{"id":1272,"date":"2025-07-07T20:16:00","date_gmt":"2025-07-07T12:16:00","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=1272"},"modified":"2025-09-17T20:18:03","modified_gmt":"2025-09-17T12:18:03","slug":"agentic-deep-research%e6%96%b0%e8%8c%83%e5%bc%8f%ef%bc%8c%e6%8e%a8%e7%90%86%e8%83%bd%e5%8a%9b%e5%86%8d%e7%aa%81%e7%a0%b4%ef%bc%8c%e5%8f%af%e4%bf%a1%e5%ba%a6%e5%a2%9e%e5%8a%a0","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/07\/07\/agentic-deep-research%e6%96%b0%e8%8c%83%e5%bc%8f%ef%bc%8c%e6%8e%a8%e7%90%86%e8%83%bd%e5%8a%9b%e5%86%8d%e7%aa%81%e7%a0%b4%ef%bc%8c%e5%8f%af%e4%bf%a1%e5%ba%a6%e5%a2%9e%e5%8a%a0\/","title":{"rendered":"Agentic Deep Research\u65b0\u8303\u5f0f\uff0c\u63a8\u7406\u80fd\u529b\u518d\u7a81\u7834\uff0c\u53ef\u4fe1\u5ea6\u589e\u52a0"},"content":{"rendered":"\n<p>\u5c3d\u7ba1 LLM \u7684\u80fd\u529b\u4e0e\u65e5\u4ff1\u589e\uff0c\u4f46\u5176\u5728\u590d\u6742\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u4ecd\u53d7\u9650\u4e8e\u9759\u6001\u7684\u5185\u90e8\u77e5\u8bc6\u3002\u4e3a\u4ece\u6839\u672c\u4e0a\u89e3\u51b3\u8fd9\u4e00\u9650\u5236\uff0c\u7a81\u7834 AI \u80fd\u529b\u754c\u9650\uff0c\u4e1a\u754c\u7814\u7a76\u8005\u4eec\u63d0\u51fa\u4e86 Agentic Deep Research \u7cfb\u7edf\uff0c\u5728\u8be5\u7cfb\u7edf\u4e2d\u57fa\u4e8e LLM \u7684 Agent \u901a\u8fc7\u81ea\u4e3b\u63a8\u7406\u3001\u8c03\u7528\u641c\u7d22\u5f15\u64ce\u548c\u8fed\u4ee3\u5730\u6574\u5408\u4fe1\u606f\u6765\u7ed9\u51fa\u5168\u9762\u3001\u6709\u6df1\u5ea6\u4e14\u6b63\u786e\u6027\u6709\u4fdd\u969c\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<p>OpenAI \u548c Google \u7684\u7814\u7a76\u8005\u4eec\u603b\u7ed3\u4e86 Agentic Deep Researcher \u7684\u51e0\u5927\u4f18\u52bf\uff1a\uff081\uff09\u6df1\u5165\u7684\u95ee\u9898\u7406\u89e3\u80fd\u529b\uff08Comprehensive Understanding\uff09\uff1a\u80fd\u591f\u5904\u7406\u590d\u6742\u3001\u591a\u8df3\u7684\u7528\u6237\u63d0\u95ee\uff1b\uff082\uff09\u5f3a\u5927\u7684\u4fe1\u606f\u6574\u5408\u80fd\u529b\uff08Enhanced Synthesis\uff09\uff1a\u80fd\u591f\u5c06\u5e7f\u6cdb\u751a\u81f3\u51b2\u7a81\u7684\u4fe1\u606f\u6e90\u6574\u5408\u4e3a\u5408\u7406\u7684\u8f93\u51fa\uff1b\uff083\uff09\u51cf\u8f7b\u7528\u6237\u7684\u8ba4\u77e5\u8d1f\u62c5\uff08Reduced User Effort\uff09\uff1a\u6574\u4e2a research \u8fc7\u7a0b\u5b8c\u5168\u81ea\u4e3b\uff0c\u4e0d\u9700\u8981\u7528\u6237\u7684\u8fc7\u591a\u5e72\u9884\u3002<\/p>\n\n\n\n<p>\u73b0\u5b58\u6700\u5148\u8fdb\u7684 Agentic Deep Research \u7cfb\u7edf\u5f80\u5f80\u57fa\u4e8e\u7531\u53ef\u9a8c\u8bc1\u7ed3\u679c\u5956\u52b1\u6307\u5bfc\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0c\u5c3d\u7ba1\u8be5\u8bad\u7ec3\u8303\u5f0f\u5e26\u6765\u4e86\u663e\u8457\u7684\u6027\u80fd\u6536\u76ca\uff0c\u4f46\u4ecd\u5b58\u5728\u4ee5\u4e0b\u6838\u5fc3\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u68af\u5ea6\u51b2\u7a81\uff08Gradients