\uff0c\u8fd9\u9879\u6280\u672f\u5229\u7528\u4eba\u7c7b\u8bc4\u4f30\u5c06\u6a21\u578b\u8f93\u51fa\u4e0e\u4e3b\u89c2\u504f\u597d\u5bf9\u9f50\uff0c\u589e\u5f3a\u4e86\u5176\u5728\u5bf9\u8bdd\u8bbe\u7f6e\u4e2d\u7684\u5b9e\u7528\u6027\u3002\u52302022\u5e74\uff0cRLHF\u901a\u8fc7\u91c7\u7528Proximal Policy Optimization\uff08PPO\uff09\u800c\u6210\u719f\uff0c\u4f18\u5316\u4e86\u5bf9\u9f50\u7684\u7a33\u5b9a\u6027\uff0c\u5e76\u51cf\u8f7b\u4e86\u5bf9\u566a\u58f0\u53cd\u9988\u7684\u8fc7\u62df\u5408\u30022022\u5e74\u672b\u53d1\u5e03\u7684ChatGPT\u5de9\u56fa\u4e86\u8fd9\u4e9b\u8fdb\u5c55\uff0c\u5c55\u793a\u4e86RLHF\u5728\u521b\u5efa\u54cd\u5e94\u8fc5\u901f\u3001\u4e0e\u7528\u6237\u5bf9\u9f50\u7684LLM\u65b9\u9762\u7684\u53d8\u9769\u6f5c\u529b\uff0c\u5e76\u63a8\u52a8\u4e86PoLM\u7814\u7a76\u7684\u6fc0\u589e\u3002\u4e0e\u6b64\u540c\u65f6\uff0cChain-of-Thought (CoT) prompting \u51fa\u73b0\u4f5c\u4e3a\u4e00\u79cd\u63a8\u7406\u589e\u5f3a\u7b56\u7565\uff0c\u9f13\u52b1\u6a21\u578b\u9610\u660e\u590d\u6742\u4efb\u52a1\u4e2d\u7684\u4e2d\u95f4\u6b65\u9aa4\uff0c\u4ece\u800c\u63d0\u9ad8\u900f\u660e\u5ea6\u548c\u51c6\u786e\u6027\uff0c\u5c24\u5176\u662f\u5728\u903b\u8f91\u63a8\u7406\u548c\u95ee\u9898\u89e3\u51b3\u9886\u57df\u3002<br \/>\n\u57282022\u5e74\u81f32024\u5e74\u4e4b\u95f4\uff0cPoLM\u5b9e\u73b0\u4e86\u591a\u6837\u5316\u53d1\u5c55\uff0c\u4ee5\u89e3\u51b3\u9886\u57df\u7279\u5f02\u6027\uff08domain specificity\uff09\u3001\u4f26\u7406\u7a33\u5065\u6027\uff08ethical robustness\uff09\u548c\u591a\u6a21\u6001\u6574\u5408\uff08multi-modal integration\uff09\u95ee\u9898\uff0c\u8fd9\u53cd\u6620\u51fa\u4e00\u79cd\u5bf9LLM\u6539\u8fdb\u8d8a\u6765\u8d8a\u7ec6\u81f4\u7684\u65b9\u6cd5\u3002\u9886\u57df\u9002\u5e94\u6280\u672f\uff0c\u4f8b\u5982Retrieval-Augmented Generation (RAG) \uff0c\u6d8c\u73b0\u51fa\u6765\u4ee5\u6574\u5408\u5916\u90e8\u77e5\u8bc6\u5e93\uff0c\u4ece\u800c\u4e3a\u4e13\u4e1a\u9886\u57df\u63d0\u4f9b\u4e0a\u4e0b\u6587\u4e30\u5bcc\u7684\u8f93\u51fa\uff0c\u800c\u65e0\u9700\u8fdb\u884c\u5168\u9762\u7684\u91cd\u65b0\u8bad\u7ec3\u2014\u2014\u8fd9\u662f\u9700\u8981\u6700\u65b0\u4fe1\u606f\u7684\u4e13\u4e1a\u5e94\u7528\u7684\u5173\u952e\u8fdb\u6b65\u3002\u4f26\u7406\u5bf9\u9f50\u5de5\u4f5c\u5f97\u5230\u52a0\u5f3a\uff0cDirect Preference Optimization (DPO)\u57282023\u5e74\u901a\u8fc7\u76f4\u63a5\u9488\u5bf9\u4eba\u7c7b\u504f\u597d\u4f18\u5316\u6a21\u578b\u8f93\u51fa\u6765\u7b80\u5316RLHF\uff0c\u7ed5\u8fc7\u4e2d\u95f4\u5956\u52b1\u5efa\u6a21\u4ee5\u63d0\u9ad8\u6548\u7387\u548c\u7a33\u5065\u6027\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5bf9\u591a\u6a21\u6001\u80fd\u529b\u7684\u8ffd\u6c42\u4e5f\u83b7\u5f97\u4e86\u5173\u6ce8\uff0cPaLM-E \u548cFlamingo \u7b49\u6a21\u578b\u7387\u5148\u5b9e\u73b0\u4e86\u89c6\u89c9-\u8bed\u8a00\u7684\u6574\u5408\uff0c\u968f\u540e\u662fBLIP-2\u548cLLaVA\uff0c\u5b83\u4eec\u5c06\u8fd9\u4e9b\u52aa\u529b\u6269\u5c55\u5230\u66f4\u5e7f\u6cdb\u7684\u9886\u57df\uff0c\u5982\u533b\u5b66\u5f71\u50cf\u5b66\u3002 \u6548\u7387\u521b\u65b0\u4e0e\u8fd9\u4e9b\u53d1\u5c55\u540c\u6b65\u8fdb\u884c\uff0c\u7279\u522b\u662f\u901a\u8fc7Mixture of Experts (MoE)\u67b6\u6784\uff1bGoogle\u7684Switch-C Transformer \u57282022\u5e74\u5f15\u5165\u4e86\u7a00\u758f\u6fc0\u6d3b\uff0c\u6a2a\u8de82048\u4e2a\u4e13\u5bb6\uff0c\u62e5\u67091.6\u4e07\u4ebf\u4e2a\u53c2\u6570\uff0c\u800cMixtral \u5b8c\u5584\u4e86\u8fd9\u79cd\u8303\u5f0f\uff0c\u5e73\u8861\u4e86\u53ef\u6269\u5c55\u6027\u548c\u6027\u80fd\u3002\u5728\u6b64\u671f\u95f4\uff0c\u63a8\u7406\u589e\u5f3a\uff0c\u4f8b\u5982\u81ea\u6211\u535a\u5f08 \u548c\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff08MCTS\uff09\u4e0eCoT\u7684\u6574\u5408\uff0c\u901a\u8fc7\u6a21\u62df\u8fed\u4ee3\u63a8\u7406\u9014\u5f84\uff0c\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86LLM\u7684\u51b3\u7b56\u80fd\u529b\uff0c\u4e3a\u4ee5\u9ad8\u7ea7\u63a8\u7406\u4e3a\u91cd\u70b9\u7684\u6a21\u578b\u5960\u5b9a\u4e86\u57fa\u7840\u3002<br \/>\n\u968f\u7740Mixture of Experts (MoE)\u6a21\u578b\u7684\u5174\u8d77\uff0c\u4e00\u9879\u91cd\u5927\u7684\u67b6\u6784\u8fdb\u6b65\u5c55\u5f00\uff0c\u8be5\u6a21\u578b\u4e0e\u4f20\u7edf\u7684\u5bc6\u96c6\u67b6\u6784\u4e0d\u540c\uff0c\u901a\u8fc7\u52a8\u6001\u6fc0\u6d3b\u9009\u62e9\u6027\u53c2\u6570\u5b50\u96c6\uff0c\u4ece\u800c\u4f18\u5316\u4e86\u8ba1\u7b97\u6548\u7387\uff0c\u540c\u65f6\u5bb9\u7eb3\u4e86\u5e7f\u6cdb\u7684\u53c2\u6570\u89c4\u6a21\u3002Google\u7684Switch-C Transformer \u5728 2022 \u4e2d\u542f\u7528\u4e86\u8be5\u8303\u5f0f\uff0c\u5176\u4e2d\u5305\u542b1.6\u4e07\u4ebf\u4e2a\u53c2\u6570\uff0c\u5206\u5e03\u57282048\u540d\u4e13\u5bb6\u4e2d\uff0c\u8fd9\u662f\u4e00\u79cd\u5177\u6709\u7a81\u7834\u6027\u7684\u65b9\u6cd5\uff0c\u5e73\u8861\u4e86\u8d44\u6e90\u7684\u9700\u6c42\u4e0e\u6027\u80fd\u63d0\u9ad8\u3002\u968f\u540e\u7684\u8fed\u4ee3\uff0c\u5982 Mixtral \u548c DeepSeek V2.5\uff0c\u540e\u8005\u5229\u7528\u4e86 2360 \u4ebf\u4e2a\u603b\u53c2\u6570\uff0c\u5176\u4e2d 160 \u4e2a\u4e13\u5bb6\u4f7f\u7528\u4e86 210 \u4ebf\u4e2a\u6709\u6548\u53c2\u6570\u2014\u8fdb\u4e00\u6b65\u5b8c\u5584\u4e86\u8fd9\u4e00\u6846\u67b6\uff0c\u5728 LMSYS \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u53d6\u5f97\u4e86 state-of-the-art \u7684\u7ed3\u679c\uff0c\u5e76\u8bc1\u660e\u4e86\u7a00\u758f\u7684 MoE \u67b6\u6784\u5728\u53ef\u6269\u5c55\u6027\u548c\u6709\u6548\u6027\u65b9\u9762\u90fd\u53ef\u4ee5\u4e0e\u5bc6\u96c6\u6a21\u578b\u76f8\u5ab2\u7f8e\u3002\u8fd9\u4e9b\u53d1\u5c55\u5f3a\u8c03\u4e86\u4ee5\u6548\u7387\u4e3a\u4e2d\u5fc3\u7684POLMS\u7684\u8f6c\u53d8\uff0c\u4f7fLLM\u53ef\u4ee5\u901a\u8fc7\u51cf\u5c11\u7684\u8ba1\u7b97\u5f00\u9500\u6765\u5904\u7406\u590d\u6742\u7684\u4efb\u52a1\uff0c\u8fd9\u662f\u6269\u5927\u5176\u5b9e\u9645\u9002\u7528\u6027\u7684\u5173\u952e\u4e00\u6b65\u30022025 \uff0cDeepSeek-R1 \u5728Polms Innovation\u4e2d\u51fa\u73b0\u4e86\u5730\u6807\uff0c\u504f\u79bb\u4e86\u4f20\u7edf\u7684\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u4f9d\u8d56\u5bf9\u62e5\u62b1\u601d\u60f3\u94fe\uff08COT\uff09\u63a8\u7406\u548c\u63a2\u7d22\u6027RL\u7b56\u7565\u7684\u4f9d\u8d56\u3002 \u8be5\u6a21\u578b\u4ee5DeepSeek-R1-Zero\u4e3a\u4f8b\uff0c\u8be5\u6a21\u578b\u6574\u5408\u4e86\u81ea\u6211\u9a8c\u8bc1\uff0c\u53cd\u5c04\u548c\u6269\u5c55\u7684COT\u751f\u6210\uff0c\u9a8c\u8bc1\u4e86RL\u9a71\u52a8\u7684\u63a8\u7406\u6fc0\u52b1\u63aa\u65bd\u5728\u5f00\u653e\u7684\u7814\u7a76\u8303\u5f0f\u4e2d\uff0c\u5c06\u84b8\u998f\u6280\u672f\u5f15\u5165\u4e86\u4ece\u8f83\u5927\u7684\u8f83\u5c0f\u7684\u67b6\u6784\u8f6c\u79fb\u6210\u719f\u7684\u7406\u6027\u6a21\u5f0f\u3002 \u4e0e\u72ec\u7acb\u7684RL\u8bad\u7ec3\u76f8\u6bd4\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u53ef\u4ee5\u4ea7\u751f\u4f18\u8d8a\u7684\u6027\u80fd\uff0c\u800c\u4e14\u8fd8\u9884\u793a\u7740\u4ee5\u63a8\u7406\u4e3a\u4e2d\u5fc3\u7684LLM\u7684\u53ef\u6269\u5c55\u6027\u8303\u5f0f\uff0c\u53ef\u4ee5\u89e3\u51b3\u8ba1\u7b97\u6548\u7387\u548c\u4efb\u52a1\u9002\u5e94\u80fd\u529b\u7684\u6301\u7eed\u6311\u6218\u3002<\/p>\n<h4>PoLM\u7684\u516c\u5f0f\u57fa\u7840<\/h4>\n<h5>\u7b56\u7565\u4f18\u5316\u7684\u539f\u7406<\/h5>\n<p>\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09\u7b97\u6cd5\u662f\u4e00\u79cd\u5173\u952e\u7684\u52a0\u5f3a\u5b66\u4e60\u6280\u672f\uff0c\u5728\u8bf8\u5982\u4eba\u7c7b\u53cd\u9988\uff08RLHF\uff09 \u7b49\u73af\u5883\u4e2d\u5c24\u5176\u6709\u7528\uff0c\u5728\u7ef4\u6301\u7a33\u5b9a\u6027\u548c\u6548\u7387\u7684\u5730\u65b9\u662f\u91cd\u8981\u7684\u3002PPO\u901a\u8fc7\u7ea6\u675f\u7b56\u7565\u66f4\u65b0\u7684\u5927\u5c0f\u6765\u5b9e\u73b0\u8fd9\u4e9b\u76ee\u6807\uff0c\u4ece\u800c\u786e\u4fdd\u5bf9\u6a21\u578b\u884c\u4e3a\u7684\u53d8\u5316\u9010\u6e10\u4e14\u53d7\u63a7\uff0c\u4ece\u800c\u9632\u6b62\u6548\u679c\u7684\u707e\u96be\u6027\u53d8\u5316\u3002 \u5f53\u5bf9\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u7684\u7b56\u7565\u66f4\u65b0\u53ef\u80fd\u5bfc\u81f4\u4e0d\u826f\u6216\u4e0d\u53ef\u9884\u6d4b\u7684\u884c\u4e3a\u65f6\uff0c\u8fd9\u4e00\u70b9\u5c24\u5176\u91cd\u8981\u3002<br \/>\n<strong>\u5b9a\u4e49\uff1a<\/strong>\u5728PPO\u7684\u80cc\u666f\u4e0b\uff0c\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s_t \\in \\mathcal{S}<\/span>\u8868\u793a\u65f6\u95f4<span class=\"katex-eq\" data-katex-display=\"false\">t<\/span>\u7684\u73af\u5883\uff0c\u5176\u4e2d\u5305\u62ec\u6a21\u578b\u505a\u51fa\u51b3\u5b9a\u6240\u9700\u7684\u6240\u6709\u76f8\u5173\u4fe1\u606f\u3002\u52a8\u4f5c<span class=\"katex-eq\" data-katex-display=\"false\">a_t \\in \\mathcal{A}(s_t)<\/span>\u8868\u793a\u6a21\u578b\u5728\u7ed9\u5b9a\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s_t<\/span>\u4e0b\u505a\u51fa\u7684\u9009\u62e9\u3002\u6b64\u52a8\u4f5c\u662f\u6a21\u578b\u505a\u51fa\u7684\u4e00\u7cfb\u5217\u51b3\u7b56\u7684\u4e00\u90e8\u5206\u3002\u6267\u884c\u52a8\u4f5c\u540e\uff0c\u667a\u80fd\u4f53\u6536\u5230\u5956\u52b1<span class=\"katex-eq\" data-katex-display=\"false\">r_t \\in \\mathbb{R}<\/span>\uff0c\u8be5\u5956\u52b1\u4f5c\u4e3a\u6765\u81ea\u73af\u5883\u7684\u53cd\u9988\uff0c\u8868\u660e\u6240\u91c7\u53d6\u7684\u52a8\u4f5c\u7684\u6210\u529f\u4e0e\u5931\u8d25\u3002\u4f18\u52bf\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">A^\\pi(s, a)<\/span>\u8861\u91cf\u5728\u5f53\u524d\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span>\u4e0b\uff0c\u91c7\u53d6\u884c\u52a8<span class=\"katex-eq\" data-katex-display=\"false\">a<\/span>\u5728\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s<\/span>\u4e2d\u6bd4\u8be5\u72b6\u6001\u4e0b\u884c\u52a8\u7684\u9884\u671f\u503c\u6709\u591a\u5927\u4f18\u52bf\u3002\u5b83\u88ab\u6b63\u5f0f\u5b9a\u4e49\u4e3a\u52a8\u4f5c\u503c\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">Q^\\pi(s, a)<\/span>\u548c\u72b6\u6001\u503c\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">V^\\pi(s)<\/span>\u4e4b\u95f4\u7684\u5dee\u5f02\uff0c\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">A^\\pi(s, a) = Q^\\pi(s, a) - V^\\pi(s)<\/span><\/center>\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">Q^\\pi(s, a)<\/span>\u8868\u793a\u901a\u8fc7\u5728\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s<\/span>\u4e2d\u91c7\u53d6\u884c\u52a8<span class=\"katex-eq\" data-katex-display=\"false\">a<\/span>\u5e76\u9075\u5faa\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span>\u83b7\u5f97\u7684\u9884\u671f\u7d2f\u79ef\u5956\u52b1\uff0c\u800c<span class=\"katex-eq\" data-katex-display=\"false\">V^\\pi(s)<\/span>\u662f\u4ece\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s<\/span>\u5f00\u59cb\u5e76\u9075\u5faa\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span>\u7684\u9884\u671f\u7d2f\u79ef\u5956\u52b1\u3002\u8fd9\u4e24\u4e2a\u51fd\u6570\u90fd\u8003\u8651\u4e86\u672a\u6765\u7684\u5956\u52b1\uff0c\u5e76\u6309\u56e0\u5b50<span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span>\u6298\u6263\u3002<br \/>\n<strong>\u7b56\u7565\u66f4\u65b0\uff1a<\/strong>PPO\u7b97\u6cd5\u901a\u8fc7\u6839\u636e\u4f18\u52bf\u51fd\u6570\u8fdb\u884c\u589e\u91cf\u66f4\u65b0\u6765\u4f18\u5316\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_\\theta<\/span>\u3002\u7b56\u7565\u66f4\u65b0\u4f7f\u7528\u88c1\u526a\u76ee\u6807\u51fd\u6570\u6267\u884c\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">L^{CLIP}(\\theta) = \\hat{\\mathbb{E}}_t\\left[\\min\\left(r_t(\\theta)\\hat{A}_t,, \\text{clip}\\left(r_t(\\theta), 1 - \\varepsilon, 1 + \\varepsilon\\right)\\hat{A}_t\\right)\\right]<\/span><\/center>\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">r_t(\\theta)<\/span>\u8868\u793a\u5f53\u524d\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_\\theta<\/span>\u4e0b\u91c7\u53d6\u884c\u52a8<span class=\"katex-eq\" data-katex-display=\"false\">a_t<\/span>\u7684\u6982\u7387\u4e0e\u65e7\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\theta_{\\text{old}}}<\/span>\u4e0b\u7684\u6982\u7387\u4e4b\u6bd4\u3002\u8be5\u6bd4\u7387\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">r_t(\\theta)&gt;[katex]r_t(\\theta) = \\frac{\\pi_\\theta(a_t \\mid s_t)}{\\pi_{\\theta_{\\text{old}}}(a_t \\mid s_t)}<\/span><\/center>\u672f\u8bed<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{A}_t<\/span>\u662f\u5728\u65f6\u95f4\u6b65<span class=\"katex-eq\" data-katex-display=\"false\">t<\/span>\u4f30\u8ba1\u7684\u4f18\u52bf\uff0c\u88c1\u526a\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">\\text{clip}(r_t(\\theta), 1 - \\varepsilon, 1 + \\varepsilon)<\/span>\u5c06\u7b56\u7565\u66f4\u65b0\u9650\u5236\u5728\u4e00\u4e2a\u5b89\u5168\u8303\u56f4\u5185\uff0c\u7531\u8d85\u53c2\u6570<span class=\"katex-eq\" data-katex-display=\"false\">\\varepsilon<\/span>\u63a7\u5236\u3002\u8fd9\u79cd\u88c1\u526a\u673a\u5236\u786e\u4fdd\u66f4\u65b0\u4e0d\u4f1a\u4e0e\u4e4b\u524d\u7684\u7b56\u7565\u5dee\u5f02\u592a\u5927\uff0c\u4ece\u800c\u5728\u8bad\u7ec3\u671f\u95f4\u4fdd\u6301\u7a33\u5b9a\u6027\u3002<br \/>\n<strong>\u503c\u51fd\u6570\u66f4\u65b0\uff1a<\/strong>\u503c\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">V_\\phi<\/span>\u4f30\u8ba1\u5728\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_\\theta<\/span>\u4e0b\u4ece\u7ed9\u5b9a\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s_t<\/span>\u83b7\u5f97\u7684\u9884\u671f\u7d2f\u79ef\u5956\u52b1\u3002\u4e3a\u4e86\u786e\u4fdd\u503c\u51fd\u6570\u63d0\u4f9b\u51c6\u786e\u7684\u4f30\u8ba1\uff0c\u5b83\u901a\u8fc7\u6700\u5c0f\u5316\u9884\u6d4b\u503c\u548c\u5b9e\u9645\u5956\u52b1\u4e4b\u95f4\u7684\u5747\u65b9\u8bef\u5dee\u6765\u8fdb\u884c\u4f18\u5316\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\phi_{k+1} = \\arg\\min_\\phi , \\mathbb{E}{s_t \\sim \\pi{\\theta_k}}\\left[\\left(V_\\phi(s_t) - R(s_t)\\right)^2\\right]<\/span><\/center>\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">R(s_t)<\/span>\u662f\u4ece\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s_t<\/span>\u83b7\u5f97\u7684\u5b9e\u9645\u7d2f\u79ef\u5956\u52b1\uff0c\u800c<span class=\"katex-eq\" data-katex-display=\"false\">V_\\phi(s_t)<\/span>\u662f\u5728\u5f53\u524d\u7b56\u7565\u4e0b\u4f30\u8ba1\u7684\u503c\u3002\u76ee\u6807\u662f\u8c03\u6574\u53c2\u6570<span class=\"katex-eq\" data-katex-display=\"false\">\\phi<\/span>\u4ee5\u6700\u5c0f\u5316\u9884\u6d4b\u5956\u52b1\u548c\u5b9e\u9645\u5956\u52b1\u4e4b\u95f4\u7684\u5dee\u5f02\uff0c\u4ece\u800c\u63d0\u9ad8\u503c\u51fd\u6570\u7684\u51c6\u786e\u6027\u3002<\/p>\n<h5>RLHF\u539f\u7406<\/h5>\n<p>\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60 (RLHF) \u662f\u4e00\u79cd\u5173\u952e\u65b9\u6cd5\uff0c\u5b83\u901a\u8fc7\u5728\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u5229\u7528\u4eba\u7c7b\u751f\u6210 \u7684\u53cd\u9988\u6765\u4f7f\u6a21\u578b\u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u3002\u8fd9\u79cd\u65b9\u6cd5\u5305\u542b\u4e00\u4e2a\u660e\u786e\u6355\u83b7\u4eba\u7c7b\u8f93\u5165\u7684\u5956\u52b1\u51fd\u6570\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u66f4\u597d\u5730\u9002\u5e94\u7528\u6237\u504f\u597d\u548c\u5b9e\u9645\u5e94\u7528\u3002<br \/>\n<strong>\u5b9a\u4e49:<\/strong>\u5728 RLHF \u4e2d\uff0c\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span> \u5728\u8bcd\u6c47\u8868 <span class=\"katex-eq\" data-katex-display=\"false\">\\Sigma<\/span> \u4e2d\u751f\u6210\u6807\u8bb0\u5e8f\u5217\u7684\u6982\u7387\u5206\u5e03\u3002\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span> \u4ece\u8f93\u5165\u7a7a\u95f4 <span class=\"katex-eq\" data-katex-display=\"false\">X=\\Sigma^{\\le m}<\/span> \u4ea7\u751f\u4e00\u4e2a\u6807\u8bb0\u5e8f\u5217 <span class=\"katex-eq\" data-katex-display=\"false\">x_0,x_1,\\dots,x_{n-1}<\/span>\uff0c\u5176\u4e2d\u6bcf\u4e2a\u6807\u8bb0\u90fd\u6761\u4ef6\u6027\u5730\u4f9d\u8d56\u4e8e\u4e4b\u524d\u7684\u6807\u8bb0\u3002\u6a21\u578b\u7684\u8f93\u51fa\u7531\u4ee5\u4e0b\u6761\u4ef6\u6982\u7387\u5206\u5e03\u5b9a\u4e49\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\rho(x_0\\cdots x_{n-1})=\\prod_{0\\le k\uff1cn}\\rho(x_k\\mid x_0\\cdots x_{k-1})<\/span><\/center>\u8be5\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span> \u5728\u7531\u8f93\u5165\u7a7a\u95f4 <span class=\"katex-eq\" data-katex-display=\"false\">X<\/span>\u3001\u5173\u4e8e <span class=\"katex-eq\" data-katex-display=\"false\">X<\/span> \u7684\u6570\u636e\u5206\u5e03 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal D<\/span> \u548c\u8f93\u51fa\u7a7a\u95f4 <span class=\"katex-eq\" data-katex-display=\"false\">Y=\\Sigma^{\\le n}<\/span> \u5b9a\u4e49\u7684\u4efb\u52a1\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u4f8b\u5982\uff0c\u5728\u6587\u672c\u6458\u8981\u4e2d\uff0c\u4e00\u4e2a GPT-2 \u6a21\u578b\u4f7f\u7528 RLHF \u8fdb\u884c\u8bad\u7ec3\uff0c\u5176\u4e2d\u4efb\u52a1\u6d89\u53ca\u6839\u636e\u6570\u636e\u96c6\uff08\u5982 CNN\/DailyMail \u548c TL;DR\uff09\u9884\u6d4b\u6587\u672c\u6458\u8981\u3002<br \/>\n<strong>\u76ee\u6807\u51fd\u6570:<\/strong>\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span> \u662f\u4e00\u4e2a\u4e0e\u539f\u59cb\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span> \u5177\u6709\u76f8\u540c\u7ed3\u6784\u7684\u8bed\u8a00\u6a21\u578b\u3002\u6700\u521d\uff0c\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span> \u8bbe\u7f6e\u4e3a\u7b49\u4e8e <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span>\u3002\u76ee\u6807\u662f\u901a\u8fc7\u4f18\u5316\u7b56\u7565\u6765\u6700\u5927\u5316\u8f93\u5165\u8f93\u51fa\u5bf9 katex[\/katex] \u7684\u671f\u671b\u5956\u52b1 <span class=\"katex-eq\" data-katex-display=\"false\">R(x,y)<\/span>\u3002\u5956\u52b1\u51fd\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">R(x,y):X\\times Y\\to\\mathbb R<\/span> \u4e3a\u6bcf\u4e2a\u8f93\u5165\u8f93\u51fa\u5bf9\u5206\u914d\u4e00\u4e2a\u6807\u91cf\u503c\uff0c\u5e76\u4e14\u901a\u8fc7\u89e3\u51b3\u4ee5\u4e0b\u6700\u5927\u5316\u95ee\u9898\u6765\u83b7\u5f97\u6700\u4f18\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*<\/span>\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\pi^*=\\arg\\max_\\pi \\mathbb E[R]=\\mathbb E_{x\\sim\\mathcal D,,y\\sim\\pi(\\cdot\\mid x)}\\Bigl[R(x,y)\\Bigr].\n\n<\/span><\/center>\u8fd9\u4e2a\u76ee\u6807\u51fd\u6570\u4ee3\u8868\u4e00\u4e2a\u6807\u51c6\u7684 RL \u95ee\u9898\uff0c\u5176\u4e2d\u6a21\u578b\u5b66\u4f1a\u901a\u8fc7\u4e0e\u73af\u5883\u4ea4\u4e92\u6765\u6700\u5927\u5316\u671f\u671b\u5956\u52b1\uff0c\u5e76\u4ee5\u4eba\u7c7b\u53cd\u9988\u4e3a\u6307\u5bfc\u3002<\/p>\n<h5>DPO\u539f\u7406<\/h5>\n<p><strong>\u76ee\u6807\u51fd\u6570\uff1a<\/strong>\u6211\u4eec\u4ece\u4e0e\u4e4b\u524d\u65b9\u6cd5\u76f8\u540c\u7684 RL \u76ee\u6807\u51fa\u53d1\uff0c\u5728\u901a\u7528\u5956\u52b1\u51fd\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">r<\/span> \u4e0b\u8003\u8651 KL-\u7ea6\u675f\u7684\u5956\u52b1\u6700\u5927\u5316\u95ee\u9898\u3002\u6700\u4f18\u7b56\u7565\u6ee1\u8db3<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">&gt;[katex]\n\n\\pi_r(y\\mid x)=\\frac{1}{Z(x)}\\pi_{\\text{ref}}(y\\mid x)\\exp!\\Bigl(\\frac{1}{\\beta}r(x,y)\\Bigr),\n\n<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">Z(x)<\/span> \u662f\u5206\u533a\u51fd\u6570\uff0c\u786e\u4fdd\u5bf9\u6240\u6709\u53ef\u80fd\u8f93\u51fa\u5f52\u4e00\u5316\u3002\u5373\u4f7f\u4ec5\u7528 MLE \u4f30\u8ba1\u7684\u5956\u52b1 <span class=\"katex-eq\" data-katex-display=\"false\">r_\\phi\\approx r^*<\/span>\uff0c\u4e5f\u53ef\u8fd1\u4f3c <span class=\"katex-eq\" data-katex-display=\"false\">Z(x)<\/span>\uff0c\u4ece\u800c\u7b80\u5316\u4f18\u5316\uff0c\u4f7f\u7b56\u7565\u80fd\u76f4\u63a5\u4f9d\u636e\u4eba\u7c7b\u504f\u597d\u8c03\u6574\u3002<br \/>\n<strong>\u504f\u597d\u6a21\u578b:<\/strong>\u91c7\u7528 Bradley-Terry \u6a21\u578b\u523b\u753b\u4e24\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">y_1,y_2<\/span> \u95f4\u7684\u504f\u597d\uff1a\u6700\u4f18\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*<\/span> \u6ee1\u8db3<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">&gt;[katex]\n\np^(y_1\\succ y_2\\mid x)=\\frac{1}{1+\\exp!\\Bigl(\\beta\\log\\frac{\\pi^(y_2\\mid x)}{\\pi_{\\text{ref}}(y_2\\mid x)}-\\beta\\log\\frac{\\pi^*(y_1\\mid x)}{\\pi_{\\text{ref}}(y_1\\mid x)}\\Bigr)},\n\n<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">p^*(y_1\\succ y_2\\mid x)<\/span> \u8868\u793a\u5728\u8f93\u5165 <span class=\"katex-eq\" data-katex-display=\"false\">x<\/span> \u4e0b\u4eba\u7c7b\u66f4\u504f\u597d <span class=\"katex-eq\" data-katex-display=\"false\">y_1<\/span> \u800c\u975e <span class=\"katex-eq\" data-katex-display=\"false\">y_2<\/span> \u7684\u6982\u7387\uff0c\u4ece\u800c\u5c06\u4eba\u7c7b\u504f\u597d\u663e\u5f0f\u878d\u5165\u7b56\u7565\u4f18\u5316\u3002<\/p>\n<h5>GRPO\u539f\u7406<\/h5>\n<p><strong>\u5b9a\u4e49\uff1a<\/strong>\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08Group Relative Policy Optimization, GRPO\uff09\u7b97\u6cd5\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09\u7b97\u6cd5\u7684\u53d8\u4f53\uff0c\u9996\u6b21\u5728DeepSeek\u4e4b\u524d\u7684\u5de5\u4f5cDeepSeekMath\u4e2d\u5f15\u5165\u3002GRPO\u7701\u7565\u4e86\u8bc4\u8bba\u5458\u6a21\u578b\uff0c\u800c\u662f\u4f7f\u7528\u7ec4\u5f97\u5206\u4f30\u8ba1\u57fa\u7ebf\uff0c\u4e0ePPO\u76f8\u6bd4\uff0c\u8fd9\u663e\u7740\u51cf\u5c11\u4e86\u8bad\u7ec3\u8d44\u6e90\u6d88\u8017\u3002<br \/>\n<strong>\u76ee\u6807\u51fd\u6570:<\/strong>\u5bf9\u4e8e\u6bcf\u4e2a\u95ee\u9898 <span class=\"katex-eq\" data-katex-display=\"false\">q<\/span>\uff0cGRPO \u4ece\u65e7\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\theta_{\\mathrm{old}}}<\/span> \u91c7\u6837\u4e00\u7ec4\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">{o_1,\\dots,o_G}<\/span>\uff0c\u968f\u540e\u901a\u8fc7\u6700\u5927\u5316\u4ee5\u4e0b\u76ee\u6807\u66f4\u65b0\u7b56\u7565\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">&gt;[katex]\n\n\\mathcal{J}^{\\mathrm{GRPO}}(\\theta)=\n\n\\mathbb{E}{\\substack{q\\sim P(Q)\\ {o_i}{i=1}^G\\sim\\pi_{\\theta_{\\mathrm{old}}}(\\cdot|q)}}\n\n\\frac{1}{G}\\sum_{i=1}^G\\frac{1}{|o_i|}\\sum_{t=1}^{|o_i|}\n\n\\Bigl{\n\n\\min\\Bigl[\n\n\\frac{\\pi_\\theta(o_{i,t}|q,o_{i,\uff1ct})}{\\pi_{\\theta_{\\mathrm{old}}}(o_{i,t}|q,o_{i,\uff1ct})}\\hat{A}{i,t},;\n\n\\mathrm{clip}!\\left(\\frac{\\pi\\theta(o_{i,t}|q,o_{i,\uff1ct})}{\\pi_{\\theta_{\\mathrm{old}}}(o_{i,t}|q,o_{i,\uff1ct})},1!-!\\varepsilon,1!+!\\varepsilon\\right)\\hat{A}{i,t}\n\n\\Bigr]\n\n-\\beta D{\\mathrm{KL}}!\\bigl[\\pi_\\theta|\\pi_{\\mathrm{ref}}\\bigr]\n\n\\Bigr},\n\n<\/span><\/center>\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\varepsilon,\\beta<\/span> \u4e3a\u8d85\u53c2\u6570\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{A}_{i,t}<\/span> \u4e3a\u7ec4\u5185\u76f8\u5bf9\u4f18\u52bf\uff0c\u7531\u540c\u4e00\u7ec4\u8f93\u51fa\u4e4b\u95f4\u7684\u76f8\u5bf9\u5956\u52b1\u8ba1\u7b97\u5f97\u51fa\u3002<\/p>\n<h3>\u7528\u4e8e\u5fae\u8c03\u7684PoLM<\/h3>\n<p>\u5fae\u8c03\u6784\u6210\u4e86\u5c06\u9884\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b (LLM) \u9002\u914d\u5230\u7279\u5b9a\u4efb\u52a1\u7684\u57fa\u77f3\uff0c\u901a\u8fc7\u6709\u9488\u5bf9\u6027\u7684\u53c2\u6570\u8c03\u6574\u6765\u63d0\u5347\u5176\u80fd\u529b\u3002\u6b64\u8fc7\u7a0b\u5229\u7528\u5df2\u6807\u6ce8\u7684\u6216\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6570\u636e\u96c6\u6765\u4f18\u5316\u6027\u80fd\uff0c\u5f25\u5408\u4e86\u901a\u7528\u9884\u8bad\u7ec3\u548c\u7279\u5b9a\u4e8e\u9886\u57df\u7684\u9700\u6c42\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002\u672c\u7ae0\u63a2\u8ba8\u4e86\u4e09\u79cd\u4e3b\u8981\u7684\u5fae\u8c03\u8303\u5f0f\uff1a\u76d1\u7763\u5fae\u8c03 (Fine-Tuning)\uff0c\u5b83\u4f7f\u7528\u6807\u6ce8\u6570\u636e\u96c6\u6765\u63d0\u9ad8\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u51c6\u786e\u6027\uff1b\u81ea\u9002\u5e94\u5fae\u8c03 (Adaptive Fine-Tuning)\uff0c\u5b83\u901a\u8fc7\u6307\u4ee4\u8c03\u4f18\u548c\u57fa\u4e8e\u63d0\u793a\u7684\u65b9\u6cd5\u5b9a\u5236\u6a21\u578b\u884c\u4e3a\uff1b\u4ee5\u53ca \u5f3a\u5316\u5fae\u8c03 (Reinforcement Fine-Tuning)\uff0c\u5b83\u96c6\u6210\u4e86\u5f3a\u5316\u5b66\u4e60\uff0c\u57fa\u4e8e\u5956\u52b1\u4fe1\u53f7\u8fed\u4ee3\u5730\u4f18\u5316\u8f93\u51fa\uff0c\u901a\u8fc7\u52a8\u6001\u4ea4\u4e92\u4fc3\u8fdb\u6301\u7eed\u6539\u8fdb\u3002<\/p>\n<h4>\u76d1\u7763\u5fae\u8c03<\/h4>\n<p>\u76d1\u7763\u5fae\u8c03 (SFT) \u901a\u8fc7\u5229\u7528\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6807\u6ce8\u6570\u636e\u96c6\uff0c\u5c06\u9884\u8bad\u7ec3 LLM \u9002\u914d\u5230\u7279\u5b9a\u4efb\u52a1\u3002\u4e0e\u4f9d\u8d56\u6307\u4ee4\u63d0\u793a\u7684\u6307\u4ee4\u8c03\u4f18\u4e0d\u540c\uff0cSFT \u4f7f\u7528\u6807\u6ce8\u6570\u636e\u76f4\u63a5\u8c03\u6574\u6a21\u578b\u53c2\u6570\uff0c\u4ece\u800c\u4ea7\u751f\u65e2\u7cbe\u786e\u53c8\u5177\u6709\u4e0a\u4e0b\u6587\u5173\u8054\u7684\u6a21\u578b\uff0c\u540c\u65f6\u4fdd\u6301\u5e7f\u6cdb\u7684\u6cdb\u5316\u80fd\u529b\u3002 SFT \u5f25\u5408\u4e86\u9884\u8bad\u7ec3\u671f\u95f4\u7f16\u7801\u7684\u5e7f\u6cdb\u8bed\u8a00\u77e5\u8bc6\u548c\u76ee\u6807\u5e94\u7528\u7684\u7ec6\u5fae\u9700\u6c42\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002 \u9884\u8bad\u7ec3 LLM \u901a\u8fc7\u63a5\u89e6\u5927\u91cf\u7684\u8bed\u6599\u5e93\uff0c\u83b7\u5f97\u4e86\u5e7f\u4e49\u7684\u8bed\u8a00\u6a21\u5f0f\uff0c\u51cf\u5c11\u4e86\u5bf9\u5927\u91cf\u7279\u5b9a\u9886\u57df\u6570\u636e\u7684\u4f9d\u8d56\u3002 \u6a21\u578b\u9009\u62e9\u81f3\u5173\u91cd\u8981\uff1a\u50cf T5 \u8fd9\u6837\u7684\u5c0f\u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u4e14\u6570\u636e\u96c6\u6709\u9650\u7684\u73af\u5883\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u800c\u50cf GPT-4 \u8fd9\u6837\u7684\u5927\u6a21\u578b\u5219\u5229\u7528\u5176\u5353\u8d8a\u7684\u80fd\u529b\u5728\u590d\u6742\u3001\u6570\u636e\u4e30\u5bcc\u7684\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\u3002<\/p>\n<h5>SFTd\u7684\u6570\u636e\u96c6\u51c6\u5907<\/h5>\n<p>\u6784\u5efa\u9ad8\u8d28\u91cf\u7684 SFT \u6570\u636e\u96c6\u662f\u4e00\u4e2a\u591a\u65b9\u9762\u7684\u8fc7\u7a0b\uff0c\u5bf9\u5fae\u8c03\u7684\u6210\u529f\u81f3\u5173\u91cd\u8981\u3002<br \/>\n<strong>SFT \u6570\u636e\u96c6\u6784\u5efa\uff1a<\/strong>SFT\u6570\u636e\u96c6\u901a\u5e38\u6784\u9020\u6210<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}&gt;[katex]\\mathcal{D}={(I_k,X_k)}_{k=1}^N<\/span>\uff0c<br \/>\n\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">I_k<\/span> \u4e3a\u6307\u4ee4\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">X_k<\/span> \u662f\u5bf9\u5e94\u7684\u671f\u671b\u8f93\u51fa\u3002Self-Instruct \u7b49\u65b9\u6cd5\u901a\u8fc7\u5408\u6210\u65b0\u6307\u4ee4-\u8f93\u51fa\u5bf9\u5e76\u91c7\u7528 ROUGE-L \u7b49\u6307\u6807\u53bb\u91cd\uff0c\u63d0\u5347\u591a\u6837\u6027\u3002<br \/>\n<strong>SFT \u6570\u636e\u96c6\u7b5b\u9009:<\/strong>\u5f15\u5165\u8d28\u91cf\u51fd\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">r(\\cdot)<\/span> \u5bf9\u6bcf\u5bf9 katex[\/katex] \u6253\u5206\uff0c\u53ea\u4fdd\u7559\u5f97\u5206\u4e0d\u4f4e\u4e8e\u9608\u503c <span class=\"katex-eq\" data-katex-display=\"false\">\\tau<\/span> \u7684\u5b50\u96c6\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">&gt;[katex]\n\n\\mathcal{D}&#039;=\\bigl{(I_k,X_k)\\in\\mathcal{D}\\mid r(I_k,X_k)\\ge\\tau\\bigr}.\n\n<\/span><\/center>\u6307\u4ee4\u9075\u5faa\u96be\u5ea6\uff08IFD\uff09\u662f\u4e00\u79cd\u5e38\u7528\u6307\u6807\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\n\nr_\\theta(Q,A)=\n\n\\frac{\\sum_{i=1}^N \\log P(w_i^A\\mid Q,w_{\uff1ci}^A;\\theta)}\n\n{\\sum_{i=1}^N \\log P(w_i^A\\mid w_{\uff1ci}^A;\\theta)},\n\n<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">Q<\/span> \u4e3a\u6307\u4ee4\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">A<\/span> \u4e3a\u671f\u671b\u54cd\u5e94\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\theta<\/span> \u662f\u6a21\u578b\u53c2\u6570\u3002IFD \u4f4e\u4e8e\u8bbe\u5b9a\u9608\u503c\u7684\u6837\u672c\u88ab\u5254\u9664\uff0c\u5f97\u5230\u7cbe\u70bc\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}&#039;<\/span>\u3002<\/p>\n<p>SFT \u6570\u636e\u96c6\u8bc4\u4f30\uff1a\u9009\u53d6\u9ad8\u8d28\u91cf\u5b50\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}_{\\text{eval}}\\subseteq\\mathcal{D}&#039;<\/span> \u4f5c\u4e3a\u6027\u80fd\u57fa\u51c6\u3002\u4f20\u7edf\u65b9\u6cd5\uff08\u5c11\u6837\u672c GPT\u3001\u5b8c\u6574\u5fae\u8c03\uff09\u8ba1\u7b97\u6602\u8d35\uff1b\u6307\u4ee4\u6316\u6398\u901a\u8fc7\u7ebf\u6027\u8d28\u91cf\u89c4\u5219\u7ed3\u5408\u54cd\u5e94\u957f\u5ea6\u3001\u5e73\u5747\u5956\u52b1\u6a21\u578b\u5206\u7b49\u6307\u6807\uff0c\u5feb\u901f\u8bc4\u4f30\u6570\u636e\u96c6\u6574\u4f53\u8d28\u91cf\uff0c\u6210\u4e3a\u66f4\u9ad8\u6548\u7684\u53ef\u884c\u65b9\u6848\u3002<\/p>\n<h5>SFT\u7684\u8fc7\u7a0b<\/h5>\n<p>\u5982\u56fe 4 \u6240\u793a\uff0c\u5fae\u8c03\u6d41\u7a0b\u59cb\u4e8e\u4e00\u4e2a\u5df2\u5728\u5927\u89c4\u6a21\u65e0\u6807\u6ce8\u8bed\u6599\u4e0a\u5b8c\u6210\u81ea\u76d1\u7763\u9884\u8bad\u7ec3\u7684 LLM\uff0c\u8be5\u9636\u6bb5\u8d4b\u4e88\u6a21\u578b\u901a\u7528\u8bed\u4e49\u8868\u5f81\u80fd\u529b\uff1b\u968f\u540e\uff0c\u501f\u52a9\u4efb\u52a1\u76f8\u5173\u7684\u6807\u6ce8\u6570\u636e\u5bf9\u53c2\u6570\u8fdb\u884c\u5fae\u8c03\uff0c\u4f7f\u6a21\u578b\u884c\u4e3a\u4e0e\u4e0b\u6e38\u9700\u6c42\u5bf9\u9f50\u3002\u5e38\u7528\u7684\u76ee\u6807\u51fd\u6570\u4e3a\u4ea4\u53c9\u71b5\u635f\u5931\u3002\u5bf9\u4e8e\u542b <span class=\"katex-eq\" data-katex-display=\"false\">N<\/span> \u6761\u6837\u672c\u3001<span class=\"katex-eq\" data-katex-display=\"false\">C<\/span> \u4e2a\u7c7b\u522b\u7684\u5206\u7c7b\u4efb\u52a1\uff0c\u5176\u5f62\u5f0f\u4e3a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">&gt;[katex]\n\n\\mathcal{L}{\\text{fine-tune}}(\\theta)=-\\frac{1}{N}\\sum{i=1}^{N}\\sum_{j=1}^{C}y_{ij}\\log P(y_j\\mid x_i;\\theta),\n\n<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">y_{ij}\\in{0,1}<\/span> \u662f\u6837\u672c <span class=\"katex-eq\" data-katex-display=\"false\">i<\/span> \u5728\u7c7b\u522b <span class=\"katex-eq\" data-katex-display=\"false\">j<\/span> \u4e0a\u7684\u771f\u5b9e\u6807\u7b7e\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">P(y_j\\mid x_i;\\theta)<\/span> \u4e3a\u6a21\u578b\u9884\u6d4b\u7684\u6982\u7387\u3002\u6700\u5c0f\u5316\u8be5\u635f\u5931\u53ef\u4ee4\u6a21\u578b\u8f93\u51fa\u903c\u8fd1\u771f\u5b9e\u5206\u5e03\uff0c\u4ece\u800c\u63d0\u5347\u76ee\u6807\u4efb\u52a1\u6027\u80fd\u3002<br \/>\n\u5178\u578b\u8303\u4f8b\u4e3a BERT\uff1a\u5176\u5728 BooksCorpus \u4e0e Wikipedia \u7b49\u5927\u578b\u8bed\u6599\u4e0a\u5b8c\u6210\u9884\u8bad\u7ec3\u540e\uff0c\u4ec5\u9700\u5728 IMDB \u7b49\u60c5\u611f\u5206\u6790\u6570\u636e\u96c6\u4e0a\u5fae\u8c03\u6570\u8f6e\uff0c\u5373\u53ef\u5feb\u901f\u4e13\u7cbe\u4e8e\u60c5\u611f\u5206\u7c7b\u3001\u95ee\u7b54\u7b49\u5177\u4f53\u4efb\u52a1\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101094051236.png\" alt=\"\" \/><\/p>\n<p><center>\u76d1\u7763\u5fae\u8c03\u7684\u8fc7\u7a0b<\/center><\/p>\n<h5>\u5168\u53c2\u6570\u5fae\u8c03<\/h5>\n<p>\u5168\u53c2\u6570\u5fae\u8c03\u6307\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u5168\u90e8\u53c2\u6570\u8fdb\u884c\u66f4\u65b0\uff0c\u4e0eLoRA\u3001Prefix-tuning \u7b49\u4ec5\u8c03\u6574\u5b50\u96c6\u7684\u53c2\u6570\u9ad8\u6548\u65b9\u6cd5\u76f8\u5bf9\u3002\u8be5\u65b9\u5f0f\u5728\u533b\u7597\u3001\u6cd5\u5f8b\u7b49\u9ad8\u7cbe\u5ea6\u573a\u666f\u4e0b\u901a\u5e38\u66f4\u5177\u4f18\u52bf\uff0c\u4f46\u4f34\u968f\u5de8\u5927\u8ba1\u7b97\u5f00\u9500\uff1a\u4f8b\u5982\u5bf9 65 B \u53c2\u6570\u7684\u6a21\u578b\u5fae\u8c03\u9700\u8d85\u8fc7 100 GB GPU \u663e\u5b58\u3002\u4e3a\u7f13\u89e3\u8d44\u6e90\u538b\u529b\uff0c\u7814\u7a76\u8005\u63d0\u51fa LOMO \u7b49\u5185\u5b58\u4f18\u5316\u6280\u672f\uff0c\u901a\u8fc7\u964d\u4f4e\u68af\u5ea6\u4e0e\u4f18\u5316\u5668\u72b6\u6001\u5360\u7528\uff0c\u4f7f\u5927\u6a21\u578b\u5728\u53d7\u9650\u786c\u4ef6\u4e0a\u4e5f\u53ef\u8bad\u7ec3\u3002\u53c2\u6570\u66f4\u65b0\u9075\u5faa<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\theta_{t+1}&gt;[katex]\\theta_{t+1}=\\theta_t-\\eta,\\nabla_\\theta L(\\theta_t)<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\theta_t<\/span> \u4e3a\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">t<\/span> \u6b21\u8fed\u4ee3\u53c2\u6570\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\eta<\/span> \u4e3a\u5b66\u4e60\u7387\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\nabla_\\theta L(\\theta_t)<\/span> \u4e3a\u635f\u5931\u68af\u5ea6\u3002\u914d\u5408\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u3001\u6fc0\u6d3b\u68c0\u67e5\u70b9\u7b49\u6280\u672f\uff0c\u53ef\u8fdb\u4e00\u6b65\u538b\u7f29\u663e\u5b58\u9700\u6c42\u3002<br \/>\n\u5178\u578b\u6848\u4f8b\u662f\u4eceGPT-3\u5230InstructGPT<br \/>\nOpenAI \u4f7f\u7528\u4e13\u4e3a\u6307\u4ee4\u8ddf\u968f\u6784\u5efa\u7684\u6570\u636e\u96c6\u5bf9 GPT-3 \u5b9e\u65bd\u5168\u53c2\u6570\u5fae\u8c03\uff0c\u5f97\u5230 InstructGPT\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6307\u4ee4\u9075\u5faa\u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u6027\u80fd\u3002\u5c3d\u7ba1\u9700\u66f4\u65b0\u5168\u90e8 175 B \u53c2\u6570\uff0c\u8ba1\u7b97\u6210\u672c\u6781\u9ad8\uff0c\u4f46\u6700\u7ec8\u6548\u679c\u4f18\u4e8e\u4ec5\u8c03\u6574\u90e8\u5206\u53c2\u6570\u7684\u65b9\u6cd5\u3002<\/p>\n<h4>\u81ea\u9002\u5e94\u5fae\u8c03<\/h4>\n<p>\u81ea\u9002\u5e94\u5fae\u8c03\u4f1a\u4fee\u6539\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u884c\u4e3a\uff0c\u4ee5\u66f4\u597d\u5730\u6ee1\u8db3\u7528\u6237\u7279\u5b9a\u9700\u6c42\u5e76\u5904\u7406\u66f4\u5e7f\u6cdb\u7684\u4efb\u52a1\u3002\u8fd9\u79cd\u65b9\u6cd5\u5f15\u5165\u4e86\u989d\u5916\u7684\u63d0\u793a\u6765\u6307\u5bfc\u6a21\u578b\u7684\u8f93\u51fa\u751f\u6210\uff0c\u4e3a\u5b9a\u5236\u6a21\u578b\u7684\u54cd\u5e94\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7075\u6d3b\u7684\u6846\u67b6\u3002\u81ea\u9002\u5e94\u5fae\u8c03\u4e2d\u7684\u663e\u8457\u65b9\u6cd5\u5305\u62ec\u6307\u4ee4\u8c03\u4f18\uff08instruction tuning\uff09\u548c\u57fa\u4e8e\u63d0\u793a\u7684\u8c03\u4f18\uff08prompt-based tuning\uff09\uff0c\u5176\u4e2d\u4e24\u8005\u90fd\u901a\u8fc7\u5f15\u5165\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6307\u5bfc\u6765\u663e\u8457\u589e\u5f3a LLM \u7684\u9002\u5e94\u6027\u3002<\/p>\n<h5>\u6307\u4ee4\u8c03\u4f18Instruction Tuning<\/h5>\n<p>\u6307\u4ee4\u8c03\u4f18\u662f\u4e00\u79cd\u901a\u8fc7\u5728\u4e13\u95e8\u6784\u5efa\u7684\u6307\u4ee4\u6570\u636e\u96c6\u4e0a\u5fae\u8c03\u57fa\u7840 LLM \u7684\u6280\u672f\u3002\u8fd9\u79cd\u65b9\u6cd5\u5927\u5927\u63d0\u9ad8\u4e86\u6a21\u578b\u5728\u5404\u79cd\u4efb\u52a1\u548c\u9886\u57df\u4e2d\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u63d0\u9ad8\u4e86\u5176\u7075\u6d3b\u6027\u548c\u51c6\u786e\u6027\u3002\u5982\u56fe\u56fe\u6240\u793a\uff0c\u8be5\u8fc7\u7a0b\u9996\u5148\u5c06\u73b0\u6709\u7684 NLP \u6570\u636e\u96c6\uff08\u4f8b\u5982\uff0c\u7528\u4e8e\u6587\u672c\u5206\u7c7b\u3001\u7ffb\u8bd1\u548c\u6458\u8981\u7684\u6570\u636e\u96c6\uff09\u8f6c\u6362\u4e3a\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\uff0c\u5176\u4e2d\u5305\u62ec\u4efb\u52a1\u63cf\u8ff0\u3001\u8f93\u5165\u793a\u4f8b\u3001\u9884\u671f\u8f93\u51fa\u548c\u8bf4\u660e\u6027\u6f14\u793a\u3002Self-Instruct\u7b49\u6280\u672f\u901a\u8fc7\u81ea\u52a8\u751f\u6210\u989d\u5916\u7684\u6307\u4ee4-\u8f93\u51fa\u5bf9\u6765\u8fdb\u4e00\u6b65\u589e\u5f3a\u8fd9\u4e9b\u6570\u636e\u96c6\u7684\u591a\u6837\u6027\uff0c\u4ece\u800c\u6269\u5927\u6a21\u578b\u5bf9\u66f4\u5e7f\u6cdb\u4efb\u52a1\u7684\u63a5\u89e6\u3002\u5fae\u8c03\u7a0b\u5e8f\u8c03\u6574\u6a21\u578b\u7684\u53c2\u6570\u4ee5\u4e0e\u8fd9\u4e9b\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6307\u4ee4\u4fdd\u6301\u4e00\u81f4\uff0c\u4ece\u800c\u4ea7\u751f\u4e00\u4e2a\u5728\u719f\u6089\u548c\u4ee5\u524d\u672a\u89c1\u8fc7\u7684\u4efb\u52a1\u4e2d\u90fd\u8868\u73b0\u7a33\u5065\u7684 