Social training reconfigures prediction errors to shape Self-Other boundaries

The study asks whether neural circuits that compute prediction errors (PEs) for Self and Other are inherently agent-specific or whether their agent-specificity adapts to social experience. Prior work shows the brain can simulate other agents’ learning signals and that Self- and Other-attributed PEs can be neurally segregated, with the degree of segregation predicting behavioral Self-Other distinction. The authors hypothesize that Self-Other boundaries are plastic and shaped by the proportion of shared experiences between agents. They test whether manipulating experience-sharing between Self and Other (high vs low shared-information contexts) induces lasting changes in behavioral and neural Self-Other distinction, and whether ventromedial prefrontal cortex (vmPFC) structure and function track this adaptation.

Prior literature indicates simulated reward prediction errors during observational learning and mentalizing in medial prefrontal regions, and signals related to others’ sensory surprise using fMRI and MEG. Distinct neural circuits can encode Self vs Other sensory PEs, and the degree of neural distinction correlates with Self-Other behavioral distinction. vmPFC is implicated in self-referential processing, representation of others’ states depending on task framing, and mapping latent contextual states; myelination in social brain networks relates to social cognitive development. Animal work links social experience to prefrontal myelination. These findings motivate examining experience-dependent plasticity of Self-Other PE representations and vmPFC’s role in tracking relational structure between agents’ computations.

Design: Three-day experiment. Day 1 included an intertemporal choice task and a baseline visual perspective-taking task. Day 2 involved training on a probabilistic false belief task (FBT) in two social contexts differing in the frequency of shared information (Hi-Share vs Lo-Share). Day 3 involved testing on the FBT during fMRI with contexts matched on trial statistics, plus resting-state and quantitative MRI; additional analyses examined transfer to perspective-taking and relations with temporal discounting.

Probabilistic False Belief Task (FBT): On each trial, a Bernoulli outcome was generated from a drifting parameter p. Trial types: privileged (seen only by Self), shared (seen by Self and Other), and decoy (a misleading sample presented to Other, visible and known to Self). Subjects intermittently reported estimates of p for Self or for the Other player (a real person represented by an avatar). Training (day 2) manipulated shared-trial frequency between contexts (Hi-Share vs Lo-Share). Testing (day 3) used identical trial statistics across contexts to assess sustained effects.

Behavioral modeling: Parallel belief-updating models for Self and Other were fit to probe responses. Models included learning rates (α), decision temperature (τ), memory decay (δ), and a Self-Other ‘leak’ parameter (λ) capturing spillover where one agent’s PE updates the other’s belief. Model spaces varied which parameters were shared vs agent-specific and by trial type. Model comparison used BIC across contexts and sessions; model recovery analyses assessed identifiability.

Perspective-taking task: A visual perspective-taking paradigm required adopting Self or Other perspective to judge presence/number of target patterns in a scene where the avatar had a restricted field of view. Congruent vs incongruent conditions manipulated overlap between perspectives. Drift-diffusion modeling (fast-dm 2.0) estimated parameters (e.g., drift rate) at baseline and after FBT training to assess transfer. A corrected change score isolated training-specific effects.

Intertemporal choice: Subjects chose between smaller-sooner and larger-later monetary rewards. One- and two-parameter hyperbolic discounting models (with power-law time perception) fit choice data via empirical Bayes (EM algorithm with MAP per subject), incorporating a softmax choice rule.

MRI acquisition and preprocessing: 3T Siemens Prisma; BOLD EPI with 64 slices (3 mm isotropic), TR = 2 s, TE = 30 ms, flip angle 90°, field maps, and structural MPRAGE. Physiological monitoring during scanning. Preprocessing in SPM12: realign/unwarp, coregistration to MT image, normalization to MNI, 8 mm FWHM smoothing; PhysIO regressors modeled physiological noise.

fMRI analyses: GLMs localized Self- and Other-attributed unsigned PEs across privileged/shared/decoy trials using model-derived regressors. Searchlight multi-voxel pattern analysis (Decoding Toolbox) estimated trial-wise decoding of PE magnitudes. Within a mask formed by PE-responsive clusters (FWE corrected), Self vs Other classification used LASSO logistic regression on PCA-reduced features with nested cross-validation. Cross-decoding trained linear models on Self PE patterns and tested on Other PE patterns (and vice versa) to quantify representational overlap.

Quantitative MRI: Magnetization transfer (MT) maps indexed myelin-related contrast. Whole-brain regression related MT to context-dependent differences in Self–Other cross-decoding, controlling for covariates (age, gender, intracranial volume).

vmPFC relational learning regressor: A model tracked perceived probability of ‘shared’ trials (P.shared) with different learning rates to generate trial-wise regressors. Small-volume corrected vmPFC ROI and whole-brain analyses tested BOLD covariation with P.shared; correlations with cross-decoding adaptation assessed links between relational tracking and representational plasticity.

