Social threat avoidance depends on action-outcome predictability.

Avoiding threatening individuals is pivotal for adaptation to our social environment. Yet, it remains unclear whether social threat avoidance is subtended by goal-directed processes, in addition to stimulus-response associations. To test this, we manipulated outcome predictability during spontaneous approach/avoidance decisions from avatars displaying angry facial expressions. Across three virtual reality experiments, we showed that participants avoided more often when they could predict the outcome of their actions, indicating goal-directed processes. However, above-chance avoidance rate when facing unpredictable outcomes suggested that stimulus-response associations also played a role. We identified two latent classes of participants: the "goal-directed class" showed above-chance avoidance only in the predictable condition, while the "stimulus-response class" showed no credible difference between conditions but had a higher overall avoidance rate. The goal-directed class exhibited greater cardiac deceleration in the predictable condition, associated with better value integration in decision-making. Computationally, this class had an increased drift-rate in the predictable condition, reflecting increased value estimation of threat avoidance. In contrast, the stimulus-response class showed higher responsiveness to threat, indicated by increased drift-rate for avoidance and increased muscular activity at response time. These results support the central role of goal-directed processes in social threat avoidance and reveal its physiological and computational correlates.