Reproduce “DSWorld: A Data Science World Model for Efficient Autonomous 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
Zherui Yang, Fan Liu, Hao Liu
DSWorld introduces a data science world model that simulates environment transitions for autonomous data science agents, enabling them to predict operation outcomes without costly real-world execution. This approach results in approximately 14x faster reinforcement learning training and 3-6x acceleration in inference, while maintaining competitive agent performance.
Key insights
- A hybrid execution-simulation mechanism effectively balances prediction accuracy and computational efficiency for autonomous data science agents.
- Large-scale, high-quality data science transition trajectories, including synthetic data with Chain-of-Thought reasoning, are crucial for training robust world models.
- Error-aware reinforcement learning, such as Reflective World Model Optimization, iteratively refines transition predictions, leading to improved accuracy and reliability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.