• Behavioral context effect in FBT: Significant main effect of context across sessions, with performance differing between Hi-Share and Lo-Share (training: F(1,39) = 12.64, p = 0.001; testing: F(1,39) = 6.78, p = 0.013). An interaction indicated larger differences in subjects who experienced more sharing during training. Probe-level effects showed differences between Self and Other probes depending on context (e.g., paired t-tests reported for specific contrasts).
• Self–Other belief merging: Model-derived Self and Other belief trajectories were more correlated in Hi-Share than Lo-Share (training: F(1,39) = 21.7, p < 0.001; testing: significant main effect reported). Simulations indicated λ (Self–Other leak) strongly drove this merging.
• Learning models: Hi-Share behavior best explained by models including the Self–Other leak parameter λ; Lo-Share by models without λ. In the best Hi-Share model, learning rates differed by trial type and were higher for Other than Self updates (paired t(39) = −2.27, p = 0.029). The Self-Other learning-rate difference correlated with Self vs Other performance differences (Spearman ρ = −0.45, p = 0.004).
• Transfer to perspective-taking: After training, corrected drift rate change was higher for trials involving the Lo-Share agent than the Hi-Share agent (n = 46; repeated-measures ANOVA main effect of avatar: F(1,45) = 7.4, p = 0.009). Baseline congruency effects replicated (t(45) = 7.6, p < 0.001).
• fMRI representational adaptation: Within PE-responsive clusters, Self vs Other classification was above chance in Lo-Share (one-sample t(39) = 2.0, p = 0.031) but not in Hi-Share (t(39) = −0.26, p = 0.81); contexts differed (paired t(39) = 2.12, p = 0.041). Cross-decoding (Self→Other and Other→Self) was above chance in Hi-Share (one-sample t(39) = 4.22, p < 0.001), not in Lo-Share (t(39) = 0.26, p = 0.951), and higher in Hi-Share than Lo-Share (paired t(39) = 2.75, p = 0.009). Cross-decoding differences correlated with the behavioral leak-to-learning-rate ratio (r = 0.41, p = 0.009).
• vmPFC myeloarchitecture: MT in right vmPFC-adjacent white matter correlated with the cross-decoding difference between contexts (cluster: 844 voxels, whole-brain FWE p < 0.001; peak x = 12.8, y = 59.2, z = 18.4). Higher MT predicted greater representational adaptation.
• vmPFC tracks shared-information probability: BOLD in bilateral vmPFC covaried with model-derived probability of shared trials (small-volume corrected p = 0.008); whole-brain analysis identified a left lateral temporal/pole cluster (FWE p = 0.006; x = 62, y = −16.2, z = −16). vmPFC and temporal cluster contrast estimates were correlated (r = 0.45, p = 0.004). vmPFC tracking correlated with representational adaptation (r = 0.35, p = 0.032; after excluding two outliers); no such relationship for the temporal cluster.
• Relation to temporal discounting: Greater Self–Other distinction (behavioral and neural) was associated with less temporal discounting. Discounting propensity could be predicted above chance using λ (p = 0.005, permutation test), and prediction improved when including fMRI cross-decoding measures (p = 0.016). Myelin-related MT in the vmPFC region was negatively associated with discount factor (p = 0.022, small-volume corrected).

Manipulating the frequency of shared experiences between Self and Other reshapes how PE signals are partitioned into agent-specific representations. High sharing promotes Self–Other merging of PE representations (greater cross-decoding, reduced separability), whereas low sharing promotes segregation (better Self vs Other classification). These neural changes persist at least 24 hours and generalize to a distinct, non-learning visual perspective-taking task, indicating the acquisition of agent-specific relational knowledge that transfers across cognitive domains. vmPFC appears to track the statistical relationship between agents’ computations (probability of shared trials) and structural variability in vmPFC white matter (MT) predicts the degree of representational adaptation, implicating vmPFC in acquiring or deploying relational knowledge about Self–Other boundaries. The link between Self–Other distinction and temporal discounting supports a domain-general relational mechanism that governs generalization across social (Self vs Other) and temporal (present vs future self) distances.

The study demonstrates that Self–Other boundaries in neural PE representations are plastic and adapt to the statistical structure of social experiences. Training with high shared information increases representational overlap between Self- and Other-attributed PEs; low shared information enhances segregation. These adaptations correlate with vmPFC myeloarchitecture and vmPFC tracking of shared-information probabilities, and they transfer to perspective-taking performance and relate to temporal discounting tendencies. The work suggests a frontotemporal network leverages relational structure between agents’ computations to flexibly set Self–Other boundaries. Future research should directly manipulate temporal contingency between Self and Other signals to identify the underlying learning algorithm, further specify vmPFC’s computational role, and test generalization to other forms of relational learning and social contexts.

• The manipulation used proportion of shared trials as a proxy for temporal contingency; the contingency between Self and Other signals was not directly manipulated, limiting mechanistic inference about the learning algorithm.
• vmPFC correlation with adaptation (P.shared regressor) required exclusion of two outliers to reach significance; robustness needs replication.
• Structural findings are cross-sectional (no longitudinal structural change measured); correlations between MT and adaptation cannot establish causality.
• Some behavioral and figure-caption details suggest inconsistencies in the direction of performance effects across contexts; replication with larger samples and preregistration would help clarify behavioral patterns.
• Tasks may not be uniquely social; observed transfer effects could reflect domain-general relational learning processes rather than exclusively social cognition.
• Sample sizes for some analyses (e.g., fMRI n ≈ 40; perspective-taking n ≈ 46) are moderate; effects should be validated across independent cohorts.