{"id":1449,"date":"2026-06-11T16:11:44","date_gmt":"2026-06-11T08:11:44","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1449"},"modified":"2026-06-11T16:11:45","modified_gmt":"2026-06-11T08:11:45","slug":"self-evolving-llm-agents-with-in-distribution-optimization","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1449","title":{"rendered":"Self-evolving LLM Agents with In-distribution Optimization"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">1. \u6458\u8981\uff08Abstract\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u672c\u6587\u7814\u7a76\u7684\u662f\u957f\u7a0b\u4ea4\u4e92\u578b LLM Agent \u7684\u8bad\u7ec3\u95ee\u9898\uff0c\u6838\u5fc3\u5173\u6ce8\u70b9\u662f\u7a00\u758f\u5ef6\u8fdf\u5956\u52b1\u4e0b\u7684\u8d21\u732e\u5f52\u56e0\u3002\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u4ece\u9759\u6001\u6587\u672c\u751f\u6210\u9010\u6e10\u8d70\u5411\u73af\u5883\u4ea4\u4e92\uff0cLLM Agent \u9700\u8981\u5728\u7f51\u9875\u8d2d\u7269\u3001\u865a\u62df\u5b9e\u9a8c\u3001\u5bb6\u5c45\u4efb\u52a1\u7b49\u590d\u6742\u73af\u5883\u4e2d\u8fdb\u884c\u8fde\u7eed\u51b3\u7b56\u3002\u7136\u800c\uff0c\u8fd9\u7c7b\u4efb\u52a1\u901a\u5e38\u53ea\u6709\u5728\u6574\u4e2a episode \u7ed3\u675f\u540e\u624d\u7ed9\u51fa\u6700\u7ec8\u5956\u52b1\uff0c\u4e2d\u95f4\u6b65\u9aa4\u7f3a\u5c11\u660e\u786e\u76d1\u7763\uff0c\u5bfc\u81f4\u6a21\u578b\u5f88\u96be\u5224\u65ad\u54ea\u4e00\u6b65\u771f\u6b63\u63a8\u52a8\u4e86\u4efb\u52a1\u6210\u529f\uff0c\u54ea\u4e00\u6b65\u9020\u6210\u4e86\u5931\u8d25\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u8bba\u6587\u63d0\u51fa Q-Evolve\uff0c\u4e00\u79cd\u9762\u5411 LLM Agent \u7684\u81ea\u8fdb\u5316\u8bad\u7ec3\u6846\u67b6\u3002\u5b83\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u8fc7\u7a0b\u5956\u52b1\u81ea\u52a8\u6807\u6ce8\u548c\u7b56\u7565\u5b66\u4e60\u7edf\u4e00\u5230\u540c\u4e00\u4e2a in-distribution learning loop \u4e2d\u3002\u5177\u4f53\u800c\u8a00\uff0cQ-Evolve \u9996\u5148\u4f7f\u7528\u4e13\u5bb6\u8f68\u8ff9\u5bf9\u6a21\u578b\u8fdb\u884c\u884c\u4e3a\u514b\u9686\u9884\u70ed\uff0c\u7136\u540e\u6536\u96c6\u6a21\u578b\u81ea\u8eab\u4e0e\u73af\u5883\u4ea4\u4e92\u4ea7\u751f\u7684\u8f68\u8ff9\uff0c\u5e76\u5c06\u4e13\u5bb6\u6570\u636e\u548c\u81ea\u751f\u6210\u6570\u636e\u5408\u5e76\u6210 hybrid offline dataset\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6846\u67b6\u8bad\u7ec3\u4e00\u4e2a in-distribution critic\uff0c\u901a\u8fc7 weighted Implicit Q-Learning \u5c06\u7ec8\u5c40\u5956\u52b1\u5411\u524d\u4f20\u64ad\uff0c\u5f97\u5230\u66f4\u53ef\u9760\u7684 step-wise process rewards\u3002\u6700\u540e\uff0c\u6a21\u578b\u518d\u5229\u7528\u8fd9\u4e9b\u8fc7\u7a0b\u5956\u52b1\uff0c\u901a\u8fc7 behavior-proximal policy optimization \u66f4\u65b0\u7b56\u7565\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u7bc7\u8bba\u6587\u7684\u5173\u952e\u521b\u65b0\u4e0d\u5728\u4e8e\u7b80\u5355\u589e\u52a0\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\uff0c\u800c\u5728\u4e8e\u5f3a\u8c03\u8fc7\u7a0b\u5956\u52b1\u5fc5\u987b\u5728\u540c\u4e00\u6570\u636e\u5206\u5e03\u5185\u88ab\u751f\u6210\u548c\u4f7f\u7528\u3002\u4f5c\u8005\u6307\u51fa\uff0c\u4f20\u7edf PRM \u65b9\u6cd5\u5bb9\u6613\u51fa\u73b0 distribution shift\uff0c\u5373\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65f6\u770b\u5230\u7684\u6570\u636e\u548c\u7b56\u7565\u4f18\u5316\u65f6\u4ea7\u751f\u7684\u6570\u636e\u4e0d\u4e00\u81f4\uff0c\u5bfc\u81f4\u4e2d\u95f4\u53cd\u9988\u5931\u771f\u3002Q-Evolve \u5219\u901a\u8fc7\u8ba9 policy\u3001critic \u548c dataset \u5728\u540c\u4e00\u95ed\u73af\u4e2d\u5171\u540c\u6f14\u5316\uff0c\u51cf\u5c11\u8fd9\u79cd\u5206\u5e03\u9519\u4f4d\uff0c\u4f7f agent \u80fd\u591f\u7a33\u5b9a\u81ea\u6211\u6539\u8fdb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u65b9\u9762\uff0c\u8bba\u6587\u5728 AlfWorld\u3001WebShop \u548c ScienceWorld \u4e09\u4e2a\u957f\u7a0b\u4ea4\u4e92\u4efb\u52a1\u4e0a\u8fdb\u884c\u8bc4\u4f30\u3002\u7ed3\u679c\u663e\u793a\uff0cQ-Evolve \u5728\u603b\u4f53\u6027\u80fd\u3001\u6837\u672c\u6548\u7387\u548c\u6cdb\u5316\u80fd\u529b\u4e0a\u5747\u4f18\u4e8e\u591a\u79cd\u5f3a\u57fa\u7ebf\u65b9\u6cd5\u3002\u5c24\u5176\u5728 AlfWorld \u4e0a\uff0cQ-Evolve \u7528\u7ea6 13K \u73af\u5883\u6b65\u6570\u5c31\u8d85\u8fc7\u4e86\u591a\u4e2a\u4f7f\u7528 320K \u73af\u5883\u6b65\u6570\u7684 online RL \u65b9\u6cd5\uff0c\u4f53\u73b0\u51fa\u8f83\u5f3a\u7684\u6837\u672c\u6548\u7387\u3002\u6574\u4f53\u6765\u770b\uff0c\u8be5\u5de5\u4f5c\u4e3a LLM Agent \u7684\u957f\u671f\u51b3\u7b56\u8bad\u7ec3\u63d0\u4f9b\u4e86\u4e00\u79cd\u8f83\u7a33\u5065\u7684\u81ea\u8fdb\u5316\u8def\u5f84\uff0c\u4e5f\u8fdb\u4e00\u6b65\u8bf4\u660e\u8fc7\u7a0b\u7ea7\u76d1\u7763\u548c\u7b56\u7565\u5b66\u4e60\u9700\u8981\u653e\u5728\u7edf\u4e00\u5206\u5e03\u4e2d\u5171\u540c\u8bbe\u8ba1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"290\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-11-1024x290.png\"  class=\"wp-image-1450\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-11-1024x290.