{"id":6322,"date":"2025-10-13T13:36:33","date_gmt":"2025-10-13T05:36:33","guid":{"rendered":"https:\/\/nullthought.net\/?p=6322"},"modified":"2025-10-13T13:36:34","modified_gmt":"2025-10-13T05:36:34","slug":"reasoningbank%ef%bc%9a%e6%9e%84%e5%bb%ba%e9%9d%a2%e5%90%91%e6%8e%a8%e7%90%86%e7%9a%84%e8%ae%b0%e5%bf%86%e6%9c%ba%e5%88%b6%ef%bc%8c%e4%bd%bf%e6%99%ba%e8%83%bd%e4%bd%93%e5%ae%9e","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6322","title":{"rendered":"ReasoningBank\uff1a\u6784\u5efa\u201c\u9762\u5411\u63a8\u7406\u7684\u8bb0\u5fc6\u201d\u673a\u5236\uff0c\u4f7f\u667a\u80fd\u4f53\u5b9e\u73b0\u81ea\u4e3b\u8fdb\u5316\u80fd\u529b"},"content":{"rendered":"\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u9a71\u52a8\u7684\u667a\u80fd\u4f53\u6b63\u88ab\u7528\u4e8e\u957f\u671f\u3001\u6301\u7eed\u7684\u771f\u5b9e\u4e16\u754c\u4efb\u52a1\uff08\u5982\u7f51\u9875\u6d4f\u89c8\u3001\u8f6f\u4ef6\u5de5\u7a0b\u81ea\u52a8\u5316\uff09\uff0c\u4f46\u4e3b\u6d41\u667a\u80fd\u4f53\u5728\u201c\u8de8\u4efb\u52a1\u5b66\u4e60\u201d\u4e0a\u5b58\u5728\u660e\u663e\u77ed\u677f\uff1a\u5b83\u4eec\u5f80\u5f80\u628a\u6bcf\u4e2a\u4efb\u52a1\u5f53\u4f5c\u5b64\u7acb\u4e8b\u4ef6\u5904\u7406\uff0c\u65e2\u65e0\u6cd5\u4ece\u5386\u53f2\u4ea4\u4e92\u4e2d\u6c72\u53d6\u53ef\u8fc1\u79fb\u7684\u63a8\u7406\u7b56\u7565\uff0c\u4e5f\u96be\u4ee5\u907f\u514d\u91cd\u590d\u72af\u9519\u3002\u8fd9\u79cd\u201c\u65e0\u8bb0\u5fc6\u6216\u5f31\u8bb0\u5fc6\u201d\u7684\u6a21\u5f0f\u4e0d\u4ec5\u6d6a\u8d39\u4e86\u5b9d\u8d35\u7684\u7ecf\u9a8c\uff0c\u8fd8\u4f7f\u667a\u80fd\u4f53\u7f3a\u4e4f\u968f\u65f6\u95f4\u81ea\u6211\u8fdb\u5316\u7684\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2509.25140\" target=\"_blank\" rel=\"noreferrer noopener\">ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory<\/a><\/strong>\u63d0\u51fa\u8981\u6784\u5efa\u4e00\u79cd\u201c\u9762\u5411\u63a8\u7406\u7684\u8bb0\u5fc6\u201d\u673a\u5236\uff0c\u4f7f\u667a\u80fd\u4f53\u80fd\u591f\u5728\u6d4b\u8bd5\u671f\u8fde\u7eed\u4efb\u52a1\u6d41\u4e2d\uff0c\u4e3b\u52a8\u4ece\u6210\u529f\u4e0e\u5931\u8d25\u7684\u7ecf\u5386\u91cc\u63d0\u70bc\u901a\u7528\u5316\u7684\u7b56\u7565\uff0c\u5e76\u5728\u540e\u7eed\u4efb\u52a1\u4e2d\u6709\u6548\u8c03\u7528\u4e0e\u8fed\u4ee3\uff0c\u4ece\u800c\u5b9e\u73b0\u201c\u8fb9\u6d4b\u8bd5\u3001\u8fb9\u8fdb\u5316\u201d\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aSiru Ouyang, Jun Yan, I-Hung Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long T. Le, Samira Daruki, Xiangru Tang, Vishy Tirumalashetty, George Lee, Mahsan Rofouei, Hangfei Lin, Jiawei Han, Chen-Yu Lee, Tomas Pfister\uff0c\u6765\u81eaUniversity of Illinois Urbana-Champaign, Yale University\u548cGoogle\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"477\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-4-1024x477.png\" alt=\"\" class=\"wp-image-6323\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-4-1024x477.png 1024w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-4-300x140.