Study Shows Retrieved Examples Boost LLM Translation for Low-Resource Languages

Researchers found that many-shot in-context learning significantly improves machine translation for ten low-resource languages, but smart example selection matters more than quantity. BM25-based retrieval reduced needed examples by 80%, suggesting practical efficiency gains for underrepresented language communities.