How Large Language Models Adapt When Asked to Personalize: A Comparative Study of Behavioral Responses to Personalization Factors

Project Description

Personalization is a central concept in usable privacy and security, but it is often described in ways that assume a system knows how to personalize. This thesis takes a different perspective: it investigates how large language models behave when they are told what to personalize for, but not how to implement that personalization. In this project, personalization is understood as a system adapting its behavior autonomously, rather than merely allowing a user to configure settings in response to an explicit intention to change the system.

The goal of the thesis is to study how different LLMs respond when prompted to personalize their outputs based on factors such as personality, context, prior knowledge, preferences, or communication style. The focus is not on whether the model can follow explicit formatting instructions, but on whether and how it modifies its behavior when personalization is requested in a more open-ended way. This includes examining whether models infer relevant user traits, how consistently they adapt across prompts, and whether their responses reflect meaningful behavioral change or only superficial adjustment.

The thesis will compare several LLMs under a controlled set of prompt conditions. The student will design prompt scenarios that vary the personalization target while keeping the instruction style constant. The resulting outputs will be analyzed qualitatively and, where possible, quantitatively to identify patterns in adaptability, stability, sensitivity to different personalization factors, and potential differences between models. The study will contribute to a better understanding of how personalization emerges in LLM behavior and what this means for transparency, user control, and privacy in interactive AI systems.

Possible Research Questions

• How do LLMs interpret vague or factor-based personalization requests?
• Which personalization factors lead to the strongest observable behavioral adaptation?
• Do different LLMs personalize in consistent and meaningful ways, or only superficially?
• What implications do these behaviors have for usable privacy and security?

If desired, this project description can also be adapted into a shorter university-style abstract or expanded into a formal thesis proposal with aims, methods, and expected contributions.