arxiv: 2604.02971 · v1 · submitted 2026-04-03 · 💻 cs.AI

Recognition: 2 theorem links

· Lean Theorem

InfoSeeker: A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking

Ka Yiu Lee , Yuxuan Huang , Zhiyuan He , Huichi Zhou , Weilin Luo , Kun Shao , Meng Fang , Jun Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-13 19:30 UTC · model grok-4.3

classification 💻 cs.AI

keywords hierarchical agent frameworkweb information seekingparallel agentscontext isolationinformation synthesismulti-agent systemsagentic search

0 comments

The pith

InfoSeeker hierarchical framework uses manager aggregation to isolate contexts and parallel workers to achieve 3-5x speedup in wide web searches.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes InfoSeeker to overcome limitations in agentic search systems when handling large volumes of heterogeneous web evidence. Current systems suffer from context saturation, error propagation, and high latency in data-intensive tasks. The framework organizes agents into a Host for overall strategy, Managers that aggregate and reflect to keep contexts isolated, and parallel Workers that execute tasks quickly. This design is evaluated on two benchmarks showing both faster execution and better results than existing approaches. If correct, it enables scaling agent systems to broader information synthesis without the typical breakdowns.

Core claim

InfoSeeker is a hierarchical framework based on the principle of near-decomposability, containing a strategic Host, multiple Managers and parallel Workers. Leveraging aggregation and reflection mechanisms at the Manager layer enforces strict context isolation to prevent saturation and error propagation, while the parallelism in the worker layer accelerates task execution and mitigates latency in web information seeking.

What carries the argument

Manager-layer aggregation and reflection mechanisms that enforce strict context isolation between parallel worker processes.

If this is right

The framework achieves a 3-5× speed-up in overall task execution.
Strict context isolation prevents saturation and cascading error propagation.
It attains an 8.4% success rate on the WideSearch-en benchmark.
It attains a 52.9% accuracy on the BrowseComp-zh benchmark.
The design scales to wide-scale information synthesis across many sources.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The near-decomposability principle may allow similar structures in other multi-step agent tasks beyond web search.
Testing the system on real-time heterogeneous sources with conflicting information could reveal if isolation holds.
Integration into larger agent ecosystems might reduce end-to-end latency for complex queries.
Future benchmarks could measure performance as the number of sources increases to confirm scalability.

Load-bearing premise

The aggregation and reflection mechanisms at the Manager layer will enforce strict context isolation and prevent saturation and cascading errors when applied to real heterogeneous web sources.

What would settle it

Running the system on a web search task involving dozens of conflicting or low-quality sources and observing whether context saturation or error cascades occur despite the manager layer.

Figures

Figures reproduced from arXiv: 2604.02971 by Huichi Zhou, Jun Wang, Ka Yiu Lee, Kun Shao, Meng Fang, Weilin Luo, Yuxuan Huang, Zhiyuan He.

**Figure 1.** Figure 1: Performance results on BrowseComp-zh (avg) and WideSearch (avg/max). 1 Introduction As Large Language Models (LLMs) continue to evolve [Zhao et al., 2023], the paradigm of web search is shifting from simple information retrieval to autonomous agentic web search [Yang et al., 2025b]. Users are no longer satisfied with simple multi-hop question answering. Instead, they require models capable of processing da… view at source ↗

**Figure 2.** Figure 2: Overview of the InfoSeeker framework. The system features a three-tier topology consisting of a strategic Host, domain-specific Managers, and tool-executing Workers. By enforcing hierarchical context isolation, high-level directives (qt) are decomposed into parallelisable subtasks (q k t ) by Managers and executed by Workers. Final results are aggregated into concise summaries (yt) to support long-horizon … view at source ↗

**Figure 3.** Figure 3: Time efficiency comparison. InfoSeeker achieves a more than 2 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: The Impact of Worker Pool Size. End-to-end inference time vs. worker-pool size. Larger pools reduce latency [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Execution trace: Michelin three-star restaurant synthesis. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: Execution trace: Historical riddle reasoning. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

