Different Perturbations, Different Mechanisms: Understanding Continued Pre-training for Zero-Shot Dialect Robustness

Dialectal variation is a known weakness of multilingual LLMs. Perturbation-based continued pre-training (CPT) is a promising fix, but prior work lacked systematic comparison. This paper evaluates six training conditions, examining the mechanisms behind different perturbations. It provides insights into why some perturbations transfer to zero-shot dialect robustness better than others. The results could inform future CPT strategies for equitable multilingual AI.