Project Grant 2536298

Award Summary

The National Science Foundation (NSF), Division of Computing and Communication Foundations, awarded a $300,000 Project Grant to the Regents of The University of California at Riverside effective September 1, 2026, through August 31, 2029, under the Computer and Information Science and Engineering (CFDA 47.070) program. This collaborative research project develops methods to enhance the reliability of Large Language Models (LLMs) by quantifying and acting on uncertainty in multi-step reasoning tasks. The research addresses a critical gap in deploying LLMs in high-stakes domains such as transportation planning and healthcare, where models currently produce erroneous outputs while expressing high confidence in incorrect answers.

The project delivers three primary research thrusts: (1) uncertainty quantification methods that represent reasoning as graphs to capture structural inconsistencies and intermediate ambiguities; (2) a unified framework characterizing distinct forms of uncertainty including semantic ambiguity, structural reasoning errors, and inconsistent world knowledge; and (3) uncertainty-aware adaptation methods that incorporate fine-grained reasoning signals into training procedures such as reinforcement learning from human feedback and direct preference optimization. These outputs aim to enable LLM systems to recognize and correct errors autonomously, thereby accelerating safe adoption of artificial intelligence technologies in risk-sensitive applications requiring stable and reliable behavior.