Recent strategies for low-resource machine translation rely on LLMs to generate synthetic data based on text in higher-resource languages. We revisit this idea for Romansh, a language with 6 distinct varieties. LLMs tend to confuse these varieties when translating into Romansh, but they are quite good at translating out of Romansh into a high-resource language such as German. Due to this asymmetry, the direction of data augmentation is a crucial choice. We find that contrary to recent strategies, creating synthetic translations into the higher-resource language is the superior approach, and only this approach allows us to surpass a Gemini 3 Pro baseline on German-Romansh translation (+23 BLEU over Gemini in the lowest-resource variety). A human evaluation confirms that our experiments yield the first model that generates fluent translations in the individual Romansh varieties.

Large Language Models (LLMs) excel at translating 16th-century letters from Latin and Early New High German to modern English and German. While they perform well at translating well-known historical city names (e.g., Lutetia –> Paris), their ability to handle person names (e.g., Theodor Bibliander) or lesser-known toponyms (e.g., Augusta Vindelicorum –> Augsburg) remains unclear. This study investigates LLM-based translations of person and place names across various frequency bands in a corpus of 16th-century letters. Our results show that LLMs struggle with person names, achieving accuracies around 60%, but perform better with place names, reaching accuracies around 90%. We further demonstrate that including a translation suggestion for the proper noun in the prompt substantially boosts accuracy, yielding highly reliable results.