{"id":622,"date":"2024-02-29T15:56:19","date_gmt":"2024-02-29T07:56:19","guid":{"rendered":"http:\/\/lingbo.online\/?p=622"},"modified":"2024-03-06T10:24:07","modified_gmt":"2024-03-06T02:24:07","slug":"charactereval","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/algorithm_learning\/llms\/charactereval\/","title":{"rendered":"[\u8bba\u6587\u7b14\u8bb0]CharacterEval: A Chinese Benchmark for Role-Playing Conversational Agent Evaluation"},"content":{"rendered":"<h1>\u6458\u8981<\/h1>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u7684\u51fa\u73b0\u6539\u53d8\u4e86\u751f\u6210\u5f0fAgent\uff0c\u5176\u4e2d\u6700\u8fd1\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b(llm)\u7684\u51fa\u73b0\u5f7b\u5e95\u6539\u53d8\u4e86\u751f\u6210\u4ee3\u7406\u3002\u5176\u4e2d\uff0c<strong>\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4ee3\u7406(Role-Playing Conversational Agents, RPCAs)<\/strong>\u56e0\u5176\u80fd\u591f\u4e0e\u7528\u6237\u8fdb\u884c\u60c5\u611f\u4e92\u52a8\u800c\u5907\u53d7\u5173\u6ce8\u3002\u7136\u800c\uff0c\u7f3a\u4e4f\u4e00\u4e2a\u5168\u9762\u7684\u57fa\u51c6\u963b\u788d\u4e86\u8fd9\u4e00\u9886\u57df\u7684\u8fdb\u5c55\u3002\u4e3a\u4e86\u5f25\u8865\u8fd9\u4e00\u5dee\u8ddd\uff0c\u6211\u4eec\u5f15\u5165\u4e86CharacterEval\uff0c\u8fd9\u662f\u4e00\u4e2a\u5168\u9762\u8bc4\u4f30RPCA\u7684\u4e2d\u56fd\u57fa\u51c6\uff0c\u5e76\u8f85\u4ee5\u91cf\u8eab\u5b9a\u5236\u7684\u9ad8\u8d28\u91cf\u6570\u636e\u96c6\u3002<\/p>\n<p>\u6570\u636e\u96c6\u5305\u62ec1785\u4e2a\u591a\u8f6e\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\uff0c\u5176\u4e2d\u5171\u670911376\u4e2a\u4f8b\uff0c\u548c77\u4e2a\u6765\u81ea\u4e2d\u6587\u5c0f\u8bf4\u548c\u5267\u672c\u4e2d\u7684\u89d2\u8272\u3002<\/p>\n<p>\u4f7f\u7528GPT-4\u63d0\u53d6\u6700\u521d\u5bf9\u8bdd\uff0c\u7136\u540e\u8fdb\u884c\u4e25\u683c\u7684\u4eba\u5de5\u5ba1\u6838\uff0c\u5e76\u901a\u8fc7\u767e\u5ea6\u767e\u79d1\u63d0\u4f9b\u7684\u4eba\u7269\u8d44\u6599\u8fdb\u884c\u589e\u5f3a\u3002\u3002CharacterEval\u91c7\u7528\u591a\u65b9\u9762\u7684\u8bc4\u4f30\u65b9\u6cd5\uff0c\u5305\u62ec\u56db\u4e2a\u7ef4\u5ea6\u4e0a\u768413\u4e2a\u76ee\u6807\u6307\u6807\u3002\u4e3a\u4e86\u4fbf\u4e8e\u5728CharacterEval\u4e2d\u65b9\u4fbf\u5730\u8bc4\u4f30\u8fd9\u4e9b\u4e3b\u89c2\u6307\u6807\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u5f00\u53d1\u4e86\u57fa\u4e8e\u4eba\u7c7b\u6ce8\u91ca\u7684\u89d2\u8272\u626e\u6f14\u5956\u52b1\u6a21\u578bCharacterRM\uff0c\u4e0eGPT-4\u76f8\u6bd4\uff0c\u8be5\u6a21\u578b\u4e0e\u4eba\u7c7b\u5224\u65ad\u7684\u76f8\u5173\u6027\u66f4\u9ad8\u3002\u6027\u683c\u8bc4\u4ef7\u7684\u7efc\u5408\u5b9e\u9a8c\u8868\u660e\uff0c\u6c49\u8bedllm\u5728\u6c49\u8bed\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4e2d\u8868\u73b0\u51fa\u6bd4GPT-4\u66f4\u6709\u524d\u666f\u7684\u80fd\u529b\u3002\u6e90\u4ee3\u7801\uff0c\u6570\u636e\u6e90\u548c\u5956\u52b1\u6a21\u578b\u88ab\u516c\u5f00\u5728\uff1a<a href=\"https:\/\/github.com\/morecry\/CharacterEval\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/morecry\/CharacterEval<\/a><\/p>\n<h1>\u5f15\u8a00<\/h1>\n<p>\u76ee\u524d\u5df2\u6709\u7684RPCA benchmark\uff0c\u4e3b\u8981\u7684\u95ee\u9898\u662f\u6570\u636e\u96c6\u7684\u6784\u5efa\u3002\u867d\u7136\u5df2\u6709\u90e8\u5206\u6570\u636e\u96c6\uff0c\u4f46\u662f\u8fd9\u4e9b\u6570\u636e\u96c6\u7684\u8d28\u91cf\u4ee4\u4eba\u62c5\u5fe7\uff0c\u5b83\u4eec\u8981\u4e0d\u7136\u662f\u7531llm\u4ea7\u751f\u7684\uff0c\u8981\u4e48\u662f\u7531\u4e8e\u63d0\u53d6\u65b9\u6cd5\u800c\u53d7\u5230\u663e\u8457\u7684\u566a\u58f0\u5f71\u54cd\u3002\u8fd9\u4e9b\u9650\u5236\u4f7f\u5f97\u8bc4\u4f30\u7ed3\u679c\u5bf9\u4e8eRPCA\u7684\u5b9e\u9645\u80fd\u529b\u662f\u4e0d\u53ef\u9760\u7684\u3002\u56e0\u6b64\uff0c\u4f5c\u8005\u6784\u5efa\u4e86\u4e00\u4e2a\u4e2d\u6587\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u6570\u636e\u96c6\uff0c\u5305\u542b1785\u4e2a\u591a\u8f6e\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\uff0c11376\u4e2a\u4f1a\u8bdd\u793a\