Abstract: Landslide susceptibility assessment is crucial for the disaster mitigation in mountainous regions; however, traditional models often have limited generalization on small-sample datasets, and linear approaches struggle to capture the complex interactions among influencing factors. This study introduces the TabPFN machine learning model, which, with its hyperparameter-free nature and strong generalization advantage under small-sample conditions, provides a novel approach for overcoming the modeling bottlenecks. Taking Huangshan City as the study area, based on 736 landslide points, an equal number of non-landslide samples were selected from areas beyond 1 km of these points to construct a balanced dataset with a 1:1 ratio of positive to negative samples,this research establishes a landslide susceptibility indicator system by systematically selecting 14 evaluation factors across four categories: topography, geology, hydrology, and human activity. Building upon this foundation,The study introduces the TabPFN model for landslide susceptibility evaluation, and employs SHAP to analyze the underlying driving mechanisms. The results show that, compared with baseline models, TabPFN achieves the best overall performance in terms of accuracy (0.903), precision (0.965), recall (0.841), F1-score (0.899), and AUC (0.964), demonstrating stronger generalization capability and advantages in modeling complex relationships. Furthermore, SHAP analysis reveals that distance to roads, NDVI, and lithology are the primary driving factors influencing the spatial distribution of landslides, reflecting the fundamental pattern wherein human disturbance, vegetation cover, and geological conditions collectively govern landslide development, Specifically, landslide susceptibility increases significantly in areas proximal to roads, reflecting the prominent disturbance of human engineering activities on slope stability. Higher vegetation coverage exerts a notable inhibitory effect on landslide occurrence, while landslide sensitivity varies markedly across different lithological conditions, underscoring the fundamental controlling role of the geological setting. This study integrates the TabPFN machine learning model with SHAP-based interpretability to achieve both high-accuracy prediction and interpretability analysis of landslide susceptibility, It overcomes the shortcomings of traditional "black box" models in revealing the mechanisms of disaster formation, providing a technical pathway for landslide risk assessment in mountainous cities that ensures both precision and explanatory insight.