Project Grant 2553892

The National Science Foundation Division of Computing and Communication Foundations awarded the University of Maryland, College Park $400,000 on August 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop methods for efficient multilingual large language model development through cross-lingual alignment transfer.

The project develops technologies to reduce redundancy in multilingual AI model adaptation by reimagining multilingual model development as a cumulative, community-driven process rather than isolated training efforts repeated independently for each language and model version. Specifically, the award funds three research thrusts: methods for capturing multilingual adaptations as reusable capability updates that transfer across successive versions of language models; a modular adaptation framework enabling lightweight adaptation modules to be developed independently of large backbone models, allowing distributed collaboration among multilingual research communities without large-scale retraining; and investigation of shared latent representations that align different models and languages to enable capabilities learned in one setting to be projected and reused in others. The project will release an open library of reusable components to support the multilingual research community.

Performance occurs at the University of Maryland, College Park campus in College Park, Maryland, with a period of performance from August 1, 2026, through July 31, 2029. The award is a project grant under NSF's Computer and Information Science and Engineering program, which supports investigator-initiated research and education in computing, communications, and information science and engineering.