{"id":6345,"date":"2025-10-15T12:16:44","date_gmt":"2025-10-15T04:16:44","guid":{"rendered":"https:\/\/nullthought.net\/?p=6345"},"modified":"2025-10-15T12:16:45","modified_gmt":"2025-10-15T04:16:45","slug":"deepmmsearch-r1%ef%bc%9a%e9%80%9a%e8%bf%87%e5%a4%9a%e5%b7%a5%e5%85%b7%e3%80%81%e5%a4%9a%e8%bd%ae%e4%ba%a4%e4%ba%92%e6%a3%80%e7%b4%a2%e5%a2%9e%e5%bc%ba%e6%8e%a8%e7%90%86%e5%be%aa","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6345","title":{"rendered":"DeepMMSearch-R1\uff1a\u901a\u8fc7\u201c\u591a\u5de5\u5177\u3001\u591a\u8f6e\u4ea4\u4e92\u201d\u68c0\u7d22\u589e\u5f3a\u63a8\u7406\u5faa\u73af\uff0c\u5b9e\u73b0\u9762\u5411\u771f\u5b9e\u7f51\u9875\u7684\u591a\u6a21\u6001\u68c0\u7d22\/\u63a8\u7406\u4e00\u4f53\u5316"},"content":{"rendered":"\n<p>\u73b0\u5b9e\u5e94\u7528\u4e2d\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u5728\u77e5\u8bc6\u5bc6\u96c6\u4e0e\u4fe1\u606f\u68c0\u7d22\u578b\u89c6\u89c9\u95ee\u7b54\u4efb\u52a1\u4e0a\u5e38\u53d7\u9650\u4e8e\u9759\u6001\u8bad\u7ec3\u8bed\u6599\u4e0e\u957f\u5c3e\u77e5\u8bc6\u5206\u5e03\uff0c\u96be\u4ee5\u53ca\u65f6\u83b7\u53d6\u6700\u65b0\u4e8b\u5b9e\u4e0e\u5f00\u653e\u4e16\u754c\u77e5\u8bc6\uff1b\u4f20\u7edf<a href=\"https:\/\/nullthought.net\/?tag=rag\" target=\"_blank\" rel=\"noreferrer noopener\">RAG<\/a>\u4e0e\u201c\u641c\u7d22\u4ee3\u7406\u201d\u65b9\u6848\u53c8\u5e38\u5b58\u5728\u68c0\u7d22\u7ba1\u7ebf\u50f5\u786c\u3001\u67e5\u8be2\u6784\u9020\u6b20\u4f73\u3001\u8fc7\u5ea6\u68c0\u7d22\u4e0e\u566a\u58f0\u6ce8\u5165\u7b49\u95ee\u9898\u3002\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2510.12801\" target=\"_blank\" rel=\"noreferrer noopener\">DeepMMSearch-R1: Empowering Multimodal LLMs in Multimodal Web Search<\/a><\/strong>\u63d0\u51fa\u7684 DeepMMSearch-R1 \u9488\u5bf9\u8fd9\u4e9b\u75db\u70b9\uff0c\u5f3a\u8c03\u201c\u6309\u9700\u3001\u591a\u8f6e\u201d\u7684\u7f51\u9875\u641c\u7d22\u3001\u57fa\u4e8e\u68c0\u7d22\u53cd\u9988\u7684\u81ea\u53cd\u601d\/\u81ea\u7ea0\u9519\uff0c\u4ee5\u53ca\u4ee5\u201c\u88c1\u526a\u2014\u68c0\u7d22\u201d\u534f\u540c\u7684\u56fe\u50cf\u68c0\u7d22\uff0c\u4ece\u800c\u5728\u771f\u5b9e\u7f51\u7edc\u73af\u5883\u4e2d\u66f4\u7a33\u5065\u5730\u5b8c\u6210\u591a\u6a21\u6001\u4fe1\u606f\u5bfb\u5740\u4e0e\u63a8\u7406\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aKartik