{"id":1177314,"date":"2026-07-01T03:15:22","date_gmt":"2026-07-01T10:15:22","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1177314"},"modified":"2026-07-01T03:15:23","modified_gmt":"2026-07-01T10:15:23","slug":"%e5%bc%80%e6%ba%90%e4%b8%8a%e6%96%b0-%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%98%af%e5%90%a6%e8%bf%98%e5%9c%a8%e9%87%91%e9%b1%bc%e8%ae%b0%e5%bf%86%ef%bc%9f%e5%85%a8%e6%96%b0%e5%9f%ba%e5%87%86-rhelm","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/%e5%bc%80%e6%ba%90%e4%b8%8a%e6%96%b0-%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%98%af%e5%90%a6%e8%bf%98%e5%9c%a8%e9%87%91%e9%b1%bc%e8%ae%b0%e5%bf%86%ef%bc%9f%e5%85%a8%e6%96%b0%e5%9f%ba%e5%87%86-rhelm\/","title":{"rendered":"\u5f00\u6e90\u4e0a\u65b0 | \u5927\u6a21\u578b\u662f\u5426\u8fd8\u5728”\u91d1\u9c7c\u8bb0\u5fc6”\uff1f\u5168\u65b0\u57fa\u51c6 RHELM \u6d4b\u51fa\u201c\u771f\u5b9e\u957f\u671f\u8bb0\u5fc6\u201d\u5929\u82b1\u677f"},"content":{"rendered":"\n
\u6211\u4eec\u5df2\u7ecf\u4e60\u60ef\u8ba9\u5927\u6a21\u578b\u56de\u7b54\u95ee\u9898\u3001\u751f\u6210\u4ee3\u7801\u3001\u64b0\u5199\u62a5\u544a\uff0c\u4f46\u5f53\u5b83\u771f\u6b63\u8981\u6210\u4e3a\u4e00\u540d”\u8d34\u8eab\u7684\u4e2a\u4eba\u52a9\u7406”\u65f6\uff0c\u6700\u96be\u7684\u53cd\u800c\u662f\u4e00\u4ef6\u6700\u6734\u7d20\u7684\u4e8b\u2014\u2014\u8bb0\u5f97\u4f4f\u3002<\/p>\n\n\n\n
\u5b83\u9700\u8981\u8bb0\u5f97\u4f4f\u4f60\u4e09\u4e2a\u6708\u524d\u63d0\u8fc7\u7684\u8fc7\u654f\u836f\u7269\uff0c\u8bb0\u5f97\u4f4f\u4f60\u4e0a\u5468\u624d\u642c\u5bb6\u5230\u65b0\u7684\u57ce\u5e02\uff0c\u8bb0\u5f97\u4f4f\u4f60\u6628\u5929\u4ea4\u7ed9\u5b83\u7684\u90a3\u4efd\u957f\u8fbe 30 \u9875\u7684\u9879\u76ee\u62a5\u544a\u2026\u2026<\/p>\n\n\n\n
\u4e3a\u4e86\u653b\u514b\u8fd9\u4e00\u96be\u9898\uff0c\u6765\u81ea\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u3001\u5fae\u8f6fAI\u4e0e\u4e2d\u56fd\u4eba\u6c11\u5927\u5b66\u7684\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86RHELM<\/strong>\uff08Realistic, Heterogeneous, and Evolving Long-term Memory\uff09\uff0c\u4e00\u4e2a\u4e13\u4e3a\u8bc4\u6d4b\u5927\u6a21\u578b\u201c\u771f\u5b9e\u3001\u5f02\u8d28\u3001\u52a8\u6001\u957f\u671f\u8bb0\u5fc6\u80fd\u529b\u201d\u800c\u8bbe\u8ba1\u7684\u5168\u65b0\u57fa\u51c6<\/strong>\u3002\u4e0d\u540c\u4e8e\u4ee5\u5f80\u9759\u6001\u3001\u62fc\u51d1\u7684\u6d4b\u8bd5\u96c6\uff0cRHELM \u9996\u6b21\u901a\u8fc7\u6a21\u62df\u957f\u8fbe\u4e00\u5e74\u7684\u52a8\u6001\u865a\u62df\u4eba\u751f\u8f68\u8ff9\uff0c\u4e3a\u5927\u6a21\u578b\u6784\u7b51\u4e86\u4e00\u4e2a\u9ad8\u5ea6\u8fd8\u539f\u73b0\u5b9e\u4e16\u754c\u7684\u201c\u8bb0\u5fc6\u8003\u573a\u201d\u3002<\/strong>\u76ee\u524d\uff0c\u8be5\u7814\u7a76\u7684\u8bba\u6587\u4e0e\u6570\u636e\u5df2\u5168\u90e8\u5f00\u6e90\u3002<\/p>\n\n\n\n \u76f8\u5173\u94fe\u63a5\u5df2\u6574\u7406\u4e8e\u6587\u672b\uff0c\u6b22\u8fce\u70b9\u51fb\u4e86\u89e3\u66f4\u591a\u6280\u672f\u8be6\u60c5\u3002<\/em><\/strong><\/p>\n\n\n\n \u8fc7\u53bb\u51e0\u5e74\uff0c\u4e1a\u754c\u867d\u7136\u5df2\u7ecf\u51fa\u73b0\u4e86 LongMemEval\u3001LoCoMo\u3001PerLTQA\u3001PersonaMem \u7b49\u591a\u4e2a\u957f\u671f\u8bb0\u5fc6\u57fa\u51c6\uff0c\u4f46\u8fd9\u4e9b\u4f20\u7edf\u57fa\u51c6\u666e\u904d\u5b58\u5728\u201c\u4e09\u4e2a\u7ed3\u6784\u6027\u7f3a\u9677<\/strong>\u201d\uff1a<\/p>\n\n\n\n \u9996\u5148\u662f\u8bed\u4e49\u4e0d\u8fde\u8d2f\uff0c\u4eba\u8bbe\u8fc7\u4e8e”\u6241\u5e73”<\/strong>\u3002\u4e3a\u4e86\u4eba\u4e3a\u62c9\u957f\u4e0a\u4e0b\u6587\uff0c\u5f88\u591a\u57fa\u51c6\u5f80\u5f80\u4f1a\u5c06\u5f7c\u6b64\u65e0\u5173\u7684\u5bf9\u8bdd\u7247\u6bb5\u5f3a\u884c\u62fc\u63a5\u5728\u4e00\u8d77\u3002\u8fd9\u79cd”\u957f\u5bf9\u8bdd”\u5728\u8bed\u4e49\u4e0a\u662f\u65ad\u88c2\u7684\uff0c\u5176\u80cc\u540e\u7684\u7528\u6237\u753b\u50cf\u4e5f\u53ea\u662f\u51e0\u4e2a\u9759\u6001\u6807\u7b7e\uff0c\u65e0\u6cd5\u4f53\u73b0”\u4e00\u4e2a\u4eba\u4f1a\u968f\u7740\u65f6\u95f4\u6162\u6162\u53d8\u5316”\u7684\u57fa\u672c\u4e8b\u5b9e\u3002<\/p>\n\n\n\n \u5176\u6b21\u662f\u4fe1\u606f\u6e90\u5355\u4e00\uff0c\u53ea\u6709\u5bf9\u8bdd<\/strong>\u3002\u5728\u771f\u5b9e\u7684\u573a\u666f\u4e2d\uff0cAI \u52a9\u624b\u9762\u5bf9\u7684\u4e0d\u53ea\u6709\u804a\u5929\u8bb0\u5f55\uff0c\u8fd8\u6709\u90ae\u4ef6\u3001\u65e5\u62a5\u3001\u9879\u76ee\u6587\u6863\u3001\u4e2a\u4eba\u65e5\u8bb0\u7b49\u7ed3\u6784\u5404\u5f02\u7684\u6587\u672c\u3002\u8fd9\u4e9b”\u975e\u5bf9\u8bdd”\u6750\u6599\u7684\u4fe1\u606f\u5bc6\u5ea6\u66f4\u9ad8\uff0c\u4e5f\u66f4\u8d34\u8fd1\u771f\u5b9e\u5de5\u4f5c\u6d41\uff0c\u7136\u800c\u7edd\u5927\u591a\u6570\u73b0\u6709\u57fa\u51c6\u4f9d\u7136\u505c\u7559\u5728”\u7eaf\u804a\u5929”\u7684\u8bbe\u5b9a\u91cc\u3002<\/p>\n\n\n\n \u6700\u540e\u662f\u8bc4\u6d4b\u9898\u76ee\u8fc7\u4e8e”\u8001\u5b9e”<\/strong>\u3002\u73b0\u6709\u7684\u8bc4\u6d4b\u9898\u76ee\u5927\u591a\u662f”\u5927\u6d77\u635e\u9488”\u5f0f\u7684\u4e8b\u5b9e\u62bd\u53d6\uff0c\u53ea\u8981\u6a21\u578b\u80fd\u628a\u7b54\u6848\u4ece\u5386\u53f2\u8bb0\u5f55\u91cc\u627e\u51fa\u6765\u5c31\u7b97\u8fc7\u5173\u3002\u4f46\u771f\u5b9e\u7528\u6237\u5f80\u5f80\u4f1a\u63d0\u51fa\u4e0e\u81ea\u8eab\u72b6\u6001\u76f8\u77db\u76fe\u7684\u8bf7\u6c42\uff0c\u6bd4\u5982