Complex collective behaviors such as schooling are believed to emerge from simple, individual-level computations that translate incoming information from conspecifics into actions. Recently, it has been proposed that discrete behavioral modes, or internal states, may modulate these computations, affecting the resulting collective behaviors. Direct evidence for such hierarchical control remains limited due to challenges in inferring hidden perception-action computations and uncovering discrete behavioral modes from continuous behaviors. To address this, we analyzed swimming behaviors of medaka fish (Oryzias latipes) throughout development. At the group level, medaka exhibit synchronized swimming formations that develop early, emerging around 2 weeks of age and stabilizing within 1 month. Unlike many teleost species that use burst-and-coast swim patterns, medaka exhibit continuous tail and body undulations. We show that this continuous behavior can be segmented into three distinct kinematic states: acceleration, deceleration, and prolonged constant-speed swimming. Using state-dependent computational models, we tested how medaka translate social information from neighbors into actions across these kinematic states. The models revealed distinct computations governing social information processing and movement responses in each state. Moreover, social responsiveness varied significantly between states: it was strongest during constant-speed epochs, intermediate during accelerations, and lowest during decelerations. These findings highlight discrete behavioral modes as key modulators of social interaction computations underlying collective behavior. Copyright © 2026 The Authors. Published by Elsevier Inc. All rights reserved.