WebSwarm Recursive Multi-Agent Framework Beats Single-Agent and Multi-Agent Baselines on Deep-and-Wide Web Search

Researchers introduce WebSwarm, a progressive recursive delegation framework for LLM-based web search that jointly constructs task decomposition, recursive expansion, and agent collaboration during inference. Unlike single ReAct-style agents constrained by one long trajectory and limited context, or existing multi-agent systems that parallelize but still struggle with recursive depth and adaptive collaboration, WebSwarm dynamically instantiates agentic search nodes. Each node couples a local objective with a search mode specifying how it should organize search and collaboration. Nodes can solve their objective directly or further delegate child nodes; after solving, they return evidence and results upward, enabling parent nodes to further expand, revise, or aggregate the search process. To guide this, WebSwarm first probes how task-relevant information is organized on the web to ground subsequent node expansion, and reuses process-level experience across homogeneous sibling nodes. Experiments on BrowseComp-Plus, WideSearch, DeepWideSearch, and GISA show WebSwarm consistently outperforms single-agent and multi-agent baselines on deep, wide, and interleaved deep-and-wide tasks. Ablation analyses, task difficulty breakdowns, web tool efficiency measurements, and model generalization tests explain WebSwarm's effectiveness and offer insights for multi-agent search systems.