Self-referential judgments from the same personality trait scales show increased representational similarity in mPFC Authors: Keise Izuma, Ayahito Ito, Kazuki Yoshida, Ryuta Aoki Published: 2025 ABSTRACT Self-report questionnaires are widely used across psychology and related disciplines, yet the cognitive and neural processes underlying how individuals generate responses to such items remain poorly understood. Here, we investigated whether items from the same psychological scale evoke similar neural activation patterns in the medial prefrontal cortex (mPFC), a region consistently implicated in self-referential processing. While undergoing functional magnetic resonance imaging (fMRI), 32 participants completed a self-reference task in which they judged how well 72 personality-related questionnaire items (e.g., from the Big Five, emotion regulation, and well-being scales) described themselves. Using representational similarity analysis (RSA), we found that items from the same scale elicited more similar multivoxel activation patterns in the mPFC compared to items from different scales. This effect was specific to the self-reference task and was not observed during a semantic judgment control task using the same items. Furthermore, the mPFC encoded not only categorical scale membership but also a small but consistent graded component of psychological similarity among scales, as reflected in inter-scale behavioral correlations. Importantly, these effects remained significant even after controlling for sentence-level semantic similarity using multiple regression RSA, indicating that the observed neural structure reflects psychological rather than linguistic similarity. These findings suggest that the mPFC integrates internally constructed evidence in a construct-sensitive manner during self-report. They also provide insight into how psychological assessment corresponds to neural representation. Introduction The ability to report one’s own inner thoughts and feelings is a hallmark of the human species. Across the social sciences, self-report has long served as a foundational research tool, offering direct access to individuals’ internal experiences and shaping our understanding of human psychology and behavior. Beyond psychology, self-report remains a primary method of data collection in a wide range of disciplines, including medicine, education, political science, economics, public health, and marketing. Its widespread use reflects its unique capacity to capture subjective experiences that are otherwise inaccessible through external observation alone1–3. Although the psychological processes underlying self-report have been a central topic in psychology2, the neural mechanisms supporting this process remain poorly understood. Social and cognitive neuroscience studies have employed a wide range of tasks in which participants report their attitudes, preferences, feelings, and beliefs (often using Likert scales) during brain imaging. Yet these findings have not been synthesized to clarify the neural mechanisms of self-report, and as a result, our understanding of how the brain generates responses to questionnaire items remains limited. One widely used approach in social neuroscience to investigate self-referential processing is the self-reference task4,5, in which individuals are presented with personality trait adjectives and asked to judge whether each adjective describes them—a format similar to personality questionnaires. This task has consistently been shown to activate a network of brain regions, most notably the medial prefrontal cortex (mPFC) and posterior cingulate cortex (PCC)6,7. These regions are recognized as core hubs of the brain’s default mode network8 and are commonly engaged in tasks that require judgment based on internally constructed representations, such as theory of mind, autobiographical memory, spatial navigation, moral reasoning, and self-related evaluation9–14. Importantly, answering items on personality questionnaires can be seen as another instance of this broader class of tasks, insofar as it requires evaluating one’s own dispositions based on internally constructed representations. Our previous study15 demonstrated that, during the self-reference task, individuals—consciously or unconsciously—engaged in multiple cognitive processes to gather the information necessary to evaluate whether a personality trait described them. This multifaceted evaluative process was reflected in distinctive activation patterns within the mPFC, suggesting that the mPFC plays a central role in integrating internally constructed representations to support self-referential judgments. Building on this finding, the present study aims to further investigate the integrative function of the mPFC by examining whether its activation patterns during self-report questionnaire responding reflect the psychological constructs measured by each item. While undergoing functional magnetic resonance imaging (fMRI), participants completed several psychological trait scales. As a control task, participants also rated the social desirability of the same questionnaire items, allowing us to distinguish neural activity associated with self-referential evaluation from that involved in more general semantic or normative judgments. Whereas the self-reference task required participants to retrieve and integrate personally relevant information to evaluate whether each item described them, the desirability judgment task emphasized reasoning based on socially shared norms and expectations (i.e., externally guided decision-making; see ref. 9). We focused on the Big Five personality dimensions as the primary experimental stimuli because they provide a theoretically grounded and empirically validated framework for describing stable individual differences in personality. Their breadth and centrality within personality research make them a strong basis for investigating how psychological constructs are represented in the brain. The Big Five framework is among the most widely used and empirically supported models of personality, offering a parsimonious yet comprehensive taxonomy of trait-like individual differences16,17. Its enduring influence is evident not only in its widespread use in psychological assessment, but also in its integration into related fields such as behavioral genetics18,19, developmental psychology20,21, and organizational behavior22,23. Recent work further supports the Big Five’s utility as an organizing framework for psychological trait constructs: Bainbridge et al.24, for example, demonstrated that many commonly used stand-alone trait measures—including constructs such as self-esteem, grit, and empathy—can be meaningfully located within the Big Five space, often correlating as strongly with Big Five domains as with their originally associated scales. This supports the use of Big Five items for probing the neural representation of psychological traits, as they provide both theoretical clarity and empirical breadth. Accordingly, by focusing on the Big Five, the present study leverages a widely accepted framework to examine how psychological constructs are instantiated in brain activation patterns during self-referential evaluation. To analyze the neural data, we applied representational similarity analysis (RSA)25, a multivariate approach that tests whether local patterns of neural activity reflect specific psychological structures. In the present study, we asked whether multivoxel patterns in the mPFC reflect item-level similarity within or across Big Five trait domains. By comparing these neural patterns to model representational similarity structures—defined theoretically or derived from behavioral data—we sought to clarify the type of information encoded in the mPFC during questionnaire responding. RSA and related multivariate approaches have been applied to perceptual, semantic, and affective domains, including