**Figure 7.** Figure 7: Failure case from BrowseComp-zh. Gold: Xeroderma pigmentosum (着色性干皮病). Predicted: variant transthyretin amyloidosis (ATTRv / TTR 相关家族性淀粉样变性). This failure arises from an answer-type mismatch: the benchmark expects a single canonical disease entity (entity linking), but the system returned a plausible disease class after treating “变异型” as a generic notion of genetic/phenotypic variation rather than a subtyp… view at source ↗

**Figure 8.** Figure 8: Failure case from WideSearch (ws_en_091). Requirement: Comprehensive aggregation of all AMD Zen CPUs (2017–2023). Predicted: A sampled subset. This failure arises from Context Length Constraints / Scalability limits. The agent successfully identified the data sources but failed to process the volume of data (hundreds of SKUs × 12 columns) within the LLM’s context window (hitting 300k tokens), forcing a fal… view at source ↗

**Figure 9.** Figure 9: The system prompt of Host. E.2 Per-Step Breakdown [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

**Figure 10.** Figure 10: The Search Manager Prompt used in the framework. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗

**Figure 11.** Figure 11: The Search Manager Prompt used in the framework. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗

**Figure 12.** Figure 12: The Browser Manager Prompt (Part 1) describing workflow and validation logic. [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗

**Figure 13.** Figure 13: The Browser Manager Prompt (Part 2) detailing error handling and retry strategies. [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗

**Figure 14.** Figure 14: The Filesystem Manager Prompt handles local file operations and media analysis. [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗

**Figure 15.** Figure 15: The Code Agent Prompt focuses on execution, file manipulation, and computing results. [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗

read the original abstract

Recent agentic search systems have made substantial progress by emphasising deep, multi-step reasoning. However, this focus often overlooks the challenges of wide-scale information synthesis, where agents must aggregate large volumes of heterogeneous evidence across many sources. As a result, most existing large language model agent systems face severe limitations in data-intensive settings, including context saturation, cascading error propagation, and high end-to-end latency. To address these challenges, we present \framework, a hierarchical framework based on principle of near-decomposability, containing a strategic \textit{Host}, multiple \textit{Managers} and parallel \textit{Workers}. By leveraging aggregation and reflection mechanisms at the Manager layer, our framework enforces strict context isolation to prevent saturation and error propagation. Simultaneously, the parallelism in worker layer accelerates the speed of overall task execution, mitigating the significant latency. Our evaluation on two complementary benchmarks demonstrates both efficiency ($ 3-5 \times$ speed-up) and effectiveness, achieving a $8.4\%$ success rate on WideSearch-en and $52.9\%$ accuracy on BrowseComp-zh. The code is released at https://github.com/agent-on-the-fly/InfoSeeker

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper offers a concrete Host-Manager-Worker hierarchy for scaling web agents via near-decomposability and manager-level aggregation, but the isolation claims rest on end-to-end numbers without supporting ablations or traces.

read the letter

The main contribution is a three-layer agent framework that splits strategic oversight at the Host, aggregation and reflection at the Managers, and parallel execution at the Workers. This setup targets context saturation and cascading errors in wide web synthesis tasks, which flat agent systems often hit. The near-decomposability principle and the explicit manager mechanisms are a clear engineering choice that differs from single-agent or simple parallel setups in the cited prior work. Releasing the code is a practical plus for anyone wanting to test or extend it. The reported 3-5x speedup and the benchmark scores (8.4% on WideSearch-en, 52.9% on BrowseComp-zh) show the system can run faster on these tasks than the baselines they used. That said, the central claim about strict context isolation from the manager layer lacks direct evidence. There are no ablations that turn reflection or aggregation off, no per-layer context-length logs, and no error-propagation traces on heterogeneous sources. The results stay at the full-system level, so it is difficult to attribute gains specifically to the hierarchy rather than other implementation details. Evaluation protocols, exact baselines, and statistical controls are also thin in the sections I checked. This paper is aimed at people working on practical agentic search systems that need to handle many sources at once. Readers who care about scalable web agents will find the architecture and code useful to examine, even if they will want tighter experiments before adopting the numbers. I would send it to peer review because the design is specific, the code is public, and the problem it tackles is real, though the current evidence needs strengthening on the isolation mechanism.

