“Opening the Black Box: New Frameworks for Understanding LLM Internal Behavior” by Dr. Mengnan Du
Join us Monday, September 22 at 4:00 PM in Manc 024! Dr. Mengnan Du, Assistant Professor of Data Science at the New Jersey Institute of Technology (NJIT), will be presenting his research titled Opening the Black Box: New Frameworks for Understanding LLM Internal Behavior.
Abstract: As Large Language Models (LLMs) become increasingly influential in various applications, understanding their internal mechanisms and decision-making processes is crucial for ensuring their reliable and responsible deployment. In this talk, I will explore three critical dimensions of LLM explainability: single-concept understanding, multilingual capabilities, and multimodal alignment. These perspectives together provide a comprehensive view of LLM explainability, from concept representation to cross-lingual and cross-modal understanding. First, we introduce Gaussian Concept Subspace (GCS), a novel framework that moves beyond single-vector representations to model concepts as multidimensional subspaces in LLMs. Second, we examine how LLMs process and understand different languages, introducing a quantitative framework called Language Ranker to evaluate multilingual capabilities across high-resource and low-resource languages. Finally, we investigate the alignment between LLMs and Geometric Deep Models (GDMs) in protein representation, offering insights into how different modalities can be effectively integrated.
Speaker’s Bio: Mengnan Du is an Assistant Professor of Data Science at the NJIT. He received his Ph.D. in Computer Science from Texas A&M University. Prior to joining NJIT, he worked/interned with Microsoft Research, Adobe Research, Intel, Baidu Research, and JD Explore Academy. His research focuses on trustworthy machine learning, with particular emphasis on large language models (LLMs). He has published over 90 papers in top-tier conferences including NeurIPS, ICLR, and ICML, as well as prestigious journals such as TPAMI, CACM, and Cell Patterns. His work has garnered more than 8,500 Google Scholar citations. He actively contributes to the academic community by organizing workshops and tutorials at major conferences including AAAI-24, WWW-24, and COLM-25. He serves as the Senior Area Chair for EMNLP-25, Area Chairs for prestigious conferences including ICLR-26, NeurIPS-25, ICML-25, ACL-25, and AISTATS-25 & 26, and as an Associate Editor for Applied AI Letters.
All are welcome!