A-MapReduce: Executing Wide Search via Agentic MapReduce

Mingju Chen; Guibin Zhang; Heng Chang; Xiao Wang; Shiji Zhou

cs.MAFeb 1, 2026

A-MapReduce: Executing Wide Search via Agentic MapReduce

Authors: Mingju Chen, Guibin Zhang, Heng Chang, Xiao Wang, Shiji Zhou

Organizations: Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University · National University of Singapore · Tsinghua University

Abstract

Contemporary large language model (LLM)-based multi-agent systems exhibit systematic advantages in deep research tasks, which emphasize iterative, vertically structured information seeking. However, when confronted with wide search tasks characterized by large-scale, breadth-oriented retrieval, existing agentic frameworks, primarily designed around sequential, vertically structured reasoning, remain stuck in expansive search objectives and inefficient long-horizon execution. To bridge this gap, we propose A-MapReduce, a MapReduce paradigm-inspired multi-agent execution framework that recasts wide search as a horizontally structured retrieval problem. Concretely, A-MapReduce implements parallel processing of massive retrieval targets through task-adaptive decomposition and structured result aggregation. Meanwhile, it leverages experiential memory to drive the continual evolution of query-conditioned task allocation and recomposition, enabling progressive improvement in large-scale wide-search regimes. Extensive experiments on five agentic benchmarks demonstrate that A-MapReduce is (i) high-performing, achieving state-of-the-art performance on WideSearch and DeepWideSearch, and delivering 5.11% - 17.50% average Item F1 improvements compared with strong baselines with OpenAI o3 or Gemini 2.5 Pro backbones; (ii) cost-effective and efficient, delivering superior cost-performance trade-offs and reducing running time by 45.8% compared to representative multi-agent baselines. The code is available at https://github.com/mingju-c/AMapReduce.

Figures & tables

Figure 1. Structural isomorphism: A-MapReduce mirrors the MapReduce pipeline correspondingly.

Figure 2. Overall workflow of our A-MapReduce . Given a wide-search query, our framework takes lightweight sequential execution, then retrieves experiential prior as a decision anchor from experiential memory and samples a conditional MapReduce decision. It then decomposes the query into atomic tasks for batched parallel execution and reduces partial results into a single structured output. After completion, environmental feedback updates the experiential memory, enabling continual cross-task refinement of execution decisions.

Model / System	Item F1(%)		Row F1(%)		Succ. Rate(%)
	Avg@4	Max@4	Avg@4	Max@4	Avg@4	Pass@4
Single Agent †
Claude Sonnet 4 (Thinking)	57.89	66.73	31.69	41.90	2.25	5.00
Gemini 2.5 Pro	50.98	63.62	30.00	41.39	1.50	5.00
OpenAI o3-high	52.61	62.27	34.00	44.07	4.50	9.00
Kimi K2	54.36	65.08	29.67	41.35	1.12	3.50

Table 1. Main results on the WideSearch benchmark. We report Item-level F1 (Item F1) , Row-level F1 (Row F1), and Success Rate (Succ. Rate) across all systems. Entries marked with

\dagger

are as reported in the original WideSearch paper ( Wong et al., 2025 ) , included to provide a broader context for comparison, while the other systems are evaluated by us using the same evaluation protocol (App. D ).

Model / System	CE Acc.		Column F1		Item F1		Row F1		Succ. Rate
	Avg@4	Pass@4	Avg@4	Max@4	Avg@4	Max@4	Avg@4	Max@4	Avg@4	Pass@4
Closed-source LLMs †
OpenAI o3-mini	61.59	69.55	27.36	35.68	13.59	16.85	3.35	4.55	0.00	0.00
GPT-5	58.41	72.72	31.71	41.05	21.67	28.21	9.61	13.42	0.30	1.36
Claude Sonnet 4	57.95	64.09	32.63	40.16	19.94	23.38	7.31	8.97	0.90	0.90
Gemini 2.5 Pro	73.98	81.82	45.27	52.86	32.06	37.10	15.42	18.96	0.90	1.82

Table 2. Main results on the DeepWideSearch benchmark. We report core-entity accuracy (CE Acc.), Column-level F1, Item-level F1, Row-level F1, and Success Rate (Succ. Rate) across all systems. Entries marked with

\dagger

are as reported in DeepWideSearch ( Lan et al., 2025 ) and are included for comparison, while the remaining results are evaluated by us using the same evaluation protocol (App. D ).