LLM\u3002\u4f8b\u5982\uff0cInstructGPT\u548cGPT-4\u5728\u5404\u79cd\u5e94\u7528\u4e2d\u90fd\u663e\u793a\u51fa\u5728\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u65b9\u9762\u7684\u663e\u7740\u6539\u8fdb\u3002<br \/>\n\u6307\u4ee4\u8c03\u4f18\u7684\u6709\u6548\u6027\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\u6307\u4ee4\u6570\u636e\u96c6\u7684\u8d28\u91cf\u548c\u5e7f\u5ea6\u3002\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u96c6\u5e94\u5305\u542b\u5e7f\u6cdb\u7684\u8bed\u8a00\u3001\u9886\u57df\u548c\u4efb\u52a1\u590d\u6742\u6027\uff0c\u4ee5\u786e\u4fdd\u6a21\u578b\u4fdd\u6301\u5e7f\u6cdb\u9002\u7528\u6027\u3002\u6b64\u5916\uff0c\u6307\u4ee4\u7684\u6e05\u6670\u5ea6\u548c\u7ec4\u7ec7\u5728\u4f7f\u6a21\u578b\u80fd\u591f\u6709\u6548\u5730\u89e3\u91ca\u548c\u6267\u884c\u4efb\u52a1\u65b9\u9762\u8d77\u7740\u5173\u952e\u4f5c\u7528\u3002\u96c6\u6210\u6f14\u793a\u793a\u4f8b\uff08\u5305\u62ecChain-of-Thought\u63d0\u793a\uff09\u7b49\u6280\u672f\u53ef\u4ee5\u663e\u7740\u63d0\u9ad8\u5728\u9700\u8981\u590d\u6742\u63a8\u7406\u7684\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u3002\u6b64\u5916\uff0c\u786e\u4fdd\u5728\u5fae\u8c03\u9636\u6bb5\u4efb\u52a1\u7684\u5747\u8861\u5206\u5e03\u5bf9\u4e8e\u907f\u514d\u8fc7\u62df\u5408\u6216\u7531\u4e8e\u4efb\u52a1\u8986\u76d6\u4e0d\u5e73\u8861\u800c\u964d\u4f4e\u6a21\u578b\u6027\u80fd\u81f3\u5173\u91cd\u8981\u3002\u6bd4\u4f8b\u4efb\u52a1\u62bd\u6837\u6216\u52a0\u6743\u635f\u5931\u51fd\u6570\u7b49\u6280\u672f\u6709\u52a9\u4e8e\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u786e\u4fdd\u6bcf\u4e2a\u4efb\u52a1\u5bf9\u5fae\u8c03\u8fc7\u7a0b\u505a\u51fa\u516c\u5e73\u7684\u8d21\u732e\u3002\u56e0\u6b64\uff0c\u901a\u8fc7\u7cbe\u5fc3\u6784\u5efa\u548c\u7ba1\u7406\u6307\u4ee4\u6570\u636e\u96c6\uff0c\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u5927\u5927\u589e\u5f3a\u5fae\u8c03\u540eLLM\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u4f7f\u5b83\u4eec\u5728\u5e7f\u6cdb\u7684\u4efb\u52a1\u548c\u9886\u57df\u4e2d\u8868\u73b0\u51fa\u8272\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101095645498.png\" alt=\"\" \/><\/p>\n<p><center>\u6307\u4ee4\u5fae\u8c03\u7684\u5de5\u4f5c\u6d41\u7a0b\uff0c\u8bf4\u660e\u4e86\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e2d\u6307\u4ee4\u6570\u636e\u96c6\u6784\u5efa\u548c\u6307\u4ee4\u8c03\u4f18\u7684\u901a\u7528\u6d41\u7a0b<\/center><\/p>\n<h5>\u524d\u7f00\u8c03\u6574<\/h5>\n<p>Prefix-tuning\u662f\u4e00\u79cd\u53c2\u6570\u9ad8\u6548\u7684\u5fae\u8c03\u65b9\u6cd5\uff0c\u6d89\u53ca\u5728\u8bed\u8a00\u6a21\u578b\u7684\u6bcf\u4e2aTransformer\u5c42\u4e2d\u6dfb\u52a0\u4e00\u7cfb\u5217\u53ef\u8bad\u7ec3\u7684\u524d\u7f00token\uff08\u8fde\u7eed\u5411\u91cf\uff09\uff0c\u540c\u65f6\u4fdd\u6301\u6838\u5fc3\u6a21\u578b\u53c2\u6570\u56fa\u5b9a\u3002\u5982\u56fe (a)\u6240\u793a\uff0c\u8fd9\u4e9b\u524d\u7f00\u5411\u91cf\u662f\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\uff0c\u5e76\u5145\u5f53\u865a\u62dftoken\u5d4c\u5165\u3002\u4e3a\u4e86\u4f18\u5316\u524d\u7f00\u5411\u91cf\uff0c\u4f7f\u7528\u4e86\u4e00\u79cd\u91cd\u65b0\u53c2\u6570\u5316\u6280\u5de7\uff0c\u5176\u4e2d\u5b66\u4e60\u4e00\u4e2a\u5c0f\u7684\u591a\u5c42\u611f\u77e5\u5668\uff08MLP\uff09\u51fd\u6570\u5c06\u8f83\u5c0f\u7684\u77e9\u9635\u6620\u5c04\u5230\u524d\u7f00\u53c2\u6570\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u4f18\u5316\u524d\u7f00\u5411\u91cf\u3002\u8fd9\u79cd\u65b9\u6cd5\u5df2\u88ab\u8bc1\u660e\u53ef\u4ee5\u7a33\u5b9a\u8bad\u7ec3\u8fc7\u7a0b\u3002\u4e00\u65e6\u524d\u7f00\u5411\u91cf\u88ab\u4f18\u5316\uff0c\u6620\u5c04\u51fd\u6570\u5c06\u88ab\u4e22\u5f03\uff0c\u5e76\u4e14\u4ec5\u4fdd\u7559\u6d3e\u751f\u7684\u524d\u7f00\u5411\u91cf\u4ee5\u589e\u5f3a\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6027\u80fd\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101095820814.png\" alt=\"\" \/><\/p>\n<p><center>Prefix Tuning\u548cPrompt Tuning\u7684\u6bd4\u8f83\uff0c\u63cf\u7ed8\u4e86\u5b83\u4eec\u4e0d\u540c\u7684\u53c2\u6570\u5fae\u8c03\u65b9\u6cd5\uff1aa) Prefix Tuning\u548cb) Prompt Tuning<\/center>\u901a\u8fc7\u5c06\u5b66\u4e60\u7684\u8fde\u7eed\u63d0\u793a\u6dfb\u52a0\u5230\u8f93\u5165\u5e8f\u5217\u5e76\u4f7f\u7528\u9010\u5c42\u63d0\u793a\uff0c\u6a21\u578b\u7684\u884c\u4e3a\u88ab\u5f15\u5bfc\u671d\u5411\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u8f93\u51fa\uff0c\u800c\u65e0\u9700\u8fdb\u884c\u5168\u6a21\u578b\u5fae\u8c03\u3002\u7531\u4e8e\u4ec5\u8c03\u6574\u524d\u7f00\u53c2\u6570\uff0c\u8fd9\u5bfc\u81f4\u4e86\u4e00\u79cd\u66f4\u53c2\u6570\u9ad8\u6548\u7684\u65b9\u6cd5\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0cP-Tuning v2 \u5c06\u9010\u5c42\u63d0\u793a\u5411\u91cf\u5408\u5e76\u5230Transformer\u67b6\u6784\u4e2d\uff0c\u7279\u522b\u662f\u9488\u5bf9\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u4efb\u52a1\u3002\u8fd9\u79cd\u65b9\u6cd5\u8fd8\u5229\u7528\u591a\u4efb\u52a1\u5b66\u4e60\u6765\u4f18\u5316\u8de8\u4efb\u52a1\u7684\u5171\u4eab\u63d0\u793a\uff0c\u4ece\u800c\u63d0\u9ad8\u4e0d\u540c\u53c2\u6570\u89c4\u6a21\u4e0b\u7684\u6a21\u578b\u6027\u80fd\u3002Prefix-tuning\u5728\u4fc3\u8fdb\u5927\u578b\u8bed\u8a00\u6a21\u578b\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u7684\u5feb\u901f\u9ad8\u6548\u9002\u5e94\u65b9\u9762\u7684\u6f5c\u529b\u662f\u663e\u800c\u6613\u89c1\u7684\uff0c\u8fd9\u4f7f\u5176\u6210\u4e3a\u9700\u8981\u7075\u6d3b\u6027\u548c\u6548\u7387\u7684\u5e94\u7528\u7a0b\u5e8f\u7684\u5f15\u4eba\u6ce8\u76ee\u7684\u7b56\u7565\u3002<\/p>\n<h5>\u63d0\u793a\u8c03\u6574<\/h5>\n<p>Prompt-tuning\u662f\u4e00\u79cd\u901a\u8fc7\u4f18\u5316\u8f93\u5165\u5c42\u7684\u53ef\u8bad\u7ec3\u5411\u91cf\u800c\u4e0d\u662f\u4fee\u6539\u6a21\u578b\u7684\u5185\u90e8\u53c2\u6570\u6765\u9ad8\u6548\u9002\u914d\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u65b9\u6cd5\u3002\u5982\u4e0a\u56fe(b)\u6240\u793a\uff0c\u8fd9\u9879\u6280\u672f\u5efa\u7acb\u5728\u79bb\u6563\u63d0\u793a\u65b9\u6cd5\u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165\u4e86\u8f6f\u63d0\u793aToken\uff0c\u8fd9\u4e9b Token \u53ef\u4ee5\u4ee5\u4e0d\u53d7\u9650\u5236\u7684\u683c\u5f0f\u6216\u4f5c\u4e3a\u524d\u7f00\u7ed3\u6784\u5316\u3002\u8fd9\u4e9b\u5b66\u4e60\u5230\u7684\u63d0\u793a\u5d4c\u5165\u4e0e\u8f93\u5165\u6587\u672c\u5d4c\u5165\u76f8\u7ed3\u5408\uff0c\u7136\u540e\u7531\u6a21\u578b\u5904\u7406\uff0c\u4ece\u800c\u6307\u5bfc\u6a21\u578b\u7684\u8f93\u51fa\uff0c\u540c\u65f6\u4fdd\u6301\u9884\u8bad\u7ec3\u7684\u6743\u91cd\u51bb\u7ed3\u3002\u63d0\u793a\u8c03\u4f18\u7684\u4e24\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u5b9e\u73b0\u662f P-tuning \uff0c\u5b83\u4f7f\u7528\u4e00\u79cd\u7075\u6d3b\u7684\u65b9\u6cd5\u6765\u7ec4\u5408\u4e0a\u4e0b\u6587\u3001\u63d0\u793a\u548c\u76ee\u6807 Token\uff0c\u4f7f\u5176\u9002\u7528\u4e8e\u7406\u89e3\u548c\u751f\u6210\u4efb\u52a1\u3002 \u8fd9\u79cd\u65b9\u6cd5\u901a\u8fc7\u53cc\u5411 LSTM \u67b6\u6784\u589e\u5f3a\u4e86\u8f6f\u63d0\u793a\u8868\u793a\u7684\u5b66\u4e60\u3002 \u76f8\u6bd4\u4e4b\u4e0b\uff0c\u6807\u51c6\u63d0\u793a\u8c03\u4f18\u91c7\u7528\u66f4\u7b80\u5355\u7684\u8bbe\u8ba1\uff0c\u5176\u4e2d\u524d\u7f00\u63d0\u793a\u88ab\u6dfb\u52a0\u5230\u8f93\u5165\u4e4b\u524d\uff0c\u5e76\u4e14\u4ec5\u5728\u8bad\u7ec3\u671f\u95f4\u6839\u636e\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u76d1\u7763\u6765\u66f4\u65b0\u63d0\u793a\u5d4c\u5165\u3002<br \/>\n\u7814\u7a76\u8868\u660e\uff0c\u63d0\u793a\u8c03\u4f18\u53ef\u4ee5\u5339\u914d\u8bb8\u591a\u4efb\u52a1\u4e2d\u5168\u53c2\u6570\u5fae\u8c03\u7684\u6027\u80fd\uff0c\u540c\u65f6\u9700\u8981\u66f4\u5c11\u7684\u53ef\u8bad\u7ec3\u53c2\u6570\u3002\u7136\u800c\uff0c\u5b83\u7684\u6210\u529f\u4e0e\u5e95\u5c42\u8bed\u8a00\u6a21\u578b\u7684\u80fd\u529b\u5bc6\u5207\u76f8\u5173\uff0c\u56e0\u4e3a\u63d0\u793a\u8c03\u4f18\u4ec5\u4fee\u6539\u8f93\u5165\u5c42\u7684\u4e00\u5c0f\u90e8\u5206\u53c2\u6570\u3002 \u5728\u8fd9\u4e9b\u8fdb\u6b65\u7684\u57fa\u7840\u4e0a\uff0c\u8bf8\u5982 P-Tuning v2 \u7b49\u8f83\u65b0\u7684\u65b9\u6cd5\u5df2\u7ecf\u8bc1\u660e\uff0c\u63d0\u793a\u8c03\u4f18\u7b56\u7565\u53ef\u4ee5\u6709\u6548\u5730\u6269\u5c55\u5230\u5404\u79cd\u6a21\u578b\u5927\u5c0f\uff0c\u5904\u7406\u5148\u524d\u88ab\u8ba4\u4e3a\u9700\u8981\u5b8c\u5168\u5fae\u8c03\u7684\u590d\u6742\u4efb\u52a1\u3002 \u8fd9\u4e9b\u53d1\u73b0\u5c06\u63d0\u793a\u8c03\u4f18\u786e\u7acb\u4e3a\u4f20\u7edf\u5fae\u8c03\u7684\u9ad8\u6548\u66ff\u4ee3\u65b9\u6848\uff0c\u63d0\u4f9b\u4e86\u53ef\u6bd4\u7684\u6027\u80fd\uff0c\u540c\u65f6\u964d\u4f4e\u4e86\u8ba1\u7b97\u548c\u5185\u5b58\u6210\u672c\u3002<\/p>\n<h5>\u5f3a\u5316\u5fae\u8c03 Reinforcement Fine-Tuning<\/h5>\n<p>\u5f3a\u5316\u5fae\u8c03 (ReFT) \u4ee3\u8868\u4e00\u79cd\u5148\u8fdb\u6280\u672f\uff0c\u5b83\u5c06 RL \u4e0e SFT \u96c6\u6210\uff0c\u4ee5\u589e\u5f3a\u6a21\u578b\u89e3\u51b3\u590d\u6742\u3001\u52a8\u6001\u95ee\u9898\u7684\u80fd\u529b\u3002\u4e0e\u4f20\u7edf\u7684 SFT \u4e0d\u540c\uff0c\u4f20\u7edf\u7684 SFT \u901a\u5e38\u5bf9\u6bcf\u4e2a\u95ee\u9898\u4f7f\u7528\u5355\u4e2a CoT \u6807\u6ce8\uff0cReFT \u4f7f\u6a21\u578b\u80fd\u591f\u63a2\u7d22\u591a\u4e2a\u6709\u6548\u7684\u63a8\u7406\u8def\u5f84\uff0c\u4ece\u800c\u63d0\u9ad8\u5176\u6cdb\u5316\u80fd\u529b\u548c\u89e3\u51b3\u95ee\u9898\u7684\u6280\u80fd\u3002 ReFT \u8fc7\u7a0b\u4ece\u6807\u51c6\u7684 SFT \u9636\u6bb5\u5f00\u59cb\uff0c\u5176\u4e2d\u6a21\u578b\u6700\u521d\u5728\u6807\u8bb0\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u4ee5\u901a\u8fc7\u76d1\u7763\u6807\u6ce8\u5b66\u4e60\u57fa\u672c\u7684\u89e3\u51b3\u4efb\u52a1\u7684\u80fd\u529b\u3002\u5728\u6b64\u521d\u59cb\u5fae\u8c03\u4e4b\u540e\uff0c\u6a21\u578b\u4f7f\u7528 RL \u7b97\u6cd5\uff08\u4f8b\u5982\u8fd1\u7aef\u7b56\u7565\u4f18\u5316(PPO)\uff09\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u6539\u8fdb\u3002 \u5728\u5f3a\u5316\u9636\u6bb5\uff0c\u6a21\u578b\u4e3a\u6bcf\u4e2a\u95ee\u9898\u751f\u6210\u591a\u4e2a CoT \u6807\u6ce8\uff0c\u63a2\u7d22\u4e0d\u540c\u7684\u6f5c\u5728\u63a8\u7406\u8def\u5f84\u3002\u901a\u8fc7\u5c06\u6a21\u578b\u7684\u9884\u6d4b\u7b54\u6848\u4e0e\u771f\u5b9e\u7b54\u6848\u8fdb\u884c\u6bd4\u8f83\u6765\u8bc4\u4f30\u8fd9\u4e9b\u751f\u6210\u7684\u8def\u5f84\uff0c\u5bf9\u6b63\u786e\u8f93\u51fa\u5206\u914d\u5956\u52b1\uff0c\u5bf9\u9519\u8bef\u8f93\u51fa\u8fdb\u884c\u60e9\u7f5a\u3002 \u8fd9\u4e2a\u8fed\u4ee3\u8fc7\u7a0b\u9a71\u4f7f\u6a21\u578b\u8c03\u6574\u5176\u7b56\u7565\uff0c\u6700\u7ec8\u6539\u8fdb\u5176\u63a8\u7406\u7b56\u7565\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101100224122.png\" alt=\"\" \/><\/p>\n<p><center>\u5f3a\u5316\u5fae\u8c03\uff08ReFT\uff09\u7684\u8fc7\u7a0b\uff0c\u63cf\u7ed8\u4e86\u8fed\u4ee3\u7684\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u9884\u70ed\uff0c\u7136\u540e\u5bf9\u76f8\u540c\u7684\u6570\u636e\u96c6\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3<\/center>ReFT \u8fc7\u7a0b\u5206\u4e24\u4e2a\u9636\u6bb5\u6267\u884c\u3002\u4e0a\u534a\u90e8\u5206\u4ee3\u8868 SFT \u9636\u6bb5\uff0c\u6a21\u578b\u8fed\u4ee3\u8bad\u7ec3\u6570\u636e\uff0c\u5728\u51e0\u4e2a epoch \u4e2d\u5b66\u4e60\u6bcf\u4e2a\u95ee\u9898\u7684\u6b63\u786e CoT \u6807\u6ce8\u3002\u5728\u4e0b\u534a\u90e8\u5206\uff0c\u5f15\u5165\u4e86 ReFT \u9636\u6bb5\uff1a\u4ece SFT \u8bad\u7ec3\u7684\u6a21\u578b\u5f00\u59cb\uff0c\u6a21\u578b\u6839\u636e\u5176\u5f53\u524d\u7b56\u7565\u751f\u6210\u5907\u9009\u7684 CoT \u6807\u6ce8 <span class=\"katex-eq\" data-katex-display=\"false\">e&#039;<\/span>\uff0c\u5e76\u5c06\u5176\u9884\u6d4b\u7684\u7b54\u6848 <span class=\"katex-eq\" data-katex-display=\"false\">y&#039;<\/span> \u4e0e\u771f\u5b9e\u7b54\u6848 <span class=\"katex-eq\" data-katex-display=\"false\">y<\/span> \u8fdb\u884c\u6bd4\u8f83\u3002\u6b63\u786e\u7b54\u6848\u7ed9\u4e88\u6b63\u5411\u5956\u52b1\uff0c\u9519\u8bef\u7b54\u6848\u7ed9\u4e88\u8d1f\u5411\u5956\u52b1\uff0c\u9a71\u4f7f\u6a21\u578b\u63d0\u9ad8\u5176\u6027\u80fd\u3002\u7136\u540e\uff0c\u8fd9\u4e9b\u5956\u52b1\u4fe1\u53f7\u7528\u4e8e\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u66f4\u65b0\u6a21\u578b\u7684\u7b56\u7565\uff0c\u589e\u5f3a\u5176\u751f\u6210\u51c6\u786e\u548c\u591a\u6837\u5316 CoT \u6807\u6ce8\u7684\u80fd\u529b\u3002<br \/>\n\u6700\u8fd1\u7684\u7814\u7a76\u8868\u660e\uff0cReFT \u660e\u663e\u4f18\u4e8e\u4f20\u7edf\u7684 SFT \u65b9\u6cd5\u3002\u6b64\u5916\uff0c\u63a8\u7406\u65f6\u7b56\u7565\u7684\u6574\u5408\uff0c\u5982\u591a\u6570\u6295\u7968\u548c\u91cd\u65b0\u6392\u5e8f\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6027\u80fd\uff0c\u5141\u8bb8\u6a21\u578b\u5728\u8bad\u7ec3\u540e\u5b8c\u5584\u5176\u8f93\u51fa\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cReFT \u5728\u4e0d\u9700\u8981\u989d\u5916\u6216\u589e\u5f3a\u7684\u8bad\u7ec3\u6570\u636e\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u4e86\u8fd9\u4e9b\u6539\u8fdb\uff0c\u4ec5\u4ece SFT \u9636\u6bb5\u4f7f\u7528\u7684\u73b0\u6709\u6570\u636e\u96c6\u4e2d\u5b66\u4e60\u3002\u8fd9\u7a81\u51fa\u4e86\u8be5\u6a21\u578b\u5353\u8d8a\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u56e0\u4e3a\u5b83\u4ece\u53ef\u7528\u6570\u636e\u4e2d\u66f4\u9ad8\u6548\u3001\u66f4\u6709\u6548\u5730\u5b66\u4e60\u3002<\/p>\n<h3>\u7528\u4e8e\u5bf9\u9f50\u7684\u540e\u8bad\u7ec3\u6a21\u578b<\/h3>\n<p>LLM \u4e2d\u7684\u5bf9\u9f50\u6d89\u53ca\u5f15\u5bfc\u6a21\u578b\u8f93\u51fa\u4ee5\u7b26\u5408\u4eba\u7c7b\u7684\u671f\u671b\u548c\u504f\u597d\uff0c\u7279\u522b\u662f\u5728\u5b89\u5168\u5173\u952e\u6216\u9762\u5411\u7528\u6237\u7684\u5e94\u7528\u7a0b\u5e8f\u4e2d\u3002\u672c\u7ae0\u8ba8\u8bba\u4e86\u5b9e\u73b0\u5bf9\u9f50\u7684\u4e09\u4e2a\u4e3b\u8981\u8303\u5f0f\uff1a\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning with Human Feedback\uff09\uff0c\u5b83\u4f7f\u7528\u4eba\u7c7b\u6807\u8bb0\u7684\u6570\u636e\u4f5c\u4e3a\u5956\u52b1\u4fe1\u53f7\uff1b\u57fa\u4e8e AI \u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning with AI Feedback\uff09\uff0c\u5b83\u5229\u7528 AI \u751f\u6210\u7684\u53cd\u9988\u6765\u89e3\u51b3\u53ef\u6269\u5c55\u6027\u95ee\u9898\uff1b\u4ee5\u53ca\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08Direct Preference Optimization\uff09\uff0c\u5b83\u76f4\u63a5\u4ece\u6210\u5bf9\u7684\u4eba\u7c7b\u504f\u597d\u6570\u636e\u4e2d\u5b66\u4e60\uff0c\u800c\u65e0\u9700\u663e\u5f0f\u7684\u5956\u52b1\u6a21\u578b\u3002 \u6bcf\u4e2a\u8303\u5f0f\u5728\u5176\u8ffd\u6c42\u7a33\u5065\u5bf9\u9f50\u7684\u8fc7\u7a0b\u4e2d\u90fd\u63d0\u4f9b\u4e86\u72ec\u7279\u7684\u4f18\u52bf\u3001\u6311\u6218\u548c\u6743\u8861\u3002 \u5bf9\u8fd9\u4e9b\u65b9\u6cd5\u53ca\u76f8\u5173\u65b9\u6cd5\u7684\u7b80\u8981\u6bd4\u8f83\u603b\u7ed3\u5728\u4e0b\u8868\u4e2d\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101101349438.jpg\" alt=\"\" \/><\/p>\n<p><center>\u5927\u578b\u8bed\u8a00\u6a21\u578b (2022\u20132024) \u5bf9\u9f50\u65b9\u6cd5\u7684\u6bd4\u8f83\u6982\u8ff0\u3002 \u672c\u8868\u8bc4\u4f30\u4e86\u516b\u4e2a\u6307\u6807\u7684\u4e3b\u8981\u5bf9\u9f50\u6280\u672f\uff1aRM1\uff08\u663e\u5f0f\u6216\u9690\u5f0f\u5956\u52b1\u6a21\u578bExplicit or Implicit Reward Model\uff09\uff0cRM2\uff08\u70b9\u5956\u52b1\u6216\u504f\u597d\u6982\u7387\u6a21\u578bPoint Reward or Preference Probability Model\uff09\uff0cRM3\uff08\u54cd\u5e94\u7ea7\u6216\u8bcd\u5143\u7ea7\u5956\u52b1Response- or Token-level Reward\uff09\uff0cRM4\uff08\u6b63\u5411\u6216\u8d1f\u5411\u5956\u52b1\u6a21\u578bPositive or Negative Reward Model\uff09\uff0cF\uff08\u53cd\u9988\u7c7b\u578b\uff1a\u4eba\u7c7b\u6216 AI\uff09\uff0cRL1\uff08\u53c2\u8003\u6a21\u578b\u6216\u65e0\u53c2\u8003\u6a21\u578bReference Model or Reference Model-Free RL\uff09\uff0cRL2\uff08\u540c\u7b56\u7565\u6216\u5f02\u7b56\u7565 On-policy or Off-policy RL\uff09\uff0c\u4ee5\u53ca O\uff08\u5728\u7ebf\/\u8fed\u4ee3\u6216\u79bb\u7ebf\/\u975e\u8fed\u4ee3\u4f18\u5316 Online\/Iterative or Offline\/Non-iterative Optimization\uff09<\/center><\/p>\n<h4>\u4f7f\u7528\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60<\/h4>\n<p>\u76d1\u7763\u5fae\u8c03 (SFT) \u5df2\u88ab\u7528\u4f5c\u5f15\u5bfc LLM \u9075\u5faa\u4eba\u7c7b\u6307\u4ee4\u7684\u57fa\u7840\u6280\u672f\u3002 \u7136\u800c\uff0c\u5728\u7eaf\u76d1\u7763\u60c5\u666f\u4e2d\uff0c\u6807\u6ce8\u6570\u636e\u7684\u591a\u6837\u6027\u548c\u8d28\u91cf\u53ef\u80fd\u53c2\u5dee\u4e0d\u9f50\uff0c\u5e76\u4e14\u76d1\u7763\u6a21\u578b\u6355\u83b7\u66f4\u7ec6\u5fae\u6216\u81ea\u9002\u5e94\u7684\u4eba\u7c7b\u504f\u597d\u7684\u80fd\u529b\u5f80\u5f80\u6709\u9650\u3002\u4f5c\u4e3a\u89e3\u51b3\u65b9\u6848\uff0c\u6709\u4eba\u63d0\u51fa\u4e86\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60 (RL) \u7684\u5fae\u8c03\u6765\u89e3\u51b3\u8fd9\u4e9b\u7f3a\u70b9\u3002\u5728 RL \u65b9\u6cd5\u4e2d\uff0c\u6765\u81ea\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60 (RLHF) \u8131\u9896\u800c\u51fa\uff0c\u6210\u4e3a\u6700\u65e9\u548c\u6700\u5177\u5f71\u54cd\u529b\u7684\u57fa\u4e8e RL \u7684\u5bf9\u9f50\u540e\u8bad\u7ec3\u65b9\u6cd5\u4e4b\u4e00\u3002<br \/>\n\u5982\u56fe\u6240\u793a\uff0cRLHF \u9996\u5148\u4ee5\u504f\u597d\u6807\u7b7e\u6216\u5956\u52b1\u4fe1\u53f7\u7684\u5f62\u5f0f\u805a\u5408\u4eba\u7c7b\u53cd\u9988\uff0c\u7136\u540e\u4f7f\u7528\u8be5\u4fe1\u606f\u6765\u8bad\u7ec3\u5956\u52b1\u6a21\u578b\u3002\u5728\u6b64\u5956\u52b1\u6a21\u578b\u7684\u5f15\u5bfc\u4e0b\uff0c\u7b56\u7565\u88ab\u8fed\u4ee3\u8c03\u6574\u4ee5\u66f4\u597d\u5730\u5339\u914d\u4eba\u7c7b\u504f\u597d\u3002\u4e0e SFT \u76f8\u6bd4\uff0cRLHF \u5305\u542b\u6301\u7eed\u7684\u3001\u504f\u597d\u9a71\u52a8\u7684\u66f4\u65b0\uff0c\u4ece\u800c\u5e26\u6765\u66f4\u5f3a\u7684\u5bf9\u9f50\u7ed3\u679c\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u73b0\u4ee3 LLM\uff0c\u5982 GPT-4 \u3001Claude \u548c Gemini \u5df2\u7ecf\u53d7\u76ca\u4e8e\u8fd9\u4e9b\u673a\u5236\uff0c\u5c55\u793a\u4e86\u5728\u6307\u4ee4\u9075\u5faa\u3001\u4e8b\u5b9e\u4e00\u81f4\u6027\u548c\u7528\u6237\u76f8\u5173\u6027\u65b9\u9762\u7684\u6539\u8fdb\u3002\u4e0b\u9762\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba RLHF \u7684\u4e3b\u8981\u7ec4\u6210\u90e8\u5206\uff0c\u5305\u62ec\u53cd\u9988\u673a\u5236\u3001\u5956\u52b1\u5efa\u6a21\u548c\u7b56\u7565\u5b66\u4e60\u7b56\u7565\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101102247613.png\" alt=\"\" \/><\/p>\n<p><center>\u6765\u81ea\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60 (RLHF) \u7684\u5de5\u4f5c\u6d41\u7a0b\uff0c\u63cf\u7ed8\u4e86\u5c06\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u7684\u6574\u4f53\u8bad\u7ec3\u8fc7\u7a0b<\/center><\/p>\n<h5>RLHF\u7684\u53cd\u9988\u673a\u5236<\/h5>\n<p>\u4eba\u7c7b\u53cd\u9988\u662fRLHF\u7684\u6838\u5fc3\uff0c\u5b83\u544a\u77e5\u5956\u52b1\u6a21\u578b\u7528\u6237\u504f\u597d\u5e76\u6307\u5bfc\u7b56\u7565\u66f4\u65b0\u3002\u672c\u5c0f\u8282\u91c7\u7528\u5206\u7c7b\u6cd5\uff0c\u5bf9\u5e38\u89c1\u7684\u4eba\u7c7b\u53cd\u9988\u5f62\u5f0f\u8fdb\u884c\u5206\u7c7b\u3002 \u88683\u5c55\u793a\u4e86\u8fd9\u4e9b\u53cd\u9988\u7c7b\u578b\uff0c\u6db5\u76d6\u4e86\u8bf8\u5982\u7c92\u5ea6\u3001\u53c2\u4e0e\u7a0b\u5ea6\u548c\u660e\u786e\u6027\u7b49\u7ef4\u5ea6\u3002\u6bcf\u79cd\u53cd\u9988\u6a21\u5f0f\u90fd\u5bf9\u6a21\u578b\u4f18\u5316\u7684\u4e0d\u540c\u65b9\u9762\u6709\u6240\u8d21\u732e\uff0c\u63d0\u4f9b\u4e86\u4e0d\u540c\u7a0b\u5ea6\u7684\u53ef\u89e3\u91ca\u6027\u3001\u53ef\u6269\u5c55\u6027\u548c\u566a\u58f0\u5bb9\u5fcd\u5ea6\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101102532267.jpg\" alt=\"\" \/><\/p>\n<p><center>\u7528\u4e8e\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u7684\u53cd\u9988\u7c7b\u578b\u5206\u7c7b\u3002\u672c\u8868\u6982\u8ff0\u4e86\u5e38\u89c1\u7684\u53cd\u9988\u7c7b\u522b\u53ca\u5176\u5728\u516d\u4e2a\u6307\u6807\u4e0a\u7684\u5b9a\u4e49\u5c5e\u6027\uff1a\u7c92\u5ea6\uff08\u8303\u56f4\uff1aepisode\u3001segment\u6216step\uff09\uff0c\u53c2\u4e0e\u5ea6\uff08\u53c2\u4e0e\u65b9\u5f0f\uff1a\u89c2\u5bdf\u3001\u4e3b\u52a8\u6216\u534f\u540c\u751f\u6210\uff09\uff0c\u5143\u6570\uff08\u5b9e\u4f8b\u8ba1\u6570\uff1a\u5355\u4e2a\u3001\u591a\u4e2a\u6216\u4e09\u5143\uff09\uff0c\u62bd\u8c61\uff08\u76ee\u6807\uff1a\u7279\u5f81\u6216\u5b9e\u4f8b\uff09\uff0c\u610f\u56fe\uff08\u76ee\u7684\uff1a\u8bc4\u4f30\u6027\u3001\u63cf\u8ff0\u6027\u6216\u5b57\u9762\u610f\u601d\uff09\uff0c\u4ee5\u53ca\u660e\u786e\u6027\uff08\u76f4\u63a5\u6027\uff1a\u663e\u5f0f\u6216\u9690\u5f0f\uff09<\/center><\/p>\n<ul>\n<li><strong>\u4e3b\u8981\u53cd\u9988(Primary Feedback)<\/strong>:\u6b64\u7c7b\u522b\u5305\u542b\u6700\u76f4\u63a5\u5f71\u54cdRLHF\u4e2d\u5956\u52b1\u6a21\u578b\u5851\u9020\u7684\u53cd\u9988\u7c7b\u578b\u3002\u4f8b\u5982\uff0cCritique \u4fa7\u91cd\u4e8e\u5bf9agent\u884c\u4e3a\u7684\u663e\u5f0f\u4eba\u7c7b\u8bc4\u4f30\uff0c\u901a\u5e38\u901a\u8fc7\u4e8c\u5143\u6216\u591a\u6807\u7b7e\u6807\u6ce8\u8fdb\u884c\u7ec6\u5316\u4ee5\u51cf\u8f7b\u566a\u58f0\u3002 Comparisons\u5141\u8bb8\u8bc4\u4f30\u8005\u6bd4\u8f83\u591a\u4e2a\u8f93\u51fa\u6216\u8f68\u8ff9\uff1b\u867d\u7136\u66f4\u5927\u7684\u9009\u62e9\u96c6\u53ef\u4ee5\u63d0\u4f9b\u66f4\u4e30\u5bcc\u7684\u4fe1\u53f7\uff0c\u4f46\u4e5f\u53ef\u80fd\u5bfc\u81f4\u56e0\u679c\u6df7\u6dc6\u3002\u8de8\u65f6\u95f4\u53cd\u9988\uff08Inter-Temporal Feedback\uff09\u901a\u8fc7\u5728\u4e0d\u540c\u7684\u65f6\u95f4\u6b65\u63d0\u4f9b\u5224\u65ad\u6765\u7ec6\u5316\u8f68\u8ff9\u8bc4\u4f30\uff0c\u800c\u4ee3\u7406\u5956\u52b1\uff08Proxy Rewards\uff09\u5219\u7ed3\u5408\u4e86\u8fd1\u4f3c\u5956\u52b1\u51fd\u6570\uff0c\u5c06\u6a21\u578b\u5bfc\u5411\u7528\u6237\u5b9a\u4e49\u7684\u76ee\u6807\u3002\u793e\u4f1a\u884c\u4e3a\uff08Social Behavior\uff09\u5229\u7528\u9690\u5f0f\u7ebf\u7d22\uff08\u4f8b\u5982\uff0c\u9762\u90e8\u8868\u60c5\uff09\u6765\u4f7fagent\u76ee\u6807\u4e0e\u7528\u6237\u60c5\u7eea\u4fdd\u6301\u4e00\u81f4\u3002\u6539\u8fdb\u5f3a\u8c03\u5b9e\u65f6\u7684\u4eba\u5de5\u5e72\u9884\uff0c\u4ee5\u8fdb\u884c\u589e\u91cf\u7b56\u7565\u7ec6\u5316\u3002\u6700\u540e\uff0c\u81ea\u7136\u8bed\u8a00\u53cd\u9988\uff08Natural Language Feedback\uff09\u5229\u7528\u6587\u672c\u4fe1\u606f\u6765\u4f20\u8fbe\u504f\u597d\u548c\u6539\u8fdb\u5efa\u8bae\u3002<\/li>\n<li><strong>\u8865\u5145\u53cd\u9988\uff08Supplementary Feedback\uff09<\/strong>\u3002\u9664\u4e86\u4e3b\u8981\u53cd\u9988\u4e4b\u5916\uff0c\u8fd8\u6709\u4e24\u7c7b\u8fdb\u4e00\u6b65\u52a0\u5f3a\u4e86\u5956\u52b1\u5efa\u6a21\u8fc7\u7a0b\u3002 \u7d27\u6025\u505c\u6b62 (Emergency stops\uff0ce-stops) \u5141\u8bb8\u4eba\u7c7b\u901a\u8fc7\u505c\u6b62 agent \u7684\u8f68\u8ff9\u6765\u5e72\u9884\u5176\u884c\u4e3a\uff0c\u800c\u65e0\u9700\u5efa\u8bae\u66ff\u4ee3\u65b9\u6848\u3002 \u8fd9\u79cd\u53cd\u9988\u7684\u7279\u5f81\u662f\u9690\u5f0f\u53c2\u4e0e\u548c\u5bf9\u9632\u6b62\u4e0d\u826f\u884c\u4e3a\u7684\u5355\u4e00\u5173\u6ce8\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u91cd\u8981\u6027\u6807\u7b7e\uff08importance labels\uff09\u8868\u660e\u7279\u5b9a\u89c2\u5bdf\u7ed3\u679c\u5bf9\u5b9e\u73b0\u76ee\u6807\u7684\u91cd\u8981\u6027\uff0c\u63d0\u4f9b\u4e0d\u76f4\u63a5\u6539\u53d8\u884c\u4e3a\u7684\u663e\u5f0f\u53cd\u9988\u3002\u8fd9\u79cd\u53cd\u9988\u56e0\u4e0a\u4e0b\u6587\u800c\u5f02\uff0c\u5e76\u4f5c\u4e3a\u8865\u5145\u8f93\u5165\uff0c\u5f3a\u5316\u5956\u52b1\u6a21\u578b\u7684\u6574\u4f53\u5b66\u4e60\u8fc7\u7a0b\u3002<\/li>\n<li><strong>\u7279\u5b9a\u4e8e\u8868\u793a\u7684\u53cd\u9988\uff08Representation-Specific Feedback\uff09\uff1a<\/strong>\u67d0\u4e9b\u53cd\u9988\u7c7b\u578b\u4e3b\u8981\u589e\u5f3a\u8868\u793a\u5b66\u4e60\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u5851\u9020\u5956\u52b1\u51fd\u6570\u3002 \u7279\u5f81\u8f68\u8ff9\uff08Feature Traces\uff09\u63d0\u793a\u4eba\u7c7b\u64cd\u4f5c\u5458\u6f14\u793a\u7ed9\u5b9a\u7279\u5f81\u7684\u5355\u8c03\u53d8\u5316\uff0c\u4ece\u800c\u5b9e\u73b0\u7279\u5f81\u96c6\u7684\u52a8\u6001\u6269\u5c55\u3002 \u76f8\u4f3c\u6027\u67e5\u8be2\uff08 Similarity Queries\uff09\u6bd4\u8f83\u8f68\u8ff9\u7684\u4e09\u5143\u7ec4\uff0c\u901a\u8fc7\u8f68\u8ff9\u7a7a\u95f4\u4e2d\u7684\u6210\u5bf9\u8ddd\u79bb\u5f15\u5bfc\u8868\u793a\u5b66\u4e60\u3002\u901a\u8fc7\u5229\u7528\u8fd9\u4e9b\u7279\u5b9a\u4e8e\u8868\u793a\u7684\u53cd\u9988\u5f62\u5f0f\uff0cRLHF \u53ef\u4ee5\u5b9e\u73b0\u5bf9\u65b0\u4efb\u52a1\u548c\u4e0a\u4e0b\u6587\u66f4\u5f3a\u5927\u7684\u6cdb\u5316\u3002<\/li>\n<\/ul>\n<h5>RLHF\u7684\u5956\u52b1\u6a21\u578b<\/h5>\n<p>\u771f\u5b9e\u7684\u5956\u52b1\u51fd\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">r(x,y)<\/span> \u901a\u5e38\u662f\u672a\u77e5\u7684\uff0c\u56e0\u6b64\u6709\u5fc5\u8981\u6839\u636e\u4eba\u7c7b\u63d0\u4f9b\u7684\u504f\u597d\u6784\u5efa\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u5956\u52b1\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">r_\\theta(x,y)<\/span>\u3002 \u8be5\u6a21\u578b\u9884\u6d4b\u5019\u9009\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">y<\/span> \u4e0e\u4eba\u7c7b\u5bf9\u7ed9\u5b9a\u8f93\u5165 <span class=\"katex-eq\" data-katex-display=\"false\">x<\/span> \u7684\u671f\u671b\u7684\u5bf9\u9f50\u7a0b\u5ea6\u3002 \u4e3a\u4e86\u83b7\u53d6 <span class=\"katex-eq\" data-katex-display=\"false\">r_\\theta(x,y)<\/span> \u7684\u8bad\u7ec3\u6570\u636e\uff0c\u4eba\u7c7b\u8bc4\u4f30\u8005\u6839\u636e\u5176\u76f8\u5bf9\u9002\u7528\u6027\u6bd4\u8f83\u6216\u6807\u6ce8\u8f93\u51fa\u5bf9\uff0c\u5e76\u4e14\u8be5\u6a21\u578b\u901a\u5e38\u4f7f\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u5728\u8fd9\u4e9b\u6bd4\u8f83\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002 \u4e3a\u4e86\u963b\u6b62\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span> \u504f\u79bb\u521d\u59cb\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span> \u592a\u8fdc\uff0c\u5c06\u4e00\u4e2a\u7531\u8d85\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">\\beta<\/span> \u63a7\u5236\u7684\u60e9\u7f5a\u9879\u5f15\u5165\u5956\u52b1\u51fd\u6570\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">r_\\theta(x,y)&gt;[katex]r_\\theta(x,y)=r(x,y)-\\beta\\log\\frac{\\pi(y\\mid x)}{\\rho(y\\mid x)}<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\pi(y\\mid x)<\/span> \u662f\u5fae\u8c03\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span> \u4ea7\u751f\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">y<\/span> \u5728\u7ed9\u5b9a\u8f93\u5165 <span class=\"katex-eq\" data-katex-display=\"false\">x<\/span> \u4e0b\u7684\u6982\u7387\uff0c\u5e76\u4e14 <span class=\"katex-eq\" data-katex-display=\"false\">\\rho(y\\mid x)<\/span> \u662f\u539f\u59cb\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span> \u4e0b\u7684\u5bf9\u5e94\u6982\u7387\u3002 \u8fd9\u4e2a\u672f\u8bed\u786e\u4fdd\uff0c\u5c3d\u7ba1 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi<\/span> \u9002\u5e94\u4eba\u7c7b\u53cd\u9988\uff0c\u5b83\u4ecd\u7136\u53d7\u5230\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">\\rho<\/span> \u4e2d\u6355\u83b7\u7684\u5148\u9a8c\u77e5\u8bc6\u7684\u7ea6\u675f\u3002<br \/>\n\u8bc4\u4f30\u5956\u52b1\u51fd\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">r_\\theta(x,y)<\/span> \u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u5b83\u76f4\u63a5\u5f71\u54cd\u5b66\u4e60\u6548\u679c\u548c\u7b56\u7565\u6027\u80fd\u3002 \u51c6\u786e\u8bc4\u4f30\u6b64\u51fd\u6570\u6709\u52a9\u4e8e\u786e\u5b9a\u5408\u9002\u7684\u5956\u52b1\u7ed3\u6784\uff0c\u4ee5\u4f7f\u6a21\u578b\u8f93\u51fa\u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\u3002 \u7136\u800c\uff0c\u5728\u5b89\u5168\u654f\u611f\u7684\u9886\u57df\u4e2d\uff0c\u6807\u51c6\u7684\u5c55\u5f00\u65b9\u6cd5\u548c\u79bb\u7b56\u7565\u8bc4\u4f30\u53ef\u80fd\u7531\u4e8e\u4e0e\u5728\u7ebf\u4ea4\u4e92\u3001\u504f\u5dee\u548c\u5bf9\u771f\u5b9e\u5956\u52b1\u7684\u9700\u6c42\u76f8\u5173\u7684\u98ce\u9669\u800c\u4e0d\u53ef\u884c\u3002 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\uff0c\u901a\u5e38\u91c7\u7528\u4e24\u79cd\u7a81\u51fa\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li>\u8ddd\u79bb\u51fd\u6570(Distance Functions)\uff1a\u6700\u8fd1\u7684\u7814\u7a76\u96c6\u4e2d\u4e8e\u5956\u52b1\u8bc4\u4f30\u8ddd\u79bb\u51fd\u6570\uff0c\u8fd9\u4e9b\u51fd\u6570\u8003\u8651\u4e86\u6f5c\u5728\u7684\u53d8\u6362\uff0c\u4f8b\u5982\u6f5c\u5728 shaping\u3002 \u4f8b\u5982\uff0cEPIC \u6d4b\u91cf\u5404\u79cd\u53d8\u6362\u4e0b\u7684\u5956\u52b1\u51fd\u6570\u7b49\u4ef7\u6027\uff0c\u800c DARD \u63d0\u70bc\u4e86\u89c4\u8303\u5316\uff0c\u4ee5\u786e\u4fdd\u8bc4\u4f30\u4fdd\u6301\u5728\u53ef\u884c\u7684\u8f6c\u6362\u4e2d\u3002 \u7c7b\u4f3c\u4e8e EPIC \u7684\u8ddd\u79bb\u901a\u8fc7\u5141\u8bb8\u89c4\u8303\u5316\u3001\u6807\u51c6\u5316\u548c\u5ea6\u91cf\u51fd\u6570\u7684\u53d8\u5f02\u6027\u6765\u6982\u62ec EPIC \u7684\u65b9\u6cd5\uff0c\u5e76\u4e14 STARC \u4fdd\u7559\u4e86 EPIC \u7684\u7406\u8bba\u7279\u6027\uff0c\u540c\u65f6\u63d0\u4f9b\u4e86\u989d\u5916\u7684\u7075\u6d3b\u6027\u3002<\/li>\n<li>\u89c6\u89c9\u548c\u4eba\u5de5\u68c0\u67e5(Visual and Human Inspection):\u5176\u4ed6\u65b9\u6cd5\u4f9d\u8d56\u4e8e\u53ef\u89e3\u91ca\u6027\u548c\u7cbe\u5fc3\u7b56\u5212\u7684\u6570\u636e\u96c6\u6765\u8861\u91cf\u5b66\u4e60\u5230\u7684\u5956\u52b1\u51fd\u6570\u7684\u6709\u6548\u6027\u3002 PRFI \u4f7f\u7528\u9884\u5904\u7406\u6b65\u9aa4\u6765\u7b80\u5316\u5956\u52b1\u51fd\u6570\uff0c\u540c\u65f6\u4fdd\u7559\u7b49\u4ef7\u6027\uff0c\u4ece\u800c\u589e\u5f3a\u5176\u900f\u660e\u5ea6\u3002 \u540c\u65f6\uff0cCONVEXDA \u548c REWARDFUSION \u63d0\u51fa\u4e86\u65e8\u5728\u6d4b\u8bd5\u5956\u52b1\u6a21\u578b\u5982\u4f55\u4e00\u81f4\u5730\u54cd\u5e94\u63d0\u793a\u4e2d\u8bed\u4e49\u53d8\u5316\u7684\u7684\u6570\u636e\u96c6\u3002 \u603b\u4e4b\uff0c\u8fd9\u4e9b\u6280\u672f\u6709\u52a9\u4e8e\u66f4\u53ef\u9760\u5730\u8bc4\u4f30\u5956\u52b1\u51fd\u6570\uff0c\u4ece\u800c\u52a0\u5f3a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e0e\u4eba\u7c7b\u504f\u597d\u7684\u5bf9\u9f50\u3002<\/li>\n<\/ul>\n<h5>RLHF\u7684\u7b56\u7565\u5b66\u4e60<\/h5>\n<p>RLHF\u7684\u7b56\u7565\u5b66\u4e60\uff0c\u5982\u56fe\u6240\u793a\u6d89\u53ca\u901a\u8fc7\u5728\u7ebf\u548c\u79bb\u7ebf\u8bbe\u7f6e\u4e2d\u7684\u4eba\u7c7b\u53cd\u9988\u6765\u4f18\u5316\u653f\u7b56\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104090716320.png\" alt=\"\" \/><\/p>\n<p><center>\u5728\u7ebf\u548c\u79bb\u7ebfRLHF\u7684\u6bd4\u8f83\uff0c\u8bf4\u660e\u5728\u7ebfRLHF\u4e2d\u7684\u7b56\u7565\u6267\u884c\u8fc7\u7a0b\u4e2d\u7684\u8fde\u7eed\u53cd\u9988\u6536\u96c6\u4e0e\u79bb\u7ebfRLHF\u4e2d\u7684\u9884\u91c7\u7528\u7684\u8f68\u8ff9\u5229\u7528\u7387\u76f8\u6bd4<\/center><\/p>\n<ul>\n<li><strong>\u5728\u7ebf\u5b66\u4e60:(Online Learning)<\/strong>:\u5728\u5728\u7ebfRLHF\u4e2d\uff0c\u7cfb\u7edf\u4f1a\u6536\u96c6\u6709\u5173\u65b0\u751f\u6210\u7684\u6a21\u578b\u8f68\u8ff9\u7684\u5b9e\u65f6\u4eba\u7c7b\u504f\u597d\u3002 DPS \u4f7f\u7528\u8d1d\u53f6\u65af\u66f4\u65b0\u6765\u7ba1\u7406\u51b3\u6597\u8fc7\u7a0b\uff0c\u800cPPS\u548cPEPS \u96c6\u6210\u52a8\u6001\u7f16\u7a0b\u548c\u5f3a\u76d7\u60f3\u6cd5\u6765\u5b8c\u5584\u653f\u7b56\u884c\u4e3a\u3002 \u5728lpbrl \u4e2d\uff0c\u7279\u5f81\u5d4c\u5165\u5f0f\u6355\u83b7\u4e0d\u65ad\u53d1\u5c55\u7684\u5956\u52b1\u7ed3\u6784\uff0c\u800cPBOP \u90fd\u96c6\u6210\u4e86\u8fc7\u6e21\u52a8\u529b\u5b66\u548c\u9996\u9009\u9879\u4fe1\u53f7\u7684\u6700\u5c0f\u4e8c\u4e58\u4f30\u8ba1\u503c\u3002 \u6700\u8fd1\uff0cPARL \u901a\u8fc7\u5c06\u53cd\u9988\u91c7\u96c6\u89c6\u4e3a\u7b56\u7565\u4f18\u5316\u7684\u7ec4\u6210\u90e8\u5206\uff0c\u4ee5\u6570\u636e\u6536\u96c6\u6548\u7387\u4e3a\u76ee\u6807\u3002<\/li>\n<li><strong>\u79bb\u7ebf\u5b66\u4e60(Offline Learning):<\/strong>\u5728\u79bb\u7ebfRLHF\u4e2d\uff0c\u4ee5\u524d\u6536\u96c6\u7684\u504f\u597d\u6807\u8bb0\u8f68\u8ff9\u7528\u4e8e\u5b66\u4e60\u6216\u5b8c\u5584\u653f\u7b56\u3002 \u4f8b\u5982\uff0c \u7814\u7a76\u5bf9\u7b56\u7565\u5b66\u4e60\u7684\u60b2\u89c2\u6700\u5927\u4f3c\u7136\u4f30\u8ba1\uff0c\u5e76\u901a\u8fc7\u6210\u5bf9\u6bd4\u8f83\u6570\u636e\uff0c\u5efa\u7acb\u4e86\u7ee9\u6548\u7684\u754c\u9650\u3002 \u5f92\u624b \u548cdcppo \u7b49\u6269\u5c55\uff0c\u63a2\u7d22\u4e86\u79bb\u7ebf\u6570\u636e\u8986\u76d6\u8303\u56f4\u548c\u7b56\u7565\u6982\u62ec\u4e4b\u95f4\u7684\u76f8\u4e92\u4f5c\u7528\u3002 \u6b64\u5916\uff0c \u5730\u5740\u5728Boltzmann\u6a21\u578b\u4e2d\u8fdb\u884c\u8fc7\u5ea6\u62df\u5408\uff0c\u4ee5\u8fdb\u884c\u6210\u5bf9\u6bd4\u8f83\uff0c\u800cDCPPO \u8fdb\u4e00\u6b65\u7814\u7a76\u4e86\u63d0\u9ad8\u53cd\u9988\u6548\u7387\u7684\u52a8\u6001\u79bb\u6563\u9009\u62e9\u6a21\u578b\u3002<\/li>\n<li><strong>\u5728\u7ebf\u548c\u79bb\u7ebf\u5b66\u4e60\u878d\u5408(Blending Online and Offline Learning):<\/strong>\u6df7\u5408\u65b9\u6cd5\u5c06\u79bb\u7ebf\u9884\u5904\u7406\u4e0e\u5728\u7ebf\u504f\u597d\u6c47\u603b\u7ed3\u5408\u5728\u4e00\u8d77\uff0c\u5229\u7528\u9884\u6536\u53d6\u7684\u6570\u636e\uff0c\u540c\u65f6\u4ecd\u5728\u5408\u5e76\u5b9e\u65f6\u66f4\u65b0\u3002 PFERL \u91c7\u7528\u4e86\u4e24\u9636\u6bb5\u7684\u65b9\u6cd5\u6765\u6700\u5927\u7a0b\u5ea6\u5730\u51cf\u5c11\u4eba\u7c7b\u67e5\u8be2\uff0c\u800cPerl \u63a2\u7d22\u4e86\u79ef\u6781\u63a2\u7d22\u7684\u4e50\u89c2\u6700\u5c0f\u4e8c\u4e58\u7b56\u7565\u3002 Dueling RL \u53ca\u5176\u6269\u5c55 (\u4f8b\u5982\uff0cPRPRL \u4e2d\u7684 REGIME) \u901a\u8fc7\u4ed4\u7ec6\u5212\u5206\u6570\u636e\u83b7\u53d6\u4e0e\u53cd\u9988\u6536\u96c6\u6765\u51cf\u5c11\u4eba\u5de5\u6807\u6ce8\u9700\u6c42\uff0c\u4ece\u800c\u4f18\u5316\u6837\u672c\u6548\u7387\u3001\u6807\u6ce8\u6210\u672c\u548c\u7b56\u7565\u6027\u80fd\u4e4b\u95f4\u7684\u6743\u8861\u3002<\/li>\n<\/ul>\n<h4>\u4f7f\u7528AI\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60<\/h4>\n<p>\u4f7f\u7528 AI \u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60 (Reinforcement Learning with AI Feedback,RLAIF) \u901a\u8fc7\u4f7f\u7528 LLM \u751f\u6210\u53cd\u9988\u4fe1\u53f7\u6765\u6269\u5c55 RLHF \u8303\u5f0f\u3002 \u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u8865\u5145\u6216\u53d6\u4ee3\u4eba\u5de5\u53cd\u9988\uff0c\u5728\u4eba\u5de5\u6807\u6ce8\u7a00\u7f3a\u3001\u6210\u672c\u9ad8\u6216\u4e0d\u4e00\u81f4\u7684\u4efb\u52a1\u4e2d\u63d0\u4f9b\u66f4\u5177\u53ef\u6269\u5c55\u6027\u3001\u66f4\u4f4e\u6210\u672c\u7684\u504f\u597d\u6570\u636e\u3002<\/p>\n<h5>RLAIF vs RLHF<\/h5>\n<p>\u5728\u5927\u89c4\u6a21\u5e94\u7528 RLHF \u7684\u4e3b\u8981\u6311\u6218\u5728\u4e8e\u5b83\u4f9d\u8d56\u4e8e\u4eba\u5de5\u751f\u6210\u7684\u504f\u597d\u6807\u7b7e\uff0c\u8fd9\u9700\u8981\u5927\u91cf\u8d44\u6e90\u6765\u6536\u96c6\u3001\u6574\u7406\u548c\u6807\u6ce8\u6570\u636e\u3002 \u6807\u6ce8\u6570\u636e\u7684\u8fc7\u7a0b\u65e2\u8017\u65f6\u53c8\u6602\u8d35\uff0c\u800c\u4e14\u4eba\u5de5\u8bc4\u4f30\u8005\u53ef\u80fd\u4f1a\u5f15\u5165\u4e0d\u4e00\u81f4\u6027\uff0c\u4ece\u800c\u4f7f\u5927\u89c4\u6a21\u3001\u4e00\u81f4\u7684\u6807\u7b7e\u5728\u6240\u6709\u6a21\u578b\u8f93\u51fa\u4e2d\u53d8\u5f97\u590d\u6742\u3002 \u8fd9\u4e9b\u9650\u5236\u6781\u5927\u5730\u9650\u5236\u4e86 RLHF \u7684\u53ef\u6269\u5c55\u6027\u548c\u6548\u7387\u3002 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\uff0c\u5b66\u8005\u63d0\u51fa\u4e86 RLAIF\uff0c\u5b83\u5c06\u4eba\u5de5\u53cd\u9988\u4e0e AI \u751f\u6210\u7684\u53cd\u9988\u76f8\u7ed3\u5408\uff0c\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u6a21\u578b\u3002 \u901a\u8fc7\u5229\u7528 LLM \u4f5c\u4e3a\u53cd\u9988\u7684\u6765\u6e90\uff0cRLAIF \u51cf\u5c11\u4e86\u5bf9\u4eba\u5de5\u6807\u6ce8\u8005\u7684\u4f9d\u8d56\uff0c\u4e3a\u4f20\u7edf\u7684 RLHF \u63d0\u4f9b\u4e86\u4e00\u79cd\u53ef\u884c\u7684\u66ff\u4ee3\u65b9\u6848\u3002 \u8fd9\u79cd\u65b9\u6cd5\u652f\u6301\u6301\u7eed\u7684\u53cd\u9988\u751f\u6210\uff0c\u663e\u8457\u63d0\u9ad8\u4e86\u53ef\u6269\u5c55\u6027\uff0c\u540c\u65f6\u4fdd\u7559\u4e86\u4eba\u7c7b\u6307\u5bfc\u7684\u6a21\u578b\u4f18\u5316\u7684\u7075\u6d3b\u6027\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104091601784.png\" alt=\"\" \/><\/p>\n<p><center>\u6bd4\u8f83 RLHF \u548c RLAIF \u65b9\u6cd5\uff0c\u63cf\u7ed8\u4e86\u5b83\u4eec\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e2d\u504f\u597d\u5bf9\u9f50\u7684\u4e0d\u540c\u65b9\u6cd5<\/center>\u5982\u56fe\u6240\u793a\uff0cRLHF \u548c RLAIF \u4e4b\u95f4\u7684\u5173\u952e\u533a\u522b\u5728\u4e8e\u53cd\u9988\u7684\u6765\u6e90\uff1aRLHF \u4f9d\u8d56\u4e8e\u4eba\u5de5\u751f\u6210\u7684\u504f\u597d\uff0c\u800c RLAIF \u4f7f\u7528 AI \u751f\u6210\u7684\u53cd\u9988\u6765\u6307\u5bfc\u7b56\u7565\u66f4\u65b0\u3002 \u7ecf\u7814\u7a76\uff0cRLAIF \u53ef\u4ee5\u5b9e\u73b0\u4e0e RLHF \u76f8\u5f53\u751a\u81f3\u66f4\u4f18\u7684\u6027\u80fd\uff0c\u7531\u4eba\u7c7b\u8bc4\u5206\u8005\u8bc4\u4f30\u3002 \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cRLAIF \u4e0d\u4ec5\u8d85\u8d8a\u4e86\u4f20\u7edf\u7684\u76d1\u7763\u5fae\u8c03\u57fa\u7ebf\uff0c\u800c\u4e14\u4f7f\u7528\u4e86\u4e0e\u7b56\u7565\u6a21\u578b\u89c4\u6a21\u76f8\u540c\u7684 LLM \u504f\u597d\u6807\u6ce8\u5668\uff0c\u7a81\u51fa\u4e86\u8be5\u65b9\u6cd5\u7684\u6548\u7387\u3002<\/p>\n<h5>RLAIF\u8bad\u7ec3\u6d41\u7a0b<\/h5>\n<p>RLAIF \u8bad\u7ec3\u6d41\u7a0b\u9075\u5faa\u51e0\u4e2a\u5173\u952e\u9636\u6bb5\uff0c\u5176\u4e2d\u5229\u7528 AI \u751f\u6210\u7684\u53cd\u9988\u6765\u8fed\u4ee3\u5730\u6539\u8fdb\u6a21\u578b\u7684\u884c\u4e3a\u3002 \u8be5\u6d41\u7a0b\u6709\u52a9\u4e8e\u5c06\u5927\u8bed\u8a00\u6a21\u578b (LLM) \u7684\u8f93\u51fa\u4e0e\u4eba\u7c7b\u671f\u671b\u5bf9\u9f50\uff0c\u8fd9\u79cd\u65b9\u5f0f\u53ef\u4ee5\u6269\u5c55\u5230\u5404\u79cd\u4efb\u52a1\uff0c\u9636\u6bb5\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li><strong>AI \u53cd\u9988\u6536\u96c6(AI Feedback Collection)\uff1a<\/strong>\u5728\u6b64\u9636\u6bb5\uff0cAI \u7cfb\u7edf\u6839\u636e\u9884\u5b9a\u4e49\u7684\u6807\u51c6\u751f\u6210\u53cd\u9988\uff0c\u8fd9\u4e9b\u6807\u51c6\u53ef\u80fd\u5305\u62ec\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6307\u6807\u3001\u54cd\u5e94\u7684\u6b63\u786e\u6027\u6216\u6a21\u578b\u8f93\u51fa\u7684\u9002\u5f53\u6027\u3002\u4e0e\u9700\u8981\u89e3\u91ca\u548c\u624b\u52a8\u6807\u6ce8\u7684\u4eba\u7c7b\u53cd\u9988\u4e0d\u540c\uff0cAI \u53cd\u9988\u53ef\u4ee5\u5728\u5404\u79cd\u6a21\u578b\u8f93\u51fa\u4e2d\u4e00\u81f4\u5730\u751f\u6210\u3002\u8fd9\u79cd\u7279\u6027\u4f7f\u5f97 AI \u53cd\u9988\u53ef\u4ee5\u6301\u7eed\u63d0\u4f9b\uff0c\u4ece\u800c\u663e\u8457\u6269\u5927\u53cd\u9988\u5faa\u73af\u3002<\/li>\n<li><strong>\u5956\u52b1\u6a21\u578b\u8bad\u7ec3(Reward Model Training):<\/strong>\u968f\u540e\uff0c\u4f7f\u7528 AI \u751f\u6210\u7684\u53cd\u9988\u6765\u8bad\u7ec3\u6216\u4f18\u5316\u5956\u52b1\u6a21\u578b\u3002\u8be5\u6a21\u578b\u5c06\u8f93\u5165-\u8f93\u51fa\u5bf9\u6620\u5c04\u5230\u76f8\u5e94\u7684\u5956\u52b1\uff0c\u4ece\u800c\u4f7f\u6a21\u578b\u7684\u8f93\u51fa\u4e0e\u53cd\u9988\u6240\u6307\u793a\u7684\u671f\u671b\u7ed3\u679c\u5bf9\u9f50\u3002\u867d\u7136\u4f20\u7edf\u7684 RLHF \u4f9d\u8d56\u4e8e\u76f4\u63a5\u7684\u4eba\u7c7b\u53cd\u9988\u6765\u8bc4\u4f30\u8f93\u51fa\uff0c\u4f46 RLAIF \u4f7f\u7528 AI \u751f\u6210\u7684\u6807\u7b7e\uff0c\u5c3d\u7ba1\u8fd9\u53ef\u80fd\u4f1a\u5f15\u5165\u4e0e\u4e00\u81f4\u6027\u548c\u504f\u89c1\u76f8\u5173\u7684\u95ee\u9898\uff0c\u4f46\u5728\u53ef\u6269\u5c55\u6027\u548c\u72ec\u7acb\u4e8e\u4eba\u529b\u8d44\u6e90\u65b9\u9762\u5177\u6709\u4f18\u52bf\u3002<\/li>\n<li><strong>\u7b56\u7565\u66f4\u65b0(Policy Update):<\/strong>\u6700\u540e\u9636\u6bb5\u5305\u62ec\u6839\u636e\u4e0a\u4e00\u6b65\u8bad\u7ec3\u7684\u5956\u52b1\u6a21\u578b\u66f4\u65b0\u6a21\u578b\u7684\u7b56\u7565\u3002 \u91c7\u7528\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u6765\u8c03\u6574\u6a21\u578b\u7684\u53c2\u6570\uff0c\u4f18\u5316\u7b56\u7565\u4ee5\u6700\u5927\u5316\u5404\u79cd\u4efb\u52a1\u7684\u7d2f\u79ef\u5956\u52b1\u3002 \u8fd9\u4e2a\u8fc7\u7a0b\u662f\u8fed\u4ee3\u7684\uff0c\u5956\u52b1\u6a21\u578b\u5f15\u5bfc\u6a21\u578b\u7684\u8f93\u51fa\uff0c\u4f7f\u5176\u4e0e\u9884\u671f\u76ee\u6807\u4fdd\u6301\u66f4\u9ad8\u7684\u4e00\u81f4\u6027\u3002<\/li>\n<\/ul>\n<p>RLAIF \u7684\u4e3b\u8981\u4f18\u52bf\u5728\u4e8e\u5b83\u80fd\u591f\u6269\u5c55\u53cd\u9988\u5faa\u73af\uff0c\u800c\u65e0\u9700\u6301\u7eed\u7684\u4eba\u5de5\u5e72\u9884\u3002 \u901a\u8fc7\u7528 AI \u751f\u6210\u7684\u53cd\u9988\u53d6\u4ee3\u4eba\u5de5\u53cd\u9988\uff0cRLAIF \u4fc3\u8fdb\u4e86 LLM \u5728\u591a\u4e2a\u4efb\u52a1\u4e0a\u7684\u6301\u7eed\u6539\u8fdb\uff0c\u7f13\u89e3\u4e86\u4eba\u5de5\u6807\u6ce8\u5de5\u4f5c\u5e26\u6765\u7684\u74f6\u9888\u3002<\/p>\n<h4>\u76f4\u63a5\u504f\u597d\u4f18\u5316<\/h4>\n<p>\u5982\u524d\u6240\u8ff0\uff0cRLHF \u901a\u5e38\u7531\u4e09\u4e2a\u9636\u6bb5\u7ec4\u6210\uff1a\u76d1\u7763\u5fae\u8c03\u3001\u5956\u52b1\u5efa\u6a21\u548c\u5f3a\u5316\u5b66\u4e60\uff08\u901a\u5e38\u901a\u8fc7\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff0cPPO\uff09\u5b9e\u73b0\u3002\u5c3d\u7ba1 RLHF \u5177\u6709\u6709\u6548\u6027\uff0c\u4f46\u5b83\u53ef\u80fd\u5f88\u590d\u6742\u4e14\u5bb9\u6613\u51fa\u73b0\u4e0d\u7a33\u5b9a\uff0c\u7279\u522b\u662f\u5728\u62df\u5408\u5956\u52b1\u6a21\u578b\u7136\u540e\u7528\u4e8e\u5fae\u8c03\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u9636\u6bb5\u3002 \u96be\u70b9\u5728\u4e8e\u521b\u5efa\u80fd\u591f\u51c6\u786e\u53cd\u6620\u4eba\u7c7b\u504f\u597d\u7684\u5956\u52b1\u6a21\u578b\uff0c\u4ee5\u53ca\u5728\u5fae\u8c03\u8bed\u8a00\u6a21\u578b\u4ee5\u4f18\u5316\u8fd9\u4e2a\u4f30\u8ba1\u5956\u52b1\u7684\u540c\u65f6\u4fdd\u6301\u63a5\u8fd1\u539f\u59cb\u6a21\u578b\u7684\u6311\u6218\u3002 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u76f4\u63a5\u504f\u597d\u4f18\u5316 (DPO)\u88ab\u5f15\u5165\u4f5c\u4e3a\u4e00\u79cd\u66f4\u7a33\u5b9a\u4e14\u8ba1\u7b97\u6548\u7387\u66f4\u9ad8\u7684\u66ff\u4ee3\u65b9\u6848\u3002 DPO \u901a\u8fc7\u5c06\u5956\u52b1\u51fd\u6570\u76f4\u63a5\u94fe\u63a5\u5230\u6700\u4f18\u7b56\u7565\u6765\u7b80\u5316\u5956\u52b1\u4f18\u5316\u8fc7\u7a0b\u3002\u5b83\u5c06\u5956\u52b1\u6700\u5927\u5316\u95ee\u9898\u89c6\u4e3a\u57fa\u4e8e\u4eba\u7c7b\u504f\u597d\u6570\u636e\u7684\u5355\u9636\u6bb5\u7b56\u7565\u8bad\u7ec3\u95ee\u9898\uff0c\u4ece\u800c\u907f\u514d\u4e86\u5956\u52b1\u6a21\u578b\u62df\u5408\u7684\u590d\u6742\u6027\u4ee5\u53ca Bradley-Terry \u6a21\u578b\u7684\u4f9d\u8d56\u6027\u3002<\/p>\n<h5>DPO\u7684\u57fa\u7840<\/h5>\n<p>RLHF \u6d89\u53ca\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u5956\u52b1\u6a21\u578b (RM) \u548c\u5fae\u8c03\u8bed\u8a00\u6a21\u578b (LM)\u3002 DPO \u901a\u8fc7\u4f7f\u7528\u4eba\u7c7b\u504f\u597d\u6570\u636e\u76f4\u63a5\u8bad\u7ec3 LM \u6765\u7b80\u5316\u6b64\u8fc7\u7a0b\uff0c\u5728\u7b56\u7565\u672c\u8eab\u4e2d\u9690\u5f0f\u5730\u6355\u83b7\u4e86\u5956\u52b1\u6a21\u578b\u3002<br \/>\nKL \u6b63\u5219\u5316\u5956\u52b1\u6700\u5927\u5316\u76ee\u6807\u3002 DPO \u4ece\u5df2\u5efa\u7acb\u7684 KL \u6b63\u5219\u5316\u5956\u52b1\u6700\u5927\u5316\u6846\u67b6\u5f00\u59cb\uff0c\u5982\u4e0b\u9762\u7684\u76ee\u6807\u6240\u793a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*&gt;[katex]\\pi^* = \\arg\\max_\\pi \\mathbb{E}{x\\sim\\mathcal{D},y\\sim\\pi(\\cdot\\mid x)}!\\Bigl[r(x,y)-\\beta,\\text{KL}!\\bigl(\\pi(\\cdot\\mid x)|\\pi{\\text{ref}}(\\cdot\\mid x)\\bigr)\\Bigr]<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">r(x,y)<\/span> \u4ee3\u8868\u5956\u52b1\u51fd\u6570\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\beta&gt;0<\/span> \u662f\u4e00\u4e2a\u63a7\u5236\u4e0e\u53c2\u8003\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span> \u63a5\u8fd1\u7a0b\u5ea6\u7684\u7cfb\u6570\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\text{KL}(\\cdot|\\cdot)<\/span> \u8868\u793a Kullback-Leibler \u6563\u5ea6\u3002 \u5728\u8fd9\u91cc\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">x\\sim\\mathcal{D}<\/span> \u4ee3\u8868\u4ece\u6570\u636e\u5206\u5e03\u4e2d\u63d0\u53d6\u7684\u8f93\u5165\uff0c\u800c <span class=\"katex-eq\" data-katex-display=\"false\">y\\sim\\pi(\\cdot\\mid x)<\/span> \u8868\u793a\u4ece\u7b56\u7565\u4e2d\u91c7\u6837\u7684\u8f93\u51fa\u3002<br \/>\n\u5bfc\u51fa\u6700\u4f18\u7b56\u7565\u3002 \u5728\u9002\u5f53\u7684\u5047\u8bbe\u4e0b\uff0c\u5f0f (14) \u7684\u89e3\u4ee5 Boltzmann \u5206\u5e03\u7684\u5f62\u5f0f\u5bfc\u51fa\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*(y\\mid&gt;[katex]\\pi^*(y\\mid x)=\\frac{1}{Z(x)},\\pi_{\\text{ref}}(y\\mid x),\\exp!\\Bigl(\\frac{1}{\\beta},r(x,y)\\Bigr)<\/span><\/center>\u5176\u4e2d\u914d\u5206\u51fd\u6570:<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">Z(x)&gt;[katex]Z(x)=\\sum_y\\pi_{\\text{ref}}(y\\mid x),\\exp!\\Bigl(\\frac{1}{\\beta},r(x,y)\\Bigr)<\/span><\/center>\u5145\u5f53\u4e00\u4e2a\u5f52\u4e00\u5316\u9879\uff0c\u786e\u4fdd <span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*<\/span> \u4ecd\u7136\u662f\u4e00\u4e2a\u6709\u6548\u7684\u6982\u7387\u5206\u5e03\uff08\u5373\uff0c\u5176\u6982\u7387\u603b\u548c\u4e3a 1\uff09\u3002<br \/>\n\u91cd\u65b0\u53c2\u6570\u5316\u5956\u52b1\u3002 \u5bf9\u5f0f (15) \u7684\u4e24\u8fb9\u53d6\u81ea\u7136\u5bf9\u6570\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u5956\u52b1 <span class=\"katex-eq\" data-katex-display=\"false\">r(x,y)<\/span> \u4e0e\u6700\u4f18\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*<\/span> \u8054\u7cfb\u8d77\u6765\u3002 \u8fd9\u5c31\u4ea7\u751f\u4e86\uff1a<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">r^(x,y)=\\beta,\\Bigl[\\log\\pi^(y\\mid x)-\\log\\pi_{\\text{ref}}(y\\mid x)\\Bigr]+\\beta\\log Z(x)<\/span><br \/>\n\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\beta\\log Z(x)<\/span> \u662f\u4e00\u4e2a\u5e38\u6570\uff0c\u4e0d\u5f71\u54cd\u5956\u52b1\u7684\u6210\u5bf9\u6bd4\u8f83\u3002 \u5982\u679c\u5df2\u77e5\u6700\u4f18\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi^<\/span>\uff0c\u5219\u53ef\u4ee5\u786e\u5b9a\u771f\u5b9e\u7684\u5956\u52b1 <span class=\"katex-eq\" data-katex-display=\"false\">r^(x,y)<\/span>\uff0c\u76f4\u81f3\u8be5\u5e38\u6570\u3002<br \/>\nBradley\u2013Terry \u504f\u597d\u3002 \u5728 Bradley-Terry \u6a21\u578b\u4e0b\uff0c\u4eba\u7c7b\u5bf9\u4e24\u4e2a\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">y_1<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">y_2<\/span> \u4e4b\u95f4\u7684\u504f\u597d\u53d7\u5176\u5956\u52b1\u503c\u5dee\u5f02\u7684\u63a7\u5236\u3002 \u504f\u597d <span class=\"katex-eq\" data-katex-display=\"false\">y_1<\/span> \u80dc\u8fc7 <span class=\"katex-eq\" data-katex-display=\"false\">y_2<\/span> \u7684\u6982\u7387\u7531\u4e0b\u5f0f\u7ed9\u51fa:<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">p^(y_1\\succ y_2\\mid x)=\\frac{\\exp!\\bigl(r^(x,y_1)\\bigr)}{\\exp!\\bigl(r^(x,y_1)\\bigr)+\\exp!\\bigl(r^(x,y_2)\\bigr)}<\/span><\/center>\u5c06\u5f0f (17) \u4ee3\u5165\u5f0f (18)\uff0c\u6211\u4eec\u5f97\u5230\u6700\u7ec8\u7684\u504f\u597d\u6a21\u578b\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">p^(y_1\\succ y_2\\mid x)=\\frac{1}{1+\\exp!\\Bigl(\\beta,\\Bigl[\\log\\frac{\\pi^(y_2\\mid x)}{\\pi_{\\text{ref}}(y_2\\mid x)}-\\log\\frac{\\pi^*(y_1\\mid x)}{\\pi_{\\text{ref}}(y_1\\mid x)}\\Bigr]\\Bigr)}<\/span><\/center>\u8be5\u8868\u8fbe\u5f0f\u5c06\u6210\u5bf9\u4eba\u7c7b\u504f\u597d\u6982\u7387\u4e0e\u6700\u4f18\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*<\/span> \u548c\u53c2\u8003\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span> \u7684\u6bd4\u7387\u8054\u7cfb\u8d77\u6765\u3002<br \/>\nDPO \u7684\u76ee\u6807:DPO \u901a\u8fc7\u76f4\u63a5\u4ece\u504f\u597d\u6570\u636e\u4e2d\u5b66\u4e60\u7b56\u7565\u6765\u89c4\u907f\u663e\u5f0f\u5956\u52b1\u5efa\u6a21\u3002 \u7ed9\u5b9a\u4e00\u4e2a\u504f\u597d\u4e09\u5143\u7ec4\u6570\u636e\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">{(x,y_w,y_l)}<\/span>\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">y_w<\/span> \u662f\u9996\u9009\u8f93\u51fa\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">y_l<\/span> \u662f\u9488\u5bf9\u63d0\u793a <span class=\"katex-eq\" data-katex-display=\"false\">x<\/span> \u7684\u4e0d\u592a\u9996\u9009\u7684\u8f93\u51fa\uff0cDPO \u6700\u5927\u5316\u89c2\u5bdf\u5230\u7684\u504f\u597d\u7684\u53ef\u80fd\u6027\u3002 \u5f62\u5f0f\u4e0a\uff0cDPO \u91c7\u7528\u4ee5\u4e0b\u76ee\u6807\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}{\\text{DPO}}(\\pi\\theta;\\pi_{\\text{ref}})&gt;[katex]\\mathcal{L}{\\text{DPO}}(\\pi\\theta;\\pi_{\\text{ref}})=-\\mathbb{E}{(x,y_w,y_l)\\sim\\mathcal{D}}!\\Bigl[\\log\\sigma!\\Bigl(\\beta,\\Bigl[\\log\\frac{\\pi\\theta(y_w\\mid x)}{\\pi_{\\text{ref}}(y_w\\mid x)}\\Bigr]-\\beta,\\Bigl[\\log\\frac{\\pi_\\theta(y_l\\mid x)}{\\pi_{\\text{ref}}(y_l\\mid x)}\\Bigr]\\Bigr)\\Bigr]<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\sigma(\\cdot)<\/span> \u662f logistic sigmoid \u51fd\u6570\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\beta\\log\\frac{\\pi_\\theta(y\\mid x)}{\\pi_{\\text{ref}}(y\\mid x)}<\/span> \u8868\u793a <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_\\theta<\/span> \u548c\u53c2\u8003\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span> \u4e4b\u95f4\u7684\u91cd\u65b0\u53c2\u6570\u5316\u7684\u5956\u52b1\u5dee\u5f02\u3002 \u901a\u8fc7\u6700\u5927\u5316 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}{\\text{DPO}}<\/span>\uff0c\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi\\theta<\/span> \u4e0e\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\uff0c\u800c\u65e0\u9700\u5355\u72ec\u7684\u5956\u52b1\u6a21\u578b\u3002 \u7531\u4e8e DPO \u76ee\u6807\u7ee7\u627f\u4e86 RLHF \u7684 KL \u6b63\u5219\u5316\u516c\u5f0f\uff0c\u5b83\u4fdd\u7559\u4e86\u57fa\u672c\u7684\u7406\u8bba\u4fdd\u8bc1\uff0c\u4f8b\u5982\u5728\u660e\u786e\u5b9a\u4e49\u7684\u504f\u597d\u5047\u8bbe\u4e0b\u7684\u4e00\u81f4\u6027\uff0c\u540c\u65f6\u5c06\u8bad\u7ec3\u8fc7\u7a0b\u7edf\u4e00\u4e3a\u5355\u4e2a\u9636\u6bb5\u3002 \u56e0\u6b64\uff0cDPO \u4fc3\u8fdb\u4e86\u5c06\u8bed\u8a00\u6a21\u578b\u4e0e\u4eba\u7c7b\u8bc4\u4f30\u5bf9\u9f50\u7684\u66f4\u76f4\u63a5\u7684\u9014\u5f84\uff0c\u964d\u4f4e\u4e86\u7cfb\u7edf\u590d\u6742\u6027\u5e76\u63d0\u9ad8\u4e86\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/p>\n<h5>DPO\u7684\u8bad\u7ec3\u7ec6\u8282<\/h5>\n<p>DPO \u6846\u67b6\u5efa\u7acb\u5728\u4e24\u4e2a\u6838\u5fc3\u6a21\u578b\u4e4b\u4e0a\uff1a\u53c2\u8003\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span> \u548c\u76ee\u6807\u7b56\u7565 <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{tar}}<\/span>\u3002 \u53c2\u8003\u7b56\u7565\u901a\u5e38\u662f\u9884\u8bad\u7ec3\u7684\u5e76\u7ecf\u8fc7\u76d1\u7763\u7684\u5fae\u8c03\u8bed\u8a00\u6a21\u578b\uff0c\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u4e0d\u53d8\u3002 \u76f8\u6bd4\u4e4b\u4e0b\uff0c\u76ee\u6807\u7b56\u7565\u4ece <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span> \u521d\u59cb\u5316\uff0c\u5e76\u4f7f\u7528\u57fa\u4e8e\u504f\u597d\u7684\u53cd\u9988\u8fdb\u884c\u8fed\u4ee3\u66f4\u65b0\uff0c\u4ece\u800c\u63d0\u9ad8\u4e0e\u4eba\u7c7b\u5224\u65ad\u7684\u5bf9\u9f50\u7a0b\u5ea6\u3002 \u56fe 11 \u63cf\u7ed8\u4e86\u6574\u4e2a\u6d41\u7a0b\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104100128281.png\" alt=\"\" \/><\/p>\n<p><center>\u76f4\u63a5\u504f\u597d\u4f18\u5316 (Direct Preference Optimization,DPO) \u7684\u5de5\u4f5c\u6d41\u7a0b\uff0c\u8bf4\u660e\u4e86\u57fa\u4e8e\u4eba\u7c7b\u504f\u597d\u4f18\u5316\u5927\u578b\u8bed\u8a00\u6a21\u578b\u8f93\u51fa\u7684\u8bad\u7ec3\u6d41\u7a0b<\/center><\/p>\n<ul>\n<li><strong>\u6570\u636e\u6536\u96c6\u548c\u51c6\u5907:<\/strong>DPO \u4f9d\u8d56\u4e8e\u901a\u8fc7\u4ece <span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span> \u4e2d\u4e3a\u6bcf\u4e2a\u63d0\u793a <span class=\"katex-eq\" data-katex-display=\"false\">x<\/span> \u91c7\u6837\u591a\u4e2a\u5019\u9009\u54cd\u5e94\u800c\u83b7\u5f97\u7684\u7cbe\u9009\u504f\u597d\u6570\u636e\u96c6\u3002 \u7136\u540e\uff0c\u4eba\u7c7b\u6807\u6ce8\u8005\u6839\u636e\u8fde\u8d2f\u6027\u3001\u76f8\u5173\u6027\u548c\u6e05\u6670\u5ea6\u7b49\u6807\u51c6\u6bd4\u8f83\u6216\u5bf9\u8fd9\u4e9b\u54cd\u5e94\u8fdb\u884c\u6392\u540d\u3002 \u4ea7\u751f\u7684\u504f\u597d\u6807\u7b7e\u4f5c\u4e3a\u4f18\u5316<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{tar}}<\/span>\u7684\u6838\u5fc3\u8bad\u7ec3\u4fe1\u53f7\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u8fc7\u7a0b:<\/strong>\u76ee\u6807\u7b56\u7565\u901a\u8fc7\u4e00\u7cfb\u5217\u57fa\u4e8e\u68af\u5ea6\u7684\u66f4\u65b0\u6765\u4f18\u5316\uff0c\u65e8\u5728\u6700\u5c0f\u5316\u635f\u5931<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}{\\text{DPO}}<\/span>\u3002 \u5177\u4f53\u6765\u8bf4\uff0c 1) \u751f\u6210\uff1a <span class=\"katex-eq\" data-katex-display=\"false\">\\pi{\\text{ref}}<\/span>\u4e3a\u6bcf\u4e2a\u63d0\u793a<span class=\"katex-eq\" data-katex-display=\"false\">x<\/span>\u4ea7\u751f\u5019\u9009\u8f93\u51fa\u3002 2) \u6807\u6ce8\uff1a\u4eba\u7c7b\u6807\u6ce8\u8005\u6bd4\u8f83\u751f\u6210\u7684\u8f93\u51fa\uff0c\u786e\u5b9a\u5b83\u4eec\u7684\u76f8\u5bf9\u504f\u597d\u3002 3) \u4f18\u5316\uff1a\u4f7f\u7528\u8fd9\u4e9b\u6210\u5bf9\u7684\u504f\u597d\uff0c\u8fed\u4ee3\u66f4\u65b0<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{tar}}<\/span>\u4ee5\u66f4\u597d\u5730\u6a21\u62df\u4eba\u7c7b\u504f\u597d\u7684\u8f93\u51fa\u3002 \u5728\u6574\u4e2a\u8fc7\u7a0b\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span>\u4fdd\u6301\u4e0d\u53d8\uff0c\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7a33\u5b9a\u7684\u57fa\u7ebf\uff0c\u7528\u4e8e\u8861\u91cf\u6539\u8fdb\u3002<\/li>\n<li><strong>\u5b9e\u9645\u8003\u8651\u56e0\u7d20:<\/strong>\u9009\u62e9\u4e00\u4e2a\u7a33\u5065\u7684\u53c2\u8003\u7b56\u7565\u901a\u5e38\u5bf9\u4e8e\u6709\u6548\u521d\u59cb\u5316DPO\u81f3\u5173\u91cd\u8981\u3002 SFT\u901a\u5e38\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_{\\text{ref}}<\/span>\u4ea7\u751f\u4e00\u4e2a\u8868\u73b0\u826f\u597d\u7684\u57fa\u7ebf\uff0c\u786e\u4fdd\u540e\u7eed\u7684\u504f\u597d\u9a71\u52a8\u66f4\u65b0\u53ef\u4ee5\u4e13\u6ce8\u4e8e\u6539\u8fdb\uff0c\u800c\u4e0d\u662f\u57fa\u672c\u7684\u6280\u80fd\u4e60\u5f97\u3002 \u6b64\u5916\uff0c\u504f\u597d\u6570\u636e\u5fc5\u987b\u8db3\u591f\u591a\u6837\u5316\uff0c\u4ee5\u6355\u6349\u7528\u6237\u671f\u671b\u7684\u5dee\u5f02\uff0c\u4ece\u800c\u4fc3\u8fdb\u6a21\u578b\u9002\u5e94\u6027\u5e76\u9632\u6b62\u5bf9\u72ed\u4e49\u5b9a\u4e49\u7684\u4efb\u52a1\u7684\u8fc7\u62df\u5408\u3002<\/li>\n<\/ul>\n<h5>DPO\u7684\u53d8\u4f53<\/h5>\n<p>\u5df2\u7ecf\u51fa\u73b0\u4e86DPO\u7684\u591a\u4e2a\u53d8\u4f53\uff0c\u4ee5\u89e3\u51b3\u7279\u5b9a\u7684\u5bf9\u9f50\u6311\u6218\u5e76\u4f18\u5316\u6587\u672c\u751f\u6210\u7684\u4e0d\u540c\u65b9\u9762\u3002 \u8868 2 \u5305\u542b\u4e86\u8fd9\u4e9b\u65b9\u6cd5\u7684\u6982\u8ff0\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u6db5\u76d6\u4e86\u4ecetoken\u7ea7\u522b\u7684\u751f\u6210\u4f18\u5316\u5230\u63a7\u5236\u5197\u4f59\u548c\u5904\u7406\u5217\u8868\u5f0f\u6216\u8d1f\u9762\u504f\u597d\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251101102532267.jpg\" alt=\"\" \/><\/p>\n<ul>\n<li><strong>\u7528\u4e8e\u4f18\u5316\u751f\u6210\u7684DPO\uff1a<\/strong>token\u7ea7\u522b\u548c\u8fed\u4ee3DPO\u7b56\u7565\u6709\u52a9\u4e8e\u4e0e\u4eba\u7c7b\u504f\u597d\u8fdb\u884c\u66f4\u7ec6\u7c92\u5ea6\u6216\u8fde\u7eed\u7684\u5bf9\u9f50\u3002 \u91cd\u65b0\u6784\u5efa\u4e3a\u8d4c\u535a\u673a\u95ee\u9898\u540e\uff0cToken \u7ea7 DPO \u91c7\u7528\u7531 <span class=\"katex-eq\" data-katex-display=\"false\">(S,A,f,r,\/rou_0)<\/span> \u5b9a\u4e49\u7684\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b (MDP)\u3002 \u8fd9\u79cd\u65b9\u6cd5\u51cf\u8f7b\u4e86\u8bf8\u5982\u5bf9\u4e8e\u4e0d\u63a8\u8350\u7684 Token \u8fc7\u5ea6 KL \u6563\u5ea6\u7b49\u6311\u6218\u3002 TDPO \u5e94\u7528\u987a\u5e8f\u6b63\u5411 KL \u6563\u5ea6\u800c\u4e0d\u662f\u53cd\u5411 KL \u6563\u5ea6\uff0c\u6539\u5584\u4e86\u6587\u672c\u751f\u6210\u4e2d\u7684\u5bf9\u9f50\u548c\u591a\u6837\u6027\u4fdd\u6301\u3002 \u8fed\u4ee3 DPO \u91c7\u7528\u591a\u8f6e\u65b9\u6cd5\uff0c\u901a\u8fc7\u91cd\u590d\u7684\u504f\u597d\u8bc4\u4f30\uff08\u901a\u5e38\u7531\u6a21\u578b\u672c\u8eab\u6267\u884c\uff09\u6765\u6301\u7eed\u4f18\u5316\u8f93\u51fa\u3002 \u6210\u5bf9\u5410\u69fd\u4f18\u5316 (PCO) \u5c06\u4e8c\u5143\u53cd\u9988\u6269\u5c55\u5230\u6210\u5bf9\u8bbe\u7f6e\uff0c\u4f7f\u7528\u8f6f\u95f4\u9694\u6765\u5e73\u8861\u63a2\u7d22\u548c\u5229\u7528\u3002 \u9010\u6b65 DPO \u5212\u5206\u504f\u597d\u6570\u636e\u96c6\u5e76\u5e94\u7528\u8fed\u4ee3\u66f4\u65b0\uff0c\u4f7f\u7528\u6bcf\u8f6e\u66f4\u65b0\u7684\u7b56\u7565\u4f5c\u4e3a\u4e0b\u4e00\u8f6e\u7684\u57fa\u7ebf\u3002<\/li>\n<li><strong>\u53ef\u63a7\u4e14\u7075\u6d3b\u7684 DPO:<\/strong>\u4e00\u4e9b DPO \u53d8\u4f53\u65e8\u5728\u7ba1\u7406\u5197\u957f\u5e76\u51cf\u5c11\u5bf9\u56fa\u5b9a\u53c2\u8003\u7b56\u7565\u7684\u9700\u6c42\u3002 R-DPO \u901a\u8fc7\u76ee\u6807\u51fd\u6570\u4e2d\u7684\u6b63\u5219\u5316\u9879\u6765\u60e9\u7f5a\u8f93\u51fa\u957f\u5ea6\uff0c\u4ece\u800c\u89e3\u51b3\u8fc7\u4e8e\u5197\u957f\u6216\u5197\u4f59\u7684\u56de\u590d\u3002 SimPO \u901a\u8fc7\u89c4\u8303\u5316\u56de\u590d\u957f\u5ea6\u548c\u7b80\u5316\u635f\u5931\u51fd\u6570\u6765\u5904\u7406\u671f\u671b\u548c\u4e0d\u671f\u671b\u7684\u8f93\u51fa\uff0c\u4ece\u800c\u6d88\u9664\u4e86\u5bf9\u53c2\u8003\u7b56\u7565\u7684\u9700\u6c42\u3002 RLOO \u5229\u7528 REINFORCE \u7b97\u6cd5\uff0c\u65e0\u9700\u8bad\u7ec3\u4ef7\u503c\u6a21\u578b\uff0c\u4ece\u800c\u5927\u5927\u51cf\u5c11\u4e86\u8ba1\u7b97\u5f00\u9500\u3002 \u5b83\u5c06\u6574\u4e2a\u56de\u590d\u89c6\u4e3a\u5355\u4e2a\u52a8\u4f5c\uff0c\u5e76\u4ece\u7a00\u758f\u5956\u52b1\u4e2d\u5b66\u4e60\uff0c\u4e0e\u4f20\u7edf\u7684\u57fa\u4e8e PPO \u7684\u65b9\u6cd5\u76f8\u6bd4\uff0c\u7b80\u5316\u4e86\u5b9e\u73b0\u3002<\/li>\n<li><strong>\u5217\u8868\u7ea7 DPO:<\/strong>\u5217\u8868\u7ea7 DPO \u65b9\u6cd5\u4e0d\u662f\u5c06\u504f\u597d\u6570\u636e\u9650\u5236\u4e3a\u6210\u5bf9\u6bd4\u8f83\uff0c\u800c\u662f\u5bf9\u8f93\u51fa\u96c6\u8fdb\u884c\u4f18\u5316\u3002\u5217\u8868\u504f\u597d\u4f18\u5316 (LiPO) \u5c06 Learning-to-Rank \u6280\u672f\u76f4\u63a5\u5e94\u7528\u4e8e\u5019\u9009\u56de\u590d\u7684\u6392\u540d\u5217\u8868\uff0c\u63d0\u9ad8\u4e86\u76f8\u5bf9\u4e8e\u91cd\u590d\u6210\u5bf9\u6bd4\u8f83\u7684\u6548\u7387\u3002 RRHF \u5c06\u504f\u597d\u5bf9\u9f50\u7eb3\u5165 SFT\uff0c\u4ece\u800c\u6d88\u9664\u4e86\u5bf9\u5355\u72ec\u53c2\u8003\u6a21\u578b\u7684\u9700\u6c42\u3002 PRO \u5c06\u5217\u8868\u504f\u597d\u5206\u89e3\u4e3a\u66f4\u7b80\u5355\u7684\u4e8c\u5143\u4efb\u52a1\uff0c\u7b80\u5316\u4e86 SFT \u671f\u95f4\u7684\u5bf9\u9f50\u3002<\/li>\n<li><strong>Negative DPO:<\/strong>\u67d0\u4e9b\u4efb\u52a1\u9700\u8981\u4ece\u4e0d\u671f\u671b\u6216\u6709\u5bb3\u7684\u8f93\u51fa\u4e2d\u5b66\u4e60\uff1aNegating Negatives (NN) \u629b\u5f03\u6b63\u9762\u54cd\u5e94\uff0c\u5e76\u6700\u5927\u7a0b\u5ea6\u5730\u504f\u79bb\u4e0d\u592a\u53d7\u6b22\u8fce\u7684\u8f93\u51fa\u3002 Negative Preference Optimization (NPO) \u5728\u8d1f\u9762\u504f\u597d\u4e0a\u91c7\u7528\u68af\u5ea6\u4e0a\u5347\uff0c\u6709\u6548\u51cf\u5c11\u6709\u5bb3\u8f93\u51fa\u5e76\u7f13\u89e3\u707e\u96be\u6027\u5d29\u6e83\u3002<\/li>\n<\/ul>\n<h3>\u7528\u4e8e\u63a8\u7406\u7684PoLM<\/h3>\n<p>\u63a8\u7406\u662f\u4f7f LLM \u80fd\u591f\u5904\u7406\u6d89\u53ca\u591a\u6b65\u903b\u8f91\u3001\u590d\u6742\u63a8\u7406\u548c\u590d\u6742\u51b3\u7b56\u5236\u5b9a\u7684\u4efb\u52a1\u7684\u6838\u5fc3\u652f\u67f1\u3002\u672c\u7ae0\u7814\u7a76\u4e86\u589e\u5f3a\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u4e24\u79cd\u6838\u5fc3\u6280\u672f\uff1aSelf-Refine for Reasoning\uff0c\u5b83\u5f15\u5bfc\u6a21\u578b\u81ea\u4e3b\u68c0\u6d4b\u548c\u7ea0\u6b63\u5176\u81ea\u8eab\u63a8\u7406\u6b65\u9aa4\u4e2d\u7684\u9519\u8bef\uff1b\u4ee5\u53ca Reinforcement Learning for Reasoning\uff0c\u5b83\u91c7\u7528\u57fa\u4e8e\u5956\u52b1\u7684\u4f18\u5316\u6765\u63d0\u9ad8\u6a21\u578b\u601d\u7ef4\u94fe\u7684\u4e00\u81f4\u6027\u548c\u6df1\u5ea6\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5171\u540c\u4f7f\u6a21\u578b\u80fd\u591f\u66f4\u7a33\u5065\u5730\u5904\u7406\u957f\u65f6\u51b3\u7b56\u5236\u5b9a\u3001\u903b\u8f91\u8bc1\u660e\u3001\u6570\u5b66\u63a8\u7406\u548c\u5176\u4ed6\u5177\u6709\u6311\u6218\u6027\u7684\u4efb\u52a1\u3002<\/p>\n<h4>Self-Refine for Reasoning<\/h4>\n<p>\u63a8\u7406\u4ecd\u7136\u662f\u4f18\u5316 LLM \u4ee5\u6267\u884c\u9700\u8981\u590d\u6742\u903b\u8f91\u63a8\u7406\u548c\u4e0a\u4e0b\u6587\u76f8\u5173\u51b3\u7b56\u5236\u5b9a\u7684\u4efb\u52a1\u7684\u6838\u5fc3\u6311\u6218\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0cself-refine \u6210\u4e3a\u4e00\u79cd\u5f3a\u5927\u7684\u673a\u5236\uff0c\u53ef\u4ee5\u5728\u6587\u672c\u751f\u6210\u671f\u95f4\u6216\u4e4b\u540e\u8fed\u4ee3\u5730\u67e5\u660e\u5e76\u7ea0\u6b63\u9519\u8bef\uff0c\u4ece\u800c\u5927\u5927\u63d0\u9ad8\u63a8\u7406\u6df1\u5ea6\u548c\u6574\u4f53\u53ef\u9760\u6027\u3002 \u5982\u56fe\uff0cself-refine \u65b9\u6cd5\u53ef\u5206\u4e3a\u56db\u7c7b\uff1aIntrinsic Self-refine\uff0c\u5b83\u4f9d\u8d56\u4e8e\u6a21\u578b\u7684\u5185\u90e8\u63a8\u7406\u5faa\u73af\uff1bExternal Self-refine\uff0c\u5b83\u7ed3\u5408\u4e86\u5916\u90e8\u53cd\u9988\u8d44\u6e90\uff1bFine-tuned Intrinsic Self-refine\uff0c\u5b83\u6839\u636e\u81ea\u751f\u6210\u7684\u66f4\u6b63\u8fed\u4ee3\u66f4\u65b0\u6a21\u578b\u7684\u63a8\u7406\u8fc7\u7a0b\uff1b\u4ee5\u53ca Fine-tuned External Self-refine\uff0c\u5b83\u5229\u7528\u5916\u90e8\u4fe1\u53f7\u548c\u5fae\u8c03\uff0c\u4ee5\u66f4\u5177\u9002\u5e94\u6027\u7684\u957f\u671f\u65b9\u5f0f\u5b8c\u5584\u63a8\u7406\u3002 \u8868 4 \u8fdb\u4e00\u6b65\u8bf4\u660e\u4e86\u6bcf\u4e2a\u7c7b\u522b\u5982\u4f55\u52a0\u5f3a LLM \u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u7684\u63a8\u7406\u80fd\u529b\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104112149489.png\" alt=\"\" \/><\/p>\n<p><center>Self-Refine \u65b9\u6cd5\u7684\u5206\u7c7b\uff0c\u63cf\u8ff0\u4e86\u7528\u4e8e\u589e\u5f3a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7684\u67b6\u6784\u53d8\u4f53<\/center><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104112323725.jpg\" alt=\"\" \/><\/p>\n<p><center>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u81ea\u6211\u4fee\u6b63\u65b9\u6cd5\u6982\u8ff0\u3002 \u672c\u8868\u603b\u7ed3\u4e86\u7a81\u51fa\u7684\u81ea\u6211\u4fee\u6b63\u6280\u672f\uff0c\u8be6\u7ec6\u8bf4\u660e\u4e86\u5b83\u4eec\u7684\u4e3b\u8981 LLM\u3001\u4efb\u52a1\u4ee5\u53ca\u8de8\u8d8a\u4e09\u4e2a\u6307\u6807\u7684\u53d1\u5e03\u65f6\u95f4\u7ebf\uff1aET\uff08\u5916\u90e8\u5de5\u5177\uff09\uff0cFT\uff08\u5fae\u8c03\uff09\uff0c\u4ee5\u53caSR\uff08\u81ea\u6211\u4fee\u6b63\uff09<\/center><\/p>\n<ul>\n<li><strong>\u5185\u6e90\u6027\u81ea\u6211\u4fee\u6b63(Intrinsic Self-Refine)<\/strong>\u3002 \u5185\u6e90\u6027\u81ea\u6211\u4fee\u6b63\u65b9\u6cd5\u4fa7\u91cd\u4e8e\u8d4b\u4e88\u6a21\u578b\u81ea\u8eab\u5728\u5185\u90e8\u68c0\u6d4b\u548c\u4fee\u590d\u9519\u8bef\u7684\u80fd\u529b\uff0c\u800c\u65e0\u9700\u501f\u52a9\u5916\u90e8\u5de5\u5177\u3002 \u4f8b\u5982\uff0cRCI Prompting \u4ec5\u5728\u8bc6\u522b\u51fa\u77db\u76fe\u6216\u9519\u8bef\u65f6\u624d\u4f1a\u89e6\u53d1\u4fee\u6b63\uff0c\u907f\u514d\u5bf9\u5fae\u5c0f\u7684\u4e0d\u786e\u5b9a\u6027\u505a\u51fa\u8fc7\u5ea6\u53cd\u5e94\u3002 CAI Revisions \u4fee\u6b63\u4e0d\u826f\u8f93\u51fa\uff08\u4f8b\u5982\uff0c\u653b\u51fb\u6027\u6587\u672c\uff09\uff0c\u540c\u65f6\u6559\u5bfc\u6a21\u578b\u81ea\u6211\u8c03\u8282\u5176\u56de\u590d\u3002\u7c7b\u4f3c\u5730\uff0cSelf-Refine \u5229\u7528\u4ece\u4f4e\u8d28\u91cf\u63d0\u793a\u5230\u9ad8\u4fdd\u771f\u6307\u4ee4\u7684\u8fc7\u6e21\uff0c\u5b8c\u5584\u4e2d\u95f4\u903b\u8f91\u4ee5\u63d0\u9ad8\u4e00\u81f4\u6027\u3002CoVe \u901a\u8fc7\u5c06\u591a\u7b54\u6848\u95ee\u9898\u5206\u89e3\u4e3a\u5b50\u4efb\u52a1\u6765\u89e3\u51b3\uff0c\u6bcf\u4e2a\u5b50\u4efb\u52a1\u5355\u72ec\u9a8c\u8bc1\u4ee5\u786e\u4fdd\u6574\u4e2a\u63a8\u7406\u94fe\u7684\u51c6\u786e\u6027\u548c\u4e00\u81f4\u6027\u3002 Weak-to-Strong Generalization (W2SG) \u65b9\u6cd5\u5229\u7528\u9ad8\u7ea7\u7b97\u6cd5\uff0c\u4f7f\u5f3a\u5927\u7684\u5b66\u751f\u6a21\u578b\u80fd\u591f\u4ece\u80fd\u529b\u8f83\u5f31\u7684\u6559\u5e08\u6a21\u578b\u4ea7\u751f\u7684\u5608\u6742\u6f14\u793a\u4e2d\u6709\u6548\u5b66\u4e60\u3002 \u8be5\u6846\u67b6\u5df2\u5728\u4e0d\u540c\u9886\u57df\u4e2d\u770b\u5230\u4e86\u4e00\u4e9b\u5173\u952e\u7684\u8fdb\u5c55\u548c\u5e94\u7528\u3002 \u6700\u8fd1\u7684\u7814\u7a76\u901a\u8fc7\u5404\u79cd\u521b\u65b0\u589e\u5f3a\u4e86 W2SG\u3002 \u4f8b\u5982\uff0c\u96c6\u6210\u5b66\u4e60\u6280\u672f\u5df2\u6210\u529f\u5e94\u7528\u4e8e\u63d0\u9ad8 W2SG \u65b9\u6cd5\u7684\u9c81\u68d2\u6027\u548c\u6709\u6548\u6027\u3002\u8fd8\u6709\u7814\u7a76\u91c7\u7528\u5f31\u5230\u5f3a\u7684\u5916\u63a8\u6765\u589e\u5f3a LLMs \u7684\u5bf9\u9f50\u3002<\/li>\n<li><strong>\u5916\u6e90\u6027\u81ea\u6211\u4fee\u6b63(External Self-Refine):<\/strong> \u8fd9\u4e9b\u65b9\u6cd5\u6d89\u53ca\u5916\u90e8\u53cd\u9988\u6765\u6e90\u6216\u8ba1\u7b97\u5de5\u5177\u6765\u6307\u5bfc\u548c\u7ea0\u6b63\u6a21\u578b\u7684\u63a8\u7406\u3002 CRITIC \u7cfb\u7edf\u5730\u68c0\u67e5\u9010\u6b65\u8f93\u51fa\uff0c\u4ece\u800c\u589e\u5f3a\u590d\u6742\u63a8\u7406\u4efb\u52a1\u7684\u53ef\u9760\u6027\u3002Reflexion \u548c Self-Debug \u5206\u522b\u5c06\u751f\u6210\u7684\u7b54\u6848\u4e0e\u53c2\u8003\u89e3\u51b3\u65b9\u6848\u6216\u5c11\u6837\u672c\u793a\u4f8b\u8fdb\u884c\u6bd4\u8f83\uff0c\u8fed\u4ee3\u5730\u6539\u8fdb\u903b\u8f91\u3002FLARE \u548c Logic-LM \u7b49\u6280\u672f\u7ed3\u5408\u4e86\u6765\u81ea\u5916\u90e8\u6587\u6863\u6216\u7b26\u53f7\u6c42\u89e3\u5668\u7684\u53c2\u8003\uff0c\u4ece\u800c\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u903b\u8f91\u9519\u8bef\u3002 RARR \u548c SelfEvolve \u8868\u660e\uff0c\u9a8c\u8bc1\u4e2d\u95f4\u72b6\u6001\uff08\u4f8b\u5982\uff0c\u7f16\u8bd1\u5668\u6d88\u606f\u6216\u76f8\u5173\u77e5\u8bc6\u6765\u6e90\uff09\u662f\u5c3d\u65e9\u4fee\u526a\u9519\u8bef\u8def\u5f84\u5e76\u5c06\u6a21\u578b\u5b8c\u5584\u4e3a\u6b63\u786e\u89e3\u51b3\u65b9\u6848\u7684\u6709\u6548\u65b9\u6cd5\u3002\u7814\u7a76\u63d0\u51fa\u4e86\u4ece\u4eba\u7c7b\u53cd\u9988\u4e2d\u8fdb\u884c\u8fed\u4ee3\u504f\u597d\u5b66\u4e60\uff0c\u5176\u4e2d\u5305\u62ec\u7528\u4e8e\u5728\u7ebf\u8bbe\u7f6e\u7684\u76f4\u63a5\u504f\u597d\u4f18\u5316 (DPO) \u7b97\u6cd5\u7684\u8fed\u4ee3\u7248\u672c\uff0c\u4ee5\u53ca\u7528\u4e8e\u79bb\u7ebf\u573a\u666f\u7684\u591a\u6b65\u62d2\u7edd\u62bd\u6837\u7b56\u7565\u3002PIT \u9690\u5f0f\u5730\u4ece\u4eba\u7c7b\u504f\u597d\u6570\u636e\u4e2d\u5b66\u4e60\u6539\u8fdb\u76ee\u6807\u3002<\/li>\n<li><strong>\u5fae\u8c03\u7684\u5185\u5728\u81ea\u6211\u5b8c\u5584(Fine-Tuned Intrinsic Self-Refine)<\/strong>:\u901a\u8fc7\u4e13\u95e8\u9488\u5bf9\u5185\u90e8\u4fee\u8ba2\u5bf9\u57fa\u7840\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u7cfb\u7edf\u5730\u52a0\u5f3a\u4e86 LLM \u7684\u81ea\u6211\u4fee\u6b63\u5faa\u73af\u3002Self-Critique \u65e8\u5728\u901a\u8fc7\u81ea\u6211\u5ba1\u67e5\u6765\u6539\u8fdb\u6458\u8981\uff0c\u800c SelFee \u4f7f\u7528\u8fed\u4ee3\u53cd\u9988\u5faa\u73af\u6765\u786e\u4fdd\u66f4\u9ad8\u6c34\u5e73\u7684\u903b\u8f91\u4e00\u81f4\u6027\u3002 Volcano \u901a\u8fc7\u5728 LLM \u7684\u67b6\u6784\u5185\u5fae\u8c03\u4e00\u4e2a\u4e13\u95e8\u7684\u6821\u6b63\u5668\u6a21\u5757\u6765\u51cf\u5c11\u591a\u6a21\u6001\u5e7b\u89c9\uff0c\u800c RL4F \u5229\u7528\u57fa\u4e8e RL \u7684\u6279\u8bc4\u5faa\u73af\uff0c\u5728\u9700\u8981\u6df1\u5165\u63a8\u7406\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5c06\u6027\u80fd\u5e73\u5747\u63d0\u9ad8\u4e86 10%\u3002 REFINER \u540c\u6837\u4e13\u6ce8\u4e8e\u4e2d\u95f4\u63a8\u7406\u8def\u5f84\uff0c\u800c\u65e0\u9700\u66f4\u6539\u6a21\u578b\u7684\u539f\u59cb\u751f\u6210\u8fc7\u7a0b\uff0c\u8fd9\u8868\u660e\u53ef\u4ee5\u901a\u8fc7\u8bad\u7ec3\u6a21\u578b\u4ed4\u7ec6\u91cd\u65b0\u68c0\u67e5\u5176\u90e8\u5206\u8f93\u51fa\u6765\u5b9e\u73b0\u6301\u7eed\u6539\u8fdb\u3002 \u6b64\u5916\uff0c\u6613\u4e8e\u5230\u96be\u6cdb\u5316\u7684\u6982\u5ff5\u5df2\u7ecf\u6210\u4e3a W2SG \u7684\u4e00\u4e2a\u6709\u524d\u666f\u7684\u53d8\u4f53\uff0c\u5176\u4e2d\u6a21\u578b\u6700\u521d\u5728\u6613\u4e8e\u9a8c\u8bc1\u7684\u793a\u4f8b\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u7136\u540e\u518d\u5904\u7406\u66f4\u590d\u6742\u7684\u4efb\u52a1\u3002 \u8fd9\u79cd\u65b9\u6cd5\u7684\u4e00\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u5b9e\u73b0\u6d89\u53ca\u5728\u4eba\u7c7b\u53ef\u9a8c\u8bc1\u7684\u793a\u4f8b\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u5f3a\u5927\u7684\u5956\u52b1\u6a21\u578b\uff0c\u7136\u540e\u6307\u5bfc\u5728\u5177\u6709\u6311\u6218\u6027\u4efb\u52a1\u4e0a\u5bf9\u66f4\u5f3a\u5927\u7684\u6a21\u578b\u7684\u76d1\u7763\u3002\u6b64\u5916\uff0cW2SG \u7684\u6709\u6548\u6027\u8d85\u51fa\u4e86 LLM \u7684\u8303\u56f4\uff0c\u5e76\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u4e2d\u4e5f\u5f97\u5230\u4e86\u6210\u529f\u7684\u5e94\u7528\u3002<\/li>\n<li><strong>\u5fae\u8c03\u7684\u5916\u90e8\u81ea\u6211\u5b8c\u5584(Fine-Tuned External Self-Refine)<\/strong>:\u5728\u9700\u8981\u957f\u671f\u6539\u8fdb\u7684\u573a\u666f\u4e2d\uff0c\u6a21\u578b\u7684\u53c2\u6570\u901a\u8fc7\u5916\u90e8\u53cd\u9988\u673a\u5236\u8fdb\u884c\u66f4\u65b0\u3002\u4f8b\u5982\uff0cSelf-Edit \u57fa\u4e8e\u6267\u884c\u7ed3\u679c\u91cd\u65b0\u751f\u6210\u4ee3\u7801\u8f93\u51fa\uff0c\u4ece\u800c\u5b9e\u73b0\u6b63\u786e\u6027\u7684\u8fed\u4ee3\u6539\u8fdb\u3002Baldur \u901a\u8fc7\u6dfb\u52a0\u6216\u4fee\u6539\u4e0a\u4e0b\u6587\u6765\u52a0\u5f3a\u5b9a\u7406\u8bc1\u660e\uff0c\u800c CodeRL \u5219\u91c7\u7528\u57fa\u4e8e\u6d4b\u8bd5\u7684\u8bc4\u8bba\u5458\u6765\u9a8c\u8bc1\u7a0b\u5e8f\u5408\u6210\u4efb\u52a1\u4e2d\u7684\u529f\u80fd\u51c6\u786e\u6027\u3002\u603b\u800c\u8a00\u4e4b\uff0c\u8fd9\u4e9b\u6280\u672f\u8868\u660e\uff0c\u5c06\u5916\u90e8\u8d44\u6e90\u4e0e\u6709\u9488\u5bf9\u6027\u7684\u5fae\u8c03\u76f8\u7ed3\u5408\uff0c\u53ef\u4ee5\u4fc3\u8fdb\u6a21\u578b\u6574\u4f53\u63a8\u7406\u6027\u80fd\u7684\u53ef\u9760\u3001\u9010\u6b65\u63d0\u5347\u3002<\/li>\n<\/ul>\n<h4>\u7528\u4e8e\u63a8\u7406\u7684\u5f3a\u5316\u5b66\u4e60<\/h4>\n<p>\u5728\u4e0a\u4e00\u5c0f\u8282\u4e2d\uff0c\u6211\u4eec\u63a2\u7d22\u4e86self-refine\u65b9\u6cd5\uff0c\u8fd9\u662f\u4e00\u79cd\u901a\u8fc7\u5c40\u90e8\u8c03\u6574\u548c\u4f18\u5316\u6539\u5584LLM\u63a8\u7406\u7684\u5e7f\u6cdb\u4f7f\u7528\u65b9\u6cd5\u3002\u8fd9\u79cd\u6280\u672f\u901a\u5e38\u5e94\u7528\u4e8e\u5355\u6b65\u4efb\u52a1\u6216\u8f93\u51fa\u7cbe\u70bc\uff0c\u4f8b\u5982\u6587\u672c\u751f\u6210\u548c\u95ee\u7b54\uff0c\u4ece\u800c\u63d0\u4f9b\u5feb\u901f\u7684\u63a8\u7406\u6536\u76ca\u3002\u7136\u800c\uff0c\u5b83\u96be\u4ee5\u5904\u7406\u9700\u8981\u591a\u6b65\u903b\u8f91\u7684\u590d\u6742\u3001\u957f\u671f\u7684\u63a8\u7406\u4efb\u52a1\u3002 OpenAI \u7684 o1 \u7cfb\u5217\u7684\u53d1\u5e03\u7a81\u51fa\u4e86\u5f3a\u5316\u5b66\u4e60 (RL) \u4f5c\u4e3a\u4e00\u79cd\u5f3a\u5927\u7684\u66ff\u4ee3\u65b9\u6848\uff0c\u901a\u8fc7\u57fa\u4e8e\u5956\u52b1\u7684\u53cd\u9988\u6765\u7cbe\u70bc\u957f\u65f6\u95f4\u7684\u5185\u90e8 CoT\uff0c\u4ece\u800c\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\u4ee5\u8fdb\u884c\u9ad8\u7ea7\u63a8\u7406\u3002\u8fd9\u663e\u7740\u63d0\u9ad8\u4e86\u590d\u6742\u4efb\u52a1\uff08\u5982\u6570\u5b66\u8bc1\u660e\u548c\u6218\u7565\u89c4\u5212\uff09\u7684\u6027\u80fd\u3002o1 \u7684\u6210\u529f\u6fc0\u53d1\u4e86\u5bf9\u5927\u89c4\u6a21 RL \u7684\u7814\u7a76\uff0c\u5176\u4e2d QwQ-32B-Preview \u5728\u6570\u5b66\u548c\u7f16\u7a0b\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u800c DeepSeek-R1 \u5219\u4e0e o1 \u7684\u80fd\u529b\u76f8\u5339\u914d\u3002 \u672c\u5c0f\u8282\u8003\u5bdf\u4e86 RL \u5728\u589e\u5f3a\u63a8\u7406\u65b9\u9762\u7684\u4f5c\u7528\uff0c\u91cd\u70b9\u4ecb\u7ecd\u4e86 DeepSeek-R1 \u548c DeepSeek-R1-Zero\uff0c\u8fd9\u4e24\u4e2a\u9886\u5148\u7684\u5f00\u6e90\u6a21\u578b\u3002<\/p>\n<h5>\u5c06\u63a8\u7406\u8868\u8ff0\u4e3aMDP\uff08Markov Decision Process\uff09<\/h5>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u63a8\u7406\u53ef\u4ee5\u4f18\u96c5\u5730\u5efa\u6a21\u4e3a\u4e00\u4e2a\u987a\u5e8f\u51b3\u7b56\u8fc7\u7a0b\uff0c\u5176\u4e2d\u6a21\u578b\u8fed\u4ee3\u5730\u6784\u9020\u4e00\u7cfb\u5217\u4e2d\u95f4\u6b65\u9aa4 <span class=\"katex-eq\" data-katex-display=\"false\">a_1,a_2,\\dots,a_T<\/span> \u4ee5\u54cd\u5e94\u8f93\u5165\u67e5\u8be2 <span class=\"katex-eq\" data-katex-display=\"false\">x<\/span>\uff0c\u4ece\u800c\u4f18\u5316\u5230\u8fbe\u6b63\u786e\u6700\u7ec8\u7b54\u6848\u7684\u53ef\u80fd\u6027\u3002 \u8fd9\u79cd\u6982\u5ff5\u5316\u5c06\u63a8\u7406\u8f6c\u5316\u4e3a\u4e00\u4e2a\u7ed3\u6784\u5316\u6846\u67b6\uff0c\u9002\u7528\u4e8e\u5f3a\u5316\u5b66\u4e60 (RL)\uff0c\u7279\u522b\u662f\u901a\u8fc7\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b (MDP) \u7684\u89c6\u89d2\uff0c\u8868\u793a\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}=(\\mathcal{S},\\mathcal{A},P,R,\\gamma)<\/span>\u3002 MDP \u5c01\u88c5\u4e86\u72b6\u6001\u3001\u52a8\u4f5c\u3001\u8f6c\u79fb\u3001\u5956\u52b1\u548c\u65f6\u95f4\u6298\u6263\u7684\u52a8\u6001\u76f8\u4e92\u4f5c\u7528\uff0c\u4e3a\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\u4ee5\u5e94\u5bf9\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u63d0\u4f9b\u4e86\u575a\u5b9e\u7684\u6570\u5b66\u57fa\u7840\u3002 \u901a\u8fc7\u5c06\u63a8\u7406\u6784\u5efa\u4e3a\u4e00\u7cfb\u5217\u6df1\u601d\u719f\u8651\u7684\u9009\u62e9\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4f7f\u6a21\u578b\u80fd\u591f\u7cfb\u7edf\u5730\u63a2\u7d22\u548c\u5b8c\u5584\u5176\u903b\u8f91\u8def\u5f84\uff0c\u8fd9\u4e0e\u6e38\u620f\u6216\u673a\u5668\u4eba\u6280\u672f\u7b49\u9886\u57df\u7684\u51b3\u7b56\u5236\u5b9a\u76f8\u4f3c\uff0c\u4f46\u9002\u5e94\u4e86\u8bed\u8a00\u548c\u6982\u5ff5\u63a8\u7406\u7684\u72ec\u7279\u6311\u6218\u3002 \u6700\u7ec8\u76ee\u6807\u662f\u63a8\u5bfc\u51fa\u6700\u4f18\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*(a_t|s_t)<\/span>\uff0c\u6700\u5927\u5316\u671f\u671b\u7d2f\u79ef\u5956\u52b1\uff0c\u8868\u793a\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">J(\\theta)=\\mathbb{E}{\\pi\\theta}!