Neural similarity between mentalizing and live social interaction during the transition to adolescence

Abstract Social interactions are essential for human development, yet little neuroimaging research has examined their underlying neurocognitive mechanisms using socially interactive paradigms during childhood and adolescence. Recent neuroimaging research has revealed activity in the mentalizing network when children engage with a live social partner, even when mentalizing is not required. While this finding suggests that social‐interactive contexts may spontaneously engage mentalizing, it is not a direct test of how similarly the brain responds to these two contexts. The current study used representational similarity analysis on data from 8‐ to 14‐year‐olds who made mental and nonmental judgments about an abstract character and a live interaction partner during fMRI. A within‐subject, 2 (Mental/Nonmental) × 2 (Peer/Character) design enabled us to examine response pattern similarity between conditions, and estimate fit to three conceptual models of how the two contexts relate: (1) social interaction and mentalizing about an abstract character are represented similarly; (2) interactive peers and abstract characters are represented differently regardless of the evaluation type; and (3) mental and nonmental states are represented dissimilarly regardless of target. We found that the temporal poles represent mentalizing and peer interactions similarly (Model 1), suggesting a neurocognitive link between the two in these regions. Much of the rest of the social brain exhibits different representations of interactive peers and abstract characters (Model 2). Our findings highlight the importance of studying social‐cognitive processes using interactive approaches, and the utility of pattern‐based analyses for understanding how social‐cognitive processes relate to each other.


Abstract
Social interactions are essential for human development, yet little neuroimaging research has examined their underlying neurocognitive mechanisms using socially interactive paradigms during childhood and adolescence. Recent neuroimaging research has revealed activity in the mentalizing network when children engage with a live social partner, even when mentalizing is not required. While this finding suggests that social-interactive contexts may spontaneously engage mentalizing, it is not a direct test of how similarly the brain responds to these two contexts. The current study used representational similarity analysis on data from 8-to 14-year-olds who made mental and nonmental judgments about an abstract character and a live interaction partner during fMRI. A within-subject, 2 (Mental/Nonmental) Â 2 (Peer/Character) design enabled us to examine response pattern similarity between conditions, and estimate fit to three conceptual models of how the two contexts relate: (1) social interaction and mentalizing about an abstract character are represented similarly; (2) interactive peers and abstract characters are represented differently regardless of the evaluation type; and (3) mental and nonmental states are represented dissimilarly regardless of target. We found that the temporal poles represent mentalizing and peer interactions similarly (Model 1), suggesting a neurocognitive link between the two in these regions. Much of the rest of the social brain exhibits different representations of interactive peers and abstract characters (Model 2). Our findings highlight the importance of studying social-cognitive processes using interactive approaches, and the utility of pattern-based analyses for understanding how social-cognitive processes relate to each other. Mentalizing (also known as "theory of mind") is the early developing ability to attribute mental states to others that are thought to play an increasingly critical role in navigating social interactions from childhood into adulthood (Frith & Frith, 2001;Frith & Frith, 2012;Wellman, 2017). Neuroimaging research has revealed a common network of brain regions, often referred to as the "mentalizing network," that are activated across a range of tasks that require mental state reasoning. This network includes the dorsal and ventral regions of the medial prefrontal cortex (dmPFC and vmPFC), inferior frontal gyrus (IFG), dorsolateral prefrontal cortex (dlPFC), precuneus, temporoparietal junction (TPJ), anterior temporal lobe (ATL), and superior temporal sulcus (STS; Mar, 2011;Molenberghs et al., 2016;Schurz et al., 2014). However, most neuroimaging investigations of mentalizing have utilized noninteractive experimental paradigms wherein adults make third-person attributions about abstract characters or simply observe social behavior. Thus, relatively little is known about the neural substrates linking mentalizing to active engagement in social-interactive contexts, much less the neurocognitive relationship between the two during the transition to adolescence, which is a period marked by significant social development.
Despite the relative lack of neuroimaging work examining the relationship between mentalizing and social interaction, evidence from developmental research suggests that the progression from childhood to adolescence may be an important period for understanding how mentalizing and peer interaction relate, as developmental changes are seen in both. Starting in early childhood, individuals' peer relationships become increasingly important, such that they spend progressively more time with friends over parents, and this social reorienting is accompanied by changes in brain systems associated with perception, motivation/affect, and executive function (Ladd, 1999;Nelson et al., 2016;Parker et al., 2006). This trajectory aligns with neuroimaging work demonstrating that the mentalizing network is functionally distinct by the age of three (Richardson et al., 2018), but exhibits increasing functional specialization through adolescence and into adulthood (Gweon et al., 2012;Moraczewski et al., 2020;Richardson et al., 2020). There is also increased sensitivity in the brain's reward system to social contexts as children transition into adolescence (Chein et al., 2011;Moreira & Telzer, 2018;Smith et al., 2015), and activity in the reward and mentalizing networks during adolescence is indicative of a tendency to spontaneously integrate peer perspectives into self-evaluations (Jankowski et al., 2014;Pfeifer et al., 2009;Van der Cruijsen et al., 2019). In particular, this work suggests an inverted-U-shaped trajectory in the neural sensitivity to and self-report of self-conscious emotions (i.e., social emotions like embarrassment that indicate perceived evaluation by others) from middle childhood to young adulthood (Somerville et al., 2013). Together, these findings indicate that the pre-adolescent to early adolescent period is a promising age-range for understanding the neurocognitive links between mentalizing and social interaction, and how they develop with age. Until recently, however, most of our understanding of the mentalizing network comes from work on adults.
Neuroimaging studies using adult samples have demonstrated activations primarily within the mentalizing network across tasks that elicit mentalizing through a variety of noninteractive approaches. This includes tasks with explicit instructions to make inferences about an abstract character's mental state (compared to, for example, physical characteristics), as well as the assessment of spontaneous mentalizing-that is, elicited independently of task demands (i.e., without prompting) 1 -for example, when individuals use mentalstate language to describe the behavior of moving shapes that are animated to resemble agency (see Mar, 2011;Molenberghs et al., 2016;Schurz et al., 2014). The spatial convergence of brain activations across these varied experimental paradigms has been used as evidence for the mentalizing brain network and laid the foundations for much of the social cognitive neuroscience literature since. However, this foundation has a crucial limitation: it characterizes social cognition only in observational (i.e., noninteractive) contexts, leaving gaps in our understanding of the neurocognitive processes involved in real-world, social-interactive behavior.

