{"id":1292,"date":"2026-04-22T18:46:38","date_gmt":"2026-04-22T10:46:38","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1292"},"modified":"2026-04-22T18:48:10","modified_gmt":"2026-04-22T10:48:10","slug":"enhancing-llm-based-search-agents-via-contribution-weighted-group-relative-policy-optimization","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1292","title":{"rendered":"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\"><strong>1. \u6458\u8981\uff08Abstract\uff09<\/strong><\/h2>\n\n\n\n<p class=\"has-text-align-left\">\u672c\u6587\u805a\u7126\u4e8e LLM-based Search Agent \u7684\u8bad\u7ec3\u95ee\u9898\u3002\u73b0\u6709\u65b9\u6cd5\u5728\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e2d\u4e3b\u8981\u9762\u4e34\u4e00\u4e2a\u6838\u5fc3\u96be\u70b9credit assignment\uff08\u8d21\u732e\u5f52\u56e0\uff09\u3002\u4e00\u65b9\u9762\uff0c\u57fa\u4e8e\u6700\u7ec8\u7b54\u6848\u7684 outcome supervision \u867d\u7136\u8bad\u7ec3\u7a33\u5b9a\uff0c\u4f46\u5956\u52b1\u4fe1\u53f7\u8fc7\u4e8e\u7a00\u758f\uff0c\u65e0\u6cd5\u533a\u5206\u4e0d\u540c\u641c\u7d22\u8f6e\u6b21\u7684\u91cd\u8981\u6027\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u57fa\u4e8e\u4e2d\u95f4\u6b65\u9aa4\u7684 process supervision \u867d\u80fd\u63d0\u4f9b\u66f4\u7ec6\u7c92\u5ea6\u4fe1\u53f7\uff0c\u4f46\u901a\u5e38\u4f9d\u8d56\u4e0d\u7a33\u5b9a\u7684\u4ef7\u503c\u4f30\u8ba1\u6216\u566a\u58f0\u8f83\u5927\u7684\u4e2d\u95f4\u5956\u52b1\uff0c\u5bb9\u6613\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u3002\u00a0<\/p>\n\n\n\n<p>\u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u8bba\u6587\u63d0\u51fa Contribution-Weighted GRPO\uff08CW-GRPO\uff09\u3002\u8be5\u65b9\u6cd5\u5728\u4fdd\u7559 GRPO \u7a33\u5b9a\u6027\u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165\u4e00\u4e2a LLM judge \u5bf9\u6bcf\u4e00\u8f6e\u641c\u7d22\u8fdb\u884c\u8bc4\u4f30\uff0c\u4ece\u201c\u68c0\u7d22\u662f\u5426\u6709\u6548\u201d\u548c\u201c\u63a8\u7406\u662f\u5426\u6b63\u786e\u201d\u4e24\u4e2a\u7ef4\u5ea6\u751f\u6210\u8d21\u732e\u4fe1\u53f7\uff0c\u5e76\u636e\u6b64\u5bf9\u6574\u6761\u8f68\u8ff9\u7684\u4f18\u52bf\u8fdb\u884c\u91cd\u5206\u914d\u3002\u4e0d\u540c\u4e8e\u76f4\u63a5\u4f18\u5316\u8fc7\u7a0b\u5956\u52b1\uff0cCW-GRPO \u5c06\u8fc7\u7a0b\u4fe1\u606f\u7528\u4e8e\u8c03\u8282\u6700\u7ec8\u5956\u52b1\u7684\u5206\u5e03\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u7a33\u5b9a\u7684\u7ec6\u7c92\u5ea6 credit assignment\u3002\u00a0<\/p>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u8be5\u65b9\u6cd5\u5728\u591a\u4e2a\u77e5\u8bc6\u5bc6\u96c6\u578b\u95ee\u7b54\u4efb\u52a1\u4e0a\u5747\u4f18\u4e8e\u73b0\u6709\u65b9\u6cd5\uff0c\u5728 Qwen3-8B \u4e0a\u63d0\u5347 5.0%\uff0c\u5728 Qwen3-1.7B \u4e0a\u63d0\u5347 