Skip to content

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

🕒 Published (v1): 2026-07-09 16:28 UTC · Source: Arxiv · link

Why this paper was selected

WebSwarm recursive multi-agent orchestration for deep web search; directly buildable

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

WebSwarm is a recursive multi-agent web search framework that progressively constructs task decomposition and agent collaboration trees during inference, driven by bottom-up evidence feedback rather than upfront planning. It introduces four typed search modes (atom, deep, wide, entity_collect) and grounds delegation decisions in probed web information structure. On four benchmarks it consistently outperforms both single-agent ReAct and prior multi-agent baselines.

Problem

Single ReAct agents are bottlenecked by context length and cannot simultaneously handle deep multi-hop reasoning and wide coverage. Existing multi-agent systems decompose tasks statically at the root level, use a single global collaboration paradigm, and align decomposition to query surface semantics rather than to how relevant evidence is actually organized on the web—leading to shallow recursion, inflexible collaboration, and misaligned search expansion.

Method

WebSwarm constructs a recursive delegation tree \(T = (V, E)\) at inference time. Each node \(v\) receives a local objective \(q_v\) and a search mode \(m_v \in \{\texttt{atom}, \texttt{deep}, \texttt{wide}, \texttt{entity\_collect}\}\):

  • atom: direct ReAct lookup for narrow fact queries.
  • deep: serial searcher→verifier chains for unknown-target identification under multi-constraint reasoning.
  • wide: parallel divide-and-conquer over known entity/attribute sets; children can themselves be any mode, enabling nesting.
  • entity_collect: multi-path parallel recall + dedup + verification for open-set enumeration.

Nodes generate child delegations as objective–mode pairs \(C_v = \{(q_i, m_i)\}\); children return results \(r_i\) upward and the parent decides to expand, revise, or aggregate via \(r_v = \text{Aggregate}(q_v, R_v)\).

Two guidance signals reduce misaligned delegation: 1. Web-Probing Agent: before any wide expansion, a lightweight scout searches and reads pages to return a web-structure hint \(h_v = \text{WebProbing}(q_v)\) that reveals whether evidence is concentrated (hub pages) or dispersed (dimensional distribution), conditioning subsequent delegation \(C_v = \text{Delegate}(q_v, m_v, R_v, h_v)\). 2. Sibling-node experience transfer: for homogeneous child batches, a scout subset \(C_v^s\) is executed first; trajectories yield \(k_v = \text{ExtractExperience}(\{\tau_i\})\) covering useful query patterns, reliable sources, and invalid paths, which is injected into remaining sibling nodes.

Key Contributions

  • Progressive recursive delegation framework coupling local objectives with typed search modes, enabling decomposition and collaboration to co-evolve during inference.
  • Web-structure-guided expansion via a lightweight Web-Probing Agent that aligns node expansion with actual evidence organization on the web.
  • Within-instance experience transfer across homogeneous sibling nodes to reduce repeated trial-and-error.
  • Empirical evaluation on four benchmarks (BrowseComp-Plus, WideSearch, DeepWideSearch, GISA) spanning deep, wide, and interleaved search.

Results

All comparisons below use GLM-4.5 as backbone unless noted.

  • BrowseComp-Plus (deep): WebSwarm 68.00 ACC vs. ReAct 50.50 (+17.50); vs. best multi-agent baseline Swarm-Agent 64.50 (+3.50).
  • WideSearch-EN (wide): Item-F1 74.37 vs. ReAct 64.61 (+9.76); SR 7.00 vs. ReAct 4.00 (+3.00).
  • DeepWideSearch-EN (hybrid): Item-F1 58.40 vs. ReAct 46.63 (+11.77); Row-F1 29.64 vs. ReAct 20.08 (+9.56).
  • GISA (general): Overall 62.30 vs. ReAct 55.54 (+6.76); List metric 66.04 vs. ReAct 50.58 (+15.46).
  • WebSwarm with GLM-4.5 surpasses single-agent ReAct with Kimi-K2 (66.50 ACC) on BrowseComp-Plus at 68.00 ACC.

Limitations

  • Evaluation uses random 200-sample subsets of BrowseComp-Plus and English subsets only, limiting scope.
  • Experience transfer is within-instance and among homogeneous siblings only; cross-task or cross-instance generalization is explicitly excluded.
  • The paper text provided is truncated; ablation results, tool-use efficiency analysis, and full model-generalization findings are not available to assess here.
  • The recursive delegation tree grows dynamically and unbounded; computational cost and latency scaling are not fully characterized in the available text.
  • Backbone is primarily GLM-4.5; generalization to other LLMs is examined but details are truncated.

Relevance to Agentic AI / LLM Agents

WebSwarm directly advances the architecture of multi-agent orchestration systems by replacing fixed, plan-then-execute topologies with evidence-driven recursive delegation—a design principle broadly applicable to any long-horizon agentic task beyond web search. The introduction of typed search modes as first-class citizens of delegation (not just role labels) is a concrete step toward structured, composable agent collaboration protocols. The web-structure probing mechanism exemplifies how agents can ground their own decomposition decisions in external world state rather than relying solely on LLM reasoning over query text. Together these ideas are relevant to the broader open problem of dynamic task graph construction in agentic pipelines.