Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Go-Browse: Training Web Agents with Structured Exploration

Apurva Gandhi, Graham Neubig

TL;DR

Go-Browse is proposed, a method for automatically collecting diverse and realistic web agent data at scale through structured exploration of web environments, and achieves efficient exploration by framing data collection as a graph search, enabling reuse of information across exploration episodes.

Abstract

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%.

Go-Browse: Training Web Agents with Structured Exploration

TL;DR

Go-Browse is proposed, a method for automatically collecting diverse and realistic web agent data at scale through structured exploration of web environments, and achieves efficient exploration by framing data collection as a graph search, enabling reuse of information across exploration episodes.

Abstract

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%.

Paper Structure

This paper contains 22 sections, 9 figures, 9 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. LLM Web Agents
  4. Exploration Policies for Data Collection
  5. Go-Browse
  6. Data Collection
  7. Fine-tuning Setup and Results
  8. Analysis
  9. Related Work
  10. Conclusion and Limitations
  11. Web Agent Implementation Details
  12. Web Agent Action Space
  13. Prompts for LM Components
  14. Additional Analyses
  15. Design Choices for Task Proposal
  16. ...and 7 more sections

Figures (9)

  • Figure 1: Overview of the Go-Browse algorithm for web agent data collection for a website. Go-Browse's outer-loop (left) maintains an exploration frontier of discovered but not yet fully explored webpages. Go-Browse's inner loop (right) explores each webpage in the frontier by (1) Proposing tasks for that webpage that are grounded in interaction; (2) Checking the feasibility of those tasks; and (3) Sampling trajectories and discovering new webpages by solving feasible tasks.
  • Figure 2: Interaction-first Exploration
  • Figure 3: Dataset statistics on the 5 WebArena domains (20 pages explored/domain).
  • Figure 4: Proportion of successful trajectories from each model in the Go-Browse-WA dataset.
  • Figure 5: Task diversity of the Go-Browse and NNetNav datasets. Zoom to read sub-task labels.
  • ...and 4 more figures