
{"id":10775,"date":"2026-08-05T18:44:38","date_gmt":"2026-08-05T10:44:38","guid":{"rendered":"https:\/\/infernews.com\/blog\/icml-2026-multimodal-deep-research-mllm-and-benchmark-the-first-long-horizon-mul\/"},"modified":"2026-08-05T18:44:38","modified_gmt":"2026-08-05T10:44:38","slug":"icml-2026-multimodal-deep-research-mllm-and-benchmark-the-first-long-horizon-mul","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/icml-2026-multimodal-deep-research-mllm-and-benchmark-the-first-long-horizon-mul\/","title":{"rendered":"Vision-DeepResearch\uff1a\u7531\u975c\u614b\u5716\u7247\u8d70\u5411\u9023\u7e8c\u5f71\u7247\u7684 DeepResearch Agent"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-e47ad552dd7e.jpg\" alt=\"icon\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9762\u5c0d\u9700\u8981\u7ffb\u67e5\u5f71\u7247\u5167\u5bb9\u3001\u518d\u7d50\u5408\u7db2\u7d61\u8cc7\u6599\u56de\u7b54\u7684\u554f\u984c\uff0c\u55ae\u9760\u6587\u5b57\u641c\u5c0b\u5f80\u5f80\u6703\u6f0f\u6389\u95dc\u9375\u756b\u9762\u3002Video-DeepResearch \u662f\u4e00\u500b\u591a\u6a21\u614b\u7814\u7a76\u578b Agent \u9805\u76ee\uff0c\u900f\u904e Video-DeepResearch\uff08Video-DR\uff09\u8655\u7406\u9023\u7e8c\u5f71\u7247\u4e2d\u7684\u6642\u7a7a\u8cc7\u8a0a\uff0c\u4e26\u628a\u8996\u89ba\u7406\u89e3\u8207\u7db2\u7d61\u63a2\u7d22\u5206\u958b\u5b89\u6392\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u4fee\u6b63\u4e86\u5169\u500b\u5e38\u898b\u5361\u4f4d\uff1a\u6a21\u578b\u504f\u5411\u4f7f\u7528\u6587\u5b57\u5de5\u5177\uff0c\u8f03\u5c11\u4e3b\u52d5\u6aa2\u8996\u5f71\u7247\uff1b\u6a21\u578b\u4ea6\u53ef\u80fd\u76f4\u63a5\u4f9d\u8cf4\u5167\u90e8\u8a18\u61b6\uff0c\u672a\u6709\u771f\u6b63\u5b8c\u6210\u5de5\u5177\u8f14\u52a9\u641c\u5c0b\u3002Pipeline \u5148\u9010\u968e\u6bb5\u89e3\u9396\u8996\u89ba\u5de5\u5177\uff0c\u8981\u6c42\u6a21\u578b\u5b8c\u6210\u8de8\u756b\u9762 grounding\uff0c\u518d\u9032\u5165\u7db2\u7d61\u6aa2\u7d22\uff0c\u53d6\u6368\u662f\u6d41\u7a0b\u8f03\u56b4\u8b39\uff0c\u4f46\u63a8\u7406\u6210\u672c\u548c\u57f7\u884c\u6642\u9593\u4ea6\u53ef\u80fd\u589e\u52a0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8a13\u7df4\u6d41\u7a0b\u5148\u4ee5 Supervised Fine-Tuning\uff08SFT\uff09\u5efa\u7acb\u57fa\u672c\u80fd\u529b\uff0c\u518d\u4f7f\u7528 Group Relative Policy Optimization\uff08GRPO\uff09\u5f37\u5316\u81ea\u4e3b\u63a2\u7d22\u3002\u9805\u76ee\u540c\u6642\u63d0\u4f9b 30 K \u500b video-grounded QA pairs\u30017 K \u689d\u6574\u7406\u5f8c\u7684 trajectories\uff0c\u4ee5\u53ca\u7a0b\u5f0f\u78bc\u3001\u8cc7\u6599\u96c6\u548c\u6a21\u578b\u6b0a\u91cd\uff0c\u7814\u7a76\u5718\u968a\u53ef\u6309\u9700\u8981\u6e2c\u8a66\u57fa\u6e96\u3001\u91cd\u73fe\u8a13\u7df4\u6216\u76f4\u63a5\u8f09\u5165\u6a21\u578b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Video-DR-35B-A3B \u5728 Video-DR \u9054\u5230 68.0% accuracy<\/li>\n<li>\u6bd4 Claude-4.5-Sonnet \u7684 63.0% \u9ad8 5.0 \u500b\u767e\u5206\u9ede<\/li>\n<li>GPT-5 \u548c Gemini 2.5 Pro \u5206\u5225\u70ba 57.0% \u548c 62.0%<\/li>\n<li>Vision-DeepResearch-30B-A3B \u5ef6\u7e8c\u540c\u4e00\u7814\u7a76\u65b9\u5411\uff0c\u53e6\u6709 SFT-only \u7684 8B \u7248\u672c<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u7d50\u679c\u53cd\u6620\u5b83\u5728\u5f71\u7247\u8b49\u64da\u8207\u7db2\u7d61\u8cc7\u6599\u9700\u8981\u4e92\u76f8\u9a57\u8b49\u7684\u5de5\u4f5c\u8f03\u6709\u50f9\u503c\uff0c\u4f8b\u5982\u7814\u7a76\u3001\u5a92\u9ad4\u6838\u67e5\u548c\u9577\u7247\u6bb5\u554f\u7b54\uff1b\u4f46 68.0% \u4ecd\u4ee3\u8868\u90e8\u5206\u554f\u984c\u6703\u51fa\u932f\uff0c\u8f03\u9069\u5408\u4f5c\u70ba\u7814\u7a76\u5e73\u53f0\u548c\u53ef\u6aa2\u9a57\u7684 Agent \u67b6\u69cb\uff0c\u800c\u4e0d\u662f\u7121\u9700\u76e3\u7763\u7684\u5f71\u7247\u5206\u6790\u670d\u52d9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/costaliya.github.io\/Video-DeepResearch\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/Osilly\/Vision-DeepResearch\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Video-DeepResearch \u5c07\u8996\u89ba\u641c\u5c0b\u5ef6\u4f38\u81f3\u9023\u7e8c\u5f71\u7247\uff0c\u8981\u6c42\u6a21\u578b\u5148\u7406\u89e3\u8de8\u756b\u9762\u7684\u8b49\u64da\uff0c\u518d\u9032\u884c\u7db2\u7d61\u63a2\u7d22\u3002<\/p>\n","protected":false},"author":8,"featured_media":10774,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,163,164,183,140,153,116,31,119,127,191,199],"tags":[],"class_list":["post-10775","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-163","category-164","category-183","category-gemini","category-openai","category-agentic","category-video","category-119","category-127","category-anthropic","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10775","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10775"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10775\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10774"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10775"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10775"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10775"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}