Federal Grant Award Summary

The National Science Foundation's Division of Computing and Communication Foundations awarded a $600,000 Project Grant to the University of California, San Diego under the Computer and Information Science and Engineering program (CFDA 47.070) for the period July 1, 2026 through June 30, 2029. The award supports research into selective prediction techniques for large visual-language models designed to improve the trustworthiness and safety of artificial intelligence (AI) systems. The project addresses the critical challenge of hallucinations in visual-language models—instances where AI systems generate plausible but factually incorrect statements—through development of methods that enable these systems to recognize uncertainty and abstain from making unreliable predictions in safety-critical applications such as healthcare, robotics, autonomous vehicles, and education.

The research encompasses three complementary technical approaches: developing probability calibration algorithms that improve confidence estimate reliability with reduced training data requirements; creating open-vocabulary hallucination verification modules that can be integrated into existing visual-language models without retraining by leveraging retrieval-augmented methods and external reference data; and designing scalable training methodologies to enhance factual consistency between generated text and visual inputs. Beyond the research outputs, the project contributes to the national interest through advancement of trustworthy AI development, safer deployment of AI technologies, and training of students in advanced machine learning and computer vision research.