This repository hosts open-source work from the Marco Translation Team of ATH-MaaS:
| Project | Description | Venue |
|---|---|---|
| Marco-MT-Algharb | The Algharb large translation model built on Qwen3-14B, our submission to the WMT 2025 General Machine Translation Shared Task. | WMT 2025 |
| M²PO | Multi-Perspective Multi-Pair Preference Optimization for Machine Translation — a preference-optimization framework that goes beyond single-reward training. | ACL 2026 (Main) |
This repository contains the system for Algharb, the submission from the Marco Translation Team of ATH-MaaS to the WMT 2025 General Machine Translation Shared Task. For setup, supported language pairs, prompt format, inference examples, and MBR decoding, see Marco-MT-Algharb/README.md.
The Algharb system is a large translation model built based on the Qwen3-14B foundation. It is designed for high-quality translation across 13 diverse language directions and demonstrates state-of-the-art performance. Our approach is centered on a multi-stage refinement pipeline that systematically enhances translation fluency and faithfulness.
🏆 WMT 2025 Dominant Performance Secured 12 medals (6 🥇, 4 🥈, 2 🥉) across 13 contested language pairs, demonstrating state-of-the-art capabilities.
🚀 Breakthrough in EN→ZH Translation
Achieved Rank #1 in the highly competitive English→Chinese direction, outperforming human translators, GPT-4.1, and Claude-4.
💡 Core Technical Innovations
- A two-stage SFT (Supervised Fine-Tuning) pipeline enhanced with
$CPO/M^2PO$ reinforcement learning. - A hybrid decoding strategy integrating word alignment and MBR (Minimum Bayes Risk).
In the WMT 2025 evaluation, our Marco-MT-Algharb system demonstrated exceptional performance. Notably, in the English-to-Chinese general translation task, our system ranked #1, outperforming leading AI systems like GPT-4.1 and Claude-4.
We open-source M²PO (Multi-Perspective Multi-Pair Preference Optimization), our ACL 2026 Main conference (long) paper on preference optimization for machine translation. Instead of relying on a single reward signal, M²PO leverages multiple preference pairs and multiple quality perspectives to better calibrate translation quality, and is built on top of the ALMA framework.
The full implementation — including data preparation, LoRA training, and vLLM-based evaluation — lives in the MMPO/ subdirectory. See MMPO/README.md for installation, training, and evaluation instructions.
@InProceedings{wang-EtAl:2025:WMT,
author = {Wang, Hao and Xu, Linlong and Liu, Heng and Liu, Yangyang and Zhao, Xiaohu and Zeng, Bo and Wang, Longyue and Luo, Weihua and Zhang, Kaifu},
title = {Marco Large Translation Model at WMT2025: Transforming Translation Capability in LLMs via Quality-Aware Training and Decoding},
booktitle = {Proceedings of the Tenth Conference on Machine Translation (WMT 2025)},
month = {November},
year = {2025},
address = {Suzhou, China},
publisher = {Association for Computational Linguistics},
pages = {587--593},
url = {https://aclanthology.org/2025.wmt-1.33}
}
@inproceedings{wang-etal-2026-m2po,
title = "{M}$^2${PO}: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation",
author = "Wang, Hao and
Xu, Linlong and
Liu, Heng and
Liu, Yangyang and
Zhao, Xiaohu and
Zeng, Bo and
Shao, Liangying and
Dong, Yichen and
Wu, Xinwei and
Zhou, Jiang and
Dong, Tianyu and
Zeng, Xiangxiang and
Wang, Longyue and
Luo, Weihua",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.469/",
doi = "10.18653/v1/2026.acl-long.469",
pages = "10315--10336"
}This model is released under the Apache License 2.0. You can find the full license text 
We used compliance checking algorithms during the training process, to ensure the compliance of the trained model(s) to the best of our ability. Due to complex data and the diversity of language model usage scenarios, we cannot guarantee that the model is completely free of copyright issues or improper content. If you believe anything infringes on your rights or generates improper content, please contact us, and we will promptly address the matter.
