Large language models increasingly understand dialectal English, yet still produce only standard, US-leaning English, leaving dialectal generation, the harder half of the problem, largely unaddressed. We introduce DiaLLM, which continually pretrains three open-weight language model families on the International Corpus of English and applies implicit and explicit post-training paradigms, each combined with three model alignment strategies, giving the first controlled comparison of these components across Australian, Indian, and Northern British English. Our results reveal a robustness–generation gap: benchmarks are shaped by continual pretraining and SFT, while alignment visibly reshapes generation in ways benchmarks do not capture. Explicit variety-targeted adaptation produces output reliably recognised as dialectal and judged more dialectal than broad alignment, yet where human judgement was directly assessed, the method that most aggressively optimises the dialectal reward is not the one judged most dialectal. Independent linguistic analysis corroborates this reward-quality gap, most clearly on two of the three families. No single alignment method dominates, and closing the gap will require richer reward designs and continued investment in dialectal resources. We release all code, checkpoints, and preference datasets.

Parameter-efficient fine-tuning with adapters (e.g., LoRA) equips LLMs with task-specific skills. However, utilizing multiple pretrained adapters for out-of-distribution (OOD) generalization remains challenging. Existing techniques for OOD generalization using multiple pretrained LoRAs, route inputs using LoRA representations (prototypes) obtained independently, assuming these representations capture complementary information. However, we observe that for existing methods, in-distribution and OOD routing entropies are often comparable, thus bringing the complementarity assumption into question. We derive the theoretical conditions that could lead to a violation of such assumptions, distilling the cause down to the presence of shared, noisy prototype subspaces. Based on this, we introduce MAPLE (Masked-Adapter Prototype LEarning), a simple learning framework that refines LoRA prototypes by masking the target task’s LoRA during prototype learning. In doing so, it encourages prototypes to discard noisy attributes, improving prototype discriminability, routing, and OOD generalization. Extensive experiments on language models of varying size, such as Phi-2 (2.7B) and LLaMA-3 (8B) equipped with heterogeneous pools of pretrained LoRAs, show that MAPLE improves the LoRA representation and thus achieves state-of-the-art performance across multiple benchmarks.