Panel of Figure 3 — Figure 3. Cost–performance trade-offs of A-MapReduce across benchmarks.

Method	Item F1 (%)	Row F1 (%)	Delay (s)
Smolagents	51.31	23.04	2617.7
Flash-Searcher	54.99	34.42	1204.7
A-MapReduce *	64.64	41.62	1460.8
A-MapReduce	67.81	45.23	953.7

Table 3. Efficiency analysis on WideSearch. We report Item F1 and Row F1 (Avg@4) together with per-task delay. A-MapReduce * denotes the variant without experience-driven evolution.

Panel of Figure 4 — Figure 4. Sensitivity analysis on WideSearch: (a) robustness across backbones; (b) sensitivity to the number of retrieved insights.

Method	Item F1(%)	Row F1(%)	Cost($)
A-MapReduce	70.12	43.65	0.60
-w/o Exemplars	68.76 ( $\downarrow$ 1.36)	40.57 ( $\downarrow$ 3.08)	0.64
-w/o Hint	66.47 ( $\downarrow$ 3.65)	40.58 ( $\downarrow$ 3.07)	0.62
-w/o Memory	64.97 ( $\downarrow$ 5.15)	37.92 ( $\downarrow$ 5.73)	1.05
-w/o $M_{q}~\&~P_{q}$	67.90 ( $\downarrow$ 2.22)	37.85 ( $\downarrow$ 5.80)	1.08
-w/o $B_{q}$	65.95 ( $\downarrow$ 4.17)	37.33 ( $\downarrow$ 6.32)	0.68

Table 4. Ablation study of the proposed experiential prior design and MapReduce decision parameters

\Theta

in A-MapReduce . We report Avg@4 for Item F1, Row F1, and per-task cost.

Figure 5. Case study on a representative WideSearch query. Left: a general MAS executes year-by-year sequential retrieval, where long-horizon context accumulation can break structured-table construction. Right: A-MapReduce instantiates the query as a schema-driven task matrix and runs MapReduce-style execution, shown for variants without and with experiential memory.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Source Benchmark	#Total	#Stage I Candidates	#Final Selected
xBench-DeepSearch	100	54	31
WebWalkerQA	680	217	120
TaskCraft	5228	429	101

Appendix

Table 5. Filtering statistics of the two-stage construction pipeline for constructing the agentic-wide subset.

Method	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Succ.(%)	Delay(s)	Cost($)
SmolAgents	66.84	47.54	51.31	30.67	21.37	23.04	4.00	2617.65	1.243
Flash-Searcher	67.57	52.16	54.99	41.08	33.00	34.42	6.40	1204.69	0.396
A-MapReduce ∗	75.63	62.34	64.64	48.14	40.26	41.62	6.75	1460.80	1.080
A-MapReduce	78.99	65.48	67.81	51.60	43.83	45.23	7.50	953.66	0.763

Appendix

Table 6. Performance comparison on WideSearch under a strictly controlled setup with GPT-5-mini as the shared backbone and an identical execution environment/configuration across all methods. We report Item-level (Precision/Recall/F1), Row-level (Precision/Recall/F1), Success Rate (Succ.), average runtime per task (Delay), and API cost per task (Cost). A-MapReduce * denotes the variant without experience-driven evolution.

$k_{\mathcal{H}}$	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Succ.(%)	Delay(s)	Cost($)
1	80.22	64.24	67.19	48.02	39.34	41.20	4.17	1506.28	0.740
2	79.11	65.28	68.13	45.08	37.51	39.25	5.00	1050.53	0.648
3	82.07	66.62	70.12	51.18	41.41	43.65	6.67	824.49	0.602
4	77.73	64.15	66.98	45.42	37.77	39.48	3.33	1312.29	0.667

Appendix

Table 7. Sensitivity to the number of retrieved insights (

k_{\mathcal{H}}

) in the experiential prior. We report Avg@4 for Item-level (Precision/Recall/F1), Row-level (Precision/Recall/F1), Success Rate (Succ.), average delay (Delay), and API cost per task (Cost).

