VEPO: Variable Entropy Policy Optimization for Low-Resource Language Foundation Models

TL;DR

VEPO enhances translation quality and tokenization efficiency for low-resource languages using reinforcement learning with verifiable rewards.

cs.CL 🔴 Advanced 2026-03-20 46 views
Chonghan Liu Yimin Du Qi An Xin He Cunqi Zhai Fei Tan Weijia Lin Xiaochun Gong Yongchao Deng Shousheng Jia Xiangzheng Zhang
reinforcement learning low-resource languages translation tokenization efficiency multilingual models

Key Findings

Methodology

This paper introduces Variable Entropy Policy Optimization (VEPO), a novel method aimed at improving translation and tokenization efficiency in low-resource language models. VEPO leverages reinforcement learning with verifiable rewards to incorporate deterministic structural constraints directly into the policy alignment process. Central to this approach is a variable entropy mechanism that allows the model to dynamically calibrate the balance between literal fidelity and semantic naturalness by modulating the exploration-exploitation trade-off. By integrating entropy-tempered advantage estimation with asymmetric clipping, VEPO maintains robust exploration while reducing the risk of policy collapse.

Key Results

  • Empirical evaluations across 90 directions in FLORES-200, COMET-22, and chrF demonstrate that VEPO significantly improves both tokenization efficiency and translation quality. Compared to existing methods, VEPO achieves a 24.9% increase in BLEU scores for low-resource language translation tasks, narrowing the performance gap with high-resource languages.
  • Through experiments on multilingual datasets, VEPO significantly reduces redundant generation and language drift while maintaining translation quality. Notably, VEPO outperforms existing commercial systems in translations of Southeast Asian languages.
  • Ablation studies show that VEPO's variable entropy mechanism performs well under different KL divergence configurations, particularly in the unconstrained No-KL regime, effectively preventing policy collapse and maintaining stable entropy levels.

Significance

The introduction of VEPO holds significant implications for both academia and industry. It addresses long-standing pain points in translation and tokenization efficiency for low-resource languages and provides new insights for developing multilingual models. By incorporating a variable entropy mechanism and verifiable rewards, VEPO enhances translation naturalness without sacrificing semantic fidelity. This method lays a solid foundation for future research in multilingual models, especially in resource-scarce language environments.

Technical Contribution

VEPO offers several technical breakthroughs. Firstly, it introduces a variable entropy mechanism that allows the model to dynamically balance exploration and exploitation. Secondly, VEPO incorporates verifiable rewards to enforce structural constraints directly within the optimization process, ensuring training stability. Additionally, VEPO's entropy-tempered advantage estimation and asymmetric clipping techniques open up new engineering possibilities, particularly in low-resource language translation tasks.

Novelty

VEPO is the first method to introduce variable entropy policy optimization in low-resource language models. Compared to existing multilingual models, VEPO not only significantly improves translation quality but also excels in tokenization efficiency and training stability. Its innovation lies in combining verifiable rewards with entropy modulation, providing a new perspective on policy optimization.

Limitations

  • VEPO may still face instability issues when dealing with extremely low-resource languages due to data scarcity. While the verifiable rewards mechanism alleviates this to some extent, model performance may still be affected in extreme cases.
  • In high-resource language translation tasks, VEPO's performance improvements are less pronounced than in low-resource languages, indicating that VEPO's advantages are primarily in resource-scarce scenarios.
  • VEPO's computational complexity is relatively high, especially during training on large-scale multilingual datasets, which may require more computational resources.

Future Work

Future research directions include further optimizing VEPO's reward model to enhance high-fidelity translation evaluation. Additionally, exploring advanced reinforcement learning methodologies to better handle linguistic diversity is crucial. The principles of dynamic entropy modulation and verifiable alignment offer a promising foundation for building more robust, inclusive, and expressive multilingual models.

AI Executive Summary

Translation of low-resource languages has long been a challenge in the field of natural language processing. Traditional large language models often underperform in these languages due to inefficient tokenization and imbalanced training data. Existing methods, while excelling in high-resource languages, struggle to achieve similar results in low-resource contexts.

To address this issue, this paper introduces a novel method called Variable Entropy Policy Optimization (VEPO). VEPO incorporates reinforcement learning with verifiable rewards to embed deterministic structural constraints directly into the policy alignment process. At its core is a variable entropy mechanism that allows the model to dynamically calibrate the balance between literal fidelity and semantic naturalness. This mechanism ensures model stability during training by modulating the exploration-exploitation trade-off.

VEPO's technical principles include entropy-tempered advantage estimation and asymmetric clipping techniques. These combined techniques enable VEPO to maintain robust exploration while reducing the risk of policy collapse. Empirical evaluations across 90 directions in FLORES-200, COMET-22, and chrF demonstrate that VEPO significantly improves both tokenization efficiency and translation quality.

