A team from arXiv proposes a post-processing framework that uses model ensembling to reconcile predictive performance with fairness in machine learning. The method is designed to operate independently of any specific model's internal structure, making it applicable to a wide range of learning tasks, model architectures, and fairness definitions.

In extensive experiments covering classification, regression, and survival analysis, the framework consistently improved fairness metrics while preserving predictive accuracy, or degrading it only slightly. This addresses a long-standing trade-off in fairness-aware machine learning, offering a flexible tool that can be layered onto existing models without retraining.

The authors emphasize the framework's generality, positioning it as a practical post-processing solution for practitioners who need to enforce fairness constraints across diverse real-world applications. The paper is available on arXiv under identifier 2603.18838.