6G networks will enable a wider range of use cases than 5G and operate in far more diverse environments. This heterogeneity challenges the use of machine learning models for prediction and automated decision-making, because models must remain reliable across different deployments, infrastructures, radio conditions, traffic mixes, and user behaviors. Retraining models from scratch for every new scenario is often impractical, costly, and slow. Causality-inspired methods and domain generalization techniques have been proposed to improve robustness under such distribution shifts by focusing learning on features that remain stable across environments. This project will reproduce and extend state-of-the-art domain generalization baselines, with particular emphasis on causality-based approaches, and assess their effectiveness for telecom-relevant prediction tasks in emerging 6G settings. The outcome will be a clearer understanding of which methods generalize best under realistic network shifts.