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-11-300x85.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-11-768x217.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-11.png 1436w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u7814\u7a76\u80cc\u666f\u4e0e\u95ee\u9898\u52a8\u673a\uff08Introduction\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd1\u5e74\u6765\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u5df2\u7ecf\u4e0d\u518d\u53ea\u662f\u6587\u672c\u751f\u6210\u5668\uff0c\u800c\u662f\u9010\u6e10\u6210\u4e3a\u4ea4\u4e92\u5f0f agent \u7684\u63a7\u5236\u5668\u3002LLM Agent \u53ef\u4ee5\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u8fdb\u884c\u63a8\u7406\u3001\u89c4\u5212\u548c\u73af\u5883\u4ea4\u4e92\uff0c\u56e0\u6b64\u88ab\u7528\u4e8e\u5bfc\u822a\u3001\u6e38\u620f\u3001\u7f51\u9875\u64cd\u4f5c\u3001\u673a\u5668\u4eba\u63a7\u5236\u7b49\u4efb\u52a1\u3002\u4e0e\u666e\u901a\u95ee\u7b54\u4e0d\u540c\uff0c\u4ea4\u4e92\u5f0f\u4efb\u52a1\u8981\u6c42\u6a21\u578b\u8fde\u7eed\u89c2\u5bdf\u73af\u5883\u3001\u9009\u62e9\u52a8\u4f5c\u3001\u63a5\u6536\u53cd\u9988\uff0c\u5e76\u5728\u591a\u4e2a\u6b65\u9aa4\u540e\u5b8c\u6210\u76ee\u6807\u3002\u8fd9\u7c7b\u4efb\u52a1\u66f4\u63a5\u8fd1\u771f\u5b9e\u4e16\u754c\u7684\u667a\u80fd\u4f53\u95ee\u9898\uff0c\u4e5f\u66f4\u80fd\u4f53\u73b0\u5927\u6a21\u578b\u4ece\u8bed\u8a00\u80fd\u529b\u8d70\u5411\u884c\u52a8\u80fd\u529b\u7684\u8d8b\u52bf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u957f\u7a0b\u4ea4\u4e92\u4efb\u52a1\u7684\u8bad\u7ec3\u96be\u5ea6\u5f88\u9ad8\u3002\u6700\u91cd\u8981\u7684\u95ee\u9898\u662f\u53cd\u9988\u7a00\u758f\u4e14\u5ef6\u8fdf\u3002Agent \u5f80\u5f80\u8981\u6267\u884c\u5f88\u591a\u6b65\u4e4b\u540e\uff0c\u624d\u5728 episode \u7ed3\u675f\u65f6\u5f97\u5230\u6210\u529f\u6216\u5931\u8d25\u7684\u5956\u52b1\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u6a21\u578b\u53ea\u80fd\u77e5\u9053\u6700\u7ec8\u7ed3\u679c\uff0c\u5374\u4e0d\u77e5\u9053\u4e2d\u95f4\u54ea\u4e00\u6b65\u662f\u5173\u952e\u8d21\u732e\uff0c\u54ea\u4e00\u6b65\u662f\u65e0\u6548\u52a8\u4f5c\u3002\u5bf9\u4e8e\u957f\u7a0b\u4efb\u52a1\u800c\u8a00\uff0c\u8fd9\u4f1a\u4e25\u91cd\u5f71\u54cd\u5b66\u4e60\u6548\u7387\uff0c\u56e0\u4e3a\u4e00\u4e2a\u5931\u8d25\u8f68\u8ff9\u53ef\u80fd\u5305\u542b\u4e00\u4e9b\u5408\u7406\u6b65\u9aa4\uff0c\u4e00\u4e2a\u6210\u529f\u8f68\u8ff9\u4e5f\u53ef\u80fd\u5305\u542b\u4e00\u4e9b\u5197\u4f59\u52a8\u4f5c\u3002\u5982\u679c\u8bad\u7ec3\u65b9\u6cd5\u4e0d\u80fd\u62c6\u89e3\u8fd9\u4e9b\u4e2d\u95f4\u8d21\u732e\uff0c\u5c31\u5f88\u96be\u771f\u6b63\u63d0\u5347 agent \u7684\u51b3\u7b56\u8d28\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u6709\u65b9\u6cd5\u5927\u81f4\u6709\u51e0\u7c7b\u3002\u4e00\u7c7b\u662f\u884c\u4e3a\u514b\u9686\u6216\u76d1\u7763\u5fae\u8c03\uff0c\u4e3b\u8981\u6a21\u4eff\u4e13\u5bb6\u8f68\u8ff9\uff0c\u4f18\u70b9\u662f\u7a33\u5b9a\uff0c\u4f46\u7f3a\u70b9\u662f\u65e0\u6cd5\u4ece\u6a21\u578b\u81ea\u5df1\u7684\u9519\u8bef\u4e2d\u6301\u7eed\u6539\u8fdb\u3002\u53e6\u4e00\u7c7b\u662f online RL\uff0c\u4f8b\u5982 PPO \u6216 GRPO\uff0c\u5b83\u4eec\u53ef\u4ee5\u5229\u7528\u73af\u5883\u4ea4\u4e92\u8fdb\u884c\u5b66\u4e60\uff0c\u4f46\u901a\u5e38\u9700\u8981\u5927\u91cf\u91c7\u6837\uff0c\u800c\u4e14\u6ca1\u6709\u4e13\u95e8\u89e3\u51b3\u7a00\u758f\u5956\u52b1\u4e0b\u7684\u8fc7\u7a0b\u7ea7\u5f52\u56e0\u3002\u8fd8\u6709\u4e00\u7c7b\u662f PRM \u6216\u641c\u7d22\u5f0f\u65b9\u6cd5\uff0c\u901a\u8fc7\u4eba\u5de5\u6807\u6ce8\u6216\u5728\u7ebf\u641c\u7d22\u751f\u6210 step-level rewards\uff0c\u4f46\u8fd9\u4e9b\u65b9\u6cd5\u5f80\u5f80\u6210\u672c\u9ad8\uff0c\u5e76\u4e14\u5bb9\u6613\u51fa\u73b0\u5956\u52b1\u6a21\u578b\u548c\u7b56\u7565\u5206\u5e03\u4e0d\u4e00\u81f4\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u672c\u6587\u7684\u95ee\u9898\u610f\u8bc6\u6b63\u662f\u5728\u8fd9\u91cc\u3002\u4f5c\u8005\u8ba4\u4e3a\uff0c\u957f\u7a0b