Knowledge in Large Language Models: Quantification, Validation and Dissection

TITLE  "Knowledge in Large Language Models: Quantification, Validation and Dissection"


ABSTRACT: Large language models are trained on large volumes of text, and can thus capture the rich factual knowledge described therein. This has great potential in terms of what these models can converse and reason about. However, we are far from understanding what this knowledge is, how it is encoded, and how to check when LLM statements are factually true or not. In this talk I will describe our work on these problems. I will begin with work on "crawling" language models, meant to systematically describe the facts they encode using a knowledge graph. I will then describe  work on understanding the internal mechanisms that underlie knowledge representation in LLMs. Finally, I will describe an approach to detect factually incorrect statements in LLMs, via a concept of "Cross Examination" where one LLM interrogates another LLM, seeking factual inconsistencies. Time permitting, I will also describe our work on understanding implicit biases in learning RNNs.


BIO: Amir Globerson received a BSc in computer science and physics from the Hebrew University, and a PhD in computational neuroscience from the Hebrew University. After his PhD, he was a postdoctoral fellow at the University of Toronto and a Rothschild postdoctoral fellow at MIT. He joined the Hebrew University school of computer science in 2008, and moved to the Tel Aviv University School of Computer Science in 2016. He is also a research scientist at Google and is currently on sabbatical at Google NYC. He served as an Associate Editor in Chief for the IEEE Transactions on Pattern Analysis And Machine Intelligence. His work has received several paper awards (at NeurIPS,UAI, and ICML). In 2018 he served as program co-chair for the UAI conference, and in 2019 he was the general co-chair for UAI in Tel Aviv. In 2019 he received the ERC consolidator grant. He is serving as program co-chair at NeurIPS 2023, and will serve as NeurIPS 2024 general chair.