Mending the Holes: Mitigating Reward Hacking in Reinforcement Learning for Multilingual Translation

TL;DR

This paper introduces WALAR, a reinforcement learning method that uses only monolingual data to improve low-resource language translation in large language models, effectively addressing reward hacking issues and outperforming existing models.

Contribution

WALAR is a novel reinforcement learning approach that mitigates reward hacking in multilingual translation models using only monolingual data, enhancing low-resource language performance.

Findings

WALAR outperforms LLaMAX on 1400 language directions.

Mitigates reward hacking in multilingual RL training.

Supports translation for 101 languages.

Abstract

Large Language Models (LLMs) have demonstrated remarkable capability in machine translation on high-resource language pairs, yet their performance on low-resource translation still lags behind. Existing post-training methods rely heavily on high-quality parallel data, which are often scarce or unavailable for low-resource languages. In this paper, we introduce WALAR, a reinforcement training method using only monolingual text to elevate LLMs' translation capabilities on massive low-resource languages while retaining their performance on high-resource languages. Our key insight is based on the observation of failure modes (or "holes") in existing source-based multilingual quality estimation (QE) models. Reinforcement learning (RL) using these QE models tends to amplify such holes, resulting in poorer multilingual LLMs. We develop techniques including word alignment and language alignment to mitigate such holes in WALAR's reward for RL training. We continually trained an LLM supporting translation of 101 languages using WALAR. The experiments show that our new model outperforms LLaMAX, one of the strongest open-source multilingual LLMs by a large margin on 1400 language directions on Flores-101 dataset.