WiFi sensing has shown great promise in applications such as activity recognition, human identification, and health monitoring. However, models trained on Channel State Information (CSI) data often suffer from poor generalizability due to high variance and inconsistent reporting practices, especially in small-data regimes. Despite increasing attention to deep learning-based approaches, the community lacks standardized guidelines on handling variance and overfitting across architectures and datasets. In this work, we first review existing robustness strategies employed in related fields to identify techniques suitable for WiFi sensing. Based on this survey, we systematically benchmark five representative methods—stability training, mixup, coupled weight decay, early stopping, and label smoothing—selected for their theoretical grounding and prior success in mitigating variance. Our evaluation spans three model types (LeNet, LSTM, CNN+GRU) and two CSI datasets (Widar, NTU-Fi), using stratified 5-fold cross-validation with repeated trials to ensure reliable variance estimation. Results show that (1) stability training consistently improves moderately performing models, (2) coupled weight decay is especially effective for LSTMs, and (3) combining techniques can harm performance in near-saturated scenarios. Our findings offer actionable, architectureand dataset-specific guidelines for improving robustness and reproducibility in WiFi sensing research.