Narayan, Yang Xu, Tian Cao, Kavya Nerella, Vishal M. Patel, Navid Shiee, Peter Grasch, Chao Jia, Yinfei Yang, Zhe Gan\uff0c\u6765\u81ea\u82f9\u679c\u516c\u53f8\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"924\" height=\"815\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-7.png\" alt=\"\" class=\"wp-image-6346\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-7.png 924w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-7-300x265.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/10\/image-7-768x677.png 768w\" sizes=\"auto, (max-width: 924px) 100vw, 924px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2510.12801\" target=\"_blank\" rel=\"noreferrer noopener\">DeepMMSearch-R1: Empowering Multimodal LLMs in Multimodal Web Search<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u4e00\u3001\u65b9\u6cd5\u603b\u89c8<br>DeepMMSearch-R1 \u7684\u6838\u5fc3\u662f\u4e00\u4e2a\u201c\u591a\u5de5\u5177\u3001\u591a\u8f6e\u4ea4\u4e92\u201d\u7684\u68c0\u7d22\u589e\u5f3a\u63a8\u7406\u5faa\u73af\uff1a<br>1\uff09\u6587\u672c\u641c\u7d22\uff1a\u6a21\u578b\u6839\u636e\u95ee\u9898\u4e0e\u5df2\u83b7\u4fe1\u606f\u52a8\u6001\u751f\u6210\/\u6539\u5199\u6587\u672c\u67e5\u8be2\uff0c\u591a\u8f6e\u8fed\u4ee3\u4ee5\u7f29\u5c0f\u68c0\u7d22\u8303\u56f4\uff0c\u5b9e\u73b0\u81ea\u53cd\u601d\u4e0e\u81ea\u7ea0\u9519\uff1b<br>2\uff09\u56fe\u50cf\u88c1\u526a+\u56fe\u50cf\u641c\u7d22\uff1a\u5148\u7531\u6a21\u578b\u751f\u6210\u201c\u6307\u4ee3\u8868\u8fbe\u201d\uff08referring expression\uff09\uff0c\u518d\u7528 Grounding DINO \u5bf9\u8f93\u5165\u56fe\u50cf\u8fdb\u884c\u5b9a\u4f4d\u4e0e\u88c1\u526a\uff0c\u6700\u540e\u4ee5\u88c1\u526a\u56fe\u4f5c\u4e3a\u68c0\u7d22\u952e\u63d0\u5347\u76f8\u4f3c\u56fe\u4e0e\u7f51\u9875\u4e0a\u4e0b\u6587\u7684\u76f8\u5173\u6027\u4e0e\u7eaf\u5ea6\uff1b<br>3\uff09\u4e24\u9636\u6bb5\u8bad\u7ec3\uff1a\u5148\u5728\u65b0\u6784\u9020\u7684\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\uff0c\u518d\u7528\u5728\u7ebf\u5f3a\u5316\u5b66\u4e60\uff08GRPO\uff09\u4f18\u5316\u5de5\u5177\u9009\u62e9\u4e0e\u641c\u7d22\u6548\u7387\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001\u6570\u636e\u96c6\uff1aDeepMMSearchVQA<br>\u4f5c\u8005\u6784\u5efa\u4e86 