<p>Human activity recognition (HAR) is a study that uses machine learning to recognize human activities. Deep learning (DL) models are frequently employed for HAR, but they are often susceptible to catastrophic forgetting, which refers to the severe performance degradation on previously learned tasks after training on new data. Continual Fine-tuning aims to prevent this forgetting while adapting the DL model to new tasks. However, most existing Continual Fine-tuning methods primarily focus on updating model parameters (e.g., weight and bias), neglecting the crucial importance of hyperparameter configuration in adapting to evolving data distributions. Recent work coined as “Model Soup” optimizes fine-tuning but it has not yet been applied to Continual Learning pipeline to prevent catastrophic forgetting. To bridge this gap, we propose a novel soup algorithm, C-Soup, which significantly leverages the performance of the proposed method. Intensive experiments were performed on three sensor-based HAR datasets. Results show that the proposed method consistently improves the performance of the DL model for Continual Fine-tuning, provided the base Continual Learning model performs sufficiently well (e.g., accuracy above 40%). Furthermore, the DL model utilizing C-Soup outperforms existing Continual Fine-tuning methods in limited data settings.</p>