Education

UF researcher develops system to keep AI agents from drifting off task

A University of Florida computer science professor, Sumit Jha, has created a tool called ATLAS that detects when an AI model begins to stray from its assigned role and corrects the behavior from within the model itself, without retraining or slowing it down. The system addresses what Jha calls “role decoherence” — a problem where AI agents handling long, complex tasks gradually lose track of their original purpose as new information accumulates. The research is detailed in a co-authored paper and comes amid growing concern about AI agents acting in unintended ways.

Sources: UF News