| SECOND-PERSON NEUROSCIENCE
A body of neuroimaging studies, collectively referred to as "secondperson neuroscience," has provided new avenues for understanding the neurocognitive processes involved in social interaction (Redcay & Schilbach, 2019;Schilbach et al., 2013). This set of approaches has utilized paradigms that involve engaging with a social partner in real time, thereby providing practical and theoretical advancements to the study of social cognition compared to traditional, third-person approaches. For instance, second-person neuroscience work using developmental samples has demonstrated that children are more motivated and rewarded when interacting with a live social partner compared to a computer or character, as indexed through behavioral responses, self-report, and neural activations (Alkire et al., 2018;Warnell et al., 2018). Second-person neuroscience studies using adult and child samples have also demonstrated that simply engaging with a live social partner recruits a more extended network of brain systems than has been previously reported in traditional social neuroscience studies relying on noninteractive, third-person approaches (Redcay et al., 2010;Redcay & Schilbach, 2019;Warnell et al., 2018). In particular, this work has demonstrated greater activations in the mentalizing network when individuals simply perceive a real-life social partner (versus an abstract or unknown social entity), even in the absence of any explicit mental state information or task demands to mentalize (Alkire et al., 2018;Redcay & Schilbach, 2019;Warnell et al., 2018). Although these findings seem to indicate that individuals are spontaneously mentalizing in the presence of a social partner, it is possible that these brain regions are part of an integrated network of systems that support "online" social behavior (i.e., when actively engaged in realtime social contexts) that have superficial overlap with regions supporting "offline" social cognition in noninteractive and observational contexts (Schilbach, 2014;Schilbach et al., 2013). 1 Throughout the manuscript, we contrast spontaneous with explicit mentalizing, which we define as being explicitly prompted by the task. This is not to be confused with implicit versus explicit cognitive processes (e.g., Frith & Frith, 2008), either of which can be engaged spontaneously.
Attempts at interpreting the meaning of activations in the mentalizing network during social engagements point to the broader limitations of reversely inferring a cognitive process from the spatial location of brain activations (Hutzler, 2014;Poldrack, 2006). For instance, Alkire et al. (2018) reported overlapping activations in the ATL, STS, and IFG when children reasoned about the mental state of an abstract character and when they engaged in social interactions that did not explicitly require mental state reasoning, and inferred that social interaction may induce spontaneous mentalizing. However, these brain regions are associated with other, higher-order cognitive processes that may be important for social interactions-the ATL and STS are associated with the representation of person-specific information (Anzellotti, 2017;Anzellotti & Caramazza, 2017;Blank et al., 2015;Collins et al., 2016;Olson et al., 2013;Perrodin et al., 2015;Simmons et al., 2010;Wang et al., 2017), and the IFG is implicated in a range of executive processes, such as working memory and behavioral inhibition (Breitling et al., 2020;Drummond et al., 2017;Hartwigsen et al., 2019). Thus, it is unclear if the overlapping activity associated with social interaction reported by Alkire et al. (2018) is specific to mentalizing, or reflects other cognitive processes utilized within social interactive contexts.
Multi-voxel, pattern-based approaches help alleviate some of the problems of reverse inference (Hebart & Baker, 2018;Poldrack, 2011). In particular, representational similarity analysis (RSA) leverages voxel-wise activity patterns to estimate the similarity of neural responses elicited by different task conditions, which can be used to infer commonalities in the underlying neurocognitive process (Kriegeskorte et al., 2008). RSA has proven to be a useful tool for assessing similarities between domaingeneral and social-cognitive functions (e.g., similar response patterns for social and physical distance in the inferior parietal lobule; Parkinson et al., 2014), and for disentangling subtle differences in response patterns across brain regions engaged by a common task that are indicative of diverging function (e.g., different trait dimen-  (Skerry & Saxe, 2015), and are involved in learning social information about unknown others (Dziura & Thompson, 2018). Importantly, RSA has revealed that activity patterns in mentalizing regions of both adults and children distinguish mental-state information and nonmental, social information, and that the distinctiveness of these patterns increased from 5 to 12 years of age, even when univariate activations do not exhibit a relationship with age (Richardson et al., 2020). Together, these findings demonstrate that RSA provides a sensitive approach to assess the similarity of neurocognitive processes associated with social interaction and mentalizing, and to assess how this relationship changed with age.

