Animal models can provide valuable insights into the mechanisms underlying stress-related disorders. Yet, significant translational challenges persist, as laboratory behavioral assays are often reductionistic, with limited attention to ethologically relevant behavioral diversity. Recent advances in high-throughput pose-estimation tools and computational ethology methods are addressing this limitation by enhancing the resolution and validity of behavioral phenotyping. In this context, it is known that early life stress (ELS) reshapes how animals handle subsequent threats later in life, but the fine-scale dynamics and ethological details of this shift remain elusive. To overcome this, we combined a deep-learning pose-estimation pipeline (DeepLabCut) with a supervised freezing classifier (SimBA) and an unsupervised behavioral motifs identification platform (keypoint MoSeq) to study in detail the diversity and dynamics of behavior in an auditory fear-conditioning (FC) paradigm in two independent cohorts of adult male mice that were exposed to ELS through the limited bedding and nesting (LBN) paradigm. We first validated the blunted freezing response after ELS in a supervised manner using SimBA. Next, keypoint MoSeq segmented the same pose-estimation data into ethologically meaningful motifs over time. When compared to control animals, ELS offspring showed an altered FC response, reduced behavioral entropy and limited diversity in their behavioral repertoire. Such response was characterized by longer active-behavior bouts and more recurrent transitions between states, indicating a more stereotyped and predictable response. Multidimensional scaling of time-binned behavioral vectors and distance metrics identified a resilient subpopulation within the ELS group that displayed a control-like behavioral profile, characterized by a steeper increase in freezing behavior during the FC task and a more diverse behavioral repertoire with reduced recurrence of stereotyped actions, less frequent and shorter active bouts and prolonged passive responses. Overall, our findings suggest that ELS shifts the balance between passive and active coping strategies and that resilience is marked by a less stereotypical yet more diverse and flexible behavioral response to a subsequent stressful demand. Finally, we further validated the unsupervised behavioral motifs with a predictive model that identified distinctive kinematic features of these responses, which could be used to build new behavioral classifiers that can be applied in other behavioral paradigms. These results demonstrate the potential of computational ethology to dissect complex behavioral patterns and improve our understanding of individual stress responses. By combining supervised and unsupervised behavioral analysis tools, we can deepen our understanding of the latent structure of stress behavior and identify objective markers of vulnerability and resilience.