Reproduce “WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiaoshuai Song, Liancheng Zhang, Kangzhi Zhao, Yutao Zhu, Zhongyuan Wang, Guanting Dong, Jinghan Yang, Han Li and 3 more

WebSwarm introduces a multi-agent orchestration framework that dynamically constructs recursive delegation trees to tackle complex web information-seeking tasks requiring deep reasoning and broad coverage. The framework improves performance on benchmarks such as BrowseComp-Plus by 17.50 accuracy points and WideSearch by 10.91 Row F1 over ReAct, by adapting search modes and incorporating web-structure guidance.

Key insights

  • Dynamic, recursive task decomposition combined with mode-guided agent collaboration is crucial for effectively navigating the deep and wide requirements of complex web searches, allowing search structures to evolve as evidence accumulates.
  • Aligning search delegation with the actual organization of information on the web through web-structure-guided expansion significantly improves search efficiency by reducing misaligned expansions and redundant retrieval.
  • Transferring process-level experience among similar subtasks (experience reuse) can consistently enhance the reliability and quality of multi-agent solutions, particularly for homogeneous sibling nodes.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.