| CURRENT STUDY
The overarching goal of the current study is to advance our understanding of the neuro-representational links between social interaction and mentalizing during the transition to adolescence. In particular, the current work was motivated by three central questions: (1) Do brain regions associated with social interaction exhibit evidence for spontaneous mentalizing? (2) How does the neural similarity between social interaction and mentalizing change from middlechildhood to early adolescence? (3) What are common neurorepresentational links between social interaction and mentalizing exhibited across social brain regions? To accomplish this, we utilized a larger sample of participants who underwent the experimental paradigm originally reported by Alkire et al. (2018), wherein participants played a guessing game with a live interaction partner who provided hints, as well as making guesses about an abstract character, the hints for which were generated by a computer. In both cases, half of the hints involved information about the target's mental state and half involved nonmental, physical information about the target to determine the right answer. This yielded a within-subject, 2 (Peer/Character) Â 2 (Mental/Nonmental) design that enabled us to quantify the similarity of brain activity patterns associated with offline mentalizing and social interaction, and assess fit with models about the underlying neurocognitive process involved ( Figure 1).
We formalized three conceptual models of the neurocognitive links between social interaction and mentalizing. Consistent with the idea that social interaction engages spontaneous mentalizing, Model 1 (the "interaction engages mentalizing" model) proposes that interactions with a peer elicit a similar pattern of brain activity as when thinking about the mental state of a character, but that is dissimilar from the pattern of brain activity elicited by thinking about nonmental characteristics of a character. Alternatively, Model 2 (the "interaction" model) proposes that the interactive peer context and the character conditions elicit dissimilar patterns of brain activity from each other, but do not distinguish mentalizing and nonmental state reasoning from either target. Finally, Model 3 (the "mentalizing" model) proposes that mentalizing and nonmental state reasoning elicit dissimilar patterns of brain activity, but does not distinguish between peer and character conditions.
To address the first question of whether brain regions associated with social interaction show evidence for spontaneous mentalizing, we calculated the fit to each of our three conceptual models within a set of a priori regions of interest (ROIs) commonly implicated in studies of social interaction that were obtained from Neurosynth (Yarkoni et al., 2011). Additionally, convergent evidence for the ROI analysis was sought through exploratory whole-brain searches for model fit using the searchlight approach (Supplemental Section 5). Because the transition to adolescence is marked by significant changes in social-interactive contexts and mentalizing capabilities, we further assessed how neural similarity between social interaction and mentalizing changed from middle childhood to early adolescence by examining the relationship between age and model fit in each of the aforementioned social interaction ROIs. Finally, because we are interested in the contribution of brain structure associated with other social cognitive processes in linking social interaction and mentalizing, we conducted model-free analyses in a larger set of "social brain" ROIs obtained from Alcalá-L opez et al. (2018). The motivation for this set of analysis was to elucidate neuro-representational links between social interaction and mentalizing that were not captured by our a priori models, and to uncover similarities across brain regions in terms of their neural similarity structure (also known as "representational connectivity").