Conflicts\uff09\uff1a\u5728\u57fa\u4e8e\u53ef\u9a8c\u8bc1\u7ed3\u679c\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\u8303\u5f0f\u4e2d\uff0c\u5373\u4f7f\u4e2d\u95f4\u7684\u63a8\u7406\u8fc7\u7a0b\u6216\u7814\u7a76\u7b56\u7565\u662f\u6709\u6548\u7684\uff0c\u53ea\u8981\u6700\u7ec8\u7b54\u6848\u9519\u8bef\uff0c\u6574\u4e2a\u63a8\u7406\u8f68\u8ff9\u90fd\u4f1a\u53d7\u5230\u60e9\u7f5a\u3002\u8fd9\u79cd\u7c97\u7c92\u5ea6\u7684\u5956\u52b1\u8bbe\u8ba1\u5728\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u4e0e\u6700\u7ec8\u7b54\u6848\u4e4b\u95f4\u5f15\u5165\u4e86\u6f5c\u5728\u7684\u68af\u5ea6\u51b2\u7a81\uff0c\u963b\u788d\u4e86\u6a21\u578b\u53d1\u73b0\u66f4\u4f18\u7684\u63a8\u7406\u80fd\u529b\u548c\u7814\u7a76\u7b56\u7565\uff0c\u4ece\u800c\u9650\u5236\u4e86\u5176\u6cdb\u5316\u80fd\u529b<\/li>\n\n\n\n<li>\u5956\u52b1\u7a00\u758f\uff08Reward sparsity\uff09\uff1a\u57fa\u4e8e\u7ed3\u679c\u7684\u5f3a\u5316\u5b66\u4e60\u4ec5\u4f9d\u8d56\u6700\u7ec8\u7b54\u6848\u751f\u6210\u5956\u52b1\uff0c\u5bfc\u81f4\u6bcf\u4e2a\u8bad\u7ec3\u6837\u672c\u53ea\u80fd\u63d0\u4f9b\u7a00\u758f\u7684\u53cd\u9988\u4fe1\u53f7\u3002\u8fd9\u4e25\u91cd\u9650\u5236\u4e86\u7b56\u7565\u4f18\u5316\u7684\u6548\u7387\uff0c\u56e0\u4e3a\u5b83\u589e\u52a0\u4e86\u5bf9\u66f4\u5927\u89c4\u6a21\u8bad\u7ec3\u6570\u636e\u548c\u66f4\u957f\u8bad\u7ec3\u5468\u671f\u7684\u4f9d\u8d56\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4ee5\u4e0a\u4e24\u4e2a\u9650\u5236\u9650\u5236\u4e86 Agentic Deep Research \u7cfb\u7edf\u7684\u6027\u80fd\u4e0a\u7ebf\uff0c\u4e3a\u51b3\u89e3\u8fd9\u4e24\u5927\u9650\u5236\uff0c\u6765\u81ea\u8682\u8681\u5b89\u5168\u4e0e\u667a\u80fd\u5b9e\u9a8c\u5ba4\u56e2\u961f\u63d0\u51fa\u4e86 Atom-Searcher\uff0c\u8fdb\u4e00\u6b65\u63a8\u52a8\u4e86 Agentic Deep Research \u7cfb\u7edf\u7684\u6027\u80fd\u8fb9\u754c\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bba\u6587\u6807\u9898\uff1aAtom-Searcher: Enhancing Agentic Deep Research via Fine-Grained Atomic Thought Reward<\/li>\n\n\n\n<li>\u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/2508.12800<\/li>\n\n\n\n<li>Github: https:\/\/github.com\/antgroup\/Research-Venus<\/li>\n\n\n\n<li>Huggingface: https:\/\/huggingface.co\/dikw\/Atom-Searcher<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"442\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-479-1024x442.jpg\" alt=\"\" class=\"wp-image-1273\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-479-1024x442.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-479-300x129.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-479-768x331.