object categories25, words and sentences26,27, and emotional states28. However, to our knowledge, no study has applied RSA to items from established psychometric questionnaires. The present study therefore examines whether items from the same psychological construct evoke similar activation patterns in the mPFC, thereby linking psychometric structure with neural representation. Tourangeau et al.2 proposed an influential model of the cognitive processes involved in self-report, identifying four key stages: comprehension, retrieval, judgment, and response. First, respondents must understand the question (comprehension stage). Second, they must recall relevant information from memory (retrieval stage), such as episodic memories (e.g., “When was the last time I felt anxious?”) or more abstract self-knowledge (e.g., “Am I generally an anxious person?”). Third, once information is retrieved, respondents must evaluate it and formulate a response (judgment stage), such as summarizing across multiple memories or integrating conflicting information. Finally, they must map this internal judgment onto the response options provided (e.g., a Likert scale) (response stage). Whereas the first and last stages are shared across the self-reference and control (desirability judgment) tasks, the retrieval and judgment stages in the self-reference task uniquely involve constructing and evaluating internally generated representations. Accordingly, we hypothesized that to the extent that well-constructed psychological scales prompt individuals to access similar kinds of internal information, items assessing the same trait dimension would evoke similar multivoxel activation patterns in the mPFC, reflecting a shared evidence integration process. Methods Participants We aimed to recruit at least 30 participants based on our previous studies15,29 that employed similar self-referential tasks and multivariate analysis approaches. The final sample consisted of 32 participants (14 female, 18 male; age 18–26 years, M = 22.09, SD = 1.71). Participants’ sex was determined based on self-report. No information on race or ethnicity was collected. A total of 38 right-handed Japanese individuals were initially recruited through the university participant pool at Hokkaido University. Data from six participants were excluded due to excessive head motion (>3 mm; one participant), scanner malfunction (four participants), and low reliability in self-reference or desirability ratings (one participant). For the scanner malfunction cases, the EPI images showed severe signal loss, most likely due to poor contact of the radiofrequency coil, especially in the ventral mPFC, and could not be successfully normalized for group-level analysis. For the participant with low rating reliability (r = 0.46 across repetitions, well below the sample average r = 0.78), exclusion followed the same criterion as our previous study29. All participants reported no history of psychiatric disorders. Written informed consent was obtained from all participants prior to the fMRI session. The study was approved by the ethics committee of Kochi University of Technology as part of a joint research project between Hokkaido University and Kochi University of Technology. Participants received 4000 Japanese yen as compensation. Preregistration This study was not preregistered. Experimental procedure Experimental stimuli Experimental stimuli consisted of 72 questionnaire items in total: 60 items from Big Five Inventory-2 (BFI-2)17, six items from the Psychological Well-Being (PWB): Purpose in Life scale30, and six items from the Emotion Regulation (ER): Reappraisal scale31. The five dimensions of the BFI-2 are known to exhibit low intercorrelations, and the two additional scales were included based on prior findings showing that they are not highly correlated with any of the Big Five dimensions24. All scales had been previously translated into Japanese and validated with Japanese samples32–34. While the original PWB: Purpose in Life scale includes seven items, one item (“Done all there is to do in life”) was excluded, as it showed divergent response patterns from the rest of the scale in a Japanese sample, resulting in a lower Cronbach’s alpha in prior research32. fMRI experiment Prior to the fMRI scan, participants received instructions regarding MRI safety and the tasks they would perform inside the scanner. During the fMRI session, participants were presented with one questionnaire item and a 4-point scale on each trial and completed two tasks: self-reference and semantic judgment (Fig. 1; see details below). In the self-reference task, participants were asked to judge how much each item described them. In the semantic judgment task, they rated how desirable each item was. For both tasks, responses were made using a 4-point scale (self-reference: not at all—very much; semantic judgment: very undesirable—very desirable).Fig. 1fMRI tasks.a Self-reference task. b Semantic judgment task. fMRI tasks. a Self-reference task. b Semantic judgment task. Each item remained on the screen until the participant responded, with a maximum response window of 10 s. If no response was made within this time limit, a warning message (“Respond faster”) was displayed for 2 s. Trials were separated by a jittered inter-trial interval (ITI) ranging from 3 to 7 s (mean = 5 s). Participants provided their responses using an MRI-compatible button box with four buttons. To prevent systematic associations between specific fingers and response values, the mapping between fingers and rating options was reversed after the fourth run (of eight total runs). Each of the 72 questionnaire items was presented four times in total—twice for each of the two tasks—so that participants made both self-reference and semantic ratings twice per item. This repetition was implemented to increase the signal-to-noise ratio35. However, to prevent participants from simply recalling and repeating their previous responses without thoughtful engagement, we instructed them that inconsistencies across repetitions were acceptable and asked them to respond as if they were encountering each item for the first time on every trial. Although item repetitions raise the possibility of memory-based responding, such effects would be expected to reduce differentiation between items from different scales (and between tasks) and thus bias against, rather than generate, the predicted scale-sensitive similarity effects specific to the self-reference task. Participants completed four runs of each task. Each run consisted of 36 trials, with one questionnaire item presented per trial, and the trial order was randomly determined within each run. The same set of 36 items was presented across both tasks and in each run. Task order was counterbalanced across participants using one of two fixed sequences: ABBAABBA or BAABBAAB. The experimental tasks were programmed using Psychtoolbox (http://psychtoolbox.org/) with MATLAB software (version 2018a, http://www.mathworks.co.uk). Behavioral data analysis To assess response reliability, we computed the correlation between the two repetitions of each item within each task for every participant. The Fisher z-transformed correlation coefficients were then compared between the two tasks using a paired t-test. To examine self-enhancement tendencies—that is, whether more desirable items were more likely to be endorsed as self-descriptive—we calculated, for each participant, the Pearson correlation between self-descriptiveness and desirability ratings (averaged across repetitions). The Fisher z-transformed correlation coefficients were then submitted to a one-sample t-test against zero. All t-tests assumed normally distributed data, but this was not formally tested. Two-tailed p-values are reported throughout. For each t-test, we also report the effect size (Cohen’s d) and its 95% confidence interval (CI). fMRI