{"id":103321,"date":"2026-09-10T09:09:02","date_gmt":"2026-09-10T01:09:02","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/103321.html"},"modified":"2026-09-10T09:09:02","modified_gmt":"2026-09-10T01:09:02","slug":"reasoningbank-scaling-agent-self-evolving-with-reasoning-memory%e8%ae%ba%e6%96%87%e9%98%85%e8%af%bb","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/103321.html","title":{"rendered":"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb"},"content":{"rendered":"

\u4eca\u5929\u5206\u4eab\u7684\u8bba\u6587\u662f\u300aReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u300b<\/p>\n

\u539f\u6587\u94fe\u63a5：https:\/\/arxiv.org\/abs\/2509.25140<\/p>\n

\n

\u8fd9\u662f\u4e00\u7bc7\u8c37\u6b4c\u7684\u8bba\u6587，\u5173\u4e8e\u81ea\u6f14\u5316\u667a\u80fd\u4f53\u7684\u3002<\/p>\n

\n

\u968f\u7740\u5927\u578b\u8bed\u8a00\u6a21\u578b（LLM）\u667a\u80fd\u4f53\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u6301\u7eed\u53d1\u6325\u4f5c\u7528\u7684\u573a\u666f\u65e5\u76ca\u589e\u591a，\u5b83\u4eec\u81ea\u7136\u4f1a\u9762\u4e34\u6e90\u6e90\u4e0d\u65ad\u7684\u4efb\u52a1\u6d41\u3002\u7136\u800c，\u5176\u5173\u952e\u5c40\u9650\u5728\u4e8e\u65e0\u6cd5\u4ece\u7d2f\u79ef\u7684\u4ea4\u4e92\u5386\u53f2\u4e2d\u5b66\u4e60，\u5bfc\u81f4\u4e0d\u5f97\u4e0d\u4e22\u5f03\u6709\u4ef7\u503c\u7684\u89c1\u89e3\u5e76\u91cd\u590d\u8fc7\u53bb\u7684\u9519\u8bef\u3002\u672c\u6587\u63d0\u51faReasoningBank\u2014\u2014\u4e00\u79cd\u65b0\u578b\u8bb0\u5fc6\u6846\u67b6，\u80fd\u591f\u4ece\u667a\u80fd\u4f53\u81ea\u6211\u5224\u65ad\u7684\u6210\u529f\u4e0e\u5931\u8d25\u7ecf\u9a8c\u4e2d\u63d0\u70bc\u53ef\u6cdb\u5316\u7684\u63a8\u7406\u7b56\u7565\u3002\u5728\u6d4b\u8bd5\u9636\u6bb5，\u667a\u80fd\u4f53\u4eceReasoningBank\u4e2d\u68c0\u7d22\u76f8\u5173\u8bb0\u5fc6\u4ee5\u6307\u5bfc\u4ea4\u4e92，\u968f\u540e\u5c06\u65b0\u5b66\u5230\u7684\u77e5\u8bc6\u6574\u5408\u56de\u6846\u67b6\u4e2d，\u4f7f\u5176\u80fd\u529b\u968f\u65f6\u95f4\u63a8\u79fb\u4e0d\u65ad\u63d0\u5347\u3002\u57fa\u4e8e\u8fd9\u4e00\u5f3a\u5927\u7684\u7ecf\u9a8c\u5b66\u4e60\u673a\u5236，\u672c\u6587\u8fdb\u4e00\u6b65\u63d0\u51fa\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u65f6\u6269\u5c55（MaTTS）：\u901a\u8fc7\u6269\u5c55\u667a\u80fd\u4f53\u7684\u4ea4\u4e92\u7ecf\u9a8c，\u52a0\u901f\u5e76\u4e30\u5bcc\u5176\u5b66\u4e60\u8fc7\u7a0b\u3002\u901a\u8fc7\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u5206\u914d\u66f4\u591a\u8ba1\u7b97\u8d44\u6e90，\u667a\u80fd\u4f53\u53ef\u751f\u6210\u4e30\u5bcc\u591a\u6837\u7684\u7ecf\u9a8c，\u4e3a\u5408\u6210\u66f4\u9ad8\u8d28\u91cf\u7684\u8bb0\u5fc6\u63d0\u4f9b\u5145\u8db3\u7684\u5bf9\u6bd4\u4fe1\u53f7；\u800c\u66f4\u4f18\u8d28\u7684\u8bb0\u5fc6\u53cd\u8fc7\u6765\u53c8\u80fd\u6307\u5bfc\u66f4\u6709\u6548\u7684\u6269\u5c55，\u6700\u7ec8\u5728\u8bb0\u5fc6\u4e0e\u6d4b\u8bd5\u65f6\u6269\u5c55\u4e4b\u95f4\u5f62\u6210\u5f3a\u5927\u7684\u534f\u540c\u6548\u5e94\u3002\u5728\u7f51\u9875\u6d4f\u89c8\u548c\u8f6f\u4ef6\u5de5\u7a0b\u57fa\u51c6\u6d4b\u8bd5\u4e2d，ReasoningBank\u6301\u7eed\u4f18\u4e8e\u73b0\u6709\u4ec5\u5b58\u50a8\u539f\u59cb\u8f68\u8ff9\u6216\u5355\u4e00\u6210\u529f\u4efb\u52a1\u6d41\u7a0b\u7684\u8bb0\u5fc6\u673a\u5236，\u5728\u6709\u6548\u6027\u548c\u6548\u7387\u4e0a\u5747\u6709\u63d0\u5347；MaTTS\u5219\u8fdb\u4e00\u6b65\u653e\u5927\u4e86\u8fd9\u4e9b\u4f18\u52bf\u3002\u8fd9\u4e9b\u53d1\u73b0\u8bc1\u5b9e，\u57fa\u4e8e\u8bb0\u5fc6\u7684\u7ecf\u9a8c\u6269\u5c55\u53ef\u4f5c\u4e3a\u667a\u80fd\u4f53\u7684\u65b0\u6269\u5c55\u7ef4\u5ea6，\u4f7f\u667a\u80fd\u4f53\u80fd\u591f\u81ea\u7136\u6d8c\u73b0\u51fa\u590d\u6742\u884c\u4e3a\u5e76\u5b9e\u73b0\u81ea\u8fdb\u5316\u3002<\/p>\n

\n

\u5927\u578b\u8bed\u8a00\u6a21\u578b（LLM）\u7684\u5feb\u901f\u53d1\u5c55\u6781\u5927\u5730\u63a8\u52a8\u4e86LLM\u667a\u80fd\u4f53\u7684\u7814\u53d1（Liu et al., 2025a; Wang et al., 2024）\u3002\u8fd9\u7c7b\u667a\u80fd\u4f53\u5bf9\u4e8e\u5904\u7406\u9700\u8981\u4e0e\u73af\u5883\u8fdb\u884c\u591a\u6b65\u4ea4\u4e92\u7684\u590d\u6742\u73b0\u5b9e\u4efb\u52a1\u81f3\u5173\u91cd\u8981，\u4f8b\u5982\u7f51\u9875\u6d4f\u89c8（Gur et al., 2024）\u548c\u8ba1\u7b97\u673a\u64cd\u4f5c（Xie et al., 2024; Yang et al., 2024）\u3002\u968f\u7740\u667a\u80fd\u4f53\u8d8a\u6765\u8d8a\u591a\u5730\u88ab\u90e8\u7f72\u5230\u957f\u671f\u3001\u6301\u7eed\u8fd0\u884c\u7684\u573a\u666f\u4e2d，\u5176\u5728\u751f\u547d\u5468\u671f\u5185\u5fc5\u7136\u4f1a\u9762\u4e34\u6e90\u6e90\u4e0d\u65ad\u7684\u4efb\u52a1\u6d41\u3002\u7136\u800c，\u76ee\u524d\u7684\u667a\u80fd\u4f53\u5927\u591a\u65e0\u6cd5\u4ece\u8de8\u4efb\u52a1\u7684\u7d2f\u79ef\u7ecf\u9a8c\u4e2d\u5b66\u4e60：\u5b83\u4eec\u5c06\u6bcf\u4e2a\u4efb\u52a1\u89c6\u4e3a\u72ec\u7acb\u4e2a\u4f53，\u8fd9\u5bfc\u81f4\u5176\u5fc5\u7136\u91cd\u590d\u8fc7\u53bb\u7684\u9519\u8bef（Yin et al., 2025）\u3001\u4e22\u5f03\u76f8\u5173\u95ee\u9898\u4e2d\u7684\u6709\u4ef7\u503c\u89c1\u89e3，\u5e76\u4e14\u7f3a\u4e4f\u4f7f\u7cfb\u7edf\u80fd\u529b\u968f\u65f6\u95f4\u63d0\u5347\u7684\u81ea\u8fdb\u5316\u80fd\u529b（Gao et al., 2025）\u3002\u8fd9\u51f8\u663e\u4e86\u6784\u5efa\u53ef\u4ece\u8fc7\u5f80\u7ecf\u9a8c\u5b66\u4e60\u7684\u8bb0\u5fc6\u611f\u77e5\u578b\u667a\u80fd\u4f53\u7cfb\u7edf\u7684\u5fc5\u8981\u6027（Zhang et al., 2024b）\u3002<\/p>\n

