About me
Hi, I’m Chimaobi.
I’m a third year PhD candidate in Computer Science and Engineering at the University of Michigan, co-advised by Professors Joyce Chai and Rada Mihalcea. Before Michigan, I earned my Bachelor’s degree in Electrical and Electronics Engineering from FUTO, Nigeria, where I specialized in Electronics Engineering (ECE) and graduated as the best student in my class.
My research interests span AI alignment, personalization, LLM safety, robustness, and multi-user agents.
Research Theme
I seek to make AI models more useful for YOU and all users. As such, I am mostly interested in AI Alignment: per-user alignment(personalization), alignment to user groups (e.g cultures) and general alignment to human core values AKA the 3Hs (Helpfulness, Harmlessness, and Honesty). My research is in two folds
Life-long personalization: People vary widely in their goals, preferences, and contexts, even within the same task or prompt. As they interact with LLM-powered conversational systems such as ChatGPT, Gemini, Claude, etc., they generate rich, evolving interaction histories that reflect these individual differences. Appropriate personalization thus requires first adequate user context/information retrieval, followed by personalized response generation. In [1], I highlight failure modes across both stages, offering new formulations and evaluations toward building better personalized systems.
Robust Personalization: In optimizing for personalization, how do we ensure robustness - such that aligning user preferences does not affect factuality of models nor raise safety concerns. I formalize the notion of robustness in the context of personalization and highlight critical issues with current evaluation approaches that solely focus on alignment [2]. By rethinking how we evaluate and design personalized AI systems, I seek to build methods that preserve truthfulness and prevent harmful failure modes, while still adapting meaningfully to diverse user goals.
TL;DR: I study robust life-long personalization of AI agents – seeking ways to better adapt LLMs to user features (implicit, explicit, latent) in a dynamic fashion without compromising safety and factuality.