data acquisition We acquired images using a 3-T Siemens MAGNETOM Prisma scanner equipped with a 20-channel head coil. For functional imaging, interleaved T2*-weighted gradient-echo echo planar imaging (EPI) sequences were used to produce 42 contiguous 3.0-mm-thick transversal slices that cover nearly the entire cerebrum (TR = 2500 ms; TE = 30 ms; flip angle = 90°; field of view = 240 mm; 80 × 80 matrix; voxel size = 3.0 × 3.0 × 3.0 mm). High-resolution anatomical T1-weighted images were also acquired for each participant (TR = 2300 ms; TE = 2.32 ms; flip angle = 8°; field of view = 240 mm; 256 × 256 matrix; voxel size = 1.0 × 1.0 × 1.0 mm). fMRI data preprocessing We carried out preprocessing and statistical analysis of the fMRI data using SPM12 (Wellcome Department of Imaging Neuroscience), implemented in MATLAB (MathWorks). The first four volumes were discarded before preprocessing and data analyses to allow for T1 equilibration. We conducted preprocessing of the fMRI data with SPM 12’s preproc_fmri.m script starting with realignment of all functional images to a common image. We spatially realigned all images within each run to the first volume of the run using 7th-degree B-spline interpolation and unwarped and corrected for motion artifacts. The T1-weighted structural image was segmented and normalized into a common stereotactic space (MNI atlas). Subsequently, the normalization parameters were applied to the functional images and resampled them to 3 × 3 × 3 mm3 isotropic voxels (i.e., original voxel size was retained) using 7th-degree B-spline interpolation. Following the normalization, the data were spatially smoothed (with a Gaussian kernel of 8 mm FWHM) for the univariate analysis. To preserve fine-grained activation patterns, smoothing was not applied prior to the first-level data analysis for the RSA. Instead, smoothing was applied to the RSA-derived maps using a 4 mm FWHM Gaussian kernel, to account for individual variability in brain structure. fMRI data analysis Univariate analysis We used two general linear models (GLMs) to analyze the fMRI data. For the univariate analysis, which aimed to replicate the robust activation of the mPFC during self-referential judgments compared to semantic judgments, we modeled all trials from each condition (self-reference and semantic judgment) with separate regressors. The six head motion parameters were included as nuisance regressors. A contrast image comparing the self-reference and semantic judgment conditions was then created for each participant and submitted to a second-level group analysis (i.e., one-sample t-test across participants). For the RSA, we estimated a separate first-level GLM in which each questionnaire item was modeled individually (72 items × 2 repetitions × 2 tasks = 288 regressors across 8 runs). Trials with missing responses were modeled as a regressor of no interest, and the six head motion parameters were again included as nuisance regressors. Contrast and t-maps were created for each item in each condition, resulting in a total of 144 maps (72 per task). Importantly, although each repetition was modeled separately at the first level, for the contrasts used in RSA we combined the two repetitions of each item (assigning a weight of 1 to both repetitions). Thus, each t-map reflected the average activation pattern across repetitions for that item. These t-maps were then used in the subsequent RSA analyses. Within the mPFC mask (see below), we set statistical threshold at p < 0.005 (uncorrected), with a cluster-level threshold of p < 0.05 (FWE corrected). We also conducted a whole brain analysis using the same thresholds without the mask; a height threshold at p < 0.005 (uncorrected), with a cluster-level threshold of p < 0.05 (FWE corrected). Data distribution was assumed to be normal, but this was not formally tested. Representational similarity analysis (RSA) Neural representational similarity matrix (RSM) To test whether questionnaire items from the same psychological scale evoke similar activation patterns, we performed RSA using a searchlight approach36. Local patterns of neural activity were extracted from spherical searchlights with a 3-voxel radius, each consisting of up to 123 voxels (fewer near the edges of the brain). For each searchlight, we computed voxel-wise correlations between all possible pairs of the 72 items, resulting in a 72 × 72 neural RSM (Fig. 2a). Each cell of the neural RSM represented the correlation in neural activation between a given pair of items within that specific searchlight. We constructed separate neural RSMs using data from the self-reference and semantic judgment tasks, enabling direct comparison of representational structures across conditions.Fig. 2Schematic illustrations of representational similarity analysis (RSA).a For each participant, a neural representational similarity matrix (RSM) was constructed by computing Pearson correlations between multivoxel activation patterns for all possible pairs of questionnaire items. Separate neural RSMs were generated for the self-reference and semantic judgment tasks. b The all-item model included all 72 items from the seven psychological scales. Magenta indicates item pairs coded as similar (1), cyan as dissimilar (0), and black cells were excluded from analysis. c The Big Five model included only the 60 items from the Big Five Inventory. d The ER & PWB model included the 12 items from the Emotion Regulation (ER) and Psychological Well-Being (PWB) scales. e The inter-scale correlation model included all 72 items. Each cell value reflects the absolute value of the correlation coefficient between the two scales to which the items belong (see Fig. 3). f The inter-scale correlation model–excl was identical to the inter-scale correlation model, except that cells corresponding to item pairs from the same scale (i.e., the magenta triangle regions in panel e) were excluded from analysis. g Neural and model RSMs in the multiple regression RSA. Semantic similarity model shown here is based on Sentence-BERT. E Extraversion, A Agreeableness, C Conscientiousness, N Neuroticism, O Openness, ER Emotion Regulation, PWB Psychological Well-Being. Schematic illustrations of representational similarity analysis (RSA). a For each participant, a neural representational similarity matrix (RSM) was constructed by computing Pearson correlations between multivoxel activation patterns for all possible pairs of questionnaire items. Separate neural RSMs were generated for the self-reference and semantic judgment tasks. b The all-item model included all 72 items from the seven psychological scales. Magenta indicates item pairs coded as similar (1), cyan as dissimilar (0), and black cells were excluded from analysis. c The Big Five model included only the 60 items from the Big Five Inventory. d The ER & PWB model included the 12 items from the Emotion Regulation (ER) and Psychological Well-Being (PWB) scales. e The inter-scale correlation model included all 72 items. Each cell value reflects the absolute value of the correlation coefficient between the two scales to which the items belong (see Fig. 3). f The inter-scale correlation model–excl was identical to the inter-scale correlation model, except that cells corresponding to item pairs from the same scale (i.e., the magenta triangle regions in panel e) were excluded from analysis. g Neural and model RSMs in the multiple regression RSA. Semantic similarity model shown here is based on Sentence-BERT. E Extraversion, A Agreeableness, C Conscientiousness, N Neuroticism, O Openness, ER Emotion Regulation, PWB Psychological Well-Being. Model RSM We tested a total of five model RSMs (Fig. 2b–f); (1) All-item model, (2) Big Five model, (3) ER & PWB model, (4) inter-scale correlation model, and (5) inter-scale correlation model-excl. The all-item model (Fig. 2b) included all 72 items and tested whether neural responses were more