LLM Agent \u7684\u5173\u952e\u96be\u70b9\u4e0d\u662f\u5355\u7eaf\u7f3a\u5c11\u5956\u52b1\uff0c\u800c\u662f\u8fc7\u7a0b\u5956\u52b1\u7684\u53ef\u9760\u6027\u4f9d\u8d56\u6570\u636e\u5206\u5e03\u3002\u5982\u679c\u4e00\u4e2a PRM \u53ea\u5728\u67d0\u4e9b\u8f68\u8ff9\u4e0a\u8bad\u7ec3\uff0c\u5374\u88ab\u62ff\u53bb\u8bc4\u4ef7\u7b56\u7565\u4f18\u5316\u540e\u4ea7\u751f\u7684\u65b0\u72b6\u6001\u548c\u65b0\u52a8\u4f5c\uff0c\u5b83\u7ed9\u51fa\u7684\u5206\u6570\u5c31\u53ef\u80fd\u4e0d\u53ef\u9760\u3002\u56e0\u6b64\uff0cQ-Evolve \u8bd5\u56fe\u89e3\u51b3\u4e00\u4e2a\u66f4\u6839\u672c\u7684\u95ee\u9898\uff0c\u5373\u5982\u4f55\u5728\u540c\u4e00\u5206\u5e03\u5185\u751f\u6210\u8fc7\u7a0b\u76d1\u7763\uff0c\u5e76\u5728\u540c\u4e00\u5206\u5e03\u5185\u5b8c\u6210\u7b56\u7565\u6539\u8fdb\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u65b9\u6cd5\u6574\u4f53\u6846\u67b6\uff08Framework Overview\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Q-Evolve \u7684\u6574\u4f53\u6846\u67b6\u53ef\u4ee5\u7406\u89e3\u4e3a\u4e00\u4e2a\u81ea\u8fdb\u5316\u95ed\u73af\u3002\u7b2c\u4e00\u6b65\u662f\u884c\u4e3a\u514b\u9686\u9884\u70ed\u3002\u4f5c\u8005\u5148\u4f7f\u7528\u4e13\u5bb6\u6570\u636e\u8bad\u7ec3\u4e00\u4e2a\u521d\u59cb agent\uff0c\u4f7f\u6a21\u578b\u5177\u5907\u57fa\u672c\u4efb\u52a1\u6267\u884c\u80fd\u529b\u3002\u7b2c\u4e8c\u6b65\u662f\u81ea\u751f\u6210\u6570\u636e\u6536\u96c6\u3002\u9884\u70ed\u540e\u7684 agent \u4e0e\u73af\u5883\u4ea4\u4e92\uff0c\u4ea7\u751f\u81ea\u5df1\u7684\u6210\u529f\u548c\u5931\u8d25\u8f68\u8ff9\u3002\u7b2c\u4e09\u6b65\u662f\u6784\u5efa hybrid offline dataset\u3002\u8be5\u6570\u636e\u96c6\u540c\u65f6\u5305\u542b\u4e13\u5bb6\u8f68\u8ff9\u548c agent \u81ea\u5df1\u751f\u6210\u7684\u8f68\u8ff9\uff0c\u56e0\u6b64\u65e2\u6709\u9ad8\u8d28\u91cf\u6210\u529f\u793a\u8303\uff0c\u4e5f\u6709\u6a21\u578b\u771f\u5b9e\u5206\u5e03\u4e0b\u7684\u9519\u8bef\u6837\u672c\u3002\u7b2c\u56db\u6b65\u662f\u5728\u8be5\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3 in-distribution critic\uff0c\u5e76\u7531 critic \u63a8\u5bfc step-wise process rewards\u3002\u7b2c\u4e94\u6b65\u662f\u5229\u7528\u8fd9\u4e9b\u8fc7\u7a0b\u5956\u52b1\u66f4\u65b0\u7b56\u7565\u3002\u66f4\u65b0\u540e\u7684\u7b56\u7565\u518d\u53bb\u73af\u5883\u4e2d\u751f\u6210\u65b0\u8f68\u8ff9\uff0c\u8fdb\u5165\u4e0b\u4e00\u8f6e\u6f14\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u6846\u67b6\u7684\u5173\u952e\u5728\u4e8e hybrid data\u3002\u4e13\u5bb6\u6570\u636e\u53ef\u4ee5\u63d0\u4f9b\u6210\u529f\u4efb\u52a1\u8def\u5f84\uff0c\u5e2e\u52a9 critic \u83b7\u5f97\u53ef\u9760\u7684\u9ad8\u4ef7\u503c\u4fe1\u53f7\u3002\u81ea\u751f\u6210\u6570\u636e\u5219\u53cd\u6620\u5f53\u524d agent \u771f\u6b63\u4f1a\u9047\u5230\u7684\u72b6\u6001\u548c\u52a8\u4f5c\uff0c\u5305\u62ec\u65e0\u6548\u52a8\u4f5c\u3001\u683c\u5f0f\u9519\u8bef\u3001\u91cd\u590d\u52a8\u4f5c\u548c\u5c40\u90e8\u5408\u7406\u4f46\u6700\u7ec8\u5931\u8d25\u7684\u884c\u4e3a\u3002\u4e24\u7c7b\u6570\u636e\u7ed3\u5408\u540e\uff0c\u8fc7\u7a0b\u5956\u52b1\u4e0d\u518d\u53ea\u670d\u52a1\u4e8e\u4e13\u5bb6\u5206\u5e03\uff0c\u800c\u662f\u80fd\u8986\u76d6\u6a21\u578b\u81ea\u8eab\u7684\u884c\u4e3a\u5206\u5e03\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u4e2a\u5173\u952e\u70b9\u662f in-distribution\u3002Q-Evolve \u5e76\u4e0d\u662f\u5148\u8bad\u7ec3\u4e00\u4e2a\u9759\u6001 PRM\uff0c\u518d\u8ba9 policy \u5230\u65b0\u5206\u5e03\u4e2d\u81ea\u7531\u63a2\u7d22\uff0c\u800c\u662f\u5728\u6bcf\u4e00\u8f6e\u6f14\u5316\u4e2d\u90fd\u628a critic learning\u3001process reward labeling \u548c policy learning \u7ea6\u675f\u5728\u540c\u4e00\u4e2a hybrid dataset \u5185\u3002\u8fd9\u6837\u505a\u53ef\u4ee5\u964d\u4f4e\u5206\u5e03\u504f\u79fb\u98ce\u9669\uff0c\u4f7f\u5956\u52b1\u6807\u6ce8\u548c\u7b56\u7565\u66f4\u65b0\u4e92\u76f8\u5339\u914d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0cQ-Evolve \u672c\u8d28\u4e0a\u4e0d\u662f\u4e00\u4e2a\u5355\u8f6e\u8bad\u7ec3\u7b97\u6cd5\uff0c\u800c\u662f\u4e00\u4e2a policy\u3001critic \u548c dataset \u5171\u540c\u6f14\u5316\u7684\u7cfb\u7edf\u3002\u6bcf\u4e00\u8f6e\u66f4\u65b0\u90fd\u6bd4\u8f83\u4fdd\u5b88\uff0c\u907f\u514d\u79bb\u5f00\u5f53\u524d\u6570\u636e\u652f\u6301\u8303\u56f4\uff0c\u4f46\u591a\u8f6e\u8fed\u4ee3\u540e\u53c8\u80fd\u9010\u6b65\u63d0\u5347 agent \u7684\u957f\u7a0b\u4efb\u52a1\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"441\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-12-1024x441.png\"  class=\"wp-image-1451\" style=\"aspect-ratio:2.3217447635499893;width:594px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-12-1024x441.