Are Reasoning and Refining Really Helpful for LLM-based Machine Translation in Low-Resource Scenarios?
Published:
Please cite:
@INPROCEEDINGS{wang2026,
author={Dezhi Wang, Mieradilijiang Maimaiti, Yi Chen, Shibo Zhang, Zhengyi Guo, Zhexin Li, and Wushouer Silamu},
booktitle={The 22nd China Conference on Machine Translation and Multilingual Technology (CCMT, 2026)},
title={Are Reasoning and Refining Really Helpful for LLM-based Machine Translation in Low-Resource Scenarios?},
year={2026},
volume={},
number={},
pages={},
keywords={Large Language Models; Machine Translation; Low-Resource Languages; Chain-of-Thought; Quality Estimation},
doi={},
}
Abstract
Large language models (LLMs) perform well on high-resource languages but struggle with low-resource ones. Direct prompting often produces translations in a single step, without leveraging relevant linguistic information or reasoning. We propose KnT (Knowledge-to-Translation), a refined chain-of-thought (CoT) approach in which the model first generates translation-relevant linguistic knowledge, including a demonstration, topic information, a dictionary chain, and an explanation of a difficult word, then incorporates it into different prompts to guide translation. Multiple candidates are generated, with the best selected via a reference-free quality estimator. Experiments on FLORES-200, IWSLT, and CCMatrix, among others, show consistent improvements over strong baselines on common translation metrics such as XCOMET-XL and BLEURT-20, demonstrating that reasoning over extracted linguistic knowledge effectively improves low-resource translation. The code and data are publicly available. https://github.com/kyriezhizhi-ops/knt
[PDF] [Code] [Poster] [Slides] [Best Paper Award]
