A celebrity-based face score tracks patterns in academic reference letters

A recommendation letter is meant to convey what a candidate can do. But its length and tone might also reflect factors that are harder to separate from professional qualifications. In a new preprint, economists Fabian Slonimczyk and Shergazy Kimsanbaev investigate whether one such factor is appearance.

Their study combines an experimental way of scoring faces with hiring records from one economics department. Applicants whose photographs received higher scores tended to have longer reference letters and letters expressed in more favorable language. The relationship between those scores and subsequent job placement was less consistent. Neither result establishes that appearance caused the observed differences.

Comparing faces with a reference set

The authors call their measure the relative vision embedding, or RVE, score. An embedding is a numerical representation produced by a computer vision model from an image. The researchers crop faces from photographs, pass them through a pretrained model and compare the resulting representations with those of a reference group.

That reference group consists of 776 celebrities drawn from public attractiveness rankings, separated into race and gender categories. Each celebrity is given equal weight within a chosen reference set. A person's score reflects the average proximity of their image representation to those reference faces. The researchers also tried comparisons against all celebrities, celebrities of the same gender, and celebrities in the same race and gender category.

The method does not require new ratings for each applicant or additional training of the vision models. But its benchmark ultimately depends on human choices: the celebrities came from lists ranked by website users. The score measures similarity under a particular model and reference set; it is not an objective or universal measure of beauty.

How well does the score match human ratings?

To check the measure, the authors used the Chicago Face Database, which contains standardized photographs and attractiveness ratings from volunteers. They examined 597 people, each rated by 28 volunteers, and compared the average human rating with scores from several vision models.

Agreement varied substantially with the model and the comparison method. The strongest individual-model rank correlation reported was 0.484: a positive relationship, but far from perfect agreement. One model, Qwen3, aligned poorly with the human ratings and was excluded from the three-model combination used in the main hiring analysis.

The combined score also performed differently across groups. In the database, women received higher average human ratings than men, while some RVE variants gave women lower average scores. Within-group correlations were generally stronger than the pooled correlation, but differed by gender and race. These differences matter when interpreting a score that can capture more than the feature researchers intend to measure. The authors also found that image representations separated celebrity photographs from database photographs; the two sets differ in photographic conditions as well as in who appears in them.

Letters and job outcomes

The hiring data cover fresh economics Ph.D. applicants to a research-oriented department in Russia between 2013 and 2021. The authors found public photographs for 630 of 645 candidates, collecting 1,394 images in total. They combined multiple image representations where available and examined 1,925 reference letters.

The authors used an existing language-model-based sentiment measure to assess the letters, alongside their word counts. Higher RVE scores were associated with longer letters and more favorable sentiment, including in analyses accounting for candidate characteristics, Ph.D. program, year and characteristics of letter writers. The authors report that the positive associations persisted across 45 alternative measurement and regression specifications.

In their main gender comparison, the association between score and letter sentiment was steeper for women than for men. The corresponding difference in the letter-length relationship was not statistically significant, and the sentiment difference was not statistically significant in every alternative specification.

For job-market outcomes, the authors defined success primarily as placement at a top-100 research academic institution or a top-50 policy institution. Higher scores were positively associated with this outcome, but the estimate weakened as controls were added and was not statistically significant once letter length and sentiment were included. The authors caution that letters may themselves be affected by appearance, so controlling for them does not settle whether appearance matters.

What remains unresolved

The results are observational. Applicant qualities, letter-writer knowledge, photographic conditions and other factors could contribute to the patterns; statistical controls cannot rule out every alternative explanation. Nor does the study demonstrate that letters transmit an effect of appearance into hiring decisions. It identifies a pattern consistent with that possibility.

The evidence also comes from applicants to a single department, with a much smaller number of women than men and only five Black candidates. Results may differ in other applicant pools or institutions. The paper is an arXiv preprint; the provided record does not establish that it has been peer reviewed.