u4f8b\u548c77\u4e2a\u4e3b\u8981\u89d2\u8272\uff0c\u8fd9\u4e9b\u6570\u636e\u6765\u81ea\u4e0d\u540c\u7684\u4e2d\u6587\u5c0f\u8bf4\u548c\u5267\u672c\u3002\u7136\u540e\uff0c\u4f7f\u7528GPT-4\u63d0\u53d6\u8fd9\u4e9b\u6765\u6e90\u7684\u5bf9\u8bdd\u573a\u666f\u3001\u8bdd\u8bed\u548c\u4e3b\u89d2\u7684\u884c\u4e3a\u3002\u5728\u57fa\u672c\u7684\u9884\u5904\u7406\u548c\u5220\u9664\u8f83\u5c11\u56de\u5408\u7684\u5bf9\u8bdd\u4e4b\u540e\uff0c\u6211\u4eec\u9080\u8bf7\u6ce8\u91ca\u8005\u6765\u8bc4\u4f30\u5bf9\u8bdd\u7684\u8d28\u91cf\u3002\u4ed6\u4eec\u7684\u4efb\u52a1\u662f\u786e\u5b9a\u548c\u4fdd\u7559\u9ad8\u8d28\u91cf\u7684\u5bf9\u8bdd\uff0c\u540c\u65f6\u4e22\u5f03\u4f4e\u8d28\u91cf\u7684\u5bf9\u8bdd\u3002\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u4ece\u767e\u5ea6\u767e\u79d1\u4e2d\u6293\u53d6\u4e86\u8be6\u7ec6\u7684\u4eba\u7269\u6863\u6848\uff0c\u7ec4\u6210\u4e86\u4e00\u4e2a\u7528\u4e8eRPCA\u8bc4\u4f30\u7684\u7efc\u5408\u6570\u636e\u96c6\u3002\u6570\u636e\u96c6\u4e2d\u7684\u793a\u4f8b\u5982\u56fe1\u6240\u793a\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg1.jpg\" alt=\"\" \/><br \/>\n\u53e6\u5916\uff0c\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u662f\u4e00\u9879\u590d\u6742\u7684\u4efb\u52a1\uff0c\u4e0d\u4ec5\u9700\u8981\u6a21\u4eff\u89d2\u8272\u7684\u884c\u4e3a\u548c\u8bdd\u8bed\uff0c\u8fd8\u9700\u8981\u4fdd\u6301\u89d2\u8272\u7684\u77e5\u8bc6\uff0c\u4ee5\u53ca\u51fa\u8272\u7684\u591a\u8f6e\u5bf9\u8bdd\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u591a\u65b9\u9762\u7684\u8bc4\u4f30\u65b9\u6cd5\uff0c\u5305\u62ec\u56db\u4e2a\u7ef4\u5ea6\u768413\u4e2a\u5177\u4f53\u6307\u6807\u3002\u8bc4\u4f30\u65b9\u6cd5\u8003\u8651\u4e86\u5bf9\u8bdd\u80fd\u529b\u3001\u6027\u683c\u4e00\u81f4\u6027\u3001\u89d2\u8272\u626e\u6f14\u5438\u5f15\u529b\uff0c\u5e76\u5229\u7528\u4eba\u683c\u56de\u6eaf\u65b9\u6cd5\u6765\u8bc4\u4f30\u89d2\u8272\u626e\u6f14\u7684\u89d2\u8272\u6027\u683c\u51c6\u786e\u6027\u3002<br \/>\n\u4e3a\u4e86\u8bc4\u4f30\u5bf9\u8bdd\u80fd\u529b\uff0c\u4ece\u53e5\u5b50\u548c\u4f1a\u8bdd\u4e24\u4e2a\u5c42\u9762\u8bc4\u4ef7\u4f1a\u8bdd\u7684\u6d41\u7545\u6027\u3001\u8fde\u8d2f\u6027\u548c\u4e00\u81f4\u6027\u3002\u5728\u89d2\u8272\u626e\u6f14\u7684\u5bf9\u8bdd\u4e2d\uff0c\u89d2\u8272\u4e00\u81f4\u6027\u662f\u6700\u91cd\u8981\u7684\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u8bc4\u4f30\u77e5\u8bc6\u548c\u89d2\u8272\u7684\u4e00\u81f4\u6027\u6765\u8861\u91cfRPCA\u5982\u4f55\u6a21\u62df\u89d2\u8272\u3002\u8fd9\u5177\u4f53\u5305\u62ec\u77e5\u8bc6\u7684\u9732\u51fa\uff08knowledge exposure\uff0cKE\uff09\u3001\u77e5\u8bc6\u7684\u51c6\u786e\u6027\uff08knowledge accuracy\uff09\u548c\u77e5\u8bc6\u5e7b\u89c9\uff08knowledge hallucination\uff09\uff0c\u4ee5\u53ca\u8bc4\u4f30\u884c\u4e3a\u4e00\u81f4\u6027\uff08behavior consistency\uff09\u548c\u8868\u8fbe\u4e00\u81f4\u6027\uff08utterance consistency\uff09\u3002<br \/>\n\u8003\u8651\u5230\u89d2\u8272\u626e\u6f14\u7684\u5a31\u4e50\u6027\uff0c\u89d2\u8272\u626e\u6f14\u7684\u5438\u5f15\u529b\u4e5f\u662f\u4e00\u4e2a\u91cd\u8981\u56e0\u7d20\u3002\u6211\u4eec\u901a\u8fc7\u4eba\u7c7b\u7684\u76f8\u4f3c\u6027\uff08human-likeness\uff09\u3001\u6c9f\u901a\u6280\u5de7\uff08human-likeness\uff09\u3001\u8868\u8fbe\u591a\u6837\u6027\uff08human-likeness\uff09\u548c\u540c\u7406\u5fc3\uff08empathy\uff09\u6765\u8bc4\u4f30\u8fd9\u4e00\u70b9\u3002\u6700\u540e\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4eba\u683c\u53cd\u6d4b\u3002\u4ee5\u6536\u96c6\u7684Myers- briggs\u7c7b\u578b\u6307\u6807(MBTI)\u4eba\u683c\u7c7b\u578b\u4e3a\u53c2\u8003\uff0c\u6211\u4eec\u8ba9RPCAs\u8fdb\u884cMBTI\u8bc4\u4f30\u5e76\u8ba1\u7b97MBTI\u51c6\u786e\u6027(\u4eba\u683c\u56de\u6d4b)\u3002<br \/>\n\u4e3a\u4e86\u65b9\u4fbf\u91cd\u65b0\u5b9e\u73b0\uff0c\u6211\u4eec\u9080\u8bf7\u4e8612\u4e2a\u6ce8\u91ca\u8005\u4e3a\u6211\u4eec\u7684\u8bc4\u4f30\u7cfb\u7edf\u4e2d\u7684\u4e3b\u89c2\u6307\u6807\u7684\u4e0d\u540c\u6a21\u578b\u751f\u6210\u7684\u54cd\u5e94\u6253\u5206\u3002\u57fa\u4e8e\u4eba\u7c7b\u7684\u5224\u65ad\uff0c\u6211\u4eec\u5f00\u53d1\u4e86\u4e00\u4e2a\u89d2\u8272\u626e\u6f14\u5956\u52b1\u6a21\u578b- CharacterRM\uff0c\u5176\u4e0e\u4eba\u7c7b\u7684\u76f8\u5173\u6027\u53ef\u4ee5\u8d85\u8fc7\u6700\u5148\u8fdb\u7684LLM