Reuse Your Rewards: Reward Model Transfer for Zero-Shot Cross-Lingual Alignment
This work addresses the problem of scaling language model alignment to diverse languages for researchers and practitioners, offering a practical solution with incremental improvements in cross-lingual transfer.
The paper tackles the challenge of aligning language models across languages without multilingual preference data by training a reward model on one language and applying it zero-shot to others, achieving human preference over unaligned models in over 70% of cases for tasks like summarization and dialog generation.
Aligning language models (LMs) based on human-annotated preference data is a crucial step in obtaining practical and performant LM-based systems. However, multilingual human preference data are difficult to obtain at scale, making it challenging to extend this framework to diverse languages. In this work, we evaluate a simple approach for zero-shot cross-lingual alignment, where a reward model is trained on preference data in one source language and directly applied to other target languages. On summarization and open-ended dialog generation, we show that this method is consistently successful under comprehensive evaluation settings, including human evaluation: cross-lingually aligned models are preferred by humans over unaligned models on up to >70% of evaluation instances. We moreover find that a different-language reward model sometimes yields better aligned models than a same-language reward model. We also identify best practices when there is no language-specific data for even supervised finetuning, another component in alignment.