[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-multi-agent-search-delegation-accuracy":3,"topics-all":38,"news-related-bd350599-7253-4f32-a874-9224bab89a49":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"bd350599-7253-4f32-a874-9224bab89a49","多智能体搜索实测：32组对照里委派多数时候反而更差","一篇 arXiv 预印本用四个模型、四个事实数据集和 32 组对照，实测了 agent 搜索里的「编排者—执行者」委派：多数情况下委派不涨反跌，越过模型知识截止线时掉分更狠。论文拆出三处漏损——编排者塞进错误假设、检索失败信息传不回、工具调用重复——并发现加提示词几乎救不回来。","过去两年 agent 产品最常见的一层架构，就是「编排者 + 执行者」：一个模型负责拆任务、派工、汇总答案，另一批模型带着搜索工具各自去查。这么做的理由很直白——上下文更短、能并行、出错时职责更清楚。\n\n但这套设计到底划不划算，一直缺一份干净的对照实验。最近一篇 arXiv 预印本（[arXiv:2611.16526](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16526)，作者来自波鸿鲁尔大学与可信 AI 研究中心）专门测了这件事，结论相当反常识：把活派出去，多数时候比不派更差。\n\n## 32 组对照，委派大多是净亏损\n\n论文固定四个模型（Claude Sonnet 4.5、GPT-5、gpt-oss-120b、Qwen3-30B-A3B-Thinking-2507，报称知识截止线从 2024 年 6 月到 2025 年 7 月不等）和四个事实类数据集（图书出版年、诺奖得主获奖时年龄、世界银行 GDP 增速、iF 设计奖产品上市年），每个数据集再按模型知识截止线切成「截止前」和「截止后」两半。检索统一走 SerpAPI，四个配置分别是：不给工具、单智能体直接搜、强制委派、可选委派。\n\n结果是：在全部 32 组「模型 × 数据集 × 时间切分」组合里，可选委派有 22 组不如单智能体直接用工具（7 组更好、3 组持平），强制委派有 21 组更差（10 组更好、1 组持平）。\n\n更刺眼的是越过知识截止线之后的落差。单智能体从截止前的 0.77 掉到截止后的 0.75，只掉 3 个百分点；可选委派从 0.74 掉到 0.64，掉 14 个百分点；强制委派从 0.77 掉到 0.66，掉 13 个百分点。也就是说，模型自己知道的事委派不吃亏，模型不知道的事委派先亏一轮——跟直接用工具的差距，从 1 个百分点扩大到 9 个百分点（强制）和 11 个百分点（可选）。\n\n落差最大的一例是 GPT-5 在截止后的 GDP 任务上：单智能体 0.89，可选委派掉到 0.43，强制委派掉到 0.35。\n\n## 三处漏损\n\n第一处，编排者会往派出去的查询里塞自己的假设，而且经常塞错。GPT-5 在强制委派下有高达 98% 的截止前查询会补上「出版形式」这类信息；到了截止后的新书上，委派配置里它补的介质信息只有 4% 到 6% 是对的。论文举了个例子：用户问 2026 年那本 The Night We Met，GPT-5 把它改写成「Lord Huron 那首同名歌曲的发行年」，执行者于是正确地回答了 2015——对被派发的问题是对的，对用户问的问题却是错的。\n\n第二处是信息边界。执行者只把最终答案交回编排者，中间的工具轨迹看不到。碰到网页打不开时，单智能体的 open_url 有 13% 返回 403，委派下升到 20%（可选）和 17%（强制），另有 3%、5%、5% 拿到没有可抽取正文的页面。可即便真的撞上 403，单智能体仍能答对约 73% 的受影响条目，强制委派只有 54%、可选委派 52%——差别不在检索能力，在编排者根本不知道发生过失败。\n\n第三处是工具用得啰嗦。40% 的条目级检索轨迹包含多次 web_search，五千五百七十七对连续查询里有 9% 被判为「同一件事再搜一遍」。委派相对直搜的重复率从 7% 升到 11%，GPT-5 自己从 8% 升到 19%。更贵的还在后面：执行者明明答对了，编排者还会质疑对方的答案或干脆重派同一任务——GPT-5 在截止前的 GDP 上有高达 54% 的任务被重新派发，哪怕执行者答案正确，重派比例也从 6.4% 升到 8.8%。\n\n## 提示词救不回来\n\n论文试了两种最省事的干预。让执行者主动上报打不开的网页和它的 URL，GPT-5 的 GDP 成绩从 0.43 升到 0.51、从 0.35 升到 0.44，但仍远低于单智能体的 0.89；GPT-OSS 反而从 0.70 掉到 0.62、从 0.73 掉到 0.64。另一种是禁止重复发等价查询，Qwen 的重复率从 19.4% 降到 17.7%，GPT-5 反而从 4.4% 升到 6.6%——搜索次数没少，重复度也没真降。\n\n## 不是孤例\n\n机制层面，另一支团队在另一片场景里得到了同方向的结论。Amazon 与华盛顿大学的对照评估（[arXiv:2611.12953](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.12953)）固定模型、任务和工具，只改架构，比较函数调用单智能体、技能增强智能体、中心化多智能体编排和分诊蜂群四种设计共 8 个配置，在 13 个企业 SOP 上跑。所有需要协调的配置，比单智能体多花 1.4 到 4.7 倍 token、多等 2.6 到 5.8 倍延迟，收益通常很小且很少显著，论文因此把单智能体称为「更划算的默认选择」。他们的第三条结论更有意思：实现质量压过了架构选择——一个搭得合理的多智能体输给单智能体，但针对同一个多智能体做定向重构之后，它是唯一显著超过基线的配置，75.6% 对 67.4%。\n\n## 所以呢\n\n「多智能体」现在更像架构默认值，而不是经过论证的选择。这篇论文的价值在于把默认值拆开测了一遍：同样的工具、预算和题目，多加一层委派不自动换来更高精度，反而会在模型无知的地方集中掉分。\n\n如果系统里已经有编排者，把「委派是否划算」当成要实测的假设，而不是前提：先量单智能体在同一批任务上的基线，再看委派后的分任务准确率和 token、延迟倍数。特别盯住两点——执行者检索失败有没有传回来，执行者答对之后编排者又重派了多少次。论文自己也说了，可能有效的方向不是加提示词，而是针对性训练、搜索进度追踪，以及选择性传递证据与失败的机制。\n\n多智能体不是不能做。但下次评审架构时，该被问的问题是「这个编排者到底赚回了什么」，而不是「我们有没有编排者」。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16526","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c10eb143-7f83-440c-835d-65aedf2eea04","en","Multi-Agent Search Bench: Delegation Loses in 32 Comparisons","An arXiv preprint: 32 controlled comparisons find delegation usually loses accuracy in agentic search, and loses more past the knowledge cutoff.","Over the past two years the most common extra layer in agent products has been an orchestrator plus executors: one model breaks the task apart, hands out the pieces and assembles the final answer, while a set of sibling models run their own web searches. The rationale is straightforward. Context stays shorter, work runs in parallel, and when something goes wrong the blame has somewhere to land.\n\nWhat has been missing is a clean controlled test of whether that arrangement actually pays for itself. A recent arXiv preprint ([arXiv:2611.16526](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16526), from Ruhr University Bochum and its Center for Trustworthy AI) does exactly that, and the conclusion runs against the grain: delegating the work usually loses accuracy rather than gaining it.