The DiaLLM study addresses the disconnect between understanding and producing dialectal English by continually pretraining three open-weight language model families on the International Corpus of English. It applies implicit and explicit post-training paradigms combined with three alignment strategies to compare Australian, Indian, and Northern British English.
- Dialectal robustness and generation are dissociated; benchmarks reflect continual pretraining and SFT, while alignment reshapes generation in ways benchmarks miss.
- Explicit variety-targeted adaptation produces output reliably recognized as dialectal and preferred over broad alignment.
- The method optimizing the dialectal reward most aggressively is not preferred by human evaluators.
- Independent linguistic analysis corroborates this reward-quality gap across two of the three model families.
The authors conclude that no single alignment method dominates and closing the gap requires richer reward designs and continued investment in dialectal resources.