similar for item pairs from the same psychological scale than for those from different scales. The Big Five model (Fig. 2c) was identical in structure but included only the 60 items from the Big Five Inventory, isolating effects specific to those dimensions. The ER & PWB model (Fig. 2d) included only the 12 items from the Emotion Regulation and Psychological Well-Being scales to examine whether effects observed in the Big Five model were also present in other psychological constructs. The inter-scale correlation model (Fig. 2e) tested whether neural similarity reflected not only category membership (same vs. different scale) but also the graded degree of behavioral correlation between scales—i.e., whether items from scales that are more strongly correlated across individuals elicited more similar activation patterns in the mPFC. To construct this model, we first computed mean ratings for each of the seven scales (five Big Five dimensions, ER, and PWB) for every participant. We then calculated Pearson correlations across participants between each pair of scales, yielding a 7 × 7 inter-scale correlation matrix (reported in Fig. 3). This matrix was expanded to the item level by assigning each pair of items the absolute correlation value of their respective scales, resulting in a 72 × 72 model RSM. The inter-scale correlation model–excl (Fig. 2f) was identical to the inter-scale correlation model, except that it excluded all within-scale item pairs (i.e., the cells along the block diagonals of the RSM), allowing us to assess whether between-scale correlations drove the observed neural similarity patterns.Fig. 3Correlations across the seven scales (n = 32).***p < 0.001, **p < 0.01, * p < 0.05 (two-tailed, uncorrected for multiple comparisons). Correlations across the seven scales (n = 32). ***p < 0.001, **p < 0.01, * p < 0.05 (two-tailed, uncorrected for multiple comparisons). The fit between the neural RSM and each model RSM was evaluated by using Kendall’s tau-a for each participant37. We repeated this analysis for every searchlight across the whole brain, resulting in a correlation map for each of the two tasks. Group analysis As mentioned above, we applied smoothing before the group analysis of the RSA outputs (with a Gaussian kernel of 4-mm FWHM). Following the smoothing, the correlation maps were submitted to the second-level analysis (i.e., one sample t-tests across participants). We applied an mPFC mask to the analysis to restrict the analysis to voxels within the a priori mPFC ROI. We created the mask in the same way as in our previous studies29 using the WFU PickAtlas toolbox for SPM38. The mPFC ROI mask included Frontal_Sup_Medial_L, Frontal_Sup_Medial_R, Frontal_Mid_Orb_L, Frontal_Mid_Orb_R, Rectus_L, Rectus_R, Cingulum_Ant_L, and Cingulum_Ant_R (dilation factor = 2), which we took from the Anatomical Automatic Labeling (AAL) masks implemented in the WFU pickatlas toolbox. Within the mPFC mask, we set statistical threshold at p < 0.005 (uncorrected), with a cluster-level threshold of p < 0.05 (FWE corrected). We also conducted the whole brain analysis using the same thresholds without the mPFC mask; a height threshold at p < 0.005 (uncorrected), with a cluster-level threshold of p < 0.05 (FWE corrected). ROI analysis RSA To further investigate the role of the mPFC, we ran a ROI analysis. We defined our mPFC ROI in the same way as in our previous studies15,29, using Neurosynth39 (https://neurosynth.org/). We downloaded an association map (thresholded at q < 0.01, False Discovery Rate–corrected) which was generated from a term-based meta-analysis with the term “self-referential” (downloaded October, 2023). The mPFC ROI included 308 voxels. The RSA was conducted within this ROI. For each participant, we computed the fit between the neural RSM and each of the five model RSMs (Fig. 2b–f) within the mPFC ROI. Statistical significance was assessed using a permutation test in which the order of the 72 questionnaire items was randomly shuffled 1000 times to generate a null distribution. We controlled for multiple comparisons across the five models using the Benjamini–Hochberg false discovery rate (FDR; q = 0.05) and report both uncorrected (pperm) and FDR-corrected p-values (pFDR). To assess differences in model performance, we compared the fit of the all-item model and the inter-scale correlation model to the neural representational similarity matrices (RSMs). For each participant, model fit was quantified as Kendall’s tau correlation between the model RSM and the neural RSM. Subject-wise differences in model fit (all-item minus inter-scale) were then analyzed using the Bayesian bootstrap40,41. The Bayesian bootstrap generates posterior distributions over summary statistics by resampling the observed data with random Dirichlet(1,…,1) weights, thereby approximating Bayesian inference without parametric assumptions. We drew 50,000 bootstrap samples of the mean difference to characterize its posterior distribution. From this posterior we report (i) the posterior mean and median of the difference, (ii) a 95% credible interval (CrI), and (iii) the posterior probability that the difference is greater than zero, P(μ > 0), which quantifies the probability that the all-item model outperforms the inter-scale correlation model. Multiple regression RSA To test whether our RSA results held after controlling for semantic similarity among scale items, we conducted a multiple regression RSA in which the mPFC neural RSM derived from activation patterns of the mPFC ROI served as the dependent variable (see Fig. 2g). The model included two predictors: the inter-scale correlation model (Fig. 2e) and the semantic similarity model, both of which were z-scored prior to analysis. Semantic similarity was first estimated using Sentence-BERT42, a transformer-based model that generates sentence embeddings such that semantically similar sentences occupy nearby positions in embedding space. For the Japanese items, we used a pre-trained Japanese BERT model (cl-tohoku/bert-base-japanese-whole-word-masking) implemented within the Sentence-BERT framework. We computed cosine similarity between all possible pairs of questionnaire items and used the resulting 72 × 72 matrix as the semantic similarity model RSM (see Fig. 2g). To ensure that our results were not dependent on the embedding method, we repeated the analysis using Google’s Universal Sentence Encoder (USE)43, a deep averaging network that encodes sentences into fixed-length vectors optimized for semantic similarity and transfer learning tasks. Specifically, we used the multilingual version of USE, which supports over a dozen languages, including Japanese. The USE-based semantic similarity RSM was substituted into the regression analysis in place of the Sentence-BERT model (Fig. 2g). The two semantic similarity RSMs were moderately correlated with one another (r(2554) = 0.34, p < 0.001), suggesting that they captured overlapping but non-identical semantic structure. Statistical significance was assessed using a permutation test, as described above. Variance Partitioning Analysis (VPA) Following the multiple regression RSA, we conducted a variance partitioning analysis (VPA) to estimate the unique and shared variance explained by different predictor models. Specifically, we ran three regressions: one including both predictors (e.g., inter-scale correlation model and Sentence-BERT semantic similarity model), and two with each predictor entered individually. By comparing the variance explained (R²) across these models, we decomposed the total explained variance into (i) unique variance attributable to each predictor and (ii) variance shared between predictors. Statistical significance of each variance component was assessed using a permutation procedure. