Skip to main content
Back to timeline
arXivSource publication:

Taste divides in shape, not just degree: the same artworks shrink their contrary-weight tail under a new rater sample

Synopsis

Individual differences in how people judge faces, photographs and artworks have mostly been measured as variance. This work writes three candidate shapes as nested generative models and tests them on ten public datasets with about 20,000 raters by predicting held-out people: face judgments need only same-sign reweighting of shared dimensions, while artwork weights spread continuously along the abstract–figurative direction and their tail crosses the population consensus. With the same artworks rated by a different sample, both the individualisation gain and the reversed-weight indicators shrank markedly, so tail size belongs to the rater sample rather than the artworks.

AI-generated editorial illustration: The shape of individual differences in judgment: shared axes, personal weights, and when their spread crosses the consensus

Interpretation

Individual differences in judgment have a shape as well as a size: face judgments differ only in how much weight each person gives shared evaluative dimensions, whereas artwork judgments include people whose weights point opposite to the population. Earlier work used variance decomposition and twin designs to quantify what proportion of variance individual differences account for, and machine-learning models reported only how accurately ratings can be predicted, without comparing what form the differences take. Across ten public datasets and about 20,000 raters, the model estimated each person's latent representation from 10–20 judgments and predicted that person's judgments of stimuli unseen in training; face data saturated at same-sign reweighting across arrangement, rating and triplet tasks.

The model compresses each person into a low-dimensional latent representation inferred by an encoder from a few judgments, and a decoder generates that person's judgment function, with four nested levels allowing no individual differences, same-sign reweighting, a linear transform and stimulus-specific offsets. INDSCAL and IDIOSCAL had already formulated the reweighting and linear-transform shapes, but they required each person's complete similarity matrix and were compared by goodness of fit on samples of a few dozen people. In the Xu et al. face arrangement data, held-out Spearman rose from 0.295 for the population baseline to 0.338 for same-sign reweighting, exceeding the participant's own split-half reliability of 0.309; on the London Set, 20 ratings raised Pearson on the remaining faces from 0.590 to 0.622.

The contrary tail for artworks is the end of a continuous distribution, not a discrete minority group. Model comparison alone cannot establish whether reversed weights exist, because a reweighting model that learns its basis can absorb part of a reversal. Among 553 LAPIS raters, weights were unimodal along the first variation axis, which accounted for 64% of between-rater variance; the population weight lay only 0.62 standard deviations from zero, 28% of raters fell on the opposite side, and 9.6% had a negative cosine with the population weight.

Perspective

The conclusions rest on a reanalysis of ten public datasets covering face arrangement, face attractiveness ratings, face similarity triplets, aesthetic ratings of artworks and photographs, funniness ratings of jokes and pleasantness ratings of odours; 10–20 judgments per person suffice to estimate a latent representation that extrapolates to stimuli not used in training. The author states that this is a reanalysis of existing public data, not a preregistered confirmatory experiment.

The author notes that the hypothesis levels were formulated after analysing the first dataset and that the interpretation of the artwork structure followed the results; the synthetic control and the second rater sample test that interpretation but do not replace preregistration. The 9.6% negative-cosine figure and the corresponding constraint loss in LAPIS should not be read as population constants, since the same artworks gave 3.1% negative cosine and a smaller constraint loss in XPASS-Vis. The two photograph datasets diverged for reasons the present data cannot determine, and outside faces the artwork evidence comes from one dataset plus one additional rater sample.