\u817f\u4f24\u672a\u6108\u5374\u5728\u8be2\u95ee\u5468\u672b\u7684\u9a91\u884c\u8def\u7ebf\uff0c\u6216\u8005\u521a\u642c\u4e86\u5bb6\u8fd8\u5728\u95ee\u8001\u623f\u5b50\u9644\u8fd1\u7684\u9910\u5385\u3002\u4e00\u4e2a\u771f\u6b63”\u6709\u8bb0\u5fc6”\u7684\u52a9\u624b\uff0c\u5e94\u5f53\u80fd\u591f\u4e3b\u52a8\u8bc6\u522b\u8fd9\u79cd\u9690\u542b\u7684\u51b2\u7a81\uff0c\u800c\u4e0d\u662f\u673a\u68b0\u5730\u7167\u529e\u3002<\/p>\n\n\n\n \u4e3a\u4e86\u540c\u65f6\u89e3\u51b3\u4e0a\u8ff0\u8fd9\u4e09\u4e2a\u95ee\u9898\uff0cRHELM \u5e94\u8fd0\u800c\u751f\u3002RHELM \u7684\u6838\u5fc3\u601d\u8def\u53ef\u4ee5\u6982\u62ec\u4e3a\uff1a\u5148\u9020\u4eba\uff0c\u518d\u9020\u751f\u6d3b\uff0c\u6700\u540e\u624d\u9020\u5bf9\u8bdd\u548c\u6587\u6863<\/strong>\u3002\u6574\u4e2a\u6570\u636e\u6784\u5efa\u6d41\u7a0b\u56f4\u7ed5\u4e09\u4e2a\u652f\u67f1\u5c55\u5f00\uff1a<\/p>\n\n\n\n \u6700\u7ec8\uff0cRHELM\u5305\u542b10\u4f4d\u753b\u50cf\u5404\u5f02\u7684\u865a\u62df\u7528\u6237\u300111,764 \u8f6e\u5bf9\u8bdd\u30012,180 \u4efd\u5916\u90e8\u6750\u6599\u3002\u5355\u4e2a\u7528\u6237\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u53ef\u8fbe500K\u20131M tokens\uff0c\u5e76\u914d\u5957\u4e861,305 \u9053\u9ad8\u96be\u5ea6\u95ee\u7b54\u9898\uff0c\u8986\u76d67\u5927\u7c7b\u300127\u9879\u7cbe\u5fc3\u5b9a\u4e49\u7684\u590d\u6742\u201c\u8bb0\u5fc6\u201d\u7279\u5f81\u3002<\/p>\n\n\n\n <\/p>\n\n\n\n \u5173\u952e\u5f15\u64ce\uff1aLOOP \u6a21\u5757\u5982\u4f55”\u517b\u6210”\u4e00\u4e2a\u201c\u6709\u8840\u6709\u8089\u201d\u7684\u865a\u62df\u4eba<\/strong><\/p>\n\n\n\n LOOP \u662f RHELM \u6700\u5177\u7279\u8272\u7684\u8bbe\u8ba1\u3002\u5b83\u7cbe\u5999\u5730\u628a”\u751f\u6210\u957f\u671f\u5bf9\u8bdd”\u7684\u4efb\u52a1\uff0c\u8f6c\u5316\u4e3a\u4e86”\u6a21\u62df\u4e00\u4e2a\u4eba\u7684\u771f\u5b9e\u751f\u6d3b”\uff1a<\/p>\n\n\n\n <\/p>\n\n\n\n \u6b63\u662f\u8fd9\u79cd\u7531\u6982\u7387\u4e0e\u4e8b\u4ef6\u9a71\u52a8\u7684\u8f68\u8ff9\u8bbe\u8ba1\uff0c\u8ba9RHELM \u7684\u6570\u636e\u5448\u73b0\u51fa\u771f\u5b9e\u751f\u6d3b\u7684\u5076\u7136\u6027\u4e0e\u957f\u5c3e\u6027<\/strong>\uff0c\u800c\u8fd9\u6070\u6070\u662f\u5f53\u524d\u6a21\u578b\u6700\u96be\u5904\u7406\u7684\u90e8\u5206\u3002\u5728 LOOP \u7684\u6bcf\u4e00\u6b65\u4e4b\u4e0a\uff0cRHELM \u8fd8\u53e0\u52a0\u4e86\u4e00\u5c42\u5f02\u8d28\u5916\u90e8\u6e90\u751f\u6210\uff0c\u628a\u65e5\u5e38\u4e8b\u4ef6\u8f6c\u5316\u4e3a\u6b63\u5f0f\u98ce\u683c\u7684\u62a5\u544a\u3001\u79c1\u4eba\u98ce\u683c\u7684\u65e5\u8bb0\u3001\u7ed3\u6784\u5316\u7684\u90ae\u4ef6\uff0c\u5e76\u501f\u52a9 Deep Research Agent \u8fdb\u4e00\u6b65\u5145\u5b9e\u7ec6\u8282\uff0c\u8ba9\u6bcf\u4efd\u6587\u6863\u90fd\u62e5\u6709”\u8db3\u4ee5\u4e71\u771f”\u7684\u590d\u6742\u5ea6\u3002<\/p>\n\n\n\n \u4e03\u5927\u9898\u578b \u00d7 27 \u9879\u7279\u5f81\uff1a\u628a”\u8bb0\u5fc6\u80fd\u529b”\u62c6\u5230\u8d85\u7ec6\u9897\u7c92\u5ea6<\/strong><\/p>\n\n\n\n RHELM \u5c06\u5bf9”\u8bb0\u5fc6”\u7684\u8003\u6838\u62c6\u89e3\u6210\u4e867 \u7c7b\u95ee\u9898\uff0c\u8986\u76d6”\u7eaf\u5bf9\u8bdd”\u548c”\u5f02\u8d28\u6e90”\u4e24\u4e2a\u5c42\u9762\u3002\u5bf9\u8bdd\u7c7b<\/strong>\u5305\u542b\u4e8b\u5b9e\u578b\uff08fact\uff09\u3001\u65f6\u5e8f\u578b\uff08temporal\uff09\u3001\u5e7b\u89c9\u578b\uff08hallucination\uff09\u3001\u805a\u5408\u578b\uff08aggregation\uff09\u3001\u8bef\u5bfc\u578b\uff08misleading\uff09\u4e94\u79cd\u9898\u578b\uff1b\u5f02\u8d28\u6e90\u7c7b<\/strong>\u5219\u5305\u542b\u7eaf\u5916\u90e8\u6e90\u95ee\u7b54\uff08external source\uff09\u4e0e\u8de8\u6e90\u6df7\u5408\u95ee\u7b54\uff08mixed\uff09\u4e24\u79cd\u9898\u578b\u3002<\/p>\n\n\n\n <\/p>\n\n\n\n \u4e3a\u4e86\u4e25\u683c\u6d4b\u8bd5\u6a21\u578b\uff0c\u6bcf\u9053\u9898\u90fd\u88ab\u5f3a\u5236\u7ed1\u5b9a\u4e86\u81f3\u5c11\u4e00\u9879\u6311\u6218\u6027\u7279\u5f81\uff08\u517127\u9879\uff09\uff0c\u5305\u62ec\u8de8\u65e5\u805a\u5408\u3001\u8de8\u6e90\u5bf9\u9f50\u3001\u6a21\u7cca\u6307\u4ee3\u3001\u9690\u542b\u72b6\u6001\u7ea6\u675f\u7b49\u3002\u8fd9\u79cd\u7ec6\u7c92\u5ea6\u7684\u6807\u7b7e\u4f53\u7cfb\uff0c\u4f7f\u5f97\u540e\u7eed\u5206\u6790\u53ef\u4ee5\u7cbe\u786e\u5b9a\u4f4d”\u6a21\u578b\u5728\u54ea\u4e00\u7c7b\u7ec6\u8282\u4e0a\u6454\u4e86\u8de4”\u3002<\/p>\n\n\n\n \u5176\u4e2d\u6700\u5177\u521b\u65b0\u6027\u7684\u662f\u8bb0\u5fc6\u6761\u4ef6\u4e0b\u7684\u8bef\u5bfc\u67e5\u8be2\uff08Memory-Conditioned Misleading Queries\uff09<\/strong>\u3002\u7814\u7a76\u5458\u4eec\u4f1a\u523b\u610f\u6311\u9009\u7528\u6237\u751f\u6d3b\u4e2d\u7684\u5173\u952e\u8f6c\u6298\u4e8b\u4ef6\uff08\u5982\u6162\u6027\u4f24\u75c5\u3001\u642c\u8fc1\u3001\u8f6c\u884c\uff09\uff0c\u5e76\u57fa\u4e8e\u6b64\u8bbe\u8ba1\u4e0e\u7528\u6237\u5f53\u524d\u72b6\u6001\u76f8\u51b2\u7a81\u7684”\u9677\u9631\u5f0f\u8bf7\u6c42”\u3002\u4f8b\u5982\uff0c\u7528\u6237\u4e0a\u4e2a\u6708\u521a\u56e0\u819d\u4f24\u88ab\u533b\u751f\u5efa\u8bae\u505c\u6b62\u8dd1\u6b65\uff0c\u672c\u6708\u5374\u5411AI\u52a9\u624b\u8be2\u95ee”\u5468\u672b\u6709\u54ea\u4e9b\u9002\u5408\u957f\u8dd1\u7684\u8def\u7ebf”\u3002\u4e00\u4e2a\u771f\u6b63\u5177\u5907\u201c\u957f\u671f\u8bb0\u5fc6\u201d\u7684AI\u52a9\u624b\uff0c\u4e0d\u8be5\u7b80\u5355\u7167\u529e\uff0c\u800c\u5e94\u4e3b\u52a8\u56de\u6eaf\u5386\u53f2\u3001\u8bc6\u522b\u51b2\u7a81\u3001\u793c\u8c8c\u6307\u51fa\u95ee\u9898\uff0c\u5e76\u7ed9\u51fa\u7b26\u5408\u5f53\u524d\u7ea6\u675f\u7684\u66ff\u4ee3\u65b9\u6848\u3002\u8fd9\u662f\u8fc7\u53bb\u7684\u57fa\u51c6\u51e0\u4e4e\u4ece\u672a\u89e6\u53ca\u7684\u8003\u5bdf\u7ef4\u5ea6\uff0c\u4e5f\u662f