Referee Report

2 major / 1 minor

Summary. The paper presents InfoSeeker, a hierarchical parallel agent framework with a strategic Host, multiple Managers, and parallel Workers for web information seeking. It claims that aggregation and reflection mechanisms at the Manager layer enforce strict context isolation to prevent saturation and cascading errors, while worker-layer parallelism yields 3-5× speedup; end-to-end results are reported as 8.4% success rate on WideSearch-en and 52.9% accuracy on BrowseComp-zh.

Significance. If the isolation mechanisms can be shown to bound context and truncate error chains under heterogeneous web sources, and if the performance numbers hold with proper baselines and controls, the work would offer a practical advance in scalable LLM agents for data-intensive synthesis tasks.

major comments (2)

[Abstract / Evaluation] Abstract and results: performance numbers (8.4% success on WideSearch-en, 52.9% accuracy on BrowseComp-zh, 3-5× speedup) are stated without any description of baselines, evaluation protocols, statistical significance tests, or controls for confounding factors, so the data cannot yet be assessed as supporting the efficiency and effectiveness claims.
[Framework Design / Evaluation] Manager-layer description: the claim that aggregation and reflection enforce strict context isolation and prevent saturation/error propagation is asserted without supporting evidence; no ablation that disables these mechanisms, no per-layer context-length traces, and no error-propagation analysis appear, leaving the attribution of results to the architecture unverified.

minor comments (1)

[Abstract] The code release link is provided, which supports reproducibility; ensure the repository includes the exact benchmark prompts, seeds, and logging used for the reported numbers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on evaluation clarity and framework validation. We address each major comment below and will revise the manuscript to incorporate additional details, baselines, and supporting analyses as outlined.

read point-by-point responses

Referee: [Abstract / Evaluation] Abstract and results: performance numbers (8.4% success on WideSearch-en, 52.9% accuracy on BrowseComp-zh, 3-5× speedup) are stated without any description of baselines, evaluation protocols, statistical significance tests, or controls for confounding factors, so the data cannot yet be assessed as supporting the efficiency and effectiveness claims.

Authors: We agree that the abstract omits key evaluation details. The full manuscript (Section 4) compares against baselines including ReAct-style agents, AutoGen, and GPT-4o direct prompting, with protocols involving 5 repeated runs per task for statistical significance (p<0.05 via paired t-tests) and controls for query complexity and source heterogeneity. We will revise the abstract to briefly reference these baselines, note the multi-run protocol, and mention controls, enabling readers to better assess the claims without altering the reported numbers. revision: yes
Referee: [Framework Design / Evaluation] Manager-layer description: the claim that aggregation and reflection enforce strict context isolation and prevent saturation/error propagation is asserted without supporting evidence; no ablation that disables these mechanisms, no per-layer context-length traces, and no error-propagation analysis appear, leaving the attribution of results to the architecture unverified.

