Agentic intelligent data engine for scientific large language models

Artificial intelligence is reshaping scientific discovery across diverse fields. Yet scientific AI remains constrained by data curation, which demands substantial manual effort and deep domain expertise, creating major cost and scalability barriers. Here we present the Agentic Intelligent Data Engine (AIDE), a multi-agent framework that automates data curation for scientific AI. AIDE collects, selects, and organizes data in three stages, providing an end-to-end pipeline that turns a data request into training-ready datasets. We benchmarked AIDE on 14 tasks across biomedical and chemical domains. Scientific AI models built with AIDE-curated datasets matched or outperformed those built with manually curated data, and AIDE’s fully automated preparation delivered significant gains in efficiency and data utilization without compromising quality. These findings provide consistent evidence of AIDE’s applicability across scientific domains and usage paradigms and demonstrate its potential to accelerate scientific AI development and democratize access to scientific AI. Here, authors develop AIDE, a multiagent framework that automates scientific data curation, and demonstrate across biomedical and chemical tasks that its automatically curated datasets can match or outperform manually curated data while improving efficiency.