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

Source: arXiv:2607.08662 · Published 2026-07-09 · By Xiaoshuai Song, Liancheng Zhang, Kangzhi Zhao, Yutao Zhu, Zhongyuan Wang, Guanting Dong et al.

TL;DR

This paper addresses the limitations of existing LLM-based web search agents, especially in handling complex queries requiring both deep multi-hop reasoning and broad coverage of information. Prior single-agent approaches are constrained by trajectory length and context windows, and existing multi-agent systems lack recursive depth, flexible collaboration modes, and alignment with actual web information structure. WebSwarm proposes a progressive recursive delegation framework that dynamically creates hierarchical search nodes with paired local objectives and search modes, enabling recursive task decomposition, adaptive collaboration, and evidence-driven expansion during inference. It further guides expansion using lightweight web-probed information structure signals and transfers process-level experience across sibling nodes. Experiments on four benchmarks (BrowseComp-Plus, WideSearch, DeepWideSearch, and GISA) show that WebSwarm consistently outperforms single-agent ReAct baselines and recent multi-agent methods on diverse deep, wide, and mixed search tasks, with gains up to +17.5 accuracy points and significant improvements in structured table F1 metrics. The analysis confirms the importance of recursive delegation, multi-mode search, and guided expansion for effective deep-and-wide web search tasks.

Key findings

• WebSwarm improves BrowseComp-Plus accuracy by 17.5 points over single-agent ReAct (from 50.5% to 68.0%) and 3.5 points over the strongest multi-agent baseline.
• On WideSearch-EN, WebSwarm yields +10.91% Row F1 and +9.76% Item F1 over ReAct, showing better coverage and aggregation of structured web information.
• For DeepWideSearch-EN, WebSwarm increases Item F1 by 9.56 points and Row F1 by 11.77 points compared to ReAct, excelling in interleaved deep-and-wide search scenarios.
• Ablation removing recursive delegation drops BrowseComp-Plus accuracy from 68.0% to 63.5%, showing recursive delegation’s importance.
• Forcing a single global search mode (all-to-wide or all-to-deep) reduces performance, demonstrating the need for heterogeneous local search modes.
• Removing Web-Probing increases web tool calls significantly (e.g., from 137 to 240 on WideSearch), indicating that web-structure guidance reduces redundant exploration.
• Removing experience reuse lowers Item F1 by up to 3 points, confirming sibling-node experience improves solving reliability.
• WebSwarm’s performance gains grow with task difficulty, notably improving Hard sample performance on BrowseComp-Plus from 0% to 35.7% accuracy.

Threat model

n/a — This paper does not focus on adversarial threats or security scenarios. The setting assumes benign information-seeking agents querying public web search tools, with no explicit capabilities or intentions related to attacking or evading defenses.

Methodology — deep read

The paper focuses on complex web search tasks that require both depth (multi-hop, iterative reasoning) and width (coverage of many related entities/attributes). The threat model involves information-seeking agents interacting with web search and browsing tools, with no explicit adversarial model since this is an ML system paper. Data provenance includes four challenging benchmarks: BrowseComp-Plus (200 samples for deep fact search), WideSearch and DeepWideSearch (English subsets, tabular outputs for wide and interleaved deep-wide tasks), and GISA covering diverse search tasks.

WebSwarm’s core is a recursive delegation tree constructed online during inference. The root node starts with the original query; each node is itself an agent with a local objective and a specified search mode (atom, deep, wide, or entity_collect). Nodes either solve their subtask directly or create child nodes with new local objectives and modes, forming a recursive tree. Returned child results serve as evidence for parents to expand, revise, or aggregate. This alternating bottom-up/top-down flow allows progressive, dynamic, evidence-grounded search planning.

Search modes represent distinct local collaboration protocols: 'atom' nodes execute simple fact lookups (similar to ReAct agents), 'deep' nodes implement serial iterative search-verification loops, 'wide' nodes spawn many child nodes in parallel and aggregate results, and 'entity_collect' nodes enumerate unknown sets using multi-path recalls and verifiers. This heterogeneous mode design supports diverse search subtask bottlenecks.

To guide delegation and avoid misaligned expansions, WebSwarm uses a lightweight Web-Probing Agent that probes how relevant evidence is organized on the web and returns hints about information concentration vs dispersion. This structural cue informs the dimension and granularity of further subtask decomposition. Additionally, for sibling nodes created by wide nodes, WebSwarm identifies a small scout subset whose search trajectories are analyzed to extract experience (e.g., successful query patterns, reliable sources). This experience is then injected into remaining sibling nodes to improve consistency and efficiency.

The backbone LLM is GLM-4.5; comparisons with other LLMs are also performed. Each agent step can perform 'search' or 'fetch_url' web tool calls. Training details are not explicit as the system mainly relies on zero-shot or few-shot prompting using ReAct methodology and dynamic orchestration.

