{"id":21283,"date":"2024-07-19T15:42:41","date_gmt":"2024-07-19T07:42:41","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=21283"},"modified":"2024-07-19T15:42:41","modified_gmt":"2024-07-19T07:42:41","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e8%87%aa%e5%ad%a6%e5%90%8e%e8%83%bd%e5%8a%9b%e5%8f%8d%e4%b8%8b%e9%99%8d%ef%bc%8cllama-mistral%e9%83%bd%e6%b2%a1%e9%80%83%e8%bf%87","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2024\/07\/19\/21283\/","title":{"rendered":"\u5927\u6a21\u578b\u201c\u81ea\u5b66\u201d\u540e\u80fd\u529b\u53cd\u4e0b\u964d\uff0cLlama\/Mistral\u90fd\u6ca1\u9003\u8fc7"},"content":{"rendered":"

AI\u7ecf\u8fc7\u591a\u8f6e\u201c\u81ea\u6211\u63d0\u5347\u201d\uff0c\u80fd\u529b\u4e0d\u589e\u53cd\u964d\uff1f<\/p>\n

\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66GAIR\u56e2\u961f\u6700\u65b0\u7814\u7a76\u8868\u660e\uff0c\u5728\u5e38\u8bc6\u7406\u89e3\u3001\u6570\u5b66\u63a8\u7406\u548c\u4ee3\u7801\u751f\u6210\u7b49\u590d\u6742\u4efb\u52a1\u4e2d\uff0cAI\u7ecf\u8fc7\u591a\u8f6e\u201c\u81ea\u6211\u63d0\u5347\u201d\u540e\uff0c\u53ef\u80fd\u4f1a\u51fa\u73b0\u4e00\u79cd\u79f0\u4e3a\u201c\u81ea\u6211\u63d0\u5347\u9006\u8f6c<\/strong>\u201d\uff08self-improvement reversal\uff09\u7684\u73b0\u8c61\u3002<\/p>\n

LLaMA-2-7B\u3001Mistral-7B\u3001LLaMA-8B\u90fd\u6ca1\u9003\u8fc7\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u8fd9\u5c31\u50cf\u4e00\u4e2a\u5b66\u751f\u5237\u9898\u5237\u5230\u201d\u8d70\u706b\u5165\u9b54\u201d\u2014\u2014\u867d\u7136\u8003\u8bd5\u6210\u7ee9\u63d0\u9ad8\u4e86\uff0c\u4f46\u5b9e\u9645\u89e3\u51b3\u95ee\u9898\u7684\u80fd\u529b\u53cd\u800c\u53ef\u80fd\u4e0b\u964d\uff01<\/p>\n

\u66f4\u503c\u5f97\u8b66\u60d5\u7684\u662f\uff0c\u8fd9\u79cd\u8bad\u7ec3\u8fd8\u53ef\u80fd\u5bfc\u81f4AI\u7684\u56de\u7b54\u53d8\u5f97\u5343\u7bc7\u4e00\u5f8b\uff0c\u5931\u53bb\u4e86\u539f\u6709\u7684\u521b\u9020\u529b\u548c\u9002\u5e94\u65b0\u60c5\u51b5\u7684\u80fd\u529b<\/strong>\u3002\u597d\u6bd4\u4e00\u4e2a\u5b66\u751f\u53ea\u4f1a\u5e94\u4ed8\u8003\u8bd5\uff0c\u9047\u5230\u771f\u5b9e\u4e16\u754c\u7684\u95ee\u9898\u5c31\u675f\u624b\u65e0\u7b56\u3002<\/p>\n

\u8981\u77e5\u9053\uff0cOpenAI\u6700\u8fd1\u88ab\u66dd\u5149\u7684\u9879\u76ee\u201c\u8349\u8393\u201d\uff0c\u636e\u8bf4\u8fd8\u5728\u4f7f\u7528post-training\u9636\u6bb5\u7684\u81ea\u6211\u63d0\u5347\u6765\u63d0\u5347\u6a21\u578b\u590d\u6742\u63a8\u7406\u80fd\u529b\u2026\u2026<\/p>\n

\u76ee\u524d\u8be5\u7814\u7a76\u300aProgress or Regress\uff1fSelf-Improvement Reversal in Post-training\u300b\u5df2\u83b7\u5f97\u4e86 ICML 2024 \uff08AI for Math Workshop\uff09 \u7684Honorable Mention Award\u3002<\/p>\n

