Computational Representation of Individual Judgment

Can a large language model (LLM)-based computational representation carry enough intentionally supplied information about a particular individual to exercise a measurable component of that individual’s judgment? We investigate this question in a bounded food-preference domain using participant-specific representations constructed from explicitly supplied preferences, constraints, and subsequent corrections. Across 99 adult participants and 4,460 participant-item judgments, the representations exhibited a mean participant-level graded agreement of 74.20% with subsequent participant judgments, together with positive ordinal association and heterogeneity across individuals. Correspondence was also observed for items absent from participants’ final explicit profiles, although the adaptive teaching procedure limits interpretation as conventional zero-shot generalization. Compared with a specified indirect-attribute baseline using demographic information and food-novelty orientation, the explicitly taught representations showed 5.24 percentage points higher mean graded agreement, with the comparative difference concentrated primarily in negative participant judgments. Simple constant-rating baselines, systematic differences in use of the rating scale, adaptive in-session teaching, and substantial individual variation constrain stronger claims. The results provide initial evidence that intentionally supplied preference information can support a computational representation exhibiting measurable correspondence with an individual’s subsequent judgments within a specific domain.