Large language models (LLMs) are unreliable for extremely low-resource languages, particularly when the writing system clashes with the high-resource data used during pretraining. A new preprint on arXiv examines this problem for Bai-Chinese translation in the Bowen script, a case where standard LLM approaches fall short. The authors argue that general-purpose models cannot be applied directly and need language-specific adaptation.

To address this, the paper proposes MAS-Bai, a multi-agent translation system that is both linguistically informed and adaptive. Instead of relying on a single LLM call, the system coordinates multiple agents that each handle part of the translation process, using linguistic knowledge about Bai to guide decisions. This design is intended to overcome the mismatch between the Bowen script and the pretraining priors of large models.

The work is a preprint and has not yet been peer-reviewed, but it highlights a growing concern in natural language processing: scaling up models does not automatically make them useful for the world's low-resource languages. The proposed framework suggests that explicit linguistic structure and adaptive multi-agent workflows may be a more promising direction than brute-force prompting.