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Results Behavioral results Participants responded within the 10-s time limit on nearly all trials: 99.93% for the self-reference task and 99.87% for the semantic judgment task. The average reaction time (RT) was 2.89 s (SD = 0.69) for the self-reference task and 2.85 s (SD = 0.74) for the semantic judgment task. This difference was not statistically significant (t(31) = 0.93, p = 0.36, d = 0.06, 95% CI [–0.07, 0.19]). Participants rated each of the 72 items on self-descriptiveness and desirability twice each during the fMRI session. Their responses were highly consistent across the two repetitions. The average within-individual correlation for the self-reference task was r = 0.79 (SD = 0.11), and for the semantic judgment task, r = 0.87 (SD = 0.07). Response consistency was significantly higher in the semantic judgment task than in the self-reference task (paired t-test on Fisher z-transformed values: t(31) = –6.01, p < 0.001, d = –0.90, 95% CI [–1.25, –0.55]). We further examined whether participants exhibited a typical self-enhancement tendency15,44—that is, whether more desirable items were more likely to be endorsed as self-descriptive. For each participant, we calculated the Pearson correlation between self-descriptiveness and desirability ratings (using the average of the two ratings per item). A one-sample t-test on Fisher z-transformed correlation coefficients revealed that the average correlation (r = 0.38, SD = 0.36) was significantly positive (t(31) = 5.37, p < 0.001, d = 0.95, 95% CI [0.52, 1.36]), indicating a reliable self-enhancement effect. Using the average of the two ratings for each item, we computed Cronbach’s α for each of the seven psychological scales. All scales demonstrated good internal consistency (Cronbach’s α > 0.80), as shown in Table 1. Correlations among the seven scales are presented in Fig. 3.Table 1Cronbach’s α for the seven scalesScale# of itemsCronbach’s αBig Five: Extraversion120.94Big Five: Agreeableness120.81Big Five: Conscientiousness120.86Big Five: Neuroticism120.90Big Five: Openness120.91ER: Reappraisal60.92PWB: Purpose in Life60.85ER Emotion Regulation, PWB Psychological Well-Being. Table 1Cronbach’s α for the seven scalesScale# of itemsCronbach’s αBig Five: Extraversion120.94Big Five: Agreeableness120.81Big Five: Conscientiousness120.86Big Five: Neuroticism120.90Big Five: Openness120.91ER: Reappraisal60.92PWB: Purpose in Life60.85ER Emotion Regulation, PWB Psychological Well-Being. Cronbach’s α for the seven scales ER Emotion Regulation, PWB Psychological Well-Being. Finally, we examined the similarity of self-descriptiveness and semantic ratings across all participant pairs for the 72 questionnaire items (see Fig. 4). Correlations for the semantic ratings were uniformly positive, ranging from r = 0.57 to r = 0.93, indicating that participants rated the desirability of items in a similar way—that is, they shared a common understanding of social norms (Fig. 4b). In contrast, correlations for self-reference ratings showed greater variability, ranging from r = –0.48 to r = 0.69, suggesting that participants differed in how they judged the self-descriptiveness of each item—that is, they had distinct personality profiles (Fig. 4a). On average, inter-participant similarity in self-descriptiveness ratings was significantly lower than that in semantic ratings (t(495) = –92.21, p < 0.001, d = –4.65, 95% CI [–4.99, –4.31]), indicating that how participants responded to each item during the self-reference task was more idiosyncratic than the semantic judgment task.Fig. 4Inter-participant similarity in item-level ratings.a Similarity in self-descriptiveness judgment. b Similarity in semantic judgment. Inter-participant similarity in item-level ratings. a Similarity in self-descriptiveness judgment. b Similarity in semantic judgment. fMRI results Univariate analysis results Replicating previous findings6,7,45,46, the contrast of self-reference versus semantic judgment significantly activated the medial prefrontal cortex (mPFC) (Fig. 5a; x = –9, y = 47, z = –4; 344 voxels). No other brain regions showed significant activation for this contrast. Notably, prior studies often report additional activation in the posterior cingulate cortex (PCC) during self-referential processing. In the current study, two clusters were observed in the PCC (cluster 1: x –9, y = –58, z = 20, 85 voxels; cluster 2: x = –6, y = –31, z = 44, 32 voxels) at an uncorrected threshold of p < 0.005. However, neither survived the cluster-level threshold of p < 0.05 (FWE-corrected). The reverse contrast (i.e., semantic > self) revealed significant activation in the bilateral ventrolateral prefrontal cortex (VLPFC) (left: x = –42, y = 41, z = –13, 207 voxels; right: x = 42, y = 41, z = –13, 148 voxels).Fig. 5Results of the univariate analysis and RSA (n = 32).a Univariate results showing regions significantly activated by the self > semantic contrast, with arrows indicating the significant cluster in the medial prefrontal cortex (mPFC). b RSA results for the all-item model, identifying regions where neural response similarity significantly depended on whether item pairs were from the same or different scales. c Overlap between univariate and RSA results, with magenta indicating activation from the univariate contrast (self > semantic), cyan indicating RSA results from the all-item model, and white indicating the overlapping region. d Results from a paired t-test comparing searchlight RSA maps from the self-reference and semantic judgment tasks. mPFC neural response similarity reflected item scale membership significantly more during the self-reference task than during the semantic judgment task. In all panels, results are displayed on sagittal (x = 4) and axial (z = 0) slices. Results of the univariate analysis and RSA (n = 32). a Univariate results showing regions significantly activated by the self > semantic contrast, with arrows indicating the significant cluster in the medial prefrontal cortex (mPFC). b RSA results for the all-item model, identifying regions where neural response similarity significantly depended on whether item pairs were from the same or different scales. c Overlap between univariate and RSA results, with magenta indicating activation from the univariate contrast (self > semantic), cyan indicating RSA results from the all-item model, and white indicating the overlapping region. d Results from a paired t-test comparing searchlight RSA maps from the self-reference and semantic judgment tasks. mPFC neural response similarity reflected item scale membership significantly more during the self-reference task than during the semantic judgment task. In all panels, results are displayed on sagittal (x = 4) and axial (z = 0) slices. Searchlight RSA result A searchlight RSA within the mPFC mask identified a significant mPFC cluster associated with the all-item model (Fig. 5b; x = 3, y = 56, z = 17; 660 voxels), indicating that neural response similarity in this region varied depending on whether item pairs originated from the same or different psychological scales. Although this mPFC cluster overlapped with the mPFC area activated by the self > semantic contrast, the RSA-identified cluster tended to be located more dorsally than the univariate activation cluster (see Fig. 5c). Outside of the mPFC mask, no regions showed a significant association with the all-item model. Clusters were also observed in the PCC (x = 0, y = –49, z = 29, 105 voxels) and the left TPJ (x = –45, y = –67, z = 38, 65 voxels) at the uncorrected voxelwise threshold (p < 0.005, no cluster-extent threshold), but none survived cluster-level correction at p < 0.05 (FWE-corrected). Thus, only the mPFC showed reliable scale-dependent pattern similarity. In contrast, when the same searchlight RSA using the all-item model was applied to data from the semantic judgment task, no significant clusters were identified. A direct comparison between the two tasks (self > semantic) revealed a significant