\\left[\\sum_{t=1}^{T}\\gamma^t,R(s_t,a_t)\\right]<\/span>\uff0c\u5229\u7528\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6280\u672f\uff0c\u5982\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09\u6216\u4f18\u52bf\u6f14\u5458-\u8bc4\u8bba\u5458\uff08A2C\uff09\uff0c\u57fa\u4e8e\u73af\u5883\u53cd\u9988\u8fed\u4ee3\u589e\u5f3a\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<ul>\n<li>\u72b6\u6001\u7a7a\u95f4\uff08State Space\uff09\uff1a\u72b6\u6001\u7a7a\u95f4<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{S}<\/span>\u6784\u6210\u4e86\u8be5MDP\uff08\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\uff09\u7684\u9aa8\u5e72\uff0c\u5176\u4e2d\u6bcf\u4e2a\u72b6\u6001<span class=\"katex-eq\" data-katex-display=\"false\">s_t\\in\\mathcal{S}<\/span>\u4ee3\u8868\u65f6\u95f4\u6b65<span class=\"katex-eq\" data-katex-display=\"false\">t<\/span>\u7684\u5f53\u524d\u63a8\u7406\u8f68\u8ff9\uff0c\u662f\u63a8\u7406\u8fc7\u7a0b\u81f3\u5173\u91cd\u8981\u7684\u8bed\u8a00\u548c\u7ed3\u6784\u5143\u7d20\u7684\u4e30\u5bcc\u7ec4\u5408\u3002 \u5177\u4f53\u6765\u8bf4\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">s_t<\/span>\u5305\u62ec\u521d\u59cb\u67e5\u8be2<span class=\"katex-eq\" data-katex-display=\"false\">x<\/span>\uff0c\u5148\u524d\u7684\u63a8\u7406\u6b65\u9aa4\u5e8f\u5217<span class=\"katex-eq\" data-katex-display=\"false\">{a_1,\\dots,a_{t-1}}<\/span>\uff0c\u4ee5\u53ca\u4e00\u4e2a\u5185\u90e8\u8bb0\u5fc6\u8868\u793a\uff0c\u8be5\u8868\u793a\u7f16\u7801\u903b\u8f91\u4f9d\u8d56\u5173\u7cfb\u548c\u4e2d\u95f4\u7ed3\u8bba\uff0c\u4f8b\u5982\u90e8\u5206\u89e3\u51b3\u65b9\u6848\u6216\u63a8\u65ad\u7684\u5173\u7cfb\u3002 \u968f\u7740\u63a8\u7406\u7684\u5c55\u5f00\uff0c\u6b64\u72b6\u6001\u52a8\u6001\u6f14\u5316\uff0c\u901a\u8fc7\u6574\u5408\u901a\u8fc7\u751f\u6210\u7684\u6b65\u9aa4\u548c\u4ece\u4e0a\u4e0b\u6587\u7406\u89e3\u4e2d\u63d0\u70bc\u7684\u6f5c\u5728\u77e5\u8bc6\u8868\u8fbe\u7684\u663e\u5f0f\u8def\u5f84\uff0c\u53cd\u6620\u4e86\u601d\u7ef4\u7684\u8fdb\u7a0b\u3002 \u4f8b\u5982\uff0c\u5728\u6570\u5b66\u8bc1\u660e\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">s_t<\/span>\u53ef\u80fd\u5305\u62ec\u95ee\u9898\u9648\u8ff0\u3001\u5148\u524d\u5bfc\u51fa\u7684\u65b9\u7a0b\u5f0f\u4ee5\u53ca\u9002\u7528\u7684\u5b9a\u7406\u7684\u8bb0\u5fc6\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u5404\u4e2a\u6b65\u9aa4\u4e2d\u4fdd\u6301\u4e00\u81f4\u6027\u3002 \u8fd9\u79cd\u591a\u65b9\u9762\u7684\u72b6\u6001\u8868\u793a\u786e\u4fdd\u4e86LLM\u80fd\u591f\u81ea\u9002\u5e94\u5730\u8ddf\u8e2a\u5176\u63a8\u7406\u4e0a\u4e0b\u6587\uff0c\u8fd9\u662f\u89e3\u51b3\u9700\u8981\u6301\u7eed\u903b\u8f91\u8fde\u8d2f\u6027\u7684\u4efb\u52a1\uff08\u4f8b\u5982\u591a\u6b65\u95ee\u9898\u89e3\u51b3\u6216\u6587\u672c\u751f\u6210\u4e2d\u7684\u53d9\u4e8b\u8fde\u8d2f\u6027\uff09\u7684\u524d\u63d0\u6761\u4ef6\u3002<\/li>\n<li><strong>\u52a8\u4f5c\u7a7a\u95f4\uff08Action Space\uff09<\/strong>\uff1a\u52a8\u4f5c\u7a7a\u95f4<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{A}<\/span>\u5b9a\u4e49\u4e86\u6bcf\u4e00\u6b65\u53ef\u80fd\u7684\u51b3\u7b56\u8303\u56f4\uff0c\u5176\u4e2d\u4e00\u4e2a\u52a8\u4f5c<span class=\"katex-eq\" data-katex-display=\"false\">a_t\\in\\mathcal{A}<\/span>\u5bf9\u5e94\u4e8e\u9009\u62e9\u4e0b\u4e00\u4e2a\u63a8\u7406\u52a8\u4f5c\uff0c\u4e3a\u63a8\u8fdb\u63a8\u7406\u8fc7\u7a0b\u63d0\u4f9b\u4e86\u901a\u7528\u7684\u5de5\u5177\u5305\u3002 \u8fd9\u4e9b\u52a8\u4f5c\u53ef\u80fd\u5305\u62ec\u7528\u81ea\u7136\u8bed\u8a00\u751f\u6210token\u6216\u77ed\u8bed\u4ee5\u9610\u660e\u63a8\u7406\u7247\u6bb5\uff0c\u5e94\u7528\u9884\u5b9a\u4e49\u7684\u903b\u8f91\u6216\u6570\u5b66\u53d8\u6362\uff08\u4f8b\u5982\uff0c\u4ee3\u6570\u7b80\u5316\uff09\uff0c\u4ece\u77e5\u8bc6\u5e93\u4e2d\u9009\u62e9\u76f8\u5173\u7684\u5b9a\u7406\u6216\u89c4\u5219\u4ee5\u6269\u5c55\u63a8\u7406\u94fe\uff0c\u6216\u5728\u8fbe\u5230\u7ed3\u8bba\u6027\u7b54\u6848\u65f6\u505c\u6b62\u8be5\u8fc7\u7a0b\u3002 \u52a8\u4f5c\u7a7a\u95f4\u7684\u6027\u8d28\u56e0\u4efb\u52a1\u800c\u5f02\uff1a\u5b83\u53ef\u4ee5\u662f\u79bb\u6563\u7684\uff0c\u5982\u5728\u5f62\u5f0f\u8bc1\u660e\u4e2d\u4ece\u6709\u9650\u7684\u903b\u8f91\u89c4\u5219\u96c6\u5408\u4e2d\u9009\u62e9\uff0c\u6216\u8005\u53ef\u4ee5\u662f\u8fde\u7eed\u7684\uff0c\u5982\u5728\u5f00\u653e\u5f0f\u63a8\u7406\u573a\u666f\u4e2d\u751f\u6210\u81ea\u7531\u6587\u672c\uff0c\u53cd\u6620\u4e86LLM\u7684\u751f\u6210\u7075\u6d3b\u6027\u3002 \u8fd9\u79cd\u4e8c\u5143\u6027\u4f7f\u6a21\u578b\u80fd\u591f\u540c\u65f6\u5904\u7406\u7ed3\u6784\u5316\u9886\u57df\uff08\u5982\u7b26\u53f7\u903b\u8f91\uff09\u548c\u975e\u7ed3\u6784\u5316\u9886\u57df\uff08\u5982\u5e38\u8bc6\u63a8\u7406\uff09\uff0c\u6839\u636e\u4efb\u52a1\u7684\u8981\u6c42\u8c03\u6574\u5176\u7b56\u7565\uff0c\u540c\u65f6\u4fdd\u6301\u5176\u671d\u5411\u89e3\u51b3\u65b9\u6848\u7684\u8fde\u8d2f\u8f68\u8ff9\u3002<\/li>\n<li><strong>\u8f6c\u79fb\u51fd\u6570\uff08Transition Function\uff09\uff1a<\/strong>\u8f6c\u79fb\u52a8\u529b\u5b66\uff0c\u7531\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">P(s_{t+1}|s_t,a_t)<\/span>\u5c01\u88c5\uff0c\u63a7\u5236\u72b6\u6001\u968f\u6bcf\u4e2a\u52a8\u4f5c\u7684\u6f14\u53d8\uff0c\u63cf\u7ed8\u4e86MDP\u6846\u67b6\u5185\u63a8\u7406\u8f68\u8ff9\u7684\u8fdb\u7a0b\u3002 \u4e0e\u968f\u673a\u6027\u6e90\u4e8e\u5916\u90e8\u53d8\u91cf\uff08\u4f8b\u5982\uff0c\u73af\u5883\u566a\u58f0\uff09\u7684\u4f20\u7edfRL\u73af\u5883\u76f8\u6bd4\uff0cLLM\u4e2d\u7684\u63a8\u7406\u8f6c\u79fb\u4e3b\u8981\u5177\u6709\u786e\u5b9a\u6027\uff0c\u7531\u6a21\u578b\u7684\u81ea\u56de\u5f52\u8f93\u51fa\u6216\u7ed3\u6784\u5316\u63a8\u7406\u89c4\u5219\u9a71\u52a8\uff0c\u4f8b\u5982\u5728\u8bc1\u660e\u4e2d\u5e94\u7528\u6f14\u7ece\u6b65\u9aa4\u3002 \u7136\u800c\uff0c\u4e0d\u786e\u5b9a\u6027\u6e90\u4e8e\u56fa\u6709\u7684\u6a21\u578b\u9650\u5236\u2014\u2014\u4f8b\u5982\uff0c\u4e0d\u5b8c\u5584\u7684\u77e5\u8bc6\u3001\u6a21\u68f1\u4e24\u53ef\u7684\u4e2d\u95f4\u72b6\u6001\u6216\u6587\u672c\u751f\u6210\u4e2d\u7684\u6982\u7387\u62bd\u6837\u2014\u2014\u5f15\u5165\u4e86RL\u5fc5\u987b\u89e3\u51b3\u7684\u53d8\u5f02\u6027\u3002 \u5bf9\u4e8e\u81ea\u56de\u5f52\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\uff0c\u8f6c\u6362\u9075\u5faa\u53ef\u9884\u6d4b\u7684\u5e8f\u5217\u751f\u6210\u8fc7\u7a0b\uff0c\u4f46\u9519\u8bef\u7d2f\u79ef\u6216\u4e0d\u540c\u89e3\u91ca\u7684\u53ef\u80fd\u6027\u9700\u8981\u4e00\u4e2a\u7a33\u5065\u7684\u8bbe\u8ba1\u6765\u786e\u4fdd\u53ef\u9760\u6027\u3002 \u8fd9\u79cd\u786e\u5b9a\u6027\u4f46\u53c8\u5145\u6ee1\u4e0d\u786e\u5b9a\u6027\u7684\u52a8\u6001\u7a81\u663e\u4e86\u5bf9\u81ea\u9002\u5e94\u7b56\u7565\u7684\u9700\u6c42\uff0c\u8fd9\u4e9b\u7b56\u7565\u53ef\u4ee5\u5728\u5404\u79cd\u73af\u5883\u4e2d\u7a33\u5b9a\u63a8\u7406\uff0c\u4ece\u7cbe\u786e\u7684\u6570\u5b66\u63a8\u5bfc\u5230\u7ec6\u81f4\u7684\u53d9\u4e8b\u6784\u5efa\u3002<\/li>\n<li><strong>\u5956\u52b1\u51fd\u6570\uff08Reward Function\uff09\uff1a<\/strong>\u5956\u52b1\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">R(s_t,a_t)<\/span>\u5145\u5f53\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\uff08MDP\uff09\u7684\u8bc4\u4f30\u6838\u5fc3\uff0c\u4e3a\u6bcf\u4e2a\u63a8\u7406\u6b65\u9aa4\u7684\u8d28\u91cf\u63d0\u4f9b\u5173\u952e\u53cd\u9988\uff0c\u4ee5\u6307\u5bfc\u6a21\u578b\u7684\u5b66\u4e60\u8fc7\u7a0b\u3002 \u4e0e\u5177\u6709\u660e\u786e\u5956\u52b1\uff08\u4f8b\u5982\uff0c\u6e38\u620f\u4e2d\u7684\u5206\u6570\uff09\u7684\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u4efb\u52a1\u4e0d\u540c\uff0c\u63a8\u7406\u5956\u52b1\u5fc5\u987b\u7ecf\u8fc7\u4ed4\u7ec6\u8bbe\u8ba1\uff0c\u4ee5\u5e73\u8861\u7a00\u758f\u6027\u548c\u5bc6\u5ea6\uff0c\u53cd\u6620\u4efb\u52a1\u7684\u590d\u6742\u6027\u548c\u76ee\u6807\u3002 \u7a00\u758f\u5956\u52b1\uff0c\u4f8b\u5982\u4ec5\u5728\u8fbe\u5230\u6b63\u786e\u7684\u6700\u7ec8\u7b54\u6848\u65f6\u624d\u5206\u914d\u4e00\u4e2a\u503c\uff0c\u63d0\u4f9b\u4e86\u7b80\u5355\u6027\uff0c\u4f46\u53ef\u80fd\u4f1a\u5ef6\u8fdf\u591a\u6b65\u573a\u666f\u4e2d\u7684\u5b66\u4e60\uff0c\u800c\u5bc6\u96c6\u5956\u52b1\u5219\u8bc4\u4f30\u9010\u6b65\u6b63\u786e\u6027\u3001\u903b\u8f91\u6709\u6548\u6027\u6216\u4e0e\u4eba\u7c7b\u504f\u597d\u7684\u5bf9\u9f50\uff0c\u63d0\u4f9b\u7ec6\u81f4\u7684\u6307\u5bfc\u3002 \u8fd9\u79cd\u7075\u6d3b\u6027\u4f7f\u5956\u52b1\u51fd\u6570\u80fd\u591f\u9002\u5e94\u4e0d\u540c\u7684\u63a8\u7406\u9700\u6c42\u2014\u2014\u65e0\u8bba\u662f\u5956\u52b1\u8bc1\u660e\u4e2d\u6709\u6548\u63a8\u7406\u89c4\u5219\u7684\u5e94\u7528\uff0c\u8fd8\u662f\u53d9\u4e8b\u7247\u6bb5\u7684\u8fde\u8d2f\u6027\u2014\u2014\u786e\u4fdd\u6a21\u578b\u63a5\u6536\u5230\u6709\u610f\u4e49\u7684\u4fe1\u53f7\uff0c\u4ee5\u5728\u5176\u5373\u65f6\u548c\u6269\u5c55\u7684\u63a8\u7406\u8303\u56f4\u5185\u5b8c\u5584\u5176\u7b56\u7565\u3002<\/li>\n<li><strong>\u6298\u6263\u56e0\u5b50\uff08Discount Factor\uff09\uff1a<\/strong><span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span>\uff1a\u4e00\u4e2a\u6807\u91cf<span class=\"katex-eq\" data-katex-display=\"false\">\\gamma\\in[0,1]<\/span>\uff0c\u7528\u4e8e\u786e\u5b9a\u5373\u65f6\u5956\u52b1\u548c\u672a\u6765\u5956\u52b1\u4e4b\u95f4\u7684\u6743\u8861\u3002 \u8f83\u9ad8\u7684<span class=\"katex-eq\" data-katex-display=\"false\">\\gamma<\/span>\u9f13\u52b1\u591a\u6b65\u63a8\u7406\u4f18\u5316\uff0c\u4fc3\u8fdb\u6df1\u5ea6\u63a8\u7406\u94fe\u800c\u4e0d\u662f\u77ed\u671f\u542f\u53d1\u5f0f\u65b9\u6cd5\u3002 \u9274\u4e8e\u8fd9\u79cd MDP \u516c\u5f0f\uff0c\u76ee\u6807\u662f\u5b66\u4e60\u4e00\u4e2a\u6700\u4f18\u63a8\u7406\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi^*(a_t|s_t)<\/span>\uff0c\u4ee5\u6700\u5927\u5316\u9884\u671f\u7684\u7d2f\u79ef\u5956\u52b1\uff1a<\/li>\n<\/ul>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">J(\\theta)&gt;[katex]J(\\theta)=\\mathbb{E}{\\pi\\theta}!\\left[\\sum_{t=1}^{T}\\gamma^t,R(s_t,a_t)\\right].<\/span><\/center>\u8be5\u6846\u67b6\u652f\u6301\u5e94\u7528\u5f3a\u5316\u5b66\u4e60\u6280\u672f\uff0c\u4f8b\u5982\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09\u6216\u4f18\u52bf\u6f14\u5458-\u8bc4\u8bba\u5bb6\uff08A2C\uff09\uff0c\u901a\u8fc7\u57fa\u4e8e\u6765\u81ea\u63a8\u7406\u73af\u5883\u7684\u53cd\u9988\u8fed\u4ee3\u8c03\u6574\u7b56\u7565<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_\\theta<\/span>\u6765\u5b8c\u5584\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<h5>\u63a8\u7406\u5956\u52b1\u8bbe\u8ba1<\/h5>\n<p>\u4e0e\u5177\u6709\u660e\u786e\u5956\u52b1\uff08\u5982\u6e38\u620f\u5206\u6570\uff09\u7684\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u4e0d\u540c\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u63a8\u7406\u9700\u8981\u7ed3\u6784\u5316\u7684\u5956\u52b1\u8bbe\u8ba1\uff0c\u53cd\u6620\u6b63\u786e\u6027\u3001\u6548\u7387\u548c\u4fe1\u606f\u91cf\u3002 \u5e38\u89c1\u65b9\u6cd5\u5305\u62ec:<\/p>\n<ul>\n<li><strong>\u4e8c\u5143\u6b63\u786e\u6027\u5956\u52b1(binary correctness rewards):<\/strong>\u4e3a\u6b63\u786e\u7684\u6700\u7ec8\u7b54\u6848\u5206\u914d<span class=\"katex-eq\" data-katex-display=\"false\">r_T=1<\/span>\uff0c\u5426\u5219\u5206\u914d<span class=\"katex-eq\" data-katex-display=\"false\">r_T=0<\/span>\uff0c\u8fd9\u5f88\u7b80\u5355\uff0c\u4f46\u7531\u4e8e\u7a00\u758f\u53cd\u9988\u5f15\u5165\u4e86\u9ad8\u65b9\u5dee\uff1b<\/li>\n<li><strong>\u9010\u6b65\u51c6\u786e\u6027\u5956\u52b1(step-wise accuracy rewards):<\/strong>\u57fa\u4e8e\u63a8\u7406\u89c4\u5219\u6709\u6548\u6027\u6216\u4e2d\u95f4\u6b65\u9aa4\u4e00\u81f4\u6027\u7b49\u6307\u6807\u63d0\u4f9b\u589e\u91cf\u53cd\u9988\uff0c\u4ee5\u6307\u5bfc\u591a\u6b65\u63a8\u7406\uff1b<\/li>\n<li><strong>\u81ea\u6d3d\u6027\u5956\u52b1(self-consistency rewards):<\/strong>\u8861\u91cf\u591a\u4e2a\u63a8\u7406\u8def\u5f84\u7684\u7a33\u5b9a\u6027\uff0c\u5e76\u4e3a\u4e00\u81f4\u6027\u5206\u914d\u66f4\u9ad8\u7684\u5956\u52b1\u4ee5\u589e\u5f3a\u9c81\u68d2\u6027\uff1b<\/li>\n<li><strong>\u57fa\u4e8e\u504f\u597d\u7684\u5956\u52b1(preference-based rewards):<\/strong>\u6e90\u81ea RLHF \u6216 RLAIF\uff0c\u5176\u4e2d\u4e00\u4e2a\u6a21\u578b<span class=\"katex-eq\" data-katex-display=\"false\">r_\\phi(s_t,a_t)<\/span>\u5728\u4eba\u7c7b\u6216\u4eba\u5de5\u667a\u80fd\u53cd\u9988\u4e0a\u8bad\u7ec3\uff0c\u8bc4\u4f30\u63a8\u7406\u8d28\u91cf\uff0c\u4e3a\u590d\u6742\u4efb\u52a1\u63d0\u4f9b\u7ec6\u81f4\u7684\u6307\u5bfc\u3002<\/li>\n<\/ul>\n<h5>\u5728\u57fa\u7840\u6a21\u578b\u4e0a\u7684\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60<\/h5>\n<p>\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u5df2\u7ecf\u6210\u4e3a\u4e00\u79cd\u53d8\u9769\u6027\u7684\u540e\u8bad\u7ec3\u8303\u5f0f\uff0c\u7528\u4e8e\u589e\u5f3a LLM \u7684\u63a8\u7406\u80fd\u529b\uff0c\u5c06\u91cd\u70b9\u4ece\u4f20\u7edf\u7684 SFT \u8f6c\u79fb\u5230\u52a8\u6001\u7684\u3001\u81ea\u6211\u8fdb\u5316\u7684\u4f18\u5316\u7b56\u7565\u3002\u8fd9\u79cd\u65b9\u6cd5\u5229\u7528\u5e7f\u6cdb\u7684\u8ba1\u7b97\u6846\u67b6\u548c\u57fa\u4e8e\u8fed\u4ee3\u5956\u52b1\u7684\u53cd\u9988\uff0c\u76f4\u63a5\u4f18\u5316\u57fa\u7840\u6a21\u578b\uff0c\u7ed5\u8fc7\u4e86\u5bf9\u9884\u6807\u6ce8\u6570\u636e\u96c6\u7684\u9700\u6c42\uff0c\u5e76\u5b9e\u73b0\u4e86\u590d\u6742\u63a8\u7406\u6280\u80fd\u7684\u81ea\u4e3b\u53d1\u5c55\u3002 \u901a\u8fc7\u96c6\u6210\u5927\u89c4\u6a21 RL\uff0cLLM \u53ef\u4ee5\u89e3\u51b3\u590d\u6742\u7684\u3001\u591a\u6b65\u63a8\u7406\u4efb\u52a1\uff08\u4f8b\u5982\uff0c\u6570\u5b66\u95ee\u9898\u6c42\u89e3\u3001\u903b\u8f91\u63a8\u7406\u548c\u6218\u7565\u89c4\u5212\uff09\uff0c\u5176\u4e2d\u4f20\u7edf\u7684 SFT \u5e38\u5e38\u7531\u4e8e\u5176\u4f9d\u8d56\u4e8e\u9759\u6001\u7684\u3001\u4eba\u5de5\u7b56\u5212\u7684\u6570\u636e\u800c\u8868\u73b0\u4e0d\u4f73\u3002 DeepSeek-R1\u6a21\u578b\u4f53\u73b0\u4e86\u8fd9\u79cd\u8303\u5f0f\uff0c\u91c7\u7528\u5148\u8fdb\u7684RL\u6280\u672f\u6765\u5b9e\u73b0\u6700\u5148\u8fdb\u7684\u63a8\u7406\u6027\u80fd\uff0c\u540c\u65f6\u4f18\u5316\u8d44\u6e90\u6548\u7387\uff0c\u5982\u56fe\u6240\u793a\u3002 \u672c\u5c0f\u8282\u9610\u8ff0\u4e86\u652f\u6301 DeepSeek-R1 \u6210\u529f\u7684\u5173\u952e\u65b9\u6cd5\uff0c\u5305\u62ec\u65b0\u9896\u7684\u4f18\u5316\u7b97\u6cd5\u3001\u81ea\u9002\u5e94\u63a2\u7d22\u548c\u8f68\u8ff9\u7ba1\u7406\uff0c\u8fd9\u4e9b\u5171\u540c\u91cd\u65b0\u5b9a\u4e49\u4e86 RL \u9a71\u52a8\u7684\u63a8\u7406\u5728 LLM \u4e2d\u7684\u6f5c\u529b\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104120832636.png\" alt=\"\" \/><\/p>\n<p><center>DeepSeek-R1 \u4e2d\u7528\u4e8e\u63a8\u7406\u7684\u5f3a\u5316\u5b66\u4e60\u5de5\u4f5c\u6d41\u7a0b\uff0c\u8bf4\u660e\u4e86\u4f18\u5316\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e2d\u63a8\u7406\u80fd\u529b\u7684\u8fc7\u7a0b<\/center><\/p>\n<ul>\n<li><strong>\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316(Group Relative Policy Optimization):<\/strong>DeepSeek-R1-Zero \u6a21\u578b\u5229\u7528\u4e86\u4e00\u79cd\u590d\u6742\u7684\u8fd1\u7aef\u7b56\u7565\u4f18\u5316 (PPO) \u7684\u53d8\u4f53\uff0c\u79f0\u4e3a\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316 (GRPO)\uff0c\u4ee5\u51cf\u8f7b\u4f20\u7edf RL \u8bad\u7ec3 LLM \u56fa\u6709\u7684\u5de8\u5927\u8ba1\u7b97\u548c\u8d44\u6e90\u9700\u6c42\u3002 \u4e0e\u4f9d\u8d56\u5e7f\u6cdb\u7684\u8bc4\u8bba\u5458\u7f51\u7edc\u7684\u6807\u51c6 PPO \u4e0d\u540c\uff0cGRPO \u91c7\u7528\u57fa\u4e8e\u7ec4\u7684\u57fa\u7ebf\u4f30\u8ba1\u6765\u7b80\u5316\u4f18\u5316\u8fc7\u7a0b\uff0c\u663e\u8457\u51cf\u5c11\u8bad\u7ec3\u5f00\u9500\uff0c\u540c\u65f6\u4fdd\u6301\u7b56\u7565\u66f4\u65b0\u7684\u7a33\u5065\u6027\u3002 \u8fd9\u79cd\u6548\u7387\u4f7f\u5f97\u5927\u89c4\u6a21 RL \u53ef\u4ee5\u5728\u8d44\u6e90\u53d7\u9650\u7684\u7cfb\u7edf\u4e0a\u90e8\u7f72\uff0c\u4fc3\u8fdb\u4e86\u8de8\u6269\u5c55\u8f68\u8ff9\u7684\u63a8\u7406\u7b56\u7565\u7684\u8fed\u4ee3\u4f18\u5316\u3002 \u901a\u8fc7\u4f18\u5316\u53ef\u7ba1\u7406\u7684\u8ba1\u7b97\u8303\u56f4\u5185\u7684\u7b56\u7565\uff0cGRPO\u5c06DeepSeek-R1-Zero\u5b9a\u4f4d\u4e3a\u53ef\u6269\u5c55\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u4ee5\u589e\u5f3a\u63a8\u7406\u80fd\u529b\uff0c\u5982\u4e0a\u56fe\u6240\u793a\u3002GRPO\u5df2\u7ecf\u6210\u4e3a\u5f53\u4ee3RL\u9a71\u52a8\u63a8\u7406\u7814\u7a76\u7684\u57fa\u77f3\u3002<\/li>\n<li><strong>DeepSeek-R1-Zero: <\/strong>DeepSeek-R1-Zero \u5c55\u73b0\u4e86\u5927\u89c4\u6a21 RL \u63d0\u5347 LLM \u63a8\u7406\u7684\u53d8\u9769\u6027\u6f5c\u529b\uff0c\u65e0\u9700\u50cf SFT \u90a3\u6837\u4f5c\u4e3a\u521d\u59cb\u6b65\u9aa4\uff0c\u800c\u662f\u91c7\u7528\u4e86\u7eaf\u7cb9\u7684 RL \u9a71\u52a8\u7684\u81ea\u6211\u8fdb\u5316\u8303\u5f0f\u3002 \u8fd9\u79cd\u65b9\u6cd5\u4f7f\u6a21\u578b\u80fd\u591f\u901a\u8fc7\u5956\u52b1\u53cd\u9988\u8fed\u4ee3\u5730\u6539\u8fdb\u5176\u5185\u90e8 CoT\uff0c\u4ece\u800c\u81ea\u4e3b\u5730\u53d1\u5c55\u590d\u6742\u7684\u63a8\u7406\u6280\u80fd\uff0c\u7ed5\u8fc7\u4e86 SFT \u4e2d\u901a\u5e38\u9700\u8981\u7684\u9884\u6807\u6ce8\u6570\u636e\u96c6\u3002\u7ed3\u679c\u662f\u5728\u590d\u6742\u7684\u3001\u591a\u6b65\u63a8\u7406\u4efb\u52a1\uff08\u4f8b\u5982\uff0c\u6570\u5b66\u95ee\u9898\u6c42\u89e3\u548c\u903b\u8f91\u63a8\u5bfc\uff09\u4e2d\u6027\u80fd\u7684\u663e\u7740\u63d0\u9ad8\uff0c\u8bc1\u660e\u4e86 RL \u4ece\u57fa\u7840\u6a21\u578b\u4e2d\u89e3\u9501\u9ad8\u7ea7\u63a8\u7406\u80fd\u529b\u7684\u80fd\u529b\u3002DeepSeek-R1-Zero \u4f5c\u4e3a\u4e00\u4e2a\u6700\u5f3a\u5927\u7684\u5f00\u6e90\u63a8\u7406\u6a21\u578b\u4e4b\u4e00\uff0c\u5176\u6210\u529f\u7a81\u663e\u4e86\u51b7\u542f\u52a8 RL \u7b56\u7565\u7684\u53ef\u884c\u6027\uff0c\u4e3a\u4f20\u7edf\u7684\u8bad\u7ec3\u6d41\u7a0b\u63d0\u4f9b\u4e86\u4e00\u79cd\u8d44\u6e90\u9ad8\u6548\u7684\u66ff\u4ee3\u65b9\u6848\uff0c\u540c\u65f6\u5b9e\u73b0\u4e86\u4e0e\u6700\u5148\u8fdb\u57fa\u51c6\u6d4b\u8bd5\u6301\u5e73\u7684\u6c34\u5e73\u3002<\/li>\n<li><strong>\u9010\u6b65\u5956\u52b1\u5efa\u6a21(Stepwise Reward Modeling):<\/strong>\u4e3a\u4e86\u6307\u5bfc\u8f68\u8ff9<span class=\"katex-eq\" data-katex-display=\"false\">\\tau=(s_1,a_1,\\dots,s_T,a_T)<\/span>\u7684\u63a8\u7406\uff0cdeepSeek-r1\u91c7\u7528\u9010\u6b65\u5956\u52b1\u6a21\u578b<span class=\"katex-eq\" data-katex-display=\"false\">f_\\theta<\/span>\uff0c\u5728\u6bcf\u4e2a\u65f6\u95f4\u6b65\u4e2d\u63d0\u4f9b\u7c92\u5ea6\u53cd\u9988\uff0c\u5b9a\u4e49\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">r_t=f_\\theta(s_t,a_t\\mid\\mathcal{D}{\\text{reasoning}})<\/span>\uff0c\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}{\\text{reasoning}}<\/span>\u6784\u6210\u4e86\u5e26\u6709\u4eba\u7c7b\u4fe1\u606f\u7684COT\u5e8f\u5217\u7684\u9010\u6b65\u6b63\u786e\u6027\u6807\u7b7e\u3002 \u8fd9\u79cd\u5bc6\u96c6\u7684\u5956\u52b1\u7ed3\u6784\u4e0e\u7a00\u758f\u7684\u5e8f\u5217\u5956\u52b1\u5f62\u6210\u9c9c\u660e\u5bf9\u6bd4\uff0c\u901a\u8fc7\u5bf9\u4e2a\u4eba\u63a8\u7406\u6b65\u9aa4\u7684\u8d28\u91cf\u63d0\u4f9b\u53ef\u884c\u7684\u53ef\u884c\u89c1\u89e3\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u7cbe\u786e\u5730\u8c03\u6574\u5176\u7b56\u7565\u3002 \u901a\u8fc7\u5229\u7528\u4e13\u5bb6\u7b56\u5212\u7684\u6570\u636e\uff0c\u5956\u52b1\u6a21\u578b\u786e\u4fdd\u53cd\u9988\u4e0e\u4eba\u7c7b\u63a8\u7406\u6807\u51c6\u4fdd\u6301\u4e00\u81f4\uff0c\u4ece\u800c\u5728\u6269\u5c55\u7684\u63a8\u7406\u94fe\u4e2d\u57f9\u517b\u4e00\u81f4\u6027\u548c\u51c6\u786e\u6027\uff0c\u8fd9\u662f\u89e3\u51b3\u9700\u8981\u957f\u671f\u903b\u8f91\u5408\u6210\u7684\u4efb\u52a1\u7684\u5173\u952e\u7279\u5f81\u3002<\/li>\n<li><strong>\u81ea\u9002\u5e94\u63a2\u7d22(Adaptive Exploration):<\/strong>DeepSeek-R1 \u901a\u8fc7\u96c6\u6210\u5230\u5176\u76ee\u6807\u4e2d\u7684\u81ea\u9002\u5e94\u63a2\u7d22\u673a\u5236\u6765\u589e\u5f3a\u7b56\u7565\u4f18\u5316:<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}{\\text{PPO}^+}&gt;[katex]\\mathcal{L}{\\text{PPO}^+}=\\mathbb{E}\\tau!\\Bigl[\\min!\\Bigl(\\frac{\\pi_\\phi(a|s)}{\\pi_{\\text{old}}(a|s)},A_t,;\\text{clip}!\\Bigl(\\frac{\\pi_\\phi(a|s)}{\\pi_{\\text{old}}(a|s)},1-\\varepsilon,1+\\varepsilon\\Bigr)A_t\\Bigr)\\Bigr]+\\lambda_t,\\mathcal{H}!\\bigl(\\pi_\\phi(\\cdot|s)\\bigr),<\/span>\u5176\u4e2d\u71b5\u9879<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{H}<\/span>\u7531\u81ea\u9002\u5e94\u7cfb\u6570<span class=\"katex-eq\" data-katex-display=\"false\">\\lambda_t=\\alpha\\cdot\\exp!\\bigl(-\\beta\\cdot\\text{Var}(R(\\tau_{1:t}))\\bigr)<\/span>\u8c03\u8282\uff0c\u6839\u636e\u6574\u4e2a\u8f68\u8ff9\u7684\u5956\u52b1\u65b9\u5dee\u52a8\u6001\u8c03\u6574\u3002 \u8fd9\u79cd\u65b9\u6cd5\u5e73\u8861\u4e86\u63a2\u7d22\u548c\u5229\u7528\uff0c\u9f13\u52b1\u6a21\u578b\u5728\u8bad\u7ec3\u65e9\u671f\u63a2\u7d22\u4e0d\u540c\u7684\u63a8\u7406\u8def\u5f84\uff0c\u540c\u65f6\u968f\u7740\u65b9\u5dee\u7684\u51cf\u5c0f\u800c\u6536\u655b\u5230\u6700\u4f73\u7b56\u7565\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u63a8\u7406\u7ec6\u5316\u7684\u9c81\u68d2\u6027\u548c\u6548\u7387\u3002<\/li>\n<li><strong>\u8f68\u8ff9\u526a\u679d(Trajectory Pruning):<\/strong>\u4e3a\u4e86\u4f18\u5316\u63a8\u7406\u671f\u95f4\u7684\u8ba1\u7b97\u6548\u7387\uff0cDeepSeek-R1 \u5b9e\u73b0\u4e86\u4e00\u4e2a\u53cc\u91cd\u6ce8\u610f\u8bc4\u8bba\u5458\uff0c\u5b83\u901a\u8fc7\u7ed3\u5408\u5c40\u90e8\u6b65\u9aa4\u8bc4\u4f30\u548c\u5168\u5c40\u8f68\u8ff9\u4e0a\u4e0b\u6587\u6765\u8bc4\u4f30\u6bcf\u4e2a\u72b6\u6001\u7684\u4ef7\u503c\u3002 \u5f53\u53d1\u751f\u526a\u679d\u65f6\uff0c\u7ec8\u6b62\u4f4e\u4ef7\u503c\u7684\u63a8\u7406\u8def\u5f84\uff0c\u5c06\u8d44\u6e90\u96c6\u4e2d\u5728\u6709\u5e0c\u671b\u7684\u8f68\u8ff9\u4e0a\u3002 \u8fd9\u79cd\u673a\u5236\u51cf\u5c11\u4e86\u65e0\u7528\u7684\u63a2\u7d22\uff0c\u52a0\u901f\u4e86\u6536\u655b\uff0c\u5e76\u786e\u4fdd\u6a21\u578b\u4f18\u5148\u8003\u8651\u9ad8\u8d28\u91cf\u7684\u63a8\u7406\u5e8f\u5217\uff0c\u8fd9\u6709\u52a9\u4e8e\u5176\u5728\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u4e2d\u53d6\u5f97\u5353\u8d8a\u7684\u6027\u80fd\u3002<\/li>\n<\/ul>\n<h5>\u4f7f\u7528\u51b7\u542f\u52a8\u7684\u63a8\u7406 RL<\/h5>\n<p>DeepSeek-R1-Zero \u901a\u8fc7\u91c7\u7528\u51b7\u542f\u52a8\u65b9\u6cd5\u8fdb\u4e00\u6b65\u63a8\u8fdb\u4e86 RL \u7684\u5e94\u7528\uff0c\u907f\u514d\u4e86 SFT\uff0c\u5b8c\u5168\u4f9d\u8d56\u4e8e\u6765\u81ea\u672a\u7ecf\u8bad\u7ec3\u7684\u57fa\u7840\u6a21\u578b\u7684\u5927\u89c4\u6a21 RL\u3002\u8fd9\u79cd\u81ea\u6211\u8fdb\u5316\u7684\u7b56\u7565\u901a\u8fc7\u8fed\u4ee3\u53cd\u9988\u6765\u5b8c\u5584\u63a8\u7406\uff0c\u4ece\u800c\u5728\u6ca1\u6709\u9884\u901a\u91cf\u7684\u6570\u636e\u4f9d\u8d56\u6027\u7684\u60c5\u51b5\u4e0b\u751f\u6210\u5065\u58ee\u7684COT\u5e8f\u5217\u3002\u901a\u8fc7\u76f4\u63a5\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0cDeepSeek-R1-Zero \u5c55\u793a\u4e86 RL \u7684\u591a\u529f\u80fd\u6027\uff0c\u5b9e\u73b0\u4e86\u4e0e\u4f7f\u7528 SFT \u521d\u59cb\u5316\uff08\u4f8b\u5982\u5176 DeepSeek-R1 \u5bf9\u5e94\u6a21\u578b\uff09\u7684\u6a21\u578b\u76f8\u5f53\u6216\u66f4\u4f18\u7684\u6027\u80fd\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u51cf\u5c11\u4e86\u5bf9\u5927\u91cf\u6807\u6ce8\u6570\u636e\u96c6\u7684\u4f9d\u8d56\uff0c\u8fd8\u5c55\u793a\u4e86 RL \u81ea\u4e3b\u5f00\u53d1\u590d\u6742\u63a8\u7406\u80fd\u529b\u7684\u53ef\u80fd\u6027\uff0c\u4e3a\u672a\u6765\u7684 LLM \u5f00\u53d1\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u6269\u5c55\u7684\u8303\u4f8b\u3002 \u603b\u800c\u8a00\u4e4b\uff0cRL \u4e3a\u589e\u5f3a\u63a8\u7406\u63d0\u4f9b\u4e86\u4e00\u4e2a\u6709\u524d\u666f\u7684\u6846\u67b6\uff0c\u6709\u6548\u7684\u5956\u52b1\u8bbe\u8ba1\u3001\u7b56\u7565\u4f18\u5316\uff08\u4f8b\u5982\uff0cGRPO\uff09\u548c\u63a2\u7d22\u7b56\u7565\u4ecd\u7136\u81f3\u5173\u91cd\u8981\u3002 \u8fdb\u4e00\u6b65\u7684\u7814\u7a76\u53ef\u4ee5\u63a2\u7d22\u6df7\u5408\u65b9\u6cd5\uff0c\u6574\u5408\u6a21\u4eff\u5b66\u4e60\u6216\u81ea\u76d1\u7763\u76ee\u6807\uff0c\u4ee5\u8fdb\u4e00\u6b65\u5b8c\u5584\u8fd9\u4e9b\u80fd\u529b\uff0c\u5de9\u56fa RL \u5728\u63a8\u8fdb LLM \u63a8\u7406\u4e2d\u7684\u4f5c\u7528\u3002<\/p>\n<h3>\u63d0\u9ad8\u6548\u7387\u7684PoLM<\/h3>\n<p>\u5728\u524d\u9762\u7ae0\u8282\u8ba8\u8bba\u7684\u540e\u8bad\u7ec3\u4f18\u5316\u6280\u672f\u7684\u57fa\u7840\u4e0a\uff0c\u540e\u8bad\u7ec3\u6548\u7387\u4e13\u95e8\u9488\u5bf9 LLM \u5728\u521d\u59cb\u9884\u8bad\u7ec3\u4e4b\u540e\u7684\u8fd0\u884c\u6027\u80fd\u3002 \u4e3b\u8981\u76ee\u6807\u662f\u4f18\u5316\u5173\u952e\u7684\u90e8\u7f72\u6307\u6807\uff08\u4f8b\u5982\uff0c\u5904\u7406\u901f\u5ea6\u3001\u5185\u5b58\u4f7f\u7528\u548c\u8d44\u6e90\u6d88\u8017\uff09\uff0c\u4ece\u800c\u4f7f LLM \u80fd\u591f\u66f4\u5b9e\u9645\u5730\u5e94\u7528\u4e8e\u73b0\u5b9e\u4e16\u754c\u7684\u5e94\u7528\u7a0b\u5e8f\u3002 \u5b9e\u73b0\u540e\u8bad\u7ec3\u6548\u7387\u7684\u65b9\u6cd5\u4e3b\u8981\u5206\u4e3a\u4e09\u7c7b\uff1a<\/p>\n<ul>\n<li><strong>\u6a21\u578b\u538b\u7f29\uff08Model Compression\uff09\uff1a<\/strong>\u901a\u8fc7\u526a\u679d\u548c\u91cf\u5316\u7b49\u6280\u672f\u6765\u51cf\u5c11\u6574\u4f53\u8ba1\u7b97\u91cf\uff1b<\/li>\n<li><strong>\u53c2\u6570\u9ad8\u6548\u7684\u5fae\u8c03\uff08Parameter-Efficient Fine-Tuning\uff09\uff1a<\/strong>\u4ec5\u66f4\u65b0\u6a21\u578b\u53c2\u6570\u7684\u4e00\u5c0f\u90e8\u5206\u6216\u4f7f\u7528\u4e13\u95e8\u7684\u6a21\u5757\uff0c\u4ece\u800c\u6700\u5927\u9650\u5ea6\u5730\u964d\u4f4e\u91cd\u65b0\u8bad\u7ec3\u6210\u672c\u5e76\u52a0\u901f\u5bf9\u65b0\u4efb\u52a1\u7684\u9002\u5e94\uff1b<\/li>\n<li><strong>\u77e5\u8bc6\u84b8\u998f (Knowledge Distillation\uff09\uff1a<\/strong>\u5c06\u77e5\u8bc6\u4ece\u4e00\u4e2a\u66f4\u5927\u7684\u3001\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u8f6c\u79fb\u5230\u4e00\u4e2a\u66f4\u5c0f\u7684\u6a21\u578b\uff0c\u4f7f\u66f4\u5c0f\u7684\u6a21\u578b\u80fd\u591f\u5728\u51cf\u5c11\u8d44\u6e90\u9700\u6c42\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u76f8\u5f53\u7684\u6027\u80fd\u3002<\/li>\n<\/ul>\n<h4>\u6a21\u578b\u538b\u7f29<\/h4>\n<p>\u6a21\u578b\u538b\u7f29\u5305\u62ec\u4e00\u7ec4\u65e8\u5728\u51cf\u5c11 LLM \u7684\u5927\u5c0f\u548c\u8ba1\u7b97\u9700\u6c42\u7684\u7684\u6280\u672f\uff0c\u5176\u4e2d\u5305\u62ec\u540e\u8bad\u7ec3\u91cf\u5316\uff08post-training quantization\uff09\u3001\u53c2\u6570\u526a\u679d\uff08parameter pruning\uff09\u548c\u4f4e\u79e9\u8fd1\u4f3c\uff08low-rank approximation\uff09\u3002<\/p>\n<h5>\u540e\u8bad\u7ec3\u91cf\u5316<\/h5>\n<p>LLM \u7684\u4e00\u4e2a\u5173\u952e\u538b\u7f29\u65b9\u6cd5\u662f\u91cf\u5316\uff0c\u5b83\u5c06\u9ad8\u7cbe\u5ea6\u6570\u636e\u7c7b\u578b <span class=\"katex-eq\" data-katex-display=\"false\">X_H<\/span> (30 \u4f4d\u6d6e\u70b9\u6570) \u8f6c\u6362\u4e3a\u8f83\u4f4e\u7cbe\u5ea6\u683c\u5f0f <span class=\"katex-eq\" data-katex-display=\"false\">X_L<\/span> (8 \u4f4d\u6574\u6570)\u3002 \u8fd9\u79cd\u8f6c\u6362\u7684\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">X_L&gt;[katex]X_L=\\text{Round}!\\left(\\frac{\\text{absmax}(X_L)}{\\text{absmax}(X_H)},X_H\\right)=\\text{Round}!\\left(\\mathcal{K}\\cdot X_H\\right),<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{K}<\/span> \u4ee3\u8868\u91cf\u5316\u5e38\u6570\uff0cabsmax \u6307\u7684\u662f\u5143\u7d20\u7684\u7edd\u5bf9\u6700\u5927\u503c\u3002 \u51fd\u6570 Round \u5c06\u6d6e\u70b9\u6570\u8f6c\u6362\u4e3a\u6574\u6570\u3002 LLM \u91cf\u5316\u5305\u62ec\u540e\u8bad\u7ec3\u91cf\u5316 (post-training quantization,PTQ) \u548c\u91cf\u5316\u611f\u77e5\u8bad\u7ec3 (quantization-aware training,QAT)\u3002 PTQ \u53ef\u4ee5\u5728\u9884\u8bad\u7ec3\u540e\u8c03\u6574\u6a21\u578b\u6743\u91cd\u548c\u6fc0\u6d3b\u503c\uff0c\u4f7f\u7528\u4e00\u4e2a\u5c0f\u578b\u7684\u6821\u51c6\u6570\u636e\u96c6\u6765\u4f18\u5316\u8ba1\u7b97\u6548\u7387\u548c\u6027\u80fd\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002 \u6b64\u5916\uff0c\u4e0b\u8868\u5c55\u793a\u4e86\u51e0\u79cd\u7528\u4e8eLLM\u7684\u7a81\u51fa\u91cf\u5316\u65b9\u6cd5\u7684\u6027\u80fd\u6307\u6807\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104123416864.png\" alt=\"\" \/><\/p>\n<p><center>LLM\u7684\u540e\u8bad\u7ec3\u91cf\u5316\u6280\u672f\u7684\u56fe\u89e3<\/center><\/p>\n<ul>\n<li><strong>\u4ec5\u6743\u91cd\u91cf\u5316 (Weight-Only Quantization,WOQ):<\/strong>WOQ \u4fa7\u91cd\u4e8e\u538b\u7f29\u6a21\u578b\u6743\u91cd\u4ee5\u63d0\u9ad8\u6548\u7387\u3002 GPTQ \u4f7f\u7528\u6700\u4f18\u8111\u91cf\u5316 (Optimal Brain Quantization\uff0cOBQ) \u5e94\u7528\u9010\u5c42\u91cf\u5316\uff0c\u5c06\u6743\u91cd\u964d\u4f4e\u52303\u62164\u4f4d\uff0c\u4ee5\u964d\u4f4e\u5185\u5b58\u4f7f\u7528\u91cf\u548c\u5904\u7406\u65f6\u95f4\u3002\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6548\u7387\uff0cQuIP \u5f15\u5165\u4e86\u9488\u5bf92\u4f4d\u91cf\u5316\u7684\u4e0d\u8fde\u8d2f\u5904\u7406\uff0c\u63d0\u4f9b\u4e86\u66f4\u7d27\u51d1\u7684\u8868\u793a\u5f62\u5f0f\u3002\u7c7b\u4f3c\u5730\uff0cAWQ \u548c OWQ \u901a\u8fc7\u5bf9\u7279\u522b\u654f\u611f\u7684\u6743\u91cd\u4fdd\u6301\u9ad8\u7cbe\u5ea6\u6765\u89e3\u51b3\u51c6\u786e\u6027\u4fdd\u6301\u95ee\u9898\uff0c\u4ece\u800c\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u63a8\u7406\u671f\u95f4\u6f5c\u5728\u7684\u7cbe\u5ea6\u635f\u5931\u3002\u6700\u540e\uff0cSpQR \u5c06\u7a00\u758f\u91cf\u5316\u4e0e\u89e3\u7801\u76f8\u7ed3\u5408\uff0c\u5b9e\u73b0\u4e86\u9ad8\u6548\u7684\u9010\u4e2atoken\u63a8\u7406\uff0c\u540c\u65f6\u4fdd\u7559\u4e86\u6a21\u578b\u7684\u54cd\u5e94\u80fd\u529b\u3002<\/li>\n<li><strong>\u6743\u91cd-\u6fc0\u6d3b\u8054\u5408\u91cf\u5316 (Weight-Activation Co-Quantization\uff0cWAQ)\uff1a<\/strong>WAQ \u6574\u5408\u4e86\u6743\u91cd\u548c\u6fc0\u6d3b\u503c\u4ee5\u63d0\u9ad8\u6548\u7387\u3002 LLM.int8() \u4f7f\u7528\u7cbe\u786e\u7684\u5b58\u50a8\u6765\u5904\u7406\u6fc0\u6d3b\u5f02\u5e38\u503c\uff0c\u5e76\u91cf\u5316\u52308\u4f4d\uff0c\u540c\u65f6\u4fdd\u6301\u6027\u80fd\u3002 SmoothQuant \u5b9e\u73b0\u4e86\u9010\u901a\u9053\u7f29\u653e\uff0c\u5c06\u91cf\u5316\u96be\u70b9\u4ece\u6fc0\u6d3b\u8f6c\u79fb\u5230\u6743\u91cd\uff0c\u4ee5\u5b9e\u73b0\u65e0\u635f\u7ed3\u679c\u3002 \u6b64\u5916\uff0cOS+ \u901a\u8fc7\u901a\u9053\u7ea7\u522b\u7684\u79fb\u4f4d\u548c\u7f29\u653e\u6765\u7f13\u89e3\u5f02\u5e38\u503c\u7684\u5f71\u54cd\uff0c\u4ece\u800c\u63d0\u9ad8\u6548\u7387\u3002 OmniQuant \u5c06\u91cf\u5316\u969c\u788d\u4ece\u6fc0\u6d3b\u8f6c\u79fb\u5230\u6743\u91cd\uff0c\u5e76\u5bf9\u6781\u7aef\u503c\u5fae\u8c03\u526a\u88c1\u9608\u503c\u3002 \u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6548\u7387\uff0cRPTQ \u5bf9\u76f8\u4f3c\u901a\u9053\u8fdb\u884c\u5206\u7ec4\uff0c\u4ee5\u786e\u4fdd\u91cf\u5316\u53c2\u6570\u7684\u5747\u5300\u6027\u3002<\/li>\n<li><strong>KV-Cache \u91cf\u5316 (KV-Cache Quantization\uff0cKVQ)\uff1a<\/strong>KV-Cache \u91cf\u5316\u89e3\u51b3\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u5185\u5b58\u4f18\u5316\u6311\u6218\uff0c\u5c24\u5176\u662f\u5728\u8f93\u5165 token \u6570\u91cf\u589e\u52a0\u65f6\u3002 KVQuant \u5f15\u5165\u4e86\u5b9a\u5236\u65b9\u6cd5\uff0c\u7528\u4e8e\u5728\u957f\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0b\u8fdb\u884c\u9ad8\u6548\u63a8\u7406\uff0c\u5728\u6700\u5c0f\u635f\u5931\u7684\u60c5\u51b5\u4e0b\u4fdd\u6301\u6027\u80fd\u3002 KIVI \u901a\u8fc7\u5bf9 key \u548c value \u7f13\u5b58\u5e94\u7528\u4e0d\u540c\u7684\u91cf\u5316\u7b56\u7565\u6765\u4f18\u5316\u5185\u5b58\u8282\u7701\uff0c\u5b9e\u73b0\u4e86 2 \u6bd4\u7279\u91cf\u5316\uff0c\u65e0\u9700\u5fae\u8c03\u3002 WKVQuant \u901a\u8fc7\u4e8c\u7ef4\u91cf\u5316\u7b56\u7565\u548c\u8de8\u5757\u6b63\u5219\u5316\u8fdb\u4e00\u6b65\u5b8c\u5584\u4e86\u8fd9\u4e00\u70b9\uff0c\u63d0\u4f9b\u4e0e\u6743\u91cd-\u6fc0\u6d3b\u91cf\u5316\u76f8\u5f53\u7684\u5185\u5b58\u6548\u7387\uff0c\u4e14\u6027\u80fd\u51e0\u4e4e\u76f8\u540c\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251104123510802.jpg\" alt=\"\" \/><br \/>\n<center>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u91cf\u5316\u65b9\u6cd5\u6982\u8ff0 (2021\u20132025)\u3002\u672c\u8868\u603b\u7ed3\u4e86\u6709\u4ee3\u8868\u6027\u7684\u91cf\u5316\u6280\u672f\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u5b83\u4eec\u7684\u4e3b\u8981LLM\u3001\u4f4d\u5bbd(bit widths)\u3001\u56f0\u60d1\u5ea6\u5dee\u5f02(perplexity differences)\u3001\u52a0\u901f(speedups)\u4ee5\u53ca\u5728\u4e09\u4e2a\u6307\u6807\u4e0a\u7684\u53d1\u5e03\u65f6\u95f4\u8868\uff1a\u4f4d\u5bbd\uff08\u6743\u91cd\u3001\u6fc0\u6d3b\u548cKV\u7f13\u5b58\u7684\u4f4d\u6570\uff09\u3001\u56f0\u60d1\u5ea6\u5dee\u5f02\uff08\u5728Wikitext-2\u548cC4\u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\u53d8\u5316\uff09\u548c\u52a0\u901f\uff08\u76f8\u5bf9\u4e8e\u57fa\u7ebf\u6a21\u578b\u7684\u8ba1\u7b97\u901f\u5ea6\u63d0\u5347\uff09<\/center><\/li>\n<\/ul>\n<h5>\u53c2\u6570\u526a\u679d<\/h5>\n<p>\u53c2\u6570\u526a\u679d\u662f\u4e00\u79cd\u901a\u8fc7\u6700\u5c0f\u5316\u6a21\u578b\u5927\u5c0f\u548c\u590d\u6742\u6027\u800c\u4e0d\u727a\u7272\u51c6\u786e\u6027\u7684\u5173\u952e\u6280\u672f\uff0c\u7528\u4e8e\u63d0\u9ad8\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6548\u7387\u3002\u5982\u56fe\u6240\u793a\uff0c\u526a\u679d\u53ef\u4ee5\u5206\u4e3a\u975e\u7ed3\u6784\u5316\u526a\u679d\u548c\u7ed3\u6784\u5316\u526a\u679d\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251105013823587.png\" alt=\"\" \/><\/p>\n<p><center>\u5927\u8bed\u8a00\u6a21\u578b\u526a\u679d\u53c2\u6570\u6280\u672f\u7684\u56fe\u793a<\/center><\/p>\n<ul>\n<li><strong>\u975e\u7ed3\u6784\u5316\u526a\u679d\uff08Unstructured