WiFi channel state information (CSI) has become a compelling sensing modality for contact-less human activity recognition. However, differences in datasets, preprocessing protocols and model configurations make consistent comparison and reproducibility challenging. This study presents a unified baseline evaluation of four widely adopted deep learning architectures: multilayer perceptron (MLP), convolutional neural network (CNN), gated recurrent unit (GRU) and a hybrid CNN-GRU model across multiple publicly available CSI datasets encompassing a range of sensing tasks. We harmonize the datasets, implement a standardized preprocessing and training pipeline to reduce experimental inconsistencies and support controlled within-dataset comparisons of model behavior. Evaluations include single-person activity recognition, fall-risk estimation, multi-person occupancy classifica-tion and localization-aware activity recognition, representing progressively higher temporal and spatial complexity. Our results show dataset-dependent trends: CNNs provide an efficient accuracy-complexity trade-off in several structured activity scenarios, whereas GRUs are advantageous when temporal dynamics are more prominent, although with greater training and inference costs. In contrast, MLPs generally underperform due to limited capacity to capture spatial and temporal dependencies. Confusion matrix analysis reveals that dynamic behaviors and low-motion states remain challenging to distinguish, underscoring the importance of temporal modeling. By releasing the complete experi-mental pipeline and benchmarking results, this work establishes a reproducible reference framework for the research community and highlights directions for future investigation, including cross-dataset generalization, hybrid model design and lightweight deployment strategies.