\u5f53\u524d\u5173\u4e8e\u667a\u80fd\u4f53\u8bb0\u5fc6\u7684\u7814\u7a76\u4e3b\u8981\u96c6\u4e2d\u4e8e\u5b58\u50a8\u8fc7\u5f80\u4ea4\u4e92\u4ee5\u5b9e\u73b0\u590d\u7528（Chen et al., 2025; Sun et al., 2025; Tang et al., 2025b）\u3002\u5c3d\u7ba1\u8fd9\u7c7b\u65b9\u6cd5\u5177\u6709\u4e00\u5b9a\u4f5c\u7528，\u4f46\u901a\u5e38\u5c40\u9650\u4e8e\u5229\u7528\u539f\u59cb\u8f68\u8ff9（Kagaya et al., 2024; Kong et al., 2025; Zheng et al., 2024）\u6216\u5e38\u89c1\u7684\u6210\u529f\u6d41\u7a0b（\u5373\u5de5\u4f5c\u6d41\u3001\u6b65\u9aa4\u6d41\u7a0b）（Fang et al., 2025; Wang et al., 2025d），\u5e76\u5b58\u5728\u4e24\u4e2a\u6839\u672c\u6027\u7f3a\u9677：\u00a0<\/p>\n

1. \u65e0\u6cd5\u63d0\u70bc\u66f4\u9ad8\u5c42\u6b21\u3001\u53ef\u8fc1\u79fb\u7684\u63a8\u7406\u6a21\u5f0f；\u00a0<\/p>\n

2. \u8fc7\u5ea6\u5f3a\u8c03\u6210\u529f\u7ecf\u9a8c，\u5927\u5e45\u5ffd\u89c6\u4e86\u667a\u80fd\u4f53\u81ea\u8eab\u5931\u8d25\u4e2d\u8574\u542b\u7684\u5b9d\u8d35\u6559\u8bad（Zhang et al., 2024a）\u3002\u00a0<\/p>\n

\n

\u56e0\u6b64，\u73b0\u6709\u8bb0\u5fc6\u8bbe\u8ba1\u5f80\u5f80\u505c\u7559\u5728\u201c\u88ab\u52a8\u8bb0\u5f55\u201d\u5c42\u9762，\u65e0\u6cd5\u4e3a\u672a\u6765\u51b3\u7b56\u63d0\u4f9b\u53ef\u6267\u884c\u3001\u53ef\u6cdb\u5316\u7684\u6307\u5bfc\u3002<\/p>\n

\u4e3a\u586b\u8865\u8fd9\u4e00\u7a7a\u767d，\u672c\u6587\u63d0\u51faReasoningBank\u2014\u2014\u4e00\u79cd\u9762\u5411\u667a\u80fd\u4f53\u7cfb\u7edf\u7684\u65b0\u578b\u8bb0\u5fc6\u6846\u67b6\u3002ReasoningBank\u65e0\u9700\u771f\u503c\u6807\u7b7e，\u4ec5\u901a\u8fc7\u667a\u80fd\u4f53\u81ea\u8eab\u7684\u5224\u65ad，\u4ece\u6210\u529f\u548c\u5931\u8d25\u7ecf\u9a8c\u4e2d\u63d0\u70bc\u5e76\u7ec4\u7ec7\u8bb0\u5fc6\u5355\u5143\u3002\u5982\u56fe1\u6240\u793a，\u8be5\u6846\u67b6\u4e0d\u4ec5\u6355\u83b7\u6210\u529f\u7ecf\u9a8c\u4e2d\u7684\u6709\u6548\u7b56\u7565，\u8fd8\u63d0\u53d6\u5931\u8d25\u7ecf\u9a8c\u4e2d\u7684\u5173\u952e\u9884\u9632\u6559\u8bad，\u5e76\u5c06\u5176\u62bd\u8c61\u4e3a\u4e00\u7ec4\u53ef\u6267\u884c\u7684\u539f\u5219\u3002\u6574\u4e2a\u8fc7\u7a0b\u4ee5\u95ed\u73af\u5f62\u5f0f\u8fd0\u884c：\u9762\u5bf9\u65b0\u4efb\u52a1\u65f6，\u667a\u80fd\u4f53\u4eceReasoningBank\u4e2d\u68c0\u7d22\u76f8\u5173\u8bb0\u5fc6\u4ee5\u6307\u5bfc\u884c\u52a8；\u4efb\u52a1\u5b8c\u6210\u540e，\u65b0\u7ecf\u9a8c\u4f1a\u88ab\u5206\u6790\u3001\u63d0\u70bc\u5e76\u6574\u5408\u56deReasoningBank，\u4f7f\u667a\u80fd\u4f53\u80fd\u591f\u6301\u7eed\u8fdb\u5316\u5e76\u63d0\u5347\u5176\u7b56\u7565\u80fd\u529b\u3002<\/p>\n

\"\"<\/p>\n

\u56fe1\u5c55\u793a\u4e86ReasoningBank\u5982\u4f55\u751f\u6210\u53ef\u590d\u7528\u7684\u63a8\u7406\u7b56\u7565，\u4f7f\u8bb0\u5fc6\u5355\u5143\u5bf9\u672a\u6765\u4efb\u52a1\u66f4\u5177\u8fc1\u79fb\u6027\u3002\u76f8\u6bd4\u201c\u65e0\u8bb0\u5fc6\u201d\u57fa\u7ebf，\u8be5\u6846\u67b6\u80fd\u8ba9\u667a\u80fd\u4f53\u5728WebArena-Admin\u5b50\u96c6\u4e0a\u5b9e\u73b0\u6301\u7eed\u8fdb\u5316，\u5e76\u83b7\u5f97\u66f4\u9ad8\u7684\u7d2f\u79ef\u6210\u529f\u7387\u3002<\/p>\n

\n

\u4ee5ReasoningBank\u8fd9\u4e00\u5f3a\u5927\u7684\u7ecf\u9a8c\u5b66\u4e60\u673a\u5236\u4e3a\u57fa\u7840，\u672c\u6587\u8fdb\u4e00\u6b65\u7814\u7a76\u7ecf\u9a8c\u6269\u5c55，\u65e8\u5728\u6784\u5efa\u8bb0\u5fc6\u4e0e\u6d4b\u8bd5\u65f6\u6269\u5c55\u4e4b\u95f4\u7684\u5f3a\u5927\u534f\u540c\u5173\u7cfb\u3002\u4e0d\u540c\u4e8e\u901a\u8fc7\u201c\u589e\u52a0\u4efb\u52a1\u6570\u91cf\u201d\u5b9e\u73b0\u5e7f\u5ea6\u6269\u5c55，\u672c\u6587\u805a\u7126\u4e8e\u901a\u8fc7\u201c\u6df1\u5316\u5355\u4e2a\u4efb\u52a1\u7684\u63a2\u7d22\u7a0b\u5ea6\u201d\u5b9e\u73b0\u6df1\u5ea6\u6269\u5c55，\u5e76\u63d0\u51fa\u4e86\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u65f6\u6269\u5c55（MaTTS）\u5305\u542b\u5e76\u884c\u548c\u4e32\u884c\u4e24\u79cd\u8bbe\u7f6e\u3002\u8be5\u65b9\u6cd5\u901a\u8fc7\u751f\u6210\u591a\u6837\u5316\u7684\u63a2\u7d22\u8fc7\u7a0b\u63d0\u4f9b\u5bf9\u6bd4\u4fe1\u53f7，\u4f7fReasoningBank\u80fd\u591f\u5408\u6210\u66f4\u5177\u6cdb\u5316\u6027\u7684\u8bb0\u5fc6\u3002\u8fd9\u4e00\u8bbe\u8ba1\u5728\u8bb0\u5fc6\u4e0e\u6d4b\u8bd5\u65f6\u6269\u5c55\u4e4b\u95f4\u5f62\u6210\u534f\u540c\u6548\u5e94：\u9ad8\u8d28\u91cf\u8bb0\u5fc6\u5f15\u5bfc\u6269\u5c55\u63a2\u7d22\u5411\u66f4\u6709\u524d\u666f\u7684\u65b9\u5411\u63a8\u8fdb，\u800c\u751f\u6210\u7684\u4e30\u5bcc\u7ecf\u9a8c\u53c8\u80fd\u53cd\u8fc7\u6765\u5f3a\u5316\u8bb0\u5fc6\u3002\u8fd9\u79cd\u6b63\u5411\u53cd\u9988\u5faa\u73af\u8868\u660e，\u57fa\u4e8e\u8bb0\u5fc6\u7684\u7ecf\u9a8c\u6269\u5c55\u53ef\u4f5c\u4e3a\u667a\u80fd\u4f53\u7684\u65b0\u6269\u5c55\u7ef4\u5ea6\u3002<\/p>\n

\u672c\u6587\u5728\u7f51\u9875\u6d4f\u89c8（WebArena\u3001Mind2Web）\u548c\u8f6f\u4ef6\u5de5\u7a0b（SWE-Bench-Verified）\u7b49\u6311\u6218\u6027\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5f00\u5c55\u4e86\u5927\u91cf\u5b9e\u9a8c\u3002\u7ed3\u679c\u8868\u660e，\u6240\u63d0\u65b9\u6cd5\u5728\u6709\u6548\u6027（\u76f8\u5bf9\u63d0\u5347\u6700\u9ad8\u8fbe34.2%）\u548c\u6548\u7387（\u4ea4\u4e92\u6b65\u9aa4\u51cf\u5c1116.0%）\u4e0a\u5747\u4f18\u4e8e\u57fa\u7ebf\u65b9\u6cd5\u3002\u5177\u4f53\u800c\u8a00，ReasoningBank\u4e0eMaTTS\u7684\u534f\u540c\u6548\u679c\u6700\u4f73，\u662f\u5b9e\u73b0\u57fa\u4e8e\u8bb0\u5fc6\u7684\u7ecf\u9a8c\u6269\u5c55\u7684\u6838\u5fc3\u7ec4\u4ef6\u3002<\/p>\n

\n

\u672c\u6587\u7684\u8d21\u732e\u4e3b\u8981\u4f53\u73b0\u5728\u4e09\u65b9\u9762：\u00a0<\/p>\n

