Research map / Multi-agent systems and games / LLM agents
Go-Browse: Training Web Agents with Structured Exploration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Go-Browse, a graph-search-based structured exploration method that collects web agent data on WebArena and fine-tunes a 7B model to 21.7% success.
Problem
Web agents lack understanding of their environments and existing unsupervised data collection methods are redundant or produce localized, infeasible tasks, limiting training data quality for GUI web agents.
Approach
Frame web data collection as graph search over discovered URLs: an outer loop resets exploration to frontier pages, while an inner loop proposes tasks (navigational via NavExplorer, local via PageExplorer), filters them with a feasibility checker, and samples prefixed and unprefixed trajectories.
Key results
- Go-Browse-7B achieves 21.7% overall success on WebArena, versus 19.3% for GPT-4o-mini and 18.8% for NNetNav-7B.
- Go-Browse-7B improves over pretrained Qwen-2.5-7B-Instruct (8.3%) by 13.4 points.
- Unique URLs visited grow from 183 (1 reset) to 260 (15 resets) across 5 domains, showing outer-loop coverage benefits.
Surprising findings
- Paired bootstrap tests show the gains over GPT-4o-mini (p=0.108) and NNetNav-7B (p=0.094) are not statistically significant at 0.05.
- On Online-Mind2Web out-of-domain tasks, NNetNav-7B performs better than Go-Browse-7B.
Limitations
- Data collected only on five WebArena domains (100 URLs), limiting breadth.
- Fine-tuning uses only successful trajectories; 39K unsuccessful trajectories unused.
- Relies on LLM-generated data, risking propagation of model and prompt biases.
Experimental setup
Datasets: WebArena NNetNav-WA
Benchmarks: WebArena Online-Mind2Web
Metrics: success rate paired bootstrap p-value unique URLs visited
Baselines: NNetNav-7B GPT-4o-mini GPT-4o Claude-3.7-Sonnet Qwen-2.5-7B-Instruct
Builds on: Go-Explore ReAct BrowserGym NNetNav WebArena
Closest papers by meaning
- REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.89)
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
- WARC-Bench: Web Archive based Benchmark for GUI Subtask Executions (ICLR 2026 · similarity 0.88)
- WebLINX: Real-World Website Navigation with Multi-Turn Dialogue (ICML 2024 · similarity 0.88)
- WALT: Web Agents that Learn Tools (ICLR 2026 · similarity 0.88)
- WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.