6.3%\u3002\u8fdb\u4e00\u6b65\u5206\u6790\u53d1\u73b0\uff0c\u6210\u529f\u8f68\u8ff9\u4e2d\u7684\u6709\u6548\u8d21\u732e\u901a\u5e38\u96c6\u4e2d\u5728\u5c11\u6570\u5173\u952e\u641c\u7d22\u8f6e\u6b21\uff0c\u800c\u975e\u5747\u5300\u5206\u5e03\u3002\u6574\u4f53\u6765\u770b\uff0c\u8be5\u5de5\u4f5c\u4e0d\u4ec5\u63d0\u51fa\u4e86\u4e00\u79cd\u6709\u6548\u7684\u8bad\u7ec3\u65b9\u6cd5\uff0c\u4e5f\u63ed\u793a\u4e86\u641c\u7d22\u578b\u667a\u80fd\u4f53\u4e2d\u7684\u91cd\u8981\u7ed3\u6784\u6027\u7279\u5f81\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"794\" height=\"666\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-4.png\"  class=\"wp-image-1293\" style=\"width:386px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-4.png 794w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-4-300x252.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-4-768x644.png 768w\" sizes=\"auto, (max-width: 794px) 100vw, 794px\" title=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe\" alt=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>2. \u7814\u7a76\u80cc\u666f\u4e0e\u95ee\u9898\u52a8\u673a\uff08Introduction\uff09<\/strong><\/h2>\n\n\n\n<p>\u968f\u7740\u5927\u6a21\u578b\u80fd\u529b\u7684\u53d1\u5c55\uff0c\u4ec5\u4f9d\u8d56\u53c2\u6570\u77e5\u8bc6\u5df2\u96be\u4ee5\u6ee1\u8db3\u5b9e\u65f6\u4fe1\u606f\u548c\u590d\u6742\u77e5\u8bc6\u4efb\u52a1\u7684\u9700\u6c42\u3002\u5f15\u5165\u641c\u7d22\u80fd\u529b\u540e\uff0c\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u591a\u8f6e\u68c0\u7d22\u4e0e\u63a8\u7406\u83b7\u53d6\u5916\u90e8\u4fe1\u606f\uff0c\u4ece\u800c\u663e\u8457\u63d0\u5347\u4e8b\u5b9e\u6027\u548c\u590d\u6742\u63a8\u7406\u80fd\u529b\u3002\u56e0\u6b64\uff0csearch agent \u6210\u4e3a\u5f53\u524d\u5927\u6a21\u578b\u7684\u91cd\u8981\u53d1\u5c55\u65b9\u5411\u3002&nbsp;<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u591a\u8f6e\u641c\u7d22\u4e5f\u5e26\u6765\u4e86\u65b0\u7684\u8bad\u7ec3\u96be\u70b9\u3002\u4e00\u4e2a\u5178\u578b\u95ee\u9898\u662f\u6a21\u578b\u6700\u7ec8\u7b54\u5bf9\uff0c\u5e76\u4e0d\u610f\u5473\u7740\u6bcf\u4e00\u8f6e\u641c\u7d22\u90fd\u540c\u6837\u91cd\u8981\u3002\u6709\u4e9b\u8f6e\u6b21\u63d0\u4f9b\u4e86\u5173\u952e\u8bc1\u636e\uff0c\u6709\u4e9b\u53ea\u662f\u91cd\u590d\u6216\u65e0\u6548\u5c1d\u8bd5\u3002\u4f46\u73b0\u6709\u65b9\u6cd5\u5f80\u5f80\u65e0\u6cd5\u533a\u5206\u8fd9\u4e9b\u5dee\u5f02\u3002<\/p>\n\n\n\n<p>\u5f53\u524d\u4e3b\u6d41\u65b9\u6cd5\u4e3b\u8981\u6709\u4e24\u7c7b\uff1a<\/p>\n\n\n\n<p>* Outcome supervision\u53ea\u4f9d\u636e\u6700\u7ec8\u7b54\u6848\u6253\u5206\uff0c\u7a33\u5b9a\u4f46\u8fc7\u4e8e\u7c97\u7c92\u5ea6<\/p>\n\n\n\n<p>* Process supervision\u5bf9\u4e2d\u95f4\u6b65\u9aa4\u6253\u5206\uff0c\u4f46\u4f9d\u8d56\u989d\u5916\u6a21\u578b\u6216\u566a\u58f0\u8f83\u5927<\/p>\n\n\n\n<p>\u4e24\u8005\u5206\u522b\u5728\u201c\u7a33\u5b9a\u6027\u201d\u548c\u201c\u7cbe\u7ec6\u5ea6\u201d\u4e4b\u95f4\u5b58\u5728\u660e\u663e\u6743\u8861\u3002&nbsp;<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u8bba\u6587\u7684\u6838\u5fc3\u52a8\u673a\u662f\u80fd\u5426\u5728\u4e0d\u7834\u574f\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u524d\u63d0\u4e0b\uff0c\u5f15\u5165\u4e2d\u95f4\u8fc7\u7a0b\u4fe1\u606f\uff0c\u5b9e\u73b0\u66f4\u5408\u7406\u7684 