The performance of machine learning models largely depends on the choice of the loss function. Traditional loss functions perform well in many tasks but may be limited when handling complex data distributions or noisy environments. To address this, this paper proposes an ensemble algorithm based on the t-tilted loss. The t-tilted loss adjusts its values to generate a series of decision makers with different preferences and decision boundaries, which are then effectively combined to enhance a model’s generalization ability in noisy datasets. The experiments cover both regression and classification datasets, incorporating different noise ratios to simulate real-world noise conditions. Experimental results show that the ensemble algorithm based on t-tilted loss significantly improves model accuracy in noisy environments and outperforms traditional ensemble methods. Therefore, this study offers new insights into noise-robust learning and provides a theoretical foundation for further improving ensemble algorithms.