Hosted by the DSI Foundations of Data Science Center; Department of Statistics, Arts and Sciences; and Columbia Engineering
Speaker: Eric Wong, Assistant Professor, Computer and Information Science, University of Pennsylvania
Registration for all CUID holders is preferred. If you do not have an active CUID, registration is required and is due at 12:00 PM the day prior to the seminar. Unfortunately, we cannot guarantee entrance to Columbia’s Morningside campus if you register following 12:00 PM the day prior to the seminar. Thank you for understanding!
REGISTER
A Mechanistic Theory of Safety: How Jailbreaking 1-Layer Transformers Taught us how to Steer LLMs
Why are LLM guardrails fundamentally so easily broken, and how can we enforce them? This talk formalizes a mechanistic theory for studying safety problems. We begin with one-layer transformers, identifying rule-breaking as an inherent architectural vulnerability in the model's attention mechanism. This mechanistic theory framework (LogicBreaks) taught us a critical lesson: if attention is the key to breaking rules, it may also be the key to enforcing them.
Building upon this insight, we expand the mechanistic theory to analyze attention-based interventions, arriving at InstaBoost: an incredibly simple yet highly effective steering method that boosts the model's attention on user-provided instructions during generation. This technique, developed from analysis on one-layer transformers, provides state-of-the-art control over large-scale LLMs with just five lines of code.