1. \u63d0\u51faReasoningBank\u65b0\u578b\u8bb0\u5fc6\u6846\u67b6：\u7a81\u7834\u73b0\u6709\u65b9\u6cd5\u4ec5\u5b58\u50a8\u539f\u59cb\u8f68\u8ff9\u6216\u5355\u4e00\u6210\u529f\u6d41\u7a0b\u7684\u5c40\u9650，\u4ece\u6210\u529f\u548c\u5931\u8d25\u7ecf\u9a8c\u4e2d\u63d0\u70bc\u53ef\u6cdb\u5316\u7684\u63a8\u7406\u7b56\u7565；\u00a0<\/p>\n

2. \u63d0\u51faMaTTS（\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u65f6\u6269\u5c55）：\u6784\u5efa\u8bb0\u5fc6\u4e0e\u6d4b\u8bd5\u65f6\u6269\u5c55\u4e4b\u95f4\u7684\u5f3a\u5927\u534f\u540c\u5173\u7cfb，\u786e\u7acb\u201c\u57fa\u4e8e\u8bb0\u5fc6\u7684\u7ecf\u9a8c\u6269\u5c55\u201d\u4e3a\u667a\u80fd\u4f53\u7684\u65b0\u6269\u5c55\u7ef4\u5ea6；\u00a0<\/p>\n

3. \u901a\u8fc7\u5927\u91cf\u5b9e\u9a8c\u9a8c\u8bc1：\u6240\u63d0\u65b9\u6cd5\u4e0d\u4ec5\u5728\u6709\u6548\u6027\u548c\u6548\u7387\u4e0a\u4f18\u4e8e\u73b0\u6709\u65b9\u6cd5，\u8fd8\u80fd\u4f7f\u667a\u80fd\u4f53\u4ece\u5931\u8d25\u4e2d\u5b66\u4e60，\u5e76\u968f\u65f6\u95f4\u53d1\u5c55\u51fa\u66f4\u590d\u6742\u3001\u66f4\u5177\u6d8c\u73b0\u6027\u7684\u63a8\u7406\u7b56\u7565\u3002<\/p>\n

\n

\n

\u76f8\u5173\u5de5\u4f5c<\/p>\n

LLM\u667a\u80fd\u4f53\u7684\u8bb0\u5fc6<\/p>\n

\u8bb0\u5fc6\u5df2\u6210\u4e3a\u73b0\u4ee3\u667a\u80fd\u4f53\u7cfb\u7edf\u7684\u6838\u5fc3\u6a21\u5757（Zhang et al., 2024b），\u901a\u8fc7\u5229\u7528\u8fc7\u5f80\u4fe1\u606f\u63d0\u5347\u7cfb\u7edf\u6027\u80fd（Zhang et al., 2024b）\u3002\u73b0\u6709\u8bb0\u5fc6\u7cfb\u7edf\u4ee5\u591a\u79cd\u5f62\u5f0f\u7ec4\u7ec7\u548c\u5b58\u50a8\u4fe1\u606f，\u5305\u62ec\u7eaf\u6587\u672c（Packer et al., 2023）\u3001\u6f5c\u5728\u77e5\u8bc6\u5d4c\u5165（Wang et al., 2025b）\u548c\u7ed3\u6784\u5316\u56fe（Chhikara et al., 2025; Li et al., 2025b; Xu et al., 2025）\u3002\u9664\u8bb0\u5fc6\u5185\u5bb9\u5916，\u8fd9\u4e9b\u65b9\u6cd5\u901a\u5e38\u8fd8\u6d89\u53ca\u68c0\u7d22\u673a\u5236（\u5982\u8bed\u4e49\u641c\u7d22）\u548c\u8bb0\u5fc6\u7ba1\u7406\u7b56\u7565（\u5982\u66f4\u65b0）（Hu et al., 2025a; Tan et al., 2025）\u3002\u00a0<\/p>\n

\u8fd1\u5e74\u6765，\u968f\u7740\u5f3a\u5316\u5b66\u4e60（RL）\u5728LLM\u667a\u80fd\u4f53\u4e2d\u7684\u5e94\u7528\u589e\u591a，RL\u4e5f\u88ab\u7528\u4e8e\u667a\u80fd\u4f53\u7cfb\u7edf\u7684\u8bb0\u5fc6\u7ba1\u7406（Yu et al., 2025a; Zhou et al., 2025）\u3002\u7136\u800c，\u73b0\u6709\u7814\u7a76\u5927\u591a\u5f3a\u8c03\u4e2a\u6027\u5316（Zhang et al., 2025; Zhong et al., 2024）\u548c\u957f\u4e0a\u4e0b\u6587\u7ba1\u7406（Hu et al., 2025b; Maharana et al., 2024; Wu et al., 2025），\u800c\u672c\u6587\u5c5e\u4e8e\u201c\u4ece\u8fc7\u5f80\u7ecf\u9a8c\u5b66\u4e60\u201d\u8fd9\u4e00\u7814\u7a76\u65b9\u5411（SU et al., 2025; Zhao et al., 2024）\u2014\u2014\u8fd9\u662f\u6784\u5efa\u81ea\u8fdb\u5316\u667a\u80fd\u4f53\u7cfb\u7edf\u7684\u5173\u952e\u73af\u8282（Gao et al., 2025; Liang et al., 2024）\u3002\u00a0<\/p>\n

\u4e0e\u73b0\u6709\u4ec5\u590d\u7528\u6210\u529f\u8f68\u8ff9（Tang et al., 2025a; Zheng et al., 2024）\u6216\u6d41\u7a0b\u5316\u5de5\u4f5c\u6d41（Fang et al., 2025; Liu et al., 2025b; Qian et al., 2024; Wang et al., 2025d）\u7684\u7814\u7a76\u4e0d\u540c，ReasoningBank\u5b58\u50a8\u7684\u662f\u9ad8\u5c42\u6b21\u7b56\u7565\u548c\u63a8\u7406\u63d0\u793a：\u901a\u8fc7\u5c06\u7ecf\u9a8c\u62bd\u8c61\u4e3a\u53ef\u590d\u7528\u7684\u63a8\u7406\u5355\u5143，ReasoningBank\u4e0d\u4ec5\u80fd\u8ba9\u667a\u80fd\u4f53\u4ece\u6210\u529f\u6848\u4f8b\u4e2d\u6cdb\u5316，\u8fd8\u80fd\u4ece\u5931\u8d25\u4e2d\u5b66\u4e60，\u4ece\u800c\u4e3a\u6d4b\u8bd5\u65f6\u5b66\u4e60\u63d0\u4f9b\u66f4\u4e30\u5bcc\u7684\u6307\u5bfc\u3002\u6b64\u5916，\u672c\u6587\u9996\u6b21\u63a2\u7d22\u4e86\u201c\u8bb0\u5fc6\u611f\u77e5\u7684\u6d4b\u8bd5\u65f6\u6269\u5c55\u201d，\u4f7fReasoningBank\u80fd\u4e0e\u4e30\u5bcc\u63a2\u7d22\u8f68\u8ff9\u4e2d\u7684\u591a\u6837\u5316\u4fe1\u53f7\u534f\u540c\u5de5\u4f5c\u3002<\/p>\n

\n

\u667a\u80fd\u4f53\u6d4b\u8bd5\u65f6\u6269\u5c55<\/p>\n

\u6d4b\u8bd5\u65f6\u6269\u5c55（TTS）（Snell et al., 2025）\u5df2\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u6709\u6548\u6027，\u6210\u4e3a\u7aef\u5230\u7aef\u95ee\u9898\u6c42\u89e3（\u5982\u4ee3\u7801\u751f\u6210（Li et al., 2025a; Yu et al., 2025c）\u548c\u6570\u5b66\u63a8\u7406（Muennighoff et al., 2025））\u4e2d\u7684\u5e38\u7528\u65b9\u6cd5，\u5e38\u89c1\u6280\u672f\u5305\u62ec\u201c\u6700\u4f18N\u90091（best-of-N）\u201d（Chow et al., 2025）\u3001\u201c\u675f\u641c\u7d22（beam search）\u201d（Wu et al., 2024b）\u548c\u201c\u9a8c\u8bc1\u5668\u5229\u7528\u201d（Setlur et al., 2025）\u7b49\u3002\u00a0<\/p>\n

\u7136\u800c，TTS\u5728\u591a\u8f6e\u4ea4\u4e92\u573a\u666f（\u5c24\u5176\u662f\u667a\u80fd\u4f53\u4efb\u52a1）\u4e2d\u7684\u5e94\u7528\u5c1a\u672a\u5145\u5206\u63a2\u7d22\u3002\u73b0\u6709\u7814\u7a76\u4e3b\u8981\u501f\u9274\u63a8\u7406\u4efb\u52a1\u7684\u7ecf\u9a8c（Zhu et al., 2025b），\u5e76\u4ece\u667a\u80fd\u4f53\u7cfb\u7edf\u7684\u4e0d\u540c\u7ef4\u5ea6\u8fdb\u884c\u6269\u5c55，\u4f8b\u5982\u6bcf\u4e2a\u52a8\u4f5c\u7684\u641c\u7d22\u7a7a\u95f4（Yu et al., 2025b）\u3001\u591a\u667a\u80fd\u4f53\u7cfb\u7edf\u4e2d\u7684\u667a\u80fd\u4f53\u6570\u91cf（Jin et al., 2025）\u4ee5\u53ca\u4e0e\u73af\u5883\u7684\u4ea4\u4e92\u6b21\u6570（Shen et al., 2025）\u3002\u4f46\u8fd9\u4e9b\u7814\u7a76\u5747\u672a\u8003\u8651\u667a\u80fd\u4f53\u8bb0\u5fc6\u5728\u6269\u5c55\u4e2d\u7684\u4f5c\u7528\u2014\u2014\u800c\u8bb0\u5fc6\u672c\u53ef\u5e2e\u52a9\u667a\u80fd\u4f53\u4ece\u8fc7\u5f80\u7ecf\u9a8c\u4e2d\u5b66\u4e60，\u8fdb\u800c\u6307\u5bfc\u672a\u6765\u51b3\u7b56\u3002\u00a0<\/p>\n