| Participants
A sample of 92 neurotypical 8-to 14-year-old participants were enrolled in a larger, multi-session project investigating the neural correlates of social interaction during middle childhood. Participants were recruited from the greater Washington, DC area, and exclusionary criteria included MRI contraindications, diagnosis of neurological or F I G U R E 1 Schematic of the socialinteractive fMRI task (a). Participants were given a half-second cue indicating if they would be answering questions provided by their interaction partner (peer) or answering questions presented by the computer about a story character (character). Hints were provided for 3.5 s and either required using mental state information about the target (mental), or nonmental, physical information (nonmental). This yielded a fully withinsubject, 2 (peer/character) Â 2 (mental/ nonmental) design ( (Table 1).

| Experimental protocol
The experimental protocol is the same as outlined by Alkire et al. (2018).
Briefly, participants were instructed that they would be chatting with a peer in a different lab who would also be undergoing a brain scan, and were shown pictures of two age-and gender-matched peers that they could choose from. In actuality, there were no live interaction peers, and every participant received the same stimuli. Participants were instructed on the "guessing game" that they played during the scan wherein participants were given a hint by their interaction partner on half the trials (Peer condition), and from the computer about a fictional story character for the other half of the trials (Character condition). It was the job of the participant to answer the question "Which will I/she/he pick?" (Mental condition) or "Which of these match?" (Nonmental condition) via button-press to select the appropriate response from the two answer choices. Mental and Nonmental items were counterbalanced across participants such that each item was presented in the Peer and Character conditions roughly an equal number of times throughout data collection to prevent unintended biases.
This yielded a fully within-subject, 2 (Peer/Character) Â 2 (Mental/ Nonmental) design. The guess phase of each trial (encompassing the target cue, hint, and response options; Figure 1a) was modeled as our event of interest and was followed by a 2-6 s jittered period before feedback about the correct answer was provided for 2 s. The task was presented using PsychoPy (Peirce, 2007)

| Behavioral data analysis
Analyses of behavioral task performance-accuracy (percent correct) and reaction time (RT) in seconds-and the postscan questionnaire were conducted using R (R Core Team, 2020) and JASP (JASP Team, 2020).
The continuous variables of age, accuracy, and RT were first assessed using the Shapiro-Wilk test for normality, and data were transformed as needed to meet the assumptions of the parametric analysis of variance.
Between-group t-tests for each condition's accuracy and RT were calculated to assess gender effects, and if either measure showed significant gender differences for any condition, gender was entered as a covariate in the subsequent analysis of variance for the measure. Correlations with age for each condition's accuracy and RT were calculated to assess age effects, and if either measure showed a significant correlation with age for any condition, age was entered as a covariate in subsequent analysis of variance for the measure. Accuracy and RT were each entered into two-way repeated measures analysis of variance (with appropriate covariate as needed) to determine main effects of social Briefly, anatomical images were segmented and normalized to MNI space; functional images were skull-stripped, susceptibility distortion corrected, realigned, slice-time corrected, coregistered, and warped to the normalized anatomical image (see Supplemental Methods 2 for full fMRIprep preprocessing pipeline). Additionally, functional data were masked using subject/run-specific masks generated by fMRIprep, and intensity normalized to a mean of 100 per voxel. Subject-level, voxelwise multiple linear regression was calculated using AFNIs (Cox, 1996) 3dREMLfit for each run separately. The guess period, which includes the target cue (Peer/Character), hint, and responses options (Figure 1a), for each of the four conditions (i.e., Peer Mental, Peer Nonmental, Character Mental, and Character Nonmental) was modeled as our events of interest. This was achieved by convolving the guess period with the canonical hemodynamic response using a duration modulated response function (AFNIs dmBlock) with RT as duration to ensure that cognition related to the guess period was captured. The feedback period was modeled separately as events of no interest, along with the six motion parameters (x, y, z, roll, pitch, and yaw), their derivatives, and volumes censored due to framewise displacement >1 mm. This process yielded subject-level tmaps for each condition, which were used for all subsequent analyses.

