Mango: Multi-Agent Web Navigation via Global-View Optimization
Researchers propose Mango, a multi-agent web navigation method that uses website structure to determine optimal starting points, avoiding inefficient root URL exploration. URL selection is formulated as a multi-armed bandit problem with Thompson Sampling for adaptive budget allocation. An episodic memory component stores navigation history for learning from past attempts. On WebVoyager, Mango with GPT-5-mini achieves 63.6% success rate, outperforming the best baseline by 7.3%. The method also attains results on WebWalkerQA.
Key facts
- Mango is a multi-agent web navigation method.
- It uses website structure to dynamically determine optimal starting points.
- URL selection is formulated as a multi-armed bandit problem.
- Thompson Sampling adaptively allocates navigation budget across candidate URLs.
- An episodic memory component stores navigation history.
- On WebVoyager, Mango with GPT-5-mini achieves 63.6% success rate.
- Mango outperforms the best baseline by 7.3% on WebVoyager.
- Mango also attains results on WebWalkerQA.
Entities
—