Gagandeep Singh

University of Illinois Urbana-Champaign

Gagandeep Singh is an Assistant Professor in the Siebel School of Computing and Data Science at the University of Illinois Urbana-Champaign (UIUC). He is also part of the Science and Technology group of the Institute of Government and Public Affairs at the University of Illinois. His research combines ideas from formal logic, machine learning, and systems research to construct intelligent systems with formal guarantees about their behavior and safety. His group at UIUC has received several awards and fellowships, including the NSF Career, Google Research Scholar, Amazon Research, and the Qualcomm Innovation Fellowship. He has served on the program committee for top conferences in machine learning, security software engineering, and programming languages, such as NeurIPS, ICML, ICLR, CVPR, ICCV, CCS, ICSE, ASPLOS, CAV, POPL, and PLDI.

Talk

Title: Can We Provide Formal Guarantees for LLM Safety?

Abstract: Formal verification can offer stronger guarantees on the safety of machine learning (ML) models than benchmarking and adversarial testing. However, this technology is generally considered infeasible for LLMs due to both the difficulty in mathematically characterizing LLM safety and the high cost of formal verification for large models. In this talk, I will discuss our recent work to address these challenges, making it possible to provide formal guarantees on the safety of state-of-the-art LLMs.