Terrarium: Growing Personal Assistants via Scalable Synthesis of App Worlds
Terrarium is a framework that creates scalable training data for general-purpose personal assistants by synthesizing executable 'App Worlds' from real-world application information. It generates training tasks with evaluator rubrics and iteratively refines them using agent rollouts. This method allows models to learn across diverse app ecosystems, significantly outperforming standard Qwen3 models.