cluster in the mPFC (Fig. 5d; x = 0, y = 50, z = 8; 372 voxels). No other significant regions were found. The reverse contrast (semantic > self) did not reveal any significant activation. These findings indicate that mPFC neural response patterns reflected whether items were from the same or different psychological scales during the self-reference task, whereas no significant effects were observed for the semantic task. ROI analysis results RSA We first confirmed that neural response similarity within the mPFC ROI (Fig. 6a) during the self-reference task significantly reflected whether item pairs were from the same psychological scale, as predicted by the all-item model (mean Kendall’s τₐ = 0.005 [SD = 0.008], pperm < 0.001, pFDR = 0.002; Fig. 6b). In contrast, the effect was not significant for the semantic judgment task (mean Kendall’s τₐ = 0.001 [SD = 0.008], pperm = 0.32), and the difference in model fit between the two tasks was itself significant (pperm = 0.008), replicating and extending the results observed in the searchlight analysis (Fig. 5b, d).Fig. 6Results of the ROI analyses (n = 32).a The medial prefrontal cortex (mPFC) ROI used in the analysis (x = –6), defined using a Neurosynth term-based meta-analysis for the term “self-referential.” Results of permutation tests assessing the fit between neural RSMs and five model RSMs: b all-item model, c Big Five model, d PWB & ER model, e inter-scale correlation model, and f inter-scale correlation model–excl. g The inter-scale correlation model showed a slightly better fit to mPFC neural response patterns compared to the all-item model. Although the Wilcoxon signed-rank test did not reach significance (p = 0.121), Bayesian bootstrap estimation indicated ~99% posterior probability that the inter-scale correlation model outperformed the all-item model (posterior mean difference = –0.0059; 95% CrI [–0.0118, –0.0007]). In b–f, red vertical lines represent the mean Kendall’s tau correlation with the actual data. Results of the ROI analyses (n = 32). a The medial prefrontal cortex (mPFC) ROI used in the analysis (x = –6), defined using a Neurosynth term-based meta-analysis for the term “self-referential.” Results of permutation tests assessing the fit between neural RSMs and five model RSMs: b all-item model, c Big Five model, d PWB & ER model, e inter-scale correlation model, and f inter-scale correlation model–excl. g The inter-scale correlation model showed a slightly better fit to mPFC neural response patterns compared to the all-item model. Although the Wilcoxon signed-rank test did not reach significance (p = 0.121), Bayesian bootstrap estimation indicated ~99% posterior probability that the inter-scale correlation model outperformed the all-item model (posterior mean difference = –0.0059; 95% CrI [–0.0118, –0.0007]). In b–f, red vertical lines represent the mean Kendall’s tau correlation with the actual data. We conducted the same analysis separately for the 60 Big Five items (Big Five model; Fig. 2d) and the 12 items from the ER and PWB scales (ER & PWB model; Fig. 2e) to assess whether the observed effect was primarily driven by the Big Five items. For both models, neural response similarity in the mPFC during the self-reference task was significantly associated with whether item pairs belonged to the same scale: the Big Five model (mean Kendall’s τₐ = 0.005 [SD = 0.012], pperm = 0.001, pFDR = 0.002; Fig. 6c), and the ER & PWB model (mean Kendall’s τₐ = 0.027 [SD = 0.081], pperm = 0.020, pFDR = 0.020; Fig. 6d). There was no significant difference in model fit between the two (p = 0.18, two-tailed Wilcoxon signed-rank test), suggesting that scale-dependent mPFC pattern similarity is not unique to the Big Five, but extends to other psychological constructs as well. Thus far, our analyses focused on whether item pairs belonged to the same or different psychological scales. However, as shown in the behavioral results (Fig. 3), the scales were themselves more or less correlated with one another. To test whether mPFC neural response similarity reflected the strength of inter-scale correlations, we constructed a model RSM based on the absolute values of correlation coefficients among the seven scales (inter-scale correlation model; Fig. 2e) and computed its fit with the neural RSM. The results revealed a significant association between the two (mean Kendall’s τₐ = 0.011 [SD = 0.019], pperm < 0.001, pFDR = 0.002; Fig. 6e). To provide stronger evidence that mPFC neural similarity reflects graded psychological similarity, we conducted the same analysis using a modified version of the model that excluded all within-scale item pairs (inter-scale correlation model–excl; Fig. 2f). Even with these within-scale cells removed, the association remained significant (mean Kendall’s τₐ = 0.008 [SD = 0.022], pperm = 0.013, pFDR = 0.016; Fig. 6f). Finaly, to test whether graded inter-scale relationships explained mPFC similarity beyond simple same/different scale membership, we compared the fits of the inter-scale correlation model to the all-item model. A Wilcoxon signed-rank test did not detect a significant difference (z = –1.55, p = 0.121). However, Bayesian bootstrap estimation indicated that the posterior distribution of the mean difference (all-item – inter-scale) was narrowly negative (posterior mean = –0.0059; 95% CrI [–0.0118, –0.0007]), with only 1.3% of posterior mass above zero. This corresponds to ~99% posterior probability that the graded inter-scale model outperforms the categorical model at the population level. Thus, although the classical test was underpowered to reject the null at α = 0.05, Bayesian estimation suggests a small but reliable graded advantage for the inter-scale model. Multiple regression RSA: controlling for semantic similarity across items Some psychological scales have been criticized for including near-synonymous items, which may artificially inflate internal consistency without broadening the construct being measured47. It is therefore unsurprising that two near-synonymous items might elicit similar neural responses. Although the BFI-217 was not explicitly developed to eliminate such redundancy, its item development emphasized clarity and trait breadth, and it appears to avoid obvious near-synonyms within each scale. Still, items measuring the same trait often exhibit semantic similarity, as they are designed to tap into related aspects of a common psychological dimension. Supporting this, a recent study using large language models found that Big Five items within the same trait domain are more semantically similar to one another than to items from other domains48. This raises the question of whether our observed neural similarity might reflect such semantic overlap rather than deeper psychological structure (although a purely semantic account would not explain the mPFC specificity of our effects, nor their restriction to the self-reference task). To directly address this issue, we conducted an additional analysis incorporating sentence-level semantic similarity, estimated using Sentence-BERT42. We constructed a semantic similarity RSM and tested whether the inter-scale correlation model explained mPFC neural similarity above and beyond semantic similarity (Fig. 2g). Because the inter-scale correlation model provided a slightly better fit than the all-item model, we focused on this model here. When we repeated the analyses with the all-item model in place of the inter-scale model, the results were broadly similar in the regression RSA. However, in the VPA, the inter-scale correlation model uniquely accounted for significant variance beyond semantics (see below), whereas the variance uniquely explained by the all-item model was not significant (see Supplementary Results for details). The results first showed that, consistent with the previous study48, within-scale semantic similarity was significantly higher than