Experimental results show that VEPO achieves a 24.9% increase in BLEU scores for low-resource language translation tasks, narrowing the performance gap with high-resource languages. Additionally, VEPO significantly reduces redundant generation and language drift in multilingual datasets, outperforming existing commercial systems, particularly in translations of Southeast Asian languages.

The introduction of VEPO has garnered significant attention in academia and provides new solutions for the industry. By incorporating a variable entropy mechanism and verifiable rewards, VEPO enhances translation naturalness without sacrificing semantic fidelity. This method lays a solid foundation for future research in multilingual models, especially in resource-scarce language environments.

However, VEPO also has its limitations. It may still face instability issues when dealing with extremely low-resource languages due to data scarcity. Additionally, VEPO's computational complexity is relatively high, which may require more computational resources. Future research directions include further optimizing VEPO's reward model to enhance high-fidelity translation evaluation and exploring advanced reinforcement learning methodologies.

Deep Analysis

Background

In recent years, with the development of deep learning technologies, large language models have made significant progress in the field of natural language processing. However, these models still underperform in low-resource languages. Low-resource languages often face challenges such as data scarcity, inefficient tokenization, and model instability. Existing multilingual models, such as GPT-4 and Qwen-max, while excelling in high-resource languages, struggle to achieve similar results in low-resource contexts. To bridge this gap, researchers have attempted to improve translation quality for low-resource languages through data augmentation and specialized model architectures. However, these methods often require substantial computational resources and lack flexibility in practical applications. Therefore, effectively enhancing language model performance in low-resource environments remains a pressing issue.

Core Problem

The core problem for low-resource language models is how to improve translation quality and tokenization efficiency in the face of data scarcity. Traditional tokenization methods, when dealing with morphologically complex languages, often lead to sequence fragmentation, affecting the model's translation performance. Additionally, existing reinforcement learning methods in low-resource environments frequently encounter entropy decay and verbosity issues. These problems not only impact translation quality but also increase training instability. Therefore, achieving efficient policy optimization in low-resource environments has become a critical research question.

Innovation

The proposed Variable Entropy Policy Optimization (VEPO) method has several key innovations:


  • �� Introduction of a variable entropy mechanism: By dynamically adjusting the balance between exploration and exploitation, the model can dynamically calibrate between literal fidelity and semantic naturalness. This mechanism effectively reduces the risk of policy collapse.

  • �� Verifiable rewards mechanism: By incorporating deterministic structural constraints directly into the policy alignment process, it ensures training stability. This mechanism significantly improves translation quality in low-resource environments.

  • �� Entropy-tempered advantage estimation and asymmetric clipping techniques: By combining these techniques, VEPO maintains robust exploration while reducing redundant generation and language drift.

Methodology

The implementation of VEPO involves several key steps:


  • �� Vocabulary expansion: By introducing dedicated tokens, it optimizes tokenization efficiency for low-resource languages, reducing sequence fragmentation.

  • �� Balanced multilingual training: A 1:1 sampling ratio is employed between English and low-resource corpora to ensure model stability in multilingual environments.

  • �� Supervised fine-tuning: Fine-tuning is conducted on high-quality bilingual translation data and instruction-following datasets to enhance translation quality and instruction-following capabilities.

  • �� Variable entropy policy optimization: Entropy-aware reinforcement learning is applied to achieve precise policy alignment while maintaining stylistic flexibility.

Experiments

The experimental design includes evaluations across 90 directions in FLORES-200, COMET-22, and chrF. Baselines include existing multilingual models and dedicated translation systems. Evaluation metrics include BLEU, COMET, and chrF scores. Key hyperparameters include entropy modulation coefficients and clipping thresholds. Ablation studies show that VEPO's variable entropy mechanism performs well under different KL divergence configurations, particularly in the unconstrained No-KL regime, effectively preventing policy collapse and maintaining stable entropy levels.

Results

Experimental results show that VEPO achieves a 24.9% increase in BLEU scores for low-resource language translation tasks, narrowing the performance gap with high-resource languages. Additionally, VEPO significantly reduces redundant generation and language drift in multilingual datasets, outperforming existing commercial systems, particularly in translations of Southeast Asian languages. Ablation studies show that VEPO's variable entropy mechanism performs well under different KL divergence configurations, particularly in the unconstrained No-KL regime, effectively preventing policy collapse and maintaining stable entropy levels.

Applications

VEPO's application scenarios include low-resource language translation tasks, cross-language information retrieval, and multilingual dialogue systems. Its excellent performance in low-resource languages makes it widely applicable in these fields. Particularly in scenarios requiring high translation quality and tokenization efficiency, VEPO can significantly enhance system performance.

