CRPO Enables Cross-Lingual Alignment via English Preference Knowledge
August 24, 2026
Cross-lingual Ranking Preference Optimization (CRPO) leverages English preference data to align LLMs in target languages. It uses a hierarchical structure to jointly optimize intra- and inter-lingual preferences, improving overall output quality across multiple languages.
HOW THIS AFFECTS YOU
●
researcherYou can improve multilingual model performance without needing massive amounts of high-quality preference data in every target language.