\u672c\u6587\u901a\u8fc7\u63d0\u51fa\u201c\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u65f6\u6269\u5c55（MaTTS）\u201d\u6269\u5c55\u4e86\u8fd9\u4e00\u7814\u7a76\u65b9\u5411\u3002\u5b9e\u9a8c\u7ed3\u679c（\u8868\u660e，\u8bb0\u5fc6\u7684\u4ef7\u503c\u8fdc\u8d85\u5355\u7eaf\u7684\u8ba1\u7b97\u6269\u5c55：\u8bb0\u5fc6\u4e0e\u6269\u5c55\u53ef\u534f\u540c\u5de5\u4f5c，\u5171\u540c\u63d0\u5347\u7cfb\u7edf\u6027\u80fd\u3002<\/p>\n

\n

ReasoningBank\u6846\u67b6<\/p>\n

\u8fc7\u5f80\u7684\u539f\u59cb\u8f68\u8ff9（\u6216\u7ecf\u9a8c）\u867d\u5b8c\u6574\u4e14\u539f\u59cb，\u4f46\u5f80\u5f80\u5197\u957f\u4e14\u5305\u542b\u566a\u58f0，\u65e0\u6cd5\u76f4\u63a5\u5e94\u7528\u4e8e\u5f53\u524d\u7528\u6237\u67e5\u8be2\u3002\u5982\u56fe2\u6240\u793a，ReasoningBank\u7684\u6838\u5fc3\u529f\u80fd\u662f\u4ece\u8fc7\u5f80\u7ecf\u9a8c\u4e2d\u63d0\u70bc\u51fa\u6709\u7528\u7684\u7b56\u7565\u548c\u63a8\u7406\u63d0\u793a，\u5c06\u5176\u8f6c\u5316\u4e3a\u7ed3\u6784\u5316\u8bb0\u5fc6\u5355\u5143\u540e\u5b58\u50a8，\u4ee5\u4f9b\u672a\u6765\u590d\u7528\u3002<\/p>\n

\"\"<\/p>\n

\u56fe2\u4e3aReasoningBank\u6846\u67b6\u6982\u8ff0：\u7ecf\u9a8c\u88ab\u63d0\u70bc\u4e3a\u5305\u542b\u201c\u6807\u9898\u3001\u63cf\u8ff0\u3001\u5185\u5bb9\u201d\u7684\u7ed3\u6784\u5316\u8bb0\u5fc6\u5355\u5143；\u9762\u5bf9\u65b0\u4efb\u52a1\u65f6，\u667a\u80fd\u4f53\u68c0\u7d22\u76f8\u5173\u5355\u5143\u4e0e\u73af\u5883\u4ea4\u4e92，\u5e76\u4ece\u6210\u529f\u548c\u5931\u8d25\u8f68\u8ff9\u4e2d\u6784\u5efa\u65b0\u8bb0\u5fc6\u5355\u5143，\u6700\u7ec8\u6574\u5408\u5230ReasoningBank\u4e2d，\u5f62\u6210\u95ed\u73af\u8bb0\u5fc6\u6d41\u7a0b\u3002\u56fe3\u4e3a\u666e\u901aTTS\u4e0eMaTTS\u7684\u5bf9\u6bd4：(a)\u666e\u901aTTS（\u65e0\u805a\u5408\u7684MaTTS）；(b)MaTTS\u5e76\u884c\u6269\u5c55（\u901a\u8fc7\u591a\u6761\u8f68\u8ff9\u7684\u81ea\u5bf9\u6bd4\u7b5b\u9009\u53ef\u9760\u8bb0\u5fc6）；(c)MaTTS\u4e32\u884c\u6269\u5c55（\u901a\u8fc7\u81ea\u4f18\u5316\u4e3a\u8bb0\u5fc6\u8865\u5145\u4e2d\u95f4\u63a8\u7406\u4fe1\u53f7）\u3002<\/p>\n

\n

\u8bb0\u5fc6\u6a21\u5f0f（Memory Schema）<\/p>\n

ReasoningBank\u4e2d\u7684\u8bb0\u5fc6\u5355\u5143\u662f\u4ece\u8fc7\u5f80\u7ecf\u9a8c\u4e2d\u8bf1\u5bfc\u751f\u6210\u7684\u7ed3\u6784\u5316\u77e5\u8bc6\u5355\u5143，\u5176\u8bbe\u8ba1\u601d\u8def\u662f\u5265\u79bb\u4f4e\u5c42\u6b21\u7684\u6267\u884c\u7ec6\u8282，\u540c\u65f6\u4fdd\u7559\u53ef\u8fc1\u79fb\u7684\u63a8\u7406\u6a21\u5f0f\u4e0e\u7b56\u7565\u3002\u6bcf\u4e2a\u8bb0\u5fc6\u5355\u5143\u5305\u542b\u4e09\u4e2a\u6838\u5fc3\u7ec4\u4ef6：\u00a0<\/p>\n

（1）\u6807\u9898（Title）：\u4f5c\u4e3a\u7b80\u6d01\u6807\u8bc6，\u6982\u62ec\u8be5\u8bb0\u5fc6\u5355\u5143\u7684\u6838\u5fc3\u7b56\u7565\u6216\u63a8\u7406\u6a21\u5f0f；\u00a0<\/p>\n

（2）\u63cf\u8ff0（Description）：\u7528\u4e00\u53e5\u8bdd\u7b80\u8981\u603b\u7ed3\u8bb0\u5fc6\u5355\u5143\u7684\u6838\u5fc3\u5185\u5bb9；\u00a0<\/p>\n

（3）\u5185\u5bb9（Content）：\u8bb0\u5f55\u4ece\u8fc7\u5f80\u7ecf\u9a8c\u4e2d\u63d0\u70bc\u7684\u63a8\u7406\u6b65\u9aa4\u3001\u51b3\u7b56\u4f9d\u636e\u6216\u64cd\u4f5c\u89c1\u89e3\u3002\u00a0<\/p>\n

\n

\u8fd9\u4e9b\u8bb0\u5fc6\u5355\u5143\u517c\u5177\u201c\u4eba\u7c7b\u53ef\u89e3\u91ca\u6027\u201d\u4e0e\u201c\u673a\u5668\u53ef\u7528\u6027\u201d，\u4fbf\u4e8e\u667a\u80fd\u4f53\u9ad8\u6548\u4f7f\u7528\u548c\u96c6\u6210\u3002<\/p>\n

\n

ReasoningBank\u4e0e\u667a\u80fd\u4f53\u7684\u96c6\u6210<\/p>\n

\u914d\u5907ReasoningBank\u7684\u667a\u80fd\u4f53\u53ef\u4ece\u4e00\u7ec4\u7ecf\u8fc7\u7b5b\u9009\u7684\u53ef\u8fc1\u79fb\u7b56\u7565\u4e2d\u83b7\u53d6\u6307\u5bfc，\u8fdb\u800c\u8f85\u52a9\u51b3\u7b56\u2014\u2014\u8fd9\u4f7f\u5f97\u667a\u80fd\u4f53\u80fd\u591f\u8c03\u7528\u6709\u6548\u7684\u8fc7\u5f80\u89c1\u89e3\u3001\u89c4\u907f\u5df2\u53d1\u73b0\u7684\u9677\u9631，\u5e76\u66f4\u7a33\u5065\u5730\u9002\u914d\u672a\u89c1\u8fc7\u7684\u67e5\u8be2\u3002\u4e8c\u8005\u7684\u96c6\u6210\u8fc7\u7a0b\u5206\u4e3a\u4e09\u4e2a\u6b65\u9aa4（\u5982\u56fe2\u6240\u793a）：\u8bb0\u5fc6\u68c0\u7d22（memory retrieval）\u3001\u8bb0\u5fc6\u6784\u5efa（memory construction）\u4e0e\u8bb0\u5fc6\u6574\u5408（memory consolidation）\u3002\u00a0<\/p>\n

1. \u8bb0\u5fc6\u68c0\u7d22：\u667a\u80fd\u4f53\u6839\u636e\u5f53\u524d\u67e5\u8be2\u4e0a\u4e0b\u6587\u67e5\u8be2ReasoningBank，\u901a\u8fc7\u57fa\u4e8e\u5d4c\u5165\u7684\u76f8\u4f3c\u5ea6\u641c\u7d22，\u7b5b\u9009\u51fa\u4e0e\u5f53\u524d\u4efb\u52a1\u6700\u76f8\u5173\u7684\u524d$k$\u6761\u7ecf\u9a8c\u53ca\u5176\u5bf9\u5e94\u7684\u8bb0\u5fc6\u5355\u5143\u3002\u68c0\u7d22\u5230\u7684\u8bb0\u5fc6\u5355\u5143\u4f1a\u88ab\u6ce8\u5165\u667a\u80fd\u4f53\u7684\u7cfb\u7edf\u6307\u4ee4，\u786e\u4fdd\u51b3\u7b56\u8fc7\u7a0b\u6709\u53ef\u7528\u7684\u8fc7\u5f80\u7ecf\u9a8c\u4f5c\u4e3a\u652f\u6491\u3002\u00a0<\/p>\n