Pruning\uff09:<\/strong>\u975e\u7ed3\u6784\u5316\u526a\u679d\u901a\u8fc7\u6d88\u9664\u4e0d\u91cd\u8981\u7684\u6743\u91cd\u6765\u589e\u5f3a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u7a00\u758f\u6027\u3002 \u88ab\u79f0\u4e3a SparseGPT \u7684\u65b9\u6cd5\u901a\u8fc7\u5c11\u6837\u672c\u526a\u679d\u5b9e\u73b0\u4e86\u9ad8\u8fbe 60% \u7684\u7a00\u758f\u5ea6\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u6700\u5c0f\u7684\u635f\u5931\u3002\u65b9\u6cd5 Wanda \u57fa\u4e8e\u6743\u91cd\u5927\u5c0f\u548c\u6fc0\u6d3b\u6267\u884c\u526a\u679d\uff0c\u65e0\u9700\u91cd\u65b0\u8bad\u7ec3\u3002 \u4e0e\u6b64\u540c\u65f6\uff0cSAMSP \u5229\u7528 Hessian \u77e9\u9635\u7684\u654f\u611f\u6027\u8fdb\u884c\u52a8\u6001\u8c03\u6574\u7a00\u758f\u5ea6\uff0c\u65e8\u5728\u6700\u5c0f\u5316\u8bef\u5dee\u3002DSnoT \u901a\u8fc7\u91c7\u7528\u8fed\u4ee3\u526a\u679d\u5faa\u73af\u6765\u63d0\u9ad8\u6027\u80fd\u3002 \u6700\u540e\uff0cFlash-LLM \u4ece\u5168\u5c40\u5185\u5b58\u4e2d\u68c0\u7d22\u7a00\u758f\u6743\u91cd\uff0c\u5e76\u5728\u82af\u7247\u4e0a\u7f13\u51b2\u533a\u4e2d\u5bc6\u96c6\u5730\u91cd\u5efa\u5b83\u4eec\uff0c\u4ee5\u4fc3\u8fdb\u9ad8\u6548\u8ba1\u7b97\u3002<\/li>\n<li><strong>\u7ed3\u6784\u5316\u526a\u679d\uff08Structured Pruning\uff09:<\/strong>\u8fd9\u79cd\u65b9\u6cd5\u4fa7\u91cd\u4e8e\u526a\u679d LLM \u4e2d\u7684\u6574\u4e2a\u53c2\u6570\u7ec4\uff0c\u4ee5\u63d0\u9ad8\u786c\u4ef6\u6548\u7387\u5e76\u7b80\u5316\u7ed3\u6784\u3002\u4f8b\u5982\uff0cLLM-runer \u8bc4\u4f30 LLaMA \u7684\u91cd\u8981\u6027\uff0c\u5e76\u4f7f\u7528 LoRA \u6765\u6062\u590d\u526a\u679d\u540e\u7684\u51c6\u786e\u6027\u3002FLAP \u4f7f\u7528\u7ed3\u6784\u5316\u6307\u6807\u4f18\u5316\u538b\u7f29\uff0c\u65e0\u9700\u5fae\u8c03\u3002\u6b64\u5916\uff0cSliceGPT \u91c7\u7528 PCA \u8fdb\u884c\u526a\u679d\uff0c\u540c\u65f6\u4fdd\u6301\u6548\u7387\u3002 Sheared LLaMA \u901a\u8fc7\u57fa\u4e8e\u6b63\u5219\u5316\u7684\u526a\u679d\u6765\u4f18\u5316\u6a21\u578b\u5f62\u72b6\u3002LoRAPrune \u901a\u8fc7\u57fa\u4e8e LoRA \u91cd\u8981\u6027\u7684\u8fed\u4ee3\u7ed3\u6784\u526a\u679d\u6765\u63d0\u9ad8\u6548\u7387\u3002\u6b64\u5916\uff0cDeja Vu \u901a\u8fc7\u9884\u6d4b\u5173\u952e\u6ce8\u610f\u529b\u5934\u548c MLP \u53c2\u6570\uff0c\u4f7f\u7528\u4e0a\u4e0b\u6587\u7a00\u758f\u6027\u6765\u964d\u4f4e\u5ef6\u8fdf\uff0c\u540c\u65f6\u4fdd\u6301\u51c6\u786e\u6027\u3002<\/li>\n<li><strong>\u4f4e\u79e9\u8fd1\u4f3c\uff08Low-Rank Approximation\uff09:<\/strong>\u4f4e\u79e9\u8fd1\u4f3c\u901a\u8fc7\u7528\u8f83\u5c0f\u7684\u77e9\u9635 <span class=\"katex-eq\" data-katex-display=\"false\">U<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">V<\/span> \u6765\u8fd1\u4f3c\u6743\u91cd\u77e9\u9635 <span class=\"katex-eq\" data-katex-display=\"false\">W<\/span>\uff0c\u4ece\u800c\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">W\\approx UV^\\top<\/span>\uff0c\u4ee5\u6b64\u6765\u538b\u7f29 LLM\u3002 \u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u51cf\u5c11\u4e86\u53c2\u6570\u6570\u91cf\uff0c\u800c\u4e14\u63d0\u9ad8\u4e86\u8fd0\u8425\u6548\u7387\u3002\u4f8b\u5982\uff0cTensorGPT \u91c7\u7528 Tensor-Train \u5206\u89e3 (TTD) \u6765\u5f00\u53d1\u66f4\u6709\u6548\u7684\u5d4c\u5165\u683c\u5f0f\u3002LoSparse \u5c06\u4f4e\u79e9\u8fd1\u4f3c\u4e0e\u526a\u679d\u76f8\u7ed3\u5408\uff0c\u4e13\u95e8\u7528\u4e8e\u538b\u7f29\u8fde\u8d2f\u7684\u795e\u7ecf\u5143\u7ec4\u4ef6\u3002FWSVD \u5b9e\u65bd\u52a0\u6743 SVD \u65b9\u6cd5\uff0c\u800c ASVD \u63d0\u4f9b\u4e86\u65e0\u8bad\u7ec3\u7684 SVD \u66ff\u4ee3\u65b9\u6848\uff0c\u4e24\u8005\u90fd\u9488\u5bf9\u8bad\u7ec3\u540e\u7684\u6548\u7387\u3002\u6700\u540e\uff0cSVD-LLM \u901a\u8fc7\u5efa\u7acb\u5947\u5f02\u503c\u4e0e\u538b\u7f29\u635f\u5931\u4e4b\u95f4\u7684\u76f4\u63a5\u5173\u7cfb\u6765\u8fdb\u4e00\u6b65\u63d0\u9ad8\u538b\u7f29\u3002<\/li>\n<\/ul>\n<h4>\u53c2\u6570\u9ad8\u6548\u5fae\u8c03<\/h4>\n<p>\u53c2\u6570\u9ad8\u6548\u5fae\u8c03(PEFT)\u7684\u8fc7\u7a0b\u5305\u62ec\u51bb\u7ed3\u5b8c\u6574\u7684 LLM \u4e3b\u5e72\uff0c\u540c\u65f6\u4ec5\u4fee\u6539\u6709\u9650\u6570\u91cf\u7684\u65b0\u589e\u53c2\u6570\u3002\u5982\u56fe\u6240\u793a\uff0cPEFT \u65b9\u6cd5\u5206\u4e3a\u56db\u7c7b\uff1a\u52a0\u6027 PEFT\uff08additive PEFT\uff09\u3001\u9009\u62e9\u6027 PEFT\uff08selective PEFT\uff09\u3001\u91cd\u53c2\u6570\u5316 PEFT\uff08reparameterized PEFT\uff09\u548c\u6df7\u5408 PEFT\uff08hybrid PEFT\uff09\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251105014715663.png\" alt=\"\" \/><\/p>\n<p><center>\u53c2\u6570\u9ad8\u6548\u5fae\u8c03 (PEFT) \u7684\u63d2\u56fe\uff0c\u8bf4\u660e\u4e86\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e2d\u8d44\u6e90\u9ad8\u6548\u9002\u5e94\u7684\u65b9\u6cd5<\/center><\/p>\n<h5>\u52a0\u6027PEFT<\/h5>\n<p>\u52a0\u6027 PEFT \u5c06\u65b0\u7684\u53ef\u8bad\u7ec3\u6a21\u5757\u5e76\u5165 LLM\uff0c\u800c\u65e0\u9700\u66f4\u6539\u539f\u59cb\u53c2\u6570\uff0c\u5141\u8bb8\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u8c03\u6574\uff0c\u540c\u65f6\u4fdd\u7559\u57fa\u672c\u6a21\u578b\u7684\u77e5\u8bc6\uff0c\u8fd9\u5bf9\u4e8e\u5fae\u8c03\u662f\u6709\u6548\u7684\u3002<\/p>\n<ul>\n<li><strong>\u9002\u914d\u5668\uff08Adapters\uff09\uff1a<\/strong>\u9002\u914d\u5668\u5728 transformer \u5757\u5185\u96c6\u6210\u7d27\u51d1\u5c42\uff0c\u5b9a\u4e49\u4e3a\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">\\text{Adapter}(x)=W_{\\text{up}},\\sigma(W_{\\text{down}},x)+x,<\/span>\u5176\u4e2d\u9002\u914d\u5668\u5c42\u5305\u62ec\u4e00\u4e2a\u4e0b\u6295\u5f71\u77e9\u9635 <span class=\"katex-eq\" data-katex-display=\"false\">W_{\\text{down}}\\in\\mathbb{R}^{r\\times d}<\/span>\uff0c\u4e00\u4e2a\u975e\u7ebf\u6027\u6fc0\u6d3b <span class=\"katex-eq\" data-katex-display=\"false\">\\sigma<\/span> \u548c\u4e00\u4e2a\u4e0a\u6295\u5f71\u77e9\u9635 <span class=\"katex-eq\" data-katex-display=\"false\">W_{\\text{up}}\\in\\mathbb{R}^{d\\times r}<\/span>\u3002 \u8fd9\u91cc\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">d<\/span> \u662f\u9690\u85cf\u5c42\u7ef4\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">r<\/span> \u662f\u74f6\u9888\u7ef4\u5ea6\uff0c\u5728\u4fdd\u6301\u6027\u80fd\u7684\u540c\u65f6\u964d\u4f4e\u4e86\u590d\u6742\u6027\u3002 \u5728\u6b64\u7ed3\u6784\u7684\u57fa\u7840\u4e0a\uff0cSerial Adapter \u5728\u6bcf\u4e2a transformer \u5757\u4e2d\u5f15\u5165\u4e86\u4e24\u4e2a\u6a21\u5757\u3002 AdapterFusion \u901a\u8fc7\u5728 Adapter &amp; Norm \u9002\u914d\u5668\u6765\u63d0\u9ad8\u6548\u7387\u3002 \u5e76\u884c\u9002\u914d\u5668 (Parallel Adapter\uff0cPA) \u5e76\u884c\u4e8e\u5b50\u5c42\u8fd0\u884c\u9002\u914d\u5668\uff0c\u800c CoDA \u901a\u8fc7\u4e0e\u5b50\u5c42\u5e76\u884c\u8fd0\u884c\u9002\u914d\u5668\u8fdb\u884c\u4f18\u5316\u3002\u4e0e AdapterFusion \u4e0d\u540c\uff0cMerA \u4f7f\u7528\u6700\u4f18\u4f20\u8f93\u6280\u672f\u7edf\u4e00\u4e86\u9002\u914d\u5668\uff0c\u7528\u4e8e\u6743\u91cd\u548c\u6fc0\u6d3b\u3002<\/li>\n<li><strong>\u8f6f\u63d0\u793a\uff08Soft Prompt\uff09\uff1a<\/strong>\u8f6f\u63d0\u793a\u901a\u8fc7\u5c06\u53ef\u8c03\u6574\u7684\u5411\u91cf\u6dfb\u52a0\u5230\u8f93\u5165\u5e8f\u5217\u6765\u589e\u5f3a\u6a21\u578b\u6027\u80fd\uff0c\u800c\u4e0d\u662f\u4f18\u5316\u79bb\u6563\u7684token\u3002 \u8fd9\u79cd\u65b9\u6cd5\u88ab\u5f62\u5f0f\u5316\u4e3a\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">X^{(l)}=\\bigl[s_1^{(l)},\\dots,s_{N_S}^{(l)},x_1^{(l)},\\dots,x_{N_X}^{(l)}\\bigr],<\/span>\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">s_i^{(l)}<\/span>\u8868\u793a\u8f6f\u63d0\u793atoken\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">x_i^{(l)}<\/span>\u8868\u793a\u539f\u59cb\u8f93\u5165token\u3002 <span class=\"katex-eq\" data-katex-display=\"false\">N_S<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">N_X<\/span>\u5206\u522b\u662f\u8f6f\u63d0\u793a\u548c\u539f\u59cb\u8f93\u5165token\u7684\u6570\u91cf\u3002\u524d\u7f00\u8c03\u4f18\uff08Prefix Tuning\uff09\u5728transformer\u5c42\u4e4b\u95f4\u5f15\u5165\u53ef\u5b66\u4e60\u7684\u5411\u91cf\uff0c\u5e76\u901a\u8fc7\u91cd\u65b0\u53c2\u6570\u5316\u6765\u7a33\u5b9a\uff0c\u5e76\u901a\u8fc7 P-Tuning v2 \u548c APT \u8fdb\u884c\u4f18\u5316\u3002 \u540c\u65f6\uff0c\u63d0\u793a\u8c03\u4f18\uff08Prompt Tuning\uff09\u4fa7\u91cd\u4e8e\u521d\u59cb\u5d4c\u5165\u5c42\uff0c\u7528\u4e8e\u4ee5\u4f4e\u8ba1\u7b97\u6210\u672c\u4f18\u5316\u5927\u578b\u6a21\u578b\u3002 Xprompt \u548c IDPG \u7b80\u5316\u4e86\u63d0\u793a\u751f\u6210\u548c\u63d2\u5165\u3002 SPoT \u548c PTP \u7b49\u65b9\u6cd5\u89e3\u51b3\u4e86\u7a33\u5b9a\u6027\u548c\u6536\u655b\u901f\u5ea6\u95ee\u9898\uff0c\u800c DePT \u548c SMoP \u901a\u8fc7\u4f18\u5316\u7684\u63d0\u793a\u7ed3\u6784\u964d\u4f4e\u4e86\u8ba1\u7b97\u9700\u6c42\u3002<\/li>\n<\/ul>\n<p>\u5176\u4ed6\u52a0\u6cd5\u65b9\u6cd5\uff1a\u9664\u4e86\u65e9\u671f\u6280\u672f\uff0c(IA)\u00b3 \u548c SSF \u7b49\u65b9\u6cd5\u4fa7\u91cd\u4e8e\u901a\u8fc7\u5bf9\u6a21\u578b\u53c2\u6570\u8fdb\u884c\u6700\u5c0f\u4f46\u5f3a\u5927\u7684\u8c03\u6574\u6765\u5b9e\u73b0\u540e\u8bad\u7ec3\u6548\u7387\u3002 \u81ea\u6ce8\u610f\u529b\u64cd\u4f5c\u548c FFN \u64cd\u4f5c\u5728\u6570\u5b66\u4e0a\u5b9a\u4e49\u4e3a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\text{SA}(x)&gt;[katex]\\text{SA}(x)=\\text{Softmax}!\\Bigl(\\frac{Q\\cdot(l_k\\odot K)^T}{\\sqrt{d_{\\text{head}}}}\\Bigr)\\cdot(l_v\\odot V),<\/span><\/center><center><span class=\"katex-eq\" data-katex-display=\"false\">\\text{FFN}{\\text{transformer}}(x)&gt;[katex]\\text{FFN}{\\text{transformer}}(x)=W{\\text{up}}\\cdot!\\bigl(l_{ff}\\odot\\sigma(W_{\\text{down}},x)\\bigr),<\/span><\/center>\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">\\odot<\/span>\u8868\u793a Hadamard \u79ef\uff0c\u7f29\u653e\u5411\u91cf<span class=\"katex-eq\" data-katex-display=\"false\">l_k<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">l_v<\/span>\u53ef\u4ee5\u5e73\u6ed1\u5730\u5e76\u5165<span class=\"katex-eq\" data-katex-display=\"false\">A_Q<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">A_W<\/span>\u7684\u6743\u91cd\u77e9\u9635\u3002 \u6b64\u5916\uff0cIPA \u5c06LLM\uff08\u5982GPT-4\uff09\u4e0e\u7528\u6237\u7279\u5b9a\u9700\u6c42\u5bf9\u9f50\u3002 \u6b64\u5916\uff0c\u5b83\u4e0d\u9700\u8981\u5bf9\u5e95\u5c42\u6a21\u578b\u8fdb\u884c\u66f4\u6539\uff0c\u56e0\u6b64\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u4e86\u6548\u7387\u3002<\/p>\n<h5>\u9009\u62e9\u6027PEFT<\/h5>\n<p>\u9009\u62e9\u6027PEFT\u901a\u8fc7\u4ec5\u5fae\u8c03\u53c2\u6570\u7684\u4e00\u4e2a\u5b50\u96c6\u6765\u63d0\u9ad8\u6548\u7387\uff0c\u5982\u4e0a\u56fe\u6240\u793a\u3002 \u8fd9\u6d89\u53ca\u5c06\u4e8c\u5143\u63a9\u7801 <span class=\"katex-eq\" data-katex-display=\"false\">M={m_1,m_2,\\dots,m_n}<\/span> \u5e94\u7528\u4e8e\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">\\theta={\\theta_1,\\theta_2,\\dots,\\theta_n}<\/span>\uff0c\u5176\u4e2d\u6bcf\u4e2a <span class=\"katex-eq\" data-katex-display=\"false\">m_i<\/span> \u6307\u793a\u662f\u5426\u9009\u62e9 <span class=\"katex-eq\" data-katex-display=\"false\">\\theta_i<\/span> \u8fdb\u884c\u5fae\u8c03\u3002 \u66f4\u65b0\u540e\u7684\u53c2\u6570\u96c6\u8868\u793a\u4e3a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\theta_i&gt;[katex]\\theta_i&#039;=\\theta_i-\\eta\\cdot m_i\\cdot\\frac{\\partial\\mathcal{L}}{\\partial\\theta_i},<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\eta<\/span> \u662f\u5b66\u4e60\u7387\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\frac{\\partial\\mathcal{L}}{\\partial\\theta_i}<\/span> \u662f\u635f\u5931\u51fd\u6570\u7684\u68af\u5ea6\u3002 \u4ec5\u66f4\u65b0\u9009\u5b9a\u7684\u53c2\u6570\uff08\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">m_i=1<\/span>\uff09\uff0c\u5728\u4fdd\u6301\u6709\u6548\u6027\u7684\u540c\u65f6\u964d\u4f4e\u8ba1\u7b97\u6210\u672c\u3002 \u65e9\u671f\u65b9\u6cd5\u5305\u62ecDiff pruning \uff0c\u5b83\u4f7f\u7528\u53ef\u5fae <span class=\"katex-eq\" data-katex-display=\"false\">L_0<\/span>-\u8303\u6570\u6765\u6b63\u5219\u5316\u53ef\u5b66\u4e60\u7684\u4e8c\u5143\u63a9\u7801\uff0c\u4ee5\u53caFishMask \uff0c\u5b83\u57fa\u4e8efisher\u4fe1\u606f\u9009\u62e9\u53c2\u6570\u4ee5\u83b7\u5f97\u66f4\u5927\u7684\u76f8\u5173\u6027\u3002 LT-SFT \u5e94\u7528\u5f69\u7968\u5047\u8bbe\u6765\u8bc6\u522b\u6709\u5f71\u54cd\u529b\u7684\u53c2\u6570\u3002 SAM \u91c7\u7528\u4e8c\u9636\u8fd1\u4f3c\u8fdb\u884c\u9009\u62e9\uff0c\u800cChild-tuning \u5728\u5b50\u7f51\u7edc\u4e2d\u52a8\u6001\u9009\u62e9\u53c2\u6570\u3002 \u6b64\u5916\uff0cFAR \u548cBitFit \u901a\u8fc7\u4e13\u6ce8\u4e8e\u4f18\u5316\u7279\u5b9a\u7684\u53c2\u6570\u7ec4\u6765\u8fdb\u4e00\u6b65\u8bf4\u660e\u9009\u62e9\u6027PEFT\u3002<\/p>\n<h5>\u91cd\u53c2\u6570\u5316PEFT<\/h5>\n<p>\u91cd\u53c2\u6570\u5316PEFT\u4e3b\u8981\u91c7\u7528\u4f4e\u79e9\u53c2\u6570\u5316\u6765\u63d0\u9ad8\u6548\u7387\uff0c\u5982\u4e0a\u56fe(c)\u6240\u793a\u3002LoRA (Low Rank Adaptation\uff0c\u4f4e\u79e9\u9002\u914d) \u5f15\u5165\u4e24\u4e2a\u53ef\u8bad\u7ec3\u7684\u77e9\u9635 <span class=\"katex-eq\" data-katex-display=\"false\">W_{\\text{up}}\\in\\mathbb{R}^{d\\times r}<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">W_{\\text{down}}\\in\\mathbb{R}^{r\\times k}<\/span>\uff0c\u4fee\u6539\u8f93\u51fa\u4e3a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">h_{\\text{out}}=W_0,h_{\\text{in}}+\\alpha,(W_{\\text{up}},W_{\\text{down}},h_{\\text{in}}),<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\alpha<\/span> \u662f\u4e00\u4e2a\u7f29\u653e\u56e0\u5b50\u3002\u8fd9\u79cd\u65b9\u6cd5\u5141\u8bb8\u6709\u6548\u5730\u9002\u5e94\u65b0\u4efb\u52a1\uff0c\u540c\u65f6\u4fdd\u7559\u6838\u5fc3\u77e5\u8bc6\u3002\u5728 LoRA \u7684\u57fa\u7840\u4e0a\uff0cIntrinsic SAID \u6700\u5c0f\u5316\u4e86\u5fae\u8c03\u53c2\u6570\u7a7a\u95f4\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u4e86\u8ba1\u7b97\u9700\u6c42\u3002 \u52a8\u6001\u53d8\u4f53\uff0c\u5305\u62ec DyLoRA \u548c AdaLoRA\uff0c\u6839\u636e\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u9700\u6c42\u52a8\u6001\u8c03\u6574\u79e9\uff0c\u5176\u4e2d AdaLoRA \u989d\u5916\u7ed3\u5408\u4e86\u57fa\u4e8e SVD \u7684\u526a\u679d\u4ee5\u63d0\u9ad8\u6548\u7387\u3002SoRA \u901a\u8fc7\u79fb\u9664\u6b63\u4ea4\u6027\u7ea6\u675f\u6765\u7b80\u5316\u6d41\u7a0b\uff0c\u800c Laplace-LoRA \u5219\u5e94\u7528\u8d1d\u53f6\u65af\u6821\u51c6\u8fdb\u884c\u5fae\u8c03\u3002 Compacter \u548c VeRA \u8fdb\u4e00\u6b65\u964d\u4f4e\u4e86\u53c2\u6570\u590d\u6742\u5ea6\u3002\u6b64\u5916\uff0cDoRA \u4f18\u5316\u4e86\u65b9\u5411\u5206\u91cf\u7684\u66f4\u65b0\uff0c\u800c HiRA \u4f7f\u7528 Hadamard \u79ef\u8fdb\u884c\u9ad8\u79e9\u66f4\u65b0\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u6548\u7387\u548c\u6027\u80fd\u3002 \u4e3a\u4e86\u5e94\u5bf9\u591a\u4efb\u52a1\u548c\u4e0d\u65ad\u53d1\u5c55\u7684\u9886\u57df\uff0cTerra \u96c6\u6210\u4e86\u4e00\u4e2a\u65f6\u53d8\u77e9\u9635\uff0cToRA \u5219\u5229\u7528 Tucker \u5206\u89e3\u6765\u8fdb\u4e00\u6b65\u6539\u8fdb LoRA \u7ed3\u6784\u3002\u9664\u4e86\u7ed3\u6784\u8bbe\u8ba1\uff0cPiSSA \u548c LoRA-GA \u4f7f\u7528 SVD \u548c\u68af\u5ea6\u5bf9\u9f50\u4f18\u5316\u4e86 LoRA \u7684\u521d\u59cb\u5316\u3002\u540c\u65f6\uff0cLoRA+\u3001LoRA-Pro \u548c CopRA \u8fdb\u4e00\u6b65\u5b8c\u5584\u4e86\u68af\u5ea6\u66f4\u65b0\u7b56\u7565\u3002 \u6b64\u5916\uff0cComLoRA \u91c7\u7528\u7ade\u4e89\u5b66\u4e60\u6765\u9009\u62e9\u6027\u80fd\u6700\u4f73\u7684 LoRA \u7ec4\u4ef6\u3002<\/p>\n<h4>\u6df7\u5408PEFT<\/h4>\n<p>\u6df7\u5408 PEFT \u65b9\u6cd5\u901a\u8fc7\u96c6\u6210\u6216\u4f18\u5316\u5404\u79cd\u5fae\u8c03\u7b56\u7565\u6765\u63d0\u9ad8\u540e\u8bad\u7ec3\u6548\u7387\u3002\u4e00\u79cd\u7a81\u51fa\u7684\u6280\u672f UniPELT \u5728 transformer \u5757\u5185\u5408\u5e76\u4e86 LoRA\u3001\u524d\u7f00\u8c03\u4f18\u548c\u9002\u914d\u5668\u3002 \u8fd9\u79cd\u65b9\u6cd5\u901a\u8fc7\u95e8\u63a7\u673a\u5236\u52a8\u6001\u6fc0\u6d3b\u7ec4\u4ef6\uff0c\u8be5\u673a\u5236\u7531\u524d\u9988\u7f51\u7edc (FFN) \u7ba1\u7406\uff0c\u524d\u9988\u7f51\u7edc\u751f\u6210\u6807\u91cf <span class=\"katex-eq\" data-katex-display=\"false\">G\\in[0,1]<\/span>\uff0c\u6700\u7ec8\u4f18\u5316\u53c2\u6570\u5229\u7528\u7387\u3002 \u53e6\u4e00\u79cd\u521b\u65b0\u65b9\u6cd5 MAM Adapter \u901a\u8fc7\u5728\u524d\u7f6e\u6ce8\u610f\u5c42\u4e2d\u7b56\u7565\u6027\u5730\u5b9a\u4f4d\u524d\u7f00\u8c03\u4f18\uff0c\u5e76\u5728\u524d\u9988\u5c42\u4e2d\u4f7f\u7528\u7f29\u653e\u7684\u5e76\u884c\u9002\u914d\u5668\u6765\u6539\u8fdb\u6b64\u6280\u672f\u3002 \u6b64\u5916\uff0c\u57fa\u4e8e NAS \u7684\u65b9\u6cd5\uff0c\u5982 NOAH \u548c AUTOPEFT\uff0c\u901a\u8fc7\u8bc6\u522b\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u91cf\u8eab\u5b9a\u5236\u7684\u6700\u4f73 PEFT \u914d\u7f6e\u6765\u63d0\u9ad8\u540e\u8bad\u7ec3\u6548\u7387\u3002 \u6b64\u5916\uff0cHeadMap \u4f7f\u7528\u8d2a\u5a6a\u65b9\u6cd5\u8bc6\u522b\u4e00\u7cfb\u5217\u6ce8\u610f\u529b\u5934\uff08\u5373\u77e5\u8bc6\u7535\u8def\uff09\uff0c\u8fd9\u4e9b\u5934\u5728\u67d0\u4e9b\u4efb\u52a1\u4e2d\u8d77\u5173\u952e\u4f5c\u7528\uff0c\u5e76\u901a\u8fc7\u5c06\u8fd9\u4e9b\u6ce8\u610f\u529b\u5934\u7684\u8f93\u51fa\u6620\u5c04\u56de LLM \u7684\u6b8b\u5dee\u6d41\u6765\u6709\u6548\u5730\u63d0\u9ad8\u6a21\u578b\u6027\u80fd\u3002 \u6700\u540e\uff0cLLM-Adapters \u63d0\u4f9b\u4e86\u4e00\u4e2a\u6846\u67b6\uff0c\u7528\u4e8e\u5728 LLM \u4e2d\u96c6\u6210\u5404\u79cd PEFT \u6280\u672f\uff0c\u786e\u4fdd\u6700\u6709\u6548\u7684\u6a21\u5757\u653e\u7f6e\u4ee5\u5728\u4e0d\u540c\u7684\u6a21\u578b\u89c4\u6a21\u4e2d\u4fdd\u6301\u6548\u7387\u3002<\/p>\n<h4>\u77e5\u8bc6\u84b8\u998f<\/h4>\n<p>\u77e5\u8bc6\u84b8\u998f (Knowledge Distillation\uff0cKD) \u6784\u6210\u4e86 LLM \u8bad\u7ec3\u540e\u4f18\u5316\u7684\u4e00\u4e2a\u91cd\u8981\u6280\u672f\uff0c\u5b83\u80fd\u591f\u5c06\u77e5\u8bc6\u4ece\u5927\u578b\u7684\u9884\u8bad\u7ec3\u6559\u5e08\u6a21\u578b\u8f6c\u79fb\u5230\u7d27\u51d1\u7684\u5b66\u751f\u6a21\u578b\uff0c\u4ee5\u63d0\u9ad8\u6548\u7387\uff0c\u800c\u4e0d\u4f1a\u727a\u7272\u6027\u80fd\u3002 KD \u6700\u521d\u662f\u5728\u6a21\u578b\u538b\u7f29\u7684\u80cc\u666f\u4e0b\u5f15\u5165\u7684\uff0c\u7531\u4e8e\u5176\u5c06\u590d\u6742\u77e5\u8bc6\u63d0\u70bc\u6210\u8d44\u6e90\u9ad8\u6548\u67b6\u6784\u7684\u80fd\u529b\u800c\u83b7\u5f97\u4e86\u5e7f\u6cdb\u5173\u6ce8\uff0c\u4ece\u800c\u80fd\u591f\u5728\u8fb9\u7f18\u8bbe\u5907\u548c\u5d4c\u5165\u5f0f\u7cfb\u7edf\u7b49\u53d7\u9650\u73af\u5883\u4e2d\u90e8\u7f72\u3002 \u901a\u8fc7\u5229\u7528\u6559\u5e08\u6a21\u578b\u7684\u7ec6\u5fae\u8f93\u51fa\u5206\u5e03\u2014\u2014\u6bd4\u4f20\u7edf\u7684\u786c\u6807\u7b7e\u66f4\u4e30\u5bcc\u2014\u2014KD \u4f7f\u5b66\u751f\u4e0d\u4ec5\u80fd\u591f\u590d\u5236\u7c7b\u522b\u9884\u6d4b\uff0c\u8fd8\u80fd\u590d\u5236\u6559\u5e08\u8868\u793a\u4e2d\u6839\u6df1\u8482\u56fa\u7684\u7c7b\u95f4\u5173\u7cfb\u548c\u5fae\u5999\u6a21\u5f0f\u3002 \u6b64\u8fc7\u7a0b\u901a\u5e38\u6d89\u53ca\u4f18\u5316\u4e00\u4e2a\u590d\u5408\u635f\u5931\u51fd\u6570\uff0c\u8be5\u51fd\u6570\u5e73\u8861\u76d1\u7763\u5b66\u4e60\u76ee\u6807\u4e0e\u7279\u5b9a\u4e8e\u84b8\u998f\u7684\u76ee\u6807\uff0c\u4ece\u800c\u663e\u7740\u964d\u4f4e\u8ba1\u7b97\u548c\u5185\u5b58\u9700\u6c42\uff0c\u540c\u65f6\u4fdd\u7559\u6cdb\u5316\u80fd\u529b\u3002<br \/>\nKD \u7684\u57fa\u672c\u673a\u5236\u4f9d\u8d56\u4e8e\u6700\u5c0f\u5316\u4e00\u4e2a\u6df7\u5408\u635f\u5931\uff0c\u8be5\u635f\u5931\u5c06\u4f20\u7edf\u7684\u5206\u7c7b\u635f\u5931\u4e0e\u84b8\u998f\u9879\u76f8\u7ed3\u5408\u3002 \u5f62\u5f0f\u4e0a\uff0c\u7ed9\u5b9a\u4e00\u4e2a\u6559\u5e08\u6a21\u578b\u7684\u8f6f\u8f93\u51fa\u6982\u7387 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{p}\\mathbf{t}<\/span> \u548c\u4e00\u4e2a\u5b66\u751f\u6a21\u578b\u7684\u9884\u6d4b <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{p}\\mathbf{s}<\/span>\uff0c\u4ee5\u53ca\u771f\u5b9e\u6807\u7b7e <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{y}<\/span> \u548c\u5b66\u751f\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{y}_\\mathbf{s}<\/span>\uff0cKD \u635f\u5931\u8868\u793a\u4e3a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}{KD}&gt;[katex]\\mathcal{L}{KD}=\\alpha,\\mathcal{L}{CE}(\\mathbf{y},\\mathbf{y}\\mathbf{s})+(1-\\alpha),\\mathcal{L}{KL}(\\mathbf{p}\\mathbf{t},\\mathbf{p}\\mathbf{s}),<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}{CE}<\/span> \u8868\u793a\u6355\u83b7\u4e0e ground truth \u5bf9\u9f50\u7684\u4ea4\u53c9\u71b5\u635f\u5931\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}{KL}<\/span> \u8868\u793a\u8861\u91cf\u6559\u5e08\u548c\u5b66\u751f\u5206\u5e03\u4e4b\u95f4\u5dee\u5f02\u7684 Kullback-Leibler \u6563\u5ea6 \uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\alpha\\in[0,1]<\/span> \u662f\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u7528\u4e8e\u8c03\u8282\u8fd9\u4e9b\u76ee\u6807\u4e4b\u95f4\u7684\u6743\u8861\u3002 \u8f6f\u76ee\u6807 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{p}\\mathbf{t}<\/span>\uff0c\u901a\u5e38\u7531\u6e29\u5ea6\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">T<\/span>\uff08\u5373\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{p}\\mathbf{t}=\\text{softmax}(\\mathbf{z}\\mathbf{t}\/T)<\/span>\uff0c\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{z}\\mathbf{t}<\/span> \u662f\u6559\u5e08 logits\uff09\u8c03\u8282\uff0c\u7f16\u7801\u66f4\u4e30\u5bcc\u7684\u6982\u7387\u4fe1\u606f\uff0c\u4f7f\u5b66\u751f\u80fd\u591f\u6a21\u4eff\u6559\u5e08\u7684\u51b3\u7b56\u7ec6\u5fae\u5dee\u522b\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u6807\u7b7e\u51c6\u786e\u6027\u3002<br \/>\nKD \u5e7f\u6cdb\u5e94\u7528\u4e8e\u8d44\u6e90\u53d7\u9650\u73af\u5883\u548c\u8fc1\u79fb\u5b66\u4e60\u7684\u6a21\u578b\u538b\u7f29\u4e2d\uff0c\u5176\u4e2d\u9884\u8bad\u7ec3\u7684\u6559\u5e08\u6307\u5bfc\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u5b66\u751f\u3002\u5176\u6709\u6548\u6027\u53d6\u51b3\u4e8e\u6559\u5e08\u80fd\u529b\u3001\u5b66\u751f\u67b6\u6784\u548c\u84b8\u998f\u635f\u5931\u8bbe\u8ba1\u7b49\u56e0\u7d20\u3002 \u6700\u8fd1\u7684\u8fdb\u5c55\u5c06 KD \u6269\u5c55\u5230\u8f93\u51fa\u84b8\u998f\u4e4b\u5916\uff0c\u4ece\u800c\u5728\u540e\u8bad\u7ec3\u4f18\u5316\u4e2d\u5b9e\u73b0\u66f4\u9ad8\u6548\u3001\u66f4\u5177\u9002\u5e94\u6027\u7684 LLM\u3002 KD \u65b9\u6cd5\u53ef\u5927\u81f4\u5206\u4e3a\u9ed1\u76d2 KD \u548c\u767d\u76d2 KD\uff0c\u8fd9\u53d6\u51b3\u4e8e\u5bf9\u6559\u5e08\u6a21\u578b\u5185\u90e8\u53c2\u6570\u548c\u4e2d\u95f4\u8868\u793a\u7684\u8bbf\u95ee\u7ea7\u522b\u3002 \u5982\u4e0b\u8868\u6240\u793a\uff0c\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u53ef\u5927\u81f4\u5206\u4e3a\u4e24\u79cd\u7c7b\u578b\uff1a\u9ed1\u76d2 KD \u548c\u767d\u76d2 KD\u3002 \u6211\u4eec\u7cfb\u7edf\u5730\u603b\u7ed3\u4e86\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u4e2d\u5404\u79cd\u77e5\u8bc6\u84b8\u998f\u6280\u672f\uff0c\u4ee5\u53ca\u5b83\u4eec\u76f8\u5e94\u7684\u6280\u80fd\u3001\u6559\u5e08\u6a21\u578b\u548c\u5b66\u751f\u6a21\u578b\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251105073706533.jpg\" alt=\"\" \/><\/p>\n<p><center>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u603b\u7ed3\uff082020\u20132025\uff09\u3002\u672c\u8868\u6982\u8ff0\u4e86\u5173\u952e\u7684\u84b8\u998f\u6280\u672f\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u5b83\u4eec\u7684\u80fd\u529b\u3001\u6559\u5e08\u548c\u5b66\u751f\u6a21\u578b\u3001\u76ee\u6807\u548c\u53d1\u5e03\u65f6\u95f4\u8868\uff0c\u5206\u4e3a \u9ed1\u76d2 KD\uff08\u8bbf\u95ee\u4ec5\u9650\u4e8e\u6559\u5e08\u8f93\u51fa\uff0c\u901a\u5e38\u6765\u81ea\u95ed\u6e90 LLM\uff09\u548c \u767d\u76d2 KD\uff08\u8bbf\u95ee\u6559\u5e08\u53c2\u6570\u6216\u5206\u5e03\uff0c\u901a\u5e38\u6765\u81ea\u5f00\u6e90 LLM\uff09\u3002 \u6307\u6807\u5305\u62ec IF\uff08\u6307\u4ee4\u9075\u5faa\uff09\u3001CoT\uff08\u601d\u7ef4\u94fe\uff09\u3001ICL\uff08\u4e0a\u4e0b\u6587\u5b66\u4e60\uff09\u3001SFT\uff08\u76d1\u7763\u5fae\u8c03\uff09\u3001D&amp;S\uff08\u5dee\u5f02\u548c\u76f8\u4f3c\u6027\uff09\u3001RL\uff08\u5f3a\u5316\u5b66\u4e60\uff09\u3001TP\uff08\u601d\u7ef4\u6a21\u5f0f\uff09\u3001NLU\uff08\u81ea\u7136\u8bed\u8a00\u7406\u89e3\uff09\u548c NLG\uff08\u81ea\u7136\u8bed\u8a00\u751f\u6210\uff09<\/center><\/p>\n<ul>\n<li><strong>\u9ed1\u76d2 KD\uff1a<\/strong>\u9ed1\u76d2 KD \u6307\u7684\u662f\u5b66\u751f\u6a21\u578b\u4ec5\u4ece\u6559\u5e08\u7684\u8f93\u51fa\u903b\u8f91\u4e2d\u5b66\u4e60\uff0c\u800c\u65e0\u6cd5\u8bbf\u95ee\u5176\u5185\u90e8\u8868\u793a\u6216\u67b6\u6784\u7ec6\u8282\u7684\u573a\u666f\u3002\u8fd9\u79cd\u65b9\u6cd5\u6700\u521d\u7531 Hinton \u63d0\u51fa\uff0c\u4e0e\u7ecf\u5178\u7684 KD \u8303\u5f0f\u4e00\u81f4\uff0c\u5e76\u4e14\u7531\u4e8e\u5176\u7075\u6d3b\u6027\u800c\u88ab\u5e7f\u6cdb\u91c7\u7528\u3002 \u9ed1\u76d2 KD \u7684\u4e00\u4e2a\u5173\u952e\u4f18\u52bf\u5728\u4e8e\u5b83\u5c06\u6559\u5e08\u6a21\u578b\u89c6\u4e3a\u4e00\u4e2a\u4e0d\u900f\u660e\u7684\u51fd\u6570\uff0c\u5373\u4f7f\u6559\u5e08\u662f\u5177\u6709\u53d7\u9650\u8bbf\u95ee\u6743\u9650\u7684\u4e13\u6709\u6216\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u4e5f\u80fd\u5b9e\u73b0\u77e5\u8bc6\u8f6c\u79fb\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u5927\u578b\u6559\u5e08 LLM\uff08\u4f8b\u5982\uff0cChatGPT \u548c GPT-4\uff09\u901a\u5e38\u7528\u4e8e\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u8f93\u51fa\u3002\u540c\u65f6\uff0c\u8f83\u5c0f\u7684\u8bed\u8a00\u6a21\u578b (SLM)\uff0c\u5305\u62ec GPT-2\u3001T5\u3001Flan-T5 \u548c CodeT5\uff0c\u5145\u5f53\u5b66\u751f\u6a21\u578b\u3002 \u8fd9\u4e9b SLM \u7ecf\u8fc7\u4f18\u5316\u4ee5\u63d0\u9ad8\u6548\u7387\uff0c\u540c\u65f6\u4fdd\u6301\u5f3a\u5927\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u4f7f\u5176\u9002\u5408\u5728\u8d44\u6e90\u53d7\u9650\u7684\u73af\u5883\u4e2d\u90e8\u7f72\u3002<\/li>\n<li><strong>\u767d\u76d2 KD\uff1a<\/strong>\u767d\u76d2 KD \u901a\u8fc7\u5229\u7528\u6765\u81ea\u6559\u5e08\u5185\u90e8\u8868\u793a\u7684\u989d\u5916\u89c1\u89e3\u6765\u6269\u5c55\u4f20\u7edf\u7684\u84b8\u998f\u8303\u5f0f\u3002 \u5f53\u6559\u5e08\u6a21\u578b\u7684\u67b6\u6784\u662f\u5df2\u77e5\u4e14\u53ef\u8bbf\u95ee\u65f6\uff0c\u8fd9\u79cd\u65b9\u6cd5\u662f\u6709\u76ca\u7684\uff0c\u5141\u8bb8\u66f4\u4e30\u5bcc\u7684\u76d1\u7763\u5f62\u5f0f\u3002 \u4e0e\u5c06\u6559\u5e08\u89c6\u4e3a\u4e0d\u900f\u660e\u51fd\u6570\u7684\u9ed1\u76d2 KD \u4e0d\u540c\uff0c\u767d\u76d2 KD \u5141\u8bb8\u5b66\u751f\u6a21\u578b\u4e0d\u4ec5\u4ece\u6559\u5e08\u7684\u8f93\u51fa logits \u5b66\u4e60\uff0c\u800c\u4e14\u8fd8\u4ece\u5176\u4e2d\u95f4\u6fc0\u6d3b\u3001\u9690\u85cf\u5c42\uff0c\u751a\u81f3\u6f5c\u5728\u7684\u6ce8\u610f\u529b\u6743\u91cd\u5b66\u4e60\u3002<\/li>\n<li><strong>DeepSeek-R1: \u76f4\u63a5\u84b8\u998f\u63a8\u7406\u6a21\u5f0f\uff1a<\/strong> DeepSeek-R1 \u901a\u8fc7\u4ece\u5927\u89c4\u6a21\u6a21\u578b\u4e2d\u84b8\u998f\u590d\u6742\u7684\u63a8\u7406\u6a21\u5f0f\u5230\u7d27\u51d1\u578b\u67b6\u6784\u4e2d\uff0c\u4f8b\u8bc1\u4e86\u77e5\u8bc6\u84b8\u998f (KD) \u7684\u53d8\u9769\u6f5c\u529b\uff0c\u663e\u8457\u589e\u5f3a\u4e86\u5c0f\u578b LLM \u7684\u63a8\u7406\u80fd\u529b\uff0c\u800c\u65e0\u9700\u5bf9\u8fd9\u4e9b\u6a21\u578b\u8fdb\u884c\u76f4\u63a5 RL \u7684\u8ba1\u7b97\u8d1f\u62c5\u3002 \u8fd9\u79cd\u65b9\u6cd5\u88ab\u79f0\u4e3a\u76f4\u63a5\u84b8\u998f\uff0c\u5229\u7528\u7531\u5927\u578b\u6559\u5e08\u6a21\u578b\u751f\u6210\u7684\u7ea6 80 \u4e07\u4e2a\u6837\u672c\u7684\u7cbe\u9009\u6570\u636e\u96c6\uff0c\u5305\u62ec\u6765\u81ea DeepSeek-V3 \u7684 20 \u4e07\u4e2a\u975e\u63a8\u7406\u5b9e\u4f8b\u548c\u7531 DeepSeek-R1-Stage1 \u68c0\u67e5\u70b9\u4ea7\u751f\u7684 60 \u4e07\u4e2a\u63a8\u7406\u5b9e\u4f8b\u3002 \u8fd9\u4e9b\u6837\u672c\u6784\u6210\u4e86\u5e94\u7528\u4e8e\u5f00\u6e90\u57fa\u7840\u6a21\u578b\uff08\u5982 Qwen \u548c LLaMA mini \u53d8\u4f53\uff09\u7684 SFT \u7684\u57fa\u7840\uff0c\u4f7f\u5b66\u751f\u6a21\u578b\u80fd\u591f\u7ee7\u627f\u901a\u5e38\u4e3a\u5176\u5927\u578b\u5bf9\u5e94\u6a21\u578b\u4fdd\u7559\u7684\u590d\u6742\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>DeepSeek-R1 \u4e2d\u7684\u76f4\u63a5\u84b8\u998f\u8fc7\u7a0b\u5728\u4e00\u4e2a\u7ed3\u6784\u5316\u7ba1\u9053\u4e2d\u5c55\u5f00\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u6700\u521d\uff0c\u5728\u5e7f\u6cdb\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u7684\u6559\u5e08\u6a21\u578b\u4f1a\u751f\u6210\u4e00\u4e2a\u5305\u542b\u63a8\u7406\u548c\u975e\u63a8\u7406\u8f93\u51fa\u7684\u591a\u5143\u8bed\u6599\u5e93\uff0c\u6355\u83b7\u4e00\u7cfb\u5217\u903b\u8f91\u6a21\u5f0f\u548c\u4e8b\u5b9e\u77e5\u8bc6\u3002 \u975e\u63a8\u7406\u6570\u636e\uff08\u7ea6 20 \u4e07\u4e2a\u6837\u672c\uff09\u63d0\u4f9b\u4e86\u901a\u7528\u77e5\u8bc6\u7684\u57fa\u7ebf\uff0c\u800c\u63a8\u7406\u6570\u636e\uff08\u7ea6 60 \u4e07\u4e2a\u6837\u672c\uff09\u5c01\u88c5\u4e86\u591a\u6b65\u63a8\u7406\u94fe\uff0c\u5e76\u901a\u8fc7\u6559\u5e08\u7684\u9ad8\u7ea7\u80fd\u529b\u8fdb\u884c\u63d0\u70bc\u3002\u7136\u540e\uff0c\u8be5\u6570\u636e\u96c6\u88ab\u7528\u4e8e SFT \u9636\u6bb5\uff0c\u5176\u4e2d\u5b66\u751f\u6a21\u578b\u88ab\u8bad\u7ec3\u4f7f\u5176\u8f93\u51fa\u5206\u5e03\u4e0e\u6559\u5e08\u7684\u8f93\u51fa\u5206\u5e03\u5bf9\u9f50\uff0c\u4f7f\u7528\u63a8\u7406\u6570\u636e\u76f4\u63a5\u5fae\u8c03\u8f83\u5c0f\u7684\u6a21\u578b\u4ee5\u84b8\u998f\u51fa\u4e00\u4e2a\u7d27\u51d1\u7684\u63a8\u7406\u6a21\u578b\u3002\u4e0e\u76f4\u63a5\u5e94\u7528\u4e8e\u5c0f\u578b\u6a21\u578b\u7684\u4f20\u7edf RL\uff08\u53ef\u80fd\u7531\u4e8e\u5bb9\u91cf\u6709\u9650\u800c\u4ea7\u751f\u6b21\u4f18\u63a8\u7406\uff09\u4e0d\u540c\uff0cDeepSeek-R1 \u7684\u76f4\u63a5\u84b8\u998f\u901a\u8fc7\u8f6c\u79fb\u9884\u4f18\u5316\u7684\u63a8\u7406\u884c\u4e3a\u6765\u89c4\u907f\u6b64\u7c7b\u9650\u5236\uff0c\u4ece\u800c\u4ee5\u66f4\u5c11\u7684\u8d44\u6e90\u9700\u6c42\u5b9e\u73b0\u5353\u8d8a\u7684\u6027\u80fd\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251105074541949.png\" alt=\"\" \/><\/p>\n<p><center>eepSeek-R1 \u4e2d\u77e5\u8bc6\u84b8\u998f\u7684\u5de5\u4f5c\u6d41\u7a0b\uff0c\u8bf4\u660e\u4e86\u5c06\u63a8\u7406\u6a21\u5f0f\u4ece\u5927\u578b\u6a21\u578b\u8f6c\u79fb\u5230\u7d27\u51d1\u578b\u6a21\u578b\u7684\u8fc7\u7a0b<\/center>DeepSeek-R1 \u7684 KD \u65b9\u6cd5\u7684\u4e00\u4e2a\u663e\u8457\u7279\u5f81\u662f\u5b83\u5f3a\u8c03\u5728\u4e0d\u540c\u6a21\u578b\u5c3a\u5ea6\u4e0a\u4fdd\u6301\u63a8\u7406\u5b8c\u6574\u6027\u3002\u901a\u8fc7\u6574\u5408\u6765\u81ea DeepSeek-R1-Stage1\uff08\u901a\u8fc7\u5927\u89c4\u6a21 RL \u63d0\u70bc\u7684\u68c0\u67e5\u70b9\uff09\u7684\u63a8\u7406\u8f68\u8ff9\uff0c\u5b66\u751f\u6a21\u578b\u4e0d\u4ec5\u590d\u5236\u4e86\u4e8b\u5b9e\u51c6\u786e\u6027\uff0c\u800c\u4e14\u6a21\u62df\u4e86\u590d\u6742\u7684\u63a8\u7406\u8fc7\u7a0b\uff0c\u4f8b\u5982\u89e3\u51b3\u6570\u5b66\u95ee\u9898\u6216\u903b\u8f91\u6f14\u7ece\u6240\u9700\u7684\u8fc7\u7a0b\u3002\u8fd9\u79cd\u6709\u9488\u5bf9\u6027\u7684\u8f6c\u79fb\u4e0e\u4f20\u7edf\u7684 KD \u5f62\u6210\u5bf9\u6bd4\uff0c\u540e\u8005\u901a\u5e38\u4f18\u5148\u8003\u8651\u5206\u7c7b\u4efb\u52a1\uff0c\u5e76\u7a81\u51fa\u4e86 DeepSeek-R1 \u5728\u9762\u5411\u63a8\u7406\u7684\u84b8\u998f\u65b9\u9762\u7684\u521b\u65b0\u3002\u6b64\u5916\uff0c\u8be5\u65b9\u6cd5\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u4e86\u5bf9\u5b66\u751f\u8fdb\u884c\u5927\u91cf RL \u8fed\u4ee3\u7684\u9700\u6c42\uff0c\u5229\u7528\u6559\u5e08\u7684\u9884\u8ba1\u7b97\u63a8\u7406\u8f93\u51fa\u6765\u7b80\u5316\u8bad\u7ec3\uff0c\u4ece\u800c\u63d0\u9ad8\u6548\u7387\u548c\u53ef\u6269\u5c55\u6027\u3002\u8fd9\u79cd\u65b9\u6cd5\u5c06 DeepSeek-R1 \u5b9a\u4f4d\u4e3a\u5c06\u9ad8\u7ea7\u63a8\u7406\u84b8\u998f\u5230\u7d27\u51d1\u578b LLM \u4e2d\u7684\u8303\u4f8b\uff0c\u4e3a\u672a\u6765\u7684\u540e\u8bad\u7ec3\u4f18\u5316\u5de5\u4f5c\u63d0\u4f9b\u4e86\u84dd\u56fe\u3002<\/p>\n<h3>\u7528\u4e8e\u96c6\u6210\u548c\u9002\u914d\u7684 PoLM<\/h3>\n<p>\u96c6\u6210\u548c\u9002\u914d\u6280\u672f\u5bf9\u4e8e\u589e\u5f3a LLM \u5728\u5404\u79cd\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u591a\u529f\u80fd\u6027\u548c\u6709\u6548\u6027\u81f3\u5173\u91cd\u8981\u3002 \u8fd9\u4e9b\u65b9\u6cd5\u4f7f LLM \u80fd\u591f\u65e0\u7f1d\u5904\u7406\u5f02\u6784\u6570\u636e\u7c7b\u578b\uff0c\u9002\u5e94\u7279\u5b9a\u9886\u57df\uff0c\u5e76\u5229\u7528\u591a\u79cd\u67b6\u6784\u4f18\u52bf\uff0c\u4ece\u800c\u89e3\u51b3\u590d\u6742\u3001\u591a\u65b9\u9762\u7684\u6311\u6218\u3002 \u672c\u7ae0\u9610\u8ff0\u4e86\u4e09\u79cd\u4e3b\u8981\u7b56\u7565\uff1a\u591a\u6a21\u6001\u96c6\u6210\uff08Multi-modal Integration\uff09\uff0c\u5b83\u4f7f\u6a21\u578b\u80fd\u591f\u5904\u7406\u5404\u79cd\u6570\u636e\u6a21\u6001\uff0c\u5982\u6587\u672c\u3001\u56fe\u50cf\u548c\u97f3\u9891\uff1b\u9886\u57df\u9002\u5e94\uff08Domain Adaptation\uff09\uff0c\u5b83\u4e3a\u7279\u5b9a\u884c\u4e1a\u6216\u7528\u4f8b\u6539\u8fdb\u6a21\u578b\uff1b\u4ee5\u53ca\u6a21\u578b\u5408\u5e76\uff08Model Merging\uff09\uff0c\u5b83\u5c06\u6765\u81ea\u4e0d\u540c\u6a21\u578b\u7684\u529f\u80fd\u7ed3\u5408\u8d77\u6765\uff0c\u4ee5\u4f18\u5316\u6574\u4f53\u6027\u80fd\u3002 \u603b\u7684\u6765\u8bf4\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u589e\u5f3a\u4e86 LLM \u7684\u9002\u5e94\u6027\u3001\u6548\u7387\u548c\u9c81\u68d2\u6027\uff0c\u6269\u5927\u4e86\u5b83\u4eec\u5728\u5404\u79cd\u4efb\u52a1\u548c\u4e0a\u4e0b\u6587\u4e2d\u7684\u9002\u7528\u6027\u3002<\/p>\n<h4>\u591a\u6a21\u6001\u96c6\u6210<\/h4>\n<p>\u5728\u524d\u51e0\u7ae0\u9610\u8ff0\u7684\u540e\u8bad\u7ec3\u4f18\u5316\u7b56\u7565\u7684\u57fa\u7840\u4e0a\uff0c\u672c\u8282\u7814\u7a76\u4e86\u65e8\u5728\u589e\u5f3a LLM \u548c\u5927\u578b\u591a\u6a21\u6001\u6a21\u578b\uff08LMM\uff09\u4ee5\u6709\u6548\u5904\u7406\u591a\u6a21\u6001\u6570\u636e\u7684\u9ad8\u7ea7\u65b9\u6cd5\u3002 \u867d\u7136\u76d1\u7763\u5fae\u8c03\u589e\u5f3a\u4e86 LLM \u5728\u7279\u5b9a\u4efb\u52a1\u73af\u5883\u4e2d\u7684\u719f\u7ec3\u7a0b\u5ea6\uff0c\u4f46\u5176\u5728\u5229\u7528\u5168\u65b9\u4f4d\u591a\u6a21\u6001\u80fd\u529b\u65b9\u9762\u7684\u5c40\u9650\u6027\u9700\u8981\u66f4 \u590d\u6742\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u3002 \u8fd9\u4e9b\u6280\u672f\u4f7f LMM \u80fd\u591f\u89e3\u51b3\u590d\u6742\u7684\u8de8\u6a21\u6001\u4efb\u52a1\uff08\u4f8b\u5982\uff0c\u6839\u636e\u89c6\u89c9\u8f93\u5165\u751f\u6210\u7f51\u9875\u4ee3\u7801 \uff0c\u89e3\u91ca\u50cf\u6a21\u56e0\u8fd9\u6837\u7684\u7ec6\u5fae\u6587\u5316\u4ea7\u7269 \uff0c\u4ee5\u53ca\u5728\u4e0d\u4f9d\u8d56\u5149\u5b66\u5b57\u7b26\u8bc6\u522b\u7684\u60c5\u51b5\u4e0b\u8fdb\u884c\u6570\u5b66\u63a8\u7406\uff09\uff0c\u901a\u8fc7\u5c06\u5404\u79cd\u6570\u636e\u7c7b\u578b\u96c6\u6210\u5230\u7edf\u4e00\u7684\u6846\u67b6\u4e2d\u3002 \u901a\u5e38\uff0cLMM \u5305\u62ec\u4e00\u4e2a\u6a21\u6001\u7f16\u7801\u5668\u3001\u4e00\u4e2a\u9884\u8bad\u7ec3\u7684 LLM \u4e3b\u5e72\u548c\u4e00\u4e2a\u6a21\u6001\u8fde\u63a5\u5668 \uff0c\u5982\u56fe\u6240\u793a\u3002 \u8fd9\u79cd\u67b6\u6784\u4e3a\u4f18\u5316\u6bcf\u4e2a\u7ec4\u4ef6\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u5960\u5b9a\u4e86\u57fa\u7840\uff0c\u4fc3\u8fdb\u4e86\u5f3a\u5927\u7684\u591a\u6a21\u6001\u96c6\u6210\u548c\u6027\u80fd\u589e\u5f3a\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251105075754964.png\" alt=\"\" \/><\/p>\n<p><center>\u5178\u578b\u5927\u578b\u591a\u6a21\u6001\u6a21\u578b\u7684\u67b6\u6784<\/center><\/p>\n<h5>\u6a21\u6001\u8fde\u63a5<\/h5>\n<p>\u6a21\u6001\u8fde\u63a5\u65b9\u6cd5\u5bf9\u4e8e\u5c06\u591a\u6a21\u6001\u6570\u636e\u5408\u6210\u4e3a\u4e00\u4e2a\u8fde\u8d2f\u7684\u8868\u5f81\u6846\u67b6\u81f3\u5173\u91cd\u8981\uff0c\u5206\u4e3a\u4e09\u79cd\u4e3b\u8981\u7b56\u7565\uff1a\u57fa\u4e8e\u6295\u5f71\uff08projection-based\uff09\u3001\u57fa\u4e8e\u67e5\u8be2\uff08query-based\uff09\u548c\u57fa\u4e8e\u878d\u5408\uff08fusion-based\uff09\u7684\u65b9\u6cd5 \uff0c\u5982\u56fe\u6240\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251105075703944.png\" alt=\"\" \/><\/p>\n<p><center>\u591a\u6a21\u6001\u878d\u5408\u4e2d\u6a21\u6001\u8fde\u63a5\u65b9\u6cd5\u7684\u5206\u7c7b\uff0c\u9610\u8ff0\u4e86\u57fa\u4e8e\u6295\u5f71\u3001\u57fa\u4e8e\u67e5\u8be2\u548c\u57fa\u4e8e\u878d\u5408\u7684\u65b9\u6cd5<\/center><\/p>\n<ul>\n<li><strong>\u57fa\u4e8e\u6295\u5f71\u7684\u6a21\u6001\u8fde\u63a5\uff1a<\/strong>\u57fa\u4e8e\u6295\u5f71\u7684\u65b9\u6cd5\u5c06\u4e0d\u540c\u7684\u6a21\u6001\u8f93\u5165\u8f6c\u6362\u4e3a\u7edf\u4e00\u7684\u6587\u672c\u5d4c\u5165\u7a7a\u95f4\uff0c\u5c06\u5176\u7279\u5f81\u4e0e LLM \u7684\u8bed\u8a00\u7ef4\u5ea6\u5bf9\u9f50\uff0c\u4ee5\u5b9e\u73b0\u65e0\u7f1d\u96c6\u6210\u3002LLaMA-Adapter \u901a\u8fc7\u6574\u5408\u56fe\u50cf\u7f16\u7801\u5668\u5c06 LLM \u6269\u5c55\u5230\u591a\u6a21\u6001\u7cfb\u7edf\uff0c\u4ece\u800c\u5b9e\u73b0\u56fe\u50cf\u6761\u4ef6\u4e0b\u7684\u6307\u4ee4\u8ddf\u8e2a\uff0c\u4ee5\u6b64\u4e3a\u4f8b\u3002\u5b83\u7684\u540e\u7ee7\u8005 LLaMA-Adapter V2 \u901a\u8fc7\u5c06\u89c6\u89c9\u6807\u7b7e\u5d4c\u5165\u5230\u65e9\u671f\u7684 LLM \u5c42\u4e2d\uff0c\u6539\u8fdb\u4e86\u8fd9\u4e2a\u8fc7\u7a0b\uff0c\u4ece\u800c\u4fc3\u8fdb\u4e86\u5bf9\u89c6\u89c9\u77e5\u8bc6\u7684\u66f4\u597d\u5438\u6536\u3002 FROMAGe \u91c7\u7528\u5728\u51bb\u7ed3\u7684 LLM \u548c\u89c6\u89c9\u7f16\u7801\u5668\u6846\u67b6\u5185\u5bf9\u8f93\u5165\u548c\u8f93\u51fa\u5c42\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u5b9e\u73b0\u8de8\u6a21\u6001\u4ea4\u4e92\uff0c\u800c LLaVA-1.5 \u5219\u5229\u7528\u53cc\u7ebf\u6027\u591a\u5c42\u611f\u77e5\u5668 (MLP) \u6765\u589e\u5f3a\u591a\u6a21\u6001\u5904\u7406\u7684\u9c81\u68d2\u6027\u3002\u6700\u8fd1\u7684\u8fdb\u5c55\uff0c\u4f8b\u5982 Shikra \uff0c\u6574\u5408\u4e86\u7a7a\u95f4\u5750\u6807\u4ee5\u589e\u5f3a\u81ea\u7136\u8bed\u8a00\u5bf9\u8bdd\uff0c\u800c VILA \u5219\u4f18\u5316\u4e86\u89c6\u89c9\u8bed\u8a00\u9884\u8bad\u7ec3\uff0c\u4ee5\u83b7\u5f97\u5353\u8d8a\u7684\u96f6\u6837\u672c\u80fd\u529b\u3002 DetGPT \u901a\u8fc7\u5c06\u57fa\u4e8e\u63a8\u7406\u7684\u76ee\u6807\u68c0\u6d4b\u4e0e\u81ea\u7136\u8bed\u8a00\u4ea4\u4e92\u76f8\u7ed3\u5408\uff0c\u8fdb\u4e00\u6b65\u63a8\u8fdb\u4e86\u8fd9\u4e00\u8303\u5f0f\uff0c\u5229\u7528\u6295\u5f71\u6280\u672f\u6765\u4fc3\u8fdb\u6709\u6548\u7684\u591a\u6a21\u6001\u901a\u4fe1\u3002 SOLO \u91c7\u7528\u5355\u4e00\u7684 Transformer \u67b6\u6784\u8fdb\u884c\u7edf\u4e00\u7684\u7aef\u5230\u7aef\u89c6\u89c9\u8bed\u8a00\u5efa\u6a21\uff0c\u901a\u8fc7\u63a5\u53d7\u539f\u59cb\u56fe\u50cf\u5757\uff08\u4ee5\u50cf\u7d20\u4e3a\u5355\u4f4d\uff09\u548c\u6587\u672c\u4f5c\u4e3a\u8f93\u5165\uff0c\u800c\u65e0\u9700\u4f7f\u7528\u5355\u72ec\u7684\u9884\u8bad\u7ec3\u89c6\u89c9\u7f16\u7801\u5668\u3002 \u4e0e\u6b64\u540c\u65f6\uff0cMiniGPT-4 \u4f7f\u7528\u5355\u4e2a\u6295\u5f71\u5c42\u5c06\u51bb\u7ed3\u7684\u89c6\u89c9\u7f16\u7801\u5668\u4e0e Vicuna \u5bf9\u9f50\uff0c\u901a\u8fc7\u4e24\u9636\u6bb5\u8bad\u7ec3\u8fc7\u7a0b\u5b9e\u73b0\u4e86\u7c7b\u4f3c GPT-4 \u7684\u80fd\u529b\u3002 Idefics \u901a\u8fc7\u81ea\u56de\u5f52\u8bbe\u8ba1\u548c\u591a\u9636\u6bb5\u9884\u8bad\u7ec3\u5728\u9ad8\u6548\u63a8\u7406\u65b9\u9762\u8868\u73b0\u51fa\u8272\u3002 LaVIT \u4f7f\u7528\u79bb\u6563\u89c6\u89c9\u5206\u8bcd\u5668\u7edf\u4e00\u89c6\u89c9\u548c\u8bed\u8a00\uff0c\u4ee5\u5b9e\u73b0\u65e0\u7f1d\u751f\u6210\u3002 DeepSeek-VL2 \u901a\u8fc7\u52a8\u6001\u74e6\u7247\u548c\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\u589e\u5f3a\u4e86\u5bf9\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u7684\u7406\u89e3\u3002 \u6700\u540e\uff0cQwen2.5-VL \u901a\u8fc7\u91cd\u65b0\u8bbe\u8ba1\u7684 Vision Transformer \u63a8\u8fdb\u4e86\u591a\u6a21\u6001\u4efb\u52a1\uff0c\u5728\u611f\u77e5\u548c\u89c6\u9891\u7406\u89e3\u65b9\u9762\u8868\u73b0\u51fa\u8272\u3002<\/li>\n<li><strong>\u57fa\u4e8e\u67e5\u8be2\u7684\u6a21\u6001\u8fde\u63a5\uff1a<\/strong>\u57fa\u4e8e\u67e5\u8be2\u7684\u65b9\u6cd5\u901a\u8fc7\u91c7\u7528\u53ef\u5b66\u4e60\u7684\u67e5\u8be2 token \u4ece\u4e0d\u540c\u7684\u6a21\u6001\u4e2d\u63d0\u53d6\u7ed3\u6784\u5316\u4fe1\u606f\uff0c\u4ece\u800c\u5f25\u5408\u4e86\u6587\u672c\u6570\u636e\u548c\u975e\u6587\u672c\u6570\u636e\u4e4b\u95f4\u7684\u5dee\u8ddd\uff0c\u4ece\u800c\u589e\u5f3a\u4e86\u591a\u6a21\u6001\u96c6\u6210\u3002 BLIP-2 \u901a\u8fc7\u67e5\u8be2 Transformer \u5f00\u521b\u4e86\u8fd9\u79cd\u65b9\u6cd5\uff0c\u6709\u6548\u5730\u6574\u5408\u6587\u672c\u548c\u89c6\u89c9\u8f93\u5165\u3002 Video-LLaMA \u901a\u8fc7\u7ed3\u5408\u89c6\u89c9\u7f16\u7801\u5668\u5c06\u6b64\u6280\u672f\u6269\u5c55\u5230\u89c6\u9891\u7406\u89e3\uff0c\u800c InstructBLIP \u6539\u8fdb\u4e86\u67e5\u8be2\u673a\u5236\uff0c\u4ee5\u786e\u4fdd\u7cbe\u786e\u5730\u9075\u5faa\u6307\u4ee4\u3002X-LLM \u901a\u8fc7\u4e13\u95e8\u7684\u63a5\u53e3\u5bf9\u9f50\u591a\u6a21\u6001\u8f93\u5165\uff0c\u540e\u7eed\u7684\u521b\u65b0\uff0c\u5982 mPLUG-Owl \u548c Qwen-VL \u4f18\u5316\u4e86 Q-Former \u67b6\u6784\uff0c\u4ee5\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u3002 LION \u901a\u8fc7\u63a8\u8fdb\u89c6\u89c9\u77e5\u8bc6\u6574\u5408\uff0c\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86\u57fa\u4e8e\u67e5\u8be2\u7684\u65b9\u6cd5\u7684\u6709\u6548\u6027\uff0c\u7a81\u51fa\u4e86\u5b83\u4eec\u5728\u589e\u5f3a\u8de8\u5404\u79cd\u4efb\u52a1\u7684 LMM \u6027\u80fd\u65b9\u9762\u7684\u5b9e\u7528\u6027\u3002Qwen-VL \u662f\u4e00\u7cfb\u5217\u57fa\u4e8e Qwen-7B \u7684\u5927\u578b\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff0c\u5b83\u7ed3\u5408\u4e86\u89c6\u89c9\u63a5\u6536\u5668\u3001\u4f4d\u7f6e\u611f\u77e5\u9002\u914d\u5668\u548c\u4e09\u9636\u6bb5\u8bad\u7ec3\u6d41\u7a0b\uff0c\u4ee5\u5b9e\u73b0\u591a\u8bed\u8a00\u3001\u7ec6\u7c92\u5ea6\u7684\u89c6\u89c9\u8bed\u8a00\u7406\u89e3\u3002Lyrics \u662f\u4e00\u4e2a\u7ec6\u7c92\u5ea6\u7684\u89c6\u89c9\u8bed\u8a00\u9884\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u6846\u67b6\uff0c\u901a\u8fc7\u89c6\u89c9\u4f18\u5316\u5668\uff08\u56fe\u50cf\u6807\u8bb0\u3001\u76ee\u6807\u68c0\u6d4b\u548c\u8bed\u4e49\u5206\u5272\uff09\u548c\u591a\u5c3a\u5ea6\u67e5\u8be2 Transformer (MQ-Former) \u6574\u5408\u8bed\u4e49\u611f\u77e5\u7684\u89c6\u89c9\u5bf9\u8c61\uff0c\u4ece\u800c\u589e\u5f3a\u5927\u578b\u89c6\u89c9\u8bed\u8a00\u6a21\u578b (LVLM)\u3002<\/li>\n<li><strong>\u57fa\u4e8e\u878d\u5408\u7684\u6a21\u6001\u8fde\u63a5\uff1a<\/strong>\u57fa\u4e8e\u878d\u5408\u7684\u6280\u672f\u901a\u8fc7\u5c06\u591a\u6a21\u6001\u7279\u5f81\u76f4\u63a5\u5d4c\u5165\u5230 LLM \u67b6\u6784\u4e2d\u6765\u6df1\u5316\u8de8\u6a21\u6001\u4ea4\u4e92\uff0c\u4ece\u800c\u5728\u63a8\u7406\u7ea7\u522b\u4fc3\u8fdb\u66f4\u4e30\u5bcc\u7684\u6574\u5408\u3002 Flamingo \u91c7\u7528\u4ea4\u53c9\u6ce8\u610f\u529b\u5c42\u5728 Token \u9884\u6d4b\u671f\u95f4\u878d\u5408\u89c6\u89c9\u7279\u5f81\uff0c\u4ece\u800c\u5b9e\u73b0\u52a8\u6001\u591a\u6a21\u6001\u5904\u7406\u3002 OpenFlamingo \u5728\u6b64\u57fa\u7840\u4e0a\u6784\u5efa\uff0c\u5141\u8bb8\u51bb\u7ed3\u7684 LLM \u5173\u6ce8\u89c6\u89c9\u7f16\u7801\u5668\u8f93\u51fa\uff0c\u4ece\u800c\u589e\u5f3a\u7075\u6d3b\u6027\u3002 Otter \u5f15\u5165\u4e86\u6307\u4ee4\u8c03\u4f18\u6765\u6539\u8fdb\u591a\u6a21\u6001\u6307\u4ee4\u9075\u5faa\uff0c\u800c CogVLM \u5728 Transformer \u5c42\u5185\u96c6\u6210\u4e86\u89c6\u89c9\u4e13\u5bb6\u6a21\u5757\uff0c\u4ee5\u5b9e\u73b0\u65e0\u7f1d\u7684\u7279\u5f81\u5408\u6210\u3002 Obelics \u5229\u7528\u4ea4\u9519\u7684\u56fe\u50cf-\u6587\u672c\u8bad\u7ec3\u6570\u636e\uff0c\u7a81\u51fa\u4e86\u57fa\u4e8e\u878d\u5408\u7684\u65b9\u6cd5\u5728\u5b9e\u73b0\u5185\u805a\u591a\u6a21\u6001\u6027\u80fd\u65b9\u9762\u7684\u7a33\u5065\u6027\u3002 InternVL \u662f\u4e00\u4e2a\u5927\u578b\u89c6\u89c9\u8bed\u8a00\u57fa\u7840\u6a21\u578b\uff0c\u5b83\u5c06\u89c6\u89c9\u7f16\u7801\u5668\u6269\u5c55\u5230 60 \u4ebf\u4e2a\u53c2\u6570\uff0c\u5e76\u4f7f\u7528\u8bed\u8a00\u4e2d\u95f4\u4ef6\uff08QLLaMA\uff09\u9010\u6b65\u5c06\u5176\u4e0e LLM \u5bf9\u9f50\u3002 Llama 3 \u662f Meta \u5f00\u53d1\u7684\u5168\u65b0\u591a\u8bed\u8a00\u3001\u5de5\u5177\u4f7f\u7528\u57fa\u7840\u6a21\u578b\u7cfb\u5217\uff0c\u6269\u5c55\u5230 405B \u53c2\u6570\uff0c\u5177\u6709 128K \u4e2a Token \u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u901a\u8fc7\u6539\u8fdb\u7684\u6570\u636e\u8d28\u91cf\u3001\u66f4\u5927\u89c4\u6a21\u7684\u8bad\u7ec3\u548c\u7ed3\u6784\u5316\u7684\u540e\u8bad\u7ec3\u7b56\u7565\u8fdb\u884c\u4f18\u5316\u3002<\/li>\n<\/ul>\n<h5>\u6a21\u6001\u7f16\u7801\u5668<\/h5>\n<p>\u6a21\u6001\u7f16\u7801\u5668\u5c06\u539f\u59cb\u7684\u591a\u6a21\u6001\u8f93\u5165\u538b\u7f29\u6210\u7d27\u51d1\u3001\u8bed\u4e49\u4e30\u5bcc\u7684\u8868\u793a\uff0c\u4ece\u800c\u5b9e\u73b0\u8de8\u5404\u79cd\u4efb\u52a1\u548c\u6a21\u6001\u7684\u6709\u6548\u5904\u7406\u3002\u8fd9\u4e9b\u7ec4\u4ef6\u5bf9\u4e8e\u5c06\u5f02\u6784\u6570\u636e\u8f6c\u6362\u4e3a\u4e0e LLM \u4e3b\u5e72\u517c\u5bb9\u7684\u683c\u5f0f\u81f3\u5173\u91cd\u8981\uff0c\u652f\u6301\u4ece\u89c6\u89c9\u63a8\u7406\u5230\u97f3\u9891\u7406\u89e3\u7684\u5e94\u7528\u3002\u4e0b\u8868\u63d0\u4f9b\u4e86\u5bf9\u89c6\u89c9\u3001\u97f3\u9891\u548c\u5176\u4ed6\u6a21\u6001\u4e2d\u4f7f\u7528\u7684\u4e3b\u8981\u7f16\u7801\u5668\u7684\u5168\u9762\u603b\u7ed3\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u5b83\u4eec\u7684\u7279\u6027\u4ee5\u53ca\u5bf9\u591a\u6a21\u6001\u96c6\u6210\u7684\u8d21\u732e\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251106031103185.jpg\" alt=\"\" \/><\/p>\n<ul>\n<li><strong>\u89c6\u89c9\u7f16\u7801\u5668:<\/strong>\u89c6\u89c9\u7f16\u7801\u5668\u662f\u591a\u6a21\u6001\u5b66\u4e60\u7684\u57fa\u7840\uff0c\u6709\u52a9\u4e8e\u5728 LMM \u4e2d\u89e3\u91ca\u548c\u751f\u6210\u89c6\u89c9\u6570\u636e\u3002 CLIP \u901a\u8fc7\u5bf9\u6bd4\u5b66\u4e60\u5efa\u7acb\u8054\u5408\u56fe\u50cf-\u6587\u672c\u8868\u793a\uff0c\u589e\u5f3a\u8de8\u6a21\u6001\u5bf9\u9f50\u3002 EVA \u6539\u8fdb\u4e86\u89c6\u89c9\u6ce8\u610f\u529b\u673a\u5236\u4ee5\u63d0\u9ad8\u6548\u7387\uff0c\u800c ImageBind \u5728\u591a\u4e2a\u6a21\u6001\u4e4b\u95f4\u521b\u5efa\u4e86\u7edf\u4e00\u7684\u5d4c\u5165\u7a7a\u95f4\uff0c\u63d0\u5347\u4e86\u96f6\u6837\u672c\u8bc6\u522b\u80fd\u529b\u3002 SigLIP \u5f15\u5165\u4e86\u4e00\u4e2a\u914d\u5bf9\u7684 sigmoid \u635f\u5931\u6765\u4f18\u5316\u56fe\u50cf-\u6587\u672c\u9884\u8bad\u7ec3\uff0c\u800c DINOv2 \u91c7\u7528\u65e0\u76d1\u7763\u5b66\u4e60\u4ece\u4e0d\u540c\u6765\u6e90\u63d0\u53d6\u9c81\u68d2\u7684\u89c6\u89c9\u7279\u5f81\u3002 LLaVA \u91c7\u7528 self-instruct \u7b56\u7565\u5c06\u56fe\u50cf\u8f6c\u6362\u4e3a\u6587\u672c\u63cf\u8ff0\uff0c\u5229\u7528\u5148\u8fdb\u7684 LLM \u751f\u6210\u65b0\u6570\u636e\u96c6\u3002 Video-ChatGPT \u901a\u8fc7\u5927\u89c4\u6a21\u6307\u4ee4\u6570\u636e\u96c6\u652f\u6301\u5bf9\u8bdd\u5f0f\u89c6\u9891\u7406\u89e3\uff0c\u800c BT-Adapter \u901a\u8fc7\u9ad8\u6548\u7684\u65f6\u95f4\u5efa\u6a21\u4f18\u5316\u89c6\u9891\u7406\u89e3\u3002 VideoChat \u4fa7\u91cd\u4e8e\u65f6\u7a7a\u63a8\u7406\uff0c\u5229\u7528\u4e13\u95e8\u7684\u6570\u636e\u96c6\uff0c\u4ee5\u53ca CoDi-2 \u548c Mipha \u7b49\u6a21\u578b\u5728\u591a\u6a21\u6001\u5904\u7406\u4e2d\u5b9e\u73b0\u4e86\u6548\u7387\u63d0\u5347\u3002 VL-Mamba \u548c Cobra \u5f15\u5165\u4e86\u72b6\u6001\u7a7a\u95f4\u6a21\u578b\u4ee5\u4f18\u5316\u63a8\u7406\uff0c\u800c SPHINX-Tiny \u5f3a\u8c03\u6570\u636e\u591a\u6837\u6027\u548c\u8bad\u7ec3\u6548\u7387\u3002<\/li>\n<li><strong>\u97f3\u9891\u7f16\u7801\u5668\uff1a<\/strong>\u97f3\u9891\u7f16\u7801\u5668\u589e\u5f3a\u4e86LMMs\u5904\u7406\u548c\u89e3\u91ca\u542c\u89c9\u8f93\u5165\u7684\u80fd\u529b\uff0c\u62d3\u5bbd\u4e86\u5b83\u4eec\u7684\u591a\u6a21\u6001\u8303\u56f4\u3002 SpeechGPT \u5c06\u5927\u89c4\u6a21\u8bed\u97f3\u6570\u636e\u96c6\u4e0e\u5377\u79ef\u548ctransformer\u67b6\u6784 \u76f8\u7ed3\u5408\uff0c\u4ee5\u5b9e\u73b0\u5f3a\u5927\u7684\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u3002 AudioPaLM \u4f7f\u7528\u901a\u7528\u8bed\u97f3\u6a21\u578b\uff08USM\uff09\u7f16\u7801\u5668 \u7ed3\u5408\u6587\u672c\u548c\u8bed\u97f3\u5904\u7406\uff0c\u64c5\u957f\u96f6\u6837\u672c\u8bed\u8a00\u7ffb\u8bd1\u7b49\u4efb\u52a1\u3002 WavCaps \u91c7\u7528CNN14 \u548cHTSAT \u6765\u7f13\u89e3\u97f3\u9891-\u8bed\u8a00\u6570\u636e\u7a00\u7f3a\u6027\uff0c\u5229\u7528\u5148\u8fdb\u7684LLMs\u6765\u63d0\u5347\u6570\u636e\u96c6\u8d28\u91cf\u5e76\u589e\u5f3a\u5b66\u4e60\u6548\u679c\uff0c\u5f3a\u8c03\u4e86\u97f3\u9891\u6a21\u6001\u5728\u591a\u6a21\u6001\u7cfb\u7edf\u4e2d\u7684\u5173\u952e\u4f5c\u7528\u3002<\/li>\n<li><strong>\u5176\u4ed6\u7f16\u7801\u5668:<\/strong>\u9664\u4e86\u89c6\u89c9\u548c\u97f3\u9891\u4e4b\u5916\uff0c\u7528\u4e8e\u5176\u4ed6\u6a21\u6001\uff08\u4f8b\u59823D\u7406\u89e3\u548c\u591a\u6a21\u6001\u878d\u5408\uff09\u7684\u7f16\u7801\u5668\u5bf9\u4e8e\u5168\u9762\u7684LMMs\u81f3\u5173\u91cd\u8981\u3002 NEXT-GPT \u4fc3\u8fdb\u8de8\u6587\u672c\u3001\u56fe\u50cf\u3001\u89c6\u9891\u548c\u97f3\u9891\u7684\u8de8\u6a21\u6001\u5185\u5bb9\u751f\u6210\uff0c\u901a\u8fc7\u6700\u5c0f\u7684\u53c2\u6570\u8c03\u6574\u6765\u63a8\u8fdb\u7c7b\u4ebaAI\u80fd\u529b\u3002 ImageBind-LLM \u5bf9\u9f50\u89c6\u89c9\u548c\u8bed\u8a00\u5d4c\u5165\uff0c\u4ee5\u6539\u5584\u8de8\u6a21\u6001\u7684\u6307\u4ee4\u9075\u5faa\u3002 LL3DA \u5904\u7406\u70b9\u4e91\u6570\u636e\u4ee5\u8fdb\u884c3D\u63a8\u7406\u548c\u89c4\u5212\uff0c\u5f15\u5165\u4e86\u65b0\u9896\u7684\u7a7a\u95f4\u7406\u89e3\u65b9\u6cd5\u3002 X-LLM \u91c7\u7528Q-Former \u7528\u4e8e\u56fe\u50cf\u548c\u89c6\u9891\u8f93\u5165\uff0c\u91c7\u7528C-Former \u7528\u4e8e\u8bed\u97f3\uff0c\u5c06\u97f3\u9891\u7279\u5f81\u538b\u7f29\u6210token\u7ea7\u5d4c\u5165\uff0c\u4ee5\u589e\u5f3a\u591a\u6a21\u6001\u5b66\u4e60\u6548\u7387\u3002<\/li>\n<\/ul>\n<h4>\u9886\u57df\u81ea\u9002\u5e94<\/h4>\n<p>\u9886\u57df\u81ea\u9002\u5e94\uff08DA\uff09\u6784\u6210\u4e86\u4f18\u5316LLMs\u7684\u5173\u952e\u7684\u8bad\u7ec3\u540e\u7b56\u7565\uff0c\u4f7f\u5176\u5728\u4e13\u4e1a\u9886\u57df\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u786e\u4fdd\u5176\u5728\u76ee\u6807\u5e94\u7528\u4e2d\u7684\u6709\u6548\u6027\u3002 \u57fa\u4e8e\u8fc1\u79fb\u5b66\u4e60\u7684\u539f\u5219\uff0cDA\u901a\u8fc7\u81ea\u9002\u5e94\u51fd\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">F_{\\text{adapt}}<\/span> \u8f6c\u6362\u521d\u59cb\u6a21\u578b\uff08\u8868\u793a\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">M_{\\text{source}}<\/span>\uff09\uff0c\u4ee5\u751f\u6210\u7279\u5b9a\u9886\u57df\u7684\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">M_{\\text{target}}<\/span>\uff0c\u5982\u4e0b\u6240\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251106034647803.png\" alt=\"\" \/><br \/>\n\u6b64\u8fc7\u7a0b\u5b9a\u5236 <span class=\"katex-eq\" data-katex-display=\"false\">M_{\\text{target}}<\/span>\uff0c\u4ee5\u6ee1\u8db3\u6307\u5b9a\u9886\u57df\u7684\u72ec\u7279\u9700\u6c42\u548c\u590d\u6742\u6027\uff0c\u4ece\u800c\u4f18\u5316\u5176\u6027\u80fd\u548c\u76f8\u5173\u6027\u3002 \u901a\u8fc7\u589e\u5f3aLLMs\u5728\u7f16\u7a0b \u548c\u6570\u5b66\u63a8\u7406 \u7b49\u9886\u57df\u7684\u719f\u7ec3\u7a0b\u5ea6\uff0cDA\u4e0d\u4ec5\u63d0\u9ad8\u4e86\u7279\u5b9a\u9886\u57df\u7684\u80fd\u529b\uff0c\u8fd8\u63d0\u9ad8\u4e86\u8ba1\u7b97\u6548\u7387\uff0c\u51cf\u8f7b\u4e86\u901a\u7528\u6a21\u578b\u7684\u5c40\u9650\u6027\uff0c\u901a\u7528\u6a21\u578b\u901a\u5e38\u96be\u4ee5\u5904\u7406\u7279\u5b9a\u9886\u57df\u7684\u672f\u8bed\u548c\u63a8\u7406\u8303\u5f0f\u3002 \u6b64\u5916\uff0cDA\u5927\u5927\u51cf\u5c11\u4e86\u5bf9\u5927\u89c4\u6a21\u6807\u8bb0\u6570\u636e\u96c6\u548c\u8ba1\u7b97\u8d44\u6e90\u7684\u4f9d\u8d56\uff0c\u8fd9\u4e9b\u8d44\u6e90\u901a\u5e38\u662f\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u7279\u5b9a\u9886\u57df\u6a21\u578b\u6240\u5fc5\u9700\u7684 \uff0c\u4f7f\u5176\u6210\u4e3a\u8bad\u7ec3\u540e\u65b9\u6cd5\u7684\u6838\u5fc3\u3002<\/p>\n<h5>\u77e5\u8bc6\u7f16\u8f91<\/h5>\n<p>\u77e5\u8bc6\u7f16\u8f91\u4ee3\u8868\u4e86\u4e00\u79cd\u7cbe\u5bc6\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\uff0c\u65e8\u5728\u4fee\u6539LLM\u4ee5\u6ee1\u8db3\u7279\u5b9a\u9886\u57df\u7684\u9700\u8981\uff0c\u800c\u4e0d\u4f1a\u635f\u5bb3\u5176\u57fa\u7840\u80fd\u529b\u3002 \u8fd9\u79cd\u6280\u672f\u6709\u52a9\u4e8e\u6709\u9488\u5bf9\u6027\u7684\u53c2\u6570\u8c03\u6574\uff0c\u5728\u6574\u5408\u65b0\u7684\u6216\u66f4\u65b0\u7684\u9886\u57df\u77e5\u8bc6\u7684\u540c\u65f6\uff0c\u4fdd\u7559\u6a21\u578b\u7684\u73b0\u6709\u6027\u80fd\u3002 \u901a\u8fc7\u5b9e\u73b0\u5bf9\u4e0d\u65ad\u53d1\u5c55\u7684\u77e5\u8bc6\u9886\u57df\u7684\u5feb\u901f\u9002\u5e94\uff0c\u77e5\u8bc6\u7f16\u8f91\u6210\u4e3a\u540e\u8bad\u7ec3\u6d41\u7a0b\u4e2d\u4e0d\u53ef\u6216\u7f3a\u7684\u7ec4\u6210\u90e8\u5206\u3002\u4e0b\u8868\u4e2d\u6982\u8ff0\u4e86\u4e3b\u8981\u65b9\u6cd5\uff08\u4f8b\u5982\uff0c\u5305\u62ec\u5916\u90e8\u77e5\u8bc6\u5229\u7528\u3001\u6574\u5408\u548c\u5185\u5728\u7f16\u8f91\uff09\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251106071715208.jpg\" alt=\"\" \/><\/p>\n<p><center>\u5bf9LLM\u4e2d\u77e5\u8bc6\u7f16\u8f91\u7684\u4ee3\u8868\u6027\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\u5206\u6790\u3002 \u7f16\u8f91\u533a\u57df\u6307\u5b9a\u4e86\u6a21\u578b\u4e2d\u8981\u4fee\u6539\u7684\u7ec4\u4ef6\uff1b \u7f16\u8f91\u5668 #\u53c2\u6570\u6307\u793a\u5728\u7f16\u8f91\u8fc7\u7a0b\u4e2d\u9700\u8981\u66f4\u65b0\u7684\u53c2\u6570\u3002<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">L<\/span>\u8868\u793a\u53d7\u5230\u4fee\u6539\u7684\u5c42\u6570\uff0c<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">d_h<\/span>\u8868\u793a Transformer \u67b6\u6784\u4e2d\u9690\u85cf\u5c42\u7684\u7ef4\u5ea6\uff0c<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">d_m<\/span>\u6307\u7684\u662f\u4e0a\u6295\u5f71\u548c\u4e0b\u6295\u5f71\u9636\u6bb5\u4e4b\u95f4\u5b58\u5728\u7684\u4e2d\u95f4\u7ef4\u5ea6\uff0c \u5e76\u4e14<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">N<\/span>\u8c61\u5f81\u7740\u5728\u6bcf\u4e2a\u5355\u72ec\u5c42\u4e2d\u8fdb\u884c\u66f4\u65b0\u7684\u795e\u7ecf\u5143\u603b\u6570\u3002<\/center><strong>\u77e5\u8bc6\u7f16\u8f91\u7684\u5f62\u5f0f\u5316\u5b9a\u4e49(Formal Definition of Knowledge Editing):<\/strong>\u8003\u8651\u4e00\u4e2a\u7531 <span class=\"katex-eq\" data-katex-display=\"false\">\\theta<\/span> \u53c2\u6570\u5316\u7684\u539f\u59cbLLM\uff0c\u8be5LLM\u5728\u6570\u636e\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}{\\text{old}}<\/span> \u4e0a\u8fdb\u884c\u4e86\u9884\u8bad\u7ec3\u3002 \u8bbe <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}{\\text{new}}<\/span> \u8868\u793a\u4e00\u4e2a\u5305\u542b\u65b0\u4fe1\u606f\u6216\u66f4\u65b0\u4fe1\u606f\u7684\u6570\u636e\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\Delta K<\/span>\u3002 \u77e5\u8bc6\u7f16\u8f91\u7684\u76ee\u6807\u662f\u901a\u8fc7\u5e94\u7528\u8c03\u6574 <span class=\"katex-eq\" data-katex-display=\"false\">\\Delta\\theta<\/span>\uff0c\u63a8\u5bfc\u51fa\u4e00\u4e2a\u4fee\u6b63\u7684\u53c2\u6570\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\theta&#039;<\/span>\uff0c\u6709\u6548\u5730\u5438\u6536 <span class=\"katex-eq\" data-katex-display=\"false\">\\Delta K<\/span>\uff0c\u540c\u65f6\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u5bf9 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}_{\\text{old}}<\/span> \u7684\u9000\u5316\u3002 \u5f62\u5f0f\u4e0a\uff0c\u8fd9\u88ab\u6846\u5b9a\u4e3a\u4e00\u4e2a\u7ea6\u675f\u4f18\u5316\u95ee\u9898\uff0c\u5176\u4e2d\u66f4\u65b0\u540e\u7684\u53c2\u6570\u5b9a\u4e49\u4e3a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\theta&gt;[katex]\\theta&#039;=\\theta+\\Delta\\theta,\\quad\\text{where}\\ \\mathcal{L}(\\theta&#039;;\\mathcal{D}_{\\text{new}})\\to\\min,<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}<\/span> \u8868\u793a\u4e00\u4e2a\u635f\u5931\u51fd\u6570\uff08\u4f8b\u5982\uff0c\u4ea4\u53c9\u71b5\uff09\uff0c\u7528\u4e8e\u8bc4\u4f30\u6a21\u578b\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}_{\\text{new}}<\/span> \u4e0a\u7684\u8d28\u91cf\u3002\u4e3a\u4e86\u4fdd\u62a4\u539f\u59cb\u6570\u636e\u96c6\u7684\u6027\u80fd\uff0c\u65bd\u52a0\u4e86\u4e00\u4e2a\u7ea6\u675f\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{L}(\\theta&gt;[katex]\\mathcal{L}(\\theta&#039;;\\mathcal{D}{\\text{old}})\\leq\\mathcal{L}(\\theta;\\mathcal{D}{\\text{old}})+\\epsilon,<\/span><\/center>\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\epsilon<\/span> \u662f\u4e00\u4e2a\u5c0f\u7684\u6b63\u6570\u5e38\u6570\uff0c\u7528\u4e8e\u9650\u5236\u5728 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}_{\\text{old}}<\/span> \u4e0a\u7684\u6027\u80fd\u635f\u5931\u3002\u8fd9\u79cd\u8868\u8ff0\u786e\u4fdd\u4e86 <span class=\"katex-eq\" data-katex-display=\"false\">\\theta&#039;<\/span> \u5305\u542b\u4e86 <span class=\"katex-eq\" data-katex-display=\"false\">\\Delta K<\/span>\uff0c\u540c\u65f6\u4fdd\u7559\u4e86\u6a21\u578b\u539f\u6709\u7684\u77e5\u8bc6\u5e93\u3002 \u5b9e\u9645\u4e0a\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\Delta\\theta<\/span> \u53ef\u4ee5\u88ab\u9650\u5236\u5728\u7279\u5b9a\u7684\u67b6\u6784\u7ec4\u4ef6\uff08\u4f8b\u5982\uff0c\u6ce8\u610f\u529b\u5c42 (<span class=\"katex-eq\" data-katex-display=\"false\">\\mathrm{Attn}<\/span>) \u6216\u524d\u9988\u7f51\u7edc (<span class=\"katex-eq\" data-katex-display=\"false\">\\mathrm{FFN}<\/span>)\uff09\uff0c\u901a\u8fc7\u907f\u514d\u5168\u9762\u7684\u91cd\u65b0\u8bad\u7ec3\uff0c\u4ece\u800c\u51cf\u5c11\u8ba1\u7b97\u5f00\u9500\u5e76\u4fdd\u6301\u6838\u5fc3\u529f\u80fd\u3002<\/p>\n<p><strong>\u77e5\u8bc6\u8bc6\u522b(Knowledge Identification):<\/strong>\u77e5\u8bc6\u7f16\u8f91\u7684\u521d\u59cb\u9636\u6bb5\u4fa7\u91cd\u4e8e\u68c0\u6d4b\u65b0\u4fe1\u606f\u5e76\u5c06\u5176\u5438\u6536\u5230\u6a21\u578b\u4e2d\u3002PokeMQA \u91c7\u7528\u53ef\u7f16\u7a0b\u8303\u56f4\u68c0\u6d4b\u5668\u548c\u77e5\u8bc6\u63d0\u793a\u6765\u5256\u6790\u67e5\u8be2\uff0c\u4ece\u800c\u6709\u6548\u5730\u68c0\u7d22\u76f8\u5173\u4e8b\u5b9e\u3002\u76f8\u53cd\uff0cSERAC \u5c06\u53cd\u4e8b\u5b9e\u6a21\u578b\u4e0e\u5206\u7c7b\u5668\u76f8\u7ed3\u5408\uff0c\u4ee5\u786e\u5b9a\u65b0\u77e5\u8bc6\u6e90\u7684\u9002\u7528\u6027\uff0c\u63d0\u4f9b\u4e86\u4e00\u79cd\u5fae\u521b\u65b9\u6cd5\uff0c\u5728\u4e0d\u9700\u5927\u89c4\u6a21\u7ed3\u6784\u4fee\u6539\u7684\u60c5\u51b5\u4e0b\uff0c\u4fdd\u6301\u4e86\u57fa\u7840\u6a21\u578b\u7684\u5b8c\u6574\u6027\u3002\u5206\u6790\u4e86 LLM \u77e5\u8bc6\u66f4\u65b0\u4ea7\u751f\u6df7\u4e71\u6d9f\u6f2a\u6548\u5e94\u7684\u539f\u56e0\u3002\u73b0\u5b9e\u4e16\u754c\u7684\u7f16\u8f91\u901a\u5e38\u6e90\u4e8e\u65b0\u5174\u4e8b\u4ef6\uff0c\u8fd9\u4e9b\u4e8b\u4ef6\u5305\u542b\u4e86\u65b0\u4e8b\u5b9e\u548c\u8fc7\u53bb\u4e8b\u5b9e\u4e4b\u95f4\u7684\u903b\u8f91\u8054\u7cfb\uff0c\u57fa\u4e8e\u8fd9\u4e00\u89c2\u5bdf\uff0cEvEdit \u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u4e8b\u4ef6\u7684\u77e5\u8bc6\u7f16\u8f91\u65b9\u6cd5\uff0c\u4ee5\u786e\u5b9a\u77e5\u8bc6\u951a\u70b9\u548c\u77e5\u8bc6\u66f4\u65b0\u8fb9\u754c\u3002<br \/>\n\u77e5\u8bc6\u5173\u8054(Knowledge Association):\u5728\u8bc6\u522b\u4e4b\u540e\uff0c\u6b64\u9636\u6bb5\u5c06\u65b0\u83b7\u53d6\u7684\u4fe1\u606f\u4e0e\u6a21\u578b\u73b0\u6709\u7684\u77e5\u8bc6\u6846\u67b6\u5173\u8054\u8d77\u6765\u3002 Transformer-Patcher \u8c03\u6574\u4e86 transformer \u67b6\u6784\u4ee5\u6574\u5408\u66f4\u65b0\u7684\u4e8b\u5b9e\uff0c\u800c CaliNET \u91cd\u65b0\u6821\u51c6\u53c2\u6570\u4ee5\u4e0e\u4e8b\u5b9e\u5185\u5bb9\u5bf9\u9f50\u3002\u8bf8\u5982 Eva-KELLM\u3001MELO \u548c REMEDI \u7b49\u65b9\u6cd5\u53ef\u4ee5\u7ec6\u5316\u7279\u5b9a\u884c\u4e3a\u4ee5\u5b9e\u73b0\u7cbe\u786e\u7684\u66f4\u65b0\uff0c\u800c GRACE \u5219\u5728\u77e5\u8bc6\u63d2\u5165\u540e\u63d0\u9ad8\u9884\u6d4b\u51c6\u786e\u6027\uff0c\u786e\u4fdd\u4e0e\u4e4b\u524d\u7684\u8868\u793a\u65e0\u7f1d\u96c6\u6210\u3002<br \/>\n<strong>\u5185\u5728\u77e5\u8bc6\u7f16\u8f91(Intrinsic Knowledge Editing):<\/strong>\u6700\u540e\u9636\u6bb5\u5c06\u76f8\u5173\u4e8b\u5b9e\u5d4c\u5165\u5230\u6a21\u578b\u7684\u5185\u90e8\u7ed3\u6784\u4e2d\uff0c\u786e\u4fdd\u5168\u9762\u5438\u6536\u3002\u867d\u7136\u4f20\u7edf\u7684 fine-tuning \u53ef\u80fd\u4f1a\u6d88\u8017\u5927\u91cf\u8d44\u6e90\uff0c\u4f46\u5148\u8fdb\u7684\u6280\u672f\u53ef\u4ee5\u51cf\u8f7b\u8fd9\u79cd\u8d1f\u62c5\u3002\u7ea6\u675f\u5fae\u8c03\u548c\u5143\u5b66\u4e60\u6700\u5c0f\u5316\u77e5\u8bc6\u635f\u5931\u548c\u8fc7\u62df\u5408\u98ce\u9669\u3002\u53ef\u7f16\u8f91\u8bad\u7ec3\u548c KnowledgeEditor \u5b9e\u73b0\u53c2\u6570\u5feb\u901f\u8c03\u6574\uff0c\u540c\u65f6\u5c06\u6027\u80fd\u5f71\u54cd\u964d\u5230\u6700\u4f4e\uff0c\u800c SLAG\u3001MEND \u548c MALMEN \u89e3\u51b3\u4e86\u7f16\u8f91\u51b2\u7a81\u5e76\u652f\u6301\u5927\u89c4\u6a21\u66f4\u65b0\uff0c\u5728\u6574\u5408\u65b0\u9886\u57df\u89c1\u89e3\u7684\u540c\u65f6\u4fdd\u6301\u57fa\u672c\u80fd\u529b\u3002 LLM \u624b\u672f \u901a\u8fc7\u5e94\u7528\u9006\u68af\u5ea6\u53bb\u9664\u8fc7\u65f6\u6570\u636e\u3001\u68af\u5ea6\u4e0b\u964d\u6574\u5408\u65b0\u4e8b\u5b9e\u4ee5\u53ca KL \u6563\u5ea6\u9879\u6765\u4fdd\u7559\u73b0\u6709\u77e5\u8bc6\uff0c\u7edf\u4e00\u4e86\u9057\u5fd8\u548c\u7f16\u8f91\uff0c\u5b9e\u73b0\u4e86\u663e\u8457\u7684\u8ba1\u7b97\u6548\u7387\u3002 KNE \u5f15\u5165\u4e86\u4e00\u79cd\u77e5\u8bc6\u795e\u7ecf\u5143\u96c6\u6210\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u4ec5\u786e\u5b9a\u5e76\u66f4\u65b0\u4e0e\u65b0\u63d2\u5165\u4e8b\u5b9e\u5bc6\u5207\u76f8\u5173\u7684\u795e\u7ecf\u5143\uff0c\u4ece\u800c\u5728\u4fdd\u7559\u4e0d\u76f8\u5173\u77e5\u8bc6\u7684\u540c\u65f6\u5b9e\u73b0\u66f4\u51c6\u786e\u7684\u7f16\u8f91\u3002 OVERTONE \u901a\u8fc7\u5f15\u5165\u4e00\u79cd token \u7ea7\u5e73\u6ed1\u6280\u672f\u6765\u89e3\u51b3\u77e5\u8bc6\u7f16\u8f91\u4e2d\u5f02\u6784 token \u8fc7\u62df\u5408\u95ee\u9898\uff0c\u8be5\u6280\u672f\u81ea\u9002\u5e94\u5730\u4f18\u5316\u8bad\u7ec3\u76ee\u6807\uff0c\u4ece\u800c\u4fdd\u7559\u9884\u8bad\u7ec3\u77e5\u8bc6\u5e76\u63d0\u9ad8\u6a21\u578b\u5bf9\u65b0\u63d2\u5165\u4e8b\u5b9e\u7684\u63a8\u7406\u80fd\u529b\u3002 \u8fd9\u4e9b\u6709\u9488\u5bf9\u6027\u7684\u6280\u672f\u786e\u4fdd\u6a21\u578b\u5728\u6574\u5408\u65b0\u83b7\u53d6\u4fe1\u606f\u7684\u540c\u65f6\u4fdd\u7559\u5176\u57fa\u672c\u80fd\u529b\u3002<\/p>\n<h5>\u68c0\u7d22\u589e\u5f3a\u751f\u6210<\/h5>\n<p>\u68c0\u7d22\u589e\u5f3a\u751f\u6210 (Retrieval-Augmented Generation,RAG) \u5c06\u4f20\u7edf\u4fe1\u606f\u68c0\u7d22\u4e0e\u5f53\u4ee3LLM\u96c6\u6210\uff0c\u4ee5\u589e\u5f3a\u751f\u6210\u8f93\u51fa\u7684\u76f8\u5173\u6027\u548c\u4e8b\u5b9e\u51c6\u786e\u6027\u3002\u901a\u8fc7\u4ece\u5916\u90e8\u6765\u6e90\u52a8\u6001\u68c0\u7d22\u76f8\u5173\u4fe1\u606f\u5e76\u5c06\u5176\u5d4c\u5165\u5230\u751f\u6210\u8fc7\u7a0b\u4e2d\uff0cRAG \u89e3\u51b3\u4e86 LLM \u7279\u5b9a\u9886\u57df\u77e5\u8bc6\u7684\u7f3a\u9677\uff0c\u5e76\u964d\u4f4e\u4e86\u51fa\u73b0\u5e7b\u89c9\u5185\u5bb9\u7684\u503e\u5411\u3002 \u8fd9\u79cd\u65b9\u6cd5\u5728\u9700\u8981\u7cbe\u786e\u3001\u6700\u65b0\u4fe1\u606f\u7684\u9886\u57df\u7279\u522b\u6709\u6548\uff0c\u4f8b\u5982\u95ee\u7b54\u7cfb\u7edf\u3001\u79d1\u5b66\u7814\u7a76\u548c\u533b\u7597\u4fdd\u5065\uff0c\u5728\u8fd9\u4e9b\u9886\u57df\uff0c\u5b83\u80fd\u591f\u719f\u7ec3\u5730\u5904\u7406\u590d\u6742\u67e5\u8be2\u548c\u77e5\u8bc6\u5bc6\u96c6\u578b\u4efb\u52a1\u3002 \u6b64\u5916\uff0cRAG \u51cf\u8f7b\u4e86\u5bf9\u8bdd\u7cfb\u7edf\u4e2d\u8bef\u5bfc\u6027\u54cd\u5e94\u7684\u666e\u904d\u6027\uff0c\u63d0\u9ad8\u4e86\u77e5\u8bc6\u9a71\u52a8\u7684\u81ea\u7136\u8bed\u8a00\u751f\u6210\u7684\u4fdd\u771f\u5ea6\u3002<br \/>\n\u672c\u5c0f\u8282\u4fa7\u91cd\u4e8e\u57fa\u4e8e\u8bad\u7ec3\u7684 RAG \u65b9\u6cd5 \uff0c\u8ba4\u8bc6\u5230\u65e0\u8bad\u7ec3\u7684 RAG \u65b9\u6cd5\u53ef\u80fd\u4f1a\u7531\u4e8e\u7f3a\u4e4f\u7279\u5b9a\u4efb\u52a1\u7684\u4f18\u5316\u800c\u635f\u5bb3\u77e5\u8bc6\u5229\u7528\u6548\u7387\u3002 \u4e09\u79cd\u4e3b\u8981\u7684\u8bad\u7ec3\u7b56\u7565\u2014\u2014\u72ec\u7acb\u8bad\u7ec3(Independent Training)\u3001\u987a\u5e8f\u8bad\u7ec3(Sequential Training)\u548c\u8054\u5408\u8bad\u7ec3(Sequential Training)\u2014\u2014\u589e\u5f3a\u4e86\u6a21\u578b\u9002\u5e94\u6027\u548c\u96c6\u6210\u80fd\u529b\uff0c\u5982\u56fe\u6240\u793a:<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251106085848204.png\" alt=\"\" \/><\/p>\n<p><center>\u68c0\u7d22\u589e\u5f3a\u751f\u6210 (RAG) \u8bad\u7ec3\u65b9\u6cd5\u7684\u5206\u7c7b\uff0c\u5bf9\u72ec\u7acb\u8bad\u7ec3\u3001\u987a\u5e8f\u8bad\u7ec3\u548c\u8054\u5408\u8bad\u7ec3\u7b56\u7565\u8fdb\u884c\u5206\u7c7b<\/center><\/p>\n<ul>\n<li><strong>\u72ec\u7acb\u8bad\u7ec3:<\/strong>\u8fd9\u79cd\u7b56\u7565\u5c06\u68c0\u7d22\u5668\u548c\u751f\u6210\u5668\u8bad\u7ec3\u6210\u4e0d\u540c\u7684\u6a21\u5757\uff0c\u4ece\u800c\u5728\u4f7f\u7528\u9488\u5bf9\u4efb\u52a1\u9700\u6c42\u91cf\u8eab\u5b9a\u5236\u7684\u7a00\u758f\u6216\u5bc6\u96c6\u68c0\u7d22\u5668\u65f6\u5177\u6709\u7075\u6d3b\u6027\u3002 \u4f8b\u5982\uff0cDPR \u4f7f\u7528\u53cc BERT \u7f51\u7edc\u5206\u522b\u5bf9\u67e5\u8be2\u548c\u6bb5\u843d\u8fdb\u884c\u7f16\u7801\uff0c\u5e94\u7528\u5bf9\u6bd4\u5b66\u4e60\u6765\u4f18\u5316\u68c0\u7d22\uff0c\u800c\u65e0\u9700\u751f\u6210\u5668\u7684\u4ea4\u4e92\u3002 \u540c\u6837\uff0c\u63d0\u51fa\u4e86 Reward-RAG\uff0c\u5b83\u5229\u7528\u5956\u52b1\u6a21\u578b\u4ec5\u6839\u636e\u57fa\u4e8e GPT \u7684\u53cd\u9988\u6765\u5fae\u8c03\u68c0\u7d22\u5668\uff0c\u800c\u751f\u6210\u5668\u4fdd\u6301\u4e0d\u53d8\u3002<\/li>\n<li><strong>\u987a\u5e8f\u8bad\u7ec3:<\/strong>\u987a\u5e8f\u8bad\u7ec3\u901a\u8fc7\u4e00\u6b21\u4f18\u5316\u4e00\u4e2a\u6a21\u5757\u6765\u63d0\u9ad8\u6548\u7387\uff0c\u4ece\u800c\u4fc3\u8fdb\u68c0\u7d22\u5668\u548c\u751f\u6210\u5668\u4e4b\u95f4\u7684\u534f\u540c\u4f5c\u7528\u3002\u5b83\u5305\u62ec Retriever-First \u65b9\u6cd5\uff0c\u5982 RETRO\uff0c\u5b83\u5728\u8bad\u7ec3\u7f16\u7801\u5668-\u89e3\u7801\u5668\u4e4b\u524d\u9884\u5148\u8bad\u7ec3\u57fa\u4e8e BERT \u7684\u68c0\u7d22\u5668\uff0c\u4ee5\u65e0\u7f1d\u96c6\u6210\u68c0\u7d22\u5230\u7684\u5185\u5bb9\uff0c\u4ece\u800c\u63d0\u9ad8\u6027\u80fd\u3002 \u6216\u8005\uff0cLLM-First \u65b9\u6cd5\uff0c\u5982 RA-DIT\uff0c\u9996\u5148\u5fae\u8c03\u8bed\u8a00\u6a21\u578b\u4ee5\u6709\u6548\u5730\u5229\u7528\u68c0\u7d22\u5230\u7684\u77e5\u8bc6\uff0c\u7136\u540e\u6539\u8fdb\u68c0\u7d22\u5668\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u5bf9\u9f50\u548c\u4e00\u81f4\u6027\u3002<\/li>\n<li><strong>\u8054\u5408\u8bad\u7ec3:<\/strong>\u8054\u5408\u8bad\u7ec3\u5728\u7aef\u5230\u7aef\u6846\u67b6\u4e2d\u540c\u6b65\u68c0\u7d22\u5668\u548c\u751f\u6210\u5668\u4f18\u5316\u3002RAG \u6700\u5c0f\u5316\u8d1f\u5bf9\u6570\u4f3c\u7136\u4ee5\u5171\u540c\u8bad\u7ec3\u8fd9\u4e24\u4e2a\u7ec4\u4ef6\uff0c\u800c REALM \u901a\u8fc7\u6700\u5927\u5185\u79ef\u641c\u7d22 (MIPS) \u63d0\u9ad8\u4e86\u68c0\u7d22\u7cbe\u5ea6\u3002\u8fd9\u4e9b\u65b9\u6cd5\u9002\u5e94\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u9700\u6c42\uff0c\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u5916\u90e8\u77e5\u8bc6\u7684\u76ca\u5904\uff0c\u5e76\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u751f\u6210\u9519\u8bef\u3002<\/li>\n<\/ul>\n<h4>\u6a21\u578b\u5408\u5e76<\/h4>\n<p>\u6a21\u578b\u5408\u5e76\u5df2\u6210\u4e3a\u4e00\u79cd\u91cd\u8981\u7684\u8bad\u7ec3\u540e\u7b56\u7565\uff0c\u7528\u4e8e\u63d0\u9ad8 LLM \u5728\u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\u7684\u6027\u80fd\u548c\u6548\u7387\u3002 \u8fd9\u79cd\u65b9\u6cd5\u5c06\u4e13\u7528\u6a21\u578b\u6574\u5408\u5230\u4e00\u4e2a\u7edf\u4e00\u7684\u67b6\u6784\u4e2d\uff0c\u907f\u514d\u4e86\u5927\u91cf\u91cd\u65b0\u8bad\u7ec3\u7684\u9700\u6c42\uff0c\u5e76\u89e3\u51b3\u4e86\u5927\u578b\u6a21\u578b\u89c4\u6a21\u548c\u8ba1\u7b97\u9700\u6c42\u5e26\u6765\u7684\u6311\u6218\u3002 \u4e0e\u5728\u5408\u5e76\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u4e0d\u540c\uff0c\u6a21\u578b\u5408\u5e76\u5c06\u5355\u4efb\u52a1\u6a21\u578b\u96c6\u6210\u5230\u80fd\u591f\u80dc\u4efb\u591a\u4efb\u52a1\u7684\u5185\u805a\u5b9e\u4f53\u4e2d\uff0c\u4e3a\u591a\u4efb\u52a1\u5b66\u4e60\u63d0\u4f9b\u4e86\u4e00\u79cd\u8d44\u6e90\u9ad8\u6548\u7684\u8303\u4f8b\u3002 \u901a\u8fc7\u7b80\u5316\u8bad\u7ec3\u6d41\u7a0b\u5e76\u4fc3\u8fdb\u5f00\u53d1\u5177\u6709\u5f3a\u5927\u8de8\u5e94\u7528\u7a0b\u5e8f\u6cdb\u5316\u7684\u901a\u7528\u6a21\u578b\uff0c\u8be5\u6280\u672f\u4f18\u5316\u4e86 LLM \u5728\u4e0d\u540c\u73af\u5883\u4e2d\u7684\u90e8\u7f72\u3002 \u7ed9\u5b9a\u4e00\u7ec4\u5019\u9009\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">M={M_1,M_2,\\dots,M_n}<\/span>\uff0c\u76ee\u6807\u662f\u8bbe\u8ba1\u4e00\u4e2a\u5408\u5e76\u51fd\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">F_{\\text{merge}}<\/span>\uff0c\u8be5\u51fd\u6570\u4ea7\u751f\u4e00\u4e2a\u7edf\u4e00\u7684\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">M&#039;<\/span>\uff0c\u53ef\u80fd\u7531\u4e00\u4e2a\u57fa\u7840\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">M_1<\/span> \u951a\u5b9a\uff0c\u5982\u56fe\u6240\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251106090526824.png\" alt=\"\" \/><\/p>\n<h5>\u5c42\u6b21\u7ea7\u522b\u4e0a\u7684\u6a21\u578b\u5408\u5e76<\/h5>\n<p>\u6a21\u578b\u5408\u5e76\u6280\u672f\u88ab\u7cfb\u7edf\u5730\u5206\u4e3a\u4e09\u4e2a\u5c42\u6b21\u7ea7\u522b\u2014\u2014\u6743\u91cd\u7ea7\u3001\u8f93\u51fa\u7ea7\u548c\u6a21\u578b\u7ea7\u5408\u5e76\u2014\u2014\u5982\u56fe\u6240\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251106090638467.png\" alt=\"\" \/><\/p>\n<p><center>\u6a21\u578b\u5408\u5e76\u6280\u672f\u7684\u5206\u7c7b\uff0c\u63cf\u8ff0\u4e86\u5305\u62ec\u6743\u91cd\u7ea7\u3001\u8f93\u51fa\u7ea7\u548c\u6a21\u578b\u7ea7\u65b9\u6cd5\u5728\u5185 \u7684\u5c42\u6b21\u7ea7\u522b\uff0c\u9002\u7528\u4e8e\u5927\u578b\u8bed\u8a00\u6a21\u578b<\/center><strong>\u6743\u91cd\u7ea7\u6a21\u578b\u5408\u5e76:<\/strong>\u6743\u91cd\u7ea7\u5408\u5e76\u76f4\u63a5\u64cd\u4f5c\u53c2\u6570\u7a7a\u95f4\uff0c\u4f7f\u5176\u5bf9\u5177\u6709\u67b6\u6784\u76f8\u4f3c\u6027\u6216\u5728\u76f8\u5173\u4efb\u52a1\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u7279\u522b\u6709\u6548\u3002 \u5f62\u5f0f\u4e0a\uff0c\u7ed9\u5b9a\u53c2\u6570\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\theta_1,\\theta_2,\\dots,\\theta_n\\in\\mathbb{R}^d<\/span>\uff0c\u4e00\u4e2a\u7ebf\u6027\u5408\u5e76\u65b9\u6848\u5c06\u8fd9\u4e9b\u805a\u5408\u5230\u4e00\u4e2a\u7edf\u4e00\u7684\u96c6\u5408 <span class=\"katex-eq\" data-katex-display=\"false\">\\theta&#039;<\/span>\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\theta&gt;[katex]\\theta&#039;=\\alpha_1\\theta_1+\\alpha_2\\theta_2+\\dots+\\alpha_n\\theta_n,\\quad\\text{subject to}\\ \\alpha_k\\ge0,\\ \\sum_{k=1}^{n}\\alpha_k=1.