Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yifeng Liu, Siqi Ouyang, Yatish Hosmane Revanasiddappa, Lei Li

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.

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

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.
Paper Structure (33 sections, 9 equations, 14 figures, 7 tables)

This paper contains 33 sections, 9 equations, 14 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Method
  4. Problem Formulation
  5. WALAR Reward
  6. Quality Estimation Score.
  7. RL Training
  8. Experiments
  9. Experimental Setup
  10. Data.
  11. Models and training details.
  12. Evaluation method.
  13. Main Results
  14. WALAR improves language consistency in translation.
  15. Analysis
  16. ...and 18 more sections

Figures (14)

  • Figure 1: Holes of source-based quality estimation metrics. RL training using these metrics will amplify the holes in LLMs.
  • Figure 2: Illustration of WALAR. On each step, the LLM is prompted to translate one monolingual sentence into another language with several different rollouts. Each output will then be evaluated by language alignment, quality estimation, and word alignment. Finally, the LLM is trained using GRPO with the reward on the previous step iteratively.
  • Figure 3: LCR on language directions. WALAR improves LLMs' translation into desired target languages.
  • Figure 4: Human evaluation results on Az-Pt and En-Kn.
  • Figure 5: Cross-lingual generalization on unseen target languages. X denotes languages in Flores-101. LLaMAX3-Alpaca, trained with WALAR, demonstrates strong generalization across unseen languages.
  • ...and 9 more figures