Reproduce “BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions” 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
Tao Yu, Zhengbo Zhang, Zhiheng Lyu, Junhao Gong, Hongzhu Yi, Xinming Wang, Yuxuan Zhou, Jiabing Yang and 3 more
BrowserAgent introduces a framework for web agents that interact directly with raw web pages via human-inspired browser actions, eliminating reliance on external parsing tools. The approach achieves scalable training and demonstrates approximately 20% performance improvement over strong baselines in multi-hop question answering tasks.
Key insights
- Direct interaction with raw web page accessibility trees, mimicking human browsing, provides a more native and cost-effective way for LLMs to acquire fine-grained information.
- Parallelizing browser automation environments can overcome the scalability bottleneck, making large-scale training datasets for interactive web agents feasible.
- An explicit memory module that summarizes necessary conclusions, combined with rejection fine-tuning that encourages richer reasoning paths, significantly improves an agent's ability to handle multi-hop and long-horizon tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.