DeepMMSearchVQA\uff0c\u7528\u4e8e\u6559\u6388\u6a21\u578b\u201c\u4f55\u65f6\u641c\u3001\u641c\u4ec0\u4e48\u3001\u7528\u54ea\u79cd\u5de5\u5177\u3001\u5982\u4f55\u878d\u5408\u68c0\u7d22\u4fe1\u606f\u5e76\u51b3\u5b9a\u4e0b\u4e00\u6b65\u52a8\u4f5c\/\u4f5c\u7b54\u201d\u3002\u8be5\u6570\u636e\u96c6\u4ee5 InfoSeek \u4e3a\u57fa\u8bed\u6599\uff0c\u501f\u52a9 Gemini-2.5-Pro \u751f\u6210\u591a\u8f6e\u5bf9\u8bdd\uff0c\u6837\u672c\u4e2d\u663e\u5f0f\u5305\u542b\u5de5\u5177\u6807\u7b7e\uff08\u5982 &lt;text_search>\u3001&lt;img_search>\u3001&lt;answer>\uff09\u4e0e\u68c0\u7d22\u4fe1\u606f\u5757 &lt;information>\uff0c\u5e76\u533a\u5206\u201c\u6574\u56fe\u68c0\u7d22\u201d\u4e0e\u201c\u88c1\u526a\u68c0\u7d22\u201d\uff1b\u6700\u7ec8\u7b5b\u9009\u51fa\u7ea6 1 \u4e07\u4e2a\u8986\u76d6\u591a\u77e5\u8bc6\u95e8\u7c7b\u3001\u4e14\u5728\u201c\u9700\u68c0\u7d22\/\u514d\u68c0\u7d22\u201d\u95ee\u9898\u4e0a\u7ea6 50:50 \u5e73\u8861\u7684 VQA \u6837\u672c\u7528\u4e8e SFT\u3002<\/p>\n\n\n\n<p>\u4e09\u3001\u8bad\u7ec3\u7b56\u7565\uff1aSFT + \u5728\u7ebfGRPO<br>\uff081\uff09SFT\uff1a\u4ee5 Qwen2.5-VL-7B-Instruct \u4e3a\u57fa\u5ea7\uff0c\u4ec5\u5fae\u8c03 LLM \u6a21\u5757\u5e76\u51bb\u7ed3\u89c6\u89c9\u7f16\u7801\u4e0e\u6295\u5f71\u5c42\uff0c\u914d\u5408 LoRA\uff08r=8\uff09\u63d0\u5347\u53c2\u6570\u6548\u7387\u3002\u8bad\u7ec3\u76ee\u6807\u4e3a\u6807\u51c6\u81ea\u56de\u5f52 LM \u635f\u5931\uff0c\u5e76\u5bf9\u68c0\u7d22\u56de\u4f20\u6587\u672c\u8fdb\u884c\u635f\u5931\u5c4f\u853d\uff0c\u805a\u7126\u4e8e\u201c\u63a8\u7406\u4e0e\u5de5\u5177\u8c03\u7528\u683c\u5f0f\u201d\u7684\u5b66\u4e60\u3002<br>\uff082\uff09GRPO\uff1a\u4ece SFT \u68c0\u70b9\u51fa\u53d1\uff0c\u5728\u771f\u5b9e\u5de5\u5177\u56de\u8def\u4e2d\u591a\u6837\u5316\u751f\u6210\u591a\u6761\u8f68\u8ff9\uff0c\u4ee5\u7ec4\u5185\u76f8\u5bf9\u4f18\u52bf\u4e3a\u6838\u5fc3\u7684 GRPO \u76ee\u6807\u7a33\u5b9a\u4f18\u5316\uff1b\u5956\u52b1\u7531\u201c\u7b54\u6848\u6b63\u786e\u6027\u201d\u4e0e\u201c\u683c\u5f0f\u5408\u89c4\u201d\u7ec4\u6210\uff0c\u5e76\u52a0\u5165 KL \u6b63\u5219\u7ea6\u675f\u7b56\u7565\u6f02\u79fb\u3002\u8bad\u7ec3\u8fd8\u9650\u5236\u5de5\u5177\u8c03\u7528\u6b21\u6570\u4e0e\u8f68\u8ff9\u957f\u5ea6\uff0c\u4ee5\u9f13\u52b1\u201c\u51c6\u786e\u2014\u9ad8\u6548\u201d\u7684\u7b56\u7565\u3002<\/p>\n\n\n\n<p>\u56db\u3001\u591a\u6a21\u6001\u641c\u7d22\u6d41\u6c34\u7ebf\u4e0e\u5de5\u5177\u534f\u4f5c<br>\u7cfb\u7edf\u5305\u542b\u4e09\u7c7b\u5de5\u5177\u4e0e\u4e24\u4e2a\u201c\u6458\u8981\u201d\u6b65\u9aa4\uff1a<br>1\uff09\u6587\u672c\u641c\u7d22\uff1a\u5185\u90e8Web\u7d22\u5f15\u2192Top-k \u6587\u6863\u2192LLM\u6458\u8981\u2192\u56de\u704c\u4e0a\u4e0b\u6587\uff1b<br>2\uff09\u56fe\u50cf\u641c\u7d22\uff1a\u6574\u56fe\u6216\u88c1\u526a\u56fe\u2192\u76f8\u4f3c\u56fe\u4e0e\u9875\u9762\u5143\u4fe1\u606f\u2192LLM\u6458\u8981\u2192\u56de\u704c\u4e0a\u4e0b\u6587\uff1b<br>3\uff09\u88c1\u526a\uff08Grounding DINO\uff09\uff1a\u6839\u636e\u6a21\u578b\u751f\u6210\u7684\u6307\u4ee3\u8868\u8fbe\u5b9a\u4f4d\u6700\u76f8\u5173\u533a\u57df\u5e76\u88c1\u526a\uff0c\u7f13\u89e3\u6574\u56fe\u566a\u58f0\u5e72\u6270\u3002\u8be5\u8bbe\u8ba1\u663e\u8457\u63d0\u5347\u68c0\u7d22\u6709\u6548\u6027\u5e76\u51cf\u5c11\u65e0\u5173\u566a\u58f0\u8fdb\u5165\u4e0a\u4e0b\u6587\u3002<\/p>\n\n\n\n<p>\u4e94\u3001\u5b9e\u9a8c\u8bbe\u7f6e\u4e0e\u8bc4\u4ef7<br>\u4f5c\u8005\u8bbe\u8ba1\u4e86\u56db\u7c7b\u8bc4\u6d4b\u5de5\u4f5c\u6d41\uff1a\u2460 \u76f4\u63a5\u4f5c\u7b54\uff1b\u2461 \u521a\u6027RAG\u4e24\u6b65\uff08\u5148\u56fe\u641c\u540e\u6587\u641c\uff0c\u6bcf\u6b65\u56fa\u5b9a\u4e00\u6b21\uff09\uff1b\u2462 \u63d0\u793a\u5f0f\u641c\u7d22\u4ee3\u7406\uff08\u6d4b\u8bd5\u65f6\u901a\u8fc7\u63d0\u793a\u5f15\u5bfc\u4f7f\u7528\u5de5\u5177\uff09\uff1b\u2463 \u5177\u5907\u7f51\u9875\u641c\u7d22\u80fd\u529b\u7684MLLM\uff08\u53ef\u591a\u8f6e\u6309\u9700\u641c\u7d22\uff09\u3002\u8bc4\u6d4b\u6570\u636e\u8986\u76d6 InfoSeek\u3001DynVQA\u3001SimpleVQA\u3001Enc-VQA\u3001OK-VQA\u3001A-OKVQA \u7b49\uff1b\u6307\u6807\u91c7\u7528\u201cLLM \u8bc4\u5ba1\u201d\u6846\u67b6\uff08gpt-5-chat-latest \u4f5c\u4e3a\u88c1\u5224\uff09\u4ee5\u66f4\u597d\u523b\u753b\u5f00\u653e\u5f0f\u591a\u6a21\u6001\u56de\u7b54\u7684\u8bed\u4e49\u6b63\u786e\u6027\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u7ed3\u679c\u4e0e\u5206\u6790<br>1\uff09\u603b\u4f53\u6548\u679c\uff1aDeepMMSearch-R1-7B\uff08RL\uff09\u5728\u516d\u9879\u57fa\u51c6\u4e0a\u5e73\u5747 