credit assignment\uff1f<\/p>\n\n\n\n<p>CW-GRPO \u7684\u63d0\u51fa\uff0c\u6b63\u662f\u5bf9\u8fd9\u4e00\u95ee\u9898\u7684\u76f4\u63a5\u56de\u5e94\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"570\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-5-1024x570.png\"  class=\"wp-image-1294\" style=\"aspect-ratio:1.797332085186052;width:550px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-5-1024x570.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-5-300x167.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-5-768x427.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-5.png 1330w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe1\" alt=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>3. \u7cfb\u7edf\u6846\u67b6\u4e0e\u6574\u4f53\u8bbe\u8ba1\uff08Framework Overview\uff09<\/strong><\/h2>\n\n\n\n<p>CW-GRPO \u7684\u6574\u4f53\u7ed3\u6784\u53ef\u4ee5\u7406\u89e3\u4e3a\u4e09\u90e8\u5206\uff1a<\/p>\n\n\n\n<p>1. GRPO \u4e3b\u4f53\u8d1f\u8d23\u57fa\u4e8e\u6700\u7ec8\u7ed3\u679c\u8ba1\u7b97\u8f68\u8ff9\u7ea7\u4f18\u52bf\uff0c\u4fdd\u8bc1\u8bad\u7ec3\u7a33\u5b9a<\/p>\n\n\n\n<p>2. LLM Judge \u6a21\u5757\u5bf9\u6bcf\u4e00\u8f6e\u641c\u7d22\u8fdb\u884c\u8d28\u91cf\u8bc4\u4f30<\/p>\n\n\n\n<p>3. \u4f18\u52bf\u91cd\u5206\u914d\u673a\u5236\u6839\u636e\u8f6e\u6b21\u8d21\u732e\u91cd\u65b0\u5206\u914d\u5b66\u4e60\u4fe1\u53f7<\/p>\n\n\n\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u95ee\u9898\uff0c\u6a21\u578b\u4f1a\u91c7\u6837\u591a\u6761\u641c\u7d22\u8f68\u8ff9\uff0c\u5e76\u6839\u636e\u6700\u7ec8\u7b54\u6848\u8ba1\u7b97 outcome reward\u3002\u4e0e\u6807\u51c6 GRPO \u4e0d\u540c\uff0cCW-GRPO \u4f1a\u8fdb\u4e00\u6b65\u5206\u6790\u8f68\u8ff9\u5185\u90e8\u7ed3\u6784\uff0c\u5c06\u6574\u4f53\u4f18\u52bf\u62c6\u5206\u5e76\u91cd\u65b0\u5206\u914d\u5230\u5404\u4e2a\u641c\u7d22\u8f6e\u6b21\u3002&nbsp;<\/p>\n\n\n\n<p>\u8fd9\u91cc\u7684\u5173\u952e\u8bbe\u8ba1\u5728\u4e8eLLM judge \u5e76\u4e0d\u76f4\u63a5\u751f\u6210\u8bad\u7ec3\u5956\u52b1\uff0c\u800c\u662f\u4f5c\u4e3a\u201c\u53c2\u8003\u4fe1\u53f7\u201d\uff0c\u7528\u4e8e\u8c03\u8282\u4f18\u52bf\u5206\u5e03\u3002\u8fd9\u79cd\u8bbe\u8ba1\u6709\u6548\u907f\u514d\u4e86\u8fc7\u7a0b\u76d1\u7763\u4e2d\u5e38\u89c1\u7684\u566a\u58f0\u653e\u5927\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u4ece\u672c\u8d28\u4e0a\u770b\uff0cCW-GRPO \u5e76\u6ca1\u6709\u6539\u53d8\u201c\u54ea\u6761\u8f68\u8ff9\u66f4\u597d\u201d\uff0c\u800c\u662f\u6539\u53d8\u4e86\u201c\u8fd9\u6761\u8f68\u8ff9\u4e2d\u54ea\u51e0\u6b65\u66f4\u503c\u5f97\u5b66\u4e60\u201d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>4. \u6838\u5fc3\u65b9\u6cd5\uff08Core Method\uff09<\/strong><\/h2>\n\n\n\n<p>\u8fd9\u7bc7\u8bba\u6587\u7684\u6838\u5fc3\u65b9\u6cd5\u5176\u5b9e\u53ef\u4ee5\u62c6\u6210\u56db\u6b65\u3002\u7b2c\u4e00\u6b65\uff0c\u50cf\u6807\u51c6 GRPO \u4e00\u6837\uff0c\u5148\u5728\u540c\u4e00\u95ee\u9898\u4e0b\u91c7\u6837\u591a\u6761\u8f68\u8ff9\uff0c\u5e76\u6839\u636e\u6700\u7ec8\u7b54\u6848\u5bf9\u9519\u7b97\u51fa\u6bcf\u6761\u8f68\u8ff9\u7684 outcome-level advantage\u3002\u8fd9\u4e00\u6b65\u505a\u7684\u662f\u201c\u6574\u6761\u8f68\u8ff9\u4e4b\u95f4\u8c01\u66f4\u597d\u201d\u7684\u6bd4\u8f83\uff0c\u5b83\u89e3\u51b3\u7684\u662f\u8f68\u8ff9\u7ea7\u522b\u7684\u4f18\u52a3\u95ee\u9898\uff0c\u4f46\u8fd8\u6ca1\u6709\u89e3\u51b3\u8f6e\u6b21\u7ea7\u522b\u7684 credit assignment\u3002&nbsp;<\/p>\n\n\n\n<p>\u7b2c\u4e8c\u6b65\uff0c\u4f5c\u8005\u5f00\u59cb\u5bf9\u6bcf\u4e00\u8f6e\u641c\u7d22\u505a\u201c\u8d21\u732e\u4f30\u8ba1\u201d\u3002\u8fd9\u91cc\u4ed6\u4eec\u6ca1\u6709\u7528\u5355\u4e00\u5206\u6570\uff0c\u800c\u662f\u628a\u8f6e\u6b21\u8d28\u91cf\u62c6\u6210\u4e24\u4e2a\u4e8c\u503c\u4fe1\u53f7\uff1a\u4e00\u4e2a\u53eb retrieval utility\uff0c\u770b\u8fd9\u8f6e\u68c0\u7d22\u5230\u7684\u6587\u6863\u662f\u4e0d\u662f\u63d0\u4f9b\u4e86\u65b0\u7684\u3001\u548c\u4efb\u52a1\u76f8\u5173\u7684\u8bc1\u636e\u3002\u53e6\u4e00\u4e2a\u53eb reasoning correctness\uff0c\u770b\u8fd9\u4e00\u8f6e\u63a8\u7406\u6709\u6ca1\u6709\u6b63\u786e\u7406\u89e3\u5f53\u524d\u4e0a\u4e0b\u6587\uff0c\u5e76\u6cbf\u7740\u5408\u7406\u8def\u5f84\u5f80\u7b54\u6848\u63a8\u8fdb\u3002\u53ea\u6709\u8fd9\u4e24\u4e2a\u6761\u4ef6\u540c\u65f6\u6ee1\u8db3\uff0c\u8fd9\u4e00\u8f6e\u624d\u4f1a\u88ab\u8ba4\u4e3a\u771f\u6b63\u201c\u6709\u8d21\u732e\u201d\u3002\u4f5c\u8005\u91c7\u7528\u7684\u662f\u4e00\u79cd\u5f88\u4fdd\u5b88\u7684\u5408\u53d6\u5f0f\u8bbe\u8ba1\uff0c\u4e5f\u5c31\u662f\u4e24\u8005\u540c\u65f6\u4e3a 1 \u624d\u7b97\u8d21\u732e\u8f6e\u3002\u8fd9\u80cc\u540e\u7684\u903b\u8f91\u5f88\u6734\u7d20\uff0c\u53ea\u641c\u5230\u4fe1\u606f\u4f46\u7406\u89e3\u9519\u4e86\uff0c\u4e0d\u7b97\u771f\u8d21\u732e\u3002\u63a8\u7406\u672c\u8eab\u903b\u8f91\u901a\u987a\u4f46\u6ca1\u6709\u5e26\u6765\u65b0\u4fe1\u606f\uff0c\u4e5f\u4e0d\u7b97\u771f\u63a8\u8fdb\u3002\u00a0<\/p>\n\n\n\n<p>\u7b2c\u4e09\u6b65\uff0c\u662f\u628a\u4e0a\u9762\u7684\u4e8c\u503c\u8d21\u732e\u4fe1\u53f7\u8f6c\u6210\u771f\u6b63\u7684\u6743\u91cd\u3002\u8fd9\u91cc\u4f5c\u8005\u5bf9\u6210\u529f\u8f68\u8ff9\u548c\u5931\u8d25\u8f68\u8ff9\u505a\u4e86\u533a\u5206\u3002\u5bf9\u4e8e\u6210\u529f\u8f68\u8ff9\uff0c\u8bba\u6587\u8ba4\u4e3a\u53ef\u4ee5\u8f83\u53ef\u9760\u5730\u8bc6\u522b\u51fa\u54ea\u4e9b\u8f6e\u6b21\u5bf9\u6700\u7ec8\u6210\u529f\u66f4\u5173\u952e\uff0c\u56e0\u6b64\u4f7f\u7528\u4e00\u4e2a\u7531\u53c2\u6570 \u03b1 \u63a7\u5236\u7684 softmax \u6765\u5f3a\u8c03\u9ad8\u8d21\u732e\u8f6e\u6b21\u3002\u03b1 \u8d8a\u5927\uff0c\u5b66\u4e60\u4fe1\u53f7\u5c31\u8d8a\u96c6\u4e2d\u3002\u5f53 \u03b1 = \u221e \u65f6\uff0c\u53ea\u6709\u540c\u65f6\u6ee1\u8db3\u68c0\u7d22\u6709\u7528\u4e14\u63a8\u7406\u6b63\u786e\u7684\u8f6e\u6b21\uff0c\u624d\u771f\u6b63\u5206\u5230\u4f18\u52bf\u4fe1\u53f7\u3002\u5bf9\u4e8e\u5931\u8d25\u8f68\u8ff9\uff0c\u4f5c\u8005\u53cd\u800c\u5f88\u8c28\u614e\uff0c\u6ca1\u6709\u505a\u7ec6\u7c92\u5ea6\u533a\u5206\uff0c\u800c\u662f\u7edf\u4e00\u5747\u5300\u5206\u914d\u3002\u539f\u56e0\u662f\u5931\u8d25\u7684\u6210\u56e0\u901a\u5e38\u5f88\u6a21\u7cca\uff1a\u53ef\u80fd\u662f\u68c0\u7d22\u5668\u6ca1\u53ec\u56de\u5230\u5173\u952e\u6587\u6863\uff0c\u53ef\u80fd\u662f\u8bed\u6599\u672c\u8eab\u6ca1\u8986\u76d6\uff0c\u53ef\u80fd\u662f\u4e2d\u95f4\u8f6e\u6b21\u5176\u5b9e\u505a\u4e86\u5408\u7406\u5c1d\u8bd5\uff0c\u53ea\u662f\u6700\u540e\u8fd8\u662f\u6ca1\u89e3\u51fa\u6765\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u786c\u7ed9\u5931\u8d25\u8f68\u8ff9\u6bcf\u4e00\u8f6e\u5206\u9ad8\u4f4e\uff0c\u53cd\u800c\u5bb9\u6613\u628a\u566a\u58f0\u5f15\u8fdb\u6765\u3002\u00a0<\/p>\n\n\n\n<p>\u7b2c\u56db\u6b65\uff0c\u5c31\u662f advantage reallocation\u3002\u4f5c\u8005\u628a\u8f68\u8ff9\u7ea7\u7684 outcome advantage \u6309\u7167\u4e0a\u4e00\u6b65\u7684\u6743\u91cd\u91cd\u65b0\u5206\u914d\u5230\u6bcf\u4e00\u8f6e\u641c\u7d22\u4e0a\uff0c\u800c\u6700\u7ec8\u56de\u7b54\u8f6e\u6b21\u4ecd\u76f4\u63a5\u4f7f\u7528 outcome advantage\u3002\u8fd9\u6837\uff0c\u6a21\u578b\u5728\u66f4\u65b0\u65f6\u5c31\u4f1a\u66f4\u5173\u6ce8\u90a3\u4e9b\u771f\u6b63\u63a8\u52a8\u6210\u529f\u7684\u641c\u7d22\u8f6e\uff0c\u800c\u4e0d\u662f\u628a\u540c\u6837\u7684\u5b66\u4e60\u5f3a\u5ea6\u5e73\u5747\u5730\u6492\u5728\u6574\u6761\u8f68\u8ff9\u4e0a\u3002\u8bba\u6587\u7279\u522b\u5f3a\u8c03\uff0c\u8fd9\u79cd\u505a\u6cd5\u4e0d\u4f1a\u6539\u53d8\u6574\u6761\u8f68\u8ff9\u603b\u7684\u5b66\u4e60\u4fe1\u53f7\u91cf\u7ea7\uff0c\u53ea\u662f\u91cd\u65b0\u5b89\u6392\u4e86\u201c\u4fe1\u53f7\u843d\u5728\u54ea\u4e9b\u8f6e\u6b21\u4e0a\u201d\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u5b83\u65e2\u80fd\u4fdd\u7559 GRPO \u7684\u4f18\u5316\u7a33\u5b9a\u6027\uff0c\u53c8\u80fd\u5f15\u5165\u66f4\u7ec6\u7c92\u5ea6\u7684\u8fc7\u7a0b\u4fe1\u606f\u3002\u00a0<\/p>\n\n\n\n<p>\u53e6\u5916\uff0c\u4f5c\u8005\u8fd8\u4e13\u95e8\u505a\u4e86 judge calibration\u3002\u4ed6\u4eec\u4eba\u5de5\u6807\u6ce8\u4e86 97 \u4e2a\u641c\u7d22\u8f6e\u6b21\uff0c\u6700\u7ec8\u8ba9 LLM judge \u4e0e\u4eba\u5de5\u5224\u65ad\u5728 retrieval utility \u548c reasoning correctness \u4e0a\u8fbe\u5230\u7ea6 95% \u7684\u4e00\u81f4\u7387\uff0c\u8bf4\u660e\u8fd9\u4e2a judge \u81f3\u5c11\u5728\u4ed6\u4eec\u7684\u8bbe\u5b9a\u91cc\u4e0d\u662f\u968f\u4fbf\u62cd\u8111\u888b\u6253\u5206\uff0c\u800c\u662f\u505a\u8fc7\u8f83\u8ba4\u771f\u6821\u51c6\u7684\u3002\u8fd9\u4e2a\u7ec6\u8282\u5f88\u91cd\u8981\uff0c\u56e0\u4e3a\u5b83\u5173\u7cfb\u5230\u65b9\u6cd5\u662f\u5426\u53ef\u4fe1\uff0c\u5982\u679c judge \u672c\u8eab\u5f88\u4e0d\u9760\u8c31\uff0c\u540e\u9762\u7684\u4f18\u52bf\u91cd\u5206\u914d\u5c31\u7ad9\u4e0d\u4f4f\u3002\u00a0<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>5. \u5b9e\u9a8c\u8bbe\u7f6e\uff08Setup\uff09<\/strong><\/h2>\n\n\n\n<p>\u5b9e\u9a8c\u6db5\u76d6\u4e24\u7c7b\u4efb\u52a1<\/p>\n\n\n\n<p>* General