Large language models are increasingly capable of understanding dialectal English, yet they predominantly generate in standard, US-leaning English, leaving dialectal generation largely unaddressed. We introduce DiaLLM, which continually pretrains three open-weight language model families on the International Corpus of English and applies both implicit and explicit post-training paradigms, combined with three model alignment strategies, providing the first controlled comparison across Australian, Indian, and Northern British English.
Our findings reveal that dialectal robustness and generation are dissociated: benchmarks are influenced by continual pretraining and SFT, while alignment reshapes generation in ways not captured by benchmarks. Explicit variety-targeted adaptation reliably produces output recognized as dialectal and preferred over broad alignment; however, the method that most aggressively optimizes dialectal rewards is not favored by human evaluators. Independent linguistic analysis corroborates this reward-quality gap, particularly evident in two of the three model families. No single alignment method dominates, and closing this gap will require richer reward designs and ongoing investment in dialectal resources. We release all code, checkpoints, and preference datasets.
Blogger's Review: The DiaLLM study highlights the intricate relationship between dialect understanding and generation, emphasizing the importance of continual pretraining and the diversity of alignment strategies. Future research must delve deeper into reward design to achieve better dialect generation outcomes.