DS-STAR: A state-of-the-art versatile data science agent
data-science-agentds-stardabstepiterative-planningllm-judgeheterogeneous-data
Abstraction: Iterative plan-verify data science agent handling heterogeneous file formats
Key points:
- DS-STAR introduces three innovations: (1) a Data File Analyzer that extracts context from varied formats (JSON, unstructured text, markdown); (2) an LLM-based Verifier judge assessing plan sufficiency at each step; (3) a Router agent that corrects existing steps rather than just appending new ones
- Framework loop: Planner creates high-level plan → Coder implements → Verifier judges sufficiency → Router refines; max 10 rounds
- Achieves SOTA on DABStep (45.2% vs 41.0%), KramaBench (44.7% vs 39.8%), DA-Code (38.5% vs 37.0%) vs best prior methods
- Hard tasks average 5.6 refinement rounds; easy tasks average 3.0, with over half solved in a single round
- Removing the Data File Analyzer drops accuracy on hard DABStep tasks to 26.98%, showing rich data context is critical
- Framework generalizes across base models: tested with both Gemini-2.5-Pro and GPT-5 with complementary strengths
Connections: Google · AI Agents · Large Language Models
Source: https://research.google/blog/ds-star-a-state-of-the-art-versatile-data-science-agent/