\n\n## Thirty-two controlled comparisons, and delegation usually loses money\n\nThe paper fixes four models (Claude Sonnet 4.5, GPT-5, gpt-oss-120b and Qwen3-30B-A3B-Thinking-2507, with reported knowledge cutoffs ranging from June 2024 to July 2025) and four factual datasets (book publication years, the ages of Nobel laureates, World Bank GDP growth rates, and iF Design Award product launch years). Each dataset is then split into a pre-cutoff and a post-cutoff half, aligned to the models' own knowledge boundaries. All retrieval goes through the same SerpAPI-backed Google search interface, and four configurations are compared: no tools at all, a single agent searching directly, forced delegation, and optional delegation.\n\nThe result: across all thirty-two model-by-dataset-by-cutoff combinations, optional delegation performs worse than a single agent with direct tool access in twenty-two of them (better in seven, tied in three), and forced delegation performs worse in twenty-one (better in ten, tied in one).\n\nThe gap widens once the questions sit beyond what the models already know. The single agent goes from 0.77 pre-cutoff to 0.75 post-cutoff, a three-point slide. Optional delegation falls from 0.74 to 0.64, down fourteen points, and forced delegation falls from 0.77 to 0.66, down thirteen. In other words, when the model already holds the answer in its parameters, delegation costs little; when the model is genuinely ignorant, delegation starts charging interest. Measured against direct tool use, the gap grows from one point to nine points for forced delegation and from three points to eleven for optional.\n\nThe widest single case is GPT-5 on the post-cutoff GDP task: the single agent reaches 0.89, optional delegation drops to 0.43, and forced delegation to 0.35.\n\n## Three places the accuracy leaks out\n\nFirst, orchestrators inject their own assumptions into the queries they hand off, and those assumptions are frequently wrong. Under forced delegation, GPT-5 adds medium information to as many as 98% of its pre-cutoff queries. On post-cutoff books, only four to six percent of the medium information it adds under delegation turns out to be correct. The paper offers a concrete example: asked for the release year of a 2026 book called The Night We Met, GPT-5 rewrote the request as the release year of the song of the same name by Lord Huron. The executor duly returned 2015, which was correct for the question it was handed and wrong for the question the user asked.\n\nSecond, there is an information boundary. The executor returns only its final answer; the orchestrator never sees the tool traces in between. When a page refuses to open, 13% of open_url calls fail with a 403 under the single agent, and that climbs to 20% under optional delegation and 17% under forced; separately, 3%, 5% and 5% of calls return pages with no extractable content. Even when a 403 does occur, the single agent still answers roughly 73% of the affected items correctly, against 54% for forced delegation and 52% for optional. The gap is not about retrieval skill. It is that the orchestrator never learns the retrieval failed.\n\nThird, tool use turns wasteful. Forty percent of item-level search trajectories contain multiple web_search calls, and across five thousand five hundred seventy-seven consecutive query pairs, nine percent were judged to be the same search asked twice. Redundant searching rises from 7% under direct tool use to 11% under delegation, and for GPT-5 alone it goes from 8% to 19%. The expensive part comes later: even when the executor answered correctly, the orchestrator would still challenge the answer or simply hand the task back out again. On pre-cutoff GDP, GPT-5 re-delegates up to 54% of tasks, and even for correct executor answers the re-delegation rate climbs from 6.4% to 8.8%.