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-479.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u65b9\u6cd5\u4ecb\u7ecd<\/strong><\/p>\n\n\n\n<p>\u672c\u7814\u7a76\u63d0\u51fa\u4e86\u4e00\u79cd\u521b\u65b0\u6027\u7684 Agentic Deep Research \u7cfb\u7edf\u8bad\u7ec3\u6846\u67b6 Atom-Searcher\uff0c\u7ed3\u5408\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u4e0e\u57fa\u4e8e\u7ec6\u7c92\u5ea6\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\u6784\u5efa\u5f3a\u5927\u7684 Agentic Deep Research \u7cfb\u7edf\u3002<\/p>\n\n\n\n<p>\u4e0e\u73b0\u5b58 Agentic Deep Research \u8bad\u7ec3\u6846\u67b6\u76f8\u6bd4\uff0cAtom-Searcher \u521b\u65b0\u5730\u63d0\u51fa\u4e86 Atomic Thought \u63a8\u7406\u8303\u5f0f\uff0c\u5f15\u5bfc LLM \u8fdb\u884c\u66f4\u52a0\u6df1\u5165\u3001\u53ef\u4fe1\u548c\u53ef\u89e3\u91ca\u7684\u63a8\u7406\uff1b\u7136\u540e\u5f15\u5165 Reasoning Reward Model\uff08RRM\uff09\u5bf9 Atomic Thought \u5f0f\u7684\u63a8\u7406\u8fc7\u7a0b\u8fdb\u884c\u76d1\u7763\uff0c\u6784\u5efa\u7ec6\u7c92\u5ea6\u7684 Atomic Thought Reward\uff08ATR\uff09\uff1b\u8fdb\u800c\u63d0\u51fa\u4e00\u79cd\u8bfe\u7a0b\u5b66\u4e60\u542f\u53d1\u7684\u5956\u52b1\u878d\u5408\u7b56\u7565\u5c06 ATR \u4e0e\u53ef\u9a8c\u8bc1\u7ed3\u679c\u5956\u52b1\u8fdb\u884c\u805a\u5408\uff1b\u6700\u540e\u57fa\u4e8e\u805a\u5408\u5956\u52b1\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"543\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-483-1024x543.jpg\" alt=\"\" class=\"wp-image-1277\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-483-1024x543.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-483-300x159.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-483-768x408.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-483.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>Atomic Thought \u63a8\u7406\u8303\u5f0f<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"759\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-482-1024x759.jpg\" alt=\"\" class=\"wp-image-1276\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-482-1024x759.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-482-300x222.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-482-768x569.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-482.jpg 1107w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u9488\u5bf9 Agentic Deep Research \u7cfb\u7edf\u4e2d LLM \u751f\u6210\u7684\u63a8\u7406\u8f68\u8ff9\uff08&lt;think&gt;\uff09\u5305\u542b\u8fc7\u591a\u5197\u4f59 tokens \u4e14\u63a8\u7406\u6df1\u5ea6\u6b20\u7f3a\u7684\u95ee\u9898\uff0cAtomic Thought \u8303\u5f0f\u5c06 &lt;think&gt; \u5206\u89e3\u4e3a\u66f4\u52a0\u7ec6\u7c92\u5ea6\u7684 \u201c\u529f\u80fd\u5355\u5143\u201d\uff0c\u5982 &lt;Verification&gt;\u3001&lt;hypothesis &gt; \u7b49\uff0c\u8be5\u8303\u5f0f\u6709\u52a9\u4e8e\u5f15\u5bfc LLM \u7684\u63a8\u7406\u8fc7\u7a0b\u66f4\u52a0\u7b26\u5408\u4eba\u7684\u8ba4\u77e5\u884c\u4e3a\uff0c\u4e14\u9ad8\u5ea6\u6a21\u5757\u5316\u7684\u65b9\u5f0f\u80fd\u5927\u5927\u51cf\u5c11\u65e0\u610f\u4e49 tokens\u3002\u66f4\u8fdb\u4e00\u6b65\uff0c\u4e3a\u6fc0\u53d1 LLM \u81ea\u4e3b\u5c06 &lt; think &gt; \u5206\u89e3\u4e3a Atomic Thoughts\uff08&lt;Verification&gt;\u3001&lt;hypothesis &gt; \u7b49\uff09\u7684\u80fd\u529b\uff0c\u4f5c\u8005\u4eec\u7cbe\u5fc3\u6784\u5efa\u4e86 1000k \u9ad8\u8d28\u91cf Atomic Thought \u6307\u4ee4\u5fae\u8c03\u6570\u636e\uff0c\u5bf9 LLM \u8fdb\u884c SFT\u3002<\/p>\n\n\n\n<p><strong>\u7ec6\u7c92\u5ea6 Atomic Thought Reward \u6784\u5efa<\/strong><\/p>\n\n\n\n<p>\u5728 Agentic Deep Research \u7cfb\u7edf\u4e2d\uff0c\u76f4\u63a5\u4f7f\u7528 Reasoning Reward Model\uff08RRM\uff09\u5bf9\u63a8\u7406\u8fc7\u7a0b\u8fdb\u884c\u76d1\u7763\uff0c\u5f80\u5f80\u56e0\u4e3a &lt; think &gt; \u4e2d\u7684\u4f4e\u4fe1\u566a\u6bd4\uff08\u8fc7\u591a\u5197\u4f59 tokens\uff09\u800c\u6548\u679c\u4e0d\u4f73\u3002\u800c Atomic Thought \u7684\u63d0\u51fa\uff0c\u5f88\u597d\u5730\u89e3\u51b3\u4e86\u8be5\u95ee\u9898\uff0c\u9664\u4e86\u51cf\u5c11\u4e86 &lt; think &gt; \u4e2d\u7684\u5197\u4f59 tokens\uff0cAtomic Thoughts\uff08&lt;Verification&gt;\u3001&lt;hypothesis &gt; \u7b49\uff09\u8fd8\u4e3a RRM \u63d0\u4f9b\u4e86\u76d1\u7763\u951a\u70b9\uff0c\u6e05\u6670\u7684\u6a21\u5757\u5316\u7ed3\u6784\u4f7f\u5f97 RRM \u80fd\u591f\u51c6\u786e\u5730\u8bc4\u4f30\u6bcf\u4e2a\u529f\u80fd\u5355\u5143\u7684\u8d28\u91cf\u3002\u56e0\u6b64\uff0c\u4f5c\u8005\u4eec\u5f15\u5165 RRM \u5bf9 Atom-Thoughts \u8fdb\u884c\u76d1\u7763\uff0c\u4ece\u800c\u5f97\u5230\u7ec6\u7c92\u5ea6\u7684 Atomic Thought Reward\uff0c\u7528\u4e8e\u7f13\u89e3\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e2d\u7684\u68af\u5ea6\u51b2\u7a81\u548c\u5956\u52b1\u7a00\u758f\u95ee\u9898\u3002<\/p>\n\n\n\n<p><strong>\u8bfe\u7a0b\u5b66\u4e60\u542f\u53d1\u7684\u5956\u52b1\u805a\u5408\u7b56\u7565<\/strong><\/p>\n\n\n\n<p>\u57fa\u4e8e\u53ef\u9a8c\u8bc1\u7ed3\u679c\u7684\u5956\u52b1\u7684 Agentic Deep Research \u7cfb\u7edf\u4e4b\u6240\u4ee5\u5b58\u5728\u68af\u5ea6\u51b2\u7a81\u95ee\u9898\uff0c\u662f\u7531\u4e8e\u57fa\u4e8e\u7ed3\u679c\u7684\u5956\u52b1\u5728 token \u7ea7\u522b\u7684\u5956\u52b1\u5206\u914d\u4e0a\u8fc7\u4e8e\u7c97\u7cd9\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b83\u5c06\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u7684\u6b63\u786e\u6027\u5b8c\u5168\u5f52\u56e0\u4e8e\u6700\u7ec8\u7b54\u6848\uff0c\u5e38\u5e38\u5728\u4e0d\u8003\u8651\u5404\u6b65\u9aa4\u5b9e\u9645\u8d21\u732e\u7684\u60c5\u51b5\u4e0b\u5bf9\u5176\u8fdb\u884c\u5956\u52b1\u6216\u60e9\u7f5a\u3002\u8fd9\u79cd\u9519\u4f4d\u5728\u4f18\u5316\u8fc7\u7a0b\u4e2d\u4f1a\u5f15\u53d1\u68af\u5ea6\u51b2\u7a81\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u5c06 