Reproduce “WebSynthesis: World-Model-Guided MCTS for Efficient WebUI-Trajectory Synthesis” 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
Yifei Gao, Junhong Ye, Jiaqi Wang, Jitao Sang
Researchers at Beijing Jiaotong University developed WebSynthesis, a framework that employs an LLM-based world model and Monte Carlo Tree Search to synthesize high-fidelity web UI interaction trajectories offline. This method enables training web agents to achieve a 14.93% Pass@1 success rate on the WebArena-Lite benchmark, surpassing models trained on larger real-world datasets and reducing data collection costs.
Key insights
- Integrating an LLM-based world model with MCTS allows for the efficient and controllable generation of high-quality, diverse, and goal-directed web UI trajectories without direct real-world interaction.
- A multi-stage curriculum, including foundational UI understanding (TextUI warm-up) and learning from both successful and error-recovery (rollback) trajectories, is crucial for developing robust and adaptable web agents.
- The informational density of trajectories synthesized through world model-guided MCTS provides high sample efficiency, enabling agents trained on smaller synthetic datasets to outperform those trained on larger volumes of real-world data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.