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-12-300x129.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-12-768x331.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-12.png 1198w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe1\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">4. \u6838\u5fc3\u65b9\u6cd5\uff08Methodology\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Q-Evolve \u7684\u6838\u5fc3\u65b9\u6cd5\u53ef\u4ee5\u5206\u6210\u56db\u4e2a\u90e8\u5206\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\u90e8\u5206\u662f\u884c\u4e3a\u514b\u9686\u9884\u70ed\u3002\u4f5c\u8005\u4f7f\u7528\u4e13\u5bb6\u8f68\u8ff9\u8bad\u7ec3\u521d\u59cb\u7b56\u7565\uff0c\u8ba9 agent \u5b66\u4f1a\u57fa\u672c\u4efb\u52a1\u683c\u5f0f\u548c\u64cd\u4f5c\u6a21\u5f0f\u3002\u5bf9\u4e8e\u957f\u7a0b\u4ea4\u4e92\u4efb\u52a1\uff0c\u8fd9\u4e00\u6b65\u5f88\u91cd\u8981\uff0c\u56e0\u4e3a\u5b8c\u5168\u968f\u673a\u63a2\u7d22\u5f88\u96be\u83b7\u5f97\u6210\u529f\u8f68\u8ff9\uff0c\u4e5f\u5f88\u96be\u4e3a\u540e\u7eed critic learning \u63d0\u4f9b\u6709\u6548\u4fe1\u53f7\u3002\u884c\u4e3a\u514b\u9686\u867d\u7136\u4e0d\u80fd\u89e3\u51b3\u6700\u7ec8\u7684\u81ea\u8fdb\u5316\u95ee\u9898\uff0c\u4f46\u53ef\u4ee5\u4e3a\u540e\u7eed\u8bad\u7ec3\u63d0\u4f9b\u4e00\u4e2a\u53ef\u7528\u7684\u8d77\u70b9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e8c\u90e8\u5206\u662f hybrid data construction \u548c retrospective reward labeling\u3002\u4f5c\u8005\u5c06\u4e13\u5bb6\u8f68\u8ff9\u548c self-collected rollouts \u5408\u5e76\u4e3a\u4e00\u4e2a\u79bb\u7ebf\u6570\u636e\u96c6\u3002\u968f\u540e\uff0c\u7cfb\u7edf\u6839\u636e\u73af\u5883\u8fd4\u56de\u7684\u6587\u672c\u53cd\u9988\uff0c\u5bf9\u6bcf\u4e00\u6b65\u8fdb\u884c\u89c4\u5219\u5316\u56de\u770b\u6807\u6ce8\u3002\u4f8b\u5982\uff0c\u5982\u679c\u52a8\u4f5c\u683c\u5f0f\u9519\u8bef\uff0c\u4f1a\u7ed9\u51fa\u8d1f\u5956\u52b1\u3002\u5982\u679c\u52a8\u4f5c\u4e0d\u88ab\u73af\u5883\u63a5\u53d7\uff0c\u4e5f\u4f1a\u7ed9\u51fa\u8d1f\u5956\u52b1\u3002\u5982\u679c\u52a8\u4f5c\u6267\u884c\u540e observation \u6ca1\u6709\u53d8\u5316\uff0c\u5219\u8bf4\u660e\u53ef\u80fd\u662f\u65e0\u610f\u4e49\u91cd\u590d\u64cd\u4f5c\uff0c\u4e5f\u4f1a\u88ab\u60e9\u7f5a\u3002\u8fd9\u4e9b\u8f85\u52a9\u4fe1\u53f7\u4e0d\u9700\u8981\u4eba\u5de5\u6807\u6ce8\uff0c\u4e5f\u4e0d\u9700\u8981\u73af\u5883\u56de\u6eaf\uff0c\u6bd4\u8f83\u9002\u5408\u5b9e\u9645\u4ea4\u4e92\u573a\u666f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e09\u90e8\u5206\u662f in-distribution critic learning\u3002\u4f5c\u8005\u91c7\u7528 weighted Implicit Q-Learning \u6765\u8bad\u7ec3 critic\u3002\u666e\u901a Bellman backup \u7406\u8bba\u4e0a\u53ef\u4ee5\u628a\u7ec8\u5c40\u5956\u52b1\u5411\u524d\u4f20\u64ad\uff0c\u4f46\u5728\u7a00\u758f\u5956\u52b1\u4efb\u52a1\u4e2d\u5bb9\u6613\u88ab\u5927\u91cf\u96f6\u5956\u52b1\u548c\u5931\u8d25\u8f68\u8ff9\u6df9\u6ca1\u3002\u56e0\u6b64\uff0cQ-Evolve \u5bf9\u4e0d\u540c\u6837\u672c\u8fdb\u884c\u52a0\u6743\uff0c\u66f4\u91cd\u89c6\u6210\u529f\u8f68\u8ff9\u548c\u9760\u8fd1\u7ec8\u70b9\u7684\u6b65\u9aa4\u3002\u8fd9\u6837\u53ef\u4ee5\u8ba9 critic \u66f4\u7a33\u5b9a\u5730\u5b66\u4e60\u54ea\u4e9b\u4e2d\u95f4\u884c\u4e3a\u66f4\u53ef\u80fd\u5e26\u6765\u6700\u7ec8\u6210\u529f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u56db\u90e8\u5206\u662f process reward estimation \u548c policy optimization\u3002\u8bad\u7ec3\u597d critic \u540e\uff0c\u4f5c\u8005\u4e0d\u76f4\u63a5\u4f7f\u7528\u7b80\u5355\u7684 Q minus V \u4f5c\u4e3a\u8fc7\u7a0b\u5956\u52b1\uff0c\u800c\u662f\u901a\u8fc7 GAE \u4f30\u8ba1 step-wise advantages\u3002\u8fd9\u6837\u53ef\u4ee5\u83b7\u5f97\u66f4\u5e73\u6ed1\u3001\u66f4\u53ef\u9760\u7684\u8fc7\u7a0b\u5956\u52b1\u3002\u968f\u540e\uff0c\u6a21\u578b\u4f7f\u7528 behavior-proximal policy optimization \u66f4\u65b0\u7b56\u7565\u3002\u4e0e\u666e\u901a advantage weighted regression \u4e0d\u540c\uff0cBPPO \u4e0d\u53ea\u662f\u63d0\u9ad8\u597d\u52a8\u4f5c\u7684\u6982\u7387\uff0c\u8fd8\u4f1a\u663e\u5f0f\u538b\u4f4e\u8d1f\u4f18\u52bf\u52a8\u4f5c\u7684\u6982\u7387\uff0c\u56e0\u6b64\u66f4\u9002\u5408\u7ea0\u6b63\u957f\u7a0b\u4efb\u52a1\u4e2d\u7684\u9519\u8bef\u884c\u4e3a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4f53\u6765\u770b\uff0cQ-Evolve \u7684\u65b9\u6cd5\u8bbe\u8ba1\u6bd4\u8f83\u514b\u5236\u3002\u5b83\u6ca1\u6709\u8ba9 policy \u5728\u672a\u77e5\u5206\u5e03\u4e2d\u5927\u5e45\u8df3\u8dc3\uff0c\u800c\u662f\u5728\u5df2\u6709\u6570\u636e\u652f\u6301\u8303\u56f4\u5185\u9010\u6b65\u6539\u8fdb\u3002\u8fd9\u79cd\u8bbe\u8ba1\u867d\u7136\u770b\u8d77\u6765\u4e0d\u6fc0\u8fdb\uff0c\u4f46\u6b63\u597d\u7b26\u5408\u957f\u7a0b\u4ea4\u4e92\u4efb\u52a1\u4e2d\u7a33\u5b9a\u6027\u548c\u5b89\u5168\u6027\u7684\u8981\u6c42\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"702\" height=\"280\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-13.png\"  class=\"wp-image-1452\" style=\"width:384px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-13.png 702w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-13-300x120.png 300w\" sizes=\"auto, (max-width: 702px) 100vw, 702px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe2\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">5. \u5b9e\u9a8c\u8bbe\u7f6e\uff08Experimental Setup\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u5728\u4e09\u4e2a\u957f\u7a0b\u4ea4\u4e92\u73af\u5883\u4e0a\u8fdb\u884c\u5b9e\u9a8c\uff0c\u5206\u522b\u662f AlfWorld\u3001WebShop \u548c ScienceWorld\u3002AlfWorld \u662f\u6587\u672c\u5316\u5bb6\u5c45\u4efb\u52a1\u73af\u5883\uff0cagent \u9700\u8981\u901a\u8fc7\u8f83\u957f\u52a8\u4f5c\u5e8f\u5217\u5b8c\u6210 household tasks\uff0c\u6700\u7ec8\u53ea\u5f97\u5230\u6210\u529f\u6216\u5931\u8d25\u7684\u4e8c\u503c\u5956\u52b1\u3002WebShop \u662f\u7f51\u9875\u8d2d\u7269\u4efb\u52a1\uff0cagent \u9700\u8981\u6839\u636e\u76ee\u6807\u9700\u6c42\u6d4f\u89c8\u5546\u54c1\u5e76\u9009\u62e9\u8d2d\u4e70\uff0c\u6700\u7ec8\u5956\u52b1\u53d6\u51b3\u4e8e\u8d2d\u4e70\u5546\u54c1\u662f\u5426\u6ee1\u8db3\u5c5e\u6027\u8981\u6c42\u3002ScienceWorld \u662f\u865a\u62df\u79d1\u5b66\u5b9e\u9a8c\u73af\u5883\uff0cagent \u9700\u8981\u5b8c\u6210\u5305\u542b\u591a\u4e2a\u5b50\u76ee\u6807\u7684\u79d1\u5b66\u4efb\u52a1\uff0c\u6700\u7ec8\u6839\u636e\u4efb\u52a1\u5b8c\u6210\u60c5\u51b5\u83b7\u5f97\u7a00\u758f\u5956\u52b1\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u65b9\u9762\uff0c\u4f5c\u8005\u4e3b\u8981\u4f7f\u7528 Llama2-7B-Chat \u6784\u5efa agent\uff0c\u5e76\u5728\u540e\u7eed\u5b9e\u9a8c\u4e2d\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u65b9\u6cd5\u80fd\u8fc1\u79fb\u5230 Llama-3-8B-Instruct\u3002\u5bf9\u4e8e self-collected data\uff0c\u4f5c\u8005\u8ba9 agent \u5728\u6bcf\u4e2a\u4efb\u52a1\u4e0a\u91c7\u6837\u591a\u6761\u8f68\u8ff9\uff0c\u7528\u4e8e\u6784\u5efa\u6df7\u5408\u79bb\u7ebf\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u7ebf\u65b9\u6cd5\u8986\u76d6\u8f83\u5168\u9762\uff0c\u5305\u62ec GPT-3.5-Turbo\u3001GPT-4\u3001Reflexion\u3001SFT\u3001RFT\u3001PPO\u3001Best-of-N\u3001ETO\u3001DMPO \u548c QLASS\u3002\u5176\u4e2d QLASS \u662f\u4e00\u4e2a\u6bd4\u8f83\u5f3a\u7684 value-based agent \u65b9\u6cd5\uff0c\u901a\u8fc7\u641c\u7d22\u6811\u4f30\u8ba1 Q-value \u6765\u6307\u5bfc\u63a8\u7406\u548c\u52a8\u4f5c\u9009\u62e9\u3002\u56e0\u6b64\uff0cQ-Evolve \u4e0e QLASS \u7684\u6bd4\u8f83\u5c24\u5176\u5173\u952e\uff0c\u56e0\u4e3a\u4e24\u8005\u90fd\u5173\u6ce8 value signal \u548c\u8fc7\u7a0b\u7ea7\u6307\u5bfc\uff0c\u4f46 Q-Evolve \u66f4\u5f3a\u8c03 in-distribution learning \u548c\u6837\u672c\u6548\u7387\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bc4\u4ef7\u6307\u6807\u65b9\u9762\uff0c\u8bba\u6587\u62a5\u544a\u5404\u73af\u5883\u4e2d\u7684\u5e73\u5747\u7d2f\u8ba1\u5956\u52b1\uff0c\u5e76\u5728 ScienceWorld \u548c AlfWorld \u4e0a\u533a\u5206 seen \u548c unseen \u4efb\u52a1\uff0c\u4ee5\u68c0\u9a8c\u6a21\u578b\u662f\u5426\u53ea\u662f\u8bb0\u4f4f\u8bad\u7ec3\u4efb\u52a1\uff0c\u8fd8\u662f\u80fd\u6cdb\u5316\u5230\u672a\u89c1\u573a\u666f\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"494\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-14-1024x494.png\"  class=\"wp-image-1453\" style=\"aspect-ratio:2.071756137038036;width:642px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-14-1024x494.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-14-300x145.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-14-768x371.