Limitations & Outlook

Despite VEPO's excellent performance in low-resource language translation tasks, its performance improvements in high-resource language translation tasks are less pronounced. Additionally, VEPO's computational complexity is relatively high, especially during training on large-scale multilingual datasets, which may require more computational resources. Future research directions include further optimizing VEPO's reward model to enhance high-fidelity translation evaluation and exploring advanced reinforcement learning methodologies.

Plain Language Accessible to non-experts

Imagine you're in a kitchen trying to cook a dish. You have some basic ingredients but lack some key spices. This is like low-resource language models, which have some basic data but lack enough training data to improve translation quality. Traditional methods are like trying to make a dish with the existing ingredients, and the taste might not be great. VEPO's method is like introducing a new cooking technique that cleverly uses the existing ingredients and spices to create a delicious dish. It dynamically adjusts the cooking process's heat and time, making the dish's flavor richer. This method not only improves the dish's quality but also reduces waste. Similarly, in low-resource language models, VEPO improves translation quality and tokenization efficiency by dynamically adjusting the policy optimization process.

ELI14 Explained like you're 14

Hey there! Did you know that when computers try to translate some less common languages, they often mess up? It's like playing a game but not having enough coins to upgrade your gear, so you keep losing to the big boss. Scientists invented a new method called VEPO to help computers translate these languages better. It's like giving you a super treasure chest full of upgrade items, making you unstoppable in the game! VEPO smartly adjusts the computer's translation strategy, just like tweaking your game tactics, making it easier to beat the big boss. So now, even those uncommon languages can be translated well by computers! Isn't that cool?

Glossary

Variable Entropy Policy Optimization

A method that optimizes policy by dynamically adjusting entropy levels to improve translation quality and tokenization efficiency in low-resource language models.

In this paper, VEPO is used to achieve efficient policy optimization in low-resource language environments.

Verifiable Rewards

A reward mechanism used in reinforcement learning that incorporates deterministic structural constraints to enhance training stability.

VEPO utilizes verifiable rewards to ensure structural consistency during policy alignment.

Entropy-Tempered Advantage Estimation

A technique combining entropy modulation and advantage estimation to maintain exploration capability in reinforcement learning.

In VEPO, this technique is used to reduce the risk of policy collapse.

Asymmetric Clipping

A technique used in optimization that asymmetrically limits gradient updates to prevent policy collapse.

VEPO uses asymmetric clipping to maintain stability during training.

FLORES-200

A multilingual translation dataset containing translation tasks in 200 language directions.

Used in this paper to evaluate VEPO's translation performance.

BLEU Score

A metric for evaluating machine translation quality by measuring the similarity between translated text and reference text.

In this paper's experiments, BLEU scores are used to evaluate VEPO's translation quality.

Multilingual Model

A machine learning model capable of handling tasks in multiple languages, typically used for translation and cross-language information retrieval.

This paper discusses the challenges of multilingual models in low-resource language environments.

Sequence Fragmentation

Improper sequence segmentation due to vocabulary mismatch during tokenization, affecting translation quality.

Mentioned in this paper as a tokenization efficiency issue related to sequence fragmentation.

Redundant Generation

The occurrence of unnecessary repetition or excessive information when a model generates text.

VEPO reduces redundant generation through entropy modulation.

Language Drift

Deviation from the target language during translation, leading to inaccurate translations.

VEPO reduces language drift through structural constraints.

Open Questions Unanswered questions from this research

  • 1 How can VEPO's stability be further enhanced in extremely low-resource environments? While the verifiable rewards mechanism alleviates data scarcity issues to some extent, model performance may still be affected in extreme cases. More effective policy optimization methods need to be explored.
  • 2 VEPO's performance improvements in high-resource language translation tasks are less pronounced than in low-resource languages. How can VEPO's performance be further optimized in high-resource environments?
  • 3 VEPO's computational complexity is relatively high, especially during training on large-scale multilingual datasets, which may require more computational resources. How can VEPO's computational complexity be reduced without sacrificing performance?
  • 4 Existing reward models may have biases when evaluating high-fidelity translations. How can reward models be further optimized to improve translation quality evaluation?
  • 5 In the development of multilingual models, how can linguistic diversity be better handled? VEPO provides a solution through dynamic entropy modulation and verifiable alignment, but more advanced methods need to be explored.

Applications

Immediate Applications

Low-Resource Language Translation

VEPO can be used to improve translation quality for low-resource languages, especially in scenarios requiring high translation accuracy, such as legal documents and technical manuals.

Cross-Language Information Retrieval

By improving tokenization efficiency and translation quality, VEPO can be used in cross-language information retrieval systems, helping users quickly find the information they need in multilingual environments.

