Large Language Models as Machines of Beauty: Cognitive Averaging, Latent Space Geometry, and the Entropic Foundations of Aesthetic Preference

Abstract: This study advances the position that large language models (LLMs) and human perceptual systems are governed by a shared computational drive toward prototypicality, entropy reduction, and aesthetic coherence. Drawing on developmental evidence that infants exhibit early preferences for facial symmetry and averageness, the analysis situates aesthetic preference within broader research on processing fluency and predictive coding, emphasizing that biological perception rewards stimuli that reduce uncertainty and support efficient information compression. This foundation is used to examine how LLMs, through cross-entropy optimization, perplexity minimization, and latent space clustering, converge on high-density representational regions that operate as statistical prototypes of linguistic and conceptual categories. The examination shows that centroids within latent space function as computational counterparts to psychological prototypes, while attention mechanisms act as filters that amplify structured regularity and suppress idiosyncratic variation. Through the integration of perspectives from cognitive psychology, computational neuroscience, and machine learning, the study reframes aesthetic qualities as emergent properties of systems optimized to stabilize input and maximize predictive coherence. This perspective also clarifies phenomena such as mode collapse and embedding drift as consequences of excessive convergence toward prototypical structure, paralleling aesthetic degradation observed when biological systems over-attenuate variability. The significance of this argument lies in demonstrating that beauty can be modeled as a measurable outcome of intelligent information processing, linking infant cognition, neural prediction dynamics, and the generative capacities of artificial systems through the common logic of prototype formation and entropy minimization.