FeaturedUndergraduate thesis · FCEN-UBA · 2026
Value Systems in LLMs
Effects of Paraphrasing and Profile Elicitation on Decision-Making Consistency and Robustness
Whether language models have real values is contested. My thesis asks a more tractable question: do they have stable ones? I crossed ten meaning-preserving prompt perturbations (typos, negation, paraphrase, translation, reversed options...) with fourteen elicited personality profiles, drawn from the Big Five, Schwartz's basic values, and Moral Foundations Theory, and measured how consistent and robust the models' moral choices remained.
