Google Introduces Simula: A Reasoning-First Framework for Generating Controllable, Scalable Synthetic Datasets Across Specialized AI Domains

Google and EPFL introduced Simula, a reasoning-first framework for generating synthetic training data from first principles, not seed examples.

Simula uses hierarchical taxonomies, meta-prompt generation, and complexity adjustment to better balance quality, diversity, and difficulty.

The approach is designed for specialized domains like cybersecurity, legal reasoning, and healthcare, where real data can be scarce or sensitive.

It gives researchers more control over dataset coverage, variation, and complexity, helping create scalable synthetic datasets for niche AI tasks.