QA\uff08NQ\u3001TriviaQA\u3001PopQA\uff09<\/p>\n\n\n\n<p>* Multi-hop QA\uff08HotpotQA\u30012Wiki\u3001Musique\u3001Bamboogle\uff09<\/p>\n\n\n\n<p>\u8bc4\u4ef7\u6307\u6807\u4e3a Avg@4 Exact Match\uff08EM\uff09\u3002&nbsp;<\/p>\n\n\n\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8bba\u6587\u91c7\u7528\u4e86\u66f4\u4e25\u683c\u7684\u6d4b\u8bd5\u8bbe\u5b9a\uff1a<\/p>\n\n\n\n<p>* \u4f7f\u7528 hard-case \u6570\u636e\u96c6\uff08\u5927\u6a21\u578b\u4e5f\u96be\u4ee5\u89e3\u51b3\uff09<\/p>\n\n\n\n<p>* \u7981\u6b62\u6a21\u578b\u4f9d\u8d56\u53c2\u6570\u77e5\u8bc6\uff0c\u5fc5\u987b\u901a\u8fc7\u68c0\u7d22\u5b8c\u6210\u4efb\u52a1<\/p>\n\n\n\n<p>\u6a21\u578b\u65b9\u9762\uff0c\u91c7\u7528 Qwen3-8B \u548c Qwen3-1.7B \u4f5c\u4e3a backbone\uff0c\u5e76\u4e0e\u591a\u79cd RL \u65b9\u6cd5\u8fdb\u884c\u5bf9\u6bd4\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"992\" height=\"360\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-6.png\"  class=\"wp-image-1295\" style=\"aspect-ratio:2.7556512378902047;width:548px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-6.png 992w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-6-300x109.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-6-768x279.png 768w\" sizes=\"auto, (max-width: 992px) 100vw, 992px\" title=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe2\" alt=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>6. \u5b9e\u9a8c\u7ed3\u679c\u4e0e\u5206\u6790\uff08Experiments\uff09<\/strong><\/h2>\n\n\n\n<p>\u6574\u4f53\u7ed3\u679c CW-GRPO \u5728\u6240\u6709\u5bf9\u6bd4\u65b9\u6cd5\u4e2d\u8868\u73b0\u6700\u4f18\uff1a<\/p>\n\n\n\n<p>* Qwen3-8B\u63d0\u5347 5.0%<\/p>\n\n\n\n<p>* Qwen3-1.7B\u63d0\u5347 6.3%<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"723\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-7-1024x723.png\"  class=\"wp-image-1296\" style=\"aspect-ratio:1.4163022287791769;width:537px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-7-1024x723.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-7-300x212.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-7-768x542.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-7.png 1286w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe3\" alt=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe3\" \/><\/figure>\n\n\n\n<p>\u4efb\u52a1\u8868\u73b0\u5206\u6790<\/p>\n\n\n\n<p>\u5728 Multi-hop QA \u4e0a\u63d0\u5347\u66f4\u660e\u663e\uff0c\u8bf4\u660e\u65b9\u6cd5\u5728\u957f\u94fe\u63a8\u7406\u4efb\u52a1\u4e2d\u4f18\u52bf\u66f4\u5927\u3002<\/p>\n\n\n\n<p>\u5728\u7b80\u5355\u4efb\u52a1\u4e0a\u4f18\u52bf\u76f8\u5bf9\u6709\u9650\u3002<\/p>\n\n\n\n<p>\u8d21\u732e\u5206\u5e03\u5206\u6790<\/p>\n\n\n\n<p>\u5b9e\u9a8c\u8868\u660e\uff0c\u968f\u7740\u53c2\u6570 \u03b1 \u589e\u5927\uff0c\u6027\u80fd\u63d0\u5347\u660e\u663e\uff0c\u6700\u4f18\u51fa\u73b0\u5728 \u03b1=\u221e\u3002<\/p>\n\n\n\n<p>\u8bf4\u660e\u6210\u529f\u8f68\u8ff9\u7684\u5173\u952e\u8d21\u732e\u96c6\u4e2d\u5728\u5c11\u6570\u8f6e\u6b21\uff0c\u800c\u975e\u5e73\u5747\u5206\u5e03\u3002\u00a0<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"387\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-8-1024x387.png\"  class=\"wp-image-1297\" style=\"aspect-ratio:2.6460599497957378;width:538px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-8-1024x387.