2. \u8bb0\u5fc6\u6784\u5efa：\u5f53\u524d\u67e5\u8be2\u4efb\u52a1\u5b8c\u6210\u540e，\u9700\u901a\u8fc7\u8bb0\u5fc6\u6784\u5efa\u63d0\u53d6\u65b0\u7684\u8bb0\u5fc6\u5355\u5143\u3002\u9996\u5148，\u9700\u83b7\u53d6\u5df2\u5b8c\u6210\u8f68\u8ff9\u6b63\u786e\u6027\u7684\u201c\u4ee3\u7406\u4fe1\u53f7\u201d：\u672c\u6587\u91c7\u7528\u201cLLM\u4f5c\u4e3a\u8bc4\u5224\u8005（LLM-as-a-judge）\u201d\u7684\u65b9\u6cd5（Gu et al., 2024），\u5728\u65e0\u771f\u503c\u7684\u60c5\u51b5\u4e0b，\u6839\u636e\u67e5\u8be2\u548c\u8f68\u8ff9\u5c06\u7ed3\u679c\u6807\u6ce8\u4e3a\u201c\u6210\u529f\u201d\u6216\u201c\u5931\u8d25\u201d\u3002\u57fa\u4e8e\u8fd9\u4e9b\u4fe1\u53f7，\u91c7\u7528\u5dee\u5f02\u5316\u7684\u63d0\u53d6\u7b56\u7565：\u6210\u529f\u7ecf\u9a8c\u8d21\u732e\u7ecf\u8fc7\u9a8c\u8bc1\u7684\u7b56\u7565，\u5931\u8d25\u7ecf\u9a8c\u5219\u63d0\u4f9b\u53cd\u4e8b\u5b9e\u4fe1\u53f7\u4e0e\u9677\u9631\u63d0\u793a，\u5e2e\u52a9\u5f3a\u5316\u667a\u80fd\u4f53\u7684\u201c\u884c\u4e3a\u8fb9\u754c\u201d\u3002\u5728\u5b9e\u9645\u64cd\u4f5c\u4e2d，\u6bcf\u6761\u8f68\u8ff9（\u6216\u7ecf\u9a8c）\u53ef\u63d0\u53d6\u591a\u4e2a\u8bb0\u5fc6\u5355\u5143（\u8be6\u89c1\u9644\u5f55A.1）\u3002\u00a0<\/p>\n

3. \u8bb0\u5fc6\u6574\u5408：\u5c06\u65b0\u6784\u5efa\u7684\u8bb0\u5fc6\u5355\u5143\u901a\u8fc7\u7b80\u5355\u7684\u201c\u6dfb\u52a0\u64cd\u4f5c\u201d\u6574\u5408\u5230ReasoningBank\u4e2d，\u4f7f\u8bb0\u5fc6\u5e93\u6301\u7eed\u8fdb\u5316\u3002\u5177\u4f53\u7ec6\u8282\u89c1\u9644\u5f55A.2\u3002\u00a0<\/p>\n

\u4e0a\u8ff0\u4e09\u4e2a\u6b65\u9aa4\u5f62\u6210\u95ed\u73af：\u667a\u80fd\u4f53\u5229\u7528\u8fc7\u5f80\u7ecf\u9a8c\u6307\u5bfc\u5f53\u524d\u4efb\u52a1，\u4ece\u5f53\u524d\u4efb\u52a1\u4e2d\u6784\u5efa\u65b0\u8bb0\u5fc6，\u5e76\u6301\u7eed\u66f4\u65b0\u8bb0\u5fc6\u5e93，\u6700\u7ec8\u5b9e\u73b0\u6d4b\u8bd5\u65f6\u5b66\u4e60\u573a\u666f\u4e0b\u7684\u6301\u7eed\u8fdb\u5316\u3002（\u6ce8：\u672c\u6587\u523b\u610f\u7b80\u5316\u4e86\u8bb0\u5fc6\u4f7f\u7528\u6d41\u7a0b，\u672a\u5728\u68c0\u7d22\u6216\u6574\u5408\u73af\u8282\u5f15\u5165\u989d\u5916\u590d\u6742\u5ea6，\u4ee5\u7a81\u51faReasoningBank\u672c\u8eab\u7684\u8d21\u732e；\u672a\u6765\u53ef\u901a\u8fc7\u66f4\u590d\u6742\u7684\u6280\u672f\u8fdb\u4e00\u6b65\u4f18\u5316\u8fd9\u4e9b\u73af\u8282，\u4ee5\u83b7\u5f97\u989d\u5916\u6536\u76ca\u3002）<\/p>\n

\n

\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u65f6\u6269\u5c55（MaTTS）<\/p>\n

ReasoningBank\u4f7f\u667a\u80fd\u4f53\u80fd\u591f\u4ece\u7ecf\u9a8c\u4e2d\u5b66\u4e60，\u800c\u201c\u66f4\u591a\u7ecf\u9a8c\u201d\u610f\u5473\u7740\u201c\u66f4\u5927\u63d0\u5347\u201d\u3002\u8fd1\u5e74\u6765，\u6d4b\u8bd5\u65f6\u6269\u5c55（TTS）（Snell et al., 2025）\u5df2\u6210\u4e3a\u63d0\u5347LLM\u667a\u80fd\u4f53\u6027\u80fd\u7684\u6709\u6548\u7b56\u7565（Zhu et al., 2025a）\u2014\u2014\u901a\u8fc7\u5206\u914d\u66f4\u591a\u63a8\u7406\u9636\u6bb5\u7684\u8ba1\u7b97\u8d44\u6e90，\u53ef\u751f\u6210\u4e30\u5bcc\u7684\u63a2\u7d22\u5386\u53f2，\u5177\u5907\u5de8\u5927\u5e94\u7528\u6f5c\u529b\u3002\u00a0<\/p>\n

ReasoningBank\u4e0e\u6d4b\u8bd5\u65f6\u6269\u5c55\u7684\u76f4\u63a5\u7ed3\u5408\u65b9\u5f0f：\u5c06\u66f4\u591a\u8f68\u8ff9\u72ec\u7acb\u8f6c\u5316\u4e3a\u66f4\u591a\u8bb0\u5fc6\u5355\u5143\u3002\u4f46\u8fd9\u79cd\u666e\u901a\u5f62\u5f0f\u7684\u7ed3\u5408\u5b58\u5728\u5c40\u9650\u6027\u2014\u2014\u5b83\u672a\u5229\u7528\u201c\u540c\u4e00\u95ee\u9898\u7684\u5197\u4f59\u63a2\u7d22\u6240\u4ea7\u751f\u7684\u5185\u5728\u5bf9\u6bd4\u4fe1\u53f7\u201d，\u5bfc\u81f4\u6d4b\u8bd5\u65f6\u6269\u5c55\u7684\u6027\u80fd\u63d0\u5347\u53d7\u9650\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898，\u672c\u6587\u63d0\u51fa\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u65f6\u6269\u5c55（MaTTS），\u5c06\u6d4b\u8bd5\u65f6\u6269\u5c55\u4e0eReasoningBank\u8fdb\u884c\u6df1\u5ea6\u96c6\u6210\u3002\u4e0e\u666e\u901a\u65b9\u6cd5\u4e0d\u540c，MaTTS\u4f1a\u523b\u610f\u4ece\u6269\u5c55\u8fc7\u7a0b\u4e2d\u751f\u6210\u7684\u5927\u91cf\u6210\u529f\u4e0e\u5931\u8d25\u8f68\u8ff9\u4e2d\u5b66\u4e60，\u4ee5\u5b9e\u73b0\u66f4\u6709\u6548\u7684\u8bb0\u5fc6\u7b5b\u9009\u3002\u672c\u6587\u4e3aMaTTS\u8bbe\u8ba1\u4e86\u4e24\u79cd\u4e92\u8865\u7684\u5b9e\u73b0\u65b9\u5f0f：\u5e76\u884c\u6269\u5c55（parallel scaling）\u4e0e\u4e32\u884c\u6269\u5c55（sequential scaling），\u5177\u4f53\u5b9e\u73b0\u7ec6\u8282\u89c1\u9644\u5f55A.3\u3002<\/p>\n

\n

\u5e76\u884c\u6269\u5c55（Parallel Scaling）<\/p>\n

\u5728\u5e76\u884c\u8bbe\u7f6e\u4e0b，\u667a\u80fd\u4f53\u5728\u68c0\u7d22\u5230\u7684\u8bb0\u5fc6\u5355\u5143\u6307\u5bfc\u4e0b，\u4e3a\u540c\u4e00\u67e5\u8be2\u751f\u6210\u591a\u6761\u8f68\u8ff9\u3002\u901a\u8fc7\u5bf9\u4e0d\u540c\u8f68\u8ff9\u8fdb\u884c\u201c\u81ea\u5bf9\u6bd4（self-contrast）\u201d（Chen et al., 2020），\u667a\u80fd\u4f53\u53ef\u8bc6\u522b\u51fa\u5177\u6709\u4e00\u81f4\u6027\u7684\u63a8\u7406\u6a21\u5f0f，\u540c\u65f6\u8fc7\u6ee4\u6389\u865a\u5047\u89e3\u51b3\u65b9\u6848\u3002\u8fd9\u4e00\u8fc7\u7a0b\u80fd\u4ece\u5355\u4e00\u67e5\u8be2\u7684\u591a\u6b21\u5c1d\u8bd5\u4e2d\u7b5b\u9009\u51fa\u66f4\u53ef\u9760\u7684\u8bb0\u5fc6，\u8fdb\u800c\u4fc3\u8fdb\u591a\u6837\u5316\u63a2\u7d22\u3002<\/p>\n

\n

\u4e32\u884c\u6269\u5c55（Sequential Scaling）<\/p>\n

\u4e32\u884c\u6269\u5c55\u9075\u5faa\u201c\u81ea\u4f18\u5316（self-refinement）\u201d\u539f\u5219（Madaan et al., 2023），\u5728\u5355\u6761\u8f68\u8ff9\u521d\u59cb\u5b8c\u6210\u540e，\u5bf9\u5176\u63a8\u7406\u8fc7\u7a0b\u8fdb\u884c\u8fed\u4ee3\u4f18\u5316\u3002\u5728\u6b64\u8fc7\u7a0b\u4e2d，\u81ea\u4f18\u5316\u751f\u6210\u7684\u4e2d\u95f4\u8bb0\u5f55\u4e5f\u4f1a\u88ab\u89c6\u4e3a\u6709\u4ef7\u503c\u7684\u8bb0\u5fc6\u4fe1\u53f7\u2014\u2014\u8fd9\u4e9b\u8bb0\u5f55\u5305\u542b\u4e86\u63a8\u7406\u5c1d\u8bd5\u3001\u4fee\u6b63\u8fc7\u7a0b\u4e0e\u89c1\u89e3，\u800c\u8fd9\u4e9b\u4fe1\u606f\u53ef\u80fd\u4e0d\u4f1a\u51fa\u73b0\u5728\u6700\u7ec8\u89e3\u51b3\u65b9\u6848\u4e2d\u3002<\/p>\n