| Representational similarity analysis
Neural representational dissimilarity matrices (RDMs) were calculated per subject for each ROI by extracting voxel-wise t-values associated with each condition, using the voxel-wise values to calculate the Euclidean distance between each pair of conditions using the CoSMoMVPA toolbox for MATLAB (Oosterhof et al., 2016), and normalizing the RDMs by subtracting the minimum Euclidean distance and dividing by the range. Neural RDMs were then tested against three conceptual models of hypothesized relationships between conditions, which were formalized as binary RDMs wherein 0 = similar and 1 = dissimilar. Model 1, or the "interaction engages mentalizing" Model fit was estimated by calculating Kendall's Tau-a rank-order correlation coefficient between the off-diagonal elements of the neural and model RDMs using the RSA toolbox for MATLAB (Nili et al., 2014). For each ROI, model fit was calculated between each participant's neural RDM and the three model RDMs, and were transformed to z value using the Kendall's Tau normal approximation formula (z = 3τ*√n(n À 1)/√2(2n + 5)). The model fit estimates were entered into a Bayesian multilevel (BML) model using AFNIs RBA F I G U R E 2 Steps for model-based representational similarity analysis: We first estimate the voxel-wise response pattern for each condition for a given brain region using unsmoothed subject-level models (a), then calculate the Euclidean distance between response patterns for each pair of conditions to construct our neural dissimilarity matrices (which can also be visualized as dendrograms; b), and estimate fit to each of our models by calculating Kendall's tau-a (c). For the representational connectivity analyses, you examine the fit between the neural dissimilarities between brain regions (d) program  to assess the strength of evidence in favor of each model per ROI. This approach confers multiple advantages over the traditional null hypothesis significance testing frameworkthe ability to get around the issue of multiple comparison corrections through the calculation of a single model that accounts for the hierarchical structure of the data, better model fit afforded by the estimation of informed priors using partial pooling of data across levels, and the full reporting of results rather than dichotomizing findings based on significance thresholds (Chen et al., 2021). As a proof-of-concept, we used the aforementioned approach to conduct a preliminary reanalysis of the brain regions reported by Alkire et al. (2018)

| Model fit analysis of social interaction ROIs
We conducted a set of theory-driven analyses utilizing a set of ROIs obtained through NeuroSynth (Yarkoni et al., 2011) to examine representational similarity within brain regions associated with social interaction. The motivation behind this set of analyses was to examine brain areas associated with social interaction that may exhibit patterns

| Model-free analysis of the social brain ROIs
Model-free analyses were conducted to evaluate if other, nonhypothesized relations between task conditions exist in the social brain, and to examine the organization of these regions based on pattern similarity structure. That is, rather than examining fit to prespecified models, "representational connectivity" analysis estimates the relationship between brain areas in terms of how similar they represent the four conditions. This allows us to uncover clusters of brain regions that are "representationally connected" in terms of how they link social-interactive and mentalizing processes. To this end, we started with a set of 36 publicly available ROIs derived from an extensive set of meta-analytic, data-driven analyses of 3972 social neuroscience studies using fMRI and/or PET imaging (Alcalá-L opez et al., 2018). The ROIs were resampled to the resolution of our functional data and inflated within a gray matter mask informed by white matter and CSF skeleton to maintain a comparable amount of brain coverage and attention to anatomical contours. The resulting ROIs had a uniform volume of 120 voxels.
Analyses of these social brain ROIs proceeded in three major steps.
Step 1: because the social brain ROIs cut across many different brain systems that may not be involved in any of our task conditions, it was necessary to calculate the lower bound of the "noise ceiling," which expresses how consistent the representations are across participants. Noise ceiling calculation was conducted for each ROI and tested against zero (i.e., no consistency across participants) to determine their inclusion in further analyses (Lage-Castellanos et al., 2019).
Noise ceiling was calculated by creating an average neural RDM across all but one participant, calculating Kendall's Tau-a between this average RDM and the neural RDM from the left-out participant, and iterating this process across participants. The Kendall's Tau-a coefficients were then tested against zero using Wilcoxon signed-rank tests, and ROIs with a noise ceiling significantly above chance were retained.
Step 2: the resulting ROIs were interrogated using exploratory factor analysis (EFA; Fabrigar & Wegener, 2011) to determine the factor structure of the response patterns across the ROIs, which elucidated the number of groups that the ROIs were separated into based on representational connectivity (i.e., shared similarity structures between ROIs). This was conducted by first averaging the neural RDMs across participants for each ROI, and entering the averaged Euclidean distances into an EFA using minimum residual extraction and oblimin rotation (though results from principal axis extraction and other rotation methods yielded similar results). A cut-off eigenvalue of 1 and a minimum of 80% explained variance were used to determine the factor solution.
Step 3: the Euclidean distances were used to calculate correlations between each pair of ROIs, which were submitted to a hierarchical cluster analysis using Ward's method (Ward, 1963) to identify clusters of regions with similar neural similarity structures, and the number of clusters were set to the number of factors determined from the EFA. The neural RDMs for each ROI within each cluster were averaged together and qualitatively examined for a better understanding of the neural representations of each cluster of ROIs.