Method	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Succ.(%)	Cost($)
A-MapReduce	82.07	66.62	70.12	51.18	41.41	43.65	6.67	0.60
-w/o Few-shot	81.31 ( $\downarrow$ 0.76)	65.38 ( $\downarrow$ 1.24)	68.76 ( $\downarrow$ 1.36)	47.31 ( $\downarrow$ 3.87)	38.61 ( $\downarrow$ 2.80)	40.57 ( $\downarrow$ 3.08)	5.83 ( $\downarrow$ 0.84)	0.64
-w/o Insights	77.44 ( $\downarrow$ 4.63)	63.44 ( $\downarrow$ 3.18)	66.47 ( $\downarrow$ 3.65)	47.37 ( $\downarrow$ 3.81)	38.65 ( $\downarrow$ 2.76)	40.58 ( $\downarrow$ 3.07)	5.00 ( $\downarrow$ 1.67)	0.62
-w/o Memory	75.12 ( $\downarrow$ 6.95)	61.92 ( $\downarrow$ 4.70)	64.97 ( $\downarrow$ 5.15)	44.37 ( $\downarrow$ 6.81)	35.89 ( $\downarrow$ 5.52)	37.92 ( $\downarrow$ 5.73)	4.16 ( $\downarrow$ 2.51)	1.05
-w/o $M_{q}~\&~P_{q}$	77.54 ( $\downarrow$ 4.53)	64.52 ( $\downarrow$ 2.10)	67.90 ( $\downarrow$ 2.22)	43.30 ( $\downarrow$ 7.88)	35.96 ( $\downarrow$ 5.45)	37.85 ( $\downarrow$ 5.80)	3.33 ( $\downarrow$ 3.34)	1.08
-w/o $B_{q}$	77.95 ( $\downarrow$ 4.12)	62.40 ( $\downarrow$ 4.22)	65.95 ( $\downarrow$ 4.17)	43.77 ( $\downarrow$ 7.41)	35.17 ( $\downarrow$ 6.24)	37.33 ( $\downarrow$ 6.32)	4.17 ( $\downarrow$ 2.50)	0.68

Appendix

Table 8. Ablation study of the proposed experiential prior design and MapReduce decision parameters

\Theta

in A-MapReduce . We report Avg@4 for Item-level (Precision/Recall/F1), Row-level (Precision/Recall/F1), Success Rate (Succ.), and API cost (Cost).

Model	XBench Subset			WebWalkerQA Subset			TaskCraft Subset
Model	Acc. (%)	Delay (s)	Cost ($)	Acc. (%)	Delay (s)	Cost ($)	Acc. (%)	Delay (s)	Cost ($)
OWL	45.16	1003.27	0.13	52.50	652.08	0.11	74.26	702.48	0.08
SmolAgents	51.61	1335.10	0.57	50.00	836.29	0.21	80.20	607.04	0.12
Flash-Searcher	67.74	679.30	0.25	59.17	326.42	0.07	87.13	275.90	0.06
A-MapReduce	74.19	707.32	0.35	62.50	297.48	0.08	90.10	384.07	0.09

Appendix

Table 9. Performance comparison on Agentic-Wide subsets using GPT-5-mini as the backbone. We report Accuracy (Acc.), Delay (end-to-end runtime), and Cost (average API cost) on the XBench, WebWalkerQA, and TaskCraft subsets.

$\delta$ Setting	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Succ.(%)	Delay(s)
Low	79.53	65.12	69.17	49.68	42.72	44.39	5.00	1045.3
Medium	82.07	66.62	70.12	51.18	41.41	43.65	6.67	953.7
High	78.41	65.49	68.74	48.35	40.79	42.97	5.58	1140.5

Appendix

Table 10. Sensitivity analysis of the utility threshold

\delta

used to split retrieved episodes into positive vs. negative sets for A-MapReduce on WideSearch. We report Avg@4 for Item-level (Precision/Recall/F1), Row-level (Precision/Recall/F1), Success Rate (Succ.), and Delay (end-to-end runtime per instance).

