The research, led by Yue Xu of the California Institute of Technology together with Umit Keles, Chujun Lin, and Ralph Adolphs, tackled a fundamental methodological worry that has haunted the field. When experimenters provide trait words in advance, they constrain what participants can report. If you only offer two words, you can only ever find two dimensions of judgement. If the words are semantically redundant, the results will look artificially simple. Worse, the dimensions recovered might reflect the quirks of the English language rather than the underlying psychology of how we actually perceive people. The team wanted to know: what happens to the structure of face judgements when you strip away language entirely?

Their solution was a task called multi-arrangement, in which participants dragged digital face photographs around a white circular arena on a computer screen. The instructions were simple: place faces that seem similar in personality close together and faces that seem different far apart. No trait words were ever mentioned. Each participant worked with a subset of 20 faces, and an adaptive algorithm chose which faces to show together on each trial, resampling pairs that had been placed close to one another so their similarity could be disambiguated in later rounds. After 20 minutes of arranging, the on-screen distances were converted into pairwise dissimilarity scores and aggregated across thousands of participants into a single face-by-face similarity matrix, a representational dissimilarity matrix that captured how the group as a whole mentally organized the faces.

The first question was whether people could even do this reliably. The answer, backed by an impressive battery of quality checks, was yes. In the first experiment, which used 100 predominantly White faces, the overall consensus among participants reached a Spearman-Brown-corrected correlation of 0.82, and in the second experiment, which used a racially diverse set of 96 faces, consensus climbed to 0.90. Individual test-retest reliability was more modest, but crucially, the most internally consistent participants also agreed most with one another, suggesting everyone was approaching the task in fundamentally the same way. Every face pair was judged by at least 59 people, and the five-dimensional solution the team ultimately reported proved robust even when they randomly decimated the number of faces or participants.

Validity checks ruled out the possibility that participants were merely sorting faces by superficial appearance. The team compared their aggregated similarity matrix against reference matrices built from random permutations, hand-measured physical features like face width-to-height ratio and skin colour, and layer-by-layer activations of a VGG16 neural network trained to identify faces. Superficial visual features, particularly gender, did correlate with the judgements, but they could not explain them away. The strongest correlations of all came from trait-word rating data collected in previous studies, exactly what you would expect if participants were genuinely inferring psychological traits rather than comparing nose lengths. A post-task questionnaire added a qualitative layer: participants who described making social inferences produced data that matched the group aggregate slightly better than those who reported focusing only on visual features, yet both subgroups yielded essentially the same dimensional structure.

When the team applied multidimensional scaling to the aggregated data, cross-validation pointed decisively to five dimensions, which together explained 76 percent of the variance in the first experiment. The first and largest dimension, accounting for 27 percent of the variance, was interpreted as a spectrum from threat to approachability, and it correlated strongly with words like “criminal” and “affectionate” from earlier rating studies. The second dimension captured warmth judgements that split along gender lines: the undesirable traits associated with male faces, such as criminality, separated from those associated with female faces, such as grumpiness. The remaining dimensions mapped onto competence, a morally charged axis of virtuousness versus condescension, and youthfulness, which tracked character-related attributes like conservatism rather than purely appearance-based cues.

The second experiment extended the findings to faces one-third each of White, Black, and Asian individuals, with a correspondingly diverse participant pool. Once again, five dimensions emerged, and once again they explained the bulk of the variance. But the content shifted in a revealing way. The dominant first dimension was now driven by White-versus-Black race bias, something entirely absent from the first experiment and from the classic word-rating literature, which had mostly used White faces. A third dimension combined stereotypes about Asian faces with youthfulness and femininity. Gender and approachability persisted as core dimensions, but the racial stereotypes surfaced precisely because the stimulus set allowed them to. This is a striking demonstration that the dimensions of face judgement are not fixed properties of the human mind but depend on the faces being judged.

Comparisons with three landmark word-based studies, including Oosterhof and Todorov’s influential two-dimensional valence-dominance model from 2008, revealed both continuity and divergence. The previous studies’ latent structures correlated more strongly with one another than with the new wordless structure, hinting that word labels do impose their own semantic organization. Yet the broad strokes agreed: a small number of dimensions capture most of the variance, and an evaluative warmth or approachability factor is essentially universal. Perhaps the most intriguing conclusion is that this first dimension integrates multiple trait facets, trustworthiness, warmth, and competence, into a single holistic evaluation. When people are not given words, they do not parse faces into tidy psychological categories; they collapse everything into one gut-level judgement of whether this is someone to approach or avoid.

The authors are careful about limitations. Participants might have covertly used words while arranging faces, though their self-reported strategies do not suggest a strongly lexical approach. The task is time-consuming, so each participant handled only 20 faces, meaning all findings rest on aggregated rather than individual-level data, which carries a risk of inflated effect sizes. The age range of the stimuli was compressed compared with some earlier datasets, and the real-world spontaneity of first impressions, which unfold in messy social contexts, can never be fully captured in a laboratory arena. Future studies with people who have aphasia, pre-verbal infants, or even nonhuman animals could help determine whether the low-dimensional structure truly exists independently of language.

Still, the implications are considerable. The study confirms that the low dimensionality reported across decades of word-rating research was not an artefact of the words themselves; a five-dimensional solution emerged whether or not language was involved. At the same time, it shows that wordless judgements have a somewhat different structure, more holistic and more sensitive to the demographics of the faces on display. The approachability-threat spectrum appears to be a universal evaluative backbone of human social perception, while race and gender stereotypes emerge or fade depending on who is in the room, or on the screen. As the authors speculate, the dimensions they uncovered may represent a coarse, rapid social evaluation that becomes more elaborate as we think longer and reach for words. The next time you form an instant opinion of a stranger’s face, you are navigating a five-dimensional psychological space, one that science is only now beginning to map without asking you to name a single trait.

Subject of Research: Trait judgements of faces measured with a wordless multi-arrangement task

Article Title: Trait judgements of faces using a task without word labels

Article References: Xu, Y., Keles, U., Lin, C., & Adolphs, R. (2026). Trait judgements of faces using a task without word labels. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02605-2

Image Credits: AI Generated

DOI: 10.1038/s41562-026-02605-2

Keywords: face perception, first impressions, trait inference, multidimensional scaling, social cognition, stereotypes, multi-arrangement task, representational similarity analysis, Nature Human Behaviour, psychology, race bias, gender stereotypes