Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

DeepTrans: Deep Reasoning Translation via Reinforcement Learning

Created by
  • Haebom

Author

Jiaan Wang, Fandong Meng, Jie Zhou

Outline

This paper presents DeepTrans, a free translation model using deep inference LLMs (e.g., OpenAI o1 and DeepSeek-R1). Pointing out that free translation is understudied in existing deep inference LLMs, we introduce DeepTrans, which learns free translation through reinforcement learning (RL). Using predefined evaluation criteria for both translation results and thought processes, we build a reward model that allows DeepTrans to learn how to reason and translate freely. Furthermore, it eliminates the need for labeled translation data, avoiding the labor-intensive and resource-intensive task of data generation. Experimental results show that DeepTrans, based on Qwen2.5-7B, improves literary translation performance by 16.3%, outperforming existing powerful deep inference LLMs. We also summarize failures and interesting findings from the RL exploration process.

Takeaways, Limitations

•
Takeaways:
◦
We present a novel approach to learning free translation without labels using reinforcement learning.
◦
Achieves improved free translation performance over existing deep inference LLMs.
◦
Shows improved performance in specialized fields such as literary translation.
◦
Presenting new possibilities in free translation research.
•
Limitations:
◦
Further research is needed on the generalization performance of the compensation model presented in this paper.
◦
Additional performance evaluations for various language pairs are needed.
◦
Research is needed to improve the efficiency of RL training.
◦
Lack of specific analysis of failure cases and interesting findings.
👍
Made with Slashpage