谷歌研究指出，单次人工翻译评估容易产生“噪音”，影响模型间的质量对比。为此，团队在MQM框架中引入二次标注环节，即由另一评估员复核已有标注。实验表明，该方法能显著提升评分一致性与可靠性，尤其在人机协作流程中可平衡质量与成本。研究同时提醒，需防范评估者过度依赖初次标注，专家监督仍不可或缺。

"Supported by a growing ecosystem of open-source models, large model technology, just like water and electricity and the Internet, is increasingly becoming a convenient and widely accessible basic ...

In early 2025, the Chinese artificial intelligence (AI) start-up DeepSeek grabbed headlines abroad with models promising high ...