| Behavioral data analysis
Shapiro-Wilk tests for normality revealed that accuracy measures violated the assumption of normality (Table S3). Arcsin square root transformation of accuracy scores (the recommended transformation for percentages) did not fix the distribution of these measures to pass the Shapiro-Wilk tests for normality (nor any other transformations).
Because there are no nonparametric equivalents of the two-way repeated-measures ANOVA, we proceeded with the parametric analyses as planned, but reported results of nonparametric tests for each of the significant effects we obtained from the accuracy ANOVA ( Nonmental questions is greater with increasing age (Figure 3a;  (Table S7).  Supplemental Section 5), however, these were uncovered using an exploratory uncorrected threshold, so interpretations must be made with caution. The searchlight analyses provided additional support for Model 2 fit in the dmPFC and bilateral TPJ, which also uncovered clusters in the visual cortex, left pSTS, left IFG, and frontal pole ( Figure S4b and Table S10b). Searchlight for Model 3 did not reveal a significant fit in any region, even at exploratory threshold values.

| Model-free analyses within the social brain
Analyses to examine the similarity structure of the social brain ROIs Exploratory factor analysis of these 16 ROIs yielded a 3-factor solution that explained 87.2% of the variance in Euclidean distances.
Hierarchical clustering using Ward's method was used to estimate three clusters that minimized the total within-cluster variance ( Figure 5b).

