This study examined profiles of change in repeated mother-child interactions over the course of a 12 week treatment period for childhood aggression. The aim of this study was to investigate whether it was possible to detect the characteristic profile of change, typical for phase transitions, over the course of treatment, and whether this profile was associated with positive treatment outcomes. Entropy values were computed for six repeated real-time observations of each mother-child dyad, using a novel application of recurrence quantification analysis for categorical time series. Subsequent latent class growth curve analysis on the sequences of entropy values revealed two distinct classes of dyads, with one class showing a clear peak in entropy over the six measurement points. The latent class membership variables showed a significant systematic relationship with observed dyad improvement (as rated by clinicians). The class with the peak in entropy over the sessions consisted largely of treatment improvers. Further analysis revealed that improvers and non-improvers could not be distinguished based on content-specific changes (e.g. more positivity or less negativity during the interaction). The present study revealed a treatment-related destabilization pattern in real-time behaviors that was related to better treatment outcomes, and underlines the value of dynamic nonlinear time-series analysis (especially RQA) in the study of dyadic interactions in clinical contexts.