Stony Brook University SUNY Korea

Title: Reliable Reasoning in Large Language Models

Abstract

Large Language Models (LLMs), such as ChatGPT, have greatly expanded what is possible with AI systems. These LLMs can answer questions, generate text, translate, summarize, chat, write code, solve simple problems, and even use tools autonomously to do tasks via external APIs. Measurements on existing benchmark datasets show models seemingly matching human performance on many of these tasks. But are these models reliable? Do they reason in ways we expect them to? Answering these is crucial for responsible and safe use of such capabilities.

In this three part talk, I will first present a brief overview of the basic technology behind models like ChatGPT and how we got there. Then, I will present examples of my team’s work on measuring and improving reliability of LLMs for Question Answering. I will show how we can formalize and measure unreliable reasoning, how to carefully construct datasets that nudge models to learn the right reasoning, and how to get LLMs to hallucinate less. In the third part, I will talk about our current work on designing a testbed for measuring the reliability of autonomous tool use in LLMs.

Bio

Niranjan is an Assistant Professor in the Computer Science department at Stony Brook University, where he heads the Language Understanding and Reasoning lab (LUNR). His team works on various problems in Natural Language Processing and Artificial Intelligence. Prior to joining Stony Brook, he was a post-doctoral researcher in the University of Washington, and was one of the early members of the Allen Institute for Artificial Intelligence. Niranjan completed his PhD in Computer Science from the University of Massachusetts Amherst.