\u5237\u5206\u66f4\u9ad8\u4f46\u662f\u80fd\u529b\u4e0b\u964d<\/h2>\n

\u5177\u4f53\u8bf4\u6765\uff0c\u8be5\u5de5\u4f5c\u5c06\u201d\u8fed\u4ee3\u540e\u8bad\u7ec3\u201d\uff08Iterative post-training\uff09\u5206\u6210\u4e09\u4e2a\u4e3b\u8981\u6b65\u9aa4\uff1a<\/p>\n

    \n
  • \u7b54\u6848\u91c7\u6837\uff1a\u8ba9AI\u56de\u7b54\u4e00\u7cfb\u5217\u95ee\u9898\uff0c\u6bcf\u4e2a\u95ee\u9898\u56de\u7b54\u591a\u6b21\u3002<\/li>\n
  • \u8bad\u7ec3\u96c6\u6784\u5efa\uff1a\u4eceAI\u7684\u56de\u7b54\u4e2d\u6311\u9009\u51fa\u597d\u7684\u7b54\u6848\u3002<\/li>\n
  • \u6a21\u578b\u540e\u8bad\u7ec3\uff1a\u7528\u8fd9\u4e9b\u597d\u7b54\u6848\u6765\u201d\u6559\u201dAI\uff0c\u8ba9\u5b83\u5b66\u4f1a\u66f4\u597d\u7684\u56de\u7b54\u65b9\u5f0f\u3002<\/li>\n<\/ul>\n

    \u7814\u7a76\u4eba\u5458\u5c1d\u8bd5\u4e86\u4e0d\u540c\u7684\u201d\u6559\u5b66\u201d\u65b9\u6cd5\uff1a<\/p>\n

      \n
    • \u8fed\u4ee3SFT\uff1a\u76f4\u63a5\u544a\u8bc9AI\u54ea\u4e9b\u7b54\u6848\u662f\u5bf9\u7684\u3002<\/li>\n
    • \u8fed\u4ee3DPO\uff1a\u8ba9AI\u5b66\u4f1a\u6bd4\u8f83\u4e0d\u540c\u7b54\u6848\u7684\u597d\u574f\u3002<\/li>\n
    • \u8fed\u4ee3SFT-DPO\uff1a\u5c06\u4e0a\u9762\u4e24\u79cd\u65b9\u6cd5\u7ed3\u5408\u8d77\u6765\u3002<\/li>\n<\/ul>\n

      \u4ed6\u4eec\u8fd8\u7814\u7a76\u4e86\u5f71\u54cdAI\u5b66\u4e60\u6548\u679c\u7684\u51e0\u4e2a\u5173\u952e\u56e0\u7d20\uff08\u5982\u56fe1\u6240\u793a\uff09\uff1a<\/p>\n

        \n
      • \u8bad\u7ec3\u6b21\u6570\uff1a\u4e00\u822c\u6765\u8bf4\uff0c\u591a\u8bad\u7ec3\u51e0\u6b21\uff0cAI\u7684\u8868\u73b0\u4f1a\u66f4\u597d\u3002\u4f46\u8bad\u7ec34-5\u6b21\u540e\uff0c\u8fdb\u6b65\u5c31\u4e0d\u660e\u663e\u4e86\u3002<\/li>\n
      • AI\u7684\u57fa\u7840\u80fd\u529b\uff1a\u6709\u8da3\u7684\u662f\uff0c\u6700\u806a\u660e\u7684AI\u5e76\u4e0d\u4e00\u5b9a\u5b66\u5f97\u6700\u5feb\uff0c\u4f46\u6700\u540e\u8868\u73b0\u901a\u5e38\u6700\u597d\u3002<\/li>\n
      • \u95ee\u9898\u7c7b\u578b\uff1a\u6709\u4e9b\u7c7b\u578b\u7684\u95ee\u9898\uff08\u5982\u5e38\u8bc6\u95ee\u7b54\u548c\u7b80\u5355\u6570\u5b66\uff09AI\u5b66\u5f97\u6bd4\u8f83\u5feb\uff0c\u800c\u590d\u6742\u7684\u6570\u5b66\u95ee\u9898\u548c\u7f16\u7a0b\u4efb\u52a1\u5c31\u6bd4\u8f83\u56f0\u96be\u3002<\/li>\n
      • \u8bad\u7ec3\u65b9\u6cd5\uff1a\u4e0d\u540c\u7684\u95ee\u9898\u53ef\u80fd\u9700\u8981\u4e0d\u540c\u7684\u8bad\u7ec3\u65b9\u6cd5\u3002\u7814\u7a76\u8005\u53d1\u73b0\uff0c\u5982\u679cAI\u4e00\u5f00\u59cb\u5c31\u8868\u73b0\u4e0d\u9519\uff0c\u7528DPO\u6216SFT-DPO\u65b9\u6cd5\u6548\u679c\u66f4\u597d\u3002<\/li>\n<\/ul>\n

