\u300aChain-of-Thought Prompting Elicits Reasoning in Large Language Models\u300b<\/p>\n\n\n\n
\u4ee5\u4e0b\u662f\u7531GPT4o\u5c0d\u300aChain-of-Thought Prompting Elicits Reasoning in Large Language Models\u300b\u9019\u7bc7\u8ad6\u6587\u7684\u7e41\u9ad4\u4e2d\u6587\u7ffb\u8b6f\u8207\u81c9\u66f8\u8cbc\u6587\uff0c\u4e26\u4fdd\u7559\u539f\u59cb\u53c3\u8003\u6587\u737b\u7684\u7db2\u5740\u3002<\/p>\n\n\n\n
\u9023\u7d50\u751f\u6210\u601d\u7dad\u93c8\u7684\u63d0\u793a (Chain-of-Thought Prompting) \u5982\u4f55\u63d0\u5347\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b<\/p>\n\n\n\n
\u6458\u8981<\/p>\n\n\n\n
\u6211\u5011\u63a2\u8a0e\u5982\u4f55\u900f\u904e\u300c\u601d\u7dad\u93c8\u63d0\u793a\u300d\uff08Chain-of-Thought Prompting\uff09\uff0c\u5373\u63d0\u4f9b\u4e00\u7cfb\u5217\u7684\u4e2d\u9593\u63a8\u7406\u6b65\u9a5f\uff0c\u986f\u8457\u63d0\u5347\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u5728\u8907\u96dc\u63a8\u7406\u4efb\u52d9\u4e0a\u7684\u8868\u73fe\u3002\u7279\u5225\u662f\uff0c\u6211\u5011\u5c55\u793a\u4e86\u7576\u6a21\u578b\u898f\u6a21\u8db3\u5920\u5927\u6642\uff0c\u9019\u7a2e\u63a8\u7406\u80fd\u529b\u80fd\u5920\u81ea\u7136\u5730\u900f\u904e\u7c21\u55ae\u7684\u63d0\u793a\u65b9\u6cd5\u2014\u2014\u63d0\u4f9b\u5e7e\u500b\u5305\u542b\u601d\u7dad\u93c8\u7684\u793a\u4f8b\u4f86\u5f15\u5c0e\u6a21\u578b\u2014\u2014\u800c\u6e67\u73fe\u51fa\u4f86\u3002<\/p>\n\n\n\n
\u6211\u5011\u5728\u4e09\u7a2e\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u4e0a\u9032\u884c\u5be6\u9a57\uff0c\u7d50\u679c\u986f\u793a\uff0c\u76f8\u8f03\u65bc\u6a19\u6e96\u63d0\u793a\u6cd5\uff0c\u601d\u7dad\u93c8\u63d0\u793a\u80fd\u5920\u986f\u8457\u63d0\u5347\u6578\u5b78\u904b\u7b97\u3001\u5e38\u8b58\u63a8\u7406\u4ee5\u53ca\u7b26\u865f\u63a8\u7406\u7b49\u591a\u9805\u4efb\u52d9\u7684\u8868\u73fe\u3002\u4f8b\u5982\uff0c\u4f7f\u7528 PaLM 540B \u4e26\u50c5\u63d0\u4f9b 8 \u500b\u601d\u7dad\u93c8\u793a\u4f8b\uff0c\u5373\u53ef\u5728 GSM8K \u6578\u5b78\u984c\u5eab\u4e0a\u9054\u5230 SOTA\uff08\u6700\u5148\u9032\uff09\u6c34\u6e96\uff0c\u751a\u81f3\u8d85\u8d8a\u7d93\u904e\u5fae\u8abf\u4e26\u5177\u9a57\u8b49\u6a5f\u5236\u7684 GPT-3\u3002<\/p>\n\n\n\n
\u4ec0\u9ebc\u662f\u601d\u7dad\u93c8\u63d0\u793a\uff1f<\/p>\n\n\n\n