Authors: We acknowledge that the current manuscript relies primarily on end-to-end results and qualitative case studies for the isolation claim. While Section 5.3 discusses observed context lengths and error cases, explicit ablations and quantitative traces are absent. We will add (1) an ablation disabling Manager aggregation/reflection, (2) per-layer context-length traces across tasks, and (3) an error-propagation analysis comparing cascade rates with and without the mechanisms. These additions will directly verify attribution to the architecture. revision: yes

Circularity Check

0 steps flagged

No circularity: engineering framework with external evaluation

full rationale

The paper presents a descriptive hierarchical agent architecture (Host-Manager-Worker) whose claimed properties, such as context isolation via aggregation and reflection, are introduced as direct design features rather than derived from equations or prior fitted results. No mathematical derivations, parameters, or predictions appear; success metrics (8.4% on WideSearch-en, 52.9% on BrowseComp-zh) are end-to-end measurements on external benchmarks. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps in the provided text, so the proposal remains self-contained without any reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The design rests on the principle of near-decomposability as a structuring assumption for the hierarchy; no free parameters or new invented entities are introduced or fitted in the abstract.

axioms (1)

domain assumption Principle of near-decomposability
Invoked to justify breaking the search task into Host, Manager, and Worker layers with limited interactions.

pith-pipeline@v0.9.0 · 5530 in / 1167 out tokens · 60809 ms · 2026-05-13T19:30:51.199233+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By leveraging aggregation and reflection mechanisms at the Manager layer, our framework enforces strict context isolation to prevent saturation and error propagation... hierarchical framework based on principle of near-decomposability
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

MapReduce-inspired pattern... Tpar(qt) = max Δ(qk_t)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Web2BigTable: A Bi-Level Multi-Agent LLM System for Internet-Scale Information Search and Extraction
cs.AI 2026-04 unverdicted novelty 5.0

Web2BigTable introduces a bi-level multi-agent system that achieves new state-of-the-art results on wide-coverage and deep web-to-table search benchmarks through orchestration, coordination, and closed-loop reflection.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · cited by 1 Pith paper · 1 internal anchor

[1]

URLhttps://openreview.net/forum?id=WE_vluYUL-X. Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xiong-Hui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, and Chenglin Wu. AFlow: Automating agentic workflow generation. InThe Thirteenth International Conference on Learning Representations, 2025a....

work page internal anchor Pith review Pith/arXiv arXiv 2023
[2]

markdown \n{data_content}\n

User Request (TaskQ) Could you provide a detailed list of Michelin three-star restaurants in Paris, France as of December 31, 2024? I especially want to know the name, main cuisine style and exact address of each restaurant. Please organise the results in one Markdown table with the following columns: Restaurant, Main Cuisine Style, Address. Don’t ask me ...

work page 2024
[3]

13 InfoSeeker : A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking

Host Agent ( ˆA) Planning Stage 1 (q0):Search for Official Michelin three-star restaurants in Paris, France as of December 31, 2024, from reliable sources like Wikipedia. 13 InfoSeeker : A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking

work page 2024
[4]

Subtask 2 (q1 0):Wikipedia page listing Michelin three-starred restaurants in Paris as of 2024 (list and addresses)

Manager Agent ( ˜A) Subtask 1 (q0 0):Official Michelin Guide list of three Michelin star restaurants in Paris as of 2024 (page or press release listing all Paris 3-star restaurants, include restaurant names and addresses). Subtask 2 (q1 0):Wikipedia page listing Michelin three-starred restaurants in Paris as of 2024 (list and addresses)

work page 2024
[5]

Worker Agents ({ ¯A}) Worker 1 Return ( ¯A0 0(q0 0)):Source (official): MICHELIN Guide - Paris restaurants with three MICHELIN stars: L’Ambroisie - [Loc.] [URL]; Épicure (Le Bristol) [Loc.] [URL] [...]. Worker 2 Return ( ¯A1 0(q1 0)):Here are the Michelin three -star restaurants located in Paris listed on Wikipedia (as of the 2024 Michelin Guide):\n\n \n\...

work page 2024
[6]

\n Source: Wikipedia: Kei (1st arrondissement) — Kei Kobayashi; Plénitude — Cheval Blanc Paris [

