Comparing graphs to identify similarities is a fundamental task in visual analytics of graph data. To support this, visual analytics systems frequently employ quantitative computational measures to provide automated guidance. However, it remains unclear how well these measures align with subjective human visual perception, thereby offering recommendations that conflict with analysts’ intuitive judgments, potentially leading to confusion rather than reducing cognitive load. Multimodal Large Language Models (MLLMs), capable of visually interpreting graphs and explaining their reasoning in natural language, have emerged as a potential alternative to address this challenge. This paper bridges the gap between human and machine assessment of graph similarity through 3 interconnected experiments using a dataset of 1881 node-link diagrams. Experiment 1 collects relative similarity judgments and rationales from 32 human participants, revealing consensus on graph similarity while prioritizing global shapes and edge densities over exact topological details. Experiment 2 benchmarks 16 computational measures against these human judgments, identifying Portrait divergence as the best-performing metric, though with only moderate alignment. Experiment 3 evaluates the potential of 3 state-of-the-art MLLMs (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5) as perceptual proxies. The results demonstrate that MLLMs, particularly GPT-5, significantly outperform traditional measures in aligning with human graph similarity perception and provide interpretable rationales for their decisions, whereas Claude Sonnet 4.5 shows the best computational efficiency. Our findings suggest that MLLMs hold significant promise not only as effective, explainable proxies for human perception but also as intelligent guides that can uncover subtle nuances that might be overlooked by human analysts in visual analytics systems.

Objects in scenes follow a hierarchical organization, with "scenes" at the top level, followed by "phrases", clusters of objects that share spatial and functional proximity. Within these phrases, "anchor" objects help predict the identity and location of smaller, dependent "local" objects. Previous research has shown that this hierarchy is reflected in the mental representations of objects in adults. The current study examined whether children's object representations already reflect this hierarchy. We implemented an odd-one-out task with 36 object images to collect pairwise similarity ratings from children ages 5 to 10 years. Two different groups of children received different similarity judgment instructions: One group received no explicit definition of similarity, but the other was told to base similarity on actions typically performed with the objects. We created a priori and data-driven scene hierarchy measures to evaluate how well they aligned with children's similarity judgments. Results showed that children's representations were clearly structured at the scene level, as indicated by strong effects in both hierarchy measures. In contrast, we found no reliable phrase-level effects and only a small data-driven object-type effect. Scene-level structure strengthened with age, whereas phrase- and object-type levels showed no reliable age-related change. Importantly, similarity patterns were highly comparable across both tasks, suggesting that children's object representations by default seem to be action based. These results suggest that children organize objects along the scene level of the hierarchy incorporating actions related to the objects in their representations, whereas finer-grained relations are more weakly represented and may be more difficult to detect reliably at this age.

Generative artificial intelligence (AI) models unlock new ways to create images, emerging as a new medium alongside paintings, photographs, physically based renderings (PBR), etc. Generative AI images can be perceptually convincing without being physically plausible, allowing to investigate the boundaries of visual perception. This study examines whether generative AI images adhere to a medium-independent perceptual space converged from previous studies. We compared the perceptual similarity of images from three generative AI models against a bidirectional reflectance distribution functions (BRDFs) PBR image dataset, using human similarity judgments. In experiment1, we used the text descriptions of 32 materials (e.g., blue acrylic) from the Mitsubishi Electric Research Laboratories (MERL) BRDF dataset, prompting two text-to-image models, DALL-E 2 and Midjourney v2, to generate 32 sphere-shaped stimuli per model. Perceptual spaces derived from similarity judgments revealed that both AI models resulted in two-dimensional spaces whereas the MERL space was confined to one dimension, probably owing to a lack of surface texture. These unrelated perceptual spaces suggest the AI models generated unique and different images from identical text prompts. In experiment2 we used the text-to-image model Stable Diffusion v1.5 with ControlNet for additional depth-map constraints. Using the same 32 descriptions, we generated 3 sets using 3 different depth maps. The three resulting perceptual spaces are all two-dimensional, exhibiting high similarity, indicating a robust and non-random structure. They also show a similar structure to the MERL space and perceptual spaces from other material studies using photographs, PBR, and depictions, suggesting AI-generated imagery may indeed be used as a new medium to explore material perception.

Investigating fine- and coarse-grained structural correspondences between deep neural networks and human object image similarity judgments using unsupervised alignment.