ATR \u4e0e\u7ed3\u679c\u5956\u52b1\u76f8\u7ed3\u5408\uff0c\u5229\u7528 ATR \u4f5c\u4e3a\u8f85\u52a9\u4fe1\u53f7\u6765\u6821\u51c6\u7ed3\u679c\u5956\u52b1\uff0c\u4ece\u800c\u7f13\u89e3\u68af\u5ea6\u51b2\u7a81\u3002<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u4f7f\u7528\u9759\u6001\u7684\u5956\u52b1\u52a0\u6743\u7cfb\u6570\u65e0\u6cd5\u4e0e\u8bad\u7ec3\u52a8\u6001\u4fdd\u6301\u4e00\u81f4\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5728\u8bad\u7ec3\u521d\u671f\uff0c\u6a21\u578b\u80fd\u529b\u5c1a\u6709\u9650\uff0c\u96be\u4ee5\u751f\u6210\u5b8c\u5168\u6b63\u786e\u7684\u7b54\u6848\uff0c\u4f46\u66f4\u6709\u53ef\u80fd\u63a2\u7d22\u51fa\u5bf9\u6700\u7ec8\u6b63\u786e\u89e3\u6709\u8d21\u732e\u7684\u6709\u7528 \u201c\u539f\u5b50\u601d\u7ef4\u201d\u3002\u5982\u679c\u6b64\u9636\u6bb5\u4ec5\u4f9d\u8d56\u57fa\u4e8e\u7ed3\u679c\u7684\u5956\u52b1\uff0c\u8fd9\u4e9b\u6709\u76ca\u7684\u539f\u5b50\u601d\u7ef4\u53ef\u80fd\u56e0\u6700\u7ec8\u7b54\u6848\u9519\u8bef\u800c\u906d\u5230\u4e0d\u516c\u6b63\u7684\u60e9\u7f5a\uff1b\u76f8\u53cd\uff0c\u4e00\u4e9b\u6709\u5bb3\u7684\u539f\u5b50\u601d\u7ef4\u4e5f\u53ef\u80fd\u88ab\u9519\u8bef\u5730\u5f3a\u5316\uff0c\u5bfc\u81f4\u4e25\u91cd\u7684\u68af\u5ea6\u51b2\u7a81\uff0c\u56e0\u800c\u9700\u8981 ATR \u8fdb\u884c\u8f83\u5f3a\u7684\u6821\u51c6\u3002\u968f\u7740\u8bad\u7ec3\u7684\u63a8\u8fdb\uff0c\u6a21\u578b\u80fd\u529b\u9010\u6b65\u63d0\u5347\uff0c\u5176\u63a8\u7406\u8f68\u8ff9\u4e0e\u6b63\u786e\u7b54\u6848\u7684\u5bf9\u9f50\u7a0b\u5ea6\u4e5f\u65e5\u76ca\u63d0\u9ad8\u3002\u56e0\u6b64\uff0c\u68af\u5ea6\u51b2\u7a81\u9010\u6e10\u51cf\u5f31\uff0c\u800c\u6765\u81ea ATR \u7684\u8fc7\u5ea6\u6821\u51c6\u53ef\u80fd\u4f1a\u5f15\u5165\u4e0d\u5fc5\u8981\u7684\u566a\u58f0\uff0c\u53cd\u800c\u635f\u5bb3\u6700\u7ec8\u7684\u51c6\u786e\u6027\u3002<\/p>\n\n\n\n<p><strong>\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3<\/strong><\/p>\n\n\n\n<p>\u57fa\u4e8e\u6df7\u5408\u5956\u52b1\uff0c\u672c\u6587\u91c7\u7528\u4e86 GRPO \u7b97\u6cd5\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002\u5e76\u4f7f\u7528\u4e86 Loss Masking \u7b56\u7565\u4fdd\u8bc1\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5728\u539f\u59cb\u7684 GRPO \u6846\u67b6\u4e2d\uff0c\u635f\u5931\u51fd\u6570\u4f1a\u8ba1\u7b97\u6574\u4e2a\u63a8\u7406\u8def\u5f84\u4e2d\u6240\u6709 token \u7684\u68af\u5ea6\u3002\u4f46\u5728 Atom-Searcher \u4e2d\uff0c\u6a21\u578b\u7684\u8f93\u51fa\u8def\u5f84\u5305\u542b\u7531\u5916\u90e8\u73af\u5883\u68c0\u7d22\u5f97\u5230\u7684\u5185\u5bb9\uff08\u5982\u641c\u7d22\u7ed3\u679c\uff09\uff0c\u8fd9\u4e9b\u5185\u5bb9\u4e0d\u662f\u6a21\u578b\u751f\u6210\u7684\uff0c\u4e5f\u4e0d\u53ef\u8bad\u7ec3\u3002\u4e3a\u4e86\u907f\u514d\u6a21\u578b\u5728\u8bad\u7ec3\u65f6\u88ab\u8fd9\u4e9b\u9759\u6001\u3001\u4e0d\u53ef\u63a7\u7684\u5185\u5bb9\u8bef\u5bfc\uff0c\u672c\u6587\u91c7\u7528\u4e86 Loss Masking \u673a\u5236\uff0c\u5c06\u68c0\u7d22\u7ed3\u679c\u90e8\u5206\u7684 token \u6392\u9664\u5728\u635f\u5931\u8ba1\u7b97\u4e4b\u5916\u3002<\/p>\n\n\n\n<p><strong>\u5b9e\u9a8c\u6548\u679c<\/strong><\/p>\n\n\n\n<p><strong>\u4e3b\u5b9e\u9a8c<\/strong><\/p>\n\n\n\n<p>Atom-Searcher \u5728 In-Domain \u548c Out-of-Domain \u4e0a\u7684\u6027\u80fd\u8868\u73b0\u5747\u5341\u5206\u4eae\u773c\u3002\u5728 In-Domain Benchmarks \uff08NQ\u3001 TQ\u3001HotpotQA\u30012Wiki\uff09\u4e0a Atom-Searcher \u76f8\u8f83\u4e8e\u6700\u4f18 baseline\u2014\u2014DeepResearcher \u53d6\u5f97\u4e86 8.5% \u7684\u5e73\u5747\u6027\u80fd\u63d0\u5347\uff0c\u5728 Out-of-Domain Benchmarks\uff08Musique\u3001 Bamboogle\u3001 PopQA\uff09\u4e0a Atom-Searcher \u76f8\u8f83\u4e8e\u6700\u4f18 baseline\u2014\u2014DeepResearcher \u53d6\u5f97\u4e86 2.5% \u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"337\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-484-1024x337.jpg\" alt=\"\" class=\"wp-image-1278\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-484-1024x337.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-484-300x99.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-484-768x253.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-484.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u6d88\u878d\u5b9e\u9a8c<\/strong><\/p>\n\n\n\n<p>\u4f5c\u8005\u4eec\u8bc1\u660e\u4e86 Atom-Searcher \u4e2d Atomic Thought \u8303\u5f0f\u548c ATR \u7684\u8d21\u732e\uff0c\u5e76\u8bc1\u660e\u4e86\u76f8\u8f83\u4e8e\u4f20\u7edf\u7684 &lt; think &gt; \u63a8\u7406\u8303\u5f0f Atomic Thought \u8303\u5f0f\u4e3a RRM \u63d0\u4f9b\u4e86\u6709\u6548\u7684\u76d1\u7763\u951a\u70b9\uff0c\u4ece\u800c\u5e26\u6765\u4e86\u6027\u80fd\u63d0\u5347<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"192\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-481-1024x192.jpg\" alt=\"\" class=\"wp-image-1275\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-481-1024x192.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-481-300x56.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-481-768x144.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-481.