Towards Agentic Reasoning, Alignment and Planning of Language Models by Xiusi Chen

Abstract:
Despite the strong capabilities of large language models, they still suffer with complex reasoning tasks, aligning with diverse human ethical standards, and interacting with the external world. Identifying these challenges, my research focuses on helping language models enhance the reasoning, alignment and acting by efficiently exploiting external knowledge and exploring external world. In the first part of the talk, I will present data augmentation and knowledge distillation algorithms that can improve the reasoning ability of language models. The second part of the talk will introduce data-driven alignment approaches that make LLMs more aligned with human value and preference. The third part will illustrate teaching language models to better act in the physical world. Finally, I will outline future research directions, including large multimodal models (LMMs) extension, trustworthy prediction in scientific precision, and out-of-distribution data generalization for scientific discovery.

Bio:
Xiusi Chen is a Postdoctoral Research Associate at University of Illinois Urbana-Champaign, working with Prof. Heng Ji. He received his Ph.D. in Computer Science at University of California, Los Angeles, advised by Prof. Wei Wang. Prior to UCLA, he received his B.Sc. degree in Computer Science from Peking University. Xiusi’s research focuses on enhancing LLM reasoning, alignment, dynamics modeling and decision making. His research has been published in top-tier venues in the fields of natural language processing, machine learning, data mining, and information retrieval. Xiusi has been awarded the SDM Best Poster Award (Honorable Mention).