Reproduce “WebEvolver: Enhancing Web Agent Self-Improvement with Coevolving World Model” 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
Tianqing Fang, Hongming Zhang, Zhisong Zhang, Kaixin Ma, Wenhao Yu, Haitao Mi, Dong Yu
The WebEvolver framework from Tencent AI Lab boosts self-improving web agents by co-evolving a world model that generates diverse synthetic training data, leading to a 10% performance gain over prior self-evolving baselines on WebVoyager and sustained learning without relying on proprietary models.
Key insights
- Integrating a co-evolving world model for synthetic data generation effectively overcomes the performance stagnation typically seen in self-improving agents by enhancing exploration.
- LLM-based world models can be actively utilized within an agent's training loop to create diverse learning experiences, extending their utility beyond passive inference-time planning.
- Even partially 'hallucinated' yet plausible observations generated by the world model are beneficial for expanding training data diversity, fostering more robust agent policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.