\n

\u672c\u6587\u5b9a\u4e49\u201c\u6269\u5c55\u56e0\u5b50k\u201d：\u5bf9\u4e8e\u5e76\u884c\u6269\u5c55，k\u8868\u793a\u8f68\u8ff9\u6570\u91cf；\u5bf9\u4e8e\u4e32\u884c\u6269\u5c55，k\u8868\u793a\u4f18\u5316\u6b65\u6570\u3002\u914d\u5907ReasoningBank\u540e，\u5e76\u884c\u4e0e\u4e32\u884c\u4e24\u79cd\u7b56\u7565\u5747\u5177\u5907\u201c\u8bb0\u5fc6\u611f\u77e5\u80fd\u529b\u201d，\u786e\u4fdd\u6d4b\u8bd5\u65f6\u5206\u914d\u7684\u989d\u5916\u8ba1\u7b97\u8d44\u6e90\u80fd\u8f6c\u5316\u4e3a\u66f4\u5177\u8fc1\u79fb\u6027\u3001\u66f4\u9ad8\u8d28\u91cf\u7684\u8bb0\u5fc6，\u4ee5\u670d\u52a1\u4e8e\u672a\u6765\u4efb\u52a1\u3002<\/p>\n

\n

\u4e3a\u9a8c\u8bc1\u65b9\u6cd5\u6709\u6548\u6027，\u672c\u6587\u8bbe\u7f6e\u4e09\u7c7b\u57fa\u7ebf\u6a21\u578b，\u8986\u76d6\u201c\u65e0\u8bb0\u5fc6-\u8f68\u8ff9\u8bb0\u5fc6-\u5de5\u4f5c\u6d41\u8bb0\u5fc6\u201d\u7684\u5b8c\u6574\u68af\u5ea6：\u00a0<\/p>\n

1. \u65e0\u8bb0\u5fc6（No Memory）：\u65e0\u4efb\u4f55\u8bb0\u5fc6\u6a21\u5757\u7684\u57fa\u7840LLM\u667a\u80fd\u4f53；\u00a0<\/p>\n

2. Synapse（\u8f68\u8ff9\u8bb0\u5fc6）（Zheng et al., 2024）：\u5c06\u8fc7\u5f80\u8f68\u8ff9\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u7684\u4ee3\u8868\u6027\u65b9\u6cd5；\u00a0<\/p>\n

3. AWM（\u5de5\u4f5c\u6d41\u8bb0\u5fc6）（Wang et al., 2025d）：\u4ece\u8f68\u8ff9\u4e2d\u62bd\u8c61\u901a\u7528\u6a21\u5f0f\u5e76\u5f62\u6210\u53ef\u590d\u7528\u5de5\u4f5c\u6d41\u7684\u65b9\u6cd5\u3002\u00a0<\/p>\n

\n

\u5b9e\u9a8c\u5b9e\u73b0\u7ec6\u8282\u5982\u4e0b：\u00a0<\/p>\n

– LLM\u9aa8\u5e72\u6a21\u578b：\u91c7\u7528Gemini-2.5\u7cfb\u5217（Comanici et al., 2025）\u548cClaude-3.7（Anthropic, 2025），\u6db5\u76d6\u8de8\u6a21\u578b\u5bb6\u65cf（Gemini\u3001Claude）\u4e0e\u540c\u5bb6\u65cf\u4e0d\u540c\u80fd\u529b\u7b49\u7ea7（Gemini-2.5-Flash\u3001Gemini-2.5-Pro）\u7684\u5bf9\u6bd4；\u00a0<\/p>\n

– \u8fd0\u884c\u73af\u5883：\u7f51\u9875\u6d4f\u89c8\u4efb\u52a1\u4f7f\u7528BrowserGym（de Chezelles et al., 2025），\u8f6f\u4ef6\u5de5\u7a0b\u4efb\u52a1\u4f7f\u7528\u4ec5\u652f\u6301bash\u547d\u4ee4\u7684\u73af\u5883；\u00a0<\/p>\n

– \u667a\u80fd\u4f53\u67b6\u6784：\u91c7\u7528ReAct\u98ce\u683c（Yao et al., 2023），\u9ed8\u8ba4\u89e3\u7801\u914d\u7f6e，WebArena\u4efb\u52a1\u5355\u67e5\u8be2\u6700\u5927\u6b65\u6570\u9650\u5236\u4e3a30\u6b65\u3002\u00a0<\/p>\n

\n

\u8bc4\u4f30\u7ef4\u5ea6\u805a\u7126\u6709\u6548\u6027\u4e0e\u6548\u7387：\u00a0<\/p>\n

– \u6709\u6548\u6027：\u4ee5\u201c\u6210\u529f\u7387（SR）\u201d\u8861\u91cf，\u5373\u667a\u80fd\u4f53\u6210\u529f\u89e3\u51b3\u7528\u6237\u67e5\u8be2\u7684\u6bd4\u4f8b，\u91c7\u7528LLM\u6a21\u7cca\u5339\u914d\u4e0e\u7cbe\u786e\u5b57\u7b26\u4e32\u5339\u914d\u7ed3\u5408\u7684\u65b9\u5f0f\u9a8c\u8bc1\u7b54\u6848\u6b63\u786e\u6027；\u00a0<\/p>\n

– \u6548\u7387：\u4ee5\u201c\u5e73\u5747\u4ea4\u4e92\u6b65\u6570（Step）\u201d\u8861\u91cf，\u53cd\u6620\u4efb\u52a1\u5b8c\u6210\u8fc7\u7a0b\u4e2d\u7684\u8ba1\u7b97\u4e0e\u4ea4\u4e92\u6210\u672c\u3002\u4e0d\u540c\u57fa\u51c6\u6d4b\u8bd5\u7684\u5177\u4f53\u6307\u6807\u4f1a\u6839\u636e\u4efb\u52a1\u7279\u6027\u8c03\u6574（\u8be6\u89c1\u9644\u5f55B）\u3002<\/p>\n

\n

ReasoningBank\u7684\u6d8c\u73b0\u884c\u4e3a<\/p>\n

\u7814\u7a76\u53d1\u73b0，ReasoningBank\u4e2d\u7684\u7b56\u7565\u5e76\u975e\u5355\u4e00\u56fa\u5b9a，\u800c\u662f\u4f1a\u968f\u65f6\u95f4\u8fdb\u5316，\u5448\u73b0\u51fa\u7c7b\u4f3c\u5f3a\u5316\u5b66\u4e60（RL）\u5b66\u4e60\u52a8\u6001\u7684\u6d8c\u73b0\u884c\u4e3a（Wang et al., 2025a）\u3002\u5982\u56fe6\u6240\u793a，ReasoningBank\u4e2d\u7684\u67d0\u4e00\u8bb0\u5fc6\u5355\u5143\u5728\u6d4b\u8bd5\u65f6\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u4f1a\u9010\u6b65\u6f14\u5316，\u5177\u4f53\u9636\u6bb5\u5982\u4e0b：\u00a0<\/p>\n

1. \u6d41\u7a0b\u5316\/\u6267\u884c\u578b\u7b56\u7565：\u521d\u59cb\u9636\u6bb5，\u7b56\u7565\u4ee5\u6267\u884c\u4e3a\u5bfc\u5411（\u5982\u201c\u5bfb\u627e\u5bfc\u822a\u94fe\u63a5\u201d），\u667a\u80fd\u4f53\u4ec5\u9075\u5faa\u7b80\u5355\u7684\u52a8\u4f5c\u89c4\u5219；\u00a0<\/p>\n

2. \u539f\u5b50\u5316\u81ea\u6211\u53cd\u601d：\u968f\u7740\u7ecf\u9a8c\u79ef\u7d2f，\u7b56\u7565\u8fdb\u5316\u4e3a\u9002\u5e94\u6027\u53cd\u601d（\u5982\u201c\u91cd\u65b0\u9a8c\u8bc1\u5143\u7d20\u6807\u8bc6\u4ee5\u51cf\u5c11\u4f4e\u7ea7\u9519\u8bef\u201d）；\u00a0<\/p>\n

3. \u9002\u5e94\u6027\u68c0\u67e5：\u8fdb\u4e00\u6b65\u79ef\u7d2f\u7ecf\u9a8c\u540e，\u7b56\u7565\u53d1\u5c55\u4e3a\u7cfb\u7edf\u6027\u68c0\u67e5（\u5982\u201c\u5229\u7528\u641c\u7d22\u6216\u7b5b\u9009\u529f\u80fd\u786e\u4fdd\u7ed3\u679c\u5b8c\u6574\u6027\u540e\u518d\u53cd\u9988\u201d）；\u00a0<\/p>\n

4. \u7ec4\u5408\u5f0f\u7b56\u7565：\u6700\u7ec8\u6210\u719f\u4e3a\u7ec4\u5408\u5f0f\u63a8\u7406（\u5982\u201c\u4ea4\u53c9\u9a8c\u8bc1\u4efb\u52a1\u9700\u6c42，\u5e76\u5c06\u4f4e\u7ea7\u52a8\u4f5c\u91cd\u65b0\u6620\u5c04\u4e3a\u9ad8\u7ea7\u63a8\u7406\u201d）\u3002\u00a0<\/p>\n

