Models found a signal between shape and noise
Six pretrained vision-language models agreed more when they ranked untextured 3D objects along affective axes, meaning word pairs intended to capture a felt quality, than when they used unrelated adjective pairs. The average Spearman rank agreement, a measure of whether models put objects in the same order, was 0.364. That was below 0.441 for geometric axes, used as positive controls, but above 0.135 for the irrelevant pairs used as an empirical null. The affective-versus-null comparison produced a common-language effect, a way of summarizing how often one set of results exceeds another, of 0.906 and a Kolmogorov-Smirnov statistic of 0.70, another measure of how far two result distributions are separated.
To produce those rankings, the researchers audited the models without fine-tuning them. They sampled 4,950 unique ShapeNetCore objects from 10 categories, up to 500 per category, and rendered each as a uniform-grey, untextured shape from eight fixed viewpoints. Eight categories reached the cap; the sample included 498 bottles and 452 bookshelves. Each model averaged the eight view embeddings into one normalized object representation. It then projected that representation onto directions made from the difference between the text embeddings of opposite adjectives, producing a continuous score for each object on each axis.
The comparison used three tiers: seven geometric adjective pairs, 32 Kansei, or affective, pairs, and 20 irrelevant pairs. The Kansei tier contained 78 category-axis instances, including three shared-core pairs tested across all 10 categories; the geometric and irrelevant tiers contained 70 and 200 instances. For each category-axis combination, the main measure averaged Spearman correlations across model pairs. The study compared tiers with one-sided Mann-Whitney U tests and used 1,000 object-bootstrap replicates to calculate 95% confidence intervals.
The object mattered as much as the label
Agreement was not uniform across the objects or the language used to describe them. On the three shared-core axes, mean convergence ranged from 0.21 for bookshelves to 0.51 for jars. With the full category-specific vocabulary, the means ranged from 0.26 to 0.53. The two approaches tracked each other, with a Spearman correlation of 0.71. In other words, some categories and affective directions produced much more consistent rankings than others.
The researchers checked whether that spread could be explained simply by unstable views. Split-half reliability, a check of how consistently rankings held across different subsets of views, averaged 0.76 and ranged from 0.61 for bookshelves to 0.89 for jars. After correcting the convergence estimates for measurement unreliability, the range was still 0.32 to 0.58. Corrected convergence had little relationship with reliability itself, with ρ = 0.10, so the correction did not remove the category differences.
Another analysis pointed to a possible reason, although it did not establish a cause. The researchers compared each semantic direction with the main patterns of shape variation within a category, represented as principal-component subspaces. Alignment with those patterns was associated with convergence at ρ = +0.72 for affective axes and +0.78 for geometric axes, compared with +0.24 for irrelevant axes; the reported significance threshold for these relationships was p < 0.001. Random directions showed no alignment-convergence relationship, at ρ = -0.04 and p = 0.56. The pattern is consistent with the idea that models agree more when a semantic direction matches variation already present in a category, but it remains a correlation.
The affective signal also did not simply collapse into geometry. Across the selected directions, the strongest reported mean absolute cosine similarity, a measure of directional overlap, was 0.24, with a maximum of 0.42. In a held-out test, agreement calculated from five encoders was associated with agreement from the sixth, with overall ρ = 0.67. That suggests some stability as a model-to-model check, but it still says nothing by itself about whether people would make the same judgments.
The wording changed the result
How the text was constructed changed the result. When the prompts set two opposite adjectives against each other, affective convergence was 0.36 versus 0.14 for the irrelevant tier, a gap of 0.23. Prompts built around a single adjective narrowed the gap to 0.06, with convergence of 0.27 and 0.21. Sentence templates restored a 0.23 gap, with affective convergence at 0.44 and the null at 0.21. The comparison shows why the wording of a semantic control matters to the audit.
Category-specific axes showed a related but imperfect pattern. On their intended category, mean convergence was 0.35, compared with 0.28 on foreign categories; the common-language effect was 0.64 and p = 0.009. The pattern reversed for bottles, clocks and tables, so the category signal was not universal.
A filter for design tools, not a human test
From this, the paper proposes a practical gate for semantic controls in design interfaces. The prototype lets a user select an object and see the six encoder projections for each axis alongside a per-object agreement score. The proposed rule exposes a control when its lower bootstrap bound is above the empirical null, marks it provisional when the interval overlaps the null, and withholds it when the estimate is at or below the null. In the example, luxurious-cheap is exposed for jars at 0.47, but withheld for bookshelves at 0.10 and cabinets at 0.14.
That gate is not a substitute for a human test. The study included no human validation of whether the recovered directions match Kansei judgments and no user evaluation of whether the prototype improves designer performance, decision quality, calibration or trust. The evidence is also limited to six pretrained encoders, 10 categories of untextured, uniform-grey ShapeNetCore objects, eight viewpoints and the selected probe vocabulary. The document is an arXiv version 2 preprint dated 8 September 2026. The authors state that the rendering pipeline, embedding extraction, analysis code and precomputed audit values will be released publicly upon publication. The work was funded by the European Union under Horizon Europe Grant Agreement No. 101226927.
Paper data and sources
Original title: Do Vision-Language Models Agree on the Affective Qualities of Shape? A Cross-Model Audit for Generative Design Interfaces
Authors: Luca Bux, Thiago Rios, Ingo Scholtes, Stefan Menzel
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text