Reproduce “Go-Browse: Training Web Agents with Structured Exploration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Apurva Gandhi, Graham Neubig
Go-Browse trains web agents by maintaining a graph of discovered pages, revisiting them to explore further, and collecting navigation and page-level tasks. A feasibility checker filters proposed tasks before trajectories are sampled. On WebArena, its 7B fine-tuned model scores 21.7% overall versus 18.8% for NNetNav and 19.3% for GPT-4o mini; it scores below GPT-4o mini on Online-Mind2Web.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.