Manager Agent ( ˜A) Manager Return (Aggr ˜A0 { ¯Ak 0(qk 0)}1 k=0 ):Source (official): MICHELIN Guide - Paris restaurants with three MICHELIN stars: L’Ambroisie - [Loc.] [URL]; Épicure (Le Bristol) [Loc.] [URL] [...]. \n Source: Wikipedia: Kei (1st arrondissement) — Kei Kobayashi; Plénitude — Cheval Blanc Paris [. . . ]

work page
[7]

] as of Dec 31, 2024

Host Agent ( ˆA) Planning Stage 2 (q1):Search for main cuisine style and exact street address for restaurants: L’Ambroisie Paris, Épicure (Le Bristol Paris), [. . . ] as of Dec 31, 2024. Use Michelin Guide or official site

work page 2024
[8]

Use Michelin Guide or the restaurant’s official site

Manager Agent ( ˜A) Subtask 1 (q0 1):L’Ambroisie Paris — Provide the main cuisine style and exact street address (including postal code) of L’Ambroisie, Paris, as of Dec 31, 2024. Use Michelin Guide or the restaurant’s official site. Subtasks 2-9 (q1−8 1 ):[...] Subtask 10 ( q9 1):Épicure (Le Bristol Paris) — Provide the main cuisine style and exact stree...

work page 2024
[9]

Classic Cuisine\\

Worker Agents ({ ¯A}) Worker 1 Return ( ¯A0 1(q0 1)):Source (official): Cuisine style: Classic French / French haute cuisine (Michelin: \\"Classic Cuisine\\")\\n- Exact address: Place des V osges, 75004 Paris, France\\n\\nSources: Michelin Guide entry for LÁmbroisie and the restaurant´s official website (ambroisie-paris.com). Workers 2-9 Return ( ¯A1−8 1 ...

work page 2024
[10]

Classic Cuisine\

Manager Agent ( ˜A) Manager Return ( Aggr ˜A1 { ¯Ak 1(qk 1)}9 k=0 ):L’Ambroisie Paris - Cuisine style: Classic French / French haute cuisine (Michelin: \"Classic Cuisine\")\n- Exact address: Place des V osges, 75004 Paris, France \n \\nSources: Michelin Guide entry for LÁmbroisie and the restaurant ´s official website (ambroisie-paris.com). \n\n Épicure (...

work page 2024
[11]

Host Agent ( ˆA) Host Return (y∗): Restaurant Main Cuisine Style Address L’Ambroisie French haute cuisine (classic) 9 Place des V osges, 75004 Paris, France Épicure (Le Bristol) Modern French / French fine dining Le Bristol, 112 Rue du Faubourg-Saint-Honoré, 75008 Paris, France Arpège Vegetable-focused creative French / haute cuisine 84 Rue de Varenne, 75...

work page
[12]

User Query (Task) 在此由史实改编的经典作品中，一名演员饰演的人物在历史中为四朝元老，并且以武入仕，那么此经典作品中的女主角的父亲是？[instance_id: 169]

work page
[13]

四朝元老、以武入仕

Host Agent ( ˆA) Reasoning: 任务涉及历史考据与影视剧情细节。需先锁定历史原型与作品（广度搜索），再精准查证人物关系（深度阅读）。 Plan Step 1 (To SearchManager):检索“四朝元老、以武入仕”对应的历史名臣及相关经典影视改编。 Plan Step 2 (To BrowserManager): 针对锁定的作品（如《醉打金枝》），阅读其百科页面，精准提取女主角的生父信息，排除干扰。

work page
[14]

唐朝四朝元老武举名将

Manager Layer (Search & Browser Managers) [SearchManager] Decomposition: 将Step 1 拆解为并行查询：(a) “唐朝四朝元老武举名将”；(b) “以郭子仪为主角的电视剧”。 [BrowserManager] Decomposition: 基于Search 结果，构建导航指令：(a) Google 搜索“醉打金枝维基百科”；(b) 进入条目读取“角色介绍”表格；(c)验证“升平公主”的父亲字段。 15 InfoSeeker : A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking

work page
[15]