\u300c\u601d\u7dad\u93c8\u300d\uff08Chain of Thought, CoT\uff09\u662f\u6307\u5728\u56de\u7b54\u554f\u984c\u7684\u904e\u7a0b\u4e2d\uff0c\u900f\u904e\u4e00\u7cfb\u5217\u7684\u4e2d\u9593\u63a8\u7406\u6b65\u9a5f\u4f86\u5e6b\u52a9\u6a21\u578b\u5f97\u51fa\u6700\u7d42\u7b54\u6848\u3002\u4f8b\u5982\uff0c\u5728\u89e3\u6578\u5b78\u984c\u6642\uff0cCoT \u6703\u8b93\u6a21\u578b\u5148\u9010\u6b65\u62c6\u89e3\u554f\u984c\uff0c\u518d\u9032\u884c\u8a08\u7b97\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u7d66\u51fa\u7b54\u6848\u3002<\/p>\n\n\n\n
\ud83d\udd39 \u7bc4\u4f8b\u6bd4\u8f03
\u2022 \u6a19\u6e96\u63d0\u793a\uff08Standard Prompting\uff09
\u2022 Q: \u7f85\u5091\u6709 5 \u9846\u7db2\u7403\uff0c\u4ed6\u53c8\u8cb7\u4e86 2 \u7f50\u7db2\u7403\uff0c\u6bcf\u7f50 3 \u9846\u3002\u4ed6\u73fe\u5728\u6709\u591a\u5c11\u9846\u7db2\u7403\uff1f
\u2022 A: 11\uff08\u76f4\u63a5\u7d66\u51fa\u7b54\u6848\uff09
\u2022 \u601d\u7dad\u93c8\u63d0\u793a\uff08Chain-of-Thought Prompting\uff09
\u2022 Q: \u7f85\u5091\u6709 5 \u9846\u7db2\u7403\uff0c\u4ed6\u53c8\u8cb7\u4e86 2 \u7f50\u7db2\u7403\uff0c\u6bcf\u7f50 3 \u9846\u3002\u4ed6\u73fe\u5728\u6709\u591a\u5c11\u9846\u7db2\u7403\uff1f
\u2022 A: \u7f85\u5091\u6700\u521d\u6709 5 \u9846\u7db2\u7403\u3002\u6bcf\u7f50 3 \u9846\u7403\uff0c2 \u7f50\u5171\u6709 6 \u9846\u30025 + 6 = 11\u3002\u6240\u4ee5\u7b54\u6848\u662f 11\u3002<\/p>\n\n\n\n
\u70ba\u4f55\u9019\u6a23\u505a\u6709\u7528\uff1f
\u7576\u6211\u5011\u89e3\u6578\u5b78\u984c\u6642\uff0c\u5927\u591a\u6578\u4eba\u6703\u81ea\u7136\u5730\u5206\u89e3\u554f\u984c\u4e26\u9010\u6b65\u8a08\u7b97\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u8df3\u5230\u7b54\u6848\u3002\u900f\u904e CoT\uff0c\u8a9e\u8a00\u6a21\u578b\u80fd\u5920\u6a21\u4eff\u9019\u7a2e\u9010\u6b65\u601d\u8003\u7684\u904e\u7a0b\uff0c\u5f9e\u800c\u63d0\u9ad8\u6e96\u78ba\u5ea6\u3002<\/p>\n\n\n\n
\u5be6\u9a57\u7d50\u679c\uff1a\u601d\u7dad\u93c8\u63d0\u793a\u986f\u8457\u63d0\u5347\u63a8\u7406\u80fd\u529b<\/p>\n\n\n\n
\u6211\u5011\u5728\u4e0d\u540c\u8a9e\u8a00\u6a21\u578b\uff08GPT-3\u3001LaMDA\u3001PaLM 540B\uff09\u4e0a\u6e2c\u8a66\u4e86\u601d\u7dad\u93c8\u63d0\u793a\uff0c\u7d50\u679c\u767c\u73fe\uff1a<\/p>\n\n\n\n
\u2705 \u6578\u5b78\u904b\u7b97\uff08Arithmetic Reasoning\uff09
\u2022 \u5728 GSM8K\uff08\u6578\u5b78\u984c\u5eab\uff09\u4e0a\uff0c\u601d\u7dad\u93c8\u63d0\u793a\u8b93\u6a21\u578b\u7684\u6e96\u78ba\u7387\u8d85\u8d8a\u4ee5\u5f80\u6240\u6709\u65b9\u6cd5\uff0c\u5305\u62ec\u5fae\u8abf\u5f8c\u7684 GPT-3\u3002<\/p>\n\n\n\n
\u2705 \u5e38\u8b58\u63a8\u7406\uff08Commonsense Reasoning\uff09
\u2022 \u5728 StrategyQA \u548c CommonsenseQA \u6e2c\u8a66\u96c6\u4e2d\uff0c\u4f7f\u7528 CoT \u63d0\u793a\u5f8c\uff0cPaLM 540B \u7684\u8868\u73fe\u8d85\u8d8a\u4e86\u4eba\u985e\u5e73\u5747\u6c34\u5e73\u3002<\/p>\n\n\n\n