Concurrent cognitive load and lexical-semantic similarity judgments for action verbs and object nouns in Persian-speaking adults who stutter.

Waste printed circuit boards (WPCBs) contain valuable metals such as gold, palladium, and silver, which are typically recovered through non-ferrous metal smelting. Currently, WPCBs are manually classified by workers, who visually compare board colors and component layouts with previously processed boards. This approach is time-consuming and prone to human error. To address these limitations, we propose an image-based algorithm for automated WPCB similarity assessment. The method extracts visual features from board images and computes similarity scores, incorporating classification strategies based on board-specific characteristics. Key features identified as effective for similarity evaluation include the hue value, coefficient of variation in terminal regions, number of line elements in terminal regions, structural complexity, and number of integrated circuits. Weighted feature contributions further improve accuracy. Our experimental results demonstrate that the proposed approach achieves 88.0% accuracy for the targeted PCB types, outperforming a comparative self-supervised contrastive learning method. This image-driven solution can significantly streamline WPCB recycling by reducing reliance on manual inspection and improving operational efficiency.

Latine Americans are stereotyped as socially inferior and culturally foreign, with darker skinned Latine individuals typically experiencing greater stereotyping than their lighter skinned counterparts. However, lighter skinned Latine individuals can also appear highly prototypical of the Latine category, raising the question of whether they experience similar stereotyping as dark-skinned Latine individuals. This research reports three preregistered experiments testing this possibility. In Studies 1 and 2, participants were randomly assigned to evaluate the image of a gender-matched Latine individual who was (a) light-skinned and low in Latine prototypicality, (b) light-skinned and high in Latine prototypicality, or (c) dark-skinned and high in Latine prototypicality. Study 3 employed a repeated-measures design with multiple faces per category to capture greater facial variability and enhance external validity. Across all studies (N = 723), prototypically appearing Latine individuals-regardless of skin tone-were perceived as more foreign and socially inferior than low-prototypically appearing Latine individuals. Self-identified White (n = 230) and Latine (n = 171) participants showed similar judgment patterns (Study 2). Differences in perceptions of highly prototypical light- versus dark-skinned Latine individuals were few and inconsistent; when differences did emerge, neither group was systematically perceived as more stereotyped than the other. These findings suggest that light-skinned Latine individuals perceived as highly prototypical may experience explicit cultural stereotyping to a greater extent than less prototypical light-skinned Latine individuals and at levels comparable to those experienced by dark-skinned Latine individuals. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

ABSTRACT This study investigates whether a genderless second language (L2; English) reshapes grammatical gender (GG) effects derived from two gendered first languages (L1s; Ukrainian and Russian) in multilinguals. Previous work showed that when GG aligns across these L1s, it influences categorisation and recall of multilinguals, whereas mismatching L1s yield effects from the more proficient L1 only on categorisation. Building on earlier evidence that L2 acquisition can modulate, or restructure, GG-driven effects on cognition”–, we examined whether English dominance attenuates effects of Ukrainian and Russian, and which experiential factors predict such restructuring. Across a similarity judgement task and an associative-learning task, English weakened L1-based gender effects when gender aligned across L1s, indicating dilution of gendered conceptual representations. However, when GG systems mismatched, effects varied across tasks, with restructuring emerging only in categorisation. Furthermore, experiential factors such as daily L2 use, use of English in family contexts, and years of schooling emerged as the strongest predictors of such restructuring. Overall, our findings suggest that a genderless L2 does not uniformly override GG-based representations but interacts dynamically with multilinguals’ L1 structures and language experience, revealing asymmetries in how conceptual systems reorganise across cognitive domains.

This research introduces a new method for constructing and training an Urban Space Embedding Model (USEM) by integrating human perceptions and street-level images (SLI) into its formulation. Traditional urban embedding models often overlook subjective human experiences, such as perceptions of safety or attractiveness. To address this gap, our method leverages similarity judgments from over 1500 participants, who compared different urban spaces based on SLI. These human judgments were then used as a supervision signal in training the USEM, allowing the model to capture both visual and perceptual information about urban spaces. The method is implemented across the Netherlands, using around one million geo-tagged SLI, and demonstrated in Rotterdam. This approach represents a significant advancement in urban computing by incorporating human-centered data into urban modeling. It offers new opportunities for city planners and policymakers to better understand how urban spaces are perceived and to consider these perceptions in efforts to design more livable and inclusive environments.

