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

2026-03-13

AI Summary

This work addresses the limitations of low-resource machine translation and the reliance of existing reinforcement learning approaches on high-quality parallel corpora, which are prone to reward model deficiencies that degrade multilingual performance. The authors propose WALAR, a method that leverages only monolingual data for reinforcement training. By identifying and rectifying failure modes in source-side multilingual quality estimation models and integrating word- and language-alignment techniques to calibrate reward signals, WALAR effectively mitigates reward hacking. Evaluated across all 1,400 language directions in Flores-101, the method significantly outperforms LLaMAX—the current strongest open-source multilingual large language model—improving translation quality for all 101 low-resource languages while maintaining competitive performance on high-resource languages.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Multiagent Systems: Multiagent Learning

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Web Mining and Content Analysis: Large pretrained models with web data

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.

Problem

• reward hacking
• reinforcement learning
• multilingual translation
• quality estimation
• low-resource languages

Innovation

• reward hacking
• reinforcement learning
• monolingual data
• quality estimation
• word alignment