across-scale semantic similarity (t(2554) = 6.25, p < 0.001, d = 0.36, 95% CI [0.24, 0.47]). Visual inspection of the similarity matrix suggested that semantic similarity for the six ER items was particularly high (see Fig. 2g), but the difference between within- and across-scale similarity remained significant even when considering only Big Five items (t(1768) = 5.24, p < 0.001, d = 0.32, 95% CI [0.20, 0.44]). The results of the multiple regression RSA showed that the inter-scale correlation model remained a significant predictor of mPFC pattern similarity even after controlling for semantic similarity (mean β = 0.002 [SD = 0.004], pperm = 0.001). This suggests that multivoxel mPFC responses during self-referential evaluation reflect the psychological structure of the traits being assessed, rather than merely tracking surface-level linguistic similarity among items. The semantic similarity model (based on Sentence-BERT) was also significantly associated with mPFC neural responses (mean β = 0.004 [SD = 0.007], pperm = 0.003), indicating that semantically similar items evoke similar activation patterns in the mPFC. Because both models were significant, we next ran a VPA. This revealed that the inter-scale correlation and Sentence-BERT models together explained significant variance in mPFC representational similarity (mean total R² = 0.0028 [SD = 0.0035], pperm = 0.020). The shared (common) variance component was highly significant (mean R² = 0.0001 [SD = 0.0003], pperm < 0.001). The variance uniquely explained by the inter-scale correlation model was also significant (mean R² = 0.0007 [SD = 0.0016], pperm = 0.012), whereas the unique contribution of the Sentence-BERT model was not (mean R² = 0.0020 [SD = 0.0031], pperm = 0.10). Although the mean R² for the unique Sentence-BERT component was numerically larger, it did not reach significance because the effect was not consistent across participants. By contrast, the inter-scale correlation model explained less variance on average but did so reliably, yielding a significant unique component. These results indicate that while sentence-level semantic similarity based on Sentence-BERT accounts for some variance in mPFC neural responses, only the inter-scale correlation model captured unique and reliable variance in mPFC patterns beyond linguistic similarity. When we repeated the multiple regression RSA using the semantic similarity model based on USE43, the inter-scale correlation model again showed a significant association with mPFC neural responses (mean β = 0.003 [SD = 0.004], pperm = 0.003). In contrast, unlike the result based on Sentence-BERT, the semantic similarity model derived from USE was not significantly associated with mPFC neural responses (mean β = 0.0011 [SD = 0.007], pperm = 0.16). Note that, in addition to the above analysis controlling for semantic similarity, we also conducted a multiple regression RSA controlling for both semantic similarity and similarity in participants’ ratings. The results remained virtually the same even after controlling for rating similarity (see Supplementary Results). To complement the ROI analyses, we conducted a whole-brain searchlight regression RSA (Fig. 2g) using the inter-scale correlation model and the Sentence-BERT semantic similarity model. The results were similar to those from the searchlight RSA reported above (Fig. 5b, d). For the inter-scale correlation model, the analysis of self-reference data revealed a significant cluster only in the mPFC (x = –3, y = 44, z = 20, 1045 voxels). Although there were clusters in the PCC (x = –3, y = –49, z = 26, 138 voxels), the left TPJ (x = –45, y = −70, z = 38, 60 voxels) and left inferior frontal gyrus (IFG; x = –48, y = 32, z = 2, 80 voxels), these three clusters did not survive the cluster-level threshold of p < 0.05 (FWE-corrected). This mPFC effect remained significant when directly compared with the semantic judgment task (x = –3, y = 56, z = 11, 271 voxels), and no other regions showed significant results. By contrast, the Sentence-BERT model was associated with widespread clusters in several brain regions, including the mPFC, cingulate cortex, the bilateral dorsolateral prefrontal cortex, the left inferior parietal lobule, and extensive regions of occipital cortex during the self-reference task. However, none of these clusters survived the self > semantic comparison, indicating that these effects reflect general semantic similarity across items rather than processes specific to self-referential evaluation. Finally, in an exploratory whole-brain analysis restricted to the 60 Big Five items, we tested whether trait-specific representational models (e.g., Extraversion, Neuroticism) predicted neural similarity. No significant effects were observed, consistent with the interpretation that our findings reflect a general representation of psychological scale membership rather than localized encoding of individual traits (see Supplementary Results and Supplementary Fig. 1). Discussion The present study used RSA to investigate how the human brain—particularly the mPFC—represents psychological traits as assessed through common personality questionnaires. By combining multivariate fMRI analyses with theoretical and behaviorally derived models of psychological structures, we demonstrated that mPFC neural response patterns during self-referential judgments are sensitive to whether items are drawn from the same psychological scale (Fig. 5b, d). No region other than the mPFC showed the same result. Importantly, this effect was observed across both Big Five and other scales (PWB, ER), despite differences in item length and complexity, and was absent in the semantic judgment task, indicating that it cannot be explained by linguistic or perceptual features. Moreover, we found that mPFC similarity patterns were not merely categorical (same vs. different scale) but reflected a small but consistent graded component, as indexed by inter-scale correlations (Fig. 6e). Although the Wilcoxon test did not reach significance, Bayesian bootstrap analysis indicated that the inter-scale model was more likely to provide a slightly better fit than the categorical model, suggesting that mPFC representations are sensitive to the degree of overlap between psychological constructs, not only their categorical boundaries. Regression RSA and variance partitioning further demonstrated that the effect was uniquely accounted for by psychological structure rather than sentence-level semantic similarity. The pattern of results observed in the mPFC likely reflects the underlying psychological similarity in how individuals respond to each scale item. Specifically, each item from a well-constructed psychological scale prompts respondents to access similar aspects of their self-concept—potentially represented in the mPFC in terms of self-importance (see ref. 29)—and/or to retrieve relevant information in a similar manner to form a judgment and select a response (e.g., drawing on similar personal memories and contextual self-knowledge)15. This process, in turn, results in similar activation patterns in the mPFC across items belonging to the same scale. Our previous study15 demonstrated that mPFC neural responses during a self-reference task with trait adjectives reflect multiple cognitive processes, some of which are shared with other tasks involving judgments based on internally constructed representations (e.g., other-reference, introspection, and autobiographical memory tasks). The present study extends this work by showing that even within the same self-reference task, mPFC responses distinguish among items from different psychological scales. Crucially, our multiple regression RSA demonstrated that this scale-sensitive pattern similarity in the mPFC could not be explained by surface-level linguistic similarity alone. Instead, it reflects the underlying psychological structure of the traits being