Exploring Universal Human Values with Large Language Models: The AWARE-Value Model

Human values play a vital role in everyday functioning, yet their underlying structure remains incompletely understood—for instance, how many core values truly exist, and how these values are organized. Traditional value theories often employ a top-down approach, relying on expert-defined constructs, whereas bottom-up methods (e.g., the psycholexical approach) are data-driven but suffer from limited scale and manual effort. In this study, we leverage advances in large language models (LLMs) to construct a comprehensive value structure. Combining top-down and bottom-up perspectives, we first automatically extracted 4,648 values from vast, diverse, and longitudinal internet corpora, which were refined into 521 key value descriptors using an efficient embedding-based semantic filtering process. Next, to uncover the structure of the value space, we applied LLMs to analyze 392,843 textual passages from 10,995 individuals, quantifying individual tendencies across these value descriptors. Our analysis successfully uncovers the latent organization of human values into a five-factor structure, which we term AWARE: Authenticity, Well-Being, Actualization, Relatedness, and Epistemic Need. This value structures showed high consistency when compared to a parallel human-validated system (N=2,300). Psychometric analysis further confirmed the robustness and reliability of the proposed value structure. We then rigorously validated this structure through data-driven validation across three cross-cultural survey datasets (N = 164,809), which established its robust predictive power. Collectively, these results provide compelling evidence for the reliability, generalizability, and practical relevance of the proposed AWARE structure of the value space.