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u6848\u4f8b\u5206\u6790<\/strong><\/p>\n\n\n\n<p>\u4f5c\u8005\u4eec\u901a\u8fc7\u6848\u4f8b\u5206\u6790\u5bf9\u6bd4\u4e86 Atom-Searcher \u4e0e\u6700\u4f18 baseline\u2014\u2014DeepResearcher \u7684\u63a8\u7406\u8fc7\u7a0b\u3002\u5c55\u793a\u4e86 Atom-Searcher \u7684\u4f18\u52bf\uff1a\uff081\uff09Atom-Searcher \u5728\u5176\u63a8\u7406\u8fc7\u7a0b\u4e2d\u81ea\u4e3b\u751f\u6210\u4e86 Atomic Thoughts\uff0c\u5c55\u73b0\u51fa\u66f4\u63a5\u8fd1\u4eba\u7c7b\u7684\u8ba4\u77e5\u884c\u4e3a\uff0c\u4f8b\u5982\u95ee\u9898\u5206\u6790\u3001\u63d0\u51fa\u89e3\u51b3\u65b9\u6848\u5047\u8bbe\u3001\u9884\u6d4b\u9519\u8bef\u4ee5\u53ca\u89c4\u5212\u4e0b\u4e00\u6b65\u64cd\u4f5c\uff0c\u4f7f\u5176\u63a8\u7406\u8fc7\u7a0b\u66f4\u52a0\u6df1\u5165\u4e14\u6e05\u6670\uff1b\uff082\uff09Atom-Searcher \u4f1a\u89e6\u53d1\u66f4\u591a\u7684\u641c\u7d22\u8c03\u7528\uff0c\u4ece\u800c\u83b7\u53d6\u66f4\u4e30\u5bcc\u7684\u5916\u90e8\u4fe1\u606f\uff0c\u4ee5\u786e\u4fdd\u7b54\u6848\u7684\u6b63\u786e\u6027\u3002\u8fd9\u4e9b\u4f18\u52bf\u8868\u660e\uff0cAtom-Searcher \u5728\u66f4\u590d\u6742\u7684 Deep Research \u4efb\u52a1\u4e2d\u5177\u6709\u5de8\u5927\u6f5c\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"849\" height=\"1024\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-485-849x1024.jpg\" alt=\"\" class=\"wp-image-1279\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-485-849x1024.jpg 849w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-485-249x300.jpg 249w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-485-768x926.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-485.jpg 1108w\" sizes=\"auto, (max-width: 849px) 100vw, 849px\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u5c3d\u7ba1 LLM \u7684\u80fd\u529b\u4e0e\u65e5\u4ff1\u589e\uff0c\u4f46\u5176\u5728\u590d\u6742\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u4ecd\u53d7\u9650\u4e8e\u9759\u6001\u7684\u5185\u90e8\u77e5\u8bc6\u3002\u4e3a\u4ece\u6839\u672c\u4e0a\u89e3\u51b3\u8fd9\u4e00\u9650\u5236\uff0c\u7a81\u7834 AI [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":1274,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[],"class_list":["post-1272","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-10"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1272","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=1272"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1272\/revisions"}],"predecessor-version":[{"id":1280,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1272\/revisions\/1280"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/1274"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=1272"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=1272"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=1272"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}