Abstract: To enhance the accuracy and generalizability of intelligent slope stability assessment, a hybrid CNN-LSTM model integrating Bayesian optimization and a multi-head attention mechanism（BCLM） is proposed. The model takes six key parameters——slope height, slope angle, cohesion, unit weight, internal friction angle, and pore water pressure ratio——as inputs. A convolutional neural network extracts spatial features, a long short-term memory network captures sequential dependencies, and a multi-head attention mechanism dynamically focuses on the most critical parameters. Based on 353 initial slope cases, outliers are removed using the 1.5 × IQR criterion in conjunction with geotechnical domain knowledge, followed by robust scaling to yield a balanced dataset of 330 high-quality samples. Internal testing demonstrates that BCLM achieves an accuracy of 95.45% and an AUC of 0.953 2, representing improvements of 8.60% and 7.44%, respectively, over the baseline CNN-LSTM model, and also outperforms random forest（90.91%） and artificial neural network（86.36%）. Both feature sensitivity analysis and SHAP evaluation consistently reveal that the model is most sensitive to unit weight, followed in descending order by cohesion, internal friction angle, slope height, slope angle, and pore water pressure ratio——a ranking that aligns closely with established geomechanical principles. On an independent external test set of 66 cases, the model attains an accuracy of 89.39% and an AUC of 0.921 9, confirming its robust generalization capability. The findings indicate that BCLM combines high predictive accuracy, strong generalization, and physical interpretability, thereby offering an effective intelligent approach for slope stability assessment.