RHELM \u60f3\u771f\u6b63\u63a8\u52a8\u4e1a\u754c\u53bb\u89e3\u51b3\u7684\u6838\u5fc3\u75db\u70b9\u3002<\/strong><\/p>\n\n\n\n \u7814\u7a76\u5458\u4eec\u5728 RHELM \u4e0a\u7cfb\u7edf\u5730\u8bc4\u6d4b\u4e86\u76ee\u524d\u5e02\u9762\u4e0a\u7684\u4e09\u7c7b\u4e3b\u6d41\u957f\u6587\u672c\u4e0e\u8bb0\u5fc6\u65b9\u6848\uff1a<\/p>\n\n\n\n \u7136\u800c\uff0c\u6700\u7ec8\u7684\u8bc4\u6d4b\u7ed3\u679c\u7ed9\u6574\u4e2a\u884c\u4e1a\u6cfc\u4e86\u4e00\u76c6\u51b7\u6c34\uff0c\u4e5f\u8bca\u65ad\u51fa\u4e86\u5f53\u524d\u5927\u6a21\u578b\u201c\u957f\u671f\u8bb0\u5fc6\u201d\u7684\u5e95\u5c42\u74f6\u9888\u3002<\/p>\n\n\n\n \u884c\u4e1a\u73b0\u72b6\uff1a\u6574\u4f53\u5206\u6570\u504f\u4f4e\uff0c\u591a\u6e90\u6df7\u5408\u6210\u4e3a\u666e\u904d\u91cd\u707e\u533a<\/strong><\/p>\n\n\n\n \u5b9e\u6d4b\u6570\u636e\u663e\u793a\uff0c\u5927\u6a21\u578b\u7684\u6574\u4f53\u5206\u6570\u666e\u904d\u504f\u4f4e<\/strong>\u3002\u8868\u73b0\u6700\u5f3a\u7684Claude Opus 4.5 \u5728\u5f15\u5165\u5916\u90e8\u6e90\u540e\u7684\u5e73\u5747\u5206\u4ec5\u4e3a 38.1\uff0c\u6ca1\u6709\u5916\u90e8\u6e90\u65f6\u4e5f\u53ea\u6709 36.2\u3002\u8fd9\u8868\u660e\u5927\u6a21\u578b\u8ddd\u79bb”\u53ef\u9760\u7684\u4e2a\u4eba\u8bb0\u5fc6\u52a9\u624b”\u8fd8\u6709\u5f88\u957f\u4e00\u6bb5\u8def\u8981\u8d70\u3002<\/p>\n\n\n\n \u4ee4\u4eba\u610f\u5916\u7684\u662f\uff0c\u52a0\u5165\u5916\u90e8\u6e90\u5e76\u4e0d\u603b\u662f\u5229\u597d<\/strong>\u3002\u4e00\u65e6\u628a\u90ae\u4ef6\u3001\u65e5\u5fd7\u3001\u62a5\u544a\u7b49\u5f02\u8d28\u6750\u6599\u585e\u8fdb\u4e0a\u4e0b\u6587\uff0cRAG \u5728\u6807\u51c6\u9898\u578b\u4e0a\u7684\u5f97\u5206\u53cd\u800c\u51fa\u73b0\u4e0b\u6ed1\uff08\u4f8b\u5982 RAG@k=50 \u7684\u5f97\u5206\u4ece 59.9 \u8dcc\u5230 54.6\uff09\u3002\u8fd9\u8bf4\u660e\u73b0\u6709\u68c0\u7d22\u673a\u5236\u8fd8\u6ca1\u6709\u5b66\u4f1a\u8de8\u6a21\u6001\u5730\u878d\u5408\u8bb0\u5fc6\u3002\u56e0\u6b64\uff0c\u8de8\u6e90\u6df7\u5408\u9898\u578b\u6210\u4e3a\u666e\u904d\u91cd\u707e\u533a<\/strong>\u3002\u65e0\u8bba\u54ea\u79cd\u8303\u5f0f\uff0c\u5728\u9700\u8981”\u5bf9\u8bdd + \u5916\u90e8\u6e90”\u534f\u540c\u63a8\u7406\u7684\u9898\u76ee\u4e0a\uff0c\u5927\u6a21\u578b\u7684\u5f97\u5206\u90fd\u8dcc\u843d\u5f97\u6700\u4e3a\u60e8\u70c8\u3002<\/p>\n\n\n\n \u540c\u65f6\uff0c\u5e7b\u89c9\u4e0e\u8bef\u5bfc\u578b\u9898\u76ee\u65e0\u60c5\u66b4\u9732\u77ed\u677f<\/strong>\u3002\u5bf9\u51e0\u4e4e\u6240\u6709\u65b9\u6cd5\u800c\u8a00\uff0c\u8bef\u5bfc\u578b\u9898\u578b\u7684\u51c6\u786e\u7387\u751a\u81f3\u4e0d\u8db35%\uff1b\u800c\u4e14 RAG \u68c0\u7d22\u5230\u7684\u8bc1\u636e\u8d8a\u591a\uff0c\u5e7b\u89c9\u578b\u9898\u578b\u7684\u5f97\u5206\u53cd\u800c\u8d8a\u4f4e\uff08\u4ece13.2\u964d\u81f311.2\uff09\u3002<\/p>\n\n\n\n \u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5f3a\u63a8\u7406\u6a21\u578b\u5c55\u73b0\u51fa\u660e\u663e\u4f18\u52bf<\/strong>\u3002Claude Opus 4.5\u548cGemini-2.5-Pro \u5728\u5e7b\u89c9\u4e0e\u8bef\u5bfc\u7ef4\u5ea6\u4e0a\u7684\u8868\u73b0\u663e\u8457\u597d\u4e8e\u5176\u4ed6\u6a21\u578b\u3002\u8fd9\u8bf4\u660e\u9ad8\u9636\u7684\u63a8\u7406\u80fd\u529b\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u8bc6\u522b\u5e76\u62b5\u5fa1”\u770b\u4f3c\u5408\u7406”\u7684\u865a\u5047\u524d\u63d0\u3002<\/p>\n\n\n\n <\/p>\n\n\n\n \u95ee\u9898\u51fa\u5728\u54ea\uff1f\u68c0\u7d22\u53ec\u56de\u7684\u5929\u82b1\u677f<\/strong><\/p>\n\n\n\n \u4e3a\u4e86\u8fdb\u4e00\u6b65\u5b9a\u4f4d\u6280\u672f\u74f6\u9888\uff0c\u7814\u7a76\u5458\u4eec\u6bd4\u8f83\u4e86 bge-large-en-v1.5\u3001bge-m3\u3001all-MiniLM-L6-v2 \u4ee5\u53ca OpenAI \u7cfb\u5217 embedding \u5728\u4e0d\u540c top-k \u4e0b\u7684\u53ec\u56de\u7387\u3002<\/p>\n\n\n\n \u5f97\u51fa\u7684\u7ed3\u8bba\u76f8\u5f53\u4e0d\u4e50\u89c2\uff1a\u5373\u4fbf\u628a\u68c0\u7d22\u9884\u7b97\u653e\u5bbd\u5230 k=50\uff0c\u53ec\u56de\u5230\u7684\u8bc1\u636e\u4f9d\u7136\u6709\u9650\uff0c\u8fdc\u4e0d\u8db3\u4ee5\u652f\u6491\u5927\u6a21\u578b\u8fdb\u884c\u7cbe\u786e\u4f5c\u7b54\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5728 RHELM \u8fd9\u79cd\u957f\u7a0b\u3001\u5f02\u8d28\u3001\u52a8\u6001\u7684\u8bed\u6599\u4e0b\uff0c\u5355\u7eaf\u901a\u8fc7\u201c\u5806\u5411\u91cf\u68c0\u7d22\u201d\u7684\u4f20\u7edf\u65b9\u6cd5\u5df2\u7ecf\u65e0\u6cd5\u6ee1\u8db3\u771f\u5b9e\u8bb0\u5fc6\u52a9\u624b\u7684\u9700\u6c42\u3002\u8fd9\u4e00\u53d1\u73b0\u628a\u77db\u5934\u76f4\u63a5\u6307\u5411\u4e86\u8bb0\u5fc6\u7cfb\u7edf\u7684\u5e95\u5c42\u67b6\u6784\u8bbe\u8ba1\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f”\u6362\u4e2a\u66f4\u5f3a\u7684 embedding\u6a21\u578b”\u90a3\u4e48\u7b80\u5355\u3002<\/p>\n\n\n\n <\/p>\n\n\n\n \u6700\u96be\u7684 10 \u9879\u7279\u5f81\uff1a\u6a21\u578b\u7a76\u7adf\u5361\u5728\u4e86\u54ea\u4e00\u6b65\uff1f<\/strong><\/p>\n\n\n\n \u7814\u7a76\u5458\u4eec\u8fdb\u4e00\u6b65\u62bd\u53d6\u51fa”\u8868\u73b0\u6700\u5dee\u7684 