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-8-300x113.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-8-768x290.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-8.png 1308w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe4\" alt=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe4\" \/><\/figure>\n\n\n\n<p>\u6d88\u878d\u5b9e\u9a8c<\/p>\n\n\n\n<p>\u53bb\u6389\u4efb\u4e00\u4fe1\u53f7\uff08\u68c0\u7d22\u6216\u63a8\u7406\uff09\u90fd\u4f1a\u964d\u4f4e\u6027\u80fd\uff0c\u8bf4\u660e\u4e24\u8005\u7f3a\u4e00\u4e0d\u53ef\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"322\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-9-1024x322.png\"  class=\"wp-image-1298\" style=\"width:528px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-9-1024x322.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-9-300x94.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-9-768x241.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-9.png 1298w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe5\" alt=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe5\" \/><\/figure>\n\n\n\n<p>\u8bad\u7ec3\u7a33\u5b9a\u6027\u4e0e\u6848\u4f8b\u5206\u6790<\/p>\n\n\n\n<p>\u8bad\u7ec3\u66f2\u7ebf\u7a33\u5b9a\u63d0\u5347\uff08Figure 3\uff09\uff0c\u8bf4\u660e\u65b9\u6cd5\u672a\u7834\u574f\u4f18\u5316\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n<p>\u6848\u4f8b\u5206\u6790\uff08Figure 4\uff09\u663e\u793a\uff0c\u6a21\u578b\u5728\u8bad\u7ec3\u540e\u80fd\u591f\u8fdb\u884c\u66f4\u5168\u9762\u641c\u7d22\u5e76\u6b63\u786e\u5229\u7528\u8bc1\u636e\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"839\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-10-1024x839.png\"  class=\"wp-image-1299\" style=\"aspect-ratio:1.2202097235462346;width:597px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-10-1024x839.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-10-300x246.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-10-768x629.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/04\/image-10.png 1330w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe6\" alt=\"Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>7. \u8d21\u732e\u4e0e\u7ed3\u8bba\uff08Conclusion\uff09<\/strong><\/h2>\n\n\n\n<p>\u672c\u6587\u7684\u4e3b\u8981\u8d21\u732e\u5305\u62ec\uff1a<\/p>\n\n\n\n<p>\uff081\uff09\u63d0\u51fa\u5c06 process supervision \u91cd\u6784\u4e3a\u4f18\u52bf\u91cd\u5206\u914d\u95ee\u9898<\/p>\n\n\n\n<p>\uff082\uff09\u8bbe\u8ba1 CW-GRPO\uff0c\u5b9e\u73b0\u7a33\u5b9a\u4e14\u7ec6\u7c92\u5ea6\u7684 credit