Evaluation of QWEN-3 for English to Ukrainian Translation

AbstractWe report the results of evaluating Qwen3 for the English-to-Ukrainian language pair of the general MT task of WMT 2025.In addition to the quantitative evaluation, we performed a qualitative evaluation, in collaboration with a native Ukrainian speaker - therefore we present an example-heavy analysis of the typical failures the LLMs still do when translating natural language, particularly into Ukrainian.We report also on the practicalities of using LLMs, such as on the difficulties of making them follow instructions, on ways to exploit the increased “smartness” of the reasoning models while simultaneously avoiding the reasoning part improperly interfering with the chain in which the LLM is just one element.