UF Team Develops ATLAS to Keep AI on Track

The lightweight system corrects large language models that drift from their assigned roles while they generate output.

Updated on Nov. 2, 2026 in Artificial Intelligence

Bold flat-color editorial illustration of a single cream dial with a red needle on a navy base, with thin lines trailing into a clean field. — University of Florida computer science professor Sumit Jha's team developed ATLAS, a lightweight system that corrects large language models drifting from their assigned roles during output generation. AI Illustration. Upload story photo >

University of Florida computer science professor Sumit Jha and his team created ATLAS, a lightweight system that keeps large language models focused on their assigned roles. ATLAS works inside the neural network during output generation rather than checking answers afterward.

Why it matters

Most current safety systems review AI answers only after they are generated, while ATLAS intervenes as the model works. Keeping autonomous AI systems on their assigned tasks remains a challenge as they take on more complex work in Gainesville and beyond.

The research describes a method rather than a measured outcome, so no quantitative results are reported in the source material. Its performance across different models and workloads has not been detailed.

The players

Sumit Jha

Jha is a University of Florida computer science professor who led the team that created ATLAS.

The details

ATLAS monitors the neural network in real time, similar to ongoing medical scans, and calculates the exact corrections needed when an agent begins to drift. The tiny adjustments redirect the agent before small distractions grow, using math to reinforce the agent's main mission, specific duties, and current facts.

November 2, 2026: Article published about the ATLAS research.
2026: Research to be presented at the Conference on Empirical Methods in Natural Language Processing in Budapest.

The Tech Race

ATLAS shifts the approach from post-generation checks to in-process monitoring, a contrast with the output-review safety systems most current deployments rely on. It positions University of Florida research within the ongoing effort to make autonomous AI agents reliable as they handle longer, multi-agent workloads.

ATLAS aims to keep AI agents on task without adding significant computing cost, which could make them more dependable in everyday tools. Its practical effect on users will depend on whether the approach is adopted in deployed systems.

The takeaway

Keeping AI agents focused during operation, rather than only checking their answers afterward, offers a path to more reliable autonomous systems. Readers following AI development can watch how such in-process methods are evaluated as they are presented at the 2026 conference.