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-4-768x358.png 768w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-4.png 1245w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2509.25140\" target=\"_blank\" rel=\"noreferrer noopener\">ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u4e00\u3001\u6838\u5fc3\u8d21\u732e\u4e0e\u6574\u4f53\u6846\u67b6<br>\u8bba\u6587\u63d0\u51fa\u7684\u6838\u5fc3\u662f ReasoningBank \u4e0e\u4e0e\u4e4b\u8026\u5408\u7684\u8bb0\u5fc6\u611f\u77e5\u6d4b\u8bd5\u671f\u6269\u5c55 MaTTS\uff08Memory-aware Test-Time Scaling\uff09\uff1a<br>1\uff09ReasoningBank\uff1a\u628a\u667a\u80fd\u4f53\u5728\u4ea4\u4e92\u8f68\u8ff9\u4e2d\u7684\u201c\u601d\u8003\u4e0e\u884c\u52a8\u201d\u84b8\u998f\u4e3a\u7ed3\u6784\u5316\u7684\u201c\u7b56\u7565\u7ea7\u8bb0\u5fc6\u9879\u201d\uff08title\/description\/content \u4e09\u6bb5\u5f0f\uff09\uff0c\u65e2\u4fdd\u7559\u6210\u529f\u7ecf\u9a8c\u4e2d\u7684\u6709\u6548\u5957\u8def\uff0c\u4e5f\u663e\u5f0f\u5438\u7eb3\u5931\u8d25\u7ecf\u5386\u4e2d\u7684\u201c\u9632\u9519\u542f\u793a\/\u53cd\u4e8b\u5b9e\u7ebf\u7d22\u201d\u3002\u4e0e\u53ea\u5b58\u539f\u59cb\u8f68\u8ff9\u6216\u53ea\u5b58\u6210\u529f\u6d41\u7a0b\u7684\u4ee5\u5f80\u65b9\u6cd5\u4e0d\u540c\uff0c\u5b83\u5f3a\u8c03\u63d0\u70bc\u53ef\u8fc1\u79fb\u7684\u201c\u9ad8\u5c42\u63a8\u7406\u6a21\u5f0f\u201d\u3002<br>2\uff09MaTTS\uff1a\u5728\u6d4b\u8bd5\u671f\u5bf9\u540c\u4e00\u4efb\u52a1\u201c\u52a0\u6df1\u63a2\u7d22\u201d\u7684\u540c\u65f6\uff0c\u8ba9 ReasoningBank \u4e3b\u5bfc\u8bb0\u5fc6\u7684\u751f\u6210\u4e0e\u7b5b\u9009\uff0c\u5f62\u6210\u201c\u63a2\u7d22\u2192\u5bf9\u6bd4\/\u81ea\u7ea0\u2192\u9ad8\u8d28\u91cf\u8bb0\u5fc6\u2192\u66f4\u6709\u6548\u63a2\u7d22\u201d\u7684\u6b63\u53cd\u9988\u56de\u8def\u3002\u4f5c\u8005\u5206\u522b\u7ed9\u51fa\u5e76\u884c\u6269\u5c55\uff08\u591a\u6761\u5e76\u884c\u8f68\u8ff9\u7684\u81ea\u5bf9\u6bd4\uff09\u4e0e\u4e32\u884c\u6269\u5c55\uff08\u540c\u4e00\u8f68\u8ff9\u7684\u591a\u8f6e\u81ea\u4fee\u6b63\uff09\u4e24\u79cd\u5b9e\u73b0\u3002<br>3\uff09\u5b9e\u9a8c\u4e0a\u6a2a\u8de8 WebArena\u3001Mind2Web \u4e0e SWE-Bench-Verified\uff0c\u591a\u79cd\u57fa\u5ea7\u6a21\u578b\uff08Gemini-2.5\/Claude-3.7\uff09\uff0c\u5728\u6210\u529f\u7387\u4e0e\u4ea4\u4e92\u6b65\u6570\u4e24\u65b9\u9762\u540c\u65f6\u83b7\u76ca\uff0c\u5e76\u5c55\u793a\u4e86\u201c\u8bb0\u5fc6\u8d28\u91cf \u00d7 \u6d4b\u8bd5\u671f\u6269\u5c55\u89c4\u6a21\u201d\u7684\u534f\u540c\u6548\u5e94\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001\u65b9\u6cd5\u7ec6\u8282\uff1aReasoningBank \u7684\u4e09\u9636\u6bb5\u95ed\u73af<br>1\uff09\u8bb0\u5fc6\u68c0\u7d22\uff08Memory Retrieval\uff09\uff1a\u9762\u5bf9\u65b0\u4efb\u52a1\uff0c\u5148\u7528\u5f53\u524d\u4e0a\u4e0b\u6587\u4ece ReasoningBank \u4e2d\u505a\u5d4c\u5165\u53ec\u56de\uff0c\u9009\u53d6 top-k \u76f8\u5173\u8bb0\u5fc6\u9879\uff0c\u4f5c\u4e3a\u7cfb\u7edf\u6307\u4ee4\u6ce8\u5165\u5230\u7b56\u7565\u751f\u6210\u4e2d\uff0c\u5e2e\u52a9\u51b3\u7b56\u201c\u5c11\u8d70\u5f2f\u8def\u3001\u89c4\u907f\u65e2\u77e5\u5751\u70b9\u201d\u3002<br>2\uff09\u8bb0\u5fc6\u6784\u5efa\uff08Memory Extraction\uff09\uff1a\u4efb\u52a1\u5b8c\u6210\u540e\uff0c\u91c7\u7528\u201cLLM-as-a-judge\u201d\u5728\u65e0\u6807\u7b7e\u6761\u4ef6\u4e0b\u81ea\u8bc4\u8f68\u8ff9\u6210\u8d25\uff1b\u5bf9\u6210\u529f\u8f68\u8ff9\u62bd\u53d6\u201c\u6709\u6548\u7b56\u7565\u201d\uff0c\u5bf9\u5931\u8d25\u8f68\u8ff9\u62bd\u53d6\u201c\u9632\u9519\/\u8fb9\u754c\/\u53cd\u4f8b\u201d\u63d0\u793a\uff1b\u540c\u4e00\u8f68\u8ff9\u53ef\u62bd\u53d6\u591a\u6761\u8bb0\u5fc6\u9879\u4ee5\u8986\u76d6\u4e0d\u540c\u7ef4\u5ea6\u7684\u7b56\u7565\u4fe1\u53f7\u3002<br>3\uff09\u8bb0\u5fc6\u56fa\u5316\uff08Memory Consolidation\uff09\uff1a\u5c06\u65b0\u62bd\u53d6\u7684\u8bb0\u5fc6\u9879\u7b80\u5355\u8ffd\u52a0\u8fdb ReasoningBank\uff0c\u6301\u7eed\u6eda\u52a8\u79ef\u7d2f\uff0c\u5f62\u6210\u201c\u53d6\u7528\u2014\u751f\u6210\u2014\u56de\u704c\u201d\u7684\u95ed\u73af\u3002\u4f5c\u8005\u523b\u610f\u4fdd\u6301\u8fd9\u4e00\u7ba1\u7ebf\u6734\u7d20\uff0c\u4ee5\u7a81\u51fa\u7b56\u7565\u7ea7\u8bb0\u5fc6\u672c\u8eab\u7684\u8d21\u732e\u3002<\/p>\n\n\n\n<p>\u4e09\u3001\u65b9\u6cd5\u7ec6\u8282\uff1aMaTTS \u7684\u4e24\u79cd\u8bb0\u5fc6\u611f\u77e5\u6269\u5c55<br>1\uff09\u5e76\u884c\u6269\u5c55\uff08Parallel Scaling\uff09\uff1a\u9488\u5bf9\u540c\u4e00\u67e5\u8be2\u5e76\u884c\u751f\u6210\u591a\u6761\u8f68\u8ff9\uff0c\u5229\u7528\u201c\u81ea\u5bf9\u6bd4\u201d\u5728\u591a\u89e3\u4e4b\u95f4\u5f52\u7eb3\u5171\u8bc6\u7b56\u7565\u3001\u5254\u9664\u5076\u7136\u6027\u8def\u5f84\uff0c\u4ece\u800c\u4ea7\u51fa\u66f4\u53ef\u9760\u7684\u8bb0\u5fc6\u3002\u8be5\u6d41\u7a0b\u4e0e ReasoningBank \u4ea4\u66ff\uff1a\u5148\u7528\u8bb0\u5fc6\u5f15\u5bfc\u591a\u6837\u63a2\u7d22\uff0c\u518d\u7528\u591a\u8f68\u5bf9\u6bd4\u53cd\u54fa\u66f4\u5f3a\u8bb0\u5fc6\u3002<br>2\uff09\u4e32\u884c\u6269\u5c55\uff08Sequential Scaling\uff09\uff1a\u5728\u5355\u6761\u8f68\u8ff9\u4e0a\u505a\u591a\u8f6e\u81ea\u6211\u4fee\u6b63\uff08self-refine\uff09\uff0c\u628a\u4e2d\u95f4\u7684\u68c0\u89c6\u3001\u7ea0\u504f\u3001\u5047\u8bbe\u68c0\u9a8c\u7b49\u201c\u601d\u7ef4\u75d5\u8ff9\u201d\u4e5f\u4f5c\u4e3a\u8bb0\u5fc6\u4fe1\u53f7\u4fdd\u5b58\uff0c\u4e30\u5bcc\u7b56\u7565\u7c92\u5ea6\u3002\u6269\u5c55\u56e0\u5b50 k \u5206\u522b\u8868\u793a\u5e76\u884c\u8f68\u8ff9\u6570\u6216\u4e32\u884c\u81ea\u4fee\u6b21\u6570\u3002<\/p>\n\n\n\n<p>\u56db\u3001\u5b9e\u9a8c\u8bbe\u7f6e\u4e0e\u5bf9\u6bd4\u57fa\u7ebf<br>\u57fa\u51c6\uff1aWebArena\uff08\u591a\u57df\u7f51\u9875\u64cd\u4f5c\uff0c\u53bb\u9664\u4e86\u5730\u56fe\u57df\u4ee5\u4fdd\u8bc1\u53ef\u7528\u6027\uff09\u3001Mind2Web\uff08\u8de8\u4efb\u52a1\/\u8de8\u7ad9\u70b9\/\u8de8\u9886\u57df\u6cdb\u5316\uff09\uff0cSWE-Bench-Verified\uff08\u4ed3\u5e93\u7ea7\u95ee\u9898\u4fee\u590d\uff09\u3002<br>\u57fa\u5ea7\u6a21\u578b\uff1aGemini-2.5-flash \/ -pro \u4e0e Claude-3.7-sonnet\u3002<br>\u5bf9\u6bd4\uff1a\u65e0\u8bb0\u5fc6\uff08No Memory\uff09\u3001\u8f68\u8ff9\u8bb0\u5fc6\uff08Synapse\uff09\u4e0e\u6d41\u7a0b\u8bb0\u5fc6\uff08AWM\/Agent Workflow Memory\uff09\u3002<br>\u6307\u6807\uff1a\u6210\u529f\u7387\uff08SR\/SSR\/Task SR \u7b49\uff09\u4e0e\u4ea4\u4e92\u6b65\u6570\uff08Step\uff09\u3002<\/p>\n\n\n\n<p>\u4e94\u3001\u4e3b\u8981\u7ed3\u679c\u4e0e\u73b0\u8c61<br>1\uff09\u6574\u4f53\u6709\u6548\u6027\u4e0e\u6548\u7387<br>\u2022 WebArena\uff1aReasoningBank \u5728\u4e09\u4e2a\u57fa\u5ea7\u4e0a\u603b\u4f53 SR \u5206\u522b\u8f83\u65e0\u8bb0\u5fc6\u63d0\u5347 +8.3\u3001+7.2\u3001+4.6\uff0c\u540c\u65f6\u5e73\u5747\u6b65\u6570\u663e\u8457\u964d\u4f4e\uff08\u4f8b\u5982\u5728\u591a\u4e2a\u5b50\u96c6\u4e0a\u8f83 No Memory \u4e0e\u5176\u4ed6\u8bb0\u5fc6\u57fa\u7ebf\u5404\u51cf\u5c11\u7ea6 1.0\u20131.6 \u6b65\uff09\u3002<br>\u2022 SWE-Bench-Verified\uff1aResolve Rate \u63d0\u5347\uff08\u4f8b\u5982 Gemini-2.5-pro\uff1a57.4 vs 54.0\uff09\uff0c\u6b65\u6570\u8fdb\u4e00\u6b65\u4e0b\u964d\uff0819.8 vs 21.1\uff09\uff0c\u8bf4\u660e\u7b56\u7565\u7ea7\u8bb0\u5fc6\u80fd\u51cf\u5c11\u201c\u65e0\u6548\u5c1d\u8bd5\u201d\u3002<br>\u2022 Mind2Web\uff1a\u5728\u8de8\u4efb\u52a1\u3001\u8de8\u7ad9\u4e0e\u8de8\u9886\u57df\u4e09\u79cd\u6cdb\u5316\u8bbe\u5b9a\u4e0b\uff0cReasoningBank \u5728 EA\/AF1\/SSR\/SR \u6307\u6807\u4e0a\u5747\u6709\u589e\u76ca\uff0c\u8de8\u9886\u57df\u589e\u76ca\u6700\u4e3a\u663e\u8457\uff0c\u4f53\u73b0\u5176\u201c\u9ad8\u53ef\u8fc1\u79fb\u7b56\u7565\u201d\u7684\u4ef7\u503c\u3002<br>2\uff09MaTTS \u7684\u6269\u5c55\u6536\u76ca<br>\u2022 \u968f\u7740\u6269\u5c55\u56e0\u5b50 k \u589e\u5927\uff0c\u5e76\u884c\u4e0e\u4e32\u884c\u4e24\u79cd MaTTS \u7684\u6210\u529f\u7387\u603b\u4f53\u4e0a\u5347\uff1b\u5728 k=5 \u65f6\uff0c\u5e76\u884c\u53ef\u8fbe 55.1\u3001\u4e32\u884c 54.5\u3002\u76f8\u6bd4\u201c\u65e0\u8bb0\u5fc6\u6269\u5c55\u201d\uff0cMaTTS \u7684\u589e\u957f\u66f4\u7a33\u5b9a\u3001\u66f4\u53ef\u89c2\uff1b\u76f8\u6bd4\u201c\u65e0\u805a\u5408\u7684\u6734\u7d20 TTS\uff08\u76f8\u5f53\u4e8e\u628a\u591a\u8f68\u8ff9\u5404\u81ea\u72ec\u7acb\u8bb0\u5fc6\u5316\uff09\u201d\uff0cMaTTS \u4e5f\u6301\u7eed\u5360\u4f18\uff0c\u8fd9\u8868\u660e\u201c\u57fa\u4e8e\u5bf9\u6bd4\/\u81ea\u4fee\u7684\u8bb0\u5fc6\u805a\u5408\u201d\u81f3\u5173\u91cd\u8981\u3002R<br>\u2022 \u5c0f\u89c4\u6a21\u6269\u5c55\u65f6\u4e32\u884c\u7565\u5360\u4f18\uff0c\u4f46\u968f\u7740 k \u589e\u5927\u5e76\u884c\u66f4\u80fd\u6301\u7eed\u4ea7\u751f\u591a\u6837\u5316\u63a2\u7d22\u4e0e\u66f4\u5f3a\u7684\u5bf9\u6bd4\u4fe1\u53f7\uff0c\u56e0\u6b64\u5728\u9ad8 k \u533a\u95f4\u53cd\u8d85\u3002<br>3\uff09\u8bb0\u5fc6 \u00d7 \u6269\u5c55\u7684\u534f\u540c<br>\u2022 \u66f4\u597d\u7684\u8bb0\u5fc6\u5e26\u6765\u66f4\u5f3a\u7684 Best-of-N\uff08BoN\uff09\uff1a\u65e0\u8bb0\u5fc6\u65f6\u6269\u5c55\u51e0\u4e4e\u4e0d\u6da8\u6216\u6ce2\u52a8\uff1bSynapse\/AWM \u6709\u4e00\u5b9a\u63d0\u5347\uff1bReasoningBank \u7684\u63d0\u5347\u6700\u5927\uff08\u4f8b\u5982 BoN \u4ece 49.7 \u5347\u81f3 52.4\uff09\u3002<br>\u2022 \u6269\u5c55\u53cd\u54fa\u8bb0\u5fc6\u8d28\u91cf\uff1a\u4ee5 Pass@1 \u8861\u91cf\u201c\u8bb0\u5fc6\u56fa\u5316\u540e\u7684\u5e73\u5747\u8f68\u8ff9\u8d28\u91cf\u201d\uff0c\u5f31\u8bb0\u5fc6\u5728\u6269\u5c55\u540e\u751a\u81f3\u4e0b\u964d\uff08\u566a\u58f0\u5927\u4e8e\u4fe1\u606f\uff09\uff0c\u800c ReasoningBank \u552f\u4e00\u5b9e\u73b0\u6b63\u5411\u63d0\u5347\uff0849.7\u219250.8\uff09\uff0c\u8bf4\u660e\u53ea\u6709\u201c\u9ad8\u8d28\u91cf\u7b56\u7565\u8bb0\u5fc6\u201d\u624d\u80fd\u628a\u6269\u5c55\u751f\u6210\u7684\u591a\u6837\u8f68\u8ff9\u8f6c\u5316\u4e3a\u6709\u6548\u7684\u5bf9\u6bd4\u4fe1\u53f7\uff0c\u5f62\u6210\u826f\u6027\u5faa\u73af\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u6df1\u5165\u5206\u6790\uff1a\u4e3a\u4f55\u201c\u5931\u8d25\u4e5f\u5e94\u88ab\u8bb0\u5fc6\u201d<br>\u4f5c\u8005\u505a\u4e86\u201c\u53ea\u7528\u6210\u529f\u8f68\u8ff9 vs \u540c\u65f6\u7eb3\u5165\u5931\u8d25\u8f68\u8ff9\u201d\u7684\u6d88\u878d\u3002\u7ed3\u679c\u663e\u793a\uff1a\u57fa\u4e8e\u6210\u529f\u8f68\u8ff9\u7684\u4f20\u7edf\u65b9\u6cd5\uff08Synapse\/AWM\uff09\u5728\u52a0\u5165\u5931\u8d25\u6570\u636e\u540e\u6536\u76ca\u6709\u9650\u751a\u81f3\u9000\u5316\uff1b\u800c ReasoningBank \u901a\u8fc7\u4e13\u95e8\u7684\u5931\u8d25\u84b8\u998f\u7b56\u7565\uff0c\u80fd\u628a\u5931\u8d25\u8f6c\u5316\u4e3a\u201c\u8fb9\u754c\/\u9632\u9519\/\u53cd\u4e8b\u5b9e\u201d\u77e5\u8bc6\uff0c\u6210\u529f\u7387\u4ece\u201c\u4ec5\u6210\u529f\u201d\u6761\u4ef6\u4e0b\u7684 46.5 \u8fdb\u4e00\u6b65\u63d0\u5347\u5230 49.7\u3002\u8fd9\u4e3a\u201c\u8ba9\u667a\u80fd\u4f53\u5b66\u4f1a\u4e0d\u53bb\u505a\u4ec0\u4e48\u201d\u63d0\u4f9b\u4e86\u5b9e\u8bc1\u3002<\/p>\n\n\n\n<p>\u4e03\u3001\u6548\u7387\u89c6\u89d2\uff1a\u51cf\u5c11\u7684\u662f\u201c\u5197\u4f59\u63a2\u7d22\u201d\uff0c\u800c\u975e\u201c\u8349\u7387\u653e\u5f03\u201d<br>\u5c06\u6b65\u6570\u5206\u522b\u5728\u201c\u6210\u529f\u7528\u4f8b\/\u5931\u8d25\u7528\u4f8b\u201d\u4e0a\u7edf\u8ba1\uff0cReasoningBank \u5728\u56db\u4e2a WebArena \u5b50\u57df\u5747\u663e\u8457\u51cf\u5c11\u6210\u529f\u7528\u4f8b\u7684\u5e73\u5747\u6b65\u6570\uff08\u6700\u9ad8\u51cf\u5c11 2.1 \u6b65\uff0c\u7ea6 26.9% \u76f8\u5bf9\u4e0b\u964d\uff09\uff0c\u800c\u5bf9\u5931\u8d25\u7528\u4f8b\u7684\u6b65\u6570\u5f71\u54cd\u8f83\u5c0f\u3002\u8fd9\u8868\u660e\u5b83\u4e3b\u8981\u662f\u5728\u201c\u627e\u5bf9\u65b9\u5411\u540e\u5c11\u8d70\u5f2f\u8def\u201d\uff0c\u800c\u4e0d\u662f\u201c\u66f4\u65e9\u653e\u5f03\u201d\u3002<\/p>\n\n\n\n<p>\u516b\u3001\u4e0e\u73b0\u6709\u8bb0\u5fc6\u8303\u5f0f\u7684\u5dee\u5f02\u4e0e\u4f18\u52bf<br>1\uff09\u76f8\u5bf9\u201c\u539f\u59cb\u8f68\u8ff9\u8bb0\u5fc6\u201d\uff08\u91cd\u653e\u957f\u4e0a\u4e0b\u6587\uff09\uff1aReasoningBank \u901a\u8fc7\u7ed3\u6784\u5316\u7684\u7b56\u7565\u9879\uff0c\u907f\u514d\u4e86\u957f\u6587\u672c\u68c0\u7d22\/\u6ce8\u5165\u7684\u566a\u58f0\u4e0e\u6210\u672c\uff0c\u66f4\u5229\u4e8e\u6cdb\u5316\u4e0e\u538b\u7f29\u3002<br>2\uff09\u76f8\u5bf9\u201c\u6d41\u7a0b\/\u5de5\u4f5c\u6d41\u8bb0\u5fc6\u201d\uff08\u53ea\u5b58\u6210\u529f\u6a21\u5757\u5316\u6d41\u7a0b\uff09\uff1a\u5b83\u663e\u5f0f\u5438\u7eb3\u5931\u8d25\u4fe1\u53f7\uff0c\u4e14\u62bd\u53d6\u7684\u662f\u201c\u53ef\u7ec4\u5408\u7684\u7b56\u7565\u4e0e\u5224\u65ad\u8981\u70b9\u201d\uff0c\u4e0d\u5c40\u9650\u4e8e\u56fa\u5b9a\u6d41\u7a0b\uff0c\u6709\u5229\u4e8e\u4e0d\u540c\u7f51\u7ad9\/\u4ed3\u5e93\/\u95ee\u9898\u57df\u95f4\u7684\u8fc1\u79fb\u3002<br>3\uff09\u4e0e\u6d4b\u8bd5\u671f\u6269\u5c55\u534f\u540c\uff1a\u7b56\u7565\u8bb0\u5fc6\u63d0\u4f9b\u4e86\u201c\u5f15\u5bfc\u63a2\u7d22\u7684\u5148\u9a8c\u201d\uff0c\u6269\u5c55\u4ea7\u751f\u201c\u5bf9\u6bd4\u5b66\u4e60\u7684\u591a\u6837\u6837\u672c\u201d\uff0c\u4e8c\u8005\u4e92\u4fc3\uff0c\u5f62\u6210\u65b0\u7684\u6269\u5c55\u7ef4\u5ea6\u2014\u2014\u201c\u7ecf\u9a8c\u9a71\u52a8\u7684\u6269\u5c55\u201d\u3002<\/p>\n\n\n\n<p>\u4e5d\u3001\u6f5c\u5728\u5c40\u9650\u4e0e\u5f00\u653e\u95ee\u9898<br>1\uff09LLM-as-judge \u7684\u81ea\u8bc4\u504f\u5dee\uff1a\u5728\u65e0\u6807\u6ce8\u573a\u666f\u4f9d\u8d56\u6a21\u578b\u81ea\u5224\uff0c\u53ef\u80fd\u5f15\u5165\u7cfb\u7edf\u6027\u504f\u5dee\uff0c\u8fdb\u800c\u5f71\u54cd\u8bb0\u5fc6\u7684\u7eaf\u5ea6\uff1b\u540e\u7eed\u53ef\u8003\u8651\u591a\u8bc4\u5ba1\u3001\u591a\u89c6\u89d2\u4e00\u81f4\u6027\u6216\u8f7b\u91cf\u5916\u90e8\u9a8c\u8bc1\u5668\u3002<br>2\uff09\u8bb0\u5fc6\u5e93\u89c4\u6a21\u4e0e\u68c0\u7d22\uff1a\u968f\u7740\u4efb\u52a1\u6d41\u589e\u957f\uff0c\u8bb0\u5fc6\u9879\u6570\u91cf\u81a8\u80c0\uff0c\u9700\u8981\u66f4\u5f3a\u7684\u7d22\u5f15\u3001\u53bb\u5197\u3001\u805a\u7c7b\u4e0e\u7b56\u7565\u5143\u5b66\u4e60\uff08\u4f8b\u5982\u6839\u636e\u4efb\u52a1\u57df\u81ea\u9002\u5e94\u9009\u62e9\u201c\u5931\u8d25\u9632\u9519\u201d\u7c7b\u8bb0\u5fc6\u7684\u6743\u91cd\uff09\u3002<br>3\uff09\u5bf9\u73af\u5883\u6f02\u79fb\/\u7ad9\u70b9\u5dee\u5f02\u7684\u9c81\u68d2\u6027\uff1a\u7b56\u7565\u5c42\u8bb0\u5fc6\u867d\u53ef\u8fc1\u79fb\uff0c\u4f46\u9875\u9762\u7ed3\u6784\u3001DOM \u53ef\u8bbf\u95ee\u6027\u4e0e\u4ee3\u7801\u4ed3\u751f\u6001\u5dee\u5f02\u4ecd\u4f1a\u9020\u6210\u201c\u7b56\u7565-\u52a8\u4f5c\u843d\u5dee\u201d\uff0c\u9700\u8981\u4e0e\u7ed3\u6784\u5316\u5de5\u5177\u89c2\u6d4b\u3001\u7a0b\u5e8f\u5316\u6280\u80fd\u8bf1\u5bfc\u5171\u540c\u6f14\u8fdb\u3002<br>4\uff09\u8ba1\u7b97-\u6536\u76ca\u66f2\u7ebf\uff1a\u5e76\u884c\/\u4e32\u884c\u6269\u5c55\u90fd\u8981\u4ed8\u51fa\u989d\u5916\u7b97\u529b\uff0c\u4f55\u65f6\u505c\u6b62\u3001\u5982\u4f55\u81ea\u9002\u5e94\u9009\u62e9 k\u3001\u5982\u4f55\u5728\u201c\u5feb\u901f\u89e3\/\u6df1\u5ea6\u63a2\u201d\u95f4\u5e73\u8861\uff0c\u4ecd\u9700\u66f4\u7cfb\u7edf\u7684\u4ee3\u4ef7\u611f\u77e5\u8c03\u5ea6\u3002<\/p>\n\n\n\n<p>\u5341\u3001\u5b9e\u8df5\u843d\u5730\u5efa\u8bae\uff08\u4ece\u5de5\u7a0b\u89c6\u89d2\uff09<br>1\uff09\u8bb0\u5fc6\u9879\u6a21\u677f\u5316\uff1a\u5f3a\u5236\u4e09\u6bb5\u5f0f\uff08\u6807\u9898\/\u4e00\u53e5\u8bdd\u6458\u8981\/\u8981\u70b9\u6e05\u5355\uff09\uff0c\u8981\u70b9\u5c3d\u91cf\u5199\u6210\u201c\u6761\u4ef6-\u52a8\u4f5c-\u68c0\u67e5\u201d\u7ed3\u6784\uff0c\u4fbf\u4e8e\u68c0\u7d22\u4e0e\u6ce8\u5165\u3002<br>2\uff09\u5931\u8d25\u4f18\u5148\u7684\u5b89\u5168\u6805\uff1a\u4e3a\u6bcf\u4e2a\u5e38\u89c1\u5931\u8d25\u6001\uff08\u5982\u201c\u53ea\u770b\u6700\u8fd1\u8ba2\u5355\u201d\u201c\u9519\u628a\u5206\u9875\u4e3a\u65e0\u9650\u6eda\u52a8\u201d\uff09\u5efa\u7acb\u9632\u9519\u8bb0\u5fc6\uff0c\u4f5c\u4e3a\u7cfb\u7edf\u63d0\u793a\u524d\u7f00\u4f18\u5148\u6ce8\u5165\u3002<br>3\uff09\u6269\u5c55\u56e0\u5b50\u81ea\u9002\u5e94\uff1a\u5c0f k \u5148\u7528\u4e32\u884c\uff08\u4f4e\u6210\u672c\u6536\u655b\/\u7ea0\u504f\uff09\uff0c\u5f53\u4e0d\u518d\u4ea7\u751f\u65b0\u6d1e\u89c1\u65f6\u5207\u6362\u5230\u5e76\u884c\u4ee5\u5f15\u5165\u591a\u6837\u6027\u3002<br>4\uff09\u8bb0\u5fc6\u5065\u5eb7\u5ea6\u5ea6\u91cf\uff1a\u5bf9\u8bb0\u5fc6\u9879\u7ef4\u62a4\u201c\u547d\u4e2d\u7387\/\u89e6\u53d1\u540e\u6210\u529f\u7387\/\u8bef\u89e6\u53d1\u7387\/\u65f6\u6548\u6027\u201d\u7b49\u6307\u6807\uff0c\u5b9a\u671f\u81ea\u52a8\u5408\u5e76\u76f8\u4f3c\u9879\u3001\u6dd8\u6c70\u9648\u65e7\u9879\u3002<\/p>\n\n\n\n<p>\u5341\u4e00\u3001\u5b66\u672f\u4e0e\u4ea7\u4e1a\u4ef7\u503c<br>\u8bba\u6587\u628a\u201c\u8bb0\u5fc6\u8d28\u91cf\u201d\u7eb3\u5165\u201c\u6269\u5c55\u7ef4\u5ea6\u201d\u7684\u7b2c\u4e00\u6027\u8981\u7d20\uff0c\u8df3\u51fa\u4e86\u5355\u7eaf\u5806\u7b97\u529b\u6216\u4ec5\u5728\u53c2\u6570\u89c4\u6a21\u4e0a\u7684\u6269\u5c55\u903b\u8f91\uff0c\u63d0\u4f9b\u4e86\u4e00\u4e2a\u201c\u7ecf\u9a8c-\u5bf9\u6bd4-\u8bb0\u5fc6\u201d\u7684\u53ef\u64cd\u4f5c\u95ed\u73af\u3002\u5b83\u5728\u7f51\u9875\u667a\u80fd\u4f53\u4e0e\u4ee3\u7801\u667a\u80fd\u4f53\u4e24\u4e2a\u5178\u578b\u573a\u666f\u540c\u65f6\u9a8c\u8bc1\uff0c\u8bf4\u660e\u8fd9\u79cd\u201c\u7b56\u7565\u7ea7\u8bb0\u5fc6 + \u6d4b\u8bd5\u671f\u6269\u5c55\u201d\u7684\u8303\u5f0f\u6709\u671b\u6210\u4e3a\u901a\u7528\u7684\u667a\u80fd\u4f53\u81ea\u6f14\u5316\u8def\u7ebf\u3002<\/p>\n\n\n\n<p>\u5341\u4e8c\u3001\u7ed3\u8bba<br>ReasoningBank \u901a\u8fc7\u628a\u6210\u529f\u4e0e\u5931\u8d25\u5171\u540c\u84b8\u998f\u4e3a\u201c\u53ef\u8fc1\u79fb\u7684\u7b56\u7565\u8bb0\u5fc6\u201d\uff0c\u5e76\u4e0e MaTTS \u7684\u5e76\u884c\/\u4e32\u884c\u6269\u5c55\u5f62\u6210\u6b63\u53cd\u9988\uff0c\u663e\u8457\u63d0\u5347\u4e86\u667a\u80fd\u4f53\u5728\u8fde\u7eed\u4efb\u52a1\u6d41\u4e2d\u7684\u6709\u6548\u6027\u4e0e\u6548\u7387\uff1b\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5b83\u8bc1\u660e\u4e86\u201c\u8bb0\u5fc6\u9a71\u52a8\u7684\u7ecf\u9a8c\u6269\u5c55\u201d\u8fd9\u4e00\u65b0\u8303\u5f0f\u7684\u53ef\u884c\u6027\uff1a\u9ad8\u8d28\u91cf\u8bb0\u5fc6\u8ba9\u6269\u5c55\u66f4\u6709\u6548\uff0c\u591a\u6837\u6269\u5c55\u53c8\u953b\u9020\u66f4\u9ad8\u8d28\u91cf\u8bb0\u5fc6\uff0c\u6700\u7ec8\u4f7f\u667a\u80fd\u4f53\u5728\u65e0\u76d1\u7763\u7684\u6d4b\u8bd5\u671f\u81ea\u7136\u6d8c\u73b0\u51fa\u66f4\u590d\u6742\u3001\u66f4\u7a33\u5065\u7684\u63a8\u7406\u884c\u4e3a\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u9a71\u52a8\u7684\u667a\u80fd\u4f53\u6b63\u88ab\u7528\u4e8e\u957f\u671f\u3001\u6301\u7eed\u7684\u771f\u5b9e\u4e16\u754c\u4efb\u52a1\uff08\u5982\u7f51\u9875\u6d4f\u89c8\u3001\u8f6f\u4ef6\u5de5\u7a0b\u81ea\u52a8\u5316\uff09\uff0c\u4f46\u4e3b\u6d41\u667a\u80fd\u4f53 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,94,86,66],"class_list":["post-6322","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-agentic-ai","tag-llm","tag-google"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u9a71\u52a8\u7684\u667a\u80fd\u4f53\u6b63\u88ab\u7528\u4e8e\u957f\u671f\u3001\u6301\u7eed\u7684\u771f\u5b9e\u4e16\u754c\u4efb\u52a1\uff08\u5982\u7f51\u9875\u6d4f\u89c8\u3001\u8f6f\u4ef6\u5de5\u7a0b\u81ea\u52a8\u5316\uff09\uff0c\u4f46\u4e3b\u6d41\u667a\u80fd\u4f53&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6322","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6322"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6322\/revisions"}],"predecessor-version":[{"id":6324,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6322\/revisions\/6324"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6322"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6322"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6322"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}