assignment<\/p>\n\n\n\n<p>\uff083\uff09\u53d1\u73b0\u641c\u7d22\u4efb\u52a1\u4e2d\u201c\u8d21\u732e\u96c6\u4e2d\u201d\u7684\u7ed3\u6784\u6027\u7279\u5f81<\/p>\n\n\n\n<p>\uff084\uff09\u5728\u591a\u4e2a\u4efb\u52a1\u4e0a\u9a8c\u8bc1\u65b9\u6cd5\u6709\u6548\u6027<\/p>\n\n\n\n<p>\u4ece\u6574\u4f53\u6765\u770b\uff0c\u8fd9\u7bc7\u5de5\u4f5c\u771f\u6b63\u89e3\u51b3\u7684\u4e0d\u662f\u201c\u5982\u4f55\u8ba9 agent \u66f4\u5f3a\u201d\uff0c\u800c\u662f\u4e00\u4e2a\u66f4\u57fa\u7840\u7684\u95ee\u9898\uff0c\u5728\u591a\u8f6e\u641c\u7d22\u4e2d\uff0c\u6210\u529f\u7a76\u7adf\u6765\u6e90\u4e8e\u54ea\u51e0\u6b65\uff0c\u4ee5\u53ca\u8bad\u7ec3\u4fe1\u53f7\u5e94\u8be5\u5982\u4f55\u5206\u914d\u3002CW-GRPO \u7ed9\u51fa\u7684\u7b54\u6848\uff1a\u4e0d\u662f\u589e\u52a0\u5956\u52b1\uff0c\u800c\u662f\u91cd\u65b0\u5206\u914d\u5956\u52b1\u3002\u8fd9\u4e00\u601d\u8def\u5728\u4fdd\u8bc1\u7a33\u5b9a\u6027\u7684\u540c\u65f6\uff0c\u5f15\u5165\u4e86\u8fc7\u7a0b\u4fe1\u606f\uff0c\u5bf9\u540e\u7eed\u641c\u7d22\u578b\u667a\u80fd\u4f53\u7684\u8bad\u7ec3\u5177\u6709\u8f83\u5f3a\u53c2\u8003\u4ef7\u503c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. \u6458\u8981\uff08Abstract\uff09 \u672c\u6587\u805a\u7126\u4e8e LLM-based Search Agent \u7684\u8bad\u7ec3\u95ee\u9898\u3002\u73b0\u6709\u65b9\u6cd5\u5728\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e2d\u4e3b\u8981\u9762\u4e34\u4e00\u4e2a\u6838\u5fc3\u96be\u70b9credit assignment\uff08\u8d21\u732e\u5f52\u56e0\uff09\u3002\u4e00\u65b9\u9762\uff0c\u57fa\u4e8e\u6700\u7ec8\u7b54\u6848\u7684 outcome supervision \u867d\u7136\u8bad\u7ec3\u7a33\u5b9a\uff0c\u4f46\u5956\u52b1\u4fe1\u53f7\u8fc7\u4e8e\u7a00\u758f\uff0c\u65e0\u6cd5\u533a\u5206\u4e0d\u540c\u641c\u7d22\u8f6e\u6b21\u7684\u91cd\u8981\u6027\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u57fa\u4e8e\u4e2d\u95f4\u6b65\u9aa4\u7684 process supervision \u867d\u80fd\u63d0\u4f9b\u66f4\u7ec6\u7c92\u5ea6\u4fe1\u53f7\uff0c\u4f46\u901a\u5e38\u4f9d\u8d56\u4e0d\u7a33\u5b9a\u7684\u4ef7\u503c\u4f30\u8ba1\u6216\u566a\u58f0\u8f83\u5927\u7684\u4e2d\u95f4\u5956\u52b1\uff0c\u5bb9\u6613\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u3002\u00a0  &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1292\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":7,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-1292","post","type-post","status-publish","format-standard","hentry","category-rengongzhineng"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1292","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1292"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1292\/revisions"}],"predecessor-version":[{"id":1300,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1292\/revisions\/1300"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1292"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1292"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1292"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}