A computational neuroimaging study tracking autistic adolescents across two testing sessions found that the majority successfully formed abstract concept prototypes rather than relying on hyper-specific memories of individual examples. Published in the journal Molecular Autism, the investigation evaluated sixty-four autistic youth aged 14 to 18 years without intellectual disability to resolve a long-standing debate in cognitive neuroscience. Prior theoretical models had hypothesized that autism is characterized by an exemplar-specific perceptual bias, in which processing favors local sensory details over global integration and thus limits the spontaneous formation of summary representations. By pairing behavioral categorization tasks with computational models and functional magnetic resonance imaging, the researchers demonstrated that abstract concept formation was not only achievable for most participants but also served as the dominant and temporally stable cognitive strategy across a nine-month interval.
To investigate how concepts are encoded and generalized, the research team designed an experimental paradigm based on artificial cartoon animal categories, including fish, bugs, and butterflies, that varied systematically across eight binary features. In this setup, two underlying category prototypes differed from each other by all eight opposing features, while individual exemplars were designated as one, two, or three features away from their corresponding prototype. During an initial training stage, participants completed 240 trials structured into five blocks of 48 trials, learning through trial-by-trial accuracy feedback to classify eight specific training exemplars that were each two features away from their prototypes. Critically, the idealized prototypes themselves were never shown during the training phase, requiring learners to infer the underlying category structure solely from encounters with individual instances.
Following training, participants completed a generalization phase inside a 3-Tesla magnetic resonance imaging scanner, where they classified 34 stimuli per run across two separate functional runs without any performance feedback. The generalization test stimuli included familiar exemplars from the training phase, the previously unseen category prototypes, and 24 novel exemplars varying across different feature distances. To evaluate how decisions were generated, the authors implemented two formal similarity-based computational models coded in MATLAB: a prototype model, which calculates choice probabilities based on similarity to an abstract central summary representation, and an exemplar model, which calculates probabilities based on summed similarity to the specific training items stored in memory. Participant-specific null distributions generated through 10,000 Monte Carlo response shuffles were then utilized to categorize individual strategy profiles into prototype-dominant, exemplar-dominant, comparable-fit, or random categories.
Behavioral results indicated robust learning and clear evidence of cognitive abstraction across the adolescent cohort. Training accuracy exhibited a significant linear increase across blocks, with final-block correct responses averaging 80.1 percent at the first timepoint and 84.0 percent at the nine-month follow-up. During generalization, mean accuracy on novel items reached 74.2 percent at timepoint one and 78.0 percent at timepoint two. Crucially, participants exhibited a pronounced typicality gradient, showing the highest categorization accuracy for the never-before-seen prototypes and a steady linear decline in accuracy as stimuli became more distant in feature space. Furthermore, accuracy did not differ significantly between familiar training exemplars and novel exemplars matched at two feature distances away, indicating that memory for specific familiar instances provided no measurable categorization advantage over abstract structural knowledge.
Formal model comparisons confirmed the predominance of abstract representations over exemplar memorization at both the group and individual levels. Hierarchical linear modeling revealed that the prototype model provided a significantly better fit to participant choices than the exemplar model at both testing sessions without significant interaction across time. At the individual level, 82 percent of participants evaluated at the initial assessment and 71 percent at the nine-month follow-up were classified as prototype dominant. Among the 39 participants who successfully completed testing across both longitudinal timepoints, strategy dominance remained stable in 69 percent of individuals who maintained a consistent prototype strategy, while only seven participants shifted from a prototype preference to an exemplar preference, supporting the view that prototype-based abstraction represents a durable cognitive trait under supportive task conditions.
Neuroimaging analyses pinpointed distinct neural circuits supporting these computational strategies across the brain. First-level general linear models incorporated trial-by-trial parametric regressors derived from the computational models to identify blood-oxygen-level-dependent signal fluctuations tracking prototype and exemplar similarity. Region-of-interest analyses focused on bilateral ventromedial prefrontal cortex and hippocampal divisions revealed that prototype model fits were significant in the ventromedial prefrontal cortex, showing significantly stronger alignment with prototype predictions than exemplar predictions. In contrast, prototype tracking in the anterior hippocampus exhibited only a weak trend, and exemplar tracking in the posterior hippocampus did not reach statistical significance, showing that medial prefrontal cortex serves as a primary site for abstract concept integration in this cohort.