\u8fd9\u4e00\u6f14\u5316\u8fc7\u7a0b\u8868\u660e，ReasoningBank\u80fd\u8ba9\u667a\u80fd\u4f53\u4ece\u201c\u673a\u68b0\u6267\u884c\u201d\u9010\u6b65\u5347\u7ea7\u4e3a\u201c\u7075\u6d3b\u63a8\u7406\u201d，\u5b9e\u73b0\u7b56\u7565\u7684\u6301\u7eed\u4f18\u5316\u3002<\/p>\n

\"\"<\/p>\n

\u56fe6\u4e3aReasoningBank\u6d8c\u73b0\u884c\u4e3a\u7684\u6848\u4f8b\u5206\u6790，\u5c55\u793a\u4e86\u67d0\u4e00\u8bb0\u5fc6\u5355\u5143\u5728\u6d4b\u8bd5\u65f6\u5b66\u4e60\u65f6\u95f4\u7ebf\u4e0a\u4ece\u201c\u6d41\u7a0b\u5316\u7b56\u7565\u201d\u5230\u201c\u7ec4\u5408\u5f0f\u7b56\u7565\u201d\u7684\u6f14\u5316\u8fc7\u7a0b\u3002<\/p>\n

\n

\u672a\u6765\u7814\u7a76\u65b9\u5411<\/p>\n

1 \u7ec4\u5408\u5f0f\u8bb0\u5fc6（Compositional Memory）<\/p>\n

\u5f53\u524d\u6846\u67b6\u5c06\u6bcf\u6761\u7ecf\u9a8c\u63d0\u70bc\u4e3a\u591a\u4e2a\u72ec\u7acb\u7684\u8bb0\u5fc6\u5355\u5143，\u9762\u5bf9\u65b0\u67e5\u8be2\u65f6，\u4ec5\u901a\u8fc7\u76f8\u4f3c\u5ea6\u68c0\u7d22\u5e76\u72ec\u7acb\u590d\u7528\u8fd9\u4e9b\u5355\u5143\u3002\u8fd9\u79cd\u8bbe\u8ba1\u867d\u80fd\u7a81\u51fa\u8bb0\u5fc6\u5185\u5bb9\u7684\u4ef7\u503c，\u4f46\u672a\u8003\u8651\u5355\u5143\u95f4\u7684\u7ec4\u5408\u6f5c\u529b\u3002\u672a\u6765\u53ef\u63a2\u7d22\u201c\u7ec4\u5408\u611f\u77e5\u7684\u68c0\u7d22\u4e0e\u6574\u5408\u201d\u2014\u2014\u8ba9\u667a\u80fd\u4f53\u5c06\u4e92\u8865\u7684\u8bb0\u5fc6\u5355\u5143\u7ec4\u5408\u6210\u66f4\u9ad8\u5c42\u6b21\u7684\u7b56\u7565，\u6216\u5f62\u6210\u53ef\u590d\u7528\u7684\u201c\u7b56\u7565\u5b8f（macro）\u201d，\u4ece\u800c\u5728\u957f\u5468\u671f\u4efb\u52a1\u4e2d\u5b9e\u73b0\u66f4\u5f3a\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n

\n

2 \u9ad8\u7ea7\u8bb0\u5fc6\u67b6\u6784（Advanced Memory Architectures）<\/p>\n

\u5f53\u524d\u7cfb\u7edf\u8bbe\u8ba1\u523b\u610f\u4fdd\u6301\u7b80\u6d01，\u4ee5\u7a81\u51fa\u6838\u5fc3\u8d21\u732e\u3002\u672a\u6765\u53ef\u6784\u5efa\u5206\u5c42\u5f0f\u3001\u4ea7\u54c1\u7ea7\u8bb0\u5fc6\u6808，\u6574\u5408\u73b0\u6709\u6210\u719f\u8bb0\u5fc6\u8303\u5f0f：\u00a0<\/p>\n

– \u60c5\u666f\u8f68\u8ff9（episodic traces）（Fountas et al., 2025）：\u7528\u4e8e\u5b58\u50a8\u5355\u4efb\u52a1\u4e0a\u4e0b\u6587；\u00a0<\/p>\n

– \u77ed\u671f\u201c\u5de5\u4f5c\u8bb0\u5fc6\u201d（working memory）（Lumer et al., 2025）：\u7528\u4e8e\u7ef4\u62a4\u4f1a\u8bdd\u5185\u72b6\u6001；\u00a0<\/p>\n

– \u957f\u671f\u5de9\u56fa\u8bb0\u5fc6（long-term consolidated memory）（Wang et al., 2025b）：\u7ed3\u5408\u201c\u9057\u5fd8\/\u5237\u65b0\u7b56\u7565\u201d\u5b58\u50a8\u957f\u671f\u77e5\u8bc6\u3002\u00a0<\/p>\n

\n

ReasoningBank\u7684\u8bbe\u8ba1\u7406\u5ff5\u4e0e\u4e0a\u8ff0\u8bb0\u5fc6\u7ef4\u5ea6\u517c\u5bb9\u3002\u6b64\u5916，\u672a\u6765\u8fd8\u53ef\u5c06\u201c\u57fa\u4e8e\u5d4c\u5165\u7684\u76f8\u4f3c\u5ea6\u68c0\u7d22\u201d\u5347\u7ea7\u4e3a\u201c\u63a8\u7406\u5bc6\u96c6\u578b\u63a7\u5236\u5668\u201d（Shao et al., 2025）\u2014\u2014\u901a\u8fc7\u62c6\u89e3\u67e5\u8be2\u3001\u89c4\u5212\u591a\u8df3\u68c0\u7d22\u3001\u7ed3\u5408\u4e0d\u786e\u5b9a\u6027\/\u65f6\u6548\u6027\/\u6210\u672c\u7b49\u56e0\u7d20\u9009\u62e9\u8bb0\u5fc6，\u751a\u81f3\u5f15\u5165\u57fa\u4e8e\u5b66\u4e60\u7684\u8def\u7531\u4e0e\u6574\u5408\u7b56\u7565\u5b9e\u73b0\u81ea\u52a8\u5316\u3002\u8fd9\u79cd\u6574\u5408\u53ef\u5c06\u201cReasoningBank+MaTTS\u201d\u8f6c\u5316\u4e3a\u53ef\u90e8\u7f72\u7684\u8bb0\u5fc6\u670d\u52a1，\u5b9e\u73b0\u8de8\u9886\u57df\u3001\u8de8\u56e2\u961f\u7684\u89c4\u6a21\u5316\u5e94\u7528\u3002<\/p>\n

\n

\u5c40\u9650\u6027<\/p>\n

\u5c3d\u7ba1ReasoningBank\u5728\u5b9e\u9a8c\u4e2d\u8868\u73b0\u51fa\u4f18\u5f02\u6027\u80fd，\u5e76\u63d0\u51fa\u4e86\u201c\u8bb0\u5fc6\u4f5c\u4e3a\u6269\u5c55\u7ef4\u5ea6\u201d\u7684\u5b9e\u7528\u8303\u5f0f，\u4f46\u4ecd\u5b58\u5728\u4ee5\u4e0b\u5c40\u9650\u6027，\u4e3a\u672a\u6765\u7814\u7a76\u6307\u660e\u65b9\u5411：<\/p>\n

1 \u805a\u7126\u8bb0\u5fc6\u5185\u5bb9\u800c\u975e\u67b6\u6784<\/p>\n

\u672c\u6587\u7684\u6838\u5fc3\u8d21\u732e\u5728\u4e8e\u201c\u8bb0\u5fc6\u5185\u5bb9\u7684\u7b5b\u9009\u4e0e\u5229\u7528\u201d（\u5982\u6574\u5408\u5931\u8d25\u8f68\u8ff9\u3001\u63d0\u70bc\u63a8\u7406\u63d0\u793a），\u800c\u975e\u8bb0\u5fc6\u67b6\u6784\u7684\u521b\u65b0\u3002\u56e0\u6b64，\u672a\u4e0e\u201c\u60c5\u666f\u8bb0\u5fc6（episodic memory）\u201d\u201c\u5206\u5c42\u8bb0\u5fc6（hierarchical memory）\u201d\u7b49\u67b6\u6784\u8fdb\u884c\u6df1\u5165\u5bf9\u6bd4\u2014\u2014\u8fd9\u4e9b\u67b6\u6784\u89e3\u51b3\u7684\u662f\u201c\u8bb0\u5fc6\u5f62\u5f0f\/\u7ed3\u6784\u201d\u7684\u95ee\u9898，\u4e0e\u672c\u6587\u201c\u8bb0\u5fc6\u5185\u5bb9\u201d\u7684\u7814\u7a76\u65b9\u5411\u6b63\u4ea4\u3002\u672a\u6765\u53ef\u63a2\u7d22\u5c06ReasoningBank\u4e0e\u8fd9\u4e9b\u67b6\u6784\u7ed3\u5408，\u8fdb\u4e00\u6b65\u63d0\u5347\u6027\u80fd\u3002<\/p>\n

2 \u8bb0\u5fc6\u68c0\u7d22\u4e0e\u6574\u5408\u7b56\u7565\u8f83\u7b80\u5355<\/p>\n