\u2705 \u7b26\u865f\u63a8\u7406\uff08Symbolic Reasoning\uff09
\u2022 \u5728\u300c\u6700\u5f8c\u4e00\u500b\u5b57\u6bcd\u9023\u7d50\u300d\u8207\u300c\u786c\u5e63\u7ffb\u8f49\u300d\u6e2c\u8a66\u4e2d\uff0c\u601d\u7dad\u93c8\u63d0\u793a\u8b93\u6a21\u578b\u80fd\u5920\u985e\u63a8\u4e26\u63a8\u5ee3\u81f3\u66f4\u9577\u7684\u5e8f\u5217\uff0c\u6bd4\u6a19\u6e96\u63d0\u793a\u6cd5\u66f4\u5177\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n
\ud83d\udca1 \u95dc\u9375\u767c\u73fe
1\ufe0f\u20e3 \u601d\u7dad\u93c8\u63a8\u7406\u662f\u96a8\u8457\u6a21\u578b\u898f\u6a21\u589e\u5927\u800c\u6e67\u73fe\u7684\u80fd\u529b\u3002\u7576\u6a21\u578b\u8d85\u904e 100B \u53c3\u6578\u6642\uff0cCoT \u63d0\u793a\u7684\u6548\u679c\u986f\u8457\u63d0\u5347\u3002
2\ufe0f\u20e3 \u6bd4\u55ae\u7d14\u63d0\u4f9b\u6578\u5b78\u65b9\u7a0b\u66f4\u6709\u6548\u3002\u76f4\u63a5\u7d66\u6a21\u578b\u6578\u5b78\u516c\u5f0f\u7684\u63d0\u793a\u6548\u679c\u4e0d\u5982 CoT\uff0c\u8aaa\u660e\u81ea\u7136\u8a9e\u8a00\u7684\u63a8\u7406\u6b65\u9a5f\u5c0d\u6a21\u578b\u6709\u5e6b\u52a9\u3002
3\ufe0f\u20e3 \u9069\u7528\u65bc\u591a\u7a2e\u63a8\u7406\u4efb\u52d9\uff0c\u4e0d\u50c5\u9650\u65bc\u6578\u5b78\uff0c\u9084\u80fd\u589e\u5f37\u5e38\u8b58\u63a8\u7406\u8207\u7b26\u865f\u63a8\u7406\u3002<\/p>\n\n\n\n
\u672a\u4f86\u767c\u5c55\u8207\u6311\u6230<\/p>\n\n\n\n
\ud83d\ude80 \u96d6\u7136 CoT \u63d0\u793a\u5c55\u73fe\u4e86\u9a5a\u4eba\u7684\u6f5b\u529b\uff0c\u4f46\u4ecd\u7136\u6709\u4e00\u4e9b\u9650\u5236\uff1a
\u2022 \u76ee\u524d\uff0c\u9019\u7a2e\u63a8\u7406\u80fd\u529b\u50c5\u5728\u8d85\u5927\u898f\u6a21\u8a9e\u8a00\u6a21\u578b\u4e0a\u624d\u80fd\u767c\u63ee\u4f5c\u7528\uff0c\u8f03\u5c0f\u7684\u6a21\u578b\u7121\u6cd5\u53d7\u76ca\u3002
\u2022 \u751f\u6210\u7684\u63a8\u7406\u6b65\u9a5f\u672a\u5fc5\u7e3d\u662f\u6b63\u78ba\u7684\uff0c\u6709\u6642\u53ef\u80fd\u6703\u51fa\u73fe\u300c\u80e1\u8aaa\u516b\u9053\u300d\u7684\u60c5\u6cc1\u3002
\u2022 \u9700\u8981\u63a2\u7d22\u5982\u4f55\u8b93\u8f03\u5c0f\u7684\u6a21\u578b\u4e5f\u80fd\u5b78\u7fd2\u9019\u7a2e\u63a8\u7406\u65b9\u6cd5\uff0c\u4ee5\u964d\u4f4e\u8a08\u7b97\u6210\u672c\u4e26\u63d0\u9ad8\u61c9\u7528\u7bc4\u570d\u3002<\/p>\n\n\n\n
\ud83d\udd17 \u53c3\u8003\u6587\u737b
\ud83d\udcc4 \u539f\u59cb\u8ad6\u6587: arXiv:2201.11903
\ud83d\udcca GSM8K \u6578\u5b78\u6e2c\u8a66\u57fa\u6e96: arXiv:2110.14168
\ud83d\udcd1 \u76f8\u95dc\u63a8\u7406\u7814\u7a76: arXiv:2203.11171<\/p>\n\n\n\n