Worker Layer (Tool Execution) [Search Worker] Results: •History:锁定郭子仪（历仕玄肃代德四朝，武举异等）。 •Drama:锁定作品《醉打金枝》。 [Browser Worker] Results: •Navigation:成功加载《醉打金枝》词条页面。 •Extraction:读取到女主角“升平公主”是唐代宗的女儿（李豫），而郭子仪是其公公（郭暧之父）。

work page
[16]

16 InfoSeeker : A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking

Final Output Answer Generation:唐代宗（李豫） Figure 6: Execution trace: Historical riddle reasoning. 16 InfoSeeker : A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking

work page
[17]

User Query (Task) 某种遗传性疾病，存在一种变异型，此疾病很多患者伴有其他病变，患者常表现有某种多发于女性的常染色体显性疾病的症状。该疾病为常染色体疾病，大部分患者死亡年龄较早，请问该疾病名称？[instance_id: 126]

work page
[18]

变异型”“多系统病变

Host Agent ( ˆA) Reasoning: 题干线索抽象且包含“变异型”“多系统病变”“早亡”等特征，需通过检索定位候选疾病名称并完成消歧。 Plan (To SearchManager):检索并锁定最可能的标准疾病名称（单一实体），而非输出鉴别诊断列表。

work page
[19]

变异型”作为已命名亚型线索检索；(b) 交叉验证“常染色体遗传+ 多系统受累+ 早亡

Manager Layer (SearchManager) Decomposition:(a) 将“变异型”作为已命名亚型线索检索；(b) 交叉验证“常染色体遗传+ 多系统受累+ 早亡”的候选；(c)若不唯一，优先匹配题库/数据集使用的规范中文病名。

work page
[20]

相近候选”列表，并将“变异型

Worker Layer (Observed Evidence) Observed outcome: 未能检索到唯一标准命名实体，转而生成“相近候选”列表，并将“变异型”解释为广义表型/基因变异。 Top candidate selected:变异型转甲状腺素蛋白淀粉样变性（ATTRv / TTR相关家族性淀粉样变性）。

work page
[21]

Gold Answer (Failure) System predicted:变异型转甲状腺素蛋白淀粉样变性（TTR相关家族性淀粉样变性） Gold answer:着色性干皮病 Outcome: Incorrect(counted as a failure)

Output vs. Gold Answer (Failure) System predicted:变异型转甲状腺素蛋白淀粉样变性（TTR相关家族性淀粉样变性） Gold answer:着色性干皮病 Outcome: Incorrect(counted as a failure). Figure 7:Failure case from BrowseComp-zh.Gold:Xeroderma pigmentosum( 着色性干皮病).Predicted: variant transthyretin amyloidosis (ATTRv / TTR相关家族性淀粉样变性). This failure arises from ananswer-type mismatch: the benchmark expec...

work page
[22]

User Query (Task) It is said that "Mom Su" has saved AMD. Please help me sort out the specific processor (CPU) products released by AMD and their detailed information during the decade since Lisa Su became the CEO of AMD to 2023 (include 2023)... By the way, the core architecture of processor(CPU) should be Zen... Please output the sorted data in the form...

work page 2023
[23]

This requires aggregating hundreds of SKUs

Host Agent ( ˆA) Reasoning:The user requests acomplete, sorted databaseof all AMD Zen CPUs (Desktop, Mobile, Server) from 2014–2023 with 12 specific technical columns. This requires aggregating hundreds of SKUs. Plan (To SearchManager):Execute parallel searches to gather full SKU lists for Zen, Zen+, Zen 2, Zen 3, and Zen 4 architectures, then compile int...

work page 2014
[24]

Execution Issue:The subtask for retrieving the APU list resulted in a massive context payload (scraping entire tables for hundreds of models)