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-14.png 1156w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe3\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe3\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">6. \u5b9e\u9a8c\u7ed3\u679c\u4e0e\u6027\u80fd\u5206\u6790\uff08Experiments\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3b\u7ed3\u679c\u663e\u793a\uff0cQ-Evolve \u5728\u6240\u6709\u4efb\u52a1\u4e0a\u7684\u5e73\u5747\u8868\u73b0\u6700\u597d\u3002\u6839\u636e Table 2\uff0cQ-Evolve \u7684\u5e73\u5747\u5206\u8fbe\u5230 79.4\uff0c\u9ad8\u4e8e QLASS \u7684 74.5\uff0c\u4e5f\u660e\u663e\u9ad8\u4e8e ETO\u3001Best-of-N\u3001RFT \u548c SFT \u7b49\u65b9\u6cd5\u3002\u5c24\u5176\u5728 AlfWorld \u4e0a\uff0cQ-Evolve \u5728 seen \u548c unseen split \u4e0a\u5206\u522b\u8fbe\u5230 90.7 \u548c 89.6\uff0c\u663e\u8457\u8d85\u8fc7\u5176\u4ed6\u65b9\u6cd5\u3002\u8fd9\u8bf4\u660e\u5b83\u4e0d\u4ec5\u80fd\u63d0\u5347\u5df2\u89c1\u4efb\u52a1\u8868\u73b0\uff0c\u4e5f\u80fd\u5728\u672a\u89c1\u4efb\u52a1\u4e0a\u4fdd\u6301\u8f83\u5f3a\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0e QLASS \u76f8\u6bd4\uff0cQ-Evolve \u7684\u4f18\u52bf\u4e0d\u53ea\u662f\u5206\u6570\u66f4\u9ad8\uff0c\u8fd8\u4f53\u73b0\u5728\u6837\u672c\u6548\u7387\u4e0a\u3002QLASS \u9700\u8981\u5927\u91cf\u5728\u7ebf\u641c\u7d22\u548c rollout \u6765\u4f30\u8ba1 Q-values\uff0c\u800c Q-Evolve \u4e3b\u8981\u4f9d\u8d56 hybrid offline dataset \u548c in-distribution critic learning\uff0c\u5728\u8f83\u5c11\u73af\u5883\u4ea4\u4e92\u4e0b\u5c31\u80fd\u83b7\u5f97\u66f4\u5f3a\u6548\u679c\u3002\u8fd9\u4e2a\u7ed3\u679c\u8bf4\u660e\uff0c\u5bf9\u4e8e\u957f\u7a0b agent \u8bad\u7ec3\u800c\u8a00\uff0c\u4e0d\u4e00\u5b9a\u8981\u9760\u5927\u91cf online exploration\uff0c\u66f4\u5173\u952e\u7684\u662f\u5982\u4f55\u7a33\u5b9a\u5730\u628a\u7a00\u758f\u7ec8\u5c40\u5956\u52b1\u8f6c\u5316\u4e3a\u53ef\u9760\u7684\u8fc7\u7a0b\u76d1\u7763\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6d88\u878d\u5b9e\u9a8c\u8fdb\u4e00\u6b65\u8bf4\u660e\u4e86\u5404\u4e2a\u6a21\u5757\u7684\u91cd\u8981\u6027\u3002Table 3 \u663e\u793a\uff0c\u53bb\u6389 retrospective relabeling\u3001weighted IQL\u3001GAE \u6216 policy improvement \u90fd\u4f1a\u5e26\u6765\u6027\u80fd\u4e0b\u964d\u3002\u5176\u4e2d\u53bb\u6389 GAE \u540e\u4e0b\u964d\u8f83\u660e\u663e\uff0c\u8bf4\u660e\u9ad8\u8d28\u91cf advantage estimation \u662f\u8fc7\u7a0b\u5956\u52b1\u53ef\u9760\u6027\u7684\u5173\u952e\u3002\u53bb\u6389 policy improvement \u540e\u6027\u80fd\u4e0b\u964d\u66f4\u4e25\u91cd\uff0c\u8bf4\u660e\u4ec5\u6709 critic \u6216\u8fc7\u7a0b\u5956\u52b1\u8fd8\u4e0d\u591f\uff0c\u5fc5\u987b\u901a\u8fc7\u5408\u9002\u7684\u7b56\u7565\u4f18\u5316\u65b9\u6cd5\u628a\u8fd9\u4e9b\u4fe1\u53f7\u8f6c\u5316\u4e3a agent \u884c\u4e3a\u63d0\u5347\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u8fd8\u6bd4\u8f83\u4e86\u4e0d\u540c\u8fc7\u7a0b\u5956\u52b1\u9009\u62e9\u3002Table 4 \u663e\u793a\uff0c\u76f4\u63a5\u4f7f\u7528 Q minus V \u6216 potential-based shaping \u7684\u6548\u679c\u90fd\u4e0d\u5982 GAE with environmental reward\u3002\u4f5c\u8005\u8fdb\u4e00\u6b65\u53d1\u73b0\uff0c\u5c06\u8f85\u52a9\u5956\u52b1\u76f4\u63a5\u52a0\u5165 GAE \u53cd\u800c\u4f1a\u635f\u5bb3\u8868\u73b0\u3002\u8fd9\u8bf4\u660e\u8f85\u52a9\u5956\u52b1\u9002\u5408\u5e2e\u52a9 critic \u8bad\u7ec3\u548c\u9519\u8bef\u8bc6\u522b\uff0c\u4f46\u6700\u7ec8\u7b56\u7565\u4f18\u5316\u4ecd\u5e94\u4e0e\u771f\u5b9e\u4efb\u52a1\u76ee\u6807\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u81ea\u8fdb\u5316\u6548\u679c\u65b9\u9762\uff0cFigure 3 \u5c55\u793a\u4e86\u4ece BC \u5230 Iter 1 \u518d\u5230 Iter 2 \u7684\u8fde\u7eed\u63d0\u5347\u3002\u591a\u4e2a\u4efb\u52a1\u4e0a\uff0c\u7b2c\u4e8c\u8f6e\u6f14\u5316\u76f8\u8f83\u7b2c\u4e00\u8f6e\u7ee7\u7eed\u63d0\u9ad8\uff0c\u8bf4\u660e Q-Evolve \u4e0d\u662f\u4e00\u6b21\u6027\u63d0\u5347\uff0c\u800c\u662f\u771f\u6b63\u80fd\u591f\u901a\u8fc7\u6570\u636e\u5237\u65b0\u3001critic \u91cd\u5b66\u548c\u7b56\u7565\u518d\u4f18\u5316\u5b9e\u73b0\u7a33\u5b9a\u8fed\u4ee3\u6539\u8fdb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6837\u672c\u6548\u7387\u5b9e\u9a8c\u4e5f\u5f88\u6709\u8bf4\u670d\u529b\u3002Table 5 \u663e\u793a\uff0c\u5728 AlfWorld \u4e0a\uff0cQ-Evolve \u53ea\u4f7f\u7528 13K \u73af\u5883\u6b65\u6570\uff0c\u5c31\u8d85\u8fc7\u4e86\u4f7f\u7528 320K \u73af\u5883\u6b65\u6570\u7684 PPO\u3001RLOO\u3001GRPO \u53ca\u5176 SFT \u53d8\u4f53\u3002\u8fd9\u4e2a\u7ed3\u679c\u8bf4\u660e\uff0c\u5728\u7a00\u758f\u5956\u52b1\u957f\u7a0b\u4efb\u52a1\u4e2d\uff0c\u5355\u7eaf\u589e\u52a0 online RL \u91c7\u6837\u5e76\u4e0d\u4e00\u5b9a\u9ad8\u6548\uff0c\u5408\u7406\u7684\u8fc7\u7a0b\u5956\u52b1\u751f\u6210\u548c