Context shapes action understanding in social interactions.

In social interactions, individuals anticipate what others are about to do next, enabling smooth social interactions and coexistence in daily life. Here, we examined how kinematic and contextual cues are weighted and integrated to predict an interactor's actions during dynamic, real-time competitive interactions. We analyzed 76 participants (38 dyads) freely playing 300 games of rock-paper-scissors each while tracking their kinematics using motion capture. Although movement kinematics consistently conveyed intention-related information, successful action selection depended primarily on how this information influenced predictions. Reinforcement-learning based contextual inference emerged as the strongest predictor of single-trial success and modulated the degree to which movement-derived predictions guided behavior. When contextual and movement-derived information afforded different actions, prioritizing contextual inference consistently outperformed movement-derived predictions. Thus, successful social action understanding depends not on maximizing sensitivity to movement cues, but on adaptively arbitrating between movement-derived and outcome-based contextual information during ongoing social interaction.