As artificial intelligence (AI) increasingly works on its own, keeping it on the right task becomes harder. University of Florida computer science professor Sumit Jha and his team have developed a system that notices when a model deviates from its assignment and returns it to the right track - without retraining and without significant computational costs.
Reports have recently emerged that AI agents (AI-based programs that independently carry out chains of actions) are gaining more autonomy and sometimes behave differently from what developers intended. Jha studies how to make such systems more controllable and understandable, and proposes simple, non-disruptive interventions that help a model avoid losing the thread of its task.
In the paper "Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-agent AI," together with co-authors Vishal Pramanik, Maisha Maliha and Susmit Jha, he described the ATLAS system (Algebraic Task-role Locking with Adaptive Spectral Memory). The work has been accepted to the EMNLP 2026 conference, which will be held October 24-29 in Budapest.
ATLAS operates inside a large language model (LLM - a neural network trained to understand and generate text) directly during response generation. If the model begins to deviate, the system gently returns it to the task, reminding it of its role and the current information. "An AI agent can start making mistakes even before its answer looks wrong," Jha explains. "Inside the model, it may already be deviating from the task, confusing its role or relying on outdated data. ATLAS corrects this before the error reaches the next message and other agents rely on it."
Existing methods usually check AI responses after they appear. ATLAS does not require retraining the model, so it does not spend additional computing resources and does not slow down operation.
Tasks in which groups of AI agents coordinate actions over a long period are especially prone to deviation. "Suppose you told an agent that it is a doctor and must treat patients," Jha gives as an example. "By the seventh day, the fact that it is a doctor is just one of millions of signals it has received. We call this the role decoherence problem: the agent received an assignment, but over time it was told so many other things that it lost the thread." Such teams are usually managed externally with prompts, retrieval systems and rules. "ATLAS adds something like an internal compass," the researcher says. "In mathematical form, it constantly reinforces three questions: what is the mission? what is this agent responsible for? and what is true right now? Keeping these answers consistent matters when a task lasts far longer than a single conversation."
According to Jha, the system acts like MRI (magnetic resonance imaging) scans for the neural network's "brain." If the model is supposed to do a medic's work, then at the start its internal state resembles a "doctor's brain." But in the next "scan" it may resemble, for example, a "race car driver's brain." ATLAS compares the states and computes minimal changes to return the system to its original task. "The agent is already running, so the intervention must be very small," Jha explains. "The scan does not require retraining, and the agent does not need to be stopped. If it is still 99% a doctor and 1% a race car driver, we remove that 1%. We do not wait for this to become a problem."
It is important to note: the study describes a laboratory development, not a finished commercial product. The authors do not claim that the system completely eliminates undesirable AI behavior - this concerns a method for correcting deviations during model operation.