Federal Project Grant Award Summary

The National Science Foundation's Division of Computing and Communication Foundations (CFDA 47.070) awarded $300,000 to the University of California, Berkeley on September 15, 2026, for a collaborative research project addressing hallucination problems in large language models (LLMs) used for health information delivery. The project, extending through August 31, 2029, develops artificial intelligence methods to reduce LLM hallucinations and improve the reliability, transparency, and uncertainty estimation capabilities of language models operating in high-stakes environments. The research focuses on women's health as a primary application domain, encompassing conditions including breast cancer, osteoporosis, cardiovascular disease, autoimmune disorders, and mental health.

The project delivers three primary products and services: (1) creation of multilingual evidence-curated knowledge repositories with reliability scoring in English, Spanish, and French, along with dynamic benchmarks testing model reasoning, attribution, abstention, and clarification capabilities; (2) development of new model training and inference methods enabling non-hallucinating long-form generation, fine-grained attribution, calibrated uncertainty estimation, abstention when confidence is low, and proactive clarification questioning when user queries are ambiguous; and (3) establishment of evaluation frameworks to measure performance improvements. These research outputs aim to accelerate adoption of AI technologies in high-risk domains such as medicine and law where stable LLM behavior is critical for safe and reliable information delivery.