Evaluation metrics align with benchmark standards: accuracy (ACC) for deep fact search, item-level F1, row-level F1, and success rate (SR) for tabular collections, plus exact match and list/set F1 for GISA. WebSwarm is compared against strong single-agent ReAct baselines and several multi-agent prior systems (Swarm-Agent, Flash-Searcher, Table-as-Search, ROMA, InfoSeeker). Ablations isolate the impact of recursive delegation, search modes, Web-Probing, and experience reuse.

One concrete example illustrated is a football player profile search on DeepWideSearch, where WebSwarm recursively delegates subtasks for entity identification, transfer history collection, and attribute extraction, coordinating diverse modes and evidence feedback guided by probing and experience sharing.

Code and model checkpoints are partially available as per the GitHub link, supporting some reproducibility, but full details on training or random seeds are not fully described.

Technical innovations

• A novel progressive recursive delegation framework coupling local objectives with distinct search modes to jointly construct dynamic task decomposition and multi-agent collaboration.
• Introduction of four differentiated search modes (atom, deep, wide, entity_collect) to tailor local solving protocols to diverse subtask bottlenecks within web search.
• Web-structure-guided recursive delegation that probes external web information organization to align subtask expansions with actual evidence distribution.
• Within-instance process-level experience transfer across homogeneous sibling nodes, extracting reusable search strategies from scout trajectories to improve reliability and efficiency.

Datasets

• BrowseComp-Plus — 200 (sampled) — public benchmark for deep fact search
• WideSearch (English subset) — unknown exact size — public benchmark for wide structured information collection
• DeepWideSearch (English subset) — unknown exact size — public benchmark for nested deep-and-wide search tasks
• GISA — unknown size — public general web information seeking benchmark

Baselines vs proposed

• ReAct GLM-4.5: BrowseComp-Plus ACC = 50.5% vs WebSwarm = 68.0% (+17.5)
• ReAct GLM-4.5: WideSearch Item F1 = 64.61% vs WebSwarm = 74.37% (+9.76)
• ReAct GLM-4.5: DeepWideSearch Item F1 = 46.63% vs WebSwarm = 58.40% (+11.77)
• Swarm-Agent GLM-4.5: WideSearch Item F1 = 68.79% vs WebSwarm = 74.37% (+5.58)
• Table-as-Search GLM-4.5: WideSearch Item F1 = 69.44% vs WebSwarm = 74.37% (+4.93)
• ROMA GLM-4.5: BrowseComp-Plus ACC = 42.5% vs WebSwarm = 68.0% (+25.5)
• InfoSeeker GLM-4.5: DeepWideSearch Item F1 = 55.10% vs WebSwarm = 58.40% (+3.3)

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2607.08662.

Fig 2: Overview of WebSwarm with a running example from the DeepWideSearch benchmark.

Limitations

• The paper does not discuss adversarial robustness or malicious manipulation of recursive delegation by attackers.
• Evaluation focuses on English benchmarks; generalization to other languages or regions is unclear.
• No reported estimates of computational overhead, latency, or scalability of recursive delegation compared to simpler agents.
• Training details and reproducibility are limited; the backbone LLM and prompts are fixed rather than trained end-to-end.
• Ablation suggests Web-Probing improves efficiency but not always accuracy, indicating some heuristic dependence.
• Real-world deployment challenges, such as web latency, dynamic content, or multi-modal data, are not addressed.

Open questions / follow-ons

• How does WebSwarm perform under adversarial or deceptive web content designed to mislead recursive delegation?
• Can the framework be extended to multi-modal search involving video, images, or domain-specific knowledge sources?
• What are the trade-offs in latency and computational cost versus accuracy gains in real-time applications?
• How effectively can experience be generalized or transferred across different queries or tasks beyond homogeneous siblings?

Why it matters for bot defense

From a bot-defense and CAPTCHA engineering perspective, WebSwarm exemplifies advanced agent orchestration strategies that a highly capable automated information-seeking system might use to bypass simple query limits or interaction patterns by decomposing tasks into recursive, collaborative sub-agents. Understanding WebSwarm's recursive delegation, adaptive collaboration modes, and evidence-grounded expansion provides insights into how sophisticated bots could orchestrate multiple queries and interactions to simulate deep and comprehensive research-style searches. Practitioners could consider detection strategies that monitor for recursive, multi-level agent activity patterns or analyze query expansions inconsistent with typical user behaviors. Additionally, the experience transfer mechanism among sibling nodes signals potential for bots to optimize repeated queries efficiently, challenging static heuristic triggers based on query volume or similarity. Thus, defense systems might leverage behavioral signatures from such recursive or hierarchical multi-agent searches to differentiate bots from humans and design CAPTCHAs that resist such recursive multi-agent orchestration.

Cite

@article{arxiv2607_08662,
  title={ WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search },
  author={ Xiaoshuai Song and Liancheng Zhang and Kangzhi Zhao and Yutao Zhu and Zhongyuan Wang and Guanting Dong and Jinghan Yang and Han Li and Kun Gai and Ji-Rong Wen and Zhicheng Dou },
  journal={arXiv preprint arXiv:2607.08662},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.08662}
}

Read the full paper

• arXiv:2607.08662 (abstract)
• arXiv:2607.08662 (PDF)