Xiangjun Wang

We propose SAGE, a self-guided hierarchical exploration framework that trains generalist foundation model web agents without human supervision, achieving state-of-the-art performance