Department of Electrical Engineering and Computer Science (EECS), MIT

Abstract: Large Language Models show amazing abilities to analyze and summarize text, to generate appropriate and even seemingly insightful answers to question, produce advice on how to accomplish tasks and write programs, etc. Yet they also make serious errors, and we cannot predict in any specific instance whether their application will lead to one kind of outcome or the other. We do (mostly) understand the mechanics of how they are trained and how they generate responses, but those mechanics fail to explain why they often work and why they sometimes fail. My interest is in their application to clinical tasks, where the stakes are high, so we need to assure near-perfect performance. So far, we have neither theoretical guarantees nor sufficient empirical evidence for such assurance.