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

2026-04-20 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Mango is a multi-agent web navigation method designed to improve efficiency on complex websites by optimizing starting points. Unlike traditional agents that begin from the root URL, Mango utilizes a global view of the website's structure to dynamically select optimal initial navigation points. It frames URL selection as a multi-armed bandit problem, employing Thompson Sampling to adaptively distribute the navigation budget across potential URLs. Additionally, Mango incorporates an episodic memory to store and learn from past navigation attempts. Evaluated on WebVoyager, Mango achieved a 63.6% success rate with GPT-5-mini, surpassing the best baseline by 7.3%. On WebWalkerQA, it reached a 52.5% success rate, outperforming the best baseline by 26.8%. The method demonstrates generalizability across various open-source and closed-source models.

Key takeaway

For research scientists developing web agents, understanding Mango's approach to global-view optimization and adaptive URL selection is crucial. Your current root-URL-centric exploration methods may be inefficient on complex sites. Consider integrating multi-armed bandit strategies and episodic memory into your agent designs to significantly improve navigation success rates and resource utilization, as demonstrated by Mango's 26.8% improvement on WebWalkerQA.

Key insights

Optimizing web agent navigation through global-view URL selection and adaptive budget allocation significantly boosts success rates.

Principles

• Global website view enhances navigation.
• Adaptive budget allocation improves efficiency.
• Episodic memory aids learning from attempts.

Method

Mango formulates URL selection as a multi-armed bandit problem, using Thompson Sampling for budget allocation and episodic memory for learning from navigation history.

In practice

• Use Thompson Sampling for URL selection.
• Implement episodic memory for agent learning.
• Leverage website structure for starting points.

Topics

• Mango
• Multi-Agent Web Navigation
• Global-View Optimization
• Multi-Armed Bandit
• Thompson Sampling

Code references

• VichyTong/Mango

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Mango: Multi-Agent Web Navigation via Global-View Optimization — Global website structure analysis and adaptive URL prioritization significantly improve web navigation efficiency for LLM agents.
• OpenWebRL: Demystifying Online Multi-turn Reinforcement Learning for Visual Web Agents — Online multi-turn reinforcement learning on live websites offers a scalable path for training capable visual web agents.
• StepGuard: Guarding Web Navigation via Single-Step Calibration — StepGuard mitigates web navigation agent fragility through dynamic policy optimization and confidence-guided single-step error calibration.
• Multi-Armed Bandits with Arriving Arms: Sequential Screening, Dynamic Regret, and Sublinear Guarantees — UCB-AA addresses dynamic regret in multi-armed bandits where new arms arrive sequentially, improving efficiency.
• Matching with Deliberation: Test-Time Evolutionary Hierarchical Multi-Agents for Zero-Shot Compositional Image Retrieval — The PDF framework introduces self-evolution and Test-Time Scaling Law to ZS-CIR, achieving SOTA performance via hierarchical multi-agents.
• COAgents: Multi-Agent Framework to Learn and Navigate Routing Problems Search Space — COAgents uses cooperative agents and a dynamic search graph to efficiently navigate complex routing problem spaces.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.