PROMPT-TIME CROSS-LINGUAL TRANSFER FOR LOW-RESOURCE LANGUAGES

Large language models (LLMs) can translate low-resource (LR) languages from a prompt alone, but whether this reflects genuine cross-lingual transfer or surface-cuedriven task recognition remains unclear. Using a 10-language subset of the LoResLM “Rosetta Stone” benchmark in a controlled regime where in-context demonstrations are high-resource (HR)!EN pairs and the LR language appears only in the final query, I evaluate four multilingual LLMs across six prompt conditions. Two findings qualify the linguistically diverse prompting recommendation: a single well-chosen close donor matches or outperforms typologically diverse mixtures, and close donors beat far donors in 82.5% of language–model cells (+5:0 chrF++). A mixed-effects regression over 12,320 observations identifies character trigram overlap—not genealogical distance—as the strongest predictor of translation quality. The results support interpreting prompt-time “transfer” as task location and surface-form activation. Practical guideline: prefer donors with high character level overlap over typologically diverse ones.