[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-b96dc717-4b15-4758-8ba9-4813d7e32a6a":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":20,"created_at":21,"modified_at":22,"is_published":23,"publish_type":24,"image_url":13,"view_count":25},"b96dc717-4b15-4758-8ba9-4813d7e32a6a","WebSwarm 用递归委派让 LLM 搜索 Agent「又深又宽」:4 个深宽检索榜稳压单 Agent 与并联多 Agent 基线","WebSwarm 把 LLM 搜索 Agent 从「单线 ReAct」改成「递归委派树」:中国人民大学宋晓帅等人 7 月 9 日挂出的论文,在 BrowseComp-Plus、DeepWideSearch、WideSearch、GISA 四个深宽交错的检索任务上稳压单 Agent ReAct 与现有并联多 Agent 基线,把「又要深又要宽」这件矛盾的事一次性解决。\n\n关键设计是把任务拆成节点图:每个节点同时持有一个局部目标和一个「搜索模式」,模式决定该自己解还是下放给子节点;子节点把结果连同证据回传,父节点再决定扩写、修订或聚合。启动阶段先用 web probe 摸清信息在网页上的分布再铺节点树,同质化兄弟节点复用过程级经验,避免重复烧 token。\n\n作者做了模型泛化、工具效率、难度消融等多维分析,确认递归委派不是堆节点就能赢,而是「搜索模式分流 + 证据驱动逐级聚合」才有效。RecAgent \u002F DeepResearch 类深搜索产品已经在卷类似能力,WebSwarm 在论文层给出可复现基线和树结构范式——后面再做 deep-search 有现成参考。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.08662","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-07-13T06:15:00Z","2026-07-12T22:18:05.310006Z","2026-07-12T22:18:05.310014Z",true,"agent",3]