Recent Awards

This project proposes a unified research framework to systematically evaluate and mitigate these risks by studying (1) reasoning safety, focusing on detecting and correcting unsafe or unreliable reasoning traces; (2) RAG safety, ensuring source faithfulness, uncertainty calibration, and robustness against adversarial retrieval; and (3) multimodal safety, developing alignment and consistency techniques for models that integrate text, vision, and other modalities. The goal is to advance principled methods and tools that enable the safe, trustworthy deployment of next-generation LLMs in real-world applications.