{"id":6840,"date":"2025-06-02T05:25:00","date_gmt":"2025-06-01T20:25:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=6840"},"modified":"2025-06-01T11:09:08","modified_gmt":"2025-06-01T02:09:08","slug":"webdancer-towards-autonomous-information-seeking-agency","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=6840","title":{"rendered":"WebDancer, EvolveSearch, Can Large Language Models Match the Conclusions of Systematic Reviews?\u00a0"},"content":{"rendered":"\n<p>\u60c5\u5831\u691c\u7d22\u30fb\u53ce\u96c6\u3067\u3082\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u306e\u6d3b\u7528\u304c\u76db\u3093\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>WebDancer: Towards Autonomous Information Seeking Agency&nbsp;<\/strong>[67.1]<br>\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u30b7\u30b9\u30c6\u30e0\u306e\u6700\u8fd1\u306e\u9032\u6b69\u306f\u3001\u81ea\u5f8b\u7684\u306a\u591a\u6bb5\u968e\u7814\u7a76\u306e\u53ef\u80fd\u6027\u3092\u5f37\u8abf\u3057\u3066\u3044\u308b\u3002 \u30c7\u30fc\u30bf\u4e2d\u5fc3\u304a\u3088\u3073\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u6bb5\u968e\u306e\u89b3\u70b9\u304b\u3089\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u3092\u63a2\u7d22\u3059\u308b\u30a8\u30f3\u30c9\u30c4\u30fc\u30a8\u30f3\u30c9\u306e\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u60c5\u5831\u3092\u69cb\u7bc9\u3059\u308b\u305f\u3081\u306e\u51dd\u96c6\u30d1\u30e9\u30c0\u30a4\u30e0\u3092\u63d0\u6848\u3059\u308b\u3002 \u6211\u3005\u306f\u3053\u306e\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3092 ReAct, WebDancer \u306b\u57fa\u3065\u3044\u305f Web \u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u3067\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u5316\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2505.22648v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2505.22648v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Wed, 28 May 2025 17:57:07 GMT)<\/li>\n\n\n\n<li>Tongyi Lab , Alibaba&nbsp;\u306b\u3088\u308b\u60c5\u5831\u63a2\u7d22\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u306e\u63d0\u6848\u3002\u30dd\u30b9\u30c8\u30c8\u30ec\u30fc\u30cb\u30f3\u30b0\u3092\u542b\u30804\u30b9\u30c6\u30fc\u30b8\u69cb\u6210\u3002\u3053\u306e\u624b\u306e\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u3092\uff08\u7c21\u6613\u3067\u306f\u306a\u304f\u672c\u6c17\u3067\uff09\u958b\u767a\u3059\u308b\u3046\u3048\u3067\u53c2\u8003\u306b\u306a\u308b\u3002\n<ul class=\"wp-block-list\">\n<li>Step I: Construct diverse and challenging deep information seeking QA pairs based on the real-world web environment (\u00a72.1); Step II: Sample high-quality trajectories from QA pairs using both LLMs and LRMs to guide the agency learning process (\u00a72.2); Step III: Perform fine-tuning to adapt the format instruction following to agentic tasks and environments (\u00a73.1); Step IV: Apply RL to optimize the agent\u2019s decision-making and generalization capabilities in real-world web environments (\u00a73.2).<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><a href=\"https:\/\/github.com\/Alibaba-NLP\/WebAgent\">GitHub &#8211; Alibaba-NLP\/WebAgent: \ud83c\udf10 WebWalker [ACL2025] &amp; WebDancer [Preprint]<\/a><\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>EvolveSearch: An Iterative Self-Evolving Search Agent&nbsp;<\/strong>[98.2]<br>\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb(LLM)\u306f\u3001\u691c\u7d22\u30a8\u30f3\u30b8\u30f3\u3084Web\u30d6\u30e9\u30a6\u30b6\u306a\u3069\u306e\u30c4\u30fc\u30eb\u3092\u7d71\u5408\u3059\u308b\u3053\u3068\u3067\u3001\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u60c5\u5831\u691c\u7d22\u6a5f\u80fd\u3092\u5909\u9769\u3057\u305f\u3002 \u672c\u7814\u7a76\u3067\u306f,SFT\u3068RL\u3092\u7d44\u307f\u5408\u308f\u305b\u305f\u65b0\u305f\u306a\u53cd\u5fa9\u7684\u81ea\u5df1\u9032\u5316\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3067\u3042\u308bEvolveSearch\u3092\u63d0\u6848\u3059\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2505.22501v1\">\u8ad6\u6587<\/a>&nbsp;&nbsp;<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2505.22501v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>&nbsp; &nbsp;(Wed, 28 May 2025 15:50:48 GMT)<\/li>\n\n\n\n<li>\u4e0a\u8a18\u3068\u540c\u3058\u304fTongyi Lab , Alibaba\u304c\u95a2\u308f\u308b\u6210\u679c<\/li>\n<\/ul>\n\n\n\n<p>\u4e00\u65b9\u3067\u4e0b\u8a18\u306e\u3088\u3046\u306a\u6307\u6458\u3082\u3042\u308b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Can Large Language Models Match the Conclusions of Systematic Reviews?