assessed. Thus, the present findings highlight the representational richness of the mPFC in encoding fine-grained aspects of self-relevance. Previous multivariate neuroimaging studies have shown that activation patterns in the mPFC differentiate between self- and other-referential processing49–52. Feng et al.52 further demonstrated that mPFC activation patterns during the self-reference task vary depending on the type of self-knowledge being accessed—for example, traits, physical attributes, or social roles. The present study builds on this work by showing that mPFC patterns not only distinguish between items from distinct psychological constructs within standard personality questionnaires, but also show a small but reliable graded component across scales—suggesting a representational structure that extends modestly beyond categorical differences. The present findings also contribute to a theoretical understanding of self-report not merely as a measurement tool but as a cognitive process in its own right. Our results suggest that responding to a questionnaire item involves more than retrieving a static belief or trait; instead, it engages an active construction of self-relevant evidence, supported by integrative processes in the mPFC. This view aligns with models of judgment that emphasize the dynamic nature of introspection and decision-making2, and extends them by providing neural evidence that different items—depending on their psychological content—invoke distinct yet systematically related patterns of internal processing. From this perspective, self-report can be seen as the behavioral outcome of a representational evaluation process, in which the brain draws upon personal knowledge, situational context, and normative expectations to produce a coherent response. Understanding self-report in these terms has implications for how we interpret questionnaire data—not as direct reflections of latent traits, but as outputs shaped by structured, construct-specific mental operations. Our study also relates to ongoing work examining potential functional distinctions between the mPFC and PCC within the default mode network. Although clusters in the PCC appeared at an uncorrected voxelwise threshold (p < 0.005), they did not survive cluster-level correction. Thus, our results primarily highlight robust involvement of the mPFC in structured psychological evaluation, while no reliable evidence for PCC involvement was observed in the present task. Taken together with our previous study that highlighted the mPFC’s specific integrative role15, the present findings further support the view that the mPFC plays a central role in integrating internally constructed evidence during self-evaluative judgments. A natural question, which remains unanswered in the present study, is whether the observed representational patterns are unique to self-referential judgments or whether they would generalize to other-reference contexts. Our theoretical perspective suggests that these effects are not unique to self-reference, but reflect a broader role of the mPFC in evaluations grounded in internally constructed representations. Importantly, an other-reference task also requires self-report based on internally constructed representations, albeit about another person’s personality. Thus, to the extent that items from the same scale elicit similar retrieval processes when forming a judgment about another individual, we would expect similar representational patterns—potentially weaker than for self—to generalize beyond the self. Future research that directly contrasts self- and other-reference tasks will be valuable to test this prediction. Another important avenue for future research is the use of neuroimaging to assess the validity of psychological scales. Traditional validation relies almost entirely on subjective sources such as expert judgment and participants’ self-reports53, whereas neural measures may offer an additional, more objective lens. For example, during scale development, one could test whether items intended to measure the same construct elicit convergent neural representations (e.g., more similar mPFC patterns), or whether scales that appear highly correlated behaviorally nonetheless show distinct neural patterns, suggesting that they capture different constructs. At present, however, there are no established benchmarks for what constitutes “similar” versus “distinct” neural patterns; just as in psychometrics, such judgments will likely depend on replication, reliability, and convergence across studies. Neural data therefore provide a distinct and complementary source of validity evidence—analogous to convergent and discriminant validity in psychometrics—rather than a replacement for established methods. Moreover, the effect sizes observed here were modest (see Fig. 6), and neuroimaging faces practical constraints of scalability, cost, and interpretability. Within this boundary, our findings highlight how neural measures could enrich construct validation, clarify overlaps and distinctions among scales, and ultimately contribute to resolving foundational challenges in personality research54 and efforts to defragment psychology55. Limitations One limitation of the present design is that our task instructions (‘how much does this describe you?’) provide a more explicit self-referential framing than the standard Likert-scale format typically used in questionnaire administration. This may encourage somewhat greater metacognitive reflection. Nevertheless, we believe that the task captures the core evaluative process underlying self-report—judging whether an item is self-descriptive—while adapting it for an fMRI paradigm. Self-reference tasks with similar instructions are widely employed in social and cognitive neuroscience and have been shown to elicit neural activity that parallels standard self-report judgments4,5. Future work may directly compare neural responses across different response formats to further assess generalizability. As a final note, we point out that the two embedding models we used to estimate semantic similarity, Sentence-BERT and USE, yielded somewhat different results: semantic similarity predicted mPFC responses with Sentence-BERT but not with USE. This divergence likely reflects differences in model architecture and sensitivity, with transformer-based embeddings capturing finer contextual meaning than averaging-based encoders. Crucially, our central finding—that mPFC similarity reflects psychological structure beyond semantic overlap—was consistent across both models. More broadly, current embedding models capture lexical and structural aspects of meaning but fall short of representing pragmatic and culturally specific semantics, or self-relevant connotations, which may be critical for psychological scales. Future work could explore whether newer multilingual transformer models (e.g., XLM-R56, LLaMA57) provide improved estimates of semantic similarity for this purpose. Conclusion In sum, our findings indicate that neural responses in the mPFC reflect the psychological constructs assessed by self-report questionnaires. By applying RSA, we demonstrated that the mPFC encodes categorical similarities among trait constructs and also shows a small but reliable graded component—observed only when judgments draw on internally generated representations (i.e., during the self-reference task). These results shed light on the neural representations that underlie self-report and highlight a promising link between neuroscience and personality assessment. Given the central role of self-report across many areas of research, understanding its neural basis may help strengthen the bridge between subjective experience and objective measurement. Supplementary information Transparent Peer Review file Supplemental Information Reporting summary Transparent Peer Review file Supplemental Information Reporting summary