Multilingual Dialogue Systems

VEPO has wide application potential in multilingual dialogue systems, improving the accuracy and naturalness of system responses and enhancing user experience.

Long-term Vision

Global Language Equality

By enhancing translation capabilities for low-resource languages, VEPO has the potential to promote global language equality in the long term, reducing communication barriers caused by language differences.

Multilingual Education

VEPO can be used in multilingual education systems, helping students better learn and understand the culture and knowledge of different languages, promoting cross-cultural communication.

Abstract

Large language models frequently exhibit suboptimal performance on low resource languages, primarily due to inefficient subword segmentation and systemic training data imbalances. In this paper, we propose Variable Entropy Policy Optimization (VEPO), which leverages Reinforcement Learning with Verifiable Rewards to incorporate deterministic structural constraints into the policy alignment process. This framework ensures prescribed sequence length, robust format consistency, and rigorous linguistic well formedness, all enforced during training. Central to our approach is a variable entropy mechanism that enables the model to dynamically calibrate the equilibrium between literal fidelity and semantic naturalness by modulating the exploration exploitation manifold. By integrating entropy tempered advantage estimation with asymmetric clipping, VEPO sustains robust exploration while mitigating policy collapse. Empirical evaluations across 90 FLORES-200, COMET-22, chrF directions demonstrate that VEPO yields substantial improvements in both tokenization efficiency and translation quality, bridging the performance gap for underrepresented languages.

cs.CL cs.AI

References (20)

DAPO: An Open-Source LLM Reinforcement Learning System at Scale

Qiying Yu, Zheng Zhang, Ruofei Zhu et al.

2025 1419 citations ⭐ Influential View Analysis →

Proximal Policy Optimization Algorithms

John Schulman, Filip Wolski, Prafulla Dhariwal et al.

2017 25938 citations ⭐ Influential View Analysis →

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Zhihong Shao, Peiyi Wang, Qihao Zhu et al.

2024 5086 citations ⭐ Influential View Analysis →

LLaMAX: Scaling Linguistic Horizons of LLM by Enhancing Translation Capabilities Beyond 100 Languages

Yinquan Lu, Wenhao Zhu, Lei Li et al.

2024 61 citations ⭐ Influential View Analysis →

COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task

Ricardo Rei, José G. C. de Souza, Duarte M. Alves et al.

2022 453 citations ⭐ Influential

Measuring Massive Multitask Language Understanding

Dan Hendrycks, Collin Burns, Steven Basart et al.

2020 7362 citations View Analysis →

Advancing Translation Preference Modeling with RLHF: A Step Towards Cost-Effective Solution

Nuo Xu, Jun Zhao, Can Zu et al.

2024 15 citations View Analysis →

Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation

Haoran Xu, Amr Sharaf, Yunmo Chen et al.

2024 426 citations View Analysis →

On the Weaknesses of Reinforcement Learning for Neural Machine Translation

Leshem Choshen, Lior Fox, Zohar Aizenbud et al.

2019 123 citations View Analysis →

A Call for Clarity in Reporting BLEU Scores

Matt Post

2018 3382 citations View Analysis →

X-ALMA: Plug & Play Modules and Adaptive Rejection for Quality Translation at Scale

Haoran Xu, Kenton Murray, Philipp Koehn et al.

2024 36 citations View Analysis →

Minimum Risk Training for Neural Machine Translation

Shiqi Shen, Yong Cheng, Zhongjun He et al.

2015 477 citations View Analysis →

IBM Research Report Bleu: a Method for Automatic Evaluation of Machine Translation

Kishore Papineni, S. Roukos, T. Ward et al.

2001 239 citations

REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization

Jian Hu, Jason Klein Liu, Haotian Xu et al.

2025 21 citations View Analysis →

HellaSwag: Can a Machine Really Finish Your Sentence?

Rowan Zellers, Ari Holtzman, Yonatan Bisk et al.

2019 3831 citations View Analysis →

Multilingual Test-Time Scaling via Initial Thought Transfer

Prasoon Bajpai, Tanmoy Chakraborty

2025 4 citations View Analysis →

Unsupervised Cross-lingual Representation Learning at Scale

Alexis Conneau, Kartikay Khandelwal, Naman Goyal et al.

2019 8020 citations View Analysis →

Beyond English-Centric Multilingual Machine Translation

Angela Fan, Shruti Bhosale, Holger Schwenk et al.

2020 1024 citations View Analysis →

Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them

Mirac Suzgun, Nathan Scales, Nathanael Scharli et al.

2022 1731 citations View Analysis →

Seed-X: Building Strong Multilingual Translation LLM with 7B Parameters

Shanbo Cheng, Yu Bao, Qian Cao et al.

2025 16 citations View Analysis →