BARTScore

BARTScore reframes evaluation as a text-generation problem: it feeds one text into a pretrained sequence-to-sequence model (BART) and scores the other as the model's log-likelihood of generating it, so more probable output scores higher (Yuan et al., 2021). Unlike BERTScore and MoverScore, which match embeddings, this is a model-scoring approach, and changing the conditioning direction (source→output, output→reference, reference→output) lets the same metric target faithfulness, precision, or recall rather than a single overlap number.