BU Computer Systems Seminar

Talk Title: Towards Faithful LLM Systems: Three Early Examples

Speaker: Cheng Tan, Assistant Professor, Northeastern University

Abstract: Large Language Models (LLMs) are becoming central to modern computing, which elevates the importance of the underlying systemsthat support their training and serving. Unlike traditional applications, LLMs are inherently stochastic and nuanced, making their behavior difficult to specify, constrain, or control. This complexity obscures misbehavior in the underlying systems, making it harder to verify, detect, or attribute faults. As a result, ensuring the faithfulness of LLM systems---whether they behave as intended or expected---has become both a pressing and technically challenging problem.

In this talk, I present a vision of building faithful LLM systems that align with user intent---securely and verifiably. I will discuss three initial efforts toward this goal: a jailbreak oracle, a near-zero-overhead GPU information flow tracking, and a verifiable distributed training system.

Bio: Cheng Tan is an assistant professor at Khoury College of Computer Sciences at Northeastern University. His research interests are in computer systems, verifiable systems, and ML systems. He is a recipient of the SOSP’17 Best Paper Award, ASPLOS'23 Distinguished Artifact Award, MSRA StarTrack Scholar, and NSF CAREER Award.