10 \u9879\u6311\u6218\u7279\u5f81”\u8fdb\u884c\u4e86\u7cbe\u7ec6\u5206\u6790\u3002\u7ed3\u679c\u6307\u5411\u4e24\u4e2a\u660e\u786e\u7684”\u91cd\u707e\u533a”\uff1a<\/p>\n\n\n\n \u4e00\u4e2a\u662f\u8de8\u6e90\u4fe1\u606f\u805a\u5408\uff08Cross-source Aggregation\uff09\u3002\u8fd9\u4e3b\u8981\u96c6\u4e2d\u5728\u8de8\u6e90\u6df7\u5408\u95ee\u7b54\u4e0e\u805a\u5408\u578b\u95ee\u9898\u4e2d\uff0c\u6a21\u578b\u5e38\u5e38\u6df7\u6dc6\u4fe1\u606f\u6765\u6e90\uff0c\u6216\u65e0\u6cd5\u6709\u6548\u89e3\u51b3\u76f8\u4e92\u77db\u76fe\u7684\u5386\u53f2\u7247\u6bb5\u3002\u53e6\u4e00\u4e2a\u662f\u771f\u5b9e\u60c5\u5883\u63a8\u7406\uff08Real-world Contextual Reasoning\uff09\u3002\u8fd9\u591a\u96c6\u4e2d\u5728\u8bef\u5bfc\u578b\u4e0e\u5e7b\u89c9\u884c\u9898\u76ee\u91cc\uff0c\u6a21\u578b\u5bb9\u6613\u865a\u6784\u4e0d\u5b58\u5728\u7684\u4e8b\u5b9e\uff0c\u6216\u5728\u63a8\u7406\u65f6\u5b8c\u5168\u5ffd\u7565\u4e86\u7528\u6237\u5f53\u524d\u7684\u771f\u5b9e\u72b6\u6001\u3002<\/p>\n\n\n\n <\/p>\n\n\n\n \u4e00\u53e5\u8bdd\u603b\u7ed3\uff1a\u4eca\u5929\u7684\u8bb0\u5fc6\u589e\u5f3a\u6a21\u578b\u5728”\u80fd\u4e0d\u80fd\u8bb0\u4f4f”\u65b9\u9762\u8868\u73b0\u5f97\u4e0d\u7b97\u592a\u5dee\uff0c\u4f46\u5728”\u8bb0\u4f4f\u540e\u80fd\u4e0d\u80fd\u6b63\u786e\u5730\u7528”\u8fd9\u4ef6\u4e8b\u4e0a\uff0c\u4ecd\u7136\u6709\u660e\u663e\u7684\u5929\u82b1\u677f\u3002<\/p>\n\n\n\n \u7eb5\u89c2\u5168\u7bc7\uff0cRHELM \u7684\u63a8\u51fa\u4e3a\u6574\u4e2a\u884c\u4e1a\u5e26\u6765\u4e86\u6df1\u523b\u7684\u542f\u793a\u3002\u4f5c\u4e3a\u9996\u4e2a\u5c06\u5bf9\u8bdd\u6d41\u4e0e\u5f02\u8d28\u5916\u90e8\u6e90\u8fdb\u884c\u6df1\u5ea6\u8026\u5408\u7684\u957f\u671f\u8bb0\u5fc6\u57fa\u51c6\uff0cRHELM\u4e0d\u4ec5\u8ba9\u8bc4\u6d4b\u573a\u666f\u771f\u6b63\u8d34\u8fd1\u4e86\u201c\u4e2a\u4eba\u52a9\u7406\u201d\u7684\u771f\u5b9e\u65e5\u5e38\uff0c\u8fd8\u901a\u8fc7 27 \u9879\u53ef\u7ec6\u7c92\u5ea6\u5f52\u56e0\u7684\u6311\u6218\u6027\u7279\u5f81\uff0c\u4e3a\u540e\u7eed\u7684\u7814\u7a76\u5de5\u4f5c\u63d0\u4f9b\u4e86\u6e05\u6670\u7684\u80fd\u529b\u7ef4\u5ea6\u6293\u624b\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5229\u7528\u8986\u76d6\u5168\u4e0a\u4e0b\u6587\u3001RAG\u3001\u8bb0\u5fc6\u6846\u67b6\u4e09\u7c7b\u8303\u5f0f\u7684\u7cfb\u7edf\u6027\u8bc4\u6d4b\uff0c\u5b83\u660e\u786e\u6307\u51fa\u4e86\u5f53\u524d SOTA \u6a21\u578b\u5728\u8de8\u6e90\u805a\u5408\u4e0e\u771f\u5b9e\u60c5\u5883\u63a8\u7406\u4e0a\u7684\u5173\u952e\u77ed\u677f\u3002<\/p>\n\n\n\n \u7814\u7a76\u5458\u4eec\u8868\u793a\uff0cRHELM \u76ee\u524d\u4e5f\u5b58\u5728\u4e00\u5b9a\u7684\u5c40\u9650\u6027\u3002\u4f8b\u5982\u5f53\u524d\u57fa\u51c6\u4ecd\u4ee5\u6587\u672c\u7c7b\u5916\u90e8\u6e90\u4e3a\u4e3b\uff0c\u5c1a\u672a\u8986\u76d6\u89c6\u9891\u3001\u97f3\u9891\u3001\u5de5\u5177\u8c03\u7528\u7b49\u590d\u6742\u591a\u6a21\u6001\uff1b\u540c\u65f6\u7531\u4e8e\u753b\u50cf\u79cd\u5b50\u6765\u81ea PersonaHub \u7684\u7cbe\u82f1\u5b50\u96c6\uff0c\u6570\u636e\u96c6\u5728\u804c\u4e1a\u4e0e\u6559\u80b2\u80cc\u666f\u4e0a\u53ef\u80fd\u5b58\u5728\u4e00\u5b9a\u7684\u504f\u7f6e\u3002\u4f46\u6b63\u662f\u8fd9\u4e9b\u672a\u7adf\u4e4b\u529f\uff0c\u4e3a\u5f00\u6e90\u793e\u533a\u7559\u4e0b\u4e86\u6e05\u6670\u7684\u5ef6\u5c55\u7a7a\u95f4\u3002<\/p>\n\n\n\n \u5982\u679c\u8bf4”\u4e0a\u4e0b\u6587\u7a97\u53e3”\u89e3\u51b3\u7684\u662f\u5927\u6a21\u578b”\u770b\u5f97\u6709\u591a\u8fdc”\uff0c\u90a3\u4e48”\u957f\u671f\u8bb0\u5fc6”\u51b3\u5b9a\u7684\u5c31\u662f\u5927\u6a21\u578b”\u61c2\u4f60\u6709\u591a\u6df1”\u3002RHELM \u628a\u201c\u957f\u671f\u8bb0\u5fc6\u201d\u8fd9\u4ef6\u4e8b\u62c6\u5f97\u8db3\u591f\u7ec6\u3001\u505a\u5f97\u8db3\u591f\u771f\u2014\u2014\u5b83\u65e2\u662f\u4e00\u9762\u7167\u51fa\u884c\u4e1a\u73b0\u72b6\u7684\u955c\u5b50\uff0c\u4e5f\u662f\u4e00\u5f20\u6307\u5f15\u672a\u6765\u7684\u8def\u7ebf\u56fe\u3002<\/p>\n\n\n\n \u63a5\u4e0b\u6765\u503c\u5f97\u671f\u5f85\u7684\uff0c\u7edd\u4e0d\u4ec5\u4ec5\u662f\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u6216\u66f4\u5f3a\u7684\u68c0\u7d22\u7b97\u6cd5\uff0c\u800c\u662f\u771f\u6b63\u80fd\u50cf\u4eba\u4e00\u6837\uff0c\u53bb\u4e3b\u52a8\u7d2f\u79ef\u3001\u6f14\u5316\u3001\u4fee\u526a\u5e76\u667a\u80fd\u8c03\u7528\u201c\u8bb0\u5fc6\u201d\u7684\u4e0b\u4e00\u4ee3 AI \u52a9\u7406\u3002<\/p>\n\n\n\n Beyond Static Dialogues: Benchmarking Realistic, Heterogeneous, and Evolving Long-Term Memory<\/strong><\/p>\n\n\n\n \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.31086 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n \u9879\u76ee\u5730\u5740\uff1ahttps:\/\/microsoft.github.io\/RHELM (opens in new tab)<\/span><\/a><\/p>\n\n\n\n \u8bc4\u6d4b\u4ee3\u7801\uff1ahttps:\/\/github.com\/microsoft\/RHELM (opens in new tab)<\/span><\/a><\/p>\n\n\n\n\u4e3a\u4ec0\u4e48\u73b0\u6709\u7684\u201c\u8bb0\u5fc6\u8bc4\u6d4b\u201d\u4e0d\u591f\u7528\uff1f<\/h2>\n\n\n\n
RHELM\uff1a\u4ece\u4eba\u51fa\u53d1\uff0c\u6784\u9020\u4e00\u5e74\u7684\u8f68\u8ff9<\/h2>\n\n\n\n
\n
\n

\n


\u4e09\u7c7b\u8bb0\u5fc6\u8303\u5f0f\u7684\u5168\u9762\u5bf9\u6bd4\u4e0e\u6df1\u5c42\u8bca\u65ad<\/h2>\n\n\n\n
\n



\u628a”\u8bb0\u5fc6”\u91cd\u65b0\u5bf9\u9f50\u5230\u771f\u5b9e\u4e16\u754c<\/h2>\n\n\n\n