A translation can be fluent and still sound as though English grammar is sitting underneath it. Researchers presenting at WMT 2026 examine that problem in a paper introducing MELT, an editing framework for language-model translations. Their tests cover Korean, Chinese, Japanese, German and French. The method checks meaning and style before accepting proposed changes, without retraining the underlying model. The authors report that verification contributes strongly to improvements in their evaluation. Those findings concern the tested datasets and language pairs, not every translation task. The October 28–29 proceedings provide the research context; an exact publication timestamp is not available.
