Character-Noised CPT Improves Multilingual Dialect Robustness in LLMs

Character-noised continued pre-training (CPT) improves zero-shot dialect robustness across German, Italian, and Arabic without degrading performance on standard varieties. Analysis shows different perturbation methods achieve robustness through distinct representational alignment and prediction repair mechanisms.