TLDR: Google Research has introduced DS-STAR, a novel state-of-the-art AI agent designed to automate a broad spectrum of data science tasks, from statistical analysis to data wrangling, across diverse and heterogeneous data formats. It achieves top performance on key benchmarks by employing innovative iterative planning and verification mechanisms.
Google Research has announced the development of DS-STAR, a sophisticated artificial intelligence agent poised to revolutionize data science workflows by automating complex tasks that traditionally demand extensive human expertise. Unveiled on November 6, 2025, DS-STAR is engineered to handle a wide array of data science challenges, including statistical analysis, data visualization, and data wrangling, across various data types, including previously challenging unstructured formats.
The agent’s introduction addresses significant limitations in current data science automation, particularly the heavy reliance of existing tools on well-structured data like CSV files. Real-world applications often involve diverse and heterogeneous data formats such as JSON, unstructured text, and markdown files, which conventional methods struggle to process effectively. Furthermore, many data science problems are open-ended, lacking clear ground-truth labels, making it difficult to verify the correctness of an agent’s reasoning.
DS-STAR, developed by researchers including Jinsung Yoon and Jaehyun Nam, introduces three core innovations to overcome these hurdles. First, it features a robust data file analysis module that automatically extracts context and key information from varied data formats, including unstructured ones. Second, it incorporates an LLM-based judge in a verification stage that assesses the sufficiency of the analysis plan at each step. Finally, DS-STAR utilizes a sequential planning process that iteratively refines its initial plan based on feedback from the verification stage, allowing it to navigate complex analyses and draw verifiable insights from multiple data sources.
The agent’s versatility and effectiveness have been rigorously tested and confirmed through its performance on challenging benchmarks. DS-STAR has achieved state-of-the-art results on DABStep, KramaBench, and DA-Code, significantly outperforming existing alternatives. For instance, it improved accuracy on the DABStep benchmark from 41.0% to 45.2%, on KramaBench from 39.8% to 44.7%, and on DA-Code from 37.0% to 38.5% compared to the best alternatives. Notably, DS-STAR secured the top rank on the public leaderboard for the DABStep benchmark as of September 18, 2025.
Also Read:
- Google Cloud Enhances Vertex AI Agent Builder with New Observability and Deployment Tools
- Google DeepMind’s AI Revolutionizes Extreme Weather Prediction with WeatherNext Models
Its superior ability to work with multiple, heterogeneous data sources, particularly in ‘hard tasks’ requiring the processing of several files, underscores its advanced capabilities. By automating these complex data science tasks, DS-STAR holds immense potential to make data science more accessible to a broader range of individuals and organizations, fostering innovation across numerous fields.
