Findings of the WMT25 General Machine Translation Shared Task

This technical report presents the findings of the General Machine Translation Task conducted during the 2025 Conference on Machine Translation (WMT25). The task invited participants to develop translation systems for 30 language pairs, with a focus on creating challenging test sets through a new difficulty sampling technique. A total of 60 systems were evaluated, including those from large language models and popular online translation providers. The evaluation process utilized professional annotators and various metrics, including the Error Span Annotation protocol and Multidimensional Quality Metrics. The report discusses the growth in participant submissions, the biases observed in automatic scoring compared to human evaluations, and the challenges faced in translating different domains, particularly in the speech domain. Additionally, it outlines the data collection methods and the strategies employed to ensure the quality and difficulty of the test sets, emphasizing the importance of document-level translation assessments.