The Bottom Line: AI agents can be trained as data scientists to automatically generate high-quality synthetic training data, which continuously improves through meta-optimization.
Researchers have developed Autodata, a method that deploys AI agents as data scientists to produce high-quality synthetic training and evaluation data. The procedure is self-trainable through meta-optimization and delivers better results than classical data creation methods.
Autodata implements a strategy called “Agentic Self-Instruct”, in which AI agents systematically construct training data. The approach replaces manual or semi-automated data engineering with an agent-driven structure that iteratively learns which data characteristics lead to better model performance.
The method was tested on three different use cases: software development tasks from computer science, legal argumentation, and mathematical problem-solving. In all scenarios, Autodata outperformed classical synthetic data generation techniques. Additionally, meta-optimization of the data scientist agent itself led to further performance gains – the system becomes better at creating better data.
Relevant for CTOs and data architects: The method transforms increased inference compute directly into better training data. Instead of investing additional compute in model training, it can flow into data generation. This opens up a new optimization parameter when scaling training infrastructure and could change the way ML teams construct and iterate on training data.
Source: arxiv.org · Published June 23, 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.1.