Abstract: Slab quality prediction is the key link to optimize steel production process and improve product performance. The traditional prediction model based on metallurgical mechanism mainly relies on physical laws and empirical formulas, which is difficult to adapt to the complex nonlinear relationship in the highly automated production environment, and has inherent limitations such as low prediction accuracy and weak generalization ability. With the rapid development of machine learning technology, the single model has shown good performance advantages in crack risk assessment, segregation prediction and other specific tasks through feature learning and complex nonlinear fitting ability, but it still faces challenges such as high-dimensional data over fitting and unbalanced sample categories. In order to overcome these limitations, the integrated learning model significantly improves the accuracy and robustness of the prediction system through the collaborative optimization of multiple weak learners, especially in the industrial scene dealing with strong coupling of process parameters and strong noise interference. Therefore, the typical modeling methods of traditional quality prediction models were outlined, along with their application limitations. The research progress of single model (e.g., neural networks) and integrated models (e.g., random forests) in slab quality prediction was reviewed. The advantages of integrated quality prediction model in prediction accuracy and engineering application were analyzed. Based on these findings, the future development direction of key technologies were explored, including dynamic modeling optimization, small sample reinforcement learning, lightweight model deployment, interpretability improvement and whole process collaborative prediction are discussed, which provides theoretical support and technical path for the construction of a new generation of slab quality intelligent prediction system with high precision and strong adaptability.