\u00a0<\/strong>[43.3]<br>\u6211\u3005\u306f\u3001\u5927\u8a00\u8a9e\u30e2\u30c7\u30eb(LLM)\u306f\u3001\u540c\u3058\u7814\u7a76\u306b\u30a2\u30af\u30bb\u30b9\u3067\u304d\u308b\u3068\u3001\u81e8\u5e8a\u5c02\u9580\u5bb6\u304c\u66f8\u3044\u305f\u4f53\u7cfb\u7684\u306a\u30ec\u30d3\u30e5\u30fc\u306e\u7d50\u8ad6\u306b\u4e00\u81f4\u3059\u308b\u3060\u308d\u3046\u304b? MedEvidence\u3067\u306f\u3001\u63a8\u8ad6\u3001\u975e\u63a8\u8ad6\u3001\u533b\u7642\u30b9\u30da\u30b7\u30e3\u30ea\u30b9\u30c8\u3001\u3055\u307e\u3056\u307e\u306a\u30b5\u30a4\u30ba(7B-700B\u304b\u3089)\u306e\u30e2\u30c7\u30eb\u3092\u542b\u308024\u306eLCM\u3092\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u3057\u307e\u3059\u3002 MedEvidence\u3067\u306f\u3001\u63a8\u8ad6\u304c\u5fc5\u305a\u3057\u3082\u6027\u80fd\u3092\u5411\u4e0a\u3057\u3066\u304a\u3089\u305a\u3001\u3088\u308a\u5927\u898f\u6a21\u306a\u30e2\u30c7\u30eb\u3067\u306f\u5e38\u306b\u5927\u304d\u306a\u5229\u5f97\u304c\u5f97\u3089\u308c\u305a\u3001\u77e5\u8b58\u306b\u57fa\u3065\u304f\u5fae\u8abf\u6574\u306f\u7cbe\u5ea6\u3092\u4f4e\u4e0b\u3055\u305b\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2505.22787v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2505.22787v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Wed, 28 May 2025 18:58:09 GMT)<\/li>\n\n\n\n<li>\u300cConsequently, given the same studies, frontier LLMs fail to match the conclusions of systematic reviews in at least 37% of evaluated cases.\u300d\u304c\u9ad8\u3044\u304b\u4f4e\u3044\u304b\u306f\u60a9\u307e\u3057\u3044\u3068\u3053\u308d\u3060\u304c\u300cunlike humans, LLMs struggle with uncertain evidence and cannot exhibit skepticism when studies present design flaws\u300d\u306f\u6c17\u306b\u306a\u308b\u3002\u300cWe identify four key factors that influence model performance on our benchmark: (1) token length, (2) dependency on treatment outcomes, (3) inability to assess the quality of evidence, and (4) lack of skepticism toward low-quality findings.\u300d\u3068\u306e\u8a18\u8f09\u304c\u3042\u308b\u304c\u3001\u300c\u5185\u5bb9\u306e\u8a55\u4fa1\u300d\u306f\u96e3\u3057\u3044\u8ab2\u984c\u306a\u306e\u3060\u3068\u601d\u3046\u3002<\/li>\n\n\n\n<li>\u307e\u305f\u3001\u300cAcross all comparisons, medical finetuning fails to improve performance (even for medical-reasoning models) and, in most cases, actually degrades it. Indeed, fine-tuning without proper calibration can harm generalization, some- times resulting in worse performance than the base model [49, 50, 51].\u300d\u3082\u9762\u767d\u3044\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/zy-f\/med-evidence\">GitHub &#8211; zy-f\/med-evidence<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u60c5\u5831\u691c\u7d22\u30fb\u53ce\u96c6\u3067\u3082\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u306e\u6d3b\u7528\u304c\u76db\u3093\u3002 \u4e00\u65b9\u3067\u4e0b\u8a18\u306e\u3088\u3046\u306a\u6307\u6458\u3082\u3042\u308b\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[42,445],"class_list":["post-6840","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-autonomous-agent","tag-webpage-information-extraction"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6840","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6840"}],"version-history":[{"count":4,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6840\/revisions"}],"predecessor-version":[{"id":6860,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/6840\/revisions\/6860"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6840"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6840"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6840"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}