University of Florida professor Sumit Jha and his team created a system called ATLAS to prevent AI agents from deviating from their assigned roles. The tool operates inside large language models to detect and correct task drift without requiring retraining or slowing down performance. Jha explains that the system acts like an internal compass, continuously reinforcing the agent's mission, responsibilities, and current context. The research addresses the challenge of maintaining control over autonomous AI systems that work over long periods. The paper describing ATLAS was accepted for presentation at the 2026 Conference on Empirical Methods in Natural Language Processing.
Read the full story at news.ufl.edu.
Photo: andreas160578 / Pixabay