| DISCUSSION
This is the first study to evaluate the assumption that mentalizing is spontaneously engaged during social interactions by assessing the similarity in brain response patterns associated with mentalizing and peer interactions. Across our analyses we found the most consistent evidence that offline mentalizing and live peer interactions elicit similar patterns of brain activity in the left ATL, but a dissimilar pattern of activation when making nonmental state inferences about an abstract character (i.e., the "interaction engages mentalizing" model or Model 1). Our findings align with traditional, offline studies of mentalizing that have demonstrated differences in brain activity in this region when making mental versus nonmental inferences about abstract others (Aichhorn et al., 2009;Andrews-Hanna et al., 2014;Dodell-Feder et al., 2011), and add to the literature by demonstrating that patterns similar to mentalizing are elicited during social interactions that do not explicitly require mental state reasoning. In addition to providing support for the idea that mentalizing is spontaneously engaged in social-interactive contexts, our results suggest that the ATL and other temporal regions may contribute to the integration of online and offline social-cognitive processes. That is, social interactions involve the dynamic interplay between social knowledge that is available offline, such as an understanding of mental state concepts, and the online demands of reciprocal social exchange. This interpretation also fits with recent reviews describing how we get to know others, which involves acquiring and updating person-specific knowledge through online engagements with them, and implicates the ATL and pSTS in this process (Anzellotti & Young, 2020;Kovács, 2020).
Consistent with the idea that the ATL supports the integration of online and offline social processes, different lines of research have demonstrated the convergence of multimodal, social information processing streams in the anterior portions of the temporal lobe.
Ventral areas of the ATL have been associated with the representations of person-specific knowledge (e.g., names, faces, and biographical information; Anzellotti, 2017;Blank et al., 2015;Borghesani et al., 2019;Kriegeskorte et al., 2007;Wang et al., 2017), while more dorsal, anterior portions of the ATL are considered to be a semantic hub for abstract social concepts (including mental-state concepts; Arioli et al., 2020;Olson et al., 2007Olson et al., , 2013Skipper et al., 2011;Wang et al., 2019;Zahn et al., 2007). Conceptually, our findings span the middle ground between the aforementioned lines of research in that our results suggest that peer-specific and abstract knowledge about mental states are integrated and represented similarly in the ATL. This interpretation is supported by research indicating that the interplay between ventral and dorsal ATL contributes to the successful encoding and retrieval of information associated with a person's identity (Perrodin et al., 2015;Rice et al., 2018;Tsukiura et al., 2010).
Moreover, this work has demonstrated left lateralization for associating person-specific information with the individual's name Borghesani et al., 2019;Olson et al., 2013;Tsukiura et al., 2008), which provides some explanation as to why our most consistent findings were in the left ATL, since our stimuli used proper names. However, even though our findings are consistent with the different functional accounts of the ATL mentioned above, additional work is needed to disentangle the specific computations engaged by the ATL that links social interaction and mentalizing.
Many of the other brain regions we examined exhibited fit to our "interaction" model that distinguishes the Peer and Character tally different from what has been revealed using third-person approaches (Redcay & Schilbach, 2019;Schilbach et al., 2013) by demonstrating differing sensitivities to second-and third-person contexts across brain regions, but confirmatory research is needed to provide support for this interpretation.
Our findings are also consistent with previous work demonstrating changes in both mentalizing and social interaction as individuals transition from middle childhood to adolescence (Richardson et al., 2020). Accuracy scores indicate that mentalizing ability improves with age (i.e., age-by-question-type interaction), and reaction time measures suggest that this might be driven by an increased efficiency in making mental state inferences specifically for peers (i.e., significant three-way interaction between age, question type, and target; Figure 3 and Table 2). These behavioral effects align with agerelated decreases in fit with the "interaction" model (Model 2) in the left caudate and right vlPFC, which was driven by increasing neural dissimilarity between the Peer Mental and Peer Nonmental conditions. Because the vlPFC and caudate are associated with reward and valuation (Bartra et al., 2013;Nejati et al., 2018), this result may relate to the ever-increasing importance of peers, and particularly the need to understand their perspectives, during this age range. Indeed, research has demonstrated heightened sensitivity of the reward system to social contexts during the transition into adolescence (Chein et al., 2011;Moreira & Telzer, 2018;Smith et al., 2015), and a related increase in the tendency to automatically integrate peer perspectives into self-evaluations (Jankowski et al., 2014;Pfeifer et al., 2009;Van der Cruijsen et al., 2019). It is worth noting that the conditioned social reward associated with "texting" might confound the inherent reward experienced in social interactive contexts since the Peer condition stimuli were presented in a format that resembles text messages, especially given that the age range of our sample is when individuals start obtaining their own mobile devices to chat with their peers.
However, this interpretation is hard to reconcile with the fact that the age-related changes we found were for increasing dissimilarity between the two Peer conditions, which would not be expected if the reward-related activity was driven by associations with texting in general. Nonetheless, the current findings add to our understanding of the development of social cognition by demonstrating that improvements in mentalizing capacity (or propensity to do so) during the transition to adolescence are particularly salient in social-interactive contexts.
