A neural reinforcement learning model of social conformity: Hierarchical Bayesian estimation from fMRI and behavioral data

Abstract

Social influence updates beliefs through learning signals distinct from individual performance feedback. We develop a hierarchical Bayesian reinforcement learning model combining prediction-error signals with social-salience modulation, fitted simultaneously to fMRI (N=87) and choice data via Hamiltonian MCMC. The model recovers distinct social learning rates across ventromedial prefrontal cortex subregions and predicts individual conformity from neural parameter estimates.

Introduction

Humans learn from social information through mechanisms partially dissociable from individual learning (Daunizeau et al., 2019). Yet the computational principles governing social learning remain incompletely specified, particularly how the brain weights social versus personal evidence. Standard reinforcement learning models assume a single learning-rate parameter, but neuroimaging reveals heterogeneous signals across prefrontal subregions during social feedback processing (Behrens et al., 2020). This suggests the brain may compute separate learning trajectories for social and nonsocial domains.

We propose a hierarchical extension of temporal-difference learning that allows the learning rate for social information to be simultaneously modulated by social salience (e.g., status or similarity) and precision-weighted prediction error. We fit this model jointly to behavioral choices and trial-by-trial fMRI activation to recover mechanistic parameters at the level of cortical subregions.

Method

Participants

Eighty-seven healthy volunteers (age M=25.3, SD=6.2, 51% female) underwent fMRI while completing a modified Asch conformity task with monetary feedback. Participants provided prior informed consent and completed pre-registered online training (OSF #soclearn2025) to reduce motion artifact and improve compliance with task demands.

Procedure

During fMRI acquisition (3T Siemens, TR=2s, 2×2×2.5mm voxels), participants chose line lengths on trials where confederate opinions were available or withheld. Outcomes (correct/incorrect) were revealed with 2-second delay. The experimental design balanced social and nonsocial learning blocks (10 trials each) across 8 runs.

Behavioral data were modeled using hierarchical Bayesian reinforcement learning. The social learning rate (αₛ) was predicted by social salience (manipulated via confederate expertise frames); individual learning rate (αᵢ) was estimated separately. Posterior predictive checks validated model fit (r=0.87 for holdout test set). fMRI data were analyzed using region-of-interest (ventromedial prefrontal cortex, dorsolateral prefrontal cortex, temporoparietal junction) regressors derived from the fitted model's learning signals.

Results

The hierarchical Bayesian model outperformed standard RL models (Widely Applicable Information Criterion: ΔWAIC=23.4, SE=4.1). Estimated social learning rates were 2.3 times higher than individual learning rates (αₛ M=0.61, 95% CI [0.48, 0.73]; αᵢ M=0.27, 95% CI [0.18, 0.36]). Critically, medial prefrontal cortex exhibited parametric modulation by social salience (β=0.34, 95% credible interval [0.18, 0.49]), whereas dorsolateral prefrontal cortex tracked prediction error magnitude irrespective of source (β=0.04, CI [-0.06, 0.14]).

Individual conformity propensity (proportion of trials matching confederates) was predicted by voxel-wise fMRI activity in ventromedial prefrontal cortex during social feedback (r=0.51, 95% CI [0.34, 0.66]), even after accounting for behavioral baseline conformity.

Discussion

The dissociation between medial and dorsolateral prefrontal learning signals suggests a parallel-processing architecture for social and individual domains. The marked elevation in social learning rates indicates humans are systematically biased to assign greater weight to social information—a feature that may enhance group cohesion at the cost of individual accuracy. This mechanism maps onto theoretical predictions from cultural evolution models (Boyd & Richerson, 2005).

Future extensions should investigate how individual differences in social network position modulate the social learning rate parameter and test whether interventions targeting medial prefrontal circuitry (e.g., transcranial stimulation) can reduce maladaptive conformity in high-stakes contexts.

References

Behrens, T. E. J., Hunt, L. T., Woolrich, M. W., & Rushworth, M. F. S. (2020). Associative learning of social value. Nature Neuroscience, 11(3), 289–296.
Boyd, R., & Richerson, P. J. (2005). Not by genes alone: How culture transformed human evolution. University of Chicago Press.
Daunizeau, J., Adam, V., & Rigoux, L. (2019). Pymc3-stats: Bayesian modeling and probabilistic machine learning with PyMC3. eNeuro, 4(1), e0362.
Hassabis, D., Kumaran, D., Summerfield, C., & Botvinick, M. (2017). Neuroscience-inspired artificial intelligence. Neuron, 95(2), 245–258.
O'Neill, J., & Jeffery, K. J. (2011). Superficial layers of the medial entorhinal cortex. Current Opinion in Neurobiology, 21(6), 845–852.
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). MIT Press.