How gaze interactions give rise to social closeness: A computational learning account

Social gaze is a fundamental cue regulating approach and avoidance in social interactions, yet it remains unclear how repeated gaze exchanges give rise to enduring social preferences and closeness. Here, we propose a computational learning account in which gaze-contingent feedback generates prediction errors that update latent social values, thereby shaping social closeness over time. According to this account, direct and averted gaze function as socially rewarding or non-rewarding outcomes whose impact depends on individual reward calibration. We tested this account using a gaze-interaction paradigm in which participants repeatedly chose between two interaction partners that predominantly (70%) responded with direct or averted gaze and rated their perceived social closeness toward each partner. Consistent with the model, participants increasingly preferred and felt closer to the partner who more frequently returned direct gaze. Computational modeling revealed that gaze direction elicited prediction errors that updated the latent value of the interaction partner. Crucially, these learning signals were calibrated by individual differences in the subjective reward value of direct gaze. Initial attractiveness and valence of the interaction partner predicted this calibration, indicating that pre-existing impressions shape how gaze feedback is valued. In turn, reward calibration was reflected in participants’ post-interaction social evaluations, demonstrating how learning experiences contributed to impression formation. Together, these findings provide a computational account of how minimal gaze interactions can give rise to social preference and closeness through reinforcement learning mechanisms. More broadly, they suggest that eye contact functions as a powerful but individually calibrated social signal that dynamically shapes social affiliation.