DS-STAR: A state-of-the-art versatile data science agent

Google Research · 8 months ago

Researchers introduced DS-STAR, an autonomous agent that solves data science problems by converting natural language questions into executable code, addressing limitations of previous agents that struggled with diverse data formats and open-ended problems. DS-STAR achieved 45.2% accuracy on the DABStep benchmark compared to 41.0% for the previous best method, and ranked first on the public leaderboard as of September 18, 2025. The system automates complex data science workflows by iteratively planning, coding, and verifying solutions across multiple heterogeneous data sources.