Our visual world consists of an immense number of unique objects and yet, we are easily able to identify, distinguish, and reason about the things we see within a few hundred milliseconds. Here, we used a large-scale and comprehensively sampled stimulus set and developed an analysis approach to capture how rich, multidimensional object representations unfold over time in the human brain. We modeled time-resolved MEG signals of four humans (two females and two males) viewing single presentations of tens of thousands of object images based on millions of behavioral judgments. Extracting behavior-derived object dimensions from similarity judgments, we developed a data-driven approach to guide our understanding of the neural representation of the object space and found that every dimension is reflected in the neural signal. Studying the temporal profiles for different object dimensions, we found that the time courses fell into two broad types, with either a distinct and early peak (∼125 ms) or a slow rise to a late peak (∼300 ms). Further, early effects were stable across participants, in contrast to later effects which showed more variability, suggesting that early peaks may carry stimulus-specific and later peaks more participant-specific information. Dimensions with early peaks appeared to be primarily visual dimensions and those with later peaks more conceptual, suggesting that conceptual representations are more variable across people. Together, these data provide a comprehensive account of how behavior-derived object properties unfold in the human brain and form the basis for the rich nature of object vision.

Dialects vary in their allophonic patterns, potentially affecting listeners' cognitive representations of language. How different exposure (in terms of size of acoustic distinction) to dialect-specific allophonic patterns for two American English vowels, /æ ai/, affects listeners' behaviors in a perceptual similarity judgments task is explored. Theories about the phonologization of allophonic patterns, which have largely relied on production data, using novel perception data are examined. Listeners from northeastern Ohio (North), a region with less /æ/ allophony and more /ai/ allophony, and listeners from central and southwestern Ohio (Midland), a region with more /æ/ allophony and less /ai/ allophony, are contrasted. Contextual allophones of the same phoneme sound more similar to listeners with less exposure to dialect-specific allophonic variation than they do to listeners with more exposure. Exposure influences the degree of phonological contrast present in listeners' cognitive representations of allophones. These results support late abstractness theories of phonologization, in which phonological abstractness gradually develops during a sound change in progress in a community.

ABSTRACT Since the article touches on both the fields of Roma studies and Alevi studies, it is related to the extensive literature in both fields. A review of the literature on Roma in Turkey reveals that research on the religious life of Roma is insufficient, and similar judgements are repeated. Roma have not found a place in the literature centred on Alevism either. Having two historically marginalised identities, Alevi Roma become the object of a multilayered mechanism of exclusion. Indeed, Alevi Roma are not only excluded by non‐Alevi groups. Non‐Alevi Roma people also do not accept Roma Alevis and approach their beliefs with suspicion. The article highlights the exclusionary practices faced by Roma Alevis living in different cities across Turkey and emphasises that exclusion within the group is not an exceptional occurrence. Based on ethnographic observations and in‐depth interviews, this article focuses on the ways in which Roma from Uşak perceive the Alevi faith and the practices of in‐group exclusion they face. The article also provides information on the settlement of Roma groups in the city centre of Uşak, the development of a 40‐year‐long cemevi practice and the Alevi ocak system to which they belong.

Theories of transfer argue that people are more likely to transfer knowledge to a new scenario the more similar the scenario is to what they have previously learned. However, prior research predominantly relies on expert- or researcher-based judgments of how similar two scenarios are, rather than learner-based similarity metrics. Two studies (N total = 483) with undergraduate students in the United States examined how learner-based similarity judgments relate to transfer. These studies also show how using learner-based metrics can help researchers explore how features of lessons (i.e., the richness of diagrams) influence transfer. Participants sorted the stimuli in the posttest based on their similarity either at the beginning (Study 1) or the end of the study (Study 2). Participants learned about metamorphosis using either perceptually rich or bland life cycle diagrams. After the lesson, they completed a posttest after the lesson and after a month. Both studies showed that participants' similarity judgments predict transfer. Using this metric also showed that participants were more likely to extend their knowledge to animals similar to the ladybug when they learned with the rich diagram, but to dissimilar animals when they learned with the bland diagram. This was consistent after the 1-month delay. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