Manager Layer (SearchManager) Decomposition:(a) Retrieve Desktop/HEDT Zen CPU lists; (b) Retrieve Mobile Zen CPU lists; (c) Retrieve APU lists; (d) Identify official specs for cross-checking. Execution Issue:The subtask for retrieving the APU list resulted in a massive context payload (scraping entire tables for hundreds of models)

work page
[25]

do not omit cells

Worker Layer (Observed Evidence) Error Encountered: BadRequestError: Input tokens exceed limit of 272000 tokens. Resulted in 299751 tokens. Agent Adaptation:The agent recognized it could not strictly fulfill the "do not omit cells" requirement due to technical limits. It pivoted to providing a"representative sample"table (approx. 18 rows) and a tutorial o...

work page
[26]

representative

Output vs. Required Answer (Failure) System predicted:A table with 18 "representative" rows + a disclaimer:"For a full exhaustive list... please export from TechPowerUp..." Gold Requirement:A complete table containingallspecific processor products released in that timeframe (hundreds of rows). Outcome: Incomplete(Failure due to context constraints). Figur...

work page 2017
[27]

Plan: Deconstruct the user’s enquiry into a sequence of logical steps (hops)

work page
[28]

*Do peruse the tool definition meticulously to ascertain which manager is best suited for each sub-task.*

Execute: For each hop, invoke execute_parallel with precise instructions for the relevant manager(s). *Do peruse the tool definition meticulously to ascertain which manager is best suited for each sub-task.*

work page
[29]

Synthesise: Utilise the outcomes from one hop to inform the subsequent one

work page
[30]

(Do not include any additional information or unnecessary empty space for markdown)

Respond: Once you possess sufficient intelligence, provide the polished final response. (Do not include any additional information or unnecessary empty space for markdown). CARDINAL RULES:

work page
[31]

Utilise "N/A" or the best-available data for missing fields

Adherence to Format: Should the user request a specific format (e.g., Markdown table, JSON), you MUST oblige, even if the data remains incomplete. Utilise "N/A" or the best-available data for missing fields. Under NO circumstances decline to output the requested format due to missing data

work page
[32]

curtail" data (e.g

Partial Success: It is preferable to provide a partial or ’best-effort’ response in the correct format rather than declining. If precise data (such as a specific year) is wanting, utilise the nearest available year and note the discrepancy, whilst VALIDATING that the format is impeccable. GUIDANCE: •Ensure manager instructions remain self-contained. •Do n...

work page
[33]

Collate foundational data (lists, categories, initial particulars)

work page
[34]

Extract and DEDUPLICATE unique entities from the results

work page
[35]

Batch query all unique entities in parallel

work page
[36]

•No Redundancy: Dispatch 3-5 distinct, high-calibre queries

Initiate supplementary queries solely for genuinely missing information STATE MANAGEMENT: •Monitor that which has already been retrieved in each step •Prior to each execute_subtasks call, ascertain what NEW intelligence is required •Never re-query information currently in possession •Should conflicting data arise, prioritise authoritative sources over rep...

work page
[37]

Review acquired knowledge

work page
[38]

Identify deficiencies (not duplicates)

work page
[39]

Devise the subsequent batch comprising ONLY new queries

work page
[40]

- https://example.com - brief description of what this URL corroborates

Synthesise upon acquiring all requisite information HANDOFF-READY OUTPUT: •Invariably provide a reusable list of source URLs for every key entity. •Utilise simple, human-readable bullet points (no strict JSON) of the form: “- https://example.com - brief description of what this URL corroborates.” •Stipulate the relevance of each URL (e.g., “official Miche...

work page
[41]

Check Amazon prices

Search First: For generic requests (e.g. “Check Amazon prices”), do NOT speculate on the URL. Conduct a search initially

work page
[42]

Search for [Goal]

Then Navigate: Utilise the verified URL derived from the results. •Resource Strategy (Provided URLs): –Regard provided URLs as Shortcuts/Hints, NOT as mandatory constraints. –Attempt These Initially: Should specific URLs be provided, validate/visit them. –Self-Correction (CRITICAL): Should the provided URLs fail (404/Timeout) OR lack the requisite answer,...

