Score: 2

Go-Browse: Training Web Agents with Structured Exploration

Published: June 4, 2025 | arXiv ID: 2506.03533v1

By: Apurva Gandhi, Graham Neubig

Potential Business Impact:

Helps robots learn to use websites better.

Business Areas:
Semantic Search Internet Services

One of the fundamental problems in digital agents is their lack of understanding of their environment. For instance, a web browsing agent may get lost in unfamiliar websites, uncertain what pages must be visited to achieve its goals. To address this, we propose Go-Browse, a method for automatically collecting diverse and realistic web agent data at scale through structured exploration of web environments. Go-Browse achieves efficient exploration by framing data collection as a graph search, enabling reuse of information across exploration episodes. We instantiate our method on the WebArena benchmark, collecting a dataset of 10K successful task-solving trajectories and 40K interaction steps across 100 URLs. Fine-tuning a 7B parameter language model on this dataset achieves a success rate of 21.7% on the WebArena benchmark, beating GPT-4o mini by 2.4% and exceeding current state-of-the-art results for sub-10B parameter models by 2.9%.

Country of Origin
🇺🇸 United States

Repos / Data Links

Page Count
14 pages

Category
Computer Science:
Computation and Language