Go-Browse: Training Web Agents with Structured Exploration
Apurva Gandhi, Graham Neubig · 2025
Read Go-Browse: Training Web Agents with Structured Exploration by Apurva Gandhi, Graham Neubig (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-03533.txt.
Description
Apurva Gandhi and Graham Neubig train web-browsing agents by letting them explore websites systematically first, in a 2025 paper. Agents fail on unfamiliar sites because they do not know the layout: which pages exist and what each button does. Their method, Go-Browse, treats a website as a graph and explores it in a structured way, reusing what earlier visits learned.
Applied to a standard benchmark of web tasks, the exploration gathered ten thousand successful task trajectories and forty thousand interaction steps across a hundred sites. A modest seven-billion-parameter model fine-tuned on this data solved about twenty-two percent of the benchmark tasks. That beat a commercial mini model and set the pace for models of its size. The lesson is that good browsing data, collected by curiosity with a map, can substitute for a bigger brain.
Catalog record
| CREATOR | Apurva Gandhi, Graham Neubig |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-03533