\n\n## Two prompt tweaks do not rescue it\n\nThe paper tried the two cheapest available interventions. Asking executors to explicitly report inaccessible pages and their URLs moved GPT-5 on the GDP task from 0.43 to 0.51 and from 0.35 to 0.44, still far below the single agent's 0.89; and for GPT-OSS the same instruction made things worse, dropping optional delegation from 0.70 to 0.62 and forced delegation from 0.73 to 0.64. The second intervention, forbidding semantically equivalent queries, barely moved anything: Qwen's redundancy fell only from 19.4% to 17.7%, while GPT-5's actually rose from 4.4% to 6.6%. The models neither searched less often nor became less repetitive.\n\n## Not an isolated result\n\nAt the mechanism level, a different team reached a similar conclusion in a different setting. A controlled evaluation from Amazon and the University of Washington ([arXiv:2611.12953](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.12953)) holds the model, the tasks and the tools fixed and varies only the architecture, comparing four designs across eight arms - a function-calling single agent, a skill-augmented agent, centralized multi-agent orchestration and a triage swarm - over thirteen enterprise SOPs. Every coordinated arm spends 1.4 to 4.7 times the tokens and 2.6 to 5.8 times the latency of the single agent for gains that are usually small and rarely significant, which is why the paper calls the single agent the efficient default. Their third finding is the interesting one: implementation quality outweighs the single- versus multi-agent choice. A plausible multi-agent implementation loses to the single agent, but after a targeted redesign it becomes the only arm that significantly beats the baseline, 75.6% against 67.4%.\n\n## So what\n\nMulti-agent is currently an architectural default rather than an argued-for decision. The value of this preprint is that it took the default apart and measured it: same tools, same budget, same hard questions, and adding one layer of delegation does not buy higher accuracy. It loses points precisely where the model knows nothing.\n\nIf your system already has an orchestrator, the practical recommendation is to treat the question of whether delegation pays as a hypothesis you test, not a premise you inherit. Measure the single-agent baseline on the same task set first, then look at per-task accuracy after delegation alongside the token and latency multiples. Track two specific signals: whether retrieval failures inside the executor ever reach the orchestrator, and how often the orchestrator re-delegates after the executor was already right. The paper is candid that the promising directions are not more prompt engineering but targeted training, search-progress tracking, and mechanisms for selectively passing evidence and failures across the boundary.\n\nMulti-agent is not off the table. But the next time someone reviews an agent architecture, the question worth asking is what the orchestrator actually bought you, not whether you have one.","multi-agent-search-delegation-accuracy","2026-11-15T09:10:00Z","2026-11-15T09:07:15.500010Z","2026-11-15T09:07:15.500039Z",true,"agent",42,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"6c8ae96a-53e0-4ef7-8609-40b345cdee7c","决策模型实测:自己演砸42.9%,教学生82.9%","decision-model-actor-judge-teacher","2026-11-15T21:45:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"a2038f05-8ad2-4b57-bb29-47ff522e09a0","失败轨迹偷走权重:ALFWorld成功率涨7到9.4点","outcome-mass-alignment-multi-turn-agent-rl","2026-11-14T11:20:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"9531b4f7-3f4d-474c-b233-c715925a3e76","8.4B轨迹压过25.3B技能语料:形式化证明中训配方","davinci-prover-agent-native-mid-training","2026-11-14T07:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"343f6cae-ae14-4617-a317-7316748e1d6e","删掉推理链再转发:同档32B模型差51分","deletion-sensitivity-reasoning-trace-routing","2026-11-13T07:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"365a0af2-d564-4eff-bf03-a138cdea0c5f","研究智能体大考:只查交付过69%,查全要求剩28%","are-bench-research-agent-context-requirements","2026-11-13T01:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"f2723003-43ac-47fe-bb59-49030446919a","28个决策模型大考:多选最高才36.81%","mmdm-decision-model-benchmark","2026-11-12T13:30:00+00:00"]