in-distribution policy learning \u53ef\u80fd\u66f4\u52a0\u5173\u952e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u540e\uff0cTable 6 \u4f7f\u7528 Llama-3-8B-Instruct \u8fdb\u4e00\u6b65\u9a8c\u8bc1\u6a21\u578b\u6cdb\u5316\u6027\u3002Q-Evolve \u5728 WebShop\u3001ScienceWorld \u548c AlfWorld \u4e0a\u90fd\u4f18\u4e8e SFT\u3001ETO\u3001KnowAgent\u3001WKM \u548c ETO plus MPO \u7b49\u65b9\u6cd5\uff0c\u8bf4\u660e\u8be5\u6846\u67b6\u4e0d\u662f\u53ea\u5bf9\u67d0\u4e00\u4e2a base model \u6709\u6548\uff0c\u800c\u662f\u5177\u6709\u4e00\u5b9a\u67b6\u6784\u548c\u89c4\u6a21\u8fc1\u79fb\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"413\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-15-1024x413.png\"  class=\"wp-image-1454\" style=\"aspect-ratio:2.482223658694247;width:623px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-15-1024x413.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-15-300x121.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-15-768x309.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-15.png 1534w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe4\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe4\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"762\" height=\"512\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-16.png\"  class=\"wp-image-1455\" style=\"width:359px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-16.png 762w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-16-300x202.png 300w\" sizes=\"auto, (max-width: 762px) 100vw, 762px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe5\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe5\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"696\" height=\"288\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-17.png\"  class=\"wp-image-1456\" style=\"width:425px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-17.png 696w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-17-300x124.png 300w\" sizes=\"auto, (max-width: 696px) 100vw, 696px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe6\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"666\" height=\"402\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-18.png\"  class=\"wp-image-1457\" style=\"width:398px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-18.png 666w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-18-300x181.png 300w\" sizes=\"auto, (max-width: 666px) 100vw, 666px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe7\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe7\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"754\" height=\"988\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-19.png\"  class=\"wp-image-1458\" style=\"aspect-ratio:0.7631631954235991;width:399px;height:auto\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-19.png 754w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/06\/image-19-229x300.png 229w\" sizes=\"auto, (max-width: 754px) 100vw, 754px\" title=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe8\" alt=\"Self-evolving LLM Agents with In-distribution Optimization\u63d2\u56fe8\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">7. \u8d21\u732e\u4e0e\u7ed3\u8bba\uff08Conclusion\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u672c\u6587\u7684\u4e3b\u8981\u8d21\u732e\u5305\u62ec\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\uff0c\u63d0\u51fa Q-Evolve\uff0c\u5c06\u81ea\u52a8\u8fc7\u7a0b\u5956\u52b1\u6807\u6ce8\u548c\u7b56\u7565\u5b66\u4e60\u7edf\u4e00\u5230\u4e00\u4e2a in-distribution self-evolving framework \u4e2d\u3002\u76f8\u6bd4\u4f20\u7edf PRM pipeline\uff0c\u5b83\u66f4\u91cd\u89c6\u8fc7\u7a0b\u5956\u52b1\u548c\u7b56\u7565\u4f18\u5316\u4e4b\u95f4\u7684\u6570\u636e\u5206\u5e03\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e8c\uff0c\u8bbe\u8ba1 hybrid offline dataset\uff0c\u5c06\u4e13\u5bb6\u793a\u8303\u548c agent \u81ea\u751f\u6210\u8f68\u8ff9\u7ed3\u5408\u8d77\u6765\u3002\u4e13\u5bb6\u6570\u636e\u63d0\u4f9b\u6210\u529f\u8def\u5f84\u548c\u9ad8\u8d28\u91cf\u6307\u5bfc\uff0c\u81ea\u751f\u6210\u6570\u636e\u66b4\u9732\u6a21\u578b\u771f\u5b9e\u9519\u8bef\u548c\u884c\u4e3a\u5206\u5e03\uff0c\u4e24\u8005\u5171\u540c\u63d0\u5347 critic learning \u7684\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e09\uff0c\u5f15\u5165 weighted IQL \u548c GAE\uff0c\u5c06\u7a00\u758f\u7ec8\u5c40\u5956\u52b1\u8f6c\u5316\u4e3a step-wise