DSWorld: AI World Model for Efficient Data Science Agents

Autonomous data science agents, despite their advanced capabilities in data understanding and decision-making, often rely on computationally expensive trial-and-error workflows. This bottleneck motivates the development of models that can predict the outcomes of data science operations before they are actually executed. Researchers propose the concept of a Data Science World Model, which simulates the data science execution environment by forecasting state transitions based on current workflow states and potential operations. They introduce DSWorld, a practical framework that integrates structured state construction, cost-aware routing, lightweight real execution, and an LLM-based simulator for costly operations. To facilitate training, an 8K-scale transition trajectory dataset was created, alongside Reflective World Model Optimization, an error-aware reinforcement learning strategy. Experiments show DSWorld significantly accelerates RL-based agent training by approximately 14 times and search-based inference by 3-6 times, all while maintaining competitive performance and outperforming strong LLM baselines on transition prediction tasks.