Trans-Zero: Self-Play Incentivizes Large Language Models for Multilingual Translation Without Parallel Data

2025-04-20

AI Summary

To address the scarcity of parallel corpora for low-resource languages and catastrophic forgetting in multilingual neural machine translation, this paper proposes the first unsupervised multilingual translation framework based on genetic Monte Carlo tree search (G-MCTS) and preference optimization via self-play. Leveraging only monolingual data, the method exploits inherent cross-lingual representations in large language models through iterative back-translation and semantic consistency constraints to achieve high-quality translation. Its core innovations include: (i) the first integration of G-MCTS into self-play translation to guide exploration of diverse, semantically coherent translations; and (ii) preference-based reinforcement learning to stabilize training and enhance generalization without parallel supervision. Experiments demonstrate competitive performance against fully supervised baselines on multilingual benchmarks, with an average +2.8 BLEU gain on non-English directions. The framework supports zero-shot translation across 50+ language pairs—entirely without parallel data.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Sampling/Simulation-based Search

Application Category

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

Abstract

The rise of Large Language Models (LLMs) has reshaped machine translation (MT), but multilingual MT still relies heavily on parallel data for supervised fine-tuning (SFT), facing challenges like data scarcity for low-resource languages and catastrophic forgetting. To address these issues, we propose TRANS-ZERO, a self-play framework that leverages only monolingual data and the intrinsic multilingual knowledge of LLM. TRANS-ZERO combines Genetic Monte-Carlo Tree Search (G-MCTS) with preference optimization, achieving strong translation performance that rivals supervised methods. Experiments demonstrate that this approach not only matches the performance of models trained on large-scale parallel data but also excels in non-English translation directions. Further analysis reveals that G-MCTS itself significantly enhances translation quality by exploring semantically consistent candidates through iterative translations, providing a robust foundation for the framework's succuss.

Problem

• Enables multilingual translation without parallel data
• Addresses data scarcity for low-resource languages
• Prevents catastrophic forgetting in large language models

Innovation

• Self-play framework with monolingual data
• Genetic Monte-Carlo Tree Search optimization
• Iterative translation for semantic consistency