Voxel-wise whole-brain analyses using random field theory correction further illuminated the broader cortical network engaged during prototype-based generalization. Significant prototype-related activation clusters emerged bilaterally in the inferior parietal lobule, extending from the supramarginal gyrus into the angular gyrus, as well as in the right lateral occipital cortex and the right frontal pole. The right frontal pole was the sole region displaying a significant main effect of timepoint, showing increased prototype-related activation at the second visit, which suggests that familiarity with the task format may enhance recruitment of anterior prefrontal control processes. In contrast to the widespread prototype network, significant exemplar tracking across the whole brain was restricted to a single anatomical cluster situated in the bilateral cuneus cortex, mirroring the behavioral dominance of abstract representations.
The study established a direct link between abstract category generalization measured in the laboratory and real-world functional capacity in everyday life. Across participants, higher mean generalization accuracy was significantly associated with higher scores on the parent-reported Adaptive Behavior Assessment System Global Adaptive Composite, confirming that the capacity to extend learned structural knowledge to unfamiliar stimuli relates directly to practical daily independence. While this overall behavioral association did not differ between participants with high versus moderate-to-low autistic traits, neural regressions revealed a significant moderating effect of trait severity in the left inferior parietal lobule. Specifically, greater prototype-related neural activation in the left inferior parietal lobule predicted significantly higher adaptive functioning among adolescents with high parent-reported autistic traits on the Social Responsiveness Scale, whereas no such relationship emerged in youth with lower trait scores.
These experimental observations provide important constraints on perceptual deficit theories of autism, such as weak central coherence, enhanced perceptual functioning, and the enhanced discrimination hypothesis. Although those frameworks suggest that heightened sensory acuity and local feature prioritization impede the synthesis of global summaries, the current findings demonstrate that autistic adolescents readily extract commonalities across discrete experiences when provided with structured feedback and clear feature dimensions. The researchers noted that while a small minority of participants failed to acquire category structure during initial training and another subset generalized without a consistent representational reliance, the predominant profile was one of intact, stable cognitive abstraction, challenging the assumption that concept formation in autism is fundamentally constrained by hyper-specific representations.
The authors identified several methodological boundaries that contextualize the generalizability of their conclusions. The study sample was restricted to autistic adolescents with Full Scale IQ scores of 80 or above and did not include a typically developing control group, leaving open the question of whether identical neural tracking and strategy proportions characterize adolescent development more broadly. Additionally, the experimental task utilized structured cartoon stimuli with well-defined binary feature dimensions and feedback-based training, experimental parameters known to encourage prototype reliance in neurotypical adults, meaning that tasks involving continuous perceptual distortions such as dot patterns might yield different representational distributions. Finally, psychotropic medication usage was highly prevalent, with 78 percent taking at least one psychotropic agent, although stimulants were withheld for 24 hours prior to scanning, and data retention requirements for computational fMRI restricted the longitudinal neuroimaging sample size.
Subject of Research: Neural and computational mechanisms of concept learning and generalization in autistic adolescents
Article Title: From cognitive abstraction to adaptive behavior: neural bases of concept learning in autistic adolescents
Article References: Chen, Y., Hawkins, B., Puckett, H., Sharp, K., Lopez, A., Zeithamova, D., Xie, H., Verbalis, A., VanMeter, A. S., Gaillard, W. D., Kenworthy, L., & Vaidya, C. J. (2026). From cognitive abstraction to adaptive behavior: neural bases of concept learning in autistic adolescents. Molecular Autism, 17(1), Article 26. https://doi.org/10.1186/s13229-026-00721-4
Image Credits: AI Generated
DOI: 10.1186/s13229-026-00721-4
Keywords: autism, concept learning, cognitive abstraction, prototype model, exemplar model, computational modeling, fMRI, generalization, adaptive functioning, inferior parietal lobule, ventromedial prefrontal cortex, adolescent development, neuroimaging, category learning, visual perception, ABAS-GAC, Social Responsiveness Scale
News Source: Colin Clarke. (November 19, 2026). Autistic Adolescents Rely on Abstract Prototypes During Category Learning, Study Finds. Scienmag.



