DS-STAR: A state-of-the-art versatile data science agent
Google Research
Google Research built DS-STAR, an AI agent that handles messy data science tasks across CSV, JSON, and text files. It beats prior top agents on three major benchmarks, showing self-checking plans work better than one-shot answers.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Most AI data science agents are built for a tidy world of clean CSV files sitting in relational databases. Real analysts don't get that luxury. They deal with PDFs, markdown docs, unstructured text, and half-documented JSON blobs, often stitched together from multiple sources with no ground truth to check their work against. Google Research built DS-STAR to close that gap, and the results suggest the fix wasn't a bigger model but a smarter process.
The system runs in two stages. First, it scans every file in a project directory and writes a plain-language summary of what's inside, giving the agent context it wouldn't get from a schema alone. Then it loops through planning, coding, and verifying: a Planner agent drafts steps, a Coder agent turns those into runnable scripts, and a Verifier agent, itself an LLM acting as judge, checks whether the plan actually solves the problem. If not, a Router agent decides whether to patch an existing step or add a new one, and the cycle repeats up to ten times. It's less like an assembly line and more like an analyst working in Colab, checking intermediate outputs before deciding what to do next.
That iterative correction turns out to matter a lot. On DABStep, DS-STAR pushed accuracy from 41.0% to 45.2% over the previous best method; on KramaBench it jumped from 39.8% to 44.7%; on DA-Code, from 37.0% to 38.5%. It also topped the public DABStep leaderboard as of September 18, 2025. Strip out the Router agent, and the system degrades into just piling on new steps instead of fixing broken ones, which tanks performance. Remove the file-analysis module entirely and accuracy on hard DABStep tasks collapses to 26.98%, a reminder that context about messy files isn't a nice-to-have, it's load-bearing.
The team also swapped in GPT-5 as the base model and found DS-STAR held up, though with a quirk: the GPT-5 version did better on easy, single-file tasks while the Gemini 2.5 Pro version won on harder, multi-file ones. Difficulty also shows up in iteration counts. Easy tasks in DABStep averaged 3.0 rounds, with over half solved in a single pass. Hard tasks needed 5.6 rounds on average, exactly the kind of grinding, revise-and-recheck behavior the framework was designed to support.
My take — AI-written commentary, not fact-checked reporting
This is a solid example of research chasing a real bottleneck instead of chasing benchmark vanity metrics; heterogeneous, undocumented data is the actual daily grind of data science, not clean CSVs. The Router ablation is the most interesting bit buried in the paper: teaching an agent to fix its own mistakes beats letting it just pile on more steps, which is a lesson plenty of agent builders outside data science still haven't absorbed. I'd watch whether this pattern of self-verifying, model-agnostic loops becomes the template for agentic AI generally, because right now most of the industry is still betting on bigger models rather than better process.
Read more about this at: Google Research