57.13 \u5206\uff0c\u663e\u8457\u4f18\u4e8e\u63d0\u793a\u5f0f\u641c\u7d22\u4ee3\u7406\u4e0e\u521a\u6027RAG\u6d41\u7a0b\uff0c\u63a5\u8fd1 o3 \u7684\u6c34\u5e73\uff1b\u540c\u4e00\u6a21\u578b\u5728 SFT\u2192RL \u8fc7\u7a0b\u4e2d\u8fdb\u4e00\u6b65\u63d0\u5347\uff0c\u4f53\u73b0\u5728\u7ebf\u4e0a\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\u5bf9\u201c\u4f55\u65f6\u641c\/\u5982\u4f55\u641c\u201d\u7684\u7b56\u7565\u4f18\u5316\u4ef7\u503c\u3002<br>2\uff09RAG\u4e0e\u4ee3\u7406\u5bf9\u6bd4\uff1a\u521a\u6027\u4e24\u6b65RAG\u5728 OK-VQA\u3001A-OKVQA \u4e0a\u660e\u663e\u52a3\u4e8e\u76f4\u63a5\u4f5c\u7b54\uff0c\u539f\u56e0\u5728\u4e8e\u8fd9\u4e24\u7c7b\u6570\u636e\u591a\u4e3a\u201c\u53ef\u514d\u68c0\u7d22\u201d\u95ee\u9898\uff0c\u5f3a\u884c\u5f15\u5165\u7f51\u7edc\u68c0\u7d22\u53cd\u800c\u5e26\u6765\u566a\u58f0\uff1b\u63d0\u793a\u5f0f\u641c\u7d22\u4ee3\u7406\u867d\u66f4\u7a33\uff0c\u4f46\u4ecd\u4e0d\u53ca\u201c\u5177\u5907\u641c\u7d22\u80fd\u529b\u7684MLLM\u201d\uff0c\u8bf4\u660e\u4ec5\u9760\u63d0\u793a\u5728\u6d4b\u8bd5\u65f6\u5b66\u4f1a\u7528\u5de5\u5177\u4e0d\u5982\u8bad\u7ec3\u65f6\u663e\u5f0f\u5bf9\u63a5\u5de5\u5177\u4e0e\u7ed3\u6784\u5316\u534f\u8bae\u3002<br>3\uff09\u88c1\u526a\u68c0\u7d22\u4e0e\u81ea\u53cd\u601d\/\u81ea\u7ea0\u9519\uff1a\u5141\u8bb8\u591a\u6b21\u6587\u672c\u68c0\u7d22\u4e0e\u88c1\u526a\u5f0f\u56fe\u50cf\u68c0\u7d22\u5e26\u6765\u7a33\u5b9a\u589e\u76ca\uff0c\u5c24\u5176\u5bf9\u9700\u8981\u9501\u5b9a\u5355\u4e00\u89c6\u89c9\u5b9e\u4f53\u7684\u95ee\u9898\u66f4\u660e\u663e\uff1b\u540c\u65f6\uff0cSFT\u2192RL \u540e\uff0c\u6a21\u578b\u5728\u5fc5\u8981\u65f6\u624d\u8fdb\u884c\u88c1\u526a\uff0c\u51cf\u5c11\u4e86\u4e0d\u5fc5\u8981\u7684\u88c1\u526a\u8c03\u7528\uff08DynVQA \u221236.81%\u3001OK-VQA \u221234.86%\uff09\uff0c\u4f46\u6574\u4f53\u6027\u80fd\u4ecd\u63d0\u5347\uff0c\u4f53\u73b0\u201c\u66f4\u5c11\u4f46\u66f4\u51c6\u201d\u7684\u5de5\u5177\u4f7f\u7528\u3002<br>4\uff09\u6570\u636e\u914d\u6bd4\u4e0e\u62bd\u6837\u7b56\u7565\uff1aSFT \u6570\u636e\u4e2d\u201c\u9700\u68c0\u7d22\uff1a\u514d\u68c0\u7d22\u201d\u4fdd\u6301\u7ea6 50:50 \u6700\u4f18\uff1b\u5e76\u4e14\u6309\u77e5\u8bc6\u95e8\u7c7b\u5747\u5300\u91c7\u6837\u4f18\u4e8e\u968f\u673a\u62bd\u6837\uff0c\u80fd\u66f4\u597d\u84b8\u998f\u201c\u4f55\u65f6\u641c\/\u5982\u4f55\u641c\u201d\u7684\u884c\u4e3a\u5206\u5e03\u3002<br>5\uff09\u4e00\u822cVQA\u80fd\u529b\u4fdd\u6301\uff1a\u5728 OCRBench\u3001MMVet\u3001AI2D\u3001MathVista\u3001MMBench\u3001DocVQA\u3001InfoVQA \u7b49\u901a\u7528\u57fa\u51c6\u4e0a\uff0cDeepMMSearch-R1-7B\uff08RL\uff09\u4e0e\u57fa\u5ea7 Qwen2.5-VL-7B \u8868\u73b0\u76f8\u5f53\uff0c\u8bf4\u660e LoRA-SFT \u4e0e\u5e26 KL \u7684\u5728\u7ebf GRPO \u5e76\u672a\u727a\u7272\u901a\u7528\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u4e03\u3001\u521b\u65b0\u70b9\u4e0e\u76f8\u5bf9\u4f4d\u7f6e<br>1\uff09\u201c\u88c1\u526a\u2014\u68c0\u7d22\u201d\u4e00\u4f53\uff1a\u533a\u522b\u4e8e\u4ec5\u652f\u6301\u6574\u56fe\u68c0\u7d22\u7684\u591a\u6a21\u6001\u68c0\u7d22\u6a21\u578b\uff0c\u5f15\u5165\u6307\u4ee3\u8868\u8fbe+Grounding DINO \u7684\u4e2d\u95f4\u88c1\u526a\u5de5\u5177\uff0c\u964d\u4f4e\u80cc\u666f\u566a\u58f0\u5bf9\u56fe\u641c\u7684\u5e72\u6270\uff1b<br>2\uff09\u81ea\u53cd\u601d\/\u81ea\u7ea0\u9519\u7684\u591a\u8f6e\u6587\u672c\u68c0\u7d22\uff1a\u7528\u68c0\u7d22\u6458\u8981\u53cd\u9988\u6765\u8fed\u4ee3\u4fee\u6b63\u67e5\u8be2\uff0c\u7f13\u89e3\u201c\u4e00\u6b21\u5b9a\u578b\u201d\u7684\u67e5\u8be2\u8106\u5f31\u6027\uff1b<br>3\uff09\u6570\u636e\u4e0e\u8bad\u7ec3\u4e00\u4f53\u5316\uff1a\u4ee5\u7ed3\u6784\u5316\u5de5\u5177\u6807\u7b7e\u4e0e\u4fe1\u606f\u5757\u6784\u6210\u7684\u591a\u8f6e VQA \u8bed\u6599\uff08DeepMMSearchVQA\uff09\u914d\u5408 SFT\u2192GRPO\uff0c\u4f7f\u201c\u4f1a\u7528\u5de5\u5177\u201d\u7684\u80fd\u529b\u6210\u4e3a\u6a21\u578b\u53c2\u6570\u7684\u4e00\u90e8\u5206\uff0c\u800c\u975e\u4ec5\u9760\u63d0\u793a\u5de5\u7a0b\u3002<\/p>\n\n\n\n<p>\u516b\u3001\u5c40\u9650\u6027\u4e0e\u6f5c\u5728\u98ce\u9669<br>1\uff09\u5bf9\u5916\u90e8\u57fa\u7840\u8bbe\u65bd\u4f9d\u8d56\uff1a\u5b9e\u9645\u6548\u679c\u53d7\u641c\u7d22API\u4e0e\u7d22\u5f15\u8986\u76d6\u5ea6\u3001\u6458\u8981\u8d28\u91cf\u5f71\u54cd\uff1b<br>2\uff09\u68c0\u7d22\u504f\u89c1\u4e0e\u7248\u6743\u5408\u89c4\uff1a\u81ea\u52a8\u6458\u8981\u53ef\u80fd\u653e\u5927\u504f\u5dee\u6216\u65e7\u95fb\uff0c\u4e14\u6d89\u53ca\u5185\u5bb9\u6765\u6e90\u5408\u89c4\u4e0e\u53ef\u8ffd\u6eaf\uff1b<br>3\uff09\u591a\u8f6e\u6210\u672c\u4e0e\u65f6\u5ef6\uff1a\u5c3d\u7ba1 