From Risk to Resilience - Risk Assessment, Safety Alignment, and Guardrails for Foundation Models

Explore a comprehensive lecture on foundation model safety and risk assessment delivered by Bo Li from the University of Illinois at Urbana-Champaign at the Simons Institute. Dive into the critical challenges of deploying foundation models in sensitive domains like healthcare and finance, where reliability is paramount. Learn about unified evaluation platforms for Large Language Models (LLMs) and multimodal foundation models, covering regulatory compliance, toxicity, bias, robustness, ethics, privacy, and fairness. Discover advanced safety enhancement techniques through knowledge integration and safety guardrails, including hallucination mitigation strategies and the benefits of RAG (Retrieval-Augmented Generation) framework. Examine practical insights from evaluating various foundation models using platforms like DecodingTrust, AIR, and MMDT, and understand how to implement efficient guardrail frameworks for real-world risk mitigation.