Is valence an intrinsic dimension of conscious experience, as different authors have suggested? If so, all conscious experiences, and hence all conscious perceptions, should be valenced, even if only minimally so, and similarity judgments should be at least partly driven by one's affective dispositions. Leveraging the concept of micro-valence, we explore the extent to which valence judgments correlate with similarity judgments and with the different stages of processing in deep neural networks (DNNs). One hundred forty-nine participants provided both similarity and valence judgments for 120 images of everyday objects, using an odd-one-out task (Study 1), a spatial arrangement task (Study 2), and the Birthday task, which asks people to choose an object they would like to keep (or give away) as their birthday gift. We also extracted activations from the layers of DNNs trained to classify objects in response to the same images. Representation similarity analysis and multidimensional scaling analyses highlight the role of micro-valence in the similarity space, suggesting that valence permeates similarity judgments. DNN analyses show that this valence-similarity relationship is not entirely mediated by stimulus perceptual features and suggest that low-level visual features play a role in the computation of valence.

Conceptual similarity as aggregation over feature sets in geometric spaces.

Large language models have shown considerable abilities across many tasks, but their capacity to detect sensitive user information from text raises significant privacy concerns. While recent approaches have explored sanitizing text to hide private features, a deeper challenge remains: distinguishing true privacy preservation from deceptive transformations. In this paper, we investigate whether LLM-based sanitization reduces private feature leakage without misleading an adversary into confidently predicting incorrect labels. Using LLM as both sanitizer and adversary, we measure leakage using two entropy-based metrics: Empirical Average Objective Leakage (E-AOL) and Empirical Average Confidence Boost (E-ACB). These allow us to quantify not only how accurate adversarial predictions are, but also how confident they remain post-sanitization. We posit that deception, while reducing adversarial accuracy, will also increase confidence in incorrect inferences, and hence reduced accuracy alone should not be interpreted as true privacy. We show that while current LLMs can hide private features, their transformations sometimes cause deception. Finally, we evaluate the semantic utility of sanitized outputs using sentence embeddings, LLM-based similarity judgments, and standard metrics like BLEU and ROUGE. Our findings emphasize the importance of explicitly distinguishing between privacy and deception in LLM-based sanitization and provide a framework for evaluating this distinction under realistic adversarial conditions.

What makes good exemplars of a scene category good? Evidence from deep neural nets.

Humans develop biases during language learning. For example, we rely more heavily on consonants than on vowels to identify words. Advances on artificial intelligence have allowed the development of proficient large language models that sometimes mimic humans’ language use. They do so by tracking regularities in natural language datasets that are used to train them. Here we test the hypothesis that tracking such regularities is enough for the emergence of responses that resemble the consonant bias. We asked ChatGPT which of two nonsense words (one with a vowel and one with a consonant change) was more similar to a target word. We observed that the model uses more the consonants than the vowels to perform similarity judgments across words in the two languages that we tested (English and Spanish).

Semantic feature norms have been foundational in the study of human conceptual knowledge, yet traditional methods face trade-offs between concept/feature coverage and verifiability of quality due to the labor-intensive nature of norming studies. Here, we introduce a novel approach that augments a dataset of human-generated feature norms with responses from large language models (LLMs) while verifying the quality of norms against reliable human judgments. We find that our AI-enhanced feature norm dataset, NOVA: Norms Optimized Via AI, shows much higher feature density and overlap among concepts while outperforming a comparable human-only norm dataset and word-embedding models in predicting people's semantic similarity judgments. Taken together, we demonstrate that human conceptual knowledge is richer than captured in previous norm datasets and show that, with proper validation, LLMs can serve as powerful tools for cognitive science research.

Abstract Work with interviews has become widespread in recent decades, and modern technologies provide the choice between using only the transcript for further evaluation or additionally taking advantage of the audio track. Given that voice can carry extra information not contained in the text, this might lead to differences in judging the content of the text elements. We experimentally compared the evaluation of statements made by contemporary witnesses from the postwar period about the development of psychology in Germany. Fifty-four subjects rated whether the 26 quotations expressed either continuity in psychology after the war or rather a new beginning on a Likert scale. We experimentally varied whether or not the witnesses’ voices were added to the transcripts. On average, quotes with vs. without voice led to similar judgments. However, individual ratings in the condition with voice added tended to be more extreme. These results are relevant for future projects combining written and audio material.