RL \u964d\u4f4e\u4e86\u4e0d\u5fc5\u8981\u5de5\u5177\u8c03\u7528\uff0c\u4f46\u5728\u9ad8\u5e76\u53d1\u5e94\u7528\u4e2d\u4ecd\u9700\u7cbe\u7ec6\u7684\u8c03\u7528\u9884\u7b97\u4e0e\u7f13\u5b58\u7b56\u7565\u3002<\/p>\n\n\n\n<p>\u4e5d\u3001\u5bf9\u5de5\u7a0b\u5b9e\u8df5\u7684\u542f\u793a<br>1\uff09\u628a\u201c\u662f\u5426\u68c0\u7d22\u201d\u7684\u51b3\u7b56\u663e\u5f0f\u5316\uff1a\u5728\u4e1a\u52a1\u5c42\u63a5\u5165\u7b56\u7565\uff08\u5982\u53ef\u7f6e\u4fe1\u5ea6\u9608\u503c\u3001\u95ee\u9898\u7c7b\u522b\u767d\u540d\u5355\uff09\u4ee5\u907f\u514d\u5bf9\u201c\u514d\u68c0\u7d22\u95ee\u9898\u201d\u5f3a\u884c\u68c0\u7d22\u9020\u6210\u7684\u566a\u58f0\uff1b<br>2\uff09\u4f18\u5148\u505a\u201c\u6307\u4ee3\u2014\u88c1\u526a\u2014\u68c0\u7d22\u201d\u7684\u56fe\u641c\u901a\u8def\uff1a\u5bf9\u201c\u8bc6\u522b\u67d0\u4e00\u5c40\u90e8\u5b9e\u4f53\u5e76\u8865\u5168\u77e5\u8bc6\u201d\u7684\u573a\u666f\uff08\u5546\u6807\u3001\u5668\u4ef6\u3001\u5fbd\u6807\u3001\u52a8\u690d\u7269\uff09\u5c24\u4e3a\u5173\u952e\uff1b<br>3\uff09\u68c0\u7d22\u7ed3\u679c\u4e8c\u6b21\u84b8\u998f\uff1a\u5c06 Top-k \u7ed3\u679c\u505a\u7ed3\u6784\u5316\u6458\u8981\uff08\u5b57\u6bb5\u5316\u8981\u70b9\u3001\u8bc1\u636e\u7247\u6bb5\u4e0e\u6765\u6e90\uff09\u518d\u56de\u704c\uff0c\u63a7\u5236\u4e0a\u4e0b\u6587\u6c61\u67d3\uff1b<br>4\uff09\u6570\u636e\u5408\u6210\u4e0e\u6bd4\u4f8b\u63a7\u5236\uff1a\u81ea\u5efa\u6307\u4ee4\u6570\u636e\u65f6\u7ef4\u6301\u201c\u9700\/\u514d\u68c0\u7d22=1:1\u201d\u4e0e\u77e5\u8bc6\u95e8\u7c7b\u5747\u8861\uff0c\u6709\u52a9\u4e8e\u5b66\u5230\u66f4\u201c\u7ecf\u6d4e\u201d\u7684\u5de5\u5177\u4f7f\u7528\u7b56\u7565\uff1b<br>5\uff09\u8054\u8bad\u4e0e\u5b88\u62a4\uff1a\u91c7\u7528\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff08LoRA\uff09\u4e0e KL \u7ea6\u675f\u7684\u5728\u7ebf\u4f18\u5316\uff0c\u53ef\u5728\u4e0d\u7834\u574f\u901a\u7528\u80fd\u529b\u7684\u524d\u63d0\u4e0b\u7a33\u6b65\u6ce8\u5165\u5de5\u5177\u4f7f\u7528\u6280\u80fd\u3002<\/p>\n\n\n\n<p>\u5341\u3001\u7ed3\u8bba<br>DeepMMSearch-R1 \u5c06\u201c\u591a\u8f6e\u81ea\u53cd\u601d\u7684\u6587\u672c\u68c0\u7d22\u201d\u4e0e\u201c\u88c1\u526a\u5148\u884c\u7684\u56fe\u50cf\u68c0\u7d22\u201d\u7eb3\u5165\u7edf\u4e00\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u6846\u67b6\uff0c\u5e76\u901a\u8fc7 SFT\u2192GRPO \u7684\u4e24\u9636\u6bb5\u8bad\u7ec3\u628a\u201c\u4f1a\u7528\u5de5\u5177\u201d\u7684\u7b56\u7565\u56fa\u5316\u8fdb\u6a21\u578b\uff0c\u5728\u591a\u9879\u77e5\u8bc6\u5bc6\u96c6\u578b\u57fa\u51c6\u4e0a\u8d85\u8d8a\u5f00\u6e90\u57fa\u7ebf\u3001\u63a5\u8fd1\u5f3a\u95ed\u6e90\u6a21\u578b\uff0c\u4e14\u666e\u9002VQA\u80fd\u529b\u57fa\u672c\u4fdd\u6301\u3002\u5b83\u4e3a\u201c\u9762\u5411\u771f\u5b9e\u7f51\u9875\u7684\u591a\u6a21\u6001\u68c0\u7d22\u2014\u63a8\u7406\u4e00\u4f53\u5316\u201d\u63d0\u4f9b\u4e86\u53ef\u590d\u7528\u8303\u5f0f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u73b0\u5b9e\u5e94\u7528\u4e2d\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u5728\u77e5\u8bc6\u5bc6\u96c6\u4e0e\u4fe1\u606f\u68c0\u7d22\u578b\u89c6\u89c9\u95ee\u7b54\u4efb\u52a1\u4e0a\u5e38\u53d7\u9650\u4e8e\u9759\u6001\u8bad\u7ec3\u8bed\u6599\u4e0e\u957f\u5c3e\u77e5\u8bc6\u5206\u5e03\uff0c\u96be [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,90,86,85],"class_list":["post-6345","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-multimodal","tag-llm","tag-apple"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u73b0\u5b9e\u5e94\u7528\u4e2d\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u5728\u77e5\u8bc6\u5bc6\u96c6\u4e0e\u4fe1\u606f\u68c0\u7d22\u578b\u89c6\u89c9\u95ee\u7b54\u4efb\u52a1\u4e0a\u5e38\u53d7\u9650\u4e8e\u9759\u6001\u8bad\u7ec3\u8bed\u6599\u4e0e\u957f\u5c3e\u77e5\u8bc6\u5206\u5e03\uff0c\u96be&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6345","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6345"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6345\/revisions"}],"predecessor-version":[{"id":6347,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6345\/revisions\/6347"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6345"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6345"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6345"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}