GPT-4\u3002\u5728CharacterEval\u65b9\u9762\uff0c\u6211\u4eec\u5bf9\u73b0\u6709\u7684llm\u8fdb\u884c\u4e86\u5168\u9762\u7684\u8bc4\u4f30\uff0c\u5305\u62ec\u5f00\u6e90\u548c\u95ed\u6e90\u6a21\u578b\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u6211\u4eec\u63d0\u51fa\u7684\u4e2d\u6587LLM\u524d\u666f\u5e7f\u9614\uff0c\u800cgpt\u7cfb\u5217\u6a21\u578b\u5728\u4e2d\u6587\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4e2d\u5e76\u4e0d\u5360\u4e3b\u5bfc\u5730\u4f4d\u3002<\/p>\n<h1>\u95ee\u9898\u63cf\u8ff0<\/h1>\n<p>\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4ee3\u7406(RPCA)\u7684\u8bbe\u8ba1\u76ee\u7684\u662f\u901a\u8fc7\u6a21\u62df\u7279\u5b9a\u7684\u89d2\u8272\u4e0e\u7528\u6237\u8fdb\u884c\u5bf9\u8bdd\u3002\u8fd9\u4e9b\u89d2\u8272\u662f\u7531\u4ed6\u4eec\u7684\u77e5\u8bc6\u3001\u884c\u4e3a\u548c\u56de\u5e94\u98ce\u683c\u6765\u5b9a\u4e49\u7684\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0cRPCA\u9700\u8981\u4f7f\u7528\u4e00\u4e2a\u89d2\u8272\u6a21\u677f<span class=\"katex-eq\" data-katex-display=\"false\">P<\/span>\u3001\u4e0a\u4e0b\u6587\u5bf9\u8bdd<span class=\"katex-eq\" data-katex-display=\"false\">C_n=[q_1,r_1,q_2,r_2,...,q_n]<\/span>\u3002<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">r_i<\/span>\u5206\u522b\u5bf9\u5e94\u5bf9\u8bdd\u4e2d\u7684\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u4e2a\u95ee\u9898\u548c\u56de\u7b54\u3002RPCA\u7684\u76ee\u6807\u662f\u751f\u6210\u4e0e\u89d2\u8272\u9884\u8bbe\u76f8\u540c\u7684\u56de\u7b54<span class=\"katex-eq\" data-katex-display=\"false\">r_n<\/span>,\u516c\u5f0f\u4e3a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_eq1.jpg\" alt=\"\" \/><br \/>\n\u5176\u4e2d\u56de\u7b54<span class=\"katex-eq\" data-katex-display=\"false\">r_n<\/span>\u7531\u4e24\u4e2a\u5143\u7d20\u7ec4\u6210\uff1a\u884c\u4e3a\u548c\u8bdd\u8bed\u3002\u884c\u4e3a\u5305\u542b\u5728\u62ec\u53f7\u4e2d\uff0c\u63d0\u4f9b\u4e86\u5173\u4e8e\u89d2\u8272\u52a8\u4f5c\u3001\u8868\u60c5\u548c\u8bed\u6c14\u7684\u8be6\u7ec6\u63cf\u8ff0\u3002\u8fd9\u79cd\u5206\u79bb\u5141\u8bb8\u5bf9RPCA\u7684\u80fd\u529b\u8fdb\u884c\u7ec6\u7c92\u5ea6\u7684\u8bc4\u4f30\uff0c\u4e0d\u4ec5\u53ef\u4ee5\u751f\u6210\u9002\u5f53\u7684\u8bdd\u8bed\uff0c\u8fd8\u53ef\u4ee5\u751f\u6210\u72ec\u7279\u7684\u884c\u4e3a\u7279\u5f81\u3002<\/p>\n<h1>\u6570\u636e\u6536\u96c6<\/h1>\n<h2>\u6570\u636e\u6536\u96c6\u539f\u5219<\/h2>\n<ul>\n<li><strong>\u5fe0\u5b9e\u4e8e\u539f\u8457(Fidelity to Source Material)\uff1a<\/strong>\u6240\u6709\u5bf9\u8bdd\u90fd\u8981\u4e0e\u539f\u8457\u4fdd\u6301\u4e00\u81f4\uff0c\u786e\u4fdd\u4efb\u52a1\u771f\u5b9e\u6027\uff1b<\/li>\n<li><strong>\u5206\u5e03\u591a\u6837\u6027(Diversity in Distribution)\uff1a<\/strong>\u6570\u636e\u96c6\u5fc5\u987b\u5305\u542b\u5e7f\u6cdb\u7684\u573a\u666f\uff0c\u7528\u4e8e\u5f7b\u5e95\u8bc4\u4f30\u89d2\u8272\u626e\u6f14\u80fd\u529b\uff1b<\/li>\n<li><strong>\u591a\u8f6e\u5bf9\u8bdd\u7279\u5f81(Multi-Turn Feature)\uff1a<\/strong>\u6570\u636e\u96c6\u5e94\u8be5\u4e3b\u8981\u7531\u591a\u56de\u5408\u5bf9\u8bdd\u7ec4\u6210\uff0c\u800c\u4e0d\u662f\u5c40\u9650\u4e8e\u5355\u56de\u5408\u5bf9\u8bdd\uff1b<\/li>\n<li><strong>\u4eba\u7c7b\u5728\u73af(Human-in-the-Loop\uff09\uff1a<\/strong>\u4e3a\u4e86\u4fdd\u8bc1\u8d28\u91cf\u3001\u4eba\u7c7b\u53c2\u4e0e\u662f\u5fc5\u8981\u7684\uff0c\u56e0\u4e3a\u4ec5\u4ec5\u4f9d\u8d56LLMs\u662f\u4e0d\u5145\u5206\u7684\u3002<\/li>\n<\/ul>\n<h2>\u6570\u636e\u6536\u96c6\u6d41\u7a0b<\/h2>\n<ul>\n<li><strong>\u60c5\u8282\u5212\u5206\uff08Plot