The developmental findings of the current project were enabled, in part, by the rapid changes in social and neural development that occur during the pre-adolescent to early adolescent ages (Kilford et al., 2016;Mills et al., 2014), but our snapshot of this narrow age range also limits the generalizability of our results to other ages and populations. Thus, it is unknown whether the age-related differences in the representational similarity structures we observed would continue in a linear trajectory as individuals transition into late adolescence and adulthood, or if they would return to pre-adolescent levels similar to the inverted-U shape trajectories observed in other socialcognitive domains (Kilford et al., 2016;Somerville et al., 2013). This unknown is particularly challenging for many of our results in cortical regions associated with social cognition that undergo substantial structural changes well into adulthood, even though whole-brain volume remains relatively stable after late childhood (Mills et al., 2014(Mills et al., , 2016. At a theoretical level, although much of the motivation of the current work was to understand if mentalizing is a constituent process of social interaction, others have argued for the primacy of early social-interactive experiences as the driving force for later-developing social-cognitive abilities, like mentalizing outside of social-interactive contexts (Schilbach, 2014). That is, early social interactions may enable the neural architecture upon which offline mentalizing abilities develop. Together, our findings add to the developmental social cognitive literature, but data from a sample with a wider age range is needed to fill in the developmental trajectory of the relationship between offline mentalizing and social interaction.
In addition to advancing our theoretical understanding of the development of social cognition, the current work also demonstrates the utility of pattern-based analyses for uncovering links between constituent processes involved in social behavior. The re-analysis of the brain regions reported by Alkire and colleagues that showed over- ing specificity to mental states compared to nonmental social information from childhood into adulthood (Gweon et al., 2012;Richardson et al., 2020), and thus the age range of our sample may have contributed to our lack of significant Model 3 fit. This interpretation is supported by our behavioral analyses which revealed better accuracy for Mental trials and quicker response for the Peer Mental trials with age ( Figure 3 and Table 2). The inclusion of a matched, but fully nonsocial condition would aid in teasing apart the effects of social development.
Another factor potentially contributing to the minimal dissimilarity between mental and nonmental conditions is that the items comprising the Mental condition were deliberately varied to cover a range of mental state categories, including wants, likes, knowledge, beliefs, and emotions, while the Nonmental items asked about activities, physical characteristics, abilities, situations, and possessions (Table S1). Therefore, it is possible that this amount of variation led to nonrepresentative response patterns when averaging across items for each condition per subject. Furthermore, hints for the Mental trials included language about mental states under the assumption that such language would elicit some degree of mentalizing, but it is possible that some participants relied on strategies for making guesses that cut across the Mental/Nonmental conditions, such as semantic associations between hint and response options. Following this interpretation, our behavioral effects for question type might reflect confounding processes related to the Mental condition, such as the inherent abstractness of the language used to describe mental states versus physical attributes, which is particularly challenging to disentangle because the representation of semantic associations, abstract concepts, and social concepts are subserved by overlapping brain regions (e.g., Binney et al., 2016;Xu et al., 2018).
The aforementioned interpretations do not align with the direction of our behavioral findings, nor with some of the activations reported by Alkire et al. (2018). That is, if linguistic abstractness associated with Mental trials was driving our behavioral effects, we would expect slower reaction time and reduced accuracy (Borghi & Zarcone, 2016), rather than the target-by-question type interactions in RT and greater accuracy for Mental trials (Table 2). Activations in the dmPFC for the Mental > Nonmental and the Peer Mental > Character Mental contrasts (Alkire et al., 2018) also do not align with prior work that generally implicate lateral cortical structures in the processing of abstract concepts (Binney et al., 2016;Conca et al., 2021;Wang et al., 2010). Moreover, careful examination of our experimental stimuli revealed that a substantial portion of the items could not easily be answered through semantic association, and that for some items, semantic associations would lead to the wrong answer (e.g., "[Target] thinks skiing is dangerous" is semantically related to "Mountain slope" even though the target picking "Hot Cocoa" better matches the target's mental state; Table S1). Nonetheless, the explanations provided above do not preclude the possibility that other, unrealized confounds may have contributed to our lack of Model 3 fit, and future research should work to disentangle the effects of linguistic processes that overlap with social cognition.
Despite these limitations, our findings demonstrate the importance of incorporating realistic, social-interactive contexts for a better understanding of the neural substrates subserving everyday social cognition, and they highlight the utility of pattern-similarity-based analyses for examining the relationship between related socialcognitive processes. Through this set of approaches, we provided additional evidence for the spontaneous engagement of mentalizing during social interactions, and spotlight the functional contributions of the ATL toward integrating offline social cognition with online social interactions. Future work could build on these findings by establishing more precise neural and behavioral measures of mentalizing and related social-cognitive processes to better establish process-level similarities and differences between them. By elucidating the links and dissociations between the various cognitive mechanisms underlying social interactions, we will gain not only a better understanding of the social behaviors that shape our world but also a more complete cognitive map of how different neurocognitive functions relate to each other.

CONFLICT OF INTEREST
The authors declare that they have no competing interests, financial or otherwise.

DATA AVAILABILITY STATEMENT
The data from consenting participants that support the findings of this