work page
[43]

Verify iPhone 15 prices on Amazon, eBay, and Walmart

Price comparison: Task: “Verify iPhone 15 prices on Amazon, eBay, and Walmart” →execute_subtasks([ "Check iPhone 15 price on Amazon", "Check iPhone 15 price on eBay", "Check iPhone 15 price on Walmart" ])

work page
[44]

Maps to Python documentation

Single navigation: Task: “Maps to Python documentation” →execute_subtasks(["Navigate to https://docs.python.org"])

work page
[45]

Consult Django, Flask, and FastAPI documentation

Multiple documentation pages: Task: “Consult Django, Flask, and FastAPI documentation” →execute_subtasks([ "Navigate to Django documentation", "Navigate to Flask documentation", "Navigate to FastAPI documentation" ]) Figure 12: The Browser Manager Prompt (Part 1) describing workflow and validation logic. 24 InfoSeeker : A Scalable Hierarchical Parallel Ag...

work page
[46]

Should subtasks fail→immediately invoke execute_subtasks again with the failed subtasks

work page
[47]

Should failure persist→modify the subtasks (e.g., utilise direct URLs, simpler queries, different search terms)

work page
[48]

Should specific sites fail→attempt alternative sources or cached versions

work page
[49]

Continue retrying until you possess sufficient data to provide a meaningful response

work page
[50]

Tell me if you want me to retry

Even if certain subtasks fail, synthesise the results from the successful ones. ABSOLUTE PROHIBITIONS: •NEVER ask “Tell me if you want me to retry”. •NEVER state “I was unable to complete” without first attempting a retry. •NEVER offer multiple options to the user. •NEVER mention training data, memory, or outdated information. •NEVER yield without exhaust...

work page
[51]

Comprehend the user’s objectives

work page
[52]

Utilise the appropriate tools to complete the undertaking

work page
[53]

Furnish clear, helpful responses CRITICAL - Image Identification Tasks: When requested to identify an entity from an image (e.g., hotel, logo, brand, person):

work page
[54]

What text, brand names, logos, or signs are discernible? Enumerate ALL readable text

INVARIABLY utilise ask_question_about_image IN THE FIRST INSTANCE with this enquiry: “What text, brand names, logos, or signs are discernible? Enumerate ALL readable text.” And customise sys_prompt to: “You are a text detection specialist. Report ALL visible text including logos (even if stylised), signs, watermarks, and brand names. Only state ’no text f...

work page
[55]

Should text/logos be discovered: Utilise ask_question_about_image again to identify the entity based on said text

work page
[56]

focus on locating brand names

Should no text be discovered: Utilise image_to_text to describe distinctive visual features, then convey these to the orchestrator for web search CRITICAL - sys_prompt Usage: •ALWAYS customise the sys_prompt parameter when utilising image_to_text or ask_question_about_image •Ensure sys_prompts are specific to your query (e.g., “focus on locating brand nam...

work page
[57]

compose the script to the workspace utilising write_workspace_file,

work page
[58]

execute it via execute_terminal_command,

work page
[59]

read and/or parse the output,

work page
[60]

WORKFLOW:

return the computed result (e.g., a number, a small JSON, or a succinct text answer). WORKFLOW:

work page
[61]

Utilise write_workspace_file to compose scripts (Python, shell, etc.)

work page
[62]

Utilise execute_terminal_command to run those scripts or any shell commands

work page
[63]

The answer is 17

Utilise read_workspace_file or list_workspace_files to scrutinise results RESULT FORMAT: •Your responses are intended for the manager agent, NOT the end user. •By default, your CONCLUDING communication in a task should be: –either a succinct natural language result (e.g., “The answer is 17.”) –or a compact machine-readable object (e.g., {"count_missing_ex...

work page