Division\uff09\uff1a<\/strong>\u5c0f\u8bf4\u548c\u5267\u672c\u7b49\u53d9\u4e8b\u6587\u672c\u7684\u60c5\u8282\u975e\u5e38\u590d\u6742\uff0c\u5f88\u96be\u5c06\u6587\u672c\u5212\u5206\u4e3a\u6709\u610f\u4e49\u7684\u5757\uff0c\u4f7f\u7528\u53e5\u5b50\u6807\u8bb0\u5de5\u5177\u800c\u4e0d\u8003\u8651\u8bed\u4e49\u53ef\u80fd\u5bfc\u81f4\u5bf9\u8bdd\u4e2d\u65ad\u3002\u56e0\u6b64\uff0c\u4f7f\u7528GPT-4\u8bc6\u522b\u60c5\u8282\u8f6c\u6298\uff0c\u5373\u8868\u793a\u8fde\u7eed\u60c5\u8282\u7ed3\u675f\u7684\u53e5\u5b50\u3002\u7136\u540e\u662f\u7531\u8fd9\u4e9b\u60c5\u8282\u8f6c\u6298\u5c06\u6587\u672c\u5206\u5272\u6210\u5757\uff0c\u6bcf\u4e00\u5757\u4e2d\u5305\u542b\u4e00\u4e2a\u5b8c\u6574\u60c5\u8282\u3002<\/li>\n<li><strong>\u5bf9\u8bdd\u62bd\u53d6\uff08Dialogue Extraction\uff09\uff1a<\/strong>\u4f7f\u7528GPT4\u4ece\u60c5\u8282\u5757\u4e2d\u62bd\u53d6\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u3002\u4e3aGPT\u8bbe\u8ba1\u63d0\u793a\u8bcd\u6765\u8fdb\u884c\u4fe1\u606f\u63d0\u53d6\uff0c\u4ece\u800c\u4fdd\u7559\u4eba\u7269\u7684\u8bdd\u8bed\uff08utterances\uff09\u3001\u884c\u4e3a\uff08behaviors\uff09\u548c\u60c5\u8282\u573a\u666f\uff08 scenes\uff09\u3002<\/li>\n<li><strong>\u8d28\u91cf\u8fc7\u6ee4\uff08Quality Filtering\uff09\uff1a<\/strong>\u5c0f\u8bf4\u548c\u5267\u672c\u4e2d\u7684\u5bf9\u8bdd\u901a\u5e38\u6d89\u53ca\u4e24\u4e2a\u4ee5\u4e0a\u7684\u4eba\u7269\u3002\u5982\u679c\u7b80\u5355\u7684\u4fdd\u7559\u4e24\u4e2a\u89d2\u8272\u7684\u5bf9\u8bdd\u800c\u5ffd\u7565\u5176\u4ed6\u89d2\u8272\u4f1a\u626d\u66f2\u5bf9\u8bdd\u7ed3\u6784\u3002\u56e0\u6b64\uff0c\u6309\u7167ABAB\u6a21\u578b\u4fdd\u7559\u5bf9\u8bdd\uff0c\u76f4\u5230\u51fa\u73b0\u7b2c\u4e09\u4e2a\u4eba\u7269\u7684\u52a0\u5165\u3002\u6b64\u5916\uff0c\u53ea\u4fdd\u7559\u8d85\u8fc7\u4e94\u8f6e\uff086\u6761\uff09\u5bf9\u8bdd\uff0c\u8fc7\u6ee4\u77ed\u5bf9\u8bdd\u3002<\/li>\n<li><strong>\u4eba\u7c7b\u6ce8\u91ca\uff08Human Annotation\uff09\uff1a<\/strong>\u5c3d\u7ba1llm\u5177\u6709\u62bd\u53d6\u4efb\u52a1\u7684\u57fa\u672c\u80fd\u529b\uff0c\u4f46\u662f\u56de\u590d\u968f\u673a\u6027\u4f1a\u5f71\u54cd\u6570\u636e\u8d28\u91cf\u3002\u56e0\u6b64\uff0c\u9700\u8981\u4eba\u5de5\u6807\u6ce8\u6765\u8bc4\u4f30\u5bf9\u8bdd\u7684\u4e00\u81f4\u6027\u548c\u8d28\u91cf\uff0c\u5220\u9664\u4efb\u4f55\u6709\u95ee\u9898\u7684\u5bf9\u8bdd\u3002<\/li>\n<\/ul>\n<h1>\u8bc4\u4ef7\u6307\u6807<\/h1>\n<p>\u5982\u56fe2\u6240\u793a\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u56db\u7ef4\u8bc4\u4ef7\u4f53\u7cfb\uff0c\u5305\u62ec\u4f1a\u8bdd\u80fd\u529b\uff08conversational ability\uff09\u3001\u89d2\u8272\u4e00\u81f4\u6027\uff08character consistency\uff09\u3001\u89d2\u8272\u626e\u6f14\u5438\u5f15\u529b\uff08 role-playing attractiveness\uff09\u548c\u4e2a\u6027\u56de\u6eaf\u6d4b\u8bd5\uff08personality back-testing\uff09\uff0c\u517113\u4e2a\u6307\u6807\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg2.jpg\" alt=\"\" \/><\/p>\n<h2>\u4f1a\u8bdd\u80fd\u529b<\/h2>\n<p>\u5728\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4e2d\uff0c\u57fa\u672c\u7684\u4f1a\u8bdd\u80fd\u529b\u662f\u9996\u5148\u8981\u8003\u8651\u7684\u3002\u6587\u7ae0\u4e13\u6ce8\u4e8e\u751f\u6210\u7b54\u6848\u7684\u4e09\u4e2a\u5173\u952e\u76ee\u6807:\u6d41\u7545\u6027\u3001\u8fde\u8d2f\u6027\u548c\u4e00\u81f4\u6027\u3002<\/p>\n<ul>\n<li>Fluency (Flu.)\uff1a\u8861\u91cf\u56de\u590d\u7684\u8bed\u6cd5\u6b63\u786e\u6027\uff0c\u8868\u660e\u56de\u590d\u662f\u5426\u53ef\u8bfb\u4e14\u6ca1\u6709\u660e\u663e\u7684\u8bed\u6cd5\u9519\u8bef\uff1b<\/li>\n<li>Coherency (Coh.) :\u8bc4\u4f30\u56de\u7b54\u548c\u4e0a\u4e0b\u6587\u4e4b\u95f4\u7684\u4e3b\u9898\u76f8\u5173\u6027\u3002\u5f53\u7528\u6237\u8be2\u95ee\u7279\u5b9a\u4e3b\u9898\u65f6\uff0cRPCA\u5e94\u8be5\u6839\u636e\u95ee\u9898\u8fdb\u884c\u56de\u7b54\uff0c\u800c\u4e0d\u662f\u63d0\u4f9b\u4e0d\u76f8\u5173\u7684\u7b54\u6848\uff1b<\/li>\n<li>Consistency (Cons.) \uff1a\u5728\u4f1a\u8bdd\u671f\u95f4\u8bc4\u4f30RPCA\u7684\u7a33\u5b9a\u6027\u3002RPCA\u7684\u56de\u7b54\u4e0d\u5e94\u8be5\u4e0e\u4e4b\u524d\u56de\u5408\u7684\u56de\u7b54\u76f8\u77db\u76fe\u3002<\/li>\n<\/ul>\n<h2>\u89d2\u8272\u4e00\u81f4\u6027<\/h2>\n<p>\u89d2\u8272\u4e00\u81f4\u6027\u662f\u8bc4\u4ef7\u89d2\u8272\u626e\u6f14\u89d2\u8272\u7684\u91cd\u8981\u6307\u6807\u3002\u5f53\u89d2\u8272\u89d2\u8272\u4e00\u81f4\u6027\u53d8\u5316\u65f6\uff0c\u4f1a\u7ed9\u7528\u6237\u5e26\u6765\u6700\u76f4\u89c2\u7684\u4f53\u9a8c\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u4ece\u77e5\u8bc6\u4e00\u81f4\u6027\u548c\u4eba\u7269\u4e00\u81f4\u6027\u4e24\u4e2a\u5c42\u9762\u6765\u8bc4\u4f30\u89d2\u8272\u4e00\u81f4\u6027\u3002\u524d\u8005\u8bc4\u4f30RPCA\u662f\u5426\u80fd\u591f\u6839\u636e\u89d2\u8272\u7684\u77e5\u8bc6\u505a\u51fa\u53cd\u5e94\uff0c\u5305\u62ec\u77e5\u8bc6\u66b4\u9732\u3001\u51c6\u786e\u6027\u548c\u5e7b\u89c9\u6307\u6807\u3002\u540e\u8005\u8bc4\u4f30RPCA\u7684\u53cd\u6620\u662f\u5426\u7b26\u5408\u89d2\u8272\u7684\u4e2a\u6027\uff0c\u5305\u62ec\u884c\u4e3a\u548c\u8bdd\u8bed\u6307\u6807\u3002<\/p>\n<ul>\n<li><strong>Knowledge-Exposure (KE)\uff1a<\/strong>\u4e3a\u4e86\u8bc4\u4f30\u56de\u7b54\u7684\u4fe1\u606f\u91cf\uff0cRPCA\u5728\u56de\u7b54\u4e2d\u53cd\u6620\u77e5\u8bc6\u662f\u81f3\u5173\u91cd\u8981\u7684\uff0c\u8fd9\u53ea\u5403\u4e86\u5176\u5bf9\u77e5\u8bc6\u8868\u8fbe\u80fd\u529b\u7684\u8bc4\u4f30\u3002<\/li>\n<li><strong>Knowledge-Accuracy (KA)\uff1a<\/strong>\u4e00\u65e6RPCA\u5c55\u793a\u4e86\u4f7f\u7528\u7279\u5b9a\u77e5\u8bc6\u751f\u6210\u56de\u7b54\u7684\u80fd\u529b\uff0c\u8bc4\u4f30\u8fd9\u4e9b\u56de\u7b54\u662f\u5426\u4e0e\u89d2\u8272\u4e00\u81f4\u3002RPCA\u7684\u76ee\u6807\u662f\u6839\u636e\u6765\u81ea\u89d2\u8272\u7684\u6982\u8981\u77e5\u8bc6\u751f\u6210\u51c6\u786e\u7684\u56de\u7b54\u3002<\/li>\n<li><strong>Knowledge-Hallucination (KH)\uff1a<\/strong>\u5c06\u77e5\u8bc6\u5e7b\u89c9\u7eb3\u5165\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u7684\u8bc4\u4f30\u3002\u4e3a\u4e86\u589e\u5f3a\u7528\u6237\u4f53\u9a8c\uff0cRPCA\u5e94\u8be5\u4fdd\u6301\u4e0e\u89d2\u8272\u8eab\u4efd\u7684\u4e00\u81f4\u6027\uff0c\u5e76\u907f\u514d\u56de\u7b54\u672a\u77e5\u77e5\u8bc6\u7684\u95ee\u9898\u3002<\/li>\n<li><strong>Persona-Behavior (PB)\uff1a<\/strong>\u89d2\u8272\u7684\u884c\u4e3a(\u901a\u5e38\u5728\u62ec\u53f7\u5185\u63cf\u8ff0)\u901a\u8fc7\u63cf\u7ed8\u7ec6\u7c92\u5ea6\u7684\u52a8\u4f5c\u3001\u8868\u60c5\u548c\u8bed\u6c14\u6765\u6539\u5584\u7528\u6237\u7684\u611f\u53d7\u3002\u4e00\u81f4\u7684\u884c\u4e3a\u662f\u6709\u6548RPCA\u7684\u6807\u5fd7<\/li>\n<li><strong>Persona-Utterance (PU)\uff1a<\/strong>\u9664\u4e86\u884c\u4e3a\u4e3e\u6b62\uff0c\u89d2\u8272\u7684\u8bf4\u8bdd\u98ce\u683c\u4e5f\u5f88\u91cd\u8981\u3002\u6bcf\u4e2a\u89d2\u8272\u90fd\u6709\u72ec\u7279\u7684\u8868\u8fbe\u4e60\u60ef\u3002\u56e0\u6b64\uff0cRPCA\u7684\u8bdd\u8bed\u5e94\u8be5\u4e0e\u8fd9\u4e9b\u4e60\u60ef\u4fdd\u6301\u4e00\u81f4\uff0c\u4ee5\u719f\u7ec3\u5730\u6a21\u4eff\u89d2\u8272\u3002<\/li>\n<\/ul>\n<h2>\u89d2\u8272\u626e\u6f14\u5438\u5f15\u529b<\/h2>\n<p>RPCA\u662f\u9700\u8981\u5bf9\u7528\u6237\u7684\u60c5\u7eea\u654f\u611f\u3002\u56e0\u6b64\uff0c\u5f15\u5165\u89d2\u8272\u5438\u5f15\u529b\u7ef4\u5ea6\u6765\u8bc4\u4f30RPCA\u5728\u5bf9\u8bdd\u4e2d\u7684\u5438\u5f15\u529b\u3002\u4ece\u7528\u6237\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u8003\u8651\u56db\u4e2a\u5173\u952e\u7ef4\u5ea6\uff1a\u9c7c\u4eba\u76f8\u4f3c\u3001\u6c9f\u901a\u6280\u5de7\u3001\u8868\u8fbe\u591a\u6837\u6027\u548c\u540c\u7406\u5fc3\u3002<\/p>\n<ul>\n<li>Human-Likeness (HL)\uff1a\u5927\u591a\u6570\u5927\u6a21\u578b\u662f\u4e3a\u4e86\u4fe1\u606f\u641c\u7d22\u800c\u8bbe\u8ba1\u7684\uff0c\u5f80\u5f80\u7070\u63d0\u4f9b\u673a\u68b0\u7684\u3001\u6ca1\u6709\u611f\u60c5\u7684\u7b54\u6848\u3002\u7136\u800c\uff0c\u5728\u89d2\u8272\u626e\u6f14\u4efb\u52a1\u4e2d\uff0cRPCA\u9700\u8981\u5c55\u73b0\u51fa\u9632\u62a4\u4e00\u4e2a\u66f4\u50cf\u4eba\u7c7b\u7684\u89d2\u8272\u5f62\u8c61\u3002<\/li>\n<li>Communication Skills (CS)\uff1a\u5728\u4eba\u7c7b\u793e\u4f1a\u4e2d\uff0c\u5de7\u5999\u6c9f\u901a\u7684\u80fd\u529b\uff0c\u901a\u5e38\u88ab\u79f0\u4e3a\u60c5\u5546(EQ)\uff0c\u5bf9\u4e00\u4e2a\u4eba\u7684\u53d7\u6b22\u8fce\u7a0b\u5ea6\u6709\u91cd\u8981\u5f71\u54cd\u3002\u56e0\u6b64\uff0c\u7528\u6237\u66f4\u6709\u53ef\u80fd\u4e0e\u60c5\u5546\u8f83\u9ad8\u7684RPCA\u4e92\u52a8\uff0c\u8fd9\u53cd\u6620\u4e86\u5728\u65e5\u5e38\u751f\u6d3b\u4e2d\u5177\u6709\u5f3a\u5927\u6c9f\u901a\u6280\u5de7\u7684\u4e2a\u4eba\u7684\u53d7\u6b22\u8fce\u7a0b\u5ea6\u3002<\/li>\n<li>Expression Diversity (ED)\uff1a\u300aCharacterEval\u300b\u4e2d\u7684\u5bf9\u8bdd\u6765\u6e90\u4e8e\u73b0\u6709\u7684\u5c0f\u8bf4\u3001\u5267\u672c\u548c\u5404\u79cd\u6587\u5b66\u4f5c\u54c1\uff0c\u4eba\u7269\u5728\u884c\u4e3a\u548c\u8bdd\u8bed\u4e0a\u90fd\u5177\u6709\u4e30\u5bcc\u591a\u6837\u7684\u8868\u8fbe\u80fd\u529b\u3002\u56e0\u6b64\uff0cRPCA\u5e94\u8be5\u52aa\u529b\u5728\u5bf9\u8bdd\u4e2d\u8868\u8fbe\u8fd9\u79cd\u591a\u6837\u6027\uff0c\u4e3a\u7528\u6237\u63d0\u4f9b\u66f4\u5177\u6c89\u6d78\u611f\u7684\u4f53\u9a8c\u3002<\/li>\n<li>Empathy (Emp.)\uff1a\u867d\u7136RPCA\u7684\u4e3b\u8981\u89d2\u8272\u4e0d\u662f\u60c5\u611f\u54a8\u8be2\u5e08\uff0c\u4f46\u8868\u8fbe\u540c\u7406\u5fc3\u53ef\u4ee5\u663e\u8457\u5f71\u54cd\u5bf9\u7528\u6237\u7684\u597d\u611f\u5ea6\u3001\u5728\u89d2\u8272\u626e\u6f14\u4efb\u52a1\u4e2d\u8bc4\u4f30\u540c\u7406\u5fc3\u53ef\u4ee5\u8ba9RPCA\u6210\u4e3a\u4e00\u4e2a\u66f4\u70ed\u60c5\u3001\u66f4\u53cb\u597d\u7684\u5bf9\u8bdd\u4f19\u4f34\u3002<\/li>\n<\/ul>\n<h2>\u4e2a\u6027\u56de\u6eaf\u6d4b\u8bd5<\/h2>\n<p>\u901a\u8fc7\u4eba\u683c\u56de\u6eaf\u6d4b\u8bd5\u8bc4\u4f30RPCA\u5728\u4eba\u683c\u7ef4\u5ea6\u80cc\u666f\u4e0b\u7684\u89d2\u8272\u626e\u6f14\u80fd\u529b\uff0c\u91c7\u7528MBTI\u8fdb\u884c\u4eba\u683c\u5206\u7c7b\u3002\u4e3a\u4e86\u83b7\u5f97\u5fc5\u8981\u7684\u6807\u7b7e\uff0c\u4ece\u7f51\u7ad9\u4e2d\u6536\u96c6\u4eba\u7269\u89d2\u8272\u7684MBTI\u4fe1\u606f\u3002<br \/>\n\u4eba\u7269\u89d2\u8272MBTI\u6570\u636e\uff1a<a href=\"https:\/\/www.personality-database.com\/\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/www.personality-database.com\/<\/a><br \/>\nMBTI\u6027\u683c\u6d4b\u8bd5\u7f51\u7ad9\uff1a<a href=\"https:\/\/www.16personalities.com\/\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/www.16personalities.com\/<\/a><\/p>\n<h1>\u6848\u4f8b\u5206\u6790<\/h1>\n<h2>\u6570\u636e\u63cf\u8ff0<\/h2>\n<p>\u5c06\u6570\u636e\u96c6\u6309\u7167\u793a\u4f8b\u62c6\u5206\u4e3a\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u96c6\uff0c\u793a\u4f8b\u6709\u5143\u7ec4\u7ec4\u6210\uff08\u4eba\u7269\u3001\u4e0a\u4e0b\u6587\u3001\u56de\u7b54\uff09\uff0c\u6570\u636e\u96c6\u7684\u7edf\u8ba1\u4fe1\u606f\u5982\u4e0b\u6240\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg3.jpg\" alt=\"\" \/><\/p>\n<h2>\u5b9e\u9a8c\u8bbe\u7f6e<\/h2>\n<p>CharacterEval\u91c7\u7528\u4e00\u5957\u5168\u9762\u7684\u7ec6\u7c92\u5ea6\u4e3b\u89c2\u6307\u6807(\u4f1a\u8bdd\u80fd\u529b\u3001\u89d2\u8272\u4e00\u81f4\u6027\u548c\u89d2\u8272\u626e\u6f14\u5438\u5f15\u529b\u7ef4\u5ea6\u4e2d\u768412\u4e2a\u6307\u6807)\u6765\u8bc4\u4f30\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u4ee3\u7406(RPCA)\u7684\u591a\u7ef4\u80fd\u529b\u3002\u7136\u800c\uff0c \u5355\u4e2a\u793a\u4f8b\u662f\u4e0d\u80fd\u591f\u5145\u5206\u4ee3\u8868RPCAs\u7684\u5168\u90e8\u7ef4\u5ea6\u7684\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4f7f\u7528\u4eba\u5de5\u5efa\u7acb\u7a00\u758f\u6027\u80fd\u8bc4\u4f30\u77e9\u9635\uff0c\u5373\u5229\u7528\u4eba\u5de5\u5bf9CharacterEval\u7684\u4e00\u4e2a\u5b50\u96c6\u8fdb\u884c\u4e3b\u89c2\u6307\u6807\u8bc4\u4ef7\uff0c\u4f7f\u5f97\u8bc4\u4ef7\u7ed3\u679c\u66f4\u6709\u5dee\u5f02\u6027\u3002<br \/>\n\u7136\u540e\uff0c\u57fa\u4e8e\u8fd9\u4e9b\u4e3a\u6bcf\u4e2a\u4f8b\u5b50\u9009\u62e9\u7684\u6307\u6807\uff0c\u62db\u52df12\u4e2a\u6ce8\u91ca\u8005\uff0c\u4ee5\u4e94\u5206\u5236\u5bf9\u4e0d\u540c\u6a21\u578b\u7684\u56de\u7b54\u8fdb\u884c\u8bc4\u5206\u3002\u4eba\u7c7b\u7684\u6807\u6ce8\u7ed3\u679c\u88ab\u7528\u6765\u8bad\u7ec3\u4e00\u4e2a\u4ee5Baichuan2-13B\u4e3a\u57fa\u6a21\u578b\u7684\u89d2\u8272\u626e\u6f14\u5956\u52b1\u6a21\u578b\uff08characterRM\uff09\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cCharacterRM\u6bd4GPT-4\u4e0e\u4eba\u7c7b\u5224\u65ad\u7ed3\u679c\u7684\u76f8\u5173\u6027\u66f4\u597d\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u867d\u7136GPT4\u7684\u6027\u80fd\u4f1a\u968f\u7740\u6f14\u793a\u6b21\u6570\u7684\u589e\u52a0\u800c\u63d0\u9ad8\uff0c\u4f46\u662f\u9ad8\u6602\u7684token\u4f7f\u5f97\u5176\u6210\u672c\u8bc4\u4f30\u96be\u4ee5\u8bc4\u4ef7\u3002\u56e0\u6b64\uff0c\u4f7f\u7528CharacterRM\u5bf9\u4e3b\u89c2\u6307\u6807\u8fdb\u884c\u540e\u7eed\u8bc4\u4f30\u3002\u5728MBTI\u6d4b\u8bd5\u4e2d\uff0c\u624b\u673a\u4e8654\u4e2a\u89d2\u8272\u7684mbti\uff0c\u4f7f\u7528\u8bad\u7ec3\u7684RPCA\u56de\u7b54MBTI\u95ee\u5377\uff0c\u7136\u540e\u8ba1\u7b97\u51c6\u786e\u6027\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg4.jpg\" alt=\"\" \/><\/p>\n<h2>\u8bc4\u4f30LLMs<\/h2>\n<p>\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\u8bc4\u4f30\u4e8610\u4e2a\u5177\u6709\u4e0d\u540c\u53c2\u6570\u7684\u5927\u8bed\u8a00\u6a21\u578bBaseline\uff0c\u5305\u62ec\u5f00\u6e90\u548c\u95ed\u6e90\u6a21\u578b\u3002\u5bf9\u4e8e\u5f00\u6e90\u6a21\u578b\uff0c\u8bc4\u4f30\u5b83\u7684\u804a\u5929\u7248\u672c\uff0c\u5bf9\u4e8e\u95ed\u6e90\u6a21\u578b\uff0c\u5229\u7528\u5b98\u65b9API\u8fdb\u884c\u5fc3\u6897\u8bc4\u4f30\u3002\u4f7f\u7528GPT-3.5-turbo-1106\u7248\u672c\u4f5c\u4e3aGPT-3.5\u3002\u5728\u88ab\u8bc4\u4f30\u7684\u6a21\u578b\u4e2d\uff0cCharacterGLM\u3001MiniMax\u3001Xingchen\u548cBC-NPC-Turbo\u662f\u4e3a\u89d2\u8272\u626e\u6f14\u91cf\u8eab\u5730\u5740\u7684\u5927\u6a21\u578b\uff0c\u5176\u4f59\u6a21\u578b\u662f\u4e3a\u4e00\u822c\u804a\u5929\u5e94\u7528\u800c\u8bbe\u8ba1\u7684\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cGPT-4\u548cGPT3.5\u662f\u552f\u4e00\u4e24\u4e2a\u5728\u4e3b\u8981\u7531\u82f1\u8bed\u8bed\u6599\u5e93\u7ec4\u6210\u7684\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u3002\u6211\u4eec\u59cb\u7ec8\u4e3a\u6bcf\u4e2a\u6a21\u578b\u4f7f\u7528\u76f8\u540c\u7684\u63d0\u793a\u7b26\uff0c\u4ec5\u5bf9\u95ed\u6e90\u6a21\u578b\u8fdb\u884c\u4e86\u5fae\u5c0f\u7684\u8c03\u6574\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg5.jpg\" alt=\"\" \/><\/p>\n<h2>\u6027\u80fd\u6982\u8ff0<\/h2>\n<p>\u5404\u4e2a\u6a21\u578b\u5728\u56db\u4e2a\u7ef4\u5ea6\u7684\u8bc4\u4ef7\u7ed3\u679c\u5982\u4e0b\u56fe\u6240\u793a\u3002BC-NPC-Turbo\u5728\u524d\u4e09\u4e2a\u7ef4\u5ea6\u8868\u73b0\u51fa\u8272\uff0c\u800cGPT-4\u5728MBTI\u6d4b\u8bd5\u4e2d\u8868\u73b0\u7a81\u51fa\u3002\u4e13\u95e8\u4e3a\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u8bbe\u8ba1\u7684\u5927\u6a21\u578b\uff0c\u5982Xingchen\u3001MinMax\u548cBC-NPC-Turbo\u7ecf\u8fc7\u9488\u5bf9\u6027\u8bad\u7ec3\uff0c\u663e\u793a\u51fa\u66f4\u597d\u7684\u7ed3\u679c\u3002\u5728\u5f00\u6e90\u6a21\u578b\u4e2d\uff0cInternLM-20B\u548cBaichuan2-13B\u7684\u6f5c\u529b\u8f83\u9ad8\u3002<br \/>\n\u5c3d\u7ba1\u7f3a\u4e4f\u9488\u5bf9\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u7684\u4e13\u95e8\u5b9a\u5236\uff0c\u5927\u90e8\u5206\u6a21\u578b\u5728\u8bc4\u4f30\u7ef4\u5ea6\u4e0a\u90fd\u5448\u73b0\u51fa\u503c\u5f97\u79f0\u9053\u7684\u7ed3\u679c\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cGPT-4\u5728\u4e2d\u6587\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u7684\u6709\u6548\u6027\u6709\u6240\u4e0b\u964d\u3002\u5176\u5728\u82f1\u8bed\u8bed\u6599\u5e93\u4e2d\u7684\u4e3b\u8981\u8bad\u7ec3\u9650\u5236\u4e86\u5176\u5728\u590d\u6742\u89d2\u8272\u626e\u6f14\u573a\u666f\u4e2d\u7684\u9002\u5e94\u6027\u548c\u5bf9\u4e2d\u56fd\u6587\u5316\u7684\u6df1\u523b\u7406\u89e3\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg6.jpg\" alt=\"\" \/><br \/>\nGPT-3.5\u5728CharacterEval\u7684\u8868\u73b0\u6700\u5dee\u3002\u4ed6\u503e\u5411\u4e8e\u4ea7\u751f\u8fc7\u4e8e\u5b89\u5168\u7684\u56de\u7b54\uff0c\u6bd4\u5982\uff1a\u201c\u6211\u53ea\u662f\u4e00\u4e2a\u4eba\u5de5\u667a\u80fd\u52a9\u624b\uff0c\u4e0d\u80fd\u8fdb\u884c\u89d2\u8272\u626e\u6f14\u201d\uff0c\u8fd9\u51f8\u663e\u4e86\u5b83\u5728\u89d2\u8272\u626e\u6f14\u5e94\u7528\u7a0b\u5e8f\u4e2d\u7684\u5c40\u9650\u6027\u3002\u8fd9\u4e2a\u95ee\u9898\u6e90\u4e8eRLHF\u7684\u8fc7\u5ea6\u5bf9\u9f50\uff0c\u4f7f\u5176\u4e0d\u9002\u5408\u52a8\u6001\u89d2\u8272\u626e\u6f14\u4ea4\u4e92\u3002<\/p>\n<h2>\u7ed3\u679c\u660e\u7ec6<\/h2>\n<p>\u4e0b\u8868\u7ed9\u51fa\u4e86\u8de813\u4e2a\u6307\u6807\u7684\u8be6\u7ec6\u6027\u80fd\u3002\u5728\u4f1a\u8bdd\u80fd\u529b\u65b9\u9762\uff0cBC-NPC-Turbo\u8868\u73b0\u51fa\u4e86\u4f18\u5f02\u7684\u8868\u73b0\uff0c\u4f1a\u8bdd\u7684\u4e00\u81f4\u6027\u4ee5\u53ca\u76f8\u5bf9\u7684\u6d41\u7545\u6027\u548c\u8fde\u8d2f\u6027\u90fd\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9\u3002\u5728\u5bf9\u6bd4\u5f00\u6e90\u548c\u95ed\u6e90\u6a21\u578b\u65f6\uff0c\u5f88\u96be\u5728\u8fd9\u4e2a\u7ef4\u5ea6\u4e2d\u5ba3\u5e03\u6700\u7ec8\u7684\u8d62\u5bb6\u3002<br \/>\n\u5728\u8fdb\u884c\u540c\u7c7b\u6a21\u578b\u7684\u6bd4\u8f83\u65f6\uff0c\u53c2\u6570\u6570\u91cf\u589e\u52a0\u53ef\u4ee5\u589e\u5f3a\u4f1a\u8bdd\u80fd\u529b\u3002\u5728\u5c0f\u4e8e10B\u7684\u6a21\u578b\u4e2d\uff0cBaichuan2-7B\u548cInternLM-7B\u8868\u73b0\u76f8\u5f53\u3002\u6b64\u5916\uff0c\u5047\u8bbe\u590d\u6742\u7684\u89d2\u8272\u626e\u6f14\u5bf9\u8bdd\u548c\u4e2d\u6587\u573a\u666f\u53ef\u80fd\u4f1a\u6311\u6218GPT\u7cfb\u5217\uff0c\u6f5c\u5728\u5730\u5bfc\u81f4\u4ed6\u4eec\u7684\u8868\u73b0\u4e0b\u964d\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg7.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/charactereval_fg8.jpg\" alt=\"\" \/><\/p>\n<h2>\u7ed3\u8bba<\/h2>\n<p>\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u81f4\u529b\u4e8e\u5efa\u7acb\u4e00\u4e2a\u5168\u9762\u7684\u57fa\u51c6\u6765\u8bc4\u4f30\u6700\u8fd1\u7684\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4ee3\u7406(RPCAs)\u3002\u6211\u4eec\u5f15\u5165GPT-4\u4ece\u73b0\u6709\u7684\u5c0f\u8bf4\u548c\u5267\u672c\u4e2d\u63d0\u53d6\u5bf9\u8bdd\uff0c\u5e76\u8fdb\u884c\u4e25\u683c\u7684\u4eba\u5de5\u8fc7\u6ee4\u3002\u7ecf\u8fc7\u4e00\u7cfb\u5217\u7684\u5904\u7406\uff0c\u6211\u4eec\u53d1\u5e03\u4e86\u4e00\u4e2a\u9ad8\u8d28\u91cf\u7684\u591a\u56de\u5408\u89d2\u8272\u626e\u6f14\u6570\u636e\u96c6\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6784\u5efa\u4e86\u4e00\u4e2a\u7efc\u5408\u8bc4\u4ef7\u4f53\u7cfb\uff0c\u5bf9rpca\u7684\u591a\u7ef4\u80fd\u529b\u8fdb\u884c\u8bc4\u4ef7\u3002\u6211\u4eec\u8fd8\u6536\u96c6\u4eba\u7c7b\u6ce8\u91ca\u6765\u8bad\u7ec3\u4e00\u4e2a\u57fa\u4e8e\u5b57\u7b26\u7684\u5956\u52b1\u6a21\u578b\u6765\u8861\u91cf\u4e3b\u89c2\u6307\u6807\uff0c\u4ee5\u4fbf\u4ee5\u540e\u65b9\u4fbf\u5730\u91cd\u65b0\u5b9e\u65bd\u3002\u5927\u91cf\u7684\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u4e2d\u6587llm\u5728\u4e2d\u6587\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4e2d\u6bd4GPT-4\u5177\u6709\u66f4\u5927\u7684\u6f5c\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6458\u8981 \u5927\u8bed\u8a00\u6a21\u578b\u7684\u51fa\u73b0\u6539\u53d8\u4e86\u751f\u6210\u5f0fAgent\uff0c\u5176\u4e2d\u6700\u8fd1\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b(llm)\u7684\u51fa\u73b0\u5f7b\u5e95\u6539\u53d8\u4e86\u751f\u6210\u4ee3\u7406\u3002\u5176\u4e2d\uff0c\u89d2\u8272\u626e\u6f14\u4f1a\u8bdd\u4ee3\u7406(R &#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":[45],"tags":[],"class_list":["post-622","post","type-post","status-publish","format-standard","hentry","category-llms"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/622","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=622"}],"version-history":[{"count":8,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/622\/revisions"}],"predecessor-version":[{"id":639,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/622\/revisions\/639"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=622"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=622"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=622"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}