Overview
WebSwarm is presented as a progressive recursive delegation framework designed to enhance web search capabilities, particularly for tasks requiring both depth and breadth of information. It addresses limitations observed in existing large language model (LLM)-based web search agents and multi-agent systems, particularly concerning handling deep recursive search and adaptive collaboration.
Research Context
Large language model (LLM)-based web search agents have evolved information seeking beyond simple factoid question answering to encompass complex, research-oriented tasks. However, single ReAct-style agents are constrained by a single long trajectory and limited context, which hinders their ability to manage depth and coverage simultaneously in search tasks. While existing multi-agent systems improve search coverage through parallel execution and aggregation, they face limitations in recursive depth, adaptability in collaboration, and evidence-grounded expansion.
Approach
WebSwarm's approach involves jointly constructing task decomposition, recursive expansion, and agent collaboration during inference. The system dynamically instantiates agentic search nodes, each associated with a local objective and a specified search mode. This search mode dictates how the node organizes its search and collaboration activities. Each node possesses the capability to either fulfill its objective independently or delegate the task to child nodes. Upon completion, evidence and results are returned upwards, allowing parent nodes to expand, revise, or aggregate the search process further.
To guide the expansion of nodes, WebSwarm initially probes how task-relevant information is structured on the web. The framework also reuses process-level experience among homogeneous sibling nodes.
Findings
Experiments were conducted using WebSwarm across several benchmarks: BrowseComp-Plus, WideSearch, DeepWideSearch, and GISA. The findings indicate that WebSwarm consistently outperforms both single-agent and multi-agent baselines. This superior performance was observed on tasks classified as deep, wide, and interleaved deep-and-wide.
Further analyses were performed, including ablation studies, assessment of task difficulty, evaluation of web tool efficiency, and examination of model generalization. These analyses provided insights into WebSwarm's effectiveness and offered considerations for multi-agent search systems.