\u4e3a\u66f4\u6e05\u6670\u5730\u5206\u79bb\u201c\u8bb0\u5fc6\u5185\u5bb9\u8d28\u91cf\u201d\u7684\u5f71\u54cd，\u672c\u6587\u523b\u610f\u91c7\u7528\u201c\u57fa\u4e8e\u5d4c\u5165\u7684\u7b80\u5355\u68c0\u7d22\u201d\u4e0e\u201c\u76f4\u63a5\u8ffd\u52a0\u7684\u6574\u5408\u7b56\u7565\u201d\u3002\u5c3d\u7ba1\u8fd9\u79cd\u8bbe\u8ba1\u80fd\u786e\u4fdd\u6027\u80fd\u589e\u76ca\u53ef\u5f52\u56e0\u4e8e\u8bb0\u5fc6\u5185\u5bb9，\u4f46\u66f4\u590d\u6742\u7684\u7b56\u7565（\u5982\u81ea\u9002\u5e94\u68c0\u7d22\u3001\u5206\u5c42\u6574\u5408）\u4e0eReasoningBank\u6846\u67b6\u517c\u5bb9，\u4e14\u53ef\u80fd\u8fdb\u4e00\u6b65\u63d0\u5347\u6027\u80fd\u2014\u2014\u8fd9\u4e9b\u7b56\u7565\u5e76\u975e\u672c\u6587\u91cd\u70b9，\u9700\u672a\u6765\u5355\u72ec\u63a2\u7d22\u3002<\/p>\n

3 \u4f9d\u8d56LLM-as-a-judge\u83b7\u53d6\u6b63\u786e\u6027\u4fe1\u53f7<\/p>\n

\u5f53\u524d\u5b9e\u73b0\u4e2d，\u8f68\u8ff9\u7684\u201c\u6210\u529f\/\u5931\u8d25\u201d\u4fe1\u53f7\u7531\u201cLLM\u4f5c\u4e3a\u8bc4\u5224\u8005（LLM-as-a-judge）\u201d\u5224\u5b9a\u3002\u8fd9\u79cd\u81ea\u52a8\u6807\u6ce8\u65b9\u6cd5\u867d\u80fd\u5728\u65e0\u771f\u503c\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u89c4\u6a21\u5316\u8bc4\u4f30，\u4f46\u5728\u4efb\u52a1\u6a21\u7cca\u6216\u8bc4\u5224\u6a21\u578b\u81ea\u8eab\u5b58\u5728\u504f\u5dee\u65f6，\u53ef\u80fd\u5f15\u5165\u566a\u58f0\u3002\u5c3d\u7ba1\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\u6846\u67b6\u5bf9\u8fd9\u7c7b\u566a\u58f0\u5177\u6709\u4e00\u5b9a\u9c81\u68d2\u6027，\u4f46\u672a\u6765\u53ef\u901a\u8fc7\u5f15\u5165\u66f4\u5f3a\u7684\u9a8c\u8bc1\u5668\u3001\u4eba\u673a\u534f\u540c\u53cd\u9988\u6216\u96c6\u6210\u8bc4\u5224\u7b49\u65b9\u5f0f，\u63d0\u5347\u8bb0\u5fc6\u8bf1\u5bfc\u7684\u53ef\u9760\u6027\u3002<\/p>\n

\n

\u505a\u4e2a\u603b\u7ed3<\/p>\n

\u672c\u6587\u63d0\u51faReasoningBank\u2014\u2014\u4e00\u79cd\u4ece\u6210\u529f\u4e0e\u5931\u8d25\u7ecf\u9a8c\u4e2d\u63d0\u70bc\u7b56\u7565\u7ea7\u63a8\u7406\u4fe1\u53f7\u7684\u8bb0\u5fc6\u6846\u67b6，\u5e76\u5c06\u5176\u4e0e\u6d4b\u8bd5\u65f6\u6269\u5c55\u7ed3\u5408，\u6784\u5efa\u4e86\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u65f6\u6269\u5c55（MaTTS）\u3002\u5927\u91cf\u5b9e\u9a8c\u8868\u660e：\u00a0<\/p>\n

1. ReasoningBank\u5728\u63d0\u5347\u6027\u80fd\u7684\u540c\u65f6，\u80fd\u51cf\u5c11\u5197\u4f59\u63a2\u7d22，\u5b9e\u73b0\u201c\u6709\u6548\u6027-\u6548\u7387\u201d\u53cc\u4f18；\u00a0<\/p>\n

2. \u8bb0\u5fc6\u4e0e\u6d4b\u8bd5\u65f6\u6269\u5c55\u5b58\u5728\u5f3a\u534f\u540c\u4f5c\u7528：ReasoningBank\u5f15\u5bfc\u6269\u5c55\u5411\u66f4\u6709\u524d\u666f\u7684\u8f68\u8ff9\u63a8\u8fdb，\u800c\u591a\u6837\u5316\u8f68\u8ff9\u5219\u4e3a\u8bb0\u5fc6\u63d0\u4f9b\u5b9d\u8d35\u7684\u5bf9\u6bd4\u4fe1\u53f7，\u5f62\u6210\u6b63\u5411\u5faa\u73af；\u00a0<\/p>\n

3. \u7ec4\u4ef6\u5206\u6790\u4e0e\u6d8c\u73b0\u884c\u4e3a\u7814\u7a76\u9a8c\u8bc1\u4e86\u201c\u5931\u8d25\u7ecf\u9a8c\u6574\u5408\u201d\u201c\u7b56\u7565\u6f14\u5316\u201d\u7684\u4ef7\u503c\u3002\u00a0<\/p>\n

\u672c\u6587\u7684\u53d1\u73b0\u4e3a\u6784\u5efa\u201c\u81ea\u9002\u5e94\u3001\u7ec8\u8eab\u5b66\u4e60\u578b\u667a\u80fd\u4f53\u201d\u63d0\u4f9b\u4e86\u5207\u5b9e\u53ef\u884c\u7684\u8def\u5f84，\u672a\u6765\u53ef\u8fdb\u4e00\u6b65\u63a2\u7d22\u66f4\u590d\u6742\u7684\u8bb0\u5fc6\u67b6\u6784\u4e0e\u6269\u5c55\u7b56\u7565\u3002<\/p>\n

\n","protected":false},"excerpt":{"rendered":"

\u4eca\u5929\u5206\u4eab\u7684\u8bba\u6587\u662f\u300aReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u300b
\n\u539f\u6587\u94fe\u63a5：https:\/\/arxiv.org\/abs\/2509.25140 \u8fd9\u662f\u4e00\u7bc7\u8c37\u6b4c\u7684\u8bba\u6587，\u5173\u4e8e\u81ea\u6f14\u5316\u667a\u80fd\u4f53\u7684\u3002 \u968f\u7740\u5927\u578b\u8bed\u8a00\u6a21\u578b（LLM）\u667a\u80fd\u4f53\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u6301\u7eed\u53d1\u6325\u4f5c\u7528\u7684\u573a\u666f\u65e5\u76ca\u589e\u591a，\u5b83\u4eec\u81ea\u7136\u4f1a\u9762\u4e34\u6e90\u6e90\u4e0d\u65ad\u7684\u4efb\u52a1\u6d41\u3002\u7136\u800c，\u5176\u5173\u952e\u5c40\u9650\u5728\u4e8e\u65e0<\/p>\n","protected":false},"author":2,"featured_media":103318,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[10848,11775,74,75,3546],"topic":[],"class_list":["post-103321","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-harness","tag-11775","tag-agent","tag-llm","tag-3546"],"yoast_head":"\nReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/103321.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4eca\u5929\u5206\u4eab\u7684\u8bba\u6587\u662f\u300aReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u300b \u539f\u6587\u94fe\u63a5：https:\/\/arxiv.org\/abs\/2509.25140 \u8fd9\u662f\u4e00\u7bc7\u8c37\u6b4c\u7684\u8bba\u6587，\u5173\u4e8e\u81ea\u6f14\u5316\u667a\u80fd\u4f53\u7684\u3002 \u968f\u7740\u5927\u578b\u8bed\u8a00\u6a21\u578b（LLM）\u667a\u80fd\u4f53\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u6301\u7eed\u53d1\u6325\u4f5c\u7528\u7684\u573a\u666f\u65e5\u76ca\u589e\u591a，\u5b83\u4eec\u81ea\u7136\u4f1a\u9762\u4e34\u6e90\u6e90\u4e0d\u65ad\u7684\u4efb\u52a1\u6d41\u3002\u7136\u800c，\u5176\u5173\u952e\u5c40\u9650\u5728\u4e8e\u65e0\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/103321.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-10T01:09:02+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260910010901-6aa2032d0d8f2.jpg\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/103321.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/103321.html\",\"name\":\"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-10T01:09:02+00:00\",\"dateModified\":\"2026-09-10T01:09:02+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/103321.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/103321.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/103321.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/103321.html","og_locale":"zh_CN","og_type":"article","og_title":"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4eca\u5929\u5206\u4eab\u7684\u8bba\u6587\u662f\u300aReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u300b \u539f\u6587\u94fe\u63a5：https:\/\/arxiv.org\/abs\/2509.25140 \u8fd9\u662f\u4e00\u7bc7\u8c37\u6b4c\u7684\u8bba\u6587，\u5173\u4e8e\u81ea\u6f14\u5316\u667a\u80fd\u4f53\u7684\u3002 \u968f\u7740\u5927\u578b\u8bed\u8a00\u6a21\u578b（LLM）\u667a\u80fd\u4f53\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u6301\u7eed\u53d1\u6325\u4f5c\u7528\u7684\u573a\u666f\u65e5\u76ca\u589e\u591a，\u5b83\u4eec\u81ea\u7136\u4f1a\u9762\u4e34\u6e90\u6e90\u4e0d\u65ad\u7684\u4efb\u52a1\u6d41\u3002\u7136\u800c，\u5176\u5173\u952e\u5c40\u9650\u5728\u4e8e\u65e0","og_url":"https:\/\/www.wsisp.com\/helps\/103321.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-10T01:09:02+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260910010901-6aa2032d0d8f2.jpg"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/103321.html","url":"https:\/\/www.wsisp.com\/helps\/103321.html","name":"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-10T01:09:02+00:00","dateModified":"2026-09-10T01:09:02+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/103321.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/103321.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/103321.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory\u8bba\u6587\u9605\u8bfb"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/103321","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=103321"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/103321\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/103318"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=103321"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=103321"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=103321"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=103321"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}