Zhivar Sourati

I am Zhivar Sourati, a fourth-year Computer Science Ph.D. student at the University of Southern California. My research focuses on the reasoning capabilities of language models, particularly how to evaluate and improve their accuracy, robustness, interpretability, and alignment with human behavior. I draw heavily on insights from cognitive science to better ground LLM reasoning in human cognitive processes, including analogical reasoning, case-based reasoning, prototype-based reasoning, and distinctions between System 1 and System 2 modes of thought. Additionally, I study the broader sociolinguistic and cognitive dynamics of human–AI collaboration, examining how language technologies shape communication and expression in human interaction.

I began my Ph.D. working with Dr. Filip Ilievski on cognitively inspired methods for reasoning in language models, which has remained a central part of my research. Under the guidance of Dr. Morteza Dehghani, I later expanded my focus to explore the interaction between linguistic behavior and emerging forms of human–AI collaboration from a sociolinguistic perspective, while also continuing my work on reasoning in LLMs and their alignment with human cognition.

Beyond my work at USC, I have collaborated with Dr. Dan Roth (Oracle) on multimodal retrieval-augmented generation (RAG) for visually rich document understanding, including dynamic agentic systems for reasoning over complex documents; with Dr. Fred Morstatter (ISI) on factors influencing LLMs’ decision-making; with Dr. Don Tuggener and Dr. Mark Cieliebak (Zurich University of Applied Sciences) on dialogue summarization; and began my research journey as a research assistant with Dr. Behnam Bahrak, focusing on permutation tests and complex networks.

My research interests further include NLP, social network analysis, knowledge graphs, and data mining, with a focus on interdisciplinary problems at the intersection of computer science, psychology, linguistics, and the social sciences.