Figure 6. Case study 1: query-conditioned template evolution on a representative WideSearch task (updating $P$ ). In Case study 1, we demonstrate query-conditioned template evolution (updating $P$ ) to generate a $5\times 5$ financial table with IFRS operating profit, margin, and free cash flow (FCF), along with cited sources. As shown in Figure 6 , the system first generates an initial decision $\Theta=(M,P,B)$ in Stage 1, where the experiential prior captures latent error patterns based on prior executions. Using these memory signals, the agent refines the template $P$ in Stage 2, incorporating accounting-aware extraction and verification rules that ensure data accuracy and consistency. In this stage, the template is enhanced to handle missing values and potential data inconsistencies more effectively, improving the overall quality and reliability of the output.

Framework	Evo.	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Delay(s)	Cost($)
A-MapReduce	✓	88.67	88.67	88.67	52.00	52.00	52.00	2065.21	1.41
A-MapReduce	$\times$	57.33	57.33	57.33	0.00	0.00	0.00	3448.17	1.76

Appendix

Table 11. The performance of case study 1. We report row-level and item-level metrics, together with wall-clock runtime and API cost. The performance improvement after template evolution is shown in Table 11 , with Row F1 increasing from 0.000 to 0.520 and Item F1 improving from 0.573 to 0.887. Additionally, the evolved execution reduces time from 57.50 minutes to 34.42 minutes and cost from 1.76 USD to 1.41 USD. These results highlight the effectiveness of memory-guided template refinement in correcting errors, improving coverage of valid cells, and optimizing execution efficiency.

Figure 7. Case study 2: batching evolution for a DeepWideSearch query (updating $B$ ). We illustrate how the batching strategy evolves during execution. The query requests a $5$ -country $\times$ $6$ -year table (2019–2024; 30 rows) from multiple sources, with a year-level tie-break rule. As shown in Fig. 7 , the original decision in Stage 1 employs an open-batching strategy that mixes countries, leading to redundant page fetches and inconsistent year-specific selections. Leveraging the experiential prior, the evolution refines the batching strategy by switching from the $open$ -batching approach to $by\_attr(Country)$ , ensuring that each search agent has one retrieval context per country for all years. Additionally, the template $P$ is refined for unit normalization.

Framework	Evo.	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Delay(s)	Cost($)
A-MapReduce	✓	96.19	96.19	96.19	73.33	73.33	73.33	897.60	0.36
A-MapReduce	$\times$	78.10	78.10	78.10	0.00	0.00	0.00	962.40	0.68

Appendix

Table 12. The performance of case study 2. We report row-level and item-level metrics, together with wall-clock runtime and API cost. The results in Table 12 show significant improvements after evolution: Row F1 increases from 0.00% to 73.33%, and Item F1 rises from 78.10% to 96.19%. Additionally, the evolved execution reduces both running time and API cost, demonstrating the effectiveness of memory-guided batching evolution in enhancing structure, performance, and efficiency.

Figure 8. Case study 3: task-matrix and template evolution for a DeepWideSearch query (updating $M$ and $P$ ). In Case study 3, we demonstrate the evolution of the task-matrix and template for stable league-season alignment (updating $M$ and $P$ ). The query requests a $5$ -league $\times$ $5$ -season table, reporting the champion club and Sofascore Player-of-the-Season signals. In Stage 1, the initial decision $\Theta=(M,P,B)$ uses a task matrix $M$ indexed by ( League , Season ) and a coarse template $P$ , which can result in repeated champion lookups and unstable or missing champion-to-award mappings for some rows. Evolution improves the alignment by adding Champion as an intermediate field in $M$ and refactoring $P$ into a conditioned procedure that first fixes champions and then retrieves award signals based on the known champion context.

Framework	Evo.	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Delay(s)	Cost($)
A-MapReduce	✓	88.00	88.00	88.00	64.00	64.00	64.00	2221.80	2.58
A-MapReduce	$\times$	44.67	44.67	44.67	16.00	16.00	16.00	3037.80	4.78

Appendix

Table 13. The performance of case study 3. We report row-level and item-level metrics, together with wall-clock runtime and API cost. The results in Table 13 show significant improvements after evolution: Row F1 increases from 16.00% to 64.00%, and Item F1 improves from 44.67% to 88.00%. Furthermore, the evolved execution reduces time from 50.63 minutes to 37.03 minutes, and cost from 4.78 USD to 2.58 USD, demonstrating the effectiveness of memory-guided task-matrix and template evolution in improving structure, performance, and efficiency.

