Reproduce “Explorer: Scaling Exploration-driven Web Trajectory Synthesis for Multimodal Web Agents” 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
Vardaan Pahuja, Yadong Lu, Corby Rosset, Boyu Gou, Arindam Mitra, Spencer Whitehead, Yu Su, Ahmed Awadallah
An automated framework named Explorer enables the scalable synthesis of 94K diverse web trajectories across 49K unique URLs, generating the largest such dataset to date at $0.28 per trajectory. Small language models fine-tuned on this synthetic data demonstrated improved performance on online web agent benchmarks, outperforming larger models like GPT-3.5.
Key insights
- Large-scale, diverse, and contextually grounded web trajectory datasets are essential for training generalist web agents capable of real-world performance.
- An exploration-driven, "bottom-up" data synthesis approach effectively curates more diverse and realistic tasks than traditional static or tutorial-based methods.
- High-quality synthetic data can empower relatively smaller multimodal models (e.g., 4B/7B parameters) to achieve state-of-the-art results, surpassing larger baselines on web agent tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.