        \"\"<\/p>\n

        \u56fe1\uff1a\u4e09\u79cd\u8fed\u4ee3\u540e\u8bad\u7ec3\u8303\u5f0f\u5728\u591a\u4e2a\u4efb\u52a1\u6c42\u89e3\u4e0a\u7684Pass@1\u6027\u80fd\uff08\u4ece\u5e38\u8bc6\u77e5\u8bc6\u3001\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u65b9\u5411\u9009\u62e9\u4e86\u56db\u4e2a\u805a\u7126\u590d\u6742\u95ee\u9898\u6c42\u89e3\u80fd\u529b\u7684\u6570\u636e\u96c6\uff1aCSQA\uff0cGSM8k\uff0cMATH\uff0cMBPP\u4ee5\u53ca\u4e09\u79cd\u4e0d\u540c\u7684\u57fa\u5ea7\u6a21\u578bLLaMA-2-7B\uff0cMistral-7B\uff0cLLaMA-8B\u3002\u6240\u6709\u5b9e\u9a8c\u7684\u8fed\u4ee3\u6b21\u6570$$$$\u8bbe\u7f6e\u4e3a5\u3002\u8bc4\u4f30\u65f6\u4f7f\u7528\u8d2a\u5a6a\u89e3\u7801\uff0c\u5e76\u4e14\u9009\u62e9Pass@1\u4f5c\u4e3a\u6027\u80fd\u8bc4\u4ef7\u6307\u6807\u3002\uff09<\/p>\n

        \u8d77\u521d\uff0c\u7ed3\u679c\u770b\u8d77\u6765\u5f88\u4ee4\u4eba\u5174\u594b\u3002AI\u5728\u6d4b\u8bd5\u4e2d\u7684\u5206\u6570\u786e\u5b9e\u63d0\u9ad8\u4e86\uff01\u4f46\u7814\u7a76\u56e2\u961f\u6df1\u5165\u89c2\u5bdf\u540e\uff0c\u53d1\u73b0\u4e86\u4e00\u4e9b\u201c\u51fa\u4eba\u610f\u6599\u201d\u7684\u73b0\u8c61\uff1a<\/p>\n

        1\u3001\u80fd\u529b\u5e7b\u89c9<\/strong>\uff1a<\/p>\n

        \u7814\u7a76\u8005\u53d1\u73b0\uff0cAI\u5e76\u6ca1\u6709\u771f\u6b63\u5b66\u4f1a\u89e3\u51b3\u66f4\u96be\u7684\u95ee\u9898\u3002\u76f8\u53cd\uff0c\u5b83\u53ea\u662f\u53d8\u5f97\u66f4\u64c5\u957f\u5728\u5df2\u77e5\u7684\u7b54\u6848\u4e2d\u6311\u9009\u6b63\u786e\u7684\u90a3\u4e2a\u3002\u4ed6\u4eec\u7528\u201d\u6b63\u786e\u7b54\u6848\u8986\u76d6\u7387\u201d\u8fd9\u4e2a\u6307\u6807\u6765\u8861\u91cf\u8fd9\u4e00\u73b0\u8c61\u3002\u7ed3\u679c\u663e\u793a\uff0c\u5373\u4f7f\u662f\u672a\u7ecf\u8fc7\u591a\u8f6e\u8bad\u7ec3\u7684AI\uff0c\u53ea\u8981\u7ed9\u5b83\u8db3\u591f\u591a\u7684\u5c1d\u8bd5\u673a\u4f1a\uff0c\u4e5f\u80fd\u5728\u90a3\u4e9b\u770b\u4f3c\u201d\u5b66\u4f1a\u201d\u7684\u95ee\u9898\u4e0a\u5f97\u5230\u6b63\u786e\u7b54\u6848<\/strong>\u3002\u8fd9\u5c31\u50cf\u4e00\u4e2a\u5b66\u751f\u901a\u8fc7\u80cc\u7b54\u6848\u63d0\u9ad8\u4e86\u8003\u8bd5\u5206\u6570\uff0c\u4f46\u5b9e\u9645\u89e3\u51b3\u95ee\u9898\u7684\u80fd\u529b\u5e76\u6ca1\u6709\u63d0\u5347\u3002<\/p>\n

