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.

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

Created by
  • Haebom

Author

Shanbo Cheng, Yu Bao, Qian Cao, Luyang Huang, Liyan Kang, Zhicheng Liu, Yu Lu, Wenhao Zhu, Jingwen Chen, Zhichao Huang, Tao Li, Yifu Li, Huiying Lin, Sitong Liu, Ningxin Peng, Shuaijie She, Lu Xu, Nuo Xu, Sen Yang, Runsheng Yu, Yiming Yu, Liehao Zou, Hang Li, Lu Lu, Yuxuan Wang, Yonghui Wu

Outline

Seed-X is an open-source large-scale language model (LLM) family with 7 billion parameters, including both directed and inferential models. Pretrained using diverse, high-quality monolingual and bilingual content from 28 languages, it is fine-tuned through Chain of Thought (CoT) inference and then generalized across multiple language pairs using reinforcement learning (RL). It achieves comparable performance across 28 languages to leading closed-loop models such as Gemini-2.5 and GPT-4o, and significantly outperforms larger open-source models in both automated and human evaluation metrics. We share best practices from our optimization process and open-source our parameters to advance translation research and applications.

Takeaways, Limitations

•
Takeaways:
◦
It achieves performance similar to state-of-the-art closed-loop models with a relatively small size of 7 billion parameters, demonstrating the potential for developing lightweight, high-performance multilingual translation models.
◦
It is released as open source and contributes to the development of multilingual translation research and applications.
◦
Improving generalization performance across diverse language pairs using chain-of-thought (CoT) inference and reinforcement learning (RL).
◦
We present an effective pre-training method using a high-quality multilingual dataset that supports various languages.
•
Limitations:
◦
The paper lacks specific references to Limitations or future research directions.
◦
7 billion parameters is still a significant model size, so research into developing models of smaller sizes may be necessary.
◦
Detailed performance analysis, such as performance deviations for specific language pairs or sentence types, may be lacking.
👍
Made with Slashpage