[HUGGINGFACE]score: 0.63
Cross-Lingual Preference Tuning on English-Only Reward Models Transfers Across 14 Languages
May 24, 2026
CroCo shows that contrastive preference tuning using self-generated responses ranked by an English-trained reward model transfers to 14 high- and low-resource languages without language-specific annotation, improving over baselines while avoiding catastrophic forgetting.
paper
HOW THIS AFFECTS YOU
●
builderYou can reduce multilingual RLHF annotation costs by using an English reward model to generate cross-lingual preference pairs, with demonstrated gains across most of 14 languages.
●
researcherThe finding that English reward model rankings generalize usefully across 14 languages challenges the assumption that multilingual preference tuning requires per-language annotation.