        2\u3001\u591a\u6837\u6027\u4e27\u5931<\/strong>\uff1a<\/p>\n

        \u968f\u7740\u8bad\u7ec3\u6b21\u6570\u589e\u52a0\uff0cAI\u7684\u56de\u7b54\u53d8\u5f97\u8d8a\u6765\u8d8a\u201c\u5343\u7bc7\u4e00\u5f8b\u201d<\/strong>\u3002\u7814\u7a76\u8005\u4ece\u4e09\u4e2a\u65b9\u9762\u6d4b\u91cf\u4e86\u8fd9\u79cd\u53d8\u5316\uff1a<\/p>\n

          \n
        • \u8bed\u6cd5\u591a\u6837\u6027\uff1a\u4f7f\u7528\u201dDistinct N-gram\u201d\u6307\u6807\u3002<\/li>\n
        • \u8bed\u4e49\u591a\u6837\u6027\uff1a\u4f7f\u7528\u53e5\u5b50\u5d4c\u5165\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u3002<\/li>\n
        • \u903b\u8f91\u591a\u6837\u6027\uff1a\u5728\u6570\u5b66\u95ee\u9898\u4e2d\u8ba1\u7b97\u4e0d\u540c\u65b9\u7a0b\u7684\u6570\u91cf\u3002\u7ed3\u679c\u663e\u793a\uff0c\u65e0\u8bba\u54ea\u79cd\u8bad\u7ec3\u65b9\u6cd5\uff0cAI\u7684\u56de\u7b54\u90fd\u53d8\u5f97\u8d8a\u6765\u8d8a\u76f8\u4f3c\uff0c\u5931\u53bb\u4e86\u539f\u6709\u7684\u521b\u610f\u548c\u591a\u6837\u6027\u3002<\/li>\n<\/ul>\n

          3\u3001\u6cdb\u5316\u80fd\u529b\u4e0b\u964d<\/strong>\uff1a<\/p>\n

          \u7814\u7a76\u8005\u8fd8\u6d4b\u8bd5\u4e86AI\u9762\u5bf9\u5168\u65b0\u7c7b\u578b\u95ee\u9898\u65f6\u7684\u8868\u73b0\u3002\u4ed6\u4eec\u5148\u8ba9AI\u5728\u4e00\u4e2a\u7b80\u5355\u7684\u6570\u5b66\u95ee\u9898\u96c6\uff08GSM8K\uff09\u4e0a\u8bad\u7ec3\uff0c\u7136\u540e\u7528\u4e00\u4e2a\u66f4\u96be\u7684\u6570\u5b66\u95ee\u9898\u96c6\uff08MATH\uff09\u6765\u6d4b\u8bd5\u5b83\u3002\u7ed3\u679c\u53d1\u73b0\uff0c\u7ecf\u8fc7\u591a\u8f6e\u201d\u81ea\u5b66\u201d\u7684AI\u5728\u9762\u5bf9\u8fd9\u4e9b\u65b0\u95ee\u9898\u65f6\uff0c\u8868\u73b0\u53cd\u800c\u66f4\u5dee\u3002\u800c\u4e14\uff0cAI\u5728\u7b80\u5355\u95ee\u9898\u548c\u56f0\u96be\u95ee\u9898\u4e0a\u7684\u8868\u73b0\u5dee\u8ddd\u8d8a\u6765\u8d8a\u5927\uff0c\u8fd9\u8bf4\u660e\u5b83\u53ef\u80fd\u53ea\u662f\u5728\u201d\u6b7b\u8bb0\u786c\u80cc\u201d\uff0c\u800c\u4e0d\u662f\u771f\u6b63\u7406\u89e3\u548c\u5b66\u4e60\u3002<\/p>\n