Figure 9. Case study 4: batching evolution and delta-patch backfilling for a DeepWideSearch query (updating $B$ ). In Case study 4, we demonstrate the evolution of batching and delta-patch backfilling for sparse fields (updating $B$ ). The query requests the Forbes Global Billionaires top-10 list for each year from 2019 to 2024 (60 rows), including net worth and Age . Evolution refines batching by switching from $per\_atom$ to $by\_attr(Year)$ , ensuring that each agent has a single year, thereby significantly increasing execution efficiency. However, this execution leaves Age cells missing. To address this, a delta-patch operator is introduced: a second MapReduce call fills in the missing Age values, followed by a keyed join.

Framework	Evo.	Item P(%)	Item R(%)	Item F1(%)	Row P(%)	Row R(%)	Row F1(%)	Delay(s)	Cost($)
A-MapReduce	✓	79.20	79.20	79.20	58.33	58.33	58.33	1917.00	1.21
A-MapReduce	$\times$	56.70	47.20	51.52	0.00	0.00	0.00	2341.20	1.56

Appendix

Table 14. The performance of case study 4. We report row-level and item-level metrics, together with wall-clock runtime and API cost.As shown in Table, the evolution results in significant improvements: Row F1 increases from 0.00% to 58.33%, and Item F1 rises from 51.52% to 79.20%. Additionally, execution time and cost decrease by 18.3% and 22.5%, respectively, demonstrating the effectiveness of delta-patch backfilling in improving performance while reducing both time and cost.

Notation	Definition	Notation	Definition
$q=(q^{\text{text}},S)$	Wide-search query: a natural-language request plus an explicit schema.	$\mathcal{D}=\{\epsilon_{i}\}_{i=1}^{N_{\mathcal{D}}}$	Set of past records; $N_{\mathcal{D}}$ is the set size.
$S=\{s_{k}\}_{k=1}^{K}$	Schema attributes (columns); $K$ is the number of requested fields.	$\epsilon_{i}=(q_{i},\Theta_{i},\tau_{i},u_{i})$	Record: query, decision, trace, and utility feedback.
$E=\{e_{i}\}_{i=1}^{N}$	Latent target entities (rows) induced by $q$ ; $N$ is the number of entities.	$\mathcal{H}$	Hint pool of reusable decision guidance (about $\Theta=(M,P,B)$ ).
$Y\in\mathcal{V}^{N\times K}$	Structured table output; $\mathcal{V}$ is the value space of cells.	$h$	A hint rule’s content used to bias planning/decision sampling.
$\Theta$	High-level execution decision sampled by the MAS (controls decomposition/scheduling/execution).	$w_{h}$	Online hint score updated from utility feedback (higher implies more reliable guidance).
$p(\Theta\mid q)$	Query-conditioned decision distribution to be refined by experience-based evolution.	$\mathcal{S}(h)$	Hint provenance: supporting record indices in $\mathcal{D}$ (for retrieval/voting).

Appendix

Table 15. Notation and Definition

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,039 papers

InfoSeeker: A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking
Apr 3, 2026Ka Yiu Lee, Yuxuan Huang, Zhiyuan He +5Multi-Agent LLM Systems Web Search Agents
Huawei Noah’s Ark Lab · University of Liverpool · University College London
WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search
Jul 9, 2026Xiaoshuai Song, Liancheng Zhang, Kangzhi Zhao +8Multi-Agent Orchestration Deep Research Agents
Gaoling School of Artificial Intelligence, Renmin University of China · Kuaishou Technology
Web2BigTable: A Bi-Level Multi-Agent LLM System for Internet-Scale Information Search and Extraction
Apr 29, 2026Yuxuan Huang, Yihang Chen, Zhiyuan He +6Multi-Agent LLM Systems Agentic Search
University of Liverpool · University College London · Huawei Noah’s Ark Lab

Apr 3, 2026J/K move · Enter open · S save

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

Ka Yiu Lee, Yuxuan Huang, Zhiyuan He, Huichi Zhou+4

Huawei Noah’s Ark Lab · University of Liverpool · University College London

Multi-Agent LLM Systems Web Search Agents

Open paper AskPDF arXiv

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 Host, multiple Managers and parallel 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

Abstract

Multi-Agent LLM Systems

Explore similar work

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

WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search

Web2BigTable: A Bi-Level Multi-Agent LLM System for Internet-Scale Information Search and Extraction