Summary of Cross-lingual Transfer Of Reward Models in Multilingual Alignment, by Jiwoo Hong et al.
Cross-lingual Transfer of Reward Models in Multilingual Alignment
by Jiwoo Hong, Noah Lee, Rodrigo Martínez-Castaño, César Rodríguez, James Thorne
First submitted to arxiv on: 23 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Reinforcement learning with human feedback (RLHF) benefits significantly from precise reward models (RMs), but recent studies in reward modeling schemes are mostly limited to English, hindering RLHF’s applicability in multilingual settings. To address this, we investigate the cross-lingual transfer of RMs trained in diverse languages, primarily from English. Our experimental results show that English RMs can be transferred across languages with an average increase of 3~4% on Multilingual RewardBench compared to target language RMs. We also analyze representation shifts facilitating this cross-lingual transfer. Additionally, we demonstrate the propagation of enhanced multilingual instruction-following capability through multilingual alignment. Our findings emphasize the importance of considering linguistic diversity in RLHF and highlight opportunities for off-the-shelf RM adaptation. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Reward learning with human feedback gets better when using precise reward models, but current studies mostly focus on English, making it hard to apply this technique in different languages. To solve this problem, we tested how well reward models trained in one language can be used in other languages. Our results show that models trained in English can work well in other languages with a 3-4% increase in performance. We also looked at how the way data is represented changes when transferring between languages. Furthermore, we showed how this transfer affects our ability to follow instructions in multiple languages. We released our code, model, and data for others to use. |
Keywords
» Artificial intelligence » Alignment » Reinforcement learning » Rlhf
