Large language models can generate fluent text, but that text tends to be less varied than human writing. This phenomenon, sometimes called model collapse, has been observed in open-ended English generation, with outputs from different models converging and individual models repeating themselves.

A new arXiv preprint asks whether this convergence forms a "multilingual hivemind" — that is, whether LLMs show the same lack of diversity in every language. The authors build on prior evidence of inter-model collapse (different models producing similar text) and intra-model collapse (a single model producing repetitive text) in English.

The abstract does not yet provide specific findings for non-English languages. Instead, it sets up the research question and signals that the analysis will compare generation across languages. For readers, the significance is in the question: if collapse is not uniform across languages, then efforts to measure or mitigate it may need to be language-specific.