          \u6beb\u65e0\u7591\u95ee\uff0cpost-training\u9636\u6bb5\u7684\u201d\u81ea\u6211\u63d0\u5347\u201c\u4ecd\u7136\u662f\u4e00\u4e2a\u5145\u6ee1\u6f5c\u529b\u7684\u7814\u7a76\u65b9\u5411\uff0c\u201d\u8ba9AI\u81ea\u6211\u63d0\u5347\u201d\u4e5f\u662f\u4e00\u4e2a\u5f88\u9177\u7684\u60f3\u6cd5\u3002\u4f46GAIR\u56e2\u961f\u7684\u53d1\u73b0\u8868\u660e\uff0cAI\u7684\u8fdb\u5316\u4e4b\u8def\u53ef\u80fd\u6bd4\u60f3\u8c61\u4e2d\u66f4\u52a0\u590d\u6742\u548c\u5145\u6ee1\u6311\u6218\u3002\u5728\u8ffd\u6c42AI\u6027\u80fd\u63d0\u5347\u7684\u540c\u65f6\uff0c\u4e5f\u9700\u8981\u66f4\u5168\u9762\u5730\u8003\u8651\uff1a<\/p>\n

            \n
          • AI\u771f\u6b63\u7684\u95ee\u9898\u89e3\u51b3\u80fd\u529b\u662f\u5426\u63d0\u5347\u4e86\uff1f<\/li>\n
          • AI\u662f\u5426\u4fdd\u6301\u4e86\u521b\u9020\u529b\u548c\u591a\u6837\u6027\uff1f<\/li>\n
          • AI\u80fd\u5426\u7075\u6d3b\u5e94\u5bf9\u65b0\u7684\u3001\u672a\u77e5\u7684\u60c5\u51b5\uff1f<\/li>\n<\/ul>\n

            \u4eba\u5de5\u667a\u80fd\u7684\u672a\u6765\u4ee4\u4eba\u671f\u5f85\uff0c\u4f46\u540c\u65f6\u4e5f\u9700\u8981\u4ee5\u66f4\u52a0\u5ba1\u614e\u548c\u5168\u9762\u7684\u89c6\u89d2\u6765\u770b\u5f85\u5176\u53d1\u5c55\u3002\u53ea\u6709\u8fd9\u6837\uff0c\u624d\u80fd\u771f\u6b63\u5b9e\u73b0AI\u7684\u6f5c\u529b\uff0c\u521b\u9020\u51fa\u65e2\u667a\u80fd\u53c8\u53ef\u9760\u7684\u7cfb\u7edf\uff0c\u4e3a\u4eba\u7c7b\u793e\u4f1a\u5e26\u6765\u771f\u6b63\u7684\u4ef7\u503c\u3002<\/p>\n

            \u6765\u6e90\uff1a\u91cf\u5b50\u4f4d<\/p>\n

            <\/div>\n

            <\/g><\/svg><\/i> \"Loading\"<\/p>\n

            <\/div>\n","protected":false},"excerpt":{"rendered":"

            AI\u7ecf\u8fc7\u591a\u8f6e\u201c\u81ea\u6211\u63d0\u5347\u201d\uff0c\u80fd\u529b\u4e0d\u589e\u53cd\u964d\uff1f \u4e0a\u6d77\u4ea4\u901a\u5927\u5b66GAIR\u56e2\u961f\u6700\u65b0\u7814\u7a76\u8868\u660e\uff0c\u5728\u5e38\u8bc6\u7406\u89e3\u3001\u6570\u5b66\u63a8\u7406\u548c\u4ee3\u7801\u751f\u6210 […]<\/p>\n

            <\/div>\n

            <\/g><\/svg><\/i> \"Loading\"<\/p>\n

            <\/div>\n","protected":false},"author":139,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23],"tags":[],"class_list":["post-21283","post","type-post","status-publish","format-standard","hentry","category-ai","category-23"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/21283","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/139"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=21283"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/21283\/revisions"}],"predecessor-version":[{"id":21285,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/21283\/revisions\/21285"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=21283"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=21283"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=21283"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}