The goal is to close a critical GAP in AI reliability science

Active Project Grant Awarded

Award ID
60NANB26D036
Recipient
Awarding agency
Obligated
$248,989

Award facts

Completion date 21% complete
Place of performance
Chapel Hill, NC 27599, USA

Award details come from USAspending.gov.

About this award

The Department of Commerce National Institute of Standards and Technology awarded the University of North Carolina at Chapel Hill, through its Office of Sponsored Research, a Project Grant of $248,989 on September 9, 2026, under CFDA 11.609, Measurement and Engineering Research and Standards, which funds scientific and engineering research and technology transfer for measurement and standards. The award funds a one-year effort to build a principled framework for AI uncertainty quantification, with an ultimate completion date of August 31, 2027 and work performed in Chapel Hill, North Carolina 27599.

The research team will carry out four lines of work. First, it will develop and evaluate uncertainty estimation methods for language models across black-box, partially accessible, and fully open systems, including models that reason through intermediate steps. Second, it will develop mathematically grounded uncertainty methods for image-generating models with formal statistical guarantees. Third, it will build a diagnostic tool that detects adversarial manipulation by measuring how model uncertainty responds to small, controlled perturbations of the input. Fourth, it will apply and validate the methods in concrete real-world settings. Expected outcomes include validated uncertainty estimators for language models covering multiple access levels and a mathematically grounded uncertainty framework for image-generating models. Performance goals also include a working adversarial detection tool.

The project trains three graduate students. The team will visit NIST in the first semester to align research directions and will maintain ongoing technical exchange throughout the year.

Description

Activities
The research team will carry out four lines of work over a one-year period. First, they will develop and evaluate uncertainty estimation methods for language models, covering black-box, partially accessible, and fully open systems, including models that reason through intermediate steps. Second, they will develop mathematically grounded uncertainty methods for image-generating models, with formal statistical guarantees derived from a principled inference framework. Third, they will build a diagnostic tool that detects adversarial manipulation by measuring how model uncertainty responds to small, controlled perturbations of the input. Fourth, the developed methods will be applied and validated in concrete real-world settings. Three graduate students will be trained across these activities. The team will visit the national institute of standards and technology in the first semester to align research directions and will maintain ongoing technical exchange throughout the year.
Expected outcomes
By the end of the award period the team expects to deliver: validated uncertainty estimators for language models covering multiple access levels; a mathematically grounded uncertainty framework for image-generating models with formal coverage properties; a prototype adversarial detection tool evaluated on standard benchmarks across multiple attack types; uncertainty methods validated in real-world application domains; open-source software implementing all developed methods; and at least three manuscripts submitted to peer-reviewed venues. Three graduate students will complete substantive training in uncertainty quantification for AI systems.
Intended beneficiaries
The immediate beneficiaries are researchers and practitioners who build or deploy AI systems and need reliable measures of model uncertainty. More broadly, the public benefits wherever AI inform consequential decisions and where overconfident model outputs currently go unchallenged. Standards bodies benefit from research results directly applicable to technical benchmarks for trustworthy AI an area of active need beyond.
Subrecipient activities
There are no planned subrecipient activities.

Quick answers

Who received award 60NANB26D036?

60NANB26D036 was awarded to University Of North Carolina At Chapel Hill by National Institute of Standards and Technology (Department of Commerce) on September 9, 2026.

How much is award 60NANB26D036 worth?

$248,989 has been obligated on 60NANB26D036.

On GovTribe

See this award's full record

  • Every transaction and subaward
  • The recipient's other awards and teaming partners
  • Similar grant awards
  • Ask GovTribe AI about this award