Mango: Multi-Agent Web Navigation via Global-View Optimization

2026-04-20

AI Summary

This work addresses the challenge that traditional web agents, starting from a fixed root URL, often become trapped in navigation dead ends on complex websites and struggle to efficiently locate target information. To overcome this limitation, the authors propose Mango, a novel approach that integrates global website structure awareness with multi-agent navigation for the first time. Mango constructs a holistic site map and formulates the selection of starting URLs as a multi-armed bandit problem, dynamically allocating navigation budgets via Thompson sampling. It further incorporates a reusable episodic memory module to enable continual learning across tasks. Experimental results demonstrate that Mango achieves a success rate of 63.6% on WebVoyager—outperforming the best baseline by 7.3%—and 52.5% on WebWalkerQA, surpassing the strongest baseline by 26.8%. The method exhibits robust generalization across both open-source and closed-source large language models.

Technology Category

• Multiagent Systems: Multiagent Planning
• Machine Learning: Online Learning & Bandits
• Search and Optimization: Learning to Search

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

Existing web agents typically initiate exploration from the root URL, which is inefficient for complex websites with deep hierarchical structures. Without a global view of the website's structure, agents frequently fall into navigation traps, explore irrelevant branches, or fail to reach target information within a limited budget. We propose Mango, a multi-agent web navigation method that leverages the website structure to dynamically determine optimal starting points. We formulate URL selection as a multi-armed bandit problem and employ Thompson Sampling to adaptively allocate the navigation budget across candidate URLs. Furthermore, we introduce an episodic memory component to store navigation history, enabling the agent to learn from previous attempts. Experiments on WebVoyager demonstrate that Mango achieves a success rate of 63.6% when using GPT-5-mini, outperforming the best baseline by 7.3%. Furthermore, on WebWalkerQA, Mango attains a 52.5% success rate, surpassing the best baseline by 26.8%. We also demonstrate the generalizability of Mango using both open-source and closed-source models as backbones. Our data and code are open-source and available at https://github.com/VichyTong/Mango.

Problem

• web navigation
• navigation traps
• website structure
• exploration efficiency
• target information retrieval

Innovation

• multi-agent web navigation
• global-view optimization
• multi-armed bandit
• Thompson Sampling
• episodic memory