<\/span><\/center>Model Soup \u901a\u8fc7\u7ebf\u6027\u7ec4\u5408\u6765\u81ea\u5728\u4e0d\u540c\u4efb\u52a1\u4e0a\u5fae\u8c03\u7684\u6a21\u578b\u7684\u6743\u91cd\u6765\u4f8b\u8bc1\u8fd9\u4e00\u70b9\uff0c\u4ece\u800c\u4ea7\u751f\u4e00\u4e2a\u5355\u4e00\u4e14\u9ad8\u6548\u7684\u6a21\u578b\u3002 \u4efb\u52a1\u7b97\u672f (Task Arithmetic, TA) \u901a\u8fc7\u5bf9\u53c2\u6570\u8fdb\u884c\u7b97\u672f\u8fd0\u7b97\u6765\u6269\u5c55\u8fd9\u79cd\u7075\u6d3b\u6027\uff0c\u4ece\u800c\u589e\u5f3a\u6027\u80fd\u9002\u5e94\u6027\u3002 \u4e3a\u4e86\u51cf\u8f7b\u5bf9\u9f50\u95ee\u9898\uff0cTIES-merging \u786e\u4fdd\u53c2\u6570\u4e00\u81f4\u6027\uff0c\u800c DARE \u901a\u8fc7\u6982\u7387\u6027\u5730\u8c03\u6574\u53c2\u6570\u589e\u91cf\u6765\u6700\u5c0f\u5316\u5e72\u6270\uff0c\u4ece\u800c\u4f18\u5316\u5408\u5e76\u8fc7\u7a0b\u4ee5\u5b9e\u73b0\u4e00\u81f4\u6027\u548c\u6548\u7387\u3002<\/p>\n<p><strong>\u8f93\u51fa\u7ea7\u6a21\u578b\u5408\u5e76:<\/strong>\u5f53\u6a21\u578b\u5728\u67b6\u6784\u6216\u521d\u59cb\u5316\u65b9\u9762\u51fa\u73b0\u5206\u6b67\u65f6\uff0c\u8f93\u51fa\u7ea7\u5408\u5e76\u53d8\u5f97\u5177\u6709\u4f18\u52bf\uff0c\u8fd9\u4f7f\u5f97\u6743\u91cd\u7ea7\u65b9\u6cd5\u53d8\u5f97\u4e0d\u5207\u5b9e\u9645\u3002 \u8fd9\u79cd\u65b9\u6cd5\u805a\u5408\u8f93\u51fa\u5206\u5e03\uff0c\u800c\u4e0d\u662f\u5185\u90e8\u53c2\u6570\uff0c\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">y&gt;[katex]y&#039;=\\alpha,y_1+(1-\\alpha),y_2,\\quad\\alpha\\in[0,1],<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">y_1<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">y_2<\/span> \u5206\u522b\u8868\u793a\u6765\u81ea\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">M_1<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">M_2<\/span> \u7684\u6982\u7387\u5206\u5e03\u3002\u7c7b\u4f3c\u4e8e\u96c6\u6210\u7b56\u7565\uff0c\u8be5\u65b9\u6cd5\u5c06\u6a21\u578b\u9884\u6d4b\u5408\u6210\u4e3a\u4e00\u4e2a\u7edf\u4e00\u7684\u8f93\u51fa\u3002LLMBlender \u901a\u8fc7\u751f\u6210\u72ec\u7acb\u8f93\u51fa\u5e76\u901a\u8fc7\u6392\u540d\u548c\u751f\u6210\u8fc7\u7a0b\u878d\u5408\u5b83\u4eec\u6765\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u800c FuseLLM \u5c06\u7ec4\u5408\u8f93\u51fa\u6982\u7387\u63d0\u70bc\u5230\u5355\u4e2a\u7f51\u7edc\u4e2d\u4ee5\u5b9e\u73b0\u5206\u5e03\u4fdd\u771f\u5ea6\u3002FuseChat \u901a\u8fc7\u5c06\u77e5\u8bc6\u4ece\u591a\u4e2a LLM \u8f6c\u79fb\u5230\u6574\u5408\u7684\u76ee\u6807\u4e2d\uff0c\u6865\u63a5\u6743\u91cd\u548c\u8f93\u51fa\u7ea7\u522b\u7684\u5408\u5e76\uff0c\u589e\u5f3a\u8de8\u6a21\u578b\u534f\u540c\u4f5c\u7528\u3002<\/p>\n<p><strong>\u6a21\u578b\u7ea7\u6a21\u578b\u5408\u5e76:<\/strong>\u6a21\u578b\u7ea7\u5408\u5e76\u901a\u8fc7\u8def\u7531\u673a\u5236\u6574\u5408\u5b50\u6a21\u578b\u6216\u5c42\uff0c\u901a\u5e38\u5728\u6df7\u5408\u4e13\u5bb6 (MoE) \u6846\u67b6\u5185\u8fdb\u884c\uff0c\u8868\u793a\u4e3a\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">M&gt;[katex]M&#039;=\\text{Merge}(M_1,M_2),<\/span><\/center>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\text{Merge}<\/span> \u8868\u793a\u786c\u6216\u8f6f\u8def\u7531\u51fd\u6570\u3002 Switch Transformer \u91c7\u7528\u79bb\u6563\u95e8\u63a7\u9009\u62e9\u6027\u5730\u6fc0\u6d3b\u4e13\u5bb6\u5c42\uff0c\u51cf\u5c11\u8ba1\u7b97\u8d1f\u8f7d\uff0c\u5c3d\u7ba1\u7531\u4e8e\u521a\u6027\u8def\u7531\u53ef\u80fd\u5b58\u5728\u6027\u80fd\u6743\u8861\u3002 SoftMoE \u548c SMEAR \u5229\u7528\u8fde\u7eed\u95e8\u63a7\u6765\u4fc3\u8fdb\u4e13\u5bb6\u4e4b\u95f4\u7684\u66f4\u5e73\u6ed1\u8fc7\u6e21\uff0c\u589e\u5f3a\u7ec4\u4ef6\u96c6\u6210\u548c\u6a21\u578b\u5185\u805a\u529b\u3002<\/p>\n<h5>\u9884\u5408\u5e76\u65b9\u6cd5<\/h5>\n<p>\u9884\u5408\u5e76\u65b9\u6cd5\u901a\u8fc7\u4f18\u5316\u72ec\u7acb\u6a21\u578b\u7684\u6743\u91cd\u7a7a\u95f4\u3001\u67b6\u6784\u4e00\u81f4\u6027\u548c\u53c2\u6570\u5bf9\u9f50\uff0c\u4e3a\u6a21\u578b\u5408\u5e76\u5960\u5b9a\u517c\u5bb9\u6027\u57fa\u7840\uff0c\u4ece\u800c\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u540e\u7eed\u878d\u5408\u9636\u6bb5\u7684\u51b2\u7a81\u548c\u5e72\u6270\u3002\u8fd9\u4e9b\u6280\u672f\u63d0\u9ad8\u4e86\u5408\u5e76\u8fc7\u7a0b\u7684\u6548\u7387\uff0c\u786e\u4fdd\u751f\u6210\u7684\u7edf\u4e00\u6a21\u578b\u4fdd\u7559\u5176\u7ec4\u6210\u90e8\u5206\u7684\u4f18\u52bf\uff0c\u540c\u65f6\u51cf\u8f7b\u6f5c\u5728\u7684\u9000\u5316\u3002<\/p>\n<ul>\n<li><strong>\u7ebf\u6027\u5316\u5fae\u8c03\uff1a<\/strong>\u8fd9\u79cd\u65b9\u6cd5\u5728\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u5207\u7ebf\u7a7a\u95f4\u5185\u4f18\u5316\u6a21\u578b\uff0c\u907f\u514d\u539f\u59cb\u7684\u975e\u7ebf\u6027\u53c2\u6570\u7a7a\u95f4\u4ee5\u5b9e\u73b0\u6743\u91cd\u89e3\u8026\uff0c\u4ece\u800c\u51cf\u5c11\u5408\u5e76\u671f\u95f4\u7684\u5e72\u6270\u3002 \u90e8\u5206\u7ebf\u6027\u5316\u9002\u914d\u5668\uff08\u4f8b\u5982\uff0cTAFT\uff09\u6216\u6ce8\u610f\u529b\u5c42\u7b49\u6280\u672f\u5c06\u6743\u91cd\u66f4\u65b0\u4e0e\u4e0d\u76f8\u4ea4\u7684\u8f93\u5165\u533a\u57df\u5bf9\u9f50\uff0c\u4fdd\u7559\u5408\u5e76\u6a21\u578b\u4e2d\u7684\u72ec\u7acb\u529f\u80fd\u3002 \u901a\u8fc7\u5c06\u66f4\u65b0\u9650\u5236\u5728\u7ebf\u6027\u5316\u6846\u67b6\u4e2d\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4fc3\u8fdb\u4e86\u8de8\u4e0d\u540c\u6a21\u578b\u7684\u65e0\u7f1d\u96c6\u6210\u3002<\/li>\n<li><strong>\u67b6\u6784\u8f6c\u6362\uff1a<\/strong>\u6b64\u7b56\u7565\u5c06\u5177\u6709\u4e0d\u540c\u67b6\u6784\u7684\u5f02\u6784\u6a21\u578b\u8f6c\u6362\u4e3a\u9002\u5408\u76f4\u63a5\u53c2\u6570\u5408\u5e76\u7684\u540c\u6784\u5f62\u5f0f\u3002 \u65b9\u6cd5\u5305\u62ec\u77e5\u8bc6\u84b8\u998f\uff0c\u5982 FuseChat \u6240\u793a\uff0c\u4ee5\u53ca\u8eab\u4efd\u5c42\u63d2\u5165\uff0c\u5982 CLAFusion \u6240\u793a\u3002 GAN Cocktail \u521d\u59cb\u5316\u76ee\u6807\u6a21\u578b\u4ee5\u5438\u6536\u6765\u81ea\u4e0d\u540c\u67b6\u6784\u7684\u8f93\u51fa\uff0c\u4ece\u800c\u5b9e\u73b0\u7edf\u4e00\u7684\u5408\u5e76\u8fc7\u7a0b\uff0c\u6709\u6548\u5730\u5f25\u5408\u7ed3\u6784\u5dee\u5f02\u3002<\/li>\n<li><strong>\u6743\u91cd\u5bf9\u9f50\uff1a<\/strong>\u8fd9\u79cd\u65b9\u6cd5\u901a\u8fc7\u6392\u5217\u5c06\u6a21\u578b\u5bf9\u9f50\u5230\u5171\u4eab\u7684\u6743\u91cd\u6c60\u4e2d\uff0c\u5229\u7528\u7ebf\u6027\u6a21\u5f0f\u8fde\u63a5 (LMC) \u5c5e\u6027\u6765\u589e\u5f3a\u517c\u5bb9\u6027\u3002 \u6280\u672f\u5305\u62ec\u6700\u4f18\u4f20\u8f93 (OTFusion)\uff0c\u542f\u53d1\u5f0f\u5339\u914d (Git re-basin) \u548c\u57fa\u4e8e\u5b66\u4e60\u7684\u5bf9\u9f50 (Deep-Align)\u3002 REPAIR \u51cf\u8f7b\u4e86\u7f3a\u4e4f\u6807\u51c6\u5316\u5c42\u7684\u6a21\u578b\u4e2d\u7684\u5bf9\u9f50\u5931\u8d25\uff0c\u786e\u4fdd\u5728\u878d\u5408\u4e4b\u524d\u8fdb\u884c\u7a33\u5065\u7684\u53c2\u6570\u6536\u655b\u3002<\/li>\n<\/ul>\n<h5>\u878d\u5408\u671f\u95f4\u7684\u65b9\u6cd5<\/h5>\n<p>\u878d\u5408\u671f\u95f4\u7684\u65b9\u6cd5\u4fa7\u91cd\u4e8e\u52a8\u6001\u4f18\u5316\u53c2\u6570\u878d\u5408\u7b56\u7565\uff0c\u4ee5\u89e3\u51b3\u4efb\u52a1\u51b2\u7a81\uff0c\u51cf\u8f7b\u5e72\u6270\uff0c\u5e76\u63d0\u9ad8\u6240\u5f97\u5408\u5e76\u6a21\u578b\u7684\u6027\u80fd\u548c\u6cdb\u5316\u80fd\u529b\u3002 \u8fd9\u4e9b\u65b9\u6cd5\u89e3\u51b3\u4e86\u5b9e\u65f6\u6574\u5408\u4e0d\u540c\u6a21\u578b\u7684\u6311\u6218\uff0c\u589e\u5f3a\u4e86\u7edf\u4e00\u67b6\u6784\u7684\u9002\u5e94\u6027\u548c\u7a33\u5065\u6027\u3002<\/p>\n<p><strong>\u57fa\u672c\u5408\u5e76(Basic Merging):<\/strong>\u8fd9\u79cd\u65b9\u6cd5\u5229\u7528\u76f4\u63a5\u7684\u53c2\u6570\u5e73\u5747\u6216\u4efb\u52a1\u5411\u91cf\u7b97\u672f\uff0c\u5c06\u4efb\u52a1\u5411\u91cf <span class=\"katex-eq\" data-katex-display=\"false\">\\tau_t<\/span> \u5b9a\u4e49\u4e3a\u5fae\u8c03\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">\\Theta^{(t)}<\/span> \u4e0e\u7b2c <span class=\"katex-eq\" data-katex-display=\"false\">t<\/span> \u4e2a\u4efb\u52a1\u7684\u521d\u59cb\u9884\u8bad\u7ec3\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">\\Theta^{(0)}<\/span> \u4e4b\u95f4\u7684\u504f\u5dee\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\tau_t&gt;[katex]\\tau_t=\\Theta^{(t)}-\\Theta^{(0)},<\/span><\/center>\u5e76\u4e14\u901a\u8fc7\u516c\u5f0f <span class=\"katex-eq\" data-katex-display=\"false\">\\Theta^{(\\text{merge})}=\\Theta^{(0)}+\\lambda\\sum_{t=1}^{T}\\tau_t<\/span> \u4fc3\u8fdb\u591a\u4efb\u52a1\u5b66\u4e60\u3002 \u867d\u7136\u8ba1\u7b97\u6548\u7387\u9ad8\u4e14\u6982\u5ff5\u4f18\u96c5\uff0c\u4f46\u8fd9\u79cd\u65b9\u6cd5\u7ecf\u5e38\u4f1a\u9047\u5230\u7531\u4e8e\u672a\u7ecf\u7f13\u89e3\u7684\u53c2\u6570\u4ea4\u4e92\u800c\u4ea7\u751f\u7684\u4efb\u52a1\u5e72\u6270\uff0c\u4ece\u800c\u9650\u5236\u4e86\u5176\u5728\u9700\u8981\u590d\u6742\u4efb\u52a1\u534f\u8c03\u7684\u573a\u666f\u4e2d\u7684\u5b9e\u7528\u6027\u3002<br \/>\n<strong>\u52a0\u6743\u5408\u5e76(Weighted Merging):<\/strong>\u8fd9\u79cd\u7b56\u7565\u6839\u636e\u5404\u4e2a\u6a21\u578b\u7684\u91cd\u8981\u6027\u52a8\u6001\u5206\u914d\u5408\u5e76\u7cfb\u6570\uff0c\u5b9a\u5236\u8d21\u732e\u4ee5\u4f18\u5316\u878d\u5408\u7ed3\u679c\u3002 MetaGPT \u901a\u8fc7\u5bf9\u6bcf\u4e2a\u4efb\u52a1\u5411\u91cf\u7684\u5e73\u65b9 L2 \u8303\u6570\u8fdb\u884c\u5f52\u4e00\u5316\u6765\u8ba1\u7b97\u6700\u4f73\u6743\u91cd\uff1a<\/p>\n<p><center><span class=\"katex-eq\" data-katex-display=\"false\">\\lambda_t^*&gt;[katex]\\lambda_t^*=\\frac{|\\tau_t|2}{\\sum{k=1}^{T}|\\tau_k|_2},<\/span><\/center>\u4ece\u800c\u5c06\u66f4\u5927\u7684\u5f71\u54cd\u5206\u914d\u7ed9\u5177\u6709\u66f4\u5b9e\u8d28\u6027\u53c2\u6570\u53d8\u5316\u7684\u4efb\u52a1\uff0c\u5982\u8f83\u9ad8\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">|\\tau_t|_2<\/span> \u6240\u793a\u3002SLERP \u91c7\u7528\u7403\u9762\u63d2\u503c\u6765\u786e\u4fdd\u5e73\u6ed1\u7684\u53c2\u6570\u8f6c\u6362\uff0c\u4fdd\u6301\u6a21\u578b\u8fde\u7eed\u6027\uff0c\u800c\u9010\u5c42 AdaMerging \u901a\u8fc7\u4f18\u5316\u6bcf\u5c42\u7c92\u5ea6\u7684\u7cfb\u6570\u6765\u5b8c\u5584\u6b64\u8fc7\u7a0b\uff0c\u589e\u5f3a\u4e86\u5408\u5e76\u67b6\u6784\u4e2d\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u7cbe\u5ea6\u3002<br \/>\n<strong>\u5b50\u7a7a\u95f4\u5408\u5e76(Subspace Merging):<\/strong>\u8fd9\u79cd\u65b9\u6cd5\u5c06\u6a21\u578b\u53c2\u6570\u6295\u5f71\u5230\u7a00\u758f\u5b50\u7a7a\u95f4\u4e2d\uff0c\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u5e72\u6270\uff0c\u540c\u65f6\u4fdd\u6301\u8ba1\u7b97\u6548\u7387\uff0c\u89e3\u51b3\u53c2\u6570\u8d21\u732e\u7684\u91cd\u53e0\u95ee\u9898\u3002TIES-Merging \u4fdd\u7559\u5e45\u5ea6\u6700\u5927\u7684\u524d 20% \u7684\u53c2\u6570\uff0c\u89e3\u51b3\u7b26\u53f7\u51b2\u7a81\u4ee5\u4fdd\u6301\u4e00\u81f4\u6027\uff0cDARE \u7f29\u653e\u7a00\u758f\u6743\u91cd\u4ee5\u51cf\u5c11\u5197\u4f59\uff0c\u800c Concrete \u5229\u7528\u53cc\u5c42\u4f18\u5316\u6765\u6784\u5efa\u81ea\u9002\u5e94\u63a9\u7801\uff0c\u786e\u4fdd\u6a21\u578b\u7ec4\u4ef6\u7684\u7cbe\u7ec6\u96c6\u6210\uff0c\u51cf\u5c11\u8de8\u4efb\u52a1\u7684\u5e72\u6270\u3002<br \/>\n<strong>\u57fa\u4e8e\u8def\u7531\u7684\u5408\u5e76(Routing-based Merging):<\/strong>\u8be5\u6280\u672f\u6839\u636e\u7279\u5b9a\u4e8e\u8f93\u5165\u5c5e\u6027\u52a8\u6001\u878d\u5408\u6a21\u578b\uff0c\u5b9e\u73b0\u4e0a\u4e0b\u6587\u54cd\u5e94\u5f0f\u96c6\u6210\u8fc7\u7a0b\u3002SMEAR \u8ba1\u7b97\u4f9d\u8d56\u4e8e\u6837\u672c\u7684\u4e13\u5bb6\u6743\u91cd\u4ee5\u4f18\u5148\u8003\u8651\u76f8\u5173\u7279\u5f81\uff0cWeight-Ensembling MoE \u91c7\u7528\u8f93\u5165\u9a71\u52a8\u7684\u7ebf\u6027\u5c42\u8def\u7531\u8fdb\u884c\u9009\u62e9\u6027\u6fc0\u6d3b\uff0c\u800c Twin-Merging \u878d\u5408\u4efb\u52a1\u5171\u4eab\u548c\u4efb\u52a1\u79c1\u6709\u77e5\u8bc6\uff0c\u4ece\u800c\u5efa\u7acb\u4e00\u4e2a\u7075\u6d3b\u7684\u5408\u5e76\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u9002\u5e94\u4e0d\u540c\u7684\u8f93\u5165\u9700\u6c42\u5e76\u589e\u5f3a\u591a\u4efb\u52a1\u9c81\u68d2\u6027\u3002<br \/>\n<strong>\u540e\u6821\u51c6(Post-calibration):<\/strong>\u8be5\u6280\u672f\u901a\u8fc7\u5c06\u7edf\u4e00\u6a21\u578b\u7684\u9690\u85cf\u8868\u793a\u4e0e\u72ec\u7acb\u7ec4\u6210\u90e8\u5206\u7684\u9690\u85cf\u8868\u793a\u5bf9\u9f50\u6765\u7ea0\u6b63\u5408\u5e76\u540e\u7684\u8868\u793a\u504f\u5dee\uff0c\u4ece\u800c\u51cf\u8f7b\u6027\u80fd\u4e0b\u964d\u3002 \u8868\u793a\u624b\u672f \u901a\u8fc7\u6539\u8fdb\u8868\u793a\u4e00\u81f4\u6027\u6765\u4f8b\u8bc1\u8fd9\u4e00\u70b9\uff0c\u4ece\u800c\u589e\u5f3a\u4e86\u5408\u5e76\u6a21\u578b\u7684\u9c81\u68d2\u6027\u548c\u51c6\u786e\u6027\u3002<\/p>\n<h3>\u6570\u636e\u96c6<\/h3>\n<p>\u7ecf\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u540e\u8bad\u7ec3\u6280\u672f\uff0c\u4ee5\u6539\u8fdb LLM \u5bf9\u4e13\u4e1a\u9886\u57df\u6216\u4efb\u52a1\u7684\u9002\u5e94\u6027\uff0c\u5229\u7528\u6570\u636e\u96c6\u4f5c\u4e3a\u6b64\u4f18\u5316\u8fc7\u7a0b\u7684\u57fa\u77f3\u3002\u5bf9\u5148\u524d\u7814\u7a76\u8fdb\u884c\u6574\u7406\uff0c\u5f3a\u8c03\u6570\u636e\u7684\u8d28\u91cf\u3001\u591a\u6837\u6027\u548c\u76f8\u5173\u6027\u6df1\u523b\u5730\u5f71\u54cd\u6a21\u578b\u7684\u6548\u679c\uff0c\u901a\u5e38\u51b3\u5b9a\u4e86\u540e\u8bad\u7ec3\u5de5\u4f5c\u7684\u6210\u529f\u3002\u4e3a\u4e86\u9610\u660e\u6570\u636e\u96c6\u5728\u6b64\u80cc\u666f\u4e0b\u7684\u5173\u952e\u4f5c\u7528\uff0c\u6211\u4eec\u5bf9\u540e\u8bad\u7ec3\u9636\u6bb5\u4e2d\u4f7f\u7528\u7684\u6570\u636e\u96c6\u8fdb\u884c\u4e86\u5168\u9762\u7684\u56de\u987e\u548c\u6df1\u5165\u5206\u6790\uff0c\u6839\u636e\u5176\u6536\u96c6\u65b9\u6cd5\u5c06\u5176\u5206\u4e3a\u4e09\u79cd\u4e3b\u8981\u7c7b\u578b\uff1a\u4eba\u5de5\u6807\u6ce8\u6570\u636e\u3001\u84b8\u998f\u6570\u636e\u548c\u5408\u6210\u6570\u636e\u3002 \u8fd9\u4e9b\u7c7b\u522b\u53cd\u6620\u4e86\u4e0d\u540c\u7684\u6570\u636e\u7ba1\u7406\u7b56\u7565\uff0c\u6a21\u578b\u91c7\u7528\u5355\u4e00\u65b9\u6cd5\u6216\u6df7\u5408\u65b9\u6cd5\uff0c\u6574\u5408\u591a\u79cd\u7c7b\u578b\u4ee5\u5e73\u8861\u53ef\u6269\u5c55\u6027\u3001\u6210\u672c\u548c\u6027\u80fd\u3002\u4e0b\u8868\u63d0\u4f9b\u4e86\u8fd9\u4e9b\u6570\u636e\u96c6\u7c7b\u578b\u7684\u8be6\u7ec6\u6982\u8ff0\uff0c\u5305\u62ec\u5b83\u4eec\u7684\u6765\u6e90\u3001\u5927\u5c0f\u3001\u8bed\u8a00\u3001\u4efb\u52a1\u548c\u540e\u8bad\u7ec3\u9636\u6bb5\uff08\u4f8b\u5982\uff0cSFT \u548c RLHF\uff09\uff0c\u6211\u4eec\u5c06\u5728\u540e\u7eed\u90e8\u5206\u4e2d\u63a2\u8ba8\u8fd9\u4e9b\u5185\u5bb9\uff0c\u4ee5\u7a81\u51fa\u5b83\u4eec\u5728\u63a8\u8fdb LLM \u529f\u80fd\u65b9\u9762\u7684\u8d21\u732e\u548c\u6311\u6218\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/11\/20251106092157128.jpg\" alt=\"\" \/><\/p>\n<p><center>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff082021-2025\uff09\u7684\u540e\u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684\u6570\u636e\u96c6\u6458\u8981\u3002 \u672c\u8868\u6982\u8ff0\u4e86\u5173\u952e\u6570\u636e\u96c6\uff0c\u8be6\u7ec6\u8bf4\u660e\u4e86\u5b83\u4eec\u7684\u5927\u5c0f\u3001\u6765\u6e90\u3001\u53d1\u5e03\u65f6\u95f4\u7ebf\u4ee5\u53ca\u5728\u4e09\u4e2a\u6307\u6807\u4e0a\u7684\u5c5e\u6027\uff1aLang\uff08\u8bed\u8a00\uff1aEN \u4ee3\u8868\u82f1\u8bed\uff0cCN \u4ee3\u8868\u4e2d\u6587\uff0cML \u4ee3\u8868\u591a\u8bed\u8a00\uff09\uff0cTask\uff08\u7c7b\u578b\uff1aMT \u4ee3\u8868\u591a\u4efb\u52a1\uff0cTS \u4ee3\u8868\u5355\u4efb\u52a1\uff09\uff0c\u4ee5\u53ca Phase\uff08\u7528\u6cd5\uff1aSFT \u4ee3\u8868\u76d1\u7763\u5fae\u8c03\uff0cRLHF \u4ee3\u8868\u4ece\u4eba\u7c7b\u53cd\u9988\u4e2d\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\uff09\u3002 \u6570\u636e\u96c6\u6db5\u76d6\u8303\u56f4\u4ece OpenAI Summarization \u5230 Magpie Reasoning V2\uff0c\u6309\u4eba\u7c7b\u6807\u6ce8\u3001\u84b8\u998f\u548c\u5408\u6210\u7c7b\u578b\u8fdb\u884c\u5206\u7c7b<\/center><\/p>\n<h4>\u4eba\u7c7b\u6807\u6ce8\u6570\u636e\u96c6<\/h4>\n<p>\u4eba\u7c7b\u6807\u6ce8\u6570\u636e\u96c6\u4ee5\u5176\u5353\u8d8a\u7684\u51c6\u786e\u6027\u548c\u4e0a\u4e0b\u6587\u4fdd\u771f\u5ea6\u800c\u8457\u79f0\uff0c\u8fd9\u4e9b\u5c5e\u6027\u6e90\u4e8e\u6807\u6ce8\u8005\u5bf9\u4efb\u52a1\u7684\u7ec6\u81f4\u7406\u89e3\u4ee5\u53ca\u4ed6\u4eec\u8fdb\u884c\u7cbe\u786e\u3001\u4e0a\u4e0b\u6587\u654f\u611f\u8c03\u6574\u7684\u80fd\u529b\u3002 \u8fd9\u4e9b\u6570\u636e\u96c6\u662f\u6539\u8fdb\u6307\u4ee4\u5fae\u8c03\u7684\u57fa\u77f3\uff0c\u901a\u8fc7\u63d0\u4f9b\u9ad8\u8d28\u91cf\u3001\u4e13\u5bb6\u7b56\u5212\u7684\u8bad\u7ec3\u4fe1\u53f7\uff0c\u663e\u8457\u63d0\u5347\u4e86 LLM \u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002\u5728\u8fd9\u4e00\u7c7b\u522b\u4e2d\uff0c\u6770\u51fa\u4ee3\u8868\u5982 Flan\u3001P3\uff08Public Pool of Prompts\uff09\u3001Sup-NatInst\uff08Super-Natural Instructions\uff09\u548c Dolly-15K \u8131\u9896\u800c\u51fa\uff0c\u6210\u4e3a LLM \u540e\u8bad\u7ec3\u4e2d\u5e7f\u6cdb\u91c7\u7528\u7684\u8d44\u6e90\uff0c\u5b83\u4eec\u5404\u81ea\u901a\u8fc7\u4eba\u7c7b\u4e13\u4e1a\u77e5\u8bc6\u4e3a\u6a21\u578b\u80fd\u529b\u7684\u4f18\u5316\u8d21\u732e\u4e86\u72ec\u7279\u4f18\u52bf\u3002<\/p>\n<ul>\n<li><strong>Flan:<\/strong> Flan \u6570\u636e\u96c6\u662f\u4e00\u9879\u57fa\u7840\u6027\u8d44\u6e90\uff0c\u6700\u521d\u5305\u542b 62 \u4e2a\u5e7f\u53d7\u8ba4\u53ef\u7684 NLP \u8bc4\u6d4b\u57fa\u51c6\u2014\u2014\u5982 HellaSwag\u3001MRPC \u548c ANLI\u2014\u2014\u51ed\u501f\u5176 180 \u4e07\u4e2a\u6837\u672c\uff0c\u4e3a\u82f1\u8bed\u73af\u5883\u4e0b\u7684\u7a33\u5065\u591a\u4efb\u52a1\u5b66\u4e60\u63d0\u4f9b\u652f\u6301\u3002\u8fd1\u671f\uff0cFlan-v2 \u4f5c\u4e3a\u66f4\u5148\u8fdb\u7684\u8fed\u4ee3\u7248\u672c\u51fa\u73b0\uff0c\u901a\u8fc7\u878d\u5408 Flan\u3001P3\u3001SUP-NATINST \u53ca\u5176\u4ed6\u6570\u636e\u7ec4\u4ef6\uff0c\u5927\u5e45\u6269\u5c55\u4e86\u5176\u5728\u591a\u79cd\u8bed\u8a00\u548c\u4efb\u52a1\u9886\u57df\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u7684\u5b9e\u7528\u6027\u3002<\/li>\n<li><strong>Sup-Natinst:<\/strong> Super-Natural Instructions (Sup-Natinst)\u63d0\u4f9b\u4e86\u6db5\u76d6 55 \u79cd\u8bed\u8a00\u7684 76 \u79cd\u4efb\u52a1\u7c7b\u578b\u7684\u5e7f\u6cdb\u800c\u591a\u6837\u7684\u6570\u7ec4\uff0c\u4ece\u800c\u786e\u7acb\u4e86\u5176\u4f5c\u4e3a\u591a\u8bed\u8a00 LLM \u540e\u8bad\u7ec3\u7684\u591a\u529f\u80fd\u8d44\u6e90\u3002\u6bcf\u4e2a\u4efb\u52a1\u90fd\u7ecf\u8fc7\u7cbe\u5fc3\u914d\u5bf9\uff0c\u5e76\u9644\u5e26\u4e00\u6761\u6307\u4ee4\uff0c\u5176\u4e2d\u5305\u62ec\u4e00\u4e2a\u6e05\u6670\u7684\u4efb\u52a1\u5b9a\u4e49\u2014\u2014\u6982\u8ff0\u4ece\u8f93\u5165\u6587\u672c\u5230\u6240\u9700\u8f93\u51fa\u7684\u6620\u5c04\u2014\u2014\u4ee5\u53ca\u4e00\u7ec4\u4f8b\u5b50\uff0c\u8fd9\u4e9b\u4f8b\u5b50\u8bf4\u660e\u4e86\u6b63\u786e\u548c\u4e0d\u6b63\u786e\u7684\u53cd\u5e94\uff0c\u4e3a\u6307\u5bfc\u6a21\u578b\u8fdb\u884c\u7cbe\u786e\u7684\u4efb\u52a1\u6267\u884c\u548c\u589e\u5f3a\u8de8\u8bed\u8a00\u9002\u5e94\u6027\u63d0\u4f9b\u4e86\u5f3a\u5927\u7684\u6846\u67b6\u3002<\/li>\n<li><strong>Dolly-15k:<\/strong>\u7531 Databricks \u5458\u5de5\u5f00\u53d1\u7684 Dolly-15K \u4ee3\u8868\u4e86\u7cbe\u5fc3\u7b56\u5212\u7684 15,000 \u4e2a\u9ad8\u8d28\u91cf\u3001\u4eba\u7c7b\u751f\u6210\u7684\u63d0\u793a-\u54cd\u5e94\u5bf9\uff0c\u4e13\u4e3a LLM \u7684\u6307\u4ee4\u5fae\u8c03\u800c\u8bbe\u8ba1\u3002 \u6db5\u76d6\u4e86\u5e7f\u6cdb\u7684\u4e3b\u9898\u548c\u573a\u666f\u2014\u2014\u5305\u62ec\u5934\u8111\u98ce\u66b4\u3001\u5185\u5bb9\u751f\u6210\u3001\u4fe1\u606f\u63d0\u53d6\u3001\u5f00\u653e\u5f0f\u95ee\u7b54\u548c\u6458\u8981\u2014\u2014\u8fd9\u4e2a\u6570\u636e\u96c6\u53cd\u6620\u4e86\u4efb\u52a1\u7c7b\u578b\u7684\u4e30\u5bcc\u591a\u6837\u6027\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u7075\u6d3b\u5730\u9002\u5e94\u5404\u79cd\u6307\u4ee4\u73af\u5883\uff0c\u5e76\u589e\u5f3a\u4e0a\u4e0b\u6587\u76f8\u5173\u6027\u3002<\/li>\n<\/ul>\n<p>\u4eba\u5de5\u6807\u6ce8\u6570\u636e\u96c6\u5728 SFT \u4e2d\u7684\u6548\u529b\u6e90\u4e8e\u5b83\u4eec\u5bf9\u4efb\u52a1\u548c\u573a\u666f\u7684\u5e7f\u6cdb\u8986\u76d6\uff0c\u4e0a\u8ff0\u8bed\u6599\u5e93\u5c31\u4f53\u73b0\u4e86\u8fd9\u4e00\u7279\u5f81\u3002 \u4f5c\u4e3a\u8865\u5145\uff0cOpenAssistant \u63d0\u4f9b\u4e86\u5927\u91cf\u7684\u591a\u8bed\u8a00\u5bf9\u8bdd\u8bed\u6599\u5e93\uff0c\u8fd9\u4e9b\u8bed\u6599\u5e93\u6e90\u4e8e\u5168\u7403\u4f17\u5305\u5de5\u4f5c\uff0c\u53ef\u514d\u8d39\u7528\u4e8e\u63a8\u8fdb\u7814\u7a76\uff0c\u800c OpenOrca \u4f7f\u7528\u6570\u767e\u4e07\u4e2a GPT-3.5 \u548c GPT-4 \u7684\u5b8c\u6210\u7ed3\u679c\u6269\u5c55\u4e86 FlanV2 \uff0c\u6784\u6210\u4e86\u7528\u4e8e\u5fae\u8c03\u548c\u4efb\u52a1\u5bf9\u9f50\u7684\u52a8\u6001\u3001\u4e0d\u65ad\u6269\u5927\u7684\u8d44\u6e90\u3002 \u7136\u800c\uff0c\u5c3d\u7ba1\u5b83\u4eec\u5bf9\u6a21\u578b\u6cdb\u5316\u505a\u51fa\u4e86\u91cd\u5927\u8d21\u732e\uff0c\u4f46\u786e\u4fdd\u4e00\u81f4\u7684\u6807\u6ce8\u8d28\u91cf\u548c\u591a\u6837\u6027\u7684\u6311\u6218\u4f9d\u7136\u5b58\u5728\uff0c\u56e0\u6b64\u9700\u8981\u4e25\u683c\u7684\u8d28\u91cf\u63a7\u5236\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u53d1\u6325\u5176\u5f71\u54cd\u3002<\/p>\n<p><strong>\u7528\u4e8e RLHF \u7684\u4eba\u5de5\u6807\u6ce8\u6570\u636e:<\/strong> \u5bf9\u4e8e RLHF\uff0cP3 \u53ca\u5176\u591a\u8bed\u8a00\u6269\u5c55 xP3 \u548c SHP \u7b49\u4eba\u5de5\u6807\u6ce8\u6570\u636e\u96c6\u63d0\u4f9b\u4e86\u5fc5\u8981\u7684\u3001\u7ecf\u8fc7\u4eba\u5de5\u6ce8\u91ca\u7684\u8bc4\u4f30\uff0c\u53ef\u4ee5\u5b8c\u5584 LLM \u4e0e\u7528\u6237\u504f\u597d\u7684\u5bf9\u9f50\uff0c\u4e3a\u5956\u52b1\u5efa\u6a21\u63d0\u4f9b\u7ec6\u81f4\u7684\u53cd\u9988\u673a\u5236\u3002<\/p>\n<ul>\n<li><strong>P3:<\/strong>P3 \u6570\u636e\u96c6 \u662f\u4e00\u4e2a\u7cbe\u5fc3\u7b56\u5212\u7684\u6307\u4ee4\u8c03\u4f18\u8d44\u6e90\uff0c\u805a\u5408\u4e86\u6765\u81ea Hugging Face Hub \u7684 2300 \u4e07\u4e2a\u591a\u4efb\u52a1\u63d0\u793a\uff0c\u6bcf\u4e2a\u63d0\u793a\u90fd\u9644\u6709\u4eba\u5de5\u8bbe\u8ba1\u7684\u6307\u4ee4\uff0c\u4ee5\u6db5\u76d6\u5404\u79cd NLP \u4efb\u52a1\uff0c\u4ece\u800c\u4e3a RLHF \u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u8d44\u6e90\uff0c\u4ee5\u589e\u5f3a LLM \u5728\u5404\u79cd\u5e94\u7528\u4e2d\u7684\u9002\u5e94\u6027\u548c\u7cbe\u786e\u6027\u3002<\/li>\n<li><strong>xP3:<\/strong> xP3 (Crosslingual Public Pool of Prompts) [463] \u5c06 P3 \u6269\u5c55\u6210\u4e00\u4e2a\u591a\u8bed\u8a00\u6846\u67b6\uff0c\u6db5\u76d6 46 \u79cd\u8bed\u8a00\u548c 16 \u4e2a NLP \u4efb\u52a1\u7684\u63d0\u793a\u548c\u76d1\u7763\u6570\u636e\uff0c\u65e8\u5728\u652f\u6301\u50cf BLOOMZ \u548c mT0 \u8fd9\u6837\u7684\u6a21\u578b\u7684\u591a\u4efb\u52a1\u63d0\u793a\u5fae\u8c03\u3002 \u5b83\u7684\u5185\u5bb9\u96c6\u6210\u4e86\u82f1\u8bed P3 \u6570\u636e\u96c6\u3001\u56db\u4e2a\u65b0\u7684\u82f1\u8bed\u4efb\u52a1\uff08\u4f8b\u5982\uff0c\u7ffb\u8bd1\u3001\u7a0b\u5e8f\u5408\u6210\uff09\u548c 30 \u4e2a\u591a\u8bed\u8a00 NLP \u6570\u636e\u96c6\uff0c\u4e3a\u8de8\u8bed\u8a00 RLHF \u4f18\u5316\u63d0\u4f9b\u4e86\u5168\u9762\u7684\u8d44\u6e90\u3002<\/li>\n<li><strong>SHP:<\/strong> SHP \u5305\u542b 349,000 \u4e2a\u4eba\u7c7b\u504f\u597d\u6807\u6ce8\uff0c\u7528\u4e8e\u56de\u7b54 18 \u4e2a\u4e3b\u9898\u9886\u57df\u7684\u63d0\u95ee\u548c\u6307\u4ee4\uff0c\u8bc4\u4f30\u54cd\u5e94\u7684\u6709\u7528\u6027\u4ee5\u8bad\u7ec3 RLHF \u5956\u52b1\u6a21\u578b\u5e76\u8bc4\u4f30\u81ea\u7136\u8bed\u8a00\u751f\u6210 (NLG) \u8d28\u91cf\uff0c\u5176\u7279\u70b9\u662f\u5b8c\u5168\u4f9d\u8d56\u4e8e\u4eba\u5de5\u7f16\u5199\u7684\u6570\u636e\uff0c\u4f7f\u5176\u4e0e HH-RLHF \u7b49\u6df7\u5408\u6570\u636e\u96c6\u533a\u5206\u5f00\u6765\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u6570\u636e\u96c6\u901a\u8fc7\u63d0\u4f9b\u591a\u6837\u7684\u4eba\u5de5\u6807\u6ce8\u8bc4\u4f30\u6765\u589e\u5f3a RLHF\uff0c\u4ece\u800c\u5b8c\u5584\u6a21\u578b\u4e0e\u7528\u6237\u504f\u597d\u7684\u5bf9\u9f50\u3002 OpenAI Summarization \u548c Webgpt \u63d0\u4f9b\u4e86\u57fa\u4e8e\u6bd4\u8f83\u7684\u7ed3\u6784\u5316\u53cd\u9988\u548c\u674e\u514b\u7279\u91cf\u8868\u8bc4\u7ea7\uff0c\u8fd9\u6709\u52a9\u4e8e\u5c06\u6a21\u578b\u8f93\u51fa\u4e0e\u4eba\u7c7b\u671f\u671b\u66f4\u7d27\u5bc6\u5730\u5bf9\u9f50\u3002 HH-RLHF \u901a\u8fc7\u5305\u542b\u5bf9\u6709\u7528\u6027\u548c\u65e0\u5bb3\u6027\u7684\u8bc4\u4f30\u6765\u8fdb\u4e00\u6b65\u52a0\u5f3a\u8be5\u6846\u67b6\uff0c\u4e3a\u65e8\u5728\u786e\u4fdd\u5b89\u5168\u548c\u5408\u4e4e\u9053\u5fb7\u7684\u54cd\u5e94\u7684\u6a21\u578b\u5960\u5b9a\u4e86\u575a\u5b9e\u7684\u57fa\u7840\u3002\u540c\u65f6\uff0cStackExchange \u8d21\u732e\u4e86\u7279\u5b9a\u9886\u57df\u7684\u3001\u7528\u6237\u751f\u6210\u7684\u5185\u5bb9\uff0c\u4e30\u5bcc\u4e86\u8bad\u7ec3\u6570\u636e\uff0c\u5c24\u5176\u662f\u6709\u5229\u4e8e\u9700\u8981\u5728\u6280\u672f\u9886\u57df\u5177\u6709\u4e13\u4e1a\u77e5\u8bc6\u7684\u6a21\u578b\u3002 \u4f46\u662f\uff0c\u8fd9\u4e9b\u6570\u636e\u96c6\u9047\u5230\u6311\u6218\uff0c\u4f8b\u5982\u53ef\u4f38\u7f29\u6027\uff0c\u4eba\u7c7b\u6ce8\u91ca\u4e2d\u7684\u6f5c\u5728\u504f\u89c1\u4ee5\u53ca\u8d85\u51fa\u5176\u7279\u5b9a\u9886\u57df\u7684\u9002\u7528\u6027\u3002 \u56e0\u6b64\uff0c\u867d\u7136\u5b83\u4eec\u5f88\u6709\u4ef7\u503c\uff0c\u4f46\u8fd9\u4e9b\u8d44\u6e90\u53ef\u80fd\u9700\u8981\u8865\u5145\u66f4\u5e7f\u6cdb\u7684\u6570\u636e\u96c6\uff0c\u4ee5\u5728\u5404\u79cd\u73b0\u5b9e\u4e16\u754c\u4efb\u52a1\u4e2d\u5b9e\u73b0\u5168\u9762\u7684\u6a21\u578b\u5bf9\u9f50\u3002<\/p>\n<h4>\u84b8\u998f\u6570\u636e\u96c6<\/h4>\n<p>\u84b8\u998f\u6570\u636e\u6e90\u4e8e\u5c06\u4f53\u91cf\u5e9e\u5927\u7684\u539f\u59cb\u6570\u636e\u96c6\u7cbe\u70bc\u4e3a\u7d27\u51d1\u3001\u4f18\u5316\u5b50\u96c6\u7684\u590d\u6742\u8fc7\u7a0b\uff0c\u8fd9\u4e9b\u5b50\u96c6\u4ecd\u80fd\u4fdd\u7559 LLM \u8bad\u7ec3\u6240\u9700\u7684\u5173\u952e\u4fe1\u606f\uff0c\u5728\u7ef4\u6301\u6027\u80fd\u7684\u540c\u65f6\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u5e76\u964d\u4f4e\u7b97\u529b\u9700\u6c42\u3002\u7531\u6b64\u5f97\u5230\u7684\u6570\u636e\u96c6\u5f80\u5f80\u53ef\u4e0e\u672a\u84b8\u998f\u7684\u539f\u59cb\u96c6\u5ab2\u7f8e\u751a\u81f3\u66f4\u80dc\u4e00\u7b79\uff0c\u80fd\u591f\u52a0\u901f\u6a21\u578b\u6536\u655b\u3001\u51cf\u5c11\u8d44\u6e90\u6d88\u8017\uff0c\u5c24\u5176\u5728 RLHF \u9636\u6bb5\u6548\u679c\u663e\u8457\u3002\u5178\u578b\u793a\u4f8b ShareGPT \u4e0e HC3\uff08Human\u2013ChatGPT Comparison Corpus\uff09\u901a\u8fc7\u63d0\u4f9b\u771f\u5b9e\u4ea4\u4e92\u4e0e\u5bf9\u6bd4\u6807\u6ce8\uff0c\u4e3a RLHF \u5fae\u8c03 LLM \u63d0\u4f9b\u4e86\u9ad8\u8d28\u91cf\u8d44\u6e90\uff0c\u88ab\u5e7f\u6cdb\u4f7f\u7528\u3002<\/p>\n<ul>\n<li><strong>ShareGPT\uff1a<\/strong>ShareGPT \u662f\u4e00\u4e2a\u52a8\u6001\u6570\u636e\u6536\u96c6\u5e73\u53f0\uff0c\u901a\u8fc7\u5176 API \u805a\u5408\u4e86\u7ea6 9 \u4e07\u6761\u7528\u6237\u4e0e ChatGPT \u6216 GPT-4 \u7684\u771f\u5b9e\u5bf9\u8bdd\u3002\u8be5\u6570\u636e\u96c6\u5305\u542b\u771f\u5b9e\u4eba\u7c7b\u6307\u4ee4\u4e0e\u5bf9\u5e94 AI \u56de\u590d\uff0c\u5c06\u81ea\u7136\u5bf9\u8bdd\u6a21\u5f0f\u6d53\u7f29\u4e3a\u9ad8\u8d28\u91cf\u8bed\u6599\uff0c\u4f7f RLHF \u80fd\u5728\u9ad8\u76f8\u5173\u3001\u9ad8\u8d28\u91cf\u7684\u57fa\u7840\u4e0a\u63d0\u5347 LLM \u7684\u5bf9\u8bdd\u6d41\u7545\u5ea6\u4e0e\u4e0a\u4e0b\u6587\u54cd\u5e94\u80fd\u529b\u3002<\/li>\n<li><strong>HC3\uff1a<\/strong>HC3 \u6570\u636e\u96c6\u65e8\u5728\u5e76\u6392\u5448\u73b0 ChatGPT \u751f\u6210\u56de\u590d\u4e0e\u4eba\u7c7b\u64b0\u5199\u7b54\u6848\uff0c\u6db5\u76d6\u5f00\u653e\u57df\u3001\u91d1\u878d\u3001\u533b\u5b66\u3001\u6cd5\u5f8b\u3001\u5fc3\u7406\u7b49\u5b66\u79d1\u7684 16.1 \u4e07\u5bf9\u95ee\u7b54\u3002\u8fd9\u4e00\u7cbe\u70bc\u8bed\u6599\u4fbf\u4e8e\u6bd4\u8f83\u4e24\u79cd\u56de\u590d\u7684\u7279\u5f81\u4e0e\u8d28\u91cf\uff0c\u5e2e\u52a9\u7814\u7a76\u8005\u5728 RLHF \u8fc7\u7a0b\u4e2d\u589e\u5f3a LLM \u8f93\u51fa\u7684\u771f\u5b9e\u6027\u4e0e\u9886\u57df\u7cbe\u5ea6\uff0c\u540c\u65f6\u51f8\u663e\u4eba\u7c7b\u4e0e AI \u5185\u5bb9\u7684\u5dee\u5f02\u3002<\/li>\n<\/ul>\n<h4>\u5408\u6210\u6570\u636e\u96c6<\/h4>\n<p>\u5408\u6210\u6570\u636e\u5728 LLM \u8bad\u7ec3\u540e\u7684 SFT \u9636\u6bb5\u88ab\u89c6\u4e3a\u53d8\u9769\u6027\u8d44\u4ea7\uff0c\u7531 AI \u6a21\u578b\u81ea\u52a8\u751f\u6210\uff0c\u53ef\u66ff\u4ee3\u4eba\u5de5\u6807\u6ce8\u8bed\u6599\uff0c\u517c\u5177\u6210\u672c\u4f4e\u5ec9\u3001\u89c4\u6a21\u6613\u6269\u3001\u9690\u79c1\u53cb\u597d\u7b49\u4f18\u52bf\u3002\u901a\u8fc7\u81ea\u52a8\u5316\u5730\u751f\u6210\u201c\u6307\u4ee4\u2013\u56de\u590d\u201d\u5bf9\u4e0e\u591a\u8f6e\u5bf9\u8bdd\uff0c\u5408\u6210\u6570\u636e\u80fd\u5feb\u901f\u6784\u5efa\u6d77\u91cf\u8bad\u7ec3\u96c6\uff0c\u63d0\u5347\u6a21\u578b\u9002\u5e94\u6027\uff1b\u5176\u4e2d Self-Instruct-52K\u3001Vicuna \u4e0e Baize \u7b49\u5df2\u6210\u4e3a\u589e\u5f3a LLM \u6307\u4ee4\u9075\u5faa\u4e0e\u5bf9\u8bdd\u751f\u6210\u80fd\u529b\u7684\u7ecf\u5178\u8d44\u6e90\u3002<br \/>\n\u57fa\u4e8e Self-Instruct \u65b9\u6cd5\u7684\u6570\u636e\u96c6<br \/>\n\u6b64\u7c7b\u6570\u636e\u96c6\u4ec5\u4ee5\u5c11\u91cf\u624b\u5de5\u79cd\u5b50\u793a\u4f8b\u8d77\u6b65\uff0c\u5229\u7528 LLM \u81ea\u751f\u6210\u5927\u89c4\u6a21\u6307\u4ee4\u9075\u5faa\u6570\u636e\uff0c\u4ece\u800c\u5f3a\u5316\u6a21\u578b\u5bf9\u5404\u79cd\u6307\u4ee4\u7684\u54cd\u5e94\u80fd\u529b\u3002\u4ee3\u8868\u5305\u62ec Self-Instruct-52K\u3001Alpaca \u53ca Magpie \u7cfb\u5217\uff0c\u5b83\u4eec\u901a\u8fc7\u53ef\u6269\u5c55\u7684\u81ea\u52a8\u5316\u6d41\u7a0b\u5171\u540c\u63a8\u52a8\u6307\u4ee4\u5fae\u8c03\u3002<\/p>\n<ul>\n<li><strong>Self-Instruct-52K\uff1a<\/strong>\u5efa\u7acb\u6307\u4ee4\u9075\u5faa\u6a21\u578b\u7684\u57fa\u51c6\uff0c\u501f\u52a9\u591a\u6837\u5316\u63d0\u793a\u6a21\u677f\u4ece 175 \u6761\u624b\u5de5\u79cd\u5b50\u6269\u5c55\u51fa 5.2 \u4e07\u793a\u4f8b\uff0c\u63d0\u5347 LLM \u7cbe\u51c6\u3001\u4e00\u81f4\u5730\u6267\u884c\u5177\u4f53\u4efb\u52a1\u6307\u4ee4\u7684\u80fd\u529b\u3002<\/li>\n<li><strong>Alpaca\uff1a<\/strong>Alpaca \u4e0e Alpaca-GPT4 \u5206\u522b\u7528 GPT-3 \u4e0e GPT-4 \u5c06 175 \u6761\u79cd\u5b50\u5bf9\u6269\u5c55\u4e3a 5.2 \u4e07\u9ad8\u8d28\u91cf\u201c\u6307\u4ee4\u2013\u56de\u590d\u201d\u5bf9\uff0c\u589e\u5f3a\u6307\u4ee4\u9075\u5faa\uff1bInstInWild \u5219\u9488\u5bf9\u591a\u8bed\u8a00\u573a\u666f\u751f\u6210\u82f1\/\u4e2d\u53cc\u8bed\u6570\u636e\uff0c\u5f3a\u5316\u8de8\u8bed\u8a00\u9002\u5e94\u6027\u3002<\/li>\n<li><strong>Magpie \u6570\u636e\u96c6\uff1a<\/strong>\u5229\u7528\u5df2\u5bf9\u9f50\u7684 LLM \u6309\u6a21\u677f\u751f\u6210\u201c\u6307\u4ee4\u2013\u56de\u590d\u201d\u5bf9\uff0c\u5f62\u6210\u591a\u4e2a\u4e13\u7528\u5b50\u96c6\uff1aMagpie-Reasoning V2\uff08\u91cd\u94fe\u5f0f\u601d\u8003\uff09\u3001Magpie-Llama-3 \/ Qwen-2 \u7cfb\u5217\uff08\u9488\u5bf9\u70ed\u95e8\u6a21\u578b\u5b9a\u5236\uff09\u3001Magpie-Gemma-2\uff08\u9002\u914d Gemma \u67b6\u6784\uff09\u53ca Magpie-Air-DPO\uff08\u878d\u5165\u504f\u597d\u4fe1\u53f7\uff09\uff0c\u5171\u540c\u63d0\u5347\u5bf9\u8bdd\u4e0e\u63a8\u7406\u4efb\u52a1\u7684 SFT \u4e0e\u6307\u4ee4\u5fae\u8c03\u6548\u679c\u3002<\/li>\n<li>\u6b64\u5916\uff0cUnnatural Instructions\uff0824 \u4e07\u793a\u4f8b\uff09\u3001Evol-Instruct\uff08\u7ecf\u8fed\u4ee3\u590d\u6742\u5ea6\u589e\u5f3a\u5f97 7\u201314.3 \u4e07\u6761\uff09\u4e0e Belle\uff08\u6765\u81ea ChatGPT \u7684 50\u2013110 \u4e07\u6761\u4e2d\u6587\u5bf9\u8bdd\uff09\u7b49\u5927\u5e45\u6269\u5145\u4e86\u6307\u4ee4\u6570\u636e\uff0c\u4f46\u5728\u8d28\u91cf\u7b5b\u9009\u3001\u590d\u6742\u5ea6\u6821\u51c6\u4e0e\u504f\u89c1\u7f13\u89e3\u65b9\u9762\u4ecd\u9762\u4e34\u6311\u6218\uff0c\u9700\u8981\u6301\u7eed\u6539\u8fdb\u4ee5\u4fdd\u8bc1\u590d\u6742\u573a\u666f\u4e0b\u7684\u53ef\u9760\u6027\u3002<\/li>\n<\/ul>\n<p><strong>\u57fa\u4e8e Self-Chat \u65b9\u6cd5\u7684\u6570\u636e\u96c6<\/strong>\uff1aSelf-Chat \u8ba9\u6a21\u578b\u5728\u5185\u90e8\u6216\u4e0e\u201c\u540c\u4f34\u201d\u6a21\u62df\u591a\u8f6e\u5bf9\u8bdd\uff0c\u5f25\u8865\u73b0\u6709\u8bed\u6599\u5728\u5bf9\u8bdd\u6df1\u5ea6\u4e0a\u7684\u4e0d\u8db3\uff0c\u63d0\u5347\u751f\u6210\u80fd\u529b\u3002Baize\u3001UltraChat \u4e0e OpenHermes \u7b49\u901a\u8fc7\u81ea\u52a8\u5316\u4ea4\u4e92\u7b56\u7565\u4f53\u73b0\u8fd9\u4e00\u601d\u8def\u3002<\/p>\n<ul>\n<li><strong>Baize\uff1a<\/strong>\u501f\u52a9 ChatGPT \u7684 Self-Chat \u6280\u672f\uff0c\u878d\u5408 Quora\u3001Stack Overflow \u4e0e Alpaca \u79cd\u5b50\uff0c\u751f\u6210 65.3 \u4e07\u6761\u591a\u8f6e\u5bf9\u8bdd\uff0c\u4e30\u5bcc\u6307\u4ee4\u8d28\u91cf\uff0c\u6539\u8fdb LLM \u5728 SFT \u4e2d\u7684\u5bf9\u8bdd\u8fde\u8d2f\u6027\u4e0e\u4efb\u52a1\u9075\u4ece\u5ea6\u3002<\/li>\n<li><strong>UltraChat\uff1a<\/strong>\u901a\u8fc7\u591a API \u5e76\u884c\u8c03\u7528 ChatGPT\uff0c\u751f\u6210\u903e 1200 \u4e07\u6761\u8986\u76d6\u5e7f\u6cdb\u4e3b\u9898\u7684\u9ad8\u8d28\u91cf\u591a\u8f6e\u8bb0\u5f55\uff0c\u514b\u670d\u4f20\u7edf\u591a\u8f6e\u6570\u636e\u96c6\u8d28\u91cf\u5dee\u3001\u6807\u6ce8\u4e0d\u51c6\u7b49\u95ee\u9898\uff0c\u4e3a\u5bf9\u8bdd\u589e\u5f3a\u63d0\u4f9b\u575a\u5b9e SFT \u8d44\u6e90\u3002<\/li>\n<li><strong>OpenHermes\uff1a<\/strong>\u7531 Teknium \u53d1\u5e03\uff0c\u542b OpenHermes-1\uff0824.3 \u4e07\u6761\uff09\u4e0e\u6269\u5c55\u7248 OpenHermes-2.5\uff08100 \u4e07\u6761\uff09\uff0c\u4e3b\u9898\u4e0e\u4efb\u52a1\u7c7b\u578b\u4e30\u5bcc\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u5bf9\u8bdd\u4e0e\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u3002<br \/>\n\u8fd9\u4e9b\u6570\u636e\u96c6\u901a\u8fc7\u6a21\u578b\u201c\u81ea\u6211\u4ea4\u8c08\u201d\u6784\u5efa\u591a\u8f6e\u573a\u666f\uff0c\u663e\u8457\u6539\u5584\u5bf9\u8bdd\u8d28\u91cf\uff0c\u586b\u8865\u8bad\u7ec3\u8bed\u6599\u7684\u5173\u952e\u7a7a\u767d\u3002<\/li>\n<\/ul>\n<p><strong>\u57fa\u4e8e\u771f\u5b9e\u7528\u6237\u4ea4\u4e92\u7684\u6570\u636e\u96c6\uff1a<\/strong>\u5229\u7528\u771f\u5b9e\u7528\u6237\u4e0e\u6a21\u578b\u7684\u5bf9\u8bdd\uff0c\u6355\u83b7\u591a\u6837\u4e14\u8d34\u8fd1\u73b0\u5b9e\u7684\u8f93\u5165\uff0c\u63d0\u5347\u6a21\u578b\u89e3\u51b3\u771f\u5b9e\u573a\u666f\u7684\u80fd\u529b\u3002Vicuna\u3001WildChat \u4e0e GenQA \u4e3a\u5178\u578b\u4ee3\u8868\u3002<\/p>\n<ul>\n<li><strong>Vicuna\uff1a<\/strong>\u4ece ShareGPT \u516c\u5f00 API \u83b7\u53d6\u7ea6 7 \u4e07\u6bb5\u7528\u6237\u5206\u4eab\u5bf9\u8bdd\uff0c\u7ecf HTML\u2192Markdown \u8f6c\u6362\u3001\u4f4e\u8d28\u6837\u672c\u8fc7\u6ee4\u4e0e\u957f\u5bf9\u8bdd\u622a\u65ad\uff0c\u4fdd\u8bc1 SFT \u6570\u636e\u8d28\u91cf\uff0c\u7528\u4e8e\u771f\u5b9e\u4ea4\u4e92\u5efa\u6a21\u3002<\/li>\n<li><strong>WildChat\uff1a<\/strong>\u5305\u542b 100 \u4e07\u6761\u591a\u8bed\u8a00\u3001\u8de8\u63d0\u793a\u7c7b\u578b\u7684\u771f\u5b9e\u7528\u6237-ChatGPT \u4ea4\u4e92\uff0c\u6db5\u76d6\u6b67\u4e49\u8bf7\u6c42\u3001\u8bed\u7801\u8f6c\u6362\u7b49\u72ec\u7279\u73b0\u8c61\uff0c\u65e2\u53ef\u7528\u4e8e SFT\uff0c\u4e5f\u53ef\u4f5c\u4e3a\u7528\u6237\u884c\u4e3a\u5206\u6790\u5de5\u5177\u3002<\/li>\n<li><strong>GenQA\uff1a<\/strong>\u63d0\u4f9b\u903e 1000 \u4e07\u6761\u7ecf\u6e05\u6d17\u8fc7\u6ee4\u7684\u6307\u4ee4\u6837\u672c\uff0c\u5b8c\u5168\u7531 LLM \u751f\u6210\uff0c\u65e0\u9700\u4eba\u5de5\u6216\u590d\u6742\u6d41\u7a0b\uff0c\u53ef\u5feb\u901f\u8865\u4f4d\u73b0\u6709\u8bed\u6599\uff0c\u89e3\u51b3\u8986\u76d6\u7f3a\u53e3\u3002<br \/>\n\u76f8\u8f83\u4eba\u5de5\u6807\u6ce8\uff0c\u5408\u6210\u6570\u636e\u5728\u6210\u672c\u3001\u89c4\u6a21\u4e0e\u9690\u79c1\u4e0a\u4f18\u52bf\u660e\u663e\uff0c\u4f46\u6df1\u5ea6\u4e0e\u771f\u5b9e\u6027\u53ef\u80fd\u4e0d\u8db3\uff0c\u5b58\u5728\u653e\u5927\u6a21\u578b\u504f\u5dee\u6216\u8fc7\u5ea6\u7b80\u5316\u7684\u98ce\u9669\u3002\u8fc7\u5ea6\u4f9d\u8d56 AI \u751f\u6210\u5185\u5bb9\u4f1a\u5c06\u56fa\u6709\u9519\u8bef\u56fa\u5316\uff0c\u56e0\u6b64\u5fc5\u987b\u5c06\u5408\u6210\u6570\u636e\u4e0e\u4eba\u5de5\u6570\u636e\u76f8\u7ed3\u5408\uff0c\u4ee5\u5168\u9762\u63d0\u5347 LLM \u7684\u9c81\u68d2\u6027\u4e0e\u573a\u666f\u9002\u7528\u6027\u3002<\/li>\n<\/ul>\n<h3>\u5e94\u7528<\/h3>\n<p>\u5c3d\u7ba1\u9884\u8bad\u7ec3\u8d4b\u4e88\u4e86\u5f3a\u5927\u7684\u57fa\u7840\u80fd\u529b\uff0c\u4f46\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u5728\u90e8\u7f72\u5230\u4e13\u4e1a\u9886\u57df\u65f6\u7ecf\u5e38\u4f1a\u9047\u5230\u6301\u7eed\u7684\u9650\u5236\uff0c\u5305\u62ec\u53d7\u9650\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3001\u4ea7\u751f\u5e7b\u89c9\u7684\u503e\u5411\u3001\u6b21\u4f18\u7684\u63a8\u7406\u80fd\u529b\u4ee5\u53ca\u6839\u6df1\u8482\u56fa\u7684\u504f\u89c1\u3002 \u8fd9\u4e9b\u7f3a\u70b9\u5728\u7cbe\u786e\u6027\u3001\u53ef\u9760\u6027\u548c\u4f26\u7406\u4e00\u81f4\u6027\u81f3\u5173\u91cd\u8981\u7684\u73b0\u5b9e\u4e16\u754c\u5e94\u7528\u4e2d\u5177\u6709\u5173\u952e\u610f\u4e49\u3002 \u8fd9\u4e9b\u6311\u6218\u4fc3\u4f7f\u4e86\u6839\u672c\u6027\u7684\u63a2\u7a76\uff1a<\/p>\n<ul>\n<li>(1) \u5982\u4f55\u7cfb\u7edf\u5730\u589e\u5f3a\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u6027\u80fd\u4ee5\u6ee1\u8db3\u7279\u5b9a\u9886\u57df\u7684\u9700\u8981\uff1f<\/li>\n<li>(2) \u4ec0\u4e48\u7b56\u7565\u80fd\u6709\u6548\u5730\u7f13\u89e3\u5e94\u7528\u573a\u666f\u4e2d\u56fa\u6709\u7684\u5b9e\u9645\u969c\u788d\uff1f \u8bad\u7ec3\u540e\u5904\u7406\uff08Post-training\uff09 \u6210\u4e3a\u4e00\u4e2a\u5173\u952e\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u901a\u8fc7\u5b8c\u5584 LLM \u5bf9\u7279\u5b9a\u9886\u57df\u672f\u8bed\u548c\u63a8\u7406\u6a21\u5f0f\u7684\u8bc6\u522b\u6765\u589e\u5f3a\u5176\u9002\u5e94\u6027\uff0c\u540c\u65f6\u4fdd\u7559\u5176\u5e7f\u6cdb\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>\u672c\u7ae0\u9610\u8ff0\u4e86\u7ecf\u8fc7\u540e\u8bad\u7ec3\u7684 LLM \u5728\u4e13\u4e1a\u3001\u6280\u672f\u548c\u4ea4\u4e92\u9886\u57df\u7684\u53d8\u9769\u6027\u5e94\u7528\uff0c\u9610\u660e\u4e86\u5b9a\u5236\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u5982\u4f55\u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\u5e76\u63d0\u9ad8\u6a21\u578b\u5728\u4e0d\u540c\u73af\u5883\u4e0b\u7684\u5b9e\u7528\u6027\u3002<\/p>\n<h4>\u4e13\u4e1a\u9886\u57df<\/h4>\n<ul>\n<li><strong>\u6cd5\u5f8b\u52a9\u624b:<\/strong>\u6cd5\u5f8b\u9886\u57df\u662f\u5229\u7528\u540e\u8bad\u7ec3\u4e3a LLM \u6ce8\u5165\u4e13\u4e1a\u77e5\u8bc6\u7684\u5178\u578b\u573a\u666f\uff0c\u4f7f\u5176\u80fd\u591f\u9a7e\u9a6d\u9519\u7efc\u590d\u6742\u7684\u6cd5\u5f8b\u77e5\u8bc6\u5e76\u5e94\u5bf9\u6cd5\u5b66\u4e2d\u7684\u591a\u91cd\u6311\u6218\u3002\u76f8\u5173\u7814\u7a76\u5df2\u6db5\u76d6\u6cd5\u5f8b\u95ee\u7b54\u3001\u5224\u51b3\u9884\u6d4b\u3001\u6587\u6863\u6458\u8981\u4ee5\u53ca\u68c0\u7d22\u589e\u5f3a\u4e0e\u53f8\u6cd5\u63a8\u7406\u7b49\u4efb\u52a1\u3002\u7ecf\u8fc7\u540e\u8bad\u7ec3\u7684\u6cd5\u5f8b\u52a9\u624b\uff08\u5982 LawGPT\u3001Lawyer-LLaMA\uff09\u4e0d\u4ec5\u5728\u5404\u7c7b\u6cd5\u5f8b\u4e8b\u52a1\u4e2d\u63d0\u4f9b\u53ef\u9760\u6307\u5f15\uff0c\u8fd8\u5728\u4e13\u4e1a\u8d44\u683c\u8003\u8bd5\u4e2d\u53d6\u5f97\u4f73\u7ee9\uff0c\u5c55\u73b0\u51fa\u5353\u8d8a\u7684\u89e3\u8bfb\u4e0e\u5206\u6790\u80fd\u529b\u3002LexiLaw \u4e0e SAUL \u7b49\u591a\u8bed\u8a00\u6a21\u578b\u8fdb\u4e00\u6b65\u5c06\u5b9e\u7528\u6027\u62d3\u5c55\u81f3\u4e2d\u82f1\u7b49\u8bed\u79cd\uff0c\u6269\u5927\u53ef\u53ca\u6027\u3002\u6838\u5fc3\u5728\u4e8e\u4ee5\u7cbe\u5fc3\u6574\u7406\u7684\u6cd5\u5f8b\u8bed\u6599\uff08\u5982 ChatLaw\uff09\u8fdb\u884c\u540e\u8bad\u7ec3\uff0c\u5c06\u5927\u91cf\u6cd5\u5f8b\u6587\u672c\u6574\u5408\u8fdb\u5bf9\u8bdd\u6570\u636e\u96c6\uff0c\u4f7f\u6a21\u578b\u5f97\u4ee5\u5b8c\u5584\u63a8\u7406\u4e0e\u672f\u8bed\u8bc6\u522b\u80fd\u529b\u3002<\/li>\n<li><strong>\u533b\u7597\u4fdd\u5065\u4e0e\u533b\u5b66:<\/strong>\u540e\u8bad\u7ec3\u501f\u52a9\u9886\u57df\u4e13\u5c5e\u6570\u636e\uff0c\u663e\u8457\u63d0\u5347 LLM \u5728\u4e34\u5e8a\u4e0e\u5b66\u672f\u573a\u666f\u4e2d\u7684\u8868\u73b0\u3002\u4e34\u5e8a\u65b9\u9762\u6db5\u76d6\u836f\u7269\u53d1\u73b0\u3001\u836f\u7269\u534f\u540c\u9884\u6d4b\u3001\u50ac\u5316\u5242\u8bbe\u8ba1\u3001\u8bca\u65ad\u652f\u6301\u3001\u75c5\u5386\u751f\u6210\u4e0e\u60a3\u8005\u4e92\u52a8\uff1b\u5b66\u672f\u65b9\u9762\u5219\u5305\u62ec\u533b\u5b66\u62a5\u544a\u5408\u6210\u4e0e\u95ee\u7b54\u3002ChatMed \u57fa\u4e8e 50 \u4e07\u6761\u533b\u7597\u54a8\u8be2\u8bb0\u5f55\u8bad\u7ec3\uff0c\u8bca\u65ad\u4e0e\u54a8\u8be2\u51c6\u786e\u6027\u589e\u5f3a\uff1bPULSE \u4f7f\u7528 400 \u4e07\u6761\u6db5\u76d6\u4e2d\u6587\u533b\u7597\u4e0e\u4e00\u822c\u9886\u57df\u6307\u4ee4\u8fdb\u884c\u5fae\u8c03\uff0c\u591a\u4efb\u52a1\u5904\u7406\u80fd\u529b\u7a81\u51fa\u3002\u901a\u8fc7\u5d4c\u5165\u7ec6\u7c92\u5ea6\u533b\u5b66\u77e5\u8bc6\u7684\u540e\u8bad\u7ec3\u8c03\u6574\uff0c\u8fd9\u4e9b\u6a21\u578b\u5168\u9762\u8d85\u8d8a\u901a\u7528\u7248\u672c\uff0c\u5f70\u663e\u5b9a\u5236\u6570\u636e\u96c6\u5bf9\u5b9e\u9645\u5e94\u7528\u7684\u4e0d\u53ef\u6216\u7f3a\u6027\uff0c\u4e5f\u4e3a LLM \u878d\u5165\u533b\u7597\u5de5\u4f5c\u6d41\u5960\u5b9a\u57fa\u7840\u3002<\/li>\n<li><strong>\u91d1\u878d\u4e0e\u7ecf\u6d4e\u5b66:<\/strong>\u5728\u91d1\u878d\u4e0e\u7ecf\u6d4e\u5b66\u9886\u57df\uff0cLLM \u5df2\u5728\u60c5\u611f\u5206\u6790\u3001\u4fe1\u606f\u62bd\u53d6\u4e0e\u95ee\u7b54\u7b49\u4efb\u52a1\u4e2d\u5c55\u73b0\u5de8\u5927\u6f5c\u529b\uff0c\u540e\u8bad\u7ec3\u5219\u901a\u8fc7\u9886\u57df\u4e13\u5c5e\u6539\u8fdb\u653e\u5927\u5176\u6548\u80fd\u3002FinGPT \u4e0e DISC-FinLLM \u7b49\u4e13\u7528\u6a21\u578b\u7ecf\u91d1\u878d\u8bed\u6599\u540e\u8bad\u7ec3\u540e\uff0c\u5728\u9700\u8981\u7ec6\u81f4\u7406\u89e3\u5e02\u573a\u52a8\u6001\u4e0e\u672f\u8bed\u7684\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02\u3002XuanYuan \u501f\u52a9\u5927\u89c4\u6a21\u91d1\u878d\u6570\u636e\u96c6\u4e0e\u5148\u8fdb\u540e\u8bad\u7ec3\u6280\u672f\uff0c\u63d0\u9ad8\u7ecf\u6d4e\u5efa\u6a21\u4e0e\u9884\u6d4b\u51c6\u786e\u6027\uff0c\u5168\u9762\u8d85\u8d8a\u672a\u8c03\u57fa\u51c6\u3002\u8fd9\u4e9b\u8fdb\u5c55\u8bf4\u660e\u540e\u8bad\u7ec3\u5728\u4f7f LLM \u9002\u5e94\u91d1\u878d\u5e94\u7528\u590d\u6742\u9700\u6c42\u4e2d\u7684\u5173\u952e\u4f5c\u7528\uff0c\u786e\u4fdd\u6a21\u578b\u5728\u89e3\u8bfb\u5b9a\u91cf\u4e0e\u5b9a\u6027\u6d1e\u5bdf\u65f6\u63d0\u4f9b\u53ef\u9760\u3001\u7b26\u5408\u884c\u4e1a\u6807\u51c6\u7684\u8f93\u51fa\u3002<\/li>\n<li><strong>MobileAgent:<\/strong>\u5927\u578b\u591a\u6a21\u6001\u6a21\u578b\uff08LMM\uff09\u7684\u6f14\u8fdb\u50ac\u751f\u4e86\u57fa\u4e8e LMM \u7684 GUI \u4ee3\u7406\u8fd9\u4e00\u65b0\u5174\u7814\u7a76\u9886\u57df\uff0c\u65e8\u5728\u5f00\u53d1\u53ef\u8de8 Web \u754c\u9762\u3001\u4e2a\u4eba\u8ba1\u7b97\u5e73\u53f0\u4e0e\u79fb\u52a8\u8bbe\u5907\u7b49\u591a\u79cd GUI \u73af\u5883\u6267\u884c\u4efb\u52a1\u7684 AI \u52a9\u624b\u3002\u5728\u79fb\u52a8\u573a\u666f\u4e0b\uff0c\u7814\u7a76\u901a\u8fc7\u5de5\u5177\u96c6\u6210\u4e0e\u989d\u5916\u63a2\u7d22\u9636\u6bb5\u589e\u5f3a\u5355\u4e2a\u4ee3\u7406\u7684\u611f\u77e5\u4e0e\u63a8\u7406\u80fd\u529b\uff1b\u8fd1\u671f\u66f4\u91c7\u7528\u591a\u4ee3\u7406\u7cfb\u7edf\u8fdb\u884c\u51b3\u7b56\u4e0e\u53cd\u601d\uff0c\u663e\u8457\u63d0\u5347\u4efb\u52a1\u6548\u7387\u3002MobileAgent-E \u5f15\u5165\u4ee3\u7406\u95f4\u5c42\u7ea7\u7ed3\u6784\uff0c\u4fc3\u8fdb\u957f\u671f\u89c4\u5212\u5e76\u63d0\u9ad8\u4f4e\u7ea7\u52a8\u4f5c\u7cbe\u5ea6\u3002\u8fd9\u4e9b\u53d1\u5c55\u51f8\u663e\u4e86\u591a\u6a21\u6001\u540e\u8bad\u7ec3\u7b56\u7565\u5728\u57f9\u80b2\u9002\u7528\u4e8e\u590d\u6742\u79fb\u52a8\u73af\u5883\u7684\u81ea\u9002\u5e94\u3001\u9ad8\u6548\u4ee3\u7406\u4e2d\u7684\u53d8\u9769\u6027\u4f5c\u7528\u3002<\/li>\n<\/ul>\n<h4>\u6280\u672f\u548c\u903b\u8f91\u63a8\u7406<\/h4>\n<ul>\n<li><strong>\u6570\u5b66\u63a8\u7406\uff1a<\/strong>\u5927\u8bed\u8a00\u6a21\u578b\u5728\u4ee3\u6570\u8fd0\u7b97\u3001\u5fae\u79ef\u5206\u4e0e\u7edf\u8ba1\u5206\u6790\u7b49\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e0a\u5c55\u73b0\u51fa\u5de8\u5927\u6f5c\u529b\uff0c\u800c\u540e\u8bad\u7ec3\u5bf9\u4e8e\u5f25\u5408\u201c\u4f1a\u7b97\u201d\u4e0e\u201c\u4eba\u7ea7\u719f\u7ec3\u201d\u4e4b\u95f4\u7684\u9e3f\u6c9f\u5c24\u4e3a\u5173\u952e\u3002GPT-4 \u5728\u6807\u51c6\u5316\u6570\u5b66\u6d4b\u8bc4\u4e2d\u53d6\u5f97\u9ad8\u5206\uff0c\u5f97\u76ca\u4e8e\u5176\u591a\u6837\u5316\u7684\u9884\u8bad\u7ec3\u8bed\u6599\uff1b\u8fdb\u4e00\u6b65\u7684 SFT \u4e0e\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09\u7b49\u540e\u8bad\u7ec3\u624b\u6bb5\u53ef\u7ee7\u7eed\u63d0\u5347\u63a8\u7406\u7cbe\u5ea6\u3002DeepSeekMath\u5229\u7528\u4e13\u95e8\u6570\u5b66\u6570\u636e\u96c6\u4e0e GRPO\uff0c\u914d\u5408\u7ed3\u6784\u5316\u601d\u7ef4\u94fe\uff08CoT\uff09\u5904\u7406\u590d\u6742\u95ee\u9898\u3002OpenAI \u7684 o1 \u5219\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u8fed\u4ee3\u4f18\u5316\u63a8\u7406\u7b56\u7565\uff0c\u5728\u591a\u6b65\u63a8\u5bfc\u4e0e\u8bc1\u660e\u4e2d\u5b9e\u73b0\u5353\u8d8a\u6027\u80fd\uff0c\u4f7f\u6a21\u578b\u8f93\u51fa\u4e0e\u4e25\u8c28\u6570\u5b66\u903b\u8f91\u5bf9\u9f50\uff0c\u6210\u4e3a\u6559\u80b2\u4e0e\u79d1\u7814\u9886\u57df\u7684\u5f97\u529b\u5de5\u5177\u3002<\/li>\n<li><strong>\u4ee3\u7801\u751f\u6210\uff1a<\/strong>\u540e\u8bad\u7ec3\u5f7b\u5e95\u9769\u65b0\u4e86\u4ee3\u7801\u751f\u6210\uff0c\u4f7f\u5927\u8bed\u8a00\u6a21\u578b\u5728\u81ea\u52a8\u7f16\u7801\u3001\u8c03\u8bd5\u4e0e\u6587\u6863\u7f16\u5199\u65b9\u9762\u8868\u73b0\u4f18\u5f02\uff0c\u663e\u8457\u6539\u53d8\u8f6f\u4ef6\u5f00\u53d1\u6d41\u7a0b\u3002Codex \u5728\u5e9e\u5927\u4e14\u591a\u6837\u7684\u4ee3\u7801\u5e93\u4e0a\u8bad\u7ec3\uff0c\u4e3a GitHub Copilot \u63d0\u4f9b\u5b9e\u65f6\u3001\u9ad8\u51c6\u786e\u5ea6\u7684\u7f16\u7801\u8f85\u52a9\uff1bCode Llama \u7b49\u4e13\u7528\u6a21\u578b\u5219\u501f\u52a9\u7f16\u7a0b\u9886\u57df\u4e13\u5c5e\u6570\u636e\u96c6\u7684\u540e\u8bad\u7ec3\uff0c\u8de8\u8bed\u8a00\u3001\u8de8\u6846\u67b6\u5730\u5e2e\u52a9\u5f00\u53d1\u8005\u3002o1 \u8fdb\u4e00\u6b65\u628a\u6570\u5b66\u63a8\u7406\u80fd\u529b\u5ef6\u4f38\u81f3\u4ee3\u7801\u751f\u6210\uff0c\u4ea7\u51fa\u9ad8\u8d28\u91cf\u3001\u4e0a\u4e0b\u6587\u611f\u77e5\u7684\u4ee3\u7801\u7247\u6bb5\uff0c\u5ab2\u7f8e\u4eba\u7c7b\u6c34\u5e73\u3002\u5f53\u524d\u7814\u7a76\u6301\u7eed\u805a\u7126\u4e2a\u6027\u5316\u589e\u5f3a\u3001\u4e0a\u4e0b\u6587\u6df1\u5ea6\u7406\u89e3\u4e0e\u4f26\u7406\u4fdd\u969c\u5d4c\u5165\uff0c\u4ee5\u9632\u8303\u4ee3\u7801\u6ee5\u7528\u7b49\u98ce\u9669\uff0c\u786e\u4fdd LLM \u5728\u63d0\u5347\u751f\u4ea7\u529b\u7684\u540c\u65f6\u606a\u5b88\u8d1f\u8d23\u4efb\u5f00\u53d1\u539f\u5219\u3002<\/li>\n<\/ul>\n<h4>\u7406\u89e3\u548c\u4ea4\u4e92<\/h4>\n<ul>\n<li><strong>\u63a8\u8350\u7cfb\u7edf\uff1a<\/strong>\u5927\u8bed\u8a00\u6a21\u578b\u5df2\u6210\u4e3a\u63a8\u8350\u7cfb\u7edf\u7684\u53d8\u9769\u6027\u5f15\u64ce\uff0c\u80fd\u591f\u5206\u6790\u7528\u6237\u4ea4\u4e92\u3001\u5546\u54c1\u63cf\u8ff0\u4e0e\u8bc4\u8bba\uff0c\u63d0\u4f9b\u524d\u6240\u672a\u6709\u7684\u7ec6\u7c92\u5ea6\u4e2a\u6027\u5316\u5efa\u8bae\u3002\u540e\u8bad\u7ec3\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86\u6a21\u578b\u6574\u5408\u60c5\u611f\u5206\u6790\u7684\u80fd\u529b\uff0c\u4f7f\u5176\u5bf9\u5185\u5bb9\u4e0e\u60c5\u611f\u5185\u6db5\u7684\u7406\u89e3\u66f4\u4e3a\u7ec6\u81f4\uff0c\u8fd9\u4e00\u70b9\u5728 GPT-4\u3001LLaRA \u4e0e AgentRec \u7b49\u4e13\u4e1a\u7cfb\u7edf\u4e2d\u5747\u5df2\u5f97\u5230\u9a8c\u8bc1\u3002\u4e9a\u9a6c\u900a\u3001\u6dd8\u5b9d\u7b49\u7535\u5546\u5de8\u5934\u5229\u7528\u8fd9\u4e9b\u80fd\u529b\u5904\u7406\u8bc4\u8bba\u60c5\u611f\u3001\u641c\u7d22\u67e5\u8be2\u4e0e\u8d2d\u4e70\u5386\u53f2\uff0c\u4ece\u800c\u5b8c\u5584\u7528\u6237\u504f\u597d\u6a21\u578b\u5e76\u9ad8\u4fdd\u771f\u5730\u9884\u6d4b\u5174\u8da3\u3002\u9664\u5546\u54c1\u6392\u5e8f\u5916\uff0c\u7ecf\u540e\u8bad\u7ec3\u7684 LLM \u8fd8\u53ef\u8fdb\u884c\u5bf9\u8bdd\u5f0f\u63a8\u8350\u3001\u89c4\u5212\u4e0e\u5185\u5bb9\u751f\u6210\uff0c\u901a\u8fc7\u52a8\u6001\u3001\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u4ea4\u4e92\u63d0\u5347\u7528\u6237\u4f53\u9a8c\uff0c\u5145\u5206\u8bc1\u660e\u4e86\u540e\u8bad\u7ec3\u5728\u8fde\u63a5\u6570\u636e\u5206\u6790\u4e0e\u5b9e\u9645\u6548\u7528\u4e4b\u95f4\u7684\u6865\u6881\u4f5c\u7528\u3002<\/li>\n<li><strong>\u8bed\u97f3\u5bf9\u8bdd\uff1a<\/strong>\u7ecf\u540e\u8bad\u7ec3\u7684 LLM \u91cd\u65b0\u5b9a\u4e49\u4e86\u8bed\u97f3\u5904\u7406\uff0c\u5c06\u8bc6\u522b\u3001\u5408\u6210\u4e0e\u7ffb\u8bd1\u63d0\u5347\u81f3\u524d\u6240\u672a\u6709\u7684\u81ea\u7136\u5ea6\u4e0e\u51c6\u786e\u5ea6\u3002\u8fd9\u4e9b\u6a21\u578b\u53ef\u5b8c\u6210\u6587\u672c\u5230\u8bed\u97f3\u3001\u6587\u672c\u5230\u97f3\u9891\u751f\u6210\u4e0e\u8bed\u97f3\u8bc6\u522b\u7b49\u4efb\u52a1\uff0c\u4e3a Alexa\u3001Siri \u4e0e\u5929\u732b\u7cbe\u7075\u7b49\u65e0\u5904\u4e0d\u5728\u7684\u5de5\u5177\u63d0\u4f9b\u652f\u6301\u3002Whisper \u4ee5\u9ad8\u4fdd\u771f\u8f6c\u5f55\u4e3a\u4f8b\uff0cGPT-4o \u5219\u5f15\u5165\u5b9e\u65f6\u8bed\u97f3\u4ea4\u4e92\uff0c\u65e0\u7f1d\u878d\u5408\u591a\u6a21\u6001\u8f93\u5165\u3002\u672a\u6765\u65b9\u5411\u5305\u62ec\u591a\u8bed\u8a00\u7ffb\u8bd1\u4e0e\u4e2a\u6027\u5316\u8bed\u97f3\u5408\u6210\uff0c\u540e\u8bad\u7ec3\u5c06\u8fdb\u4e00\u6b65\u6253\u78e8 LLM\uff0c\u4ee5\u6253\u7834\u8bed\u8a00\u969c\u788d\u5e76\u4f9d\u636e\u4e2a\u4eba\u7528\u6237\u753b\u50cf\u5b9a\u5236\u56de\u590d\uff0c\u4ece\u800c\u63d0\u5347\u5168\u7403\u8303\u56f4\u5185\u4eba\u673a\u4ea4\u4e92\u7684\u53ef\u53ca\u6027\u4e0e\u53c2\u4e0e\u5ea6\u3002<\/li>\n<li><strong>\u89c6\u9891\u7406\u89e3\uff1a<\/strong>\u5c06 LLM \u6269\u5c55\u81f3\u89c6\u9891\u7406\u89e3\u9886\u57df\u6807\u5fd7\u7740\u4e00\u4e2a\u91cd\u8981\u524d\u6cbf\uff0c\u540e\u8bad\u7ec3\u4f7f Video-LLaMA \u7b49\u6a21\u578b\u80fd\u591f\u6267\u884c\u5b57\u5e55\u751f\u6210\u3001\u6458\u8981\u4e0e\u5185\u5bb9\u5206\u6790\uff0c\u4ece\u800c\u7b80\u5316\u591a\u5a92\u4f53\u521b\u4f5c\u4e0e\u7406\u89e3\u3002Sora \u901a\u8fc7\u6587\u672c\u63d0\u793a\u751f\u6210\u590d\u6742\u89c6\u9891\uff0c\u8fdb\u4e00\u6b65\u9769\u65b0\u8be5\u9886\u57df\uff0c\u964d\u4f4e\u6280\u672f\u95e8\u69db\u5e76\u4fc3\u8fdb\u521b\u65b0\u53d9\u4e8b\uff0c\u4f7f\u5185\u5bb9\u5236\u4f5c\u5927\u4f17\u5316\u3002\u8fd9\u4e9b\u8fdb\u5c55\u4f9d\u6258\u540e\u8bad\u7ec3\u8ba9 LLM \u9002\u5e94\u89c6\u89c9-\u65f6\u5e8f\u6570\u636e\uff0c\u63d0\u5347\u5176\u5728\u6559\u80b2\u5230\u5a31\u4e50\u7b49\u5e94\u7528\u4e2d\u7684\u89e3\u91ca\u6df1\u5ea6\u4e0e\u5b9e\u7528\u6027\u3002\u7136\u800c\uff0c\u5b83\u4eec\u5728\u8ba1\u7b97\u53ef\u6269\u5c55\u6027\u3001\u9690\u79c1\u4fdd\u62a4\u4e0e\u4f26\u7406\u6cbb\u7406\u65b9\u9762\u4ea6\u5e26\u6765\u6311\u6218\uff0c\u5c24\u5176\u662f\u751f\u6210\u9519\u8bef\u5185\u5bb9\u7684\u98ce\u9669\u3002\u968f\u7740\u540e\u8bad\u7ec3\u65b9\u6cd5\u7684\u6f14\u8fdb\uff0c\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u5c06\u6210\u4e3a\u5f53\u52a1\u4e4b\u6025\uff0c\u4ee5\u786e\u4fdd\u89c6\u9891\u76f8\u5173\u5e94\u7528\u53ef\u6301\u7eed\u3001\u8d1f\u8d23\u4efb\u7684\u90e8\u7f72\uff0c\u5728\u521b\u65b0\u4e0e\u793e\u4f1a\u8003\u91cf\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/li>\n<\/ul>\n<h3>\u672a\u89e3\u51b3\u7684\u95ee\u9898\u548c\u672a\u6765\u7684\u53d1\u5c55\u65b9\u5411<\/h3>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u6279\u5224\u6027\u5730\u8bc4\u4f30\u4e86\u7528\u4e8e\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u4e2d\u672a\u89e3\u51b3\u7684\u6311\u6218\u548c\u672a\u6765\u53d1\u5c55\u65b9\u5411\uff0c\u5e76\u5c06\u6211\u4eec\u7684\u5206\u6790\u7f6e\u4e8eOpenAI\u7684o1\u548cDeepSeek-R1\u53d1\u5e03\u6240\u9884\u793a\u7684\u53d8\u9769\u6027\u8fdb\u6b65\u4e4b\u4e2d\u3002\u8fd9\u4e9b\u6a21\u578b\u5229\u7528\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\uff0c\u91cd\u65b0\u5b9a\u4e49\u4e86\u63a8\u7406\u57fa\u51c6\uff0c\u4f46\u5b83\u4eec\u7684\u51fa\u73b0\u4e5f\u7a81\u663e\u4e86\u89e3\u51b3\u540e\u8bad\u7ec3\u6280\u672f\u4e2d\u6301\u7eed\u5b58\u5728\u7684\u5c40\u9650\u6027\u7684\u7d27\u8feb\u6027\u3002\u4ee5\u4e0b\u5404\u5c0f\u8282\u9610\u8ff0\u4e86\u4e03\u4e2a\u5173\u952e\u7684\u672a\u89e3\u51b3\u95ee\u9898\uff0c\u6bcf\u4e2a\u95ee\u9898\u90fd\u5f3a\u8c03\u4e86\u5176\u5bf9\u8be5\u9886\u57df\u8fdb\u5c55\u7684\u5173\u952e\u91cd\u8981\u6027\u4ee5\u53ca\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u7684\u8feb\u5207\u9700\u8981\uff0c\u5e76\u63d0\u51fa\u4e86\u53ef\u884c\u7684\u7b56\u7565\uff0c\u4ee5\u63a8\u52a8\u672a\u6765\u7684\u7814\u7a76\uff0c\u5e76\u786e\u4fddLLM\u5728\u4e0d\u540c\u5e94\u7528\u4e2d\u7684\u8d1f\u8d23\u4efb\u53d1\u5c55\u3002<\/p>\n<ul>\n<li><strong>\u8d85\u8d8a\u5927\u89c4\u6a21RL\u7684\u63a8\u7406\u589e\u5f3a\uff1a<\/strong>o1\u548cDeepSeek-R1\u7684\u5f15\u5165\u6807\u5fd7\u7740LLM\u63a8\u7406\u80fd\u529b\u7684\u8303\u5f0f\u8f6c\u53d8\uff0c\u5229\u7528\u4e86\u5e7f\u6cdb\u7684RL\u6846\u67b6\uff08\u5982RLHF\u548c\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316GRPO\uff09\uff0c\u5728\u6570\u5b66\u8bc1\u660e\u548c\u903b\u8f91\u63a8\u5bfc\u7b49\u591a\u6b65\u95ee\u9898\u89e3\u51b3\u4e2d\u5b9e\u73b0\u4e86\u524d\u6240\u672a\u6709\u7684\u51c6\u786e\u6027\u3002\u7136\u800c\uff0c\u5bf9\u4e8c\u5143\u5956\u52b1\u4fe1\u53f7\u548c\u5927\u91cf\u4eba\u5de5\u53cd\u9988\u7684\u4f9d\u8d56\u66b4\u9732\u4e86\u4e00\u4e2a\u5173\u952e\u5c40\u9650\uff1a\u5b83\u4eec\u65e0\u6cd5\u5728\u79d1\u5b66\u5047\u8bbe\u751f\u6210\u6216\u52a8\u6001\u73af\u5883\u4e2d\u7684\u6218\u7565\u51b3\u7b56\u7b49\u590d\u6742\u3001\u5f00\u653e\u5f0f\u4efb\u52a1\u4e2d\u6709\u6548\u6cdb\u5316\u3002\u8fd9\u79cd\u5dee\u8ddd\u8feb\u5728\u7709\u776b\uff0c\u56e0\u4e3aLLM\u5728\u73b0\u5b9e\u4e16\u754c\u73af\u5883\u4e2d\u6a21\u62df\u7c7b\u4eba\u63a8\u7406\u7684\u9700\u6c42\u4e0d\u65ad\u589e\u957f\uff0c\u5176\u91cd\u8981\u6027\u5728\u4e8e\u91ca\u653e\u5b83\u4eec\u4f5c\u4e3a\u8d85\u8d8a\u5f53\u524d\u57fa\u51c6\u7684\u81ea\u4e3b\u667a\u80fd\u4f53\u7684\u6f5c\u529b\u3002\u5f53\u524d\u7684RL\u65b9\u6cd5\u96be\u4ee5\u5904\u7406\u5956\u52b1\u7a00\u758f\u6027\uff0c\u5e76\u4e14\u7f3a\u4e4f\u5bf9\u4efb\u52a1\u590d\u6742\u6027\u7684\u9002\u5e94\u6027\uff0c\u8fd9\u9700\u8981\u521b\u65b0\u7684\u6846\u67b6\u3002\u53ef\u884c\u7684\u89e3\u51b3\u65b9\u6848\u5305\u62ec\u5f00\u53d1\u591a\u76ee\u6807RL\u7cfb\u7edf\uff0c\u6574\u5408\u81ea\u6211\u76d1\u7763\u7684\u4e00\u81f4\u6027\u68c0\u67e5\uff08\u4f8b\u5982\u9a8c\u8bc1\u63a8\u7406\u6b65\u9aa4\u4e4b\u95f4\u7684\u903b\u8f91\u8fde\u8d2f\u6027\uff09\u548c\u7279\u5b9a\u9886\u57df\u7684\u5148\u9a8c\u77e5\u8bc6\uff08\u5982\u6570\u5b66\u516c\u7406\u6216\u79d1\u5b66\u539f\u7406\uff09\uff0c\u4ee5\u5728\u6ca1\u6709\u8be6\u5c3d\u4eba\u5de5\u6807\u6ce8\u7684\u60c5\u51b5\u4e0b\u6307\u5bfc\u63a8\u7406\u3002\u8fd9\u79cd\u8fdb\u6b65\u53ef\u4ee5\u51cf\u5c11\u5bf9\u4ee3\u4ef7\u9ad8\u6602\u7684\u53cd\u9988\u5faa\u73af\u7684\u4f9d\u8d56\uff0c\u63d0\u9ad8\u53ef\u6269\u5c55\u6027\uff0c\u5e76\u4f7fLLM\u80fd\u591f\u5e94\u5bf9\u672a\u77e5\u7684\u63a8\u7406\u9886\u57df\uff0cDeepSeek-R1\u7684\u51b7\u542f\u52a8RL\u521b\u65b0\u4f7f\u8fd9\u79cd\u524d\u666f\u6210\u4e3a\u53ef\u80fd\u3002<\/li>\n<li><strong>\u4e0b\u4e00\u4ee3LLM\u7684\u540e\u8bad\u7ec3\u53ef\u6269\u5c55\u6027\uff1a<\/strong>\u968f\u7740LLM\u7684\u89c4\u6a21\u548c\u590d\u6742\u6027\u4e0d\u65ad\u5347\u7ea7\uff0c\u8fd9\u5728\u4e0b\u4e00\u4ee3\u6a21\u578b\u7684\u53c2\u6570\u5bc6\u96c6\u578b\u67b6\u6784\u4e2d\u5f97\u5230\u4e86\u4f53\u73b0\uff0c\u540e\u8bad\u7ec3\u7684\u53ef\u6269\u5c55\u6027\u6210\u4e3a\u4e00\u4e2a\u4ee4\u4eba\u754f\u60e7\u4e14\u7d27\u8feb\u7684\u6311\u6218\u3002\u57fa\u4e8eRL\u7684\u65b9\u6cd5\uff08\u5982DeepSeek-R1\u7684\u51b7\u542f\u52a8\u65b9\u6cd5\uff09\u7684\u8d44\u6e90\u5bc6\u96c6\u578b\u7279\u6027\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u57fa\u7840\u8bbe\u65bd\uff0c\u8fd9\u9650\u5236\u4e86\u5bf9\u8d44\u91d1\u5145\u8db3\u7684\u5b9e\u4f53\u7684\u8bbf\u95ee\uff0c\u5e76\u5f15\u53d1\u4e86\u91cd\u5927\u7684\u53ef\u6301\u7eed\u6027\u95ee\u9898\uff0c\u5c24\u5176\u662f\u5728\u591a\u6a21\u6001\u5e94\u7528\uff08\u5982\u89c6\u9891\u5206\u6790\uff09\u548c\u5b9e\u65f6\u7cfb\u7edf\uff08\u5982\u5bf9\u8bdd\u4ee3\u7406\uff09\u4e2d\u3002\u8fd9\u4e2a\u95ee\u9898\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u5b83\u6709\u53ef\u80fd\u6269\u5927\u8d44\u6e90\u4e30\u5bcc\u548c\u8d44\u6e90\u53d7\u9650\u7684\u7814\u7a76\u793e\u533a\u4e4b\u95f4\u7684\u5dee\u8ddd\uff0c\u4ece\u800c\u963b\u788dLLM\u53d1\u5c55\u7684\u516c\u5e73\u8fdb\u5c55\u3002\u867d\u7136\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff08PEFT\uff09\u51cf\u8f7b\u4e86\u4e00\u4e9b\u5f00\u9500\uff0c\u4f46\u5176\u6027\u80fd\u901a\u5e38\u4f1a\u5728\u5927\u89c4\u6a21\u6570\u636e\u96c6\u4e0a\u4e0b\u964d\uff0c\u8fd9\u7a81\u663e\u4e86\u5bf9\u53ef\u6269\u5c55\u66ff\u4ee3\u65b9\u6848\u7684\u9700\u6c42\u3002\u53ef\u884c\u7684\u672a\u6765\u65b9\u5411\u5305\u62ec\u8bbe\u8ba1\u8f7b\u91cf\u7ea7RL\u7b97\u6cd5\uff08\u53ef\u80fd\u8c03\u6574GRPO\u4ee5\u51cf\u5c11\u5185\u5b58\u5360\u7528\uff09\u3001\u8054\u90a6\u540e\u8bad\u7ec3\u6846\u67b6\uff08\u5c06\u8ba1\u7b97\u8d1f\u8f7d\u5206\u5e03\u5728\u53bb\u4e2d\u5fc3\u5316\u7f51\u7edc\u4e0a\uff09\uff0c\u4ee5\u53ca\u5148\u8fdb\u7684\u84b8\u998f\u6280\u672f\uff08\u5728\u6700\u5927\u9650\u5ea6\u5730\u51cf\u5c11\u8d44\u6e90\u9700\u6c42\u7684\u540c\u65f6\u4fdd\u7559\u63a8\u7406\u548c\u9002\u5e94\u6027\uff09\u3002\u5982\u679c\u8fd9\u4e9b\u89e3\u51b3\u65b9\u6848\u5f97\u4ee5\u5b9e\u73b0\uff0c\u53ef\u80fd\u4f1a\u4f7f\u540e\u8bad\u7ec3\u6c11\u4e3b\u5316\uff0c\u8fd9\u4e0e\u8be5\u9886\u57df\u5bf9\u53ef\u6301\u7eed\u548c\u5305\u5bb9\u6027\u521b\u65b0\u7684\u8feb\u5207\u9700\u6c42\u76f8\u4e00\u81f4\u3002<\/li>\n<li><strong>\u57fa\u4e8eRL\u9a71\u52a8\u7684\u6a21\u578b\u7684\u4f26\u7406\u5bf9\u9f50\u548c\u504f\u5dee\u7f13\u89e3\uff1a<\/strong>\u5982o1\u7684\u8c28\u614e\u5bf9\u9f50\u7b56\u7565\u6240\u793a\uff0c\u901a\u8fc7RL\u8fdb\u884c\u540e\u8bad\u7ec3\u4f1a\u653e\u5927\u4f26\u7406\u98ce\u9669\uff0c\u56e0\u4e3a\u5b83\u53ef\u80fd\u4f1a\u5f3a\u5316\u5d4c\u5165\u5728\u8bad\u7ec3\u6570\u636e\u96c6\uff08\u5982HH-RLHF\u6216\u5408\u6210\u8bed\u6599\u5e93\uff09\u4e2d\u7684\u504f\u5dee\uff0c\u9274\u4e8eLLM\u5728\u533b\u7597\u4fdd\u5065\u8bca\u65ad\u548c\u53f8\u6cd5\u51b3\u7b56\u7b49\u654f\u611f\u9886\u57df\u4e2d\u7684\u90e8\u7f72\uff0c\u8fd9\u662f\u4e00\u4e2a\u81f3\u5173\u91cd\u8981\u7684\u7d27\u8feb\u6311\u6218\u3002\u4f26\u7406\u5bf9\u9f50\u7684\u52a8\u6001\u53ef\u53d8\u6027\uff08\u5176\u4e2d\u5728\u4e00\u4e2a\u6587\u5316\u80cc\u666f\u4e0b\u7684\u516c\u5e73\u53ef\u80fd\u5728\u53e6\u4e00\u4e2a\u6587\u5316\u80cc\u666f\u4e0b\u6784\u6210\u504f\u5dee\uff09\u5bf9\u5b9e\u73b0\u666e\u904d\u503c\u5f97\u4fe1\u8d56\u7684LLM\u6784\u6210\u4e86\u91cd\u5927\u969c\u788d\uff0c\u8fd9\u4f7f\u5f97\u8fd9\u4e2a\u95ee\u9898\u5bf9\u4e8e\u786e\u4fdd\u516c\u5e73\u548c\u5b89\u5168\u7684AI\u7cfb\u7edf\u81f3\u5173\u91cd\u8981\u3002\u5f53\u524d\u65b9\u6cd5\u5b58\u5728\u8fc7\u5ea6\u5ba1\u67e5\u7684\u98ce\u9669\uff08\u4ece\u800c\u635f\u5bb3\u5b9e\u7528\u6027\uff0c\u5982\u627c\u6740\u521b\u610f\u8f93\u51fa\uff09\u6216\u7ea0\u6b63\u4e0d\u8db3\uff08\u4ece\u800c\u5ef6\u7eed\u6709\u5bb3\u7684\u504f\u89c1\uff0c\u5982\u79cd\u65cf\u6216\u6027\u522b\u5dee\u5f02\uff09\u3002\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u9700\u8981\u5f00\u53d1\u516c\u5e73\u611f\u77e5RL\u76ee\u6807\uff0c\u7eb3\u5165\u591a\u65b9\u5229\u76ca\u76f8\u5173\u8005\u504f\u597d\u6a21\u578b\uff08\u5982\u805a\u5408\u591a\u6837\u5316\u7684\u4eba\u7c7b\u5224\u65ad\uff09\u548c\u5bf9\u6297\u6027\u53bb\u504f\u89c1\u6280\u672f\uff0c\u4ee5\u5728\u4e2d\u8bad\u7ec3\u671f\u95f4\u6d88\u9664\u6570\u636e\u96c6\u504f\u5dee\u3002\u8fd9\u4e9b\u65b9\u6cd5\u7684\u53ef\u884c\u6027\u5f97\u5230\u4e86\u53ef\u89e3\u91ca\u6027\u5de5\u5177\u548c\u591a\u76ee\u6807\u4f18\u5316\u7684\u6700\u65b0\u8fdb\u5c55\u7684\u652f\u6301\uff0c\u4ece\u800c\u80fd\u591f\u5728\u4f26\u7406\u7a33\u5065\u6027\u548c\u5b9e\u7528\u529f\u80fd\u4e4b\u95f4\u5b9e\u73b0\u5e73\u8861\u7684\u6743\u8861\uff0co1\u7684\u5b9e\u9645\u90e8\u7f72\u6311\u6218\u7a81\u663e\u4e86\u8fd9\u4e00\u5fc5\u8981\u6027\u3002<\/li>\n<li><strong>\u7528\u4e8e\u6574\u4f53\u63a8\u7406\u7684\u65e0\u7f1d\u591a\u6a21\u6001\u96c6\u6210\uff1a<\/strong>\u7531o1\u7684\u63a8\u7406\u589e\u5f3a\u548cGPT-4o\u7684\u5408\u6210\u80fd\u529b\u9884\u793a\u7684\u591a\u6a21\u6001LLM\u53d1\u5c55\u8f68\u8ff9\uff0c\u5f3a\u8c03\u4e86\u5bf9\u540e\u8bad\u7ec3\u65b9\u6cd5\u7684\u8feb\u5207\u9700\u6c42\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u53ef\u4ee5\u65e0\u7f1d\u96c6\u6210\u6587\u672c\u3001\u56fe\u50cf\u3001\u97f3\u9891\u548c\u5176\u4ed6\u6570\u636e\u7c7b\u578b\uff0c\u4ece\u800c\u5b9e\u73b0\u6574\u4f53\u63a8\u7406\u2014\u2014\u8fd9\u662f\u5bf9\u5b9e\u65f6\u89c6\u9891\u5206\u6790\u3001\u589e\u5f3a\u73b0\u5b9e\u548c\u8de8\u6a21\u6001\u79d1\u5b66\u7814\u7a76\u7b49\u5e94\u7528\u81f3\u5173\u91cd\u8981\u7684\u80fd\u529b\u3002\u7531\u4e8e\u6570\u636e\u5f02\u8d28\u6027\u548c\u5168\u9762\u7684\u591a\u6a21\u6001\u8bad\u7ec3\u8bed\u6599\u5e93\u7684\u7a00\u7f3a\u6027\uff0c\u5f53\u524d\u65b9\u6cd5\u5728\u5b9e\u73b0\u7a33\u5065\u7684\u8de8\u6a21\u6001\u5bf9\u9f50\u65b9\u9762\u6b65\u5c65\u7ef4\u8270\uff0c\u8fd9\u9650\u5236\u4e86LLM\u5728\u4e0d\u540c\u8f93\u5165\u4e4b\u95f4\u8fdb\u884c\u8fde\u8d2f\u63a8\u7406\u7684\u80fd\u529b\u3002\u8fd9\u4e00\u6311\u6218\u7684\u91cd\u8981\u6027\u5728\u4e8e\u5176\u91ca\u653e\u53d8\u9769\u6027\u5e94\u7528\u7684\u6f5c\u529b\uff0c\u7136\u800c\uff0c\u5982\u679c\u6ca1\u6709\u53ef\u6269\u5c55\u7684\u6846\u67b6\uff0c\u5176\u89e3\u51b3\u65b9\u6848\u4ecd\u7136\u96be\u4ee5\u6349\u6478\u3002DeepSeek-R1\u7684\u51b7\u542f\u52a8RL\u63d0\u4f9b\u4e86\u4e00\u4e2a\u6709\u5e0c\u671b\u7684\u8d77\u70b9\uff0c\u8fd9\u8868\u660e\u7edf\u4e00\u7684\u6a21\u6001\u7f16\u7801\u5668\uff08\u80fd\u591f\u5c06\u5f02\u6784\u6570\u636e\u7f16\u7801\u5230\u5171\u4eab\u7684\u6f5c\u5728\u7a7a\u95f4\u4e2d\uff09\u548c\u52a8\u6001RL\u7b56\u7565\uff08\u53ef\u4ee5\u81ea\u9002\u5e94\u5730\u6743\u8861\u6a21\u6001\u8d21\u732e\uff09\u4ece\u800c\u5f25\u5408\u8fd9\u4e00\u5dee\u8ddd\u3002\u672a\u6765\u7814\u7a76\u5e94\u4f18\u5148\u8003\u8651\u521b\u5efa\u591a\u6a21\u6001\u57fa\u51c6\u548c\u5408\u6210\u6570\u636e\u96c6\uff0c\u5728Magpie\u7b49\u5de5\u4f5c\u7684\u57fa\u7840\u4e0a\u63a8\u52a8\u8fdb\u5c55\uff0c\u9274\u4e8e\u591a\u6a21\u6001\u9884\u8bad\u7ec3\u548cRL\u4f18\u5316\u7684\u6700\u65b0\u8fdb\u5c55\uff0c\u8fd9\u662f\u4e00\u9879\u53ef\u884c\u7684\u5de5\u4f5c\u3002<\/li>\n<li><strong>\u4e0a\u4e0b\u6587\u81ea\u9002\u5e94\u53ef\u4fe1\u5ea6\u6846\u67b6\uff1a<\/strong>\u540e\u8bad\u7ec3LLM\u4e2d\u7684\u53ef\u4fe1\u5ea6\u8d8a\u6765\u8d8a\u88ab\u8ba4\u4e3a\u662f\u52a8\u6001\u7684\u3001\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u5c5e\u6027\uff0c\u800c\u4e0d\u662f\u9759\u6001\u7684\u8d28\u91cf\uff0co1\u5728\u6559\u80b2\u7b49\u654f\u611f\u9886\u57df\u8c28\u614e\u8f93\u51fa\uff0c\u800c\u5728\u521b\u610f\u4efb\u52a1\u4e2d\u66f4\u81ea\u7531\u7684\u54cd\u5e94\u5c31\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9\u3002\u8fd9\u79cd\u53ef\u53d8\u6027\uff08\u5176\u4e2d\u5b89\u5168\u8981\u6c42[\u5982\u907f\u514d\u5728\u6559\u80b2\u73af\u5883\u4e2d\u4f20\u64ad\u865a\u5047\u4fe1\u606f]\u53ef\u80fd\u4e0e\u6548\u7528\u9700\u6c42[\u5982\u5728\u5199\u4f5c\u4e2d\u57f9\u517b\u521b\u9020\u529b\u76f8\u51b2\u7a81])\u63d0\u51fa\u4e86\u4e00\u4e2a\u7d27\u8feb\u7684\u6311\u6218\uff0c\u56e0\u4e3a\u5b83\u5bf9\u7528\u6237\u4fe1\u4efb\u548cLLM\u5728\u4e0d\u540c\u73b0\u5b9e\u4e16\u754c\u573a\u666f\u4e2d\u7684\u9002\u7528\u6027\u81f3\u5173\u91cd\u8981\u3002\u5f53\u524d\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u901a\u5e38\u8fc7\u5206\u4f18\u5148\u8003\u8651\u5b89\u5168\u6027\uff08\u4ece\u800c\u4ea7\u751f\u964d\u4f4e\u5b9e\u9645\u4ef7\u503c\u7684\u6548\u7528\u6743\u8861\uff09\u6216\u8005\u672a\u80fd\u9002\u5e94\u7279\u5b9a\u4e8e\u4e0a\u4e0b\u6587\u7684\u9700\u6c42\uff08\u4ece\u800c\u635f\u5bb3\u53ef\u9760\u6027\uff09\u3002\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u9700\u8981\u4e0a\u4e0b\u6587\u76f8\u5173\u7684RL\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u53ef\u4ee5\u52a8\u6001\u8c03\u6574\u5b89\u5168\u6027\u548c\u6548\u7528\u4e4b\u95f4\u7684\u6743\u8861\uff0c\u5229\u7528\u5b9e\u65f6\u7528\u6237\u53cd\u9988\u548c\u53ef\u89e3\u91ca\u7684\u5b89\u5168\u6307\u6807\uff08\u5982\u751f\u6210\u8f93\u51fa\u7684\u900f\u660e\u5ea6\u5206\u6570\uff09\u6765\u786e\u4fdd\u9002\u5e94\u6027\u3002\u8fd9\u79cd\u65b9\u6cd5\u7684\u53ef\u884c\u6027\u5f97\u5230\u4e86\u81ea\u9002\u5e94\u5b66\u4e60\u7cfb\u7edf\u548c\u5b9e\u65f6\u76d1\u63a7\u7684\u8fdb\u6b65\u7684\u652f\u6301\uff0c\u4e3a\u5e73\u8861\u53ef\u4fe1\u5ea6\u4e0e\u529f\u80fd\u6027\u63d0\u4f9b\u4e86\u4e00\u6761\u9014\u5f84\uff0c\u968f\u7740o1\u7b49LLM\u6269\u5c55\u5230\u9ad8\u98ce\u9669\u5e94\u7528\uff0c\u8fd9\u662f\u4e00\u4e2a\u8feb\u5207\u7684\u9700\u6c42\u3002<\/li>\n<li><strong>\u540e\u8bad\u7ec3\u521b\u65b0\u7684\u53ef\u8bbf\u95ee\u6027\u548c\u6c11\u4e3b\u5316\uff1a<\/strong>\u5148\u8fdb\u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\uff08\u4ee5DeepSeek-R1\u7684RL\u9a71\u52a8\u65b9\u6cd5\u4e3a\u4ee3\u8868\uff09\u7684\u8ba1\u7b97\u5f3a\u5ea6\u9650\u5236\u4e86\u5b83\u4eec\u7684\u5e94\u7528\u4e8e\u8d44\u6e90\u4e30\u5bcc\u7684\u5b9e\u4f53\uff0c\u8fd9\u7ed9\u53ef\u8bbf\u95ee\u6027\u5e26\u6765\u4e86\u8feb\u8fd1\u7684\u969c\u788d\uff0c\u627c\u6740\u4e86\u8f83\u5c0f\u7814\u7a76\u793e\u533a\u548c\u884c\u4e1a\u90e8\u95e8\u7684\u521b\u65b0\uff08\u5373\u5bf9\u4e8e\u4fc3\u8fdb\u4eba\u5de5\u667a\u80fd\u9886\u57df\u516c\u5e73\u8fdb\u6b65\u81f3\u5173\u91cd\u8981\u7684\u95ee\u9898\uff09\u3002\u8fd9\u79cd\u6392\u4ed6\u6027\u4e0d\u4ec5\u9650\u5236\u4e86\u8d21\u732e\u7684\u591a\u6837\u6027\uff0c\u800c\u4e14\u963b\u788d\u4e86\u8be5\u9886\u57df\u534f\u4f5c\u5e94\u5bf9\u5168\u7403\u6311\u6218\u7684\u80fd\u529b\u3002\u5c06\u8fd9\u4e9b\u521b\u65b0\u666e\u53ca\u5316\u9700\u8981\u5f00\u53d1\u9ad8\u6548\u3001\u5f00\u6e90\u7684\u5de5\u5177\u548c\u6846\u67b6\uff0c\u8fd9\u4e9b\u5de5\u5177\u548c\u6846\u67b6\u53ef\u4ee5\u5728\u4e0d\u635f\u5bb3\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\u964d\u4f4e\u51c6\u5165\u95e8\u69db\uff0c\u901a\u8fc7\u53ef\u6269\u5c55\u7684PEFT\u9002\u5e94RL\u3001\u7528\u4e8e\u5171\u4eab\u540e\u8bad\u7ec3\u6a21\u578b\u7684\u534f\u4f5c\u5e73\u53f0\uff08\u5982Hugging Face hubs\uff09\u4ee5\u53ca\u7c7b\u4f3c\u4e8eMagpie\u7684\u7cbe\u7b80\u5408\u6210\u6570\u636e\u751f\u6210\u6d41\u7a0b\u6765\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\u3002\u672a\u6765\u7684\u52aa\u529b\u5e94\u4fa7\u91cd\u4e8e\u4f18\u5316\u8fd9\u4e9b\u89e3\u51b3\u65b9\u6848\uff0c\u4ee5\u5b9e\u73b0\u5e7f\u6cdb\u5e94\u7528\uff0c\u786e\u4fdd\u4ee5o1\u548cDeepSeek-R1\u4e3a\u4ee3\u8868\u7684\u540e\u8bad\u7ec3\u7684\u53d8\u9769\u6f5c\u529b\u5ef6\u4f38\u5230\u7cbe\u82f1\u673a\u6784\u4e4b\u5916\uff0c\u4ee5\u4e30\u5bcc\u66f4\u5e7f\u6cdb\u7684\u4eba\u5de5\u667a\u80fd\u751f\u6001\u7cfb\u7edf\u3002<\/li>\n<li><strong>\u521b\u9020\u6027\u667a\u80fd\u4e0e\u7cfb\u7edf2\u601d\u7ef4\uff1a<\/strong>\u5c06\u521b\u9020\u6027\u667a\u80fd\u878d\u5165\u7cfb\u7edf2\u63a8\u7406\u662fLLM\u53d1\u5c55\u4e2d\u7684\u4e00\u4e2a\u65b0\u5174\u524d\u6cbf\u3002\u867d\u7136\u50cfOpenAI\u7684o1\u548cDeepSeek\u7684R1\u8fd9\u6837\u7684\u63a8\u7406LLM\u5728\u6df1\u601d\u719f\u8651\u7684\u3001\u5faa\u5e8f\u6e10\u8fdb\u7684\u903b\u8f91\u5206\u6790\uff08\u6a21\u4eff\u7cfb\u7edf2\u601d\u7ef4\uff09\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u4f46\u5b83\u4eec\u5728\u521b\u9020\u6027\u667a\u80fd\u65b9\u9762\u7684\u80fd\u529b\uff08\u5305\u62ec\u751f\u6210\u65b0\u60f3\u6cd5\u3001\u7efc\u5408\u4e0d\u540c\u7684\u6982\u5ff5\u4ee5\u53ca\u7075\u6d3b\u5730\u9002\u5e94\u975e\u7ed3\u6784\u5316\u95ee\u9898\uff09\u4ecd\u6709\u5f85\u63a2\u7d22\u3002\u8fd9\u4e00\u5dee\u8ddd\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u521b\u9020\u6027\u667a\u80fd\u662f\u4eba\u7c7b\u5728\u827a\u672f\u521b\u4f5c\u3001\u79d1\u5b66\u53d1\u73b0\u548c\u6218\u7565\u521b\u65b0\u7b49\u9886\u57df\u7684\u7c7b\u4eba\u95ee\u9898\u89e3\u51b3\u7684\u57fa\u7840\uff0c\u800c\u4ec5\u4ec5\u4f9d\u9760\u50f5\u5316\u7684\u903b\u8f91\u6846\u67b6\u662f\u4e0d\u591f\u7684\u3002\u8fd9\u4e00\u6311\u6218\u7684\u7d27\u8feb\u6027\u5728\u4e8e\u5b83\u6709\u53ef\u80fd\u5c06LLM\u4ece\u5206\u6790\u5de5\u5177\u63d0\u5347\u4e3a\u81ea\u4e3b\u7684\u521b\u9020\u6027\u4e3b\u4f53\uff0c\u8fd9\u662f\u8fc8\u5411\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff08AGI\uff09\u7684\u53d8\u9769\u6027\u98de\u8dc3\u3002\u5728\u4e0b\u6587\u4e2d\uff0c\u6211\u4eec\u6982\u8ff0\u4e86\u8fd9\u4e00\u5f00\u653e\u6027\u95ee\u9898\uff0c\u5e76\u63d0\u51fa\u4e86\u672a\u6765\u7684\u65b9\u5411\uff0c\u501f\u9274\u4e86\u8be5\u8c03\u67e5\u7684\u89c1\u89e3\u3002<\/li>\n<\/ul>\n<h3>\u7ed3\u8bba<\/h3>\n<p>\u672c\u6587\u63d0\u4f9b\u4e86\u5bf9\u540e\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff08PoLM\uff09\u7684\u9996\u6b21\u8be6\u5c3d\u8c03\u67e5\uff0c\u7cfb\u7edf\u5730\u8ffd\u6eaf\u4e86\u5b83\u4eec\u4ece2018\u5e74ChatGPT\u7684\u5bf9\u9f50\u8d77\u6e90\u52302025\u5e74DeepSeek-R1\u7684\u63a8\u7406\u91cc\u7a0b\u7891\u7684\u8f68\u8ff9\uff0c\u5e76\u80af\u5b9a\u4e86\u5b83\u4eec\u5bf9\u63a8\u7406\u7cbe\u5ea6\u3001\u9886\u57df\u9002\u5e94\u6027\u548c\u4f26\u7406\u5b8c\u6574\u6027\u7684\u53d8\u9769\u6027\u5f71\u54cd\u3002 \u6211\u4eec\u8bc4\u4f30\u4e86\u5e7f\u6cdb\u7684\u6280\u672f\uff08\u5373\uff0c\u5fae\u8c03\u3001\u5bf9\u9f50\u3001\u63a8\u7406\u3001\u6548\u7387\u4ee5\u53ca\u96c6\u6210\u548c\u9002\u5e94\uff09\uff0c\u7efc\u5408\u4e86\u5b83\u4eec\u5728\u4e13\u4e1a\u3001\u6280\u672f\u548c\u4ea4\u4e92\u9886\u57df\uff08\u4ece\u6cd5\u5f8b\u5206\u6790\u5230\u591a\u6a21\u6001\u7406\u89e3\uff09\u7684\u8d21\u732e\u3002 \u6211\u4eec\u7684\u5206\u6790\u5f3a\u8c03\uff0cPoLM\u663e\u7740\u63d0\u5347\u4e86LLM\u7684\u80fd\u529b\uff0c\u4ece\u6700\u521d\u7684\u5bf9\u9f50\u521b\u65b0\u53d1\u5c55\u5230\u590d\u6742\u7684\u63a8\u7406\u6846\u67b6\uff1b\u5c3d\u7ba1\u5982\u6b64\uff0c\u5b83\u63ed\u793a\u4e86\u6301\u7eed\u7684\u6311\u6218\uff0c\u5305\u62ec\u504f\u89c1\u6301\u7eed\u5b58\u5728\u3001\u8ba1\u7b97\u53ef\u6269\u5c55\u6027\u548c\u4e0a\u4e0b\u6587\u53ef\u53d8\u4f26\u7406\u5bf9\u9f50\u3002 \u8fd9\u4e9b\u53d1\u73b0\u88ab\u5c01\u88c5\u5728\u4e00\u4e2a\u65b0\u7684\u5206\u7c7b\u6cd5\u4e2d\uff0c\u5f3a\u8c03\u4e86\u91c7\u7528\u6574\u5408\u65b9\u6cd5\u7684\u5fc5\u8981\u6027\uff0c\u5c06\u63a8\u7406\u8fdb\u5c55\u4e0e\u6548\u7387\u548c\u4f26\u7406\u8981\u6c42\u76f8\u7ed3\u5408\u3002 \u6211\u4eec\u5f97\u51fa\u7ed3\u8bba\uff0c\u6301\u7eed\u7684\u8de8\u5b66\u79d1\u5408\u4f5c\u3001\u4e25\u683c\u7684\u65b9\u6cd5\u8bc4\u4f30\u4ee5\u53ca\u81ea\u9002\u5e94\u3001\u53ef\u6269\u5c55\u6846\u67b6\u7684\u5f00\u53d1\u5bf9\u4e8e\u5c06LLM\u7684\u6f5c\u529b\u5b9e\u73b0\u4e3a\u5728\u5404\u79cd\u5e94\u7528\u4e2d\u53ef\u9760\u3001\u8d1f\u8d23\u4efb\u7684\u5de5\u5177\u81f3\u5173\u91cd\u8981\u3002 \u4f5c\u4e3a\u540c\u7c7b\u7814\u7a76\u7684\u5148\u9a71\u6027\u7efc\u8ff0\uff0c\u8fd9\u9879\u5de5\u4f5c\u6574\u5408\u4e86\u8fd1\u5e74\u6765 PoLMs \u7684\u8fdb\u5c55\uff0c\u5e76\u5960\u5b9a\u4e86\u575a\u5b9e\u7684\u5b66\u672f\u57fa\u7840\uff0c\u6fc0\u52b1\u672a\u6765\u7684\u7814\u7a76\u57f9\u517b LLM\uff0c\u4f7f\u5176\u80fd\u591f\u5de7\u5999\u5730\u6574\u5408\u7cbe\u786e\u6027\u3001\u4f26\u7406\u7a33\u5065\u6027\u548c\u591a\u529f\u80fd\u6027\uff0c\u4ee5\u6ee1\u8db3\u79d1\u5b66\u548c\u793e\u4f1a\u80cc\u666f\u4e0b\u4e0d\u65ad\u53d8\u5316\u7684\u9700\u6c42\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Meta Data \u53d1\u8868\u65f6\u95f4 2025.08.1 \u4f5c\u8005\uff1aGuiyao Tie, Zeli Zhao, Dingjie Song e &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[60,41],"tags":[62,63,61],"class_list":["post-1301","post","type-post","status-publish","format-standard","hentry","category-paper_read","category-learning_note","tag-llm","tag-63","tag-61"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1301","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1301"}],"version-history":[{"count":52,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1301\/revisions"}],"predecessor-version":[{"id":1424,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1301\/revisions\/1424"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1301"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1301"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1301"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}