\u7d50\u8ad6<\/p>\n\n\n\n
\u900f\u904e\u300c\u601d\u7dad\u93c8\u63d0\u793a\u300d\uff0c\u6211\u5011\u53ef\u4ee5\u8b93\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u5728\u6578\u5b78\u3001\u5e38\u8b58\u3001\u7b26\u865f\u63a8\u7406\u7b49\u9818\u57df\u8868\u73fe\u5f97\u66f4\u8070\u660e\u3002\u9019\u9805\u6280\u8853\u70ba AI \u63a8\u7406\u80fd\u529b\u7684\u63d0\u5347\u958b\u95e2\u4e86\u4e00\u689d\u65b0\u9053\u8def\uff0c\u672a\u4f86\u53ef\u80fd\u6703\u61c9\u7528\u65bc\u66f4\u5ee3\u6cdb\u7684\u5834\u666f\uff0c\u5982\u6cd5\u5f8b\u63a8\u7406\u3001\u91ab\u7642\u8a3a\u65b7\u3001\u91d1\u878d\u5206\u6790\u7b49\u3002<\/p>\n\n\n\n
\u4f60\u8a8d\u70ba AI \u771f\u7684\u80fd\u300c\u63a8\u7406\u300d\u55ce\uff1f\u9084\u662f\u53ea\u662f\u300c\u6a21\u4eff\u300d\u6211\u5011\u7684\u63a8\u7406\u904e\u7a0b\uff1f\u6b61\u8fce\u7559\u8a00\u8a0e\u8ad6\uff01\ud83d\udcac<\/p>\n\n\n\n
\u9019\u7bc7\u81c9\u66f8\u8cbc\u6587\u65e2\u5305\u542b\u8ad6\u6587\u7684\u6838\u5fc3\u5167\u5bb9\uff0c\u53c8\u4ee5\u8f03\u70ba\u901a\u4fd7\u7684\u65b9\u5f0f\u89e3\u91cb\u5176\u5f71\u97ff\uff0c\u9069\u5408\u63a8\u5ee3\u81f3\u66f4\u5ee3\u6cdb\u7684\u53d7\u773e\u3002\u5982\u679c\u6709\u4efb\u4f55\u9700\u8981\u8abf\u6574\u7684\u5730\u65b9\uff0c\u8acb\u544a\u8a34\u6211<\/p>\n\n\n\n
<\/p>\n","protected":false},"excerpt":{"rendered":"
\u300aChain-of-Thought Prompting Elicits Reasoning in Large […]<\/p>\n","protected":false},"author":2,"featured_media":527,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"nf_dc_page":"","iawp_total_views":1,"footnotes":""},"categories":[56,27],"tags":[34,28,116,88,76,31,121,122,74],"class_list":["post-568","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-56","category-llm","tag-ai","tag-chatgpt","tag-deepseek","tag-openai","tag-rag","tag-31","tag-121","tag-122","tag-74"],"_links":{"self":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts\/568","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=568"}],"version-history":[{"count":1,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts\/568\/revisions"}],"predecessor-version":[{"id":569,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts\/568\/revisions\/569"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/media\/527"}],"wp:attachment":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=568"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=568"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=568"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}