MAS-Bai lifts BLEU 8.11 on Bowen-script Bai translation

In Bowen-script Bai, an extremely low-resource language, Chinese characters are "frequently borrowed phonetically rather than semantically, causing severe literal-translation hallucinations." Systems that read the characters as Chinese produce confident but wrong output.

Researchers at Tianjin University's TJUNLP Lab propose MAS-Bai, a multi-agent setup that routes a sentence through one of two pipelines, retrieval-translation or a philological pathway staffed by etymological and syntactic agents, depending on what the input requires. The arxiv preprint reports 16.23 BLEU and 62.15 COMET on a held-out test set, a gain of 8.11 BLEU and 6.85 COMET over a Standard RAG baseline.

The dataset is small: 6,536 sentences in total, with 1,536 reserved for testing. The authors flag the strongest gains on "Bowen phonetic loans and divergent word order patterns," the two features that break literal translation hardest. Human evaluation and reverse-verifier analyses accompany the automatic scores. The paper is slated for Findings of EMNLP 2026, and two researchers we follow posted the link within the week.