process rewards\u3002\u8fd9\u6837\u65e2\u907f\u514d\u4eba\u5de5\u9010\u6b65\u6807\u6ce8\uff0c\u4e5f\u4e0d\u9700\u8981\u73af\u5883\u56de\u6eaf\u6216\u5927\u91cf\u5728\u7ebf\u641c\u7d22\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u56db\uff0c\u91c7\u7528 behavior-proximal policy optimization\uff0c\u5728\u6570\u636e\u652f\u6301\u8303\u56f4\u5185\u8fdb\u884c\u4fdd\u5b88\u7b56\u7565\u66f4\u65b0\u3002\u8be5\u65b9\u6cd5\u4e0d\u4ec5\u5f3a\u5316\u6b63\u4f18\u52bf\u52a8\u4f5c\uff0c\u4e5f\u538b\u4f4e\u8d1f\u4f18\u52bf\u52a8\u4f5c\uff0c\u6709\u52a9\u4e8e\u7ea0\u6b63\u957f\u7a0b\u51b3\u7b56\u4e2d\u7684\u9519\u8bef\u884c\u4e3a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e94\uff0c\u5728 AlfWorld\u3001WebShop \u548c ScienceWorld \u4e0a\u53d6\u5f97\u7a33\u5b9a\u63d0\u5347\uff0c\u5e76\u5728\u6837\u672c\u6548\u7387\u548c\u8de8\u6a21\u578b\u6cdb\u5316\u65b9\u9762\u8868\u73b0\u7a81\u51fa\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u7814\u7a76\u610f\u4e49\u6765\u770b\uff0c\u8fd9\u7bc7\u8bba\u6587\u7684\u91cd\u70b9\u4e0d\u662f\u5355\u7eaf\u63d0\u51fa\u4e00\u4e2a\u66f4\u5f3a\u7684 agent\uff0c\u800c\u662f\u56de\u7b54\u4e86\u957f\u7a0b LLM Agent \u8bad\u7ec3\u4e2d\u7684\u4e00\u4e2a\u57fa\u7840\u95ee\u9898\uff0c\u5373\u5728\u53ea\u6709\u7ec8\u5c40\u5956\u52b1\u7684\u60c5\u51b5\u4e0b\uff0c\u5982\u4f55\u53ef\u9760\u5730\u751f\u6210\u8fc7\u7a0b\u76d1\u7763\uff0c\u5e76\u8ba9 agent \u5728\u4e0d\u4e25\u91cd\u504f\u79bb\u6570\u636e\u5206\u5e03\u7684\u524d\u63d0\u4e0b\u6301\u7eed\u81ea\u6211\u6539\u8fdb\u3002Q-Evolve \u7684\u7b54\u6848\u662f\u8ba9 policy\u3001critic \u548c dataset \u5171\u540c\u6f14\u5316\uff0c\u540c\u65f6\u4fdd\u8bc1\u6bcf\u4e00\u8f6e\u66f4\u65b0\u4ecd\u7136\u5728\u5f53\u524d\u6570\u636e\u5206\u5e03\u5185\u5b8c\u6210\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4f53\u6765\u770b\uff0c\u8be5\u5de5\u4f5c\u5bf9 LLM Agent \u7684\u8bad\u7ec3\u6709\u8f83\u5f3a\u53c2\u8003\u4ef7\u503c\u3002\u5b83\u628a self-evolving agent \u4ece\u6982\u5ff5\u5c42\u9762\u63a8\u8fdb\u5230\u4e86\u4e00\u4e2a\u66f4\u5177\u4f53\u7684\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u4e2d\uff0c\u4e5f\u8bf4\u660e\u672a\u6765 agent \u8bad\u7ec3\u4e0d\u80fd\u53ea\u5173\u6ce8\u6700\u7ec8\u4efb\u52a1\u6210\u529f\u7387\uff0c\u8fd8\u9700\u8981\u5173\u6ce8\u8fc7\u7a0b\u5956\u52b1\u7684\u53ef\u9760\u6027\u3001\u6570\u636e\u5206\u5e03\u7684\u4e00\u81f4\u6027\u548c\u7b56\u7565\u66f4\u65b0\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. \u6458\u8981\uff08Abstract\uff09 \u672c\u6587\u7814\u7a76\u7684\u662f\u957f\u7a0b\u4ea4\u4e92\u578b LLM Agent \u7684\u8bad\u7ec3\u95ee\u9898\uff0c\u6838\u5fc3\u5173\u6ce8\u70b9\u662f\u7a00\u758f\u5ef6\u8fdf\u5956\u52b1\u4e0b\u7684\u8d21\u732e\u5f52\u56e0\u3002\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u4ece\u9759\u6001\u6587\u672c\u751f\u6210\u9010\u6e10\u8d70\u5411\u73af\u5883\u4ea4\u4e92\uff0cLLM Agent \u9700\u8981\u5728\u7f51\u9875\u8d2d\u7269\u3001\u865a\u62df\u5b9e\u9a8c\u3001\u5bb6\u5c45\u4efb\u52a1\u7b49\u590d\u6742\u73af\u5883\u4e2d\u8fdb\u884c\u8fde\u7eed\u51b3\u7b56\u3002\u7136\u800c\uff0c\u8fd9\u7c7b\u4efb\u52a1\u901a\u5e38\u53ea\u6709\u5728\u6574\u4e2a episode \u7ed3\u675f\u540e\u624d\u7ed9\u51fa\u6700\u7ec8\u5956\u52b1\uff0c\u4e2d\u95f4\u6b65\u9aa4\u7f3a\u5c11\u660e\u786e\u76d1\u7763\uff0c\u5bfc\u81f4\u6a21\u578b\u5f88\u96be\u5224\u65ad\u54ea\u4e00\u6b65\u771f\u6b63\u63a8\u52a8\u4e86\u4efb\u52a1\u6210\u529f\uff0c\u54ea\u4e00\u6b65\u9020\u6210\u4e86\u5931\u8d25\u3002 \u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u8bba\u6587\u63d0\u51fa Q-Evolve\uff0c\u4e00\u79cd\u9762\u5411 LLM Agent \u7684\u81ea &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1449\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":7,"featured_media":1450,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-1449","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1449","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=1449"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1449\/revisions"}],"predecessor-version":[{"id":1459,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1449\/revisions\/1459"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1450"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1449"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1449"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1449"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}