<p>Existing landslide temporal prediction models are predominantly either purely data driven or purely physics driven, each with inherent limitations. In this study, a physics–data hybrid-driven framework is developed for predicting evolution states of reservoir colluvial landslides exhibiting step-like movement patterns. The framework integrates a physics-driven pore pressure–stress coupling model calibrated via Bayesian parameter inversion, data-driven classifiers (random forest (RF) and support vector classification (SVM)) optimized using adaptive inertia weight particle swarm optimization (AIW-PSO), and a parallel coupling strategy that fuses physical and monitoring-based outputs to construct feature-enhanced datasets for an extreme gradient boosting predictor. Experimental results from the Dafengwan landslide in the Xiluodu Reservoir area, China, reveal that the hybrid model achieved strong predictive performance, with an AUC of 0.95, an F1 score of 0.95, and a balanced accuracy (BACC) of 0.91. Compared with data-driven models, improvements in MCC and Kappa reach 91.90% and 86.84%, respectively, while improvements relative to the physics-driven model are 40.82% (Kappa) and 24.56% (MCC). SHAP analysis indicates balanced contributions from physics-driven and data-driven components, while moving-block bootstrap analysis confirms model stability under temporal resampling. An additional validation at a secondary monitoring point demonstrates within-landslide applicability under varying deformation magnitudes. The proposed framework effectively captures abrupt displacement fluctuations under